1 |
Author: tomwij |
2 |
Date: 2014-01-30 16:49:47 +0000 (Thu, 30 Jan 2014) |
3 |
New Revision: 2661 |
4 |
|
5 |
Added: |
6 |
genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-BFQ-v7-3.13.patch |
7 |
genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-BFQ-v7-I-O-sched-for-3.13.patch1 |
8 |
genpatches-2.6/trunk/3.14/5000_BFQ-3-block-bfq-add-Early-Queue-Merge-EQM-to-BFQ-v7-for-3.13.0.patch |
9 |
Removed: |
10 |
genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-v6r2-3.11.patch |
11 |
genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-v6r2-I-O-sched-for-3.11.patch1 |
12 |
genpatches-2.6/trunk/3.14/5000_BFQ-3-block-add-Early-Queue-Merge-EQM-v6r2-for-3.11.0.patch1 |
13 |
genpatches-2.6/trunk/3.14/5000_BFQ-4-block-Switch-from-BFQ-v6r2-for-3.11.0-to-BFQ-v6r2-fo.patch |
14 |
Modified: |
15 |
genpatches-2.6/trunk/3.14/0000_README |
16 |
Log: |
17 |
BFQ v7 3.13. |
18 |
|
19 |
Modified: genpatches-2.6/trunk/3.14/0000_README |
20 |
=================================================================== |
21 |
--- genpatches-2.6/trunk/3.14/0000_README 2014-01-29 14:41:45 UTC (rev 2660) |
22 |
+++ genpatches-2.6/trunk/3.14/0000_README 2014-01-30 16:49:47 UTC (rev 2661) |
23 |
@@ -83,18 +83,14 @@ |
24 |
From: Tom Wijsman <TomWij@g.o> |
25 |
Desc: Add Gentoo Linux support config settings and defaults. |
26 |
|
27 |
-Patch: 5000_BFQ-1-block-cgroups-kconfig-build-bits-for-v6r2-3.11.patch |
28 |
+Patch: 5000_BFQ-1-block-cgroups-kconfig-build-bits-for-v7-3.13.patch |
29 |
From: http://algo.ing.unimo.it/people/paolo/disk_sched/ |
30 |
-Desc: BFQ v6r2 patch 1 for 3.11: Build, cgroups and kconfig bits |
31 |
+Desc: BFQ v7 patch 1 for 3.13: Build, cgroups and kconfig bits |
32 |
|
33 |
-Patch: 5000_BFQ-2-block-introduce-the-v6r2-I-O-sched-for-3.11.patch1 |
34 |
+Patch: 5000_BFQ-2-block-introduce-the-v7-I-O-sched-for-3.13.patch1 |
35 |
From: http://algo.ing.unimo.it/people/paolo/disk_sched/ |
36 |
-Desc: BFQ v6r2 patch 2 for 3.10: BFQ Scheduler |
37 |
+Desc: BFQ v7 patch 2 for 3.13: BFQ Scheduler |
38 |
|
39 |
-Patch: 5000_BFQ-3-block-add-Early-Queue-Merge-EQM-v6r2-for-3.11.0.patch1 |
40 |
+Patch: 5000_BFQ-3-block-add-Early-Queue-Merge-EQM-v7-for-3.13.0.patch |
41 |
From: http://algo.ing.unimo.it/people/paolo/disk_sched/ |
42 |
-Desc: BFQ v6r2 patch 3 for 3.10: Early Queue Merge (EQM) |
43 |
- |
44 |
-Patch: 5000_BFQ-4-block-Switch-from-BFQ-v6r2-for-3.11.0-to-BFQ-v6r2-fo.patch |
45 |
-From: http://algo.ing.unimo.it/people/paolo/disk_sched/ |
46 |
-Desc: BFQ v6r2 for 3.11.0 to BFQ v6r2 for 3.12.0. |
47 |
+Desc: BFQ v7 patch 3 for 3.13: Early Queue Merge (EQM) |
48 |
\ No newline at end of file |
49 |
|
50 |
Added: genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-BFQ-v7-3.13.patch |
51 |
=================================================================== |
52 |
--- genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-BFQ-v7-3.13.patch (rev 0) |
53 |
+++ genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-BFQ-v7-3.13.patch 2014-01-30 16:49:47 UTC (rev 2661) |
54 |
@@ -0,0 +1,104 @@ |
55 |
+From 7f029ed2a02bea57b791c032d6242129c3372a84 Mon Sep 17 00:00:00 2001 |
56 |
+From: Paolo Valente <paolo.valente@×××××××.it> |
57 |
+Date: Tue, 3 Sep 2013 16:50:42 +0200 |
58 |
+Subject: [PATCH 1/3] block: cgroups, kconfig, build bits for BFQ-v7-3.13 |
59 |
+ |
60 |
+Update Kconfig.iosched and do the related Makefile changes to include |
61 |
+kernel configuration options for BFQ. Also add the bfqio controller |
62 |
+to the cgroups subsystem. |
63 |
+ |
64 |
+Signed-off-by: Paolo Valente <paolo.valente@×××××××.it> |
65 |
+Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
66 |
+--- |
67 |
+ block/Kconfig.iosched | 32 ++++++++++++++++++++++++++++++++ |
68 |
+ block/Makefile | 1 + |
69 |
+ include/linux/cgroup_subsys.h | 4 ++++ |
70 |
+ 3 files changed, 37 insertions(+) |
71 |
+ |
72 |
+diff --git a/block/Kconfig.iosched b/block/Kconfig.iosched |
73 |
+index 421bef9..8f552ba 100644 |
74 |
+--- a/block/Kconfig.iosched |
75 |
++++ b/block/Kconfig.iosched |
76 |
+@@ -39,6 +39,27 @@ config CFQ_GROUP_IOSCHED |
77 |
+ ---help--- |
78 |
+ Enable group IO scheduling in CFQ. |
79 |
+ |
80 |
++config IOSCHED_BFQ |
81 |
++ tristate "BFQ I/O scheduler" |
82 |
++ default n |
83 |
++ ---help--- |
84 |
++ The BFQ I/O scheduler tries to distribute bandwidth among |
85 |
++ all processes according to their weights. |
86 |
++ It aims at distributing the bandwidth as desired, independently of |
87 |
++ the disk parameters and with any workload. It also tries to |
88 |
++ guarantee low latency to interactive and soft real-time |
89 |
++ applications. If compiled built-in (saying Y here), BFQ can |
90 |
++ be configured to support hierarchical scheduling. |
91 |
++ |
92 |
++config CGROUP_BFQIO |
93 |
++ bool "BFQ hierarchical scheduling support" |
94 |
++ depends on CGROUPS && IOSCHED_BFQ=y |
95 |
++ default n |
96 |
++ ---help--- |
97 |
++ Enable hierarchical scheduling in BFQ, using the cgroups |
98 |
++ filesystem interface. The name of the subsystem will be |
99 |
++ bfqio. |
100 |
++ |
101 |
+ choice |
102 |
+ prompt "Default I/O scheduler" |
103 |
+ default DEFAULT_CFQ |
104 |
+@@ -52,6 +73,16 @@ choice |
105 |
+ config DEFAULT_CFQ |
106 |
+ bool "CFQ" if IOSCHED_CFQ=y |
107 |
+ |
108 |
++ config DEFAULT_BFQ |
109 |
++ bool "BFQ" if IOSCHED_BFQ=y |
110 |
++ help |
111 |
++ Selects BFQ as the default I/O scheduler which will be |
112 |
++ used by default for all block devices. |
113 |
++ The BFQ I/O scheduler aims at distributing the bandwidth |
114 |
++ as desired, independently of the disk parameters and with |
115 |
++ any workload. It also tries to guarantee low latency to |
116 |
++ interactive and soft real-time applications. |
117 |
++ |
118 |
+ config DEFAULT_NOOP |
119 |
+ bool "No-op" |
120 |
+ |
121 |
+@@ -61,6 +92,7 @@ config DEFAULT_IOSCHED |
122 |
+ string |
123 |
+ default "deadline" if DEFAULT_DEADLINE |
124 |
+ default "cfq" if DEFAULT_CFQ |
125 |
++ default "bfq" if DEFAULT_BFQ |
126 |
+ default "noop" if DEFAULT_NOOP |
127 |
+ |
128 |
+ endmenu |
129 |
+diff --git a/block/Makefile b/block/Makefile |
130 |
+index 20645e8..cbd83fb 100644 |
131 |
+--- a/block/Makefile |
132 |
++++ b/block/Makefile |
133 |
+@@ -16,6 +16,7 @@ obj-$(CONFIG_BLK_DEV_THROTTLING) += blk-throttle.o |
134 |
+ obj-$(CONFIG_IOSCHED_NOOP) += noop-iosched.o |
135 |
+ obj-$(CONFIG_IOSCHED_DEADLINE) += deadline-iosched.o |
136 |
+ obj-$(CONFIG_IOSCHED_CFQ) += cfq-iosched.o |
137 |
++obj-$(CONFIG_IOSCHED_BFQ) += bfq-iosched.o |
138 |
+ |
139 |
+ obj-$(CONFIG_BLOCK_COMPAT) += compat_ioctl.o |
140 |
+ obj-$(CONFIG_BLK_DEV_INTEGRITY) += blk-integrity.o |
141 |
+diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h |
142 |
+index b613ffd..43c5dc9 100644 |
143 |
+--- a/include/linux/cgroup_subsys.h |
144 |
++++ b/include/linux/cgroup_subsys.h |
145 |
+@@ -39,6 +39,10 @@ SUBSYS(net_cls) |
146 |
+ SUBSYS(blkio) |
147 |
+ #endif |
148 |
+ |
149 |
++#if IS_SUBSYS_ENABLED(CONFIG_CGROUP_BFQIO) |
150 |
++SUBSYS(bfqio) |
151 |
++#endif |
152 |
++ |
153 |
+ #if IS_SUBSYS_ENABLED(CONFIG_CGROUP_PERF) |
154 |
+ SUBSYS(perf) |
155 |
+ #endif |
156 |
+-- |
157 |
+1.8.5.2 |
158 |
+ |
159 |
|
160 |
Deleted: genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-v6r2-3.11.patch |
161 |
=================================================================== |
162 |
--- genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-v6r2-3.11.patch 2014-01-29 14:41:45 UTC (rev 2660) |
163 |
+++ genpatches-2.6/trunk/3.14/5000_BFQ-1-block-cgroups-kconfig-build-bits-for-v6r2-3.11.patch 2014-01-30 16:49:47 UTC (rev 2661) |
164 |
@@ -1,97 +0,0 @@ |
165 |
-From 3728677b4d3cd39d83be87f9939328201b871c48 Mon Sep 17 00:00:00 2001 |
166 |
-From: Arianna Avanzini <avanzini.arianna@×××××.com> |
167 |
-Date: Tue, 3 Sep 2013 16:50:42 +0200 |
168 |
-Subject: [PATCH 1/3] block: cgroups, kconfig, build bits for BFQ-v6r2-3.11 |
169 |
- |
170 |
-Update Kconfig.iosched and do the related Makefile changes to include |
171 |
-kernel configuration options for BFQ. Also add the bfqio controller |
172 |
-to the cgroups subsystem. |
173 |
- |
174 |
-Signed-off-by: Paolo Valente <paolo.valente@×××××××.it> |
175 |
-Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
176 |
---- |
177 |
- block/Kconfig.iosched | 25 +++++++++++++++++++++++++ |
178 |
- block/Makefile | 1 + |
179 |
- include/linux/cgroup_subsys.h | 4 ++++ |
180 |
- 3 files changed, 30 insertions(+) |
181 |
- |
182 |
-diff --git a/block/Kconfig.iosched b/block/Kconfig.iosched |
183 |
-index 421bef9..695e064 100644 |
184 |
---- a/block/Kconfig.iosched |
185 |
-+++ b/block/Kconfig.iosched |
186 |
-@@ -39,6 +39,27 @@ config CFQ_GROUP_IOSCHED |
187 |
- ---help--- |
188 |
- Enable group IO scheduling in CFQ. |
189 |
- |
190 |
-+config IOSCHED_BFQ |
191 |
-+ tristate "BFQ I/O scheduler" |
192 |
-+ default n |
193 |
-+ ---help--- |
194 |
-+ The BFQ I/O scheduler tries to distribute bandwidth among |
195 |
-+ all processes according to their weights. |
196 |
-+ It aims at distributing the bandwidth as desired, independently of |
197 |
-+ the disk parameters and with any workload. It also tries to |
198 |
-+ guarantee low latency to interactive and soft real-time |
199 |
-+ applications. If compiled built-in (saying Y here), BFQ can |
200 |
-+ be configured to support hierarchical scheduling. |
201 |
-+ |
202 |
-+config CGROUP_BFQIO |
203 |
-+ bool "BFQ hierarchical scheduling support" |
204 |
-+ depends on CGROUPS && IOSCHED_BFQ=y |
205 |
-+ default n |
206 |
-+ ---help--- |
207 |
-+ Enable hierarchical scheduling in BFQ, using the cgroups |
208 |
-+ filesystem interface. The name of the subsystem will be |
209 |
-+ bfqio. |
210 |
-+ |
211 |
- choice |
212 |
- prompt "Default I/O scheduler" |
213 |
- default DEFAULT_CFQ |
214 |
-@@ -52,6 +73,9 @@ choice |
215 |
- config DEFAULT_CFQ |
216 |
- bool "CFQ" if IOSCHED_CFQ=y |
217 |
- |
218 |
-+ config DEFAULT_BFQ |
219 |
-+ bool "BFQ" if IOSCHED_BFQ=y |
220 |
-+ |
221 |
- config DEFAULT_NOOP |
222 |
- bool "No-op" |
223 |
- |
224 |
-@@ -61,6 +85,7 @@ config DEFAULT_IOSCHED |
225 |
- string |
226 |
- default "deadline" if DEFAULT_DEADLINE |
227 |
- default "cfq" if DEFAULT_CFQ |
228 |
-+ default "bfq" if DEFAULT_BFQ |
229 |
- default "noop" if DEFAULT_NOOP |
230 |
- |
231 |
- endmenu |
232 |
-diff --git a/block/Makefile b/block/Makefile |
233 |
-index 39b76ba..c0d20fa 100644 |
234 |
---- a/block/Makefile |
235 |
-+++ b/block/Makefile |
236 |
-@@ -15,6 +15,7 @@ obj-$(CONFIG_BLK_DEV_THROTTLING) += blk-throttle.o |
237 |
- obj-$(CONFIG_IOSCHED_NOOP) += noop-iosched.o |
238 |
- obj-$(CONFIG_IOSCHED_DEADLINE) += deadline-iosched.o |
239 |
- obj-$(CONFIG_IOSCHED_CFQ) += cfq-iosched.o |
240 |
-+obj-$(CONFIG_IOSCHED_BFQ) += bfq-iosched.o |
241 |
- |
242 |
- obj-$(CONFIG_BLOCK_COMPAT) += compat_ioctl.o |
243 |
- obj-$(CONFIG_BLK_DEV_INTEGRITY) += blk-integrity.o |
244 |
-diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h |
245 |
-index b613ffd..43c5dc9 100644 |
246 |
---- a/include/linux/cgroup_subsys.h |
247 |
-+++ b/include/linux/cgroup_subsys.h |
248 |
-@@ -39,6 +39,10 @@ SUBSYS(net_cls) |
249 |
- SUBSYS(blkio) |
250 |
- #endif |
251 |
- |
252 |
-+#if IS_SUBSYS_ENABLED(CONFIG_CGROUP_BFQIO) |
253 |
-+SUBSYS(bfqio) |
254 |
-+#endif |
255 |
-+ |
256 |
- #if IS_SUBSYS_ENABLED(CONFIG_CGROUP_PERF) |
257 |
- SUBSYS(perf) |
258 |
- #endif |
259 |
--- |
260 |
-1.8.1.4 |
261 |
- |
262 |
|
263 |
Added: genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-BFQ-v7-I-O-sched-for-3.13.patch1 |
264 |
=================================================================== |
265 |
--- genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-BFQ-v7-I-O-sched-for-3.13.patch1 (rev 0) |
266 |
+++ genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-BFQ-v7-I-O-sched-for-3.13.patch1 2014-01-30 16:49:47 UTC (rev 2661) |
267 |
@@ -0,0 +1,6008 @@ |
268 |
+From 3747f129106ce58fbad1b8f05cc836a6addd8588 Mon Sep 17 00:00:00 2001 |
269 |
+From: Paolo Valente <paolo.valente@×××××××.it> |
270 |
+Date: Thu, 9 May 2013 19:10:02 +0200 |
271 |
+Subject: [PATCH 2/3] block: introduce the BFQ-v7 I/O sched for 3.13 |
272 |
+ |
273 |
+Add the BFQ-v7 I/O scheduler to 3.13. |
274 |
+The general structure is borrowed from CFQ, as much of the code for |
275 |
+handling I/O contexts Over time, several useful features have been |
276 |
+ported from CFQ as well (details in the changelog in README.BFQ). A |
277 |
+(bfq_)queue is associated to each task doing I/O on a device, and each |
278 |
+time a scheduling decision has to be made a queue is selected and served |
279 |
+until it expires. |
280 |
+ |
281 |
+ - Slices are given in the service domain: tasks are assigned |
282 |
+ budgets, measured in number of sectors. Once got the disk, a task |
283 |
+ must however consume its assigned budget within a configurable |
284 |
+ maximum time (by default, the maximum possible value of the |
285 |
+ budgets is automatically computed to comply with this timeout). |
286 |
+ This allows the desired latency vs "throughput boosting" tradeoff |
287 |
+ to be set. |
288 |
+ |
289 |
+ - Budgets are scheduled according to a variant of WF2Q+, implemented |
290 |
+ using an augmented rb-tree to take eligibility into account while |
291 |
+ preserving an O(log N) overall complexity. |
292 |
+ |
293 |
+ - A low-latency tunable is provided; if enabled, both interactive |
294 |
+ and soft real-time applications are guaranteed a very low latency. |
295 |
+ |
296 |
+ - Latency guarantees are preserved also in the presence of NCQ. |
297 |
+ |
298 |
+ - Also with flash-based devices, a high throughput is achieved |
299 |
+ while still preserving latency guarantees. |
300 |
+ |
301 |
+ - BFQ features Early Queue Merge (EQM), a sort of fusion of the |
302 |
+ cooperating-queue-merging and the preemption mechanisms present |
303 |
+ in CFQ. EQM is in fact a unified mechanism that tries to get a |
304 |
+ sequential read pattern, and hence a high throughput, with any |
305 |
+ set of processes performing interleaved I/O over a contiguous |
306 |
+ sequence of sectors. |
307 |
+ |
308 |
+ - BFQ supports full hierarchical scheduling, exporting a cgroups |
309 |
+ interface. Since each node has a full scheduler, each group can |
310 |
+ be assigned its own weight. |
311 |
+ |
312 |
+ - If the cgroups interface is not used, only I/O priorities can be |
313 |
+ assigned to processes, with ioprio values mapped to weights |
314 |
+ with the relation weight = IOPRIO_BE_NR - ioprio. |
315 |
+ |
316 |
+ - ioprio classes are served in strict priority order, i.e., lower |
317 |
+ priority queues are not served as long as there are higher |
318 |
+ priority queues. Among queues in the same class the bandwidth is |
319 |
+ distributed in proportion to the weight of each queue. A very |
320 |
+ thin extra bandwidth is however guaranteed to the Idle class, to |
321 |
+ prevent it from starving. |
322 |
+ |
323 |
+Signed-off-by: Paolo Valente <paolo.valente@×××××××.it> |
324 |
+Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
325 |
+--- |
326 |
+ block/bfq-cgroup.c | 910 ++++++++++++++ |
327 |
+ block/bfq-ioc.c | 36 + |
328 |
+ block/bfq-iosched.c | 3268 +++++++++++++++++++++++++++++++++++++++++++++++++++ |
329 |
+ block/bfq-sched.c | 1077 +++++++++++++++++ |
330 |
+ block/bfq.h | 614 ++++++++++ |
331 |
+ 5 files changed, 5905 insertions(+) |
332 |
+ create mode 100644 block/bfq-cgroup.c |
333 |
+ create mode 100644 block/bfq-ioc.c |
334 |
+ create mode 100644 block/bfq-iosched.c |
335 |
+ create mode 100644 block/bfq-sched.c |
336 |
+ create mode 100644 block/bfq.h |
337 |
+ |
338 |
+diff --git a/block/bfq-cgroup.c b/block/bfq-cgroup.c |
339 |
+new file mode 100644 |
340 |
+index 0000000..b889acf |
341 |
+--- /dev/null |
342 |
++++ b/block/bfq-cgroup.c |
343 |
+@@ -0,0 +1,910 @@ |
344 |
++/* |
345 |
++ * BFQ: CGROUPS support. |
346 |
++ * |
347 |
++ * Based on ideas and code from CFQ: |
348 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
349 |
++ * |
350 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
351 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
352 |
++ * |
353 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
354 |
++ * |
355 |
++ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
356 |
++ */ |
357 |
++ |
358 |
++#ifdef CONFIG_CGROUP_BFQIO |
359 |
++ |
360 |
++static DEFINE_MUTEX(bfqio_mutex); |
361 |
++ |
362 |
++static bool bfqio_is_removed(struct bfqio_cgroup *bgrp) |
363 |
++{ |
364 |
++ return bgrp ? !bgrp->online : false; |
365 |
++} |
366 |
++ |
367 |
++static struct bfqio_cgroup bfqio_root_cgroup = { |
368 |
++ .weight = BFQ_DEFAULT_GRP_WEIGHT, |
369 |
++ .ioprio = BFQ_DEFAULT_GRP_IOPRIO, |
370 |
++ .ioprio_class = BFQ_DEFAULT_GRP_CLASS, |
371 |
++}; |
372 |
++ |
373 |
++static inline void bfq_init_entity(struct bfq_entity *entity, |
374 |
++ struct bfq_group *bfqg) |
375 |
++{ |
376 |
++ entity->weight = entity->new_weight; |
377 |
++ entity->orig_weight = entity->new_weight; |
378 |
++ entity->ioprio = entity->new_ioprio; |
379 |
++ entity->ioprio_class = entity->new_ioprio_class; |
380 |
++ entity->parent = bfqg->my_entity; |
381 |
++ entity->sched_data = &bfqg->sched_data; |
382 |
++} |
383 |
++ |
384 |
++static struct bfqio_cgroup *css_to_bfqio(struct cgroup_subsys_state *css) |
385 |
++{ |
386 |
++ return css ? container_of(css, struct bfqio_cgroup, css) : NULL; |
387 |
++} |
388 |
++ |
389 |
++/* |
390 |
++ * Search the bfq_group for bfqd into the hash table (by now only a list) |
391 |
++ * of bgrp. Must be called under rcu_read_lock(). |
392 |
++ */ |
393 |
++static struct bfq_group *bfqio_lookup_group(struct bfqio_cgroup *bgrp, |
394 |
++ struct bfq_data *bfqd) |
395 |
++{ |
396 |
++ struct bfq_group *bfqg; |
397 |
++ void *key; |
398 |
++ |
399 |
++ hlist_for_each_entry_rcu(bfqg, &bgrp->group_data, group_node) { |
400 |
++ key = rcu_dereference(bfqg->bfqd); |
401 |
++ if (key == bfqd) |
402 |
++ return bfqg; |
403 |
++ } |
404 |
++ |
405 |
++ return NULL; |
406 |
++} |
407 |
++ |
408 |
++static inline void bfq_group_init_entity(struct bfqio_cgroup *bgrp, |
409 |
++ struct bfq_group *bfqg) |
410 |
++{ |
411 |
++ struct bfq_entity *entity = &bfqg->entity; |
412 |
++ |
413 |
++ /* |
414 |
++ * If the weight of the entity has never been set via the sysfs |
415 |
++ * interface, then bgrp->weight == 0. In this case we initialize |
416 |
++ * the weight from the current ioprio value. Otherwise, the group |
417 |
++ * weight, if set, has priority over the ioprio value. |
418 |
++ */ |
419 |
++ if (bgrp->weight == 0) { |
420 |
++ entity->new_weight = bfq_ioprio_to_weight(bgrp->ioprio); |
421 |
++ entity->new_ioprio = bgrp->ioprio; |
422 |
++ } else { |
423 |
++ entity->new_weight = bgrp->weight; |
424 |
++ entity->new_ioprio = bfq_weight_to_ioprio(bgrp->weight); |
425 |
++ } |
426 |
++ entity->orig_weight = entity->weight = entity->new_weight; |
427 |
++ entity->ioprio = entity->new_ioprio; |
428 |
++ entity->ioprio_class = entity->new_ioprio_class = bgrp->ioprio_class; |
429 |
++ entity->my_sched_data = &bfqg->sched_data; |
430 |
++} |
431 |
++ |
432 |
++static inline void bfq_group_set_parent(struct bfq_group *bfqg, |
433 |
++ struct bfq_group *parent) |
434 |
++{ |
435 |
++ struct bfq_entity *entity; |
436 |
++ |
437 |
++ BUG_ON(parent == NULL); |
438 |
++ BUG_ON(bfqg == NULL); |
439 |
++ |
440 |
++ entity = &bfqg->entity; |
441 |
++ entity->parent = parent->my_entity; |
442 |
++ entity->sched_data = &parent->sched_data; |
443 |
++} |
444 |
++ |
445 |
++/** |
446 |
++ * bfq_group_chain_alloc - allocate a chain of groups. |
447 |
++ * @bfqd: queue descriptor. |
448 |
++ * @css: the leaf cgroup_subsys_state this chain starts from. |
449 |
++ * |
450 |
++ * Allocate a chain of groups starting from the one belonging to |
451 |
++ * @cgroup up to the root cgroup. Stop if a cgroup on the chain |
452 |
++ * to the root has already an allocated group on @bfqd. |
453 |
++ */ |
454 |
++static struct bfq_group *bfq_group_chain_alloc(struct bfq_data *bfqd, |
455 |
++ struct cgroup_subsys_state *css) |
456 |
++{ |
457 |
++ struct bfqio_cgroup *bgrp; |
458 |
++ struct bfq_group *bfqg, *prev = NULL, *leaf = NULL; |
459 |
++ |
460 |
++ for (; css != NULL; css = css->parent) { |
461 |
++ bgrp = css_to_bfqio(css); |
462 |
++ |
463 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
464 |
++ if (bfqg != NULL) { |
465 |
++ /* |
466 |
++ * All the cgroups in the path from there to the |
467 |
++ * root must have a bfq_group for bfqd, so we don't |
468 |
++ * need any more allocations. |
469 |
++ */ |
470 |
++ break; |
471 |
++ } |
472 |
++ |
473 |
++ bfqg = kzalloc(sizeof(*bfqg), GFP_ATOMIC); |
474 |
++ if (bfqg == NULL) |
475 |
++ goto cleanup; |
476 |
++ |
477 |
++ bfq_group_init_entity(bgrp, bfqg); |
478 |
++ bfqg->my_entity = &bfqg->entity; |
479 |
++ |
480 |
++ if (leaf == NULL) { |
481 |
++ leaf = bfqg; |
482 |
++ prev = leaf; |
483 |
++ } else { |
484 |
++ bfq_group_set_parent(prev, bfqg); |
485 |
++ /* |
486 |
++ * Build a list of allocated nodes using the bfqd |
487 |
++ * filed, that is still unused and will be initialized |
488 |
++ * only after the node will be connected. |
489 |
++ */ |
490 |
++ prev->bfqd = bfqg; |
491 |
++ prev = bfqg; |
492 |
++ } |
493 |
++ } |
494 |
++ |
495 |
++ return leaf; |
496 |
++ |
497 |
++cleanup: |
498 |
++ while (leaf != NULL) { |
499 |
++ prev = leaf; |
500 |
++ leaf = leaf->bfqd; |
501 |
++ kfree(prev); |
502 |
++ } |
503 |
++ |
504 |
++ return NULL; |
505 |
++} |
506 |
++ |
507 |
++/** |
508 |
++ * bfq_group_chain_link - link an allocatd group chain to a cgroup hierarchy. |
509 |
++ * @bfqd: the queue descriptor. |
510 |
++ * @css: the leaf cgroup_subsys_state to start from. |
511 |
++ * @leaf: the leaf group (to be associated to @cgroup). |
512 |
++ * |
513 |
++ * Try to link a chain of groups to a cgroup hierarchy, connecting the |
514 |
++ * nodes bottom-up, so we can be sure that when we find a cgroup in the |
515 |
++ * hierarchy that already as a group associated to @bfqd all the nodes |
516 |
++ * in the path to the root cgroup have one too. |
517 |
++ * |
518 |
++ * On locking: the queue lock protects the hierarchy (there is a hierarchy |
519 |
++ * per device) while the bfqio_cgroup lock protects the list of groups |
520 |
++ * belonging to the same cgroup. |
521 |
++ */ |
522 |
++static void bfq_group_chain_link(struct bfq_data *bfqd, |
523 |
++ struct cgroup_subsys_state *css, |
524 |
++ struct bfq_group *leaf) |
525 |
++{ |
526 |
++ struct bfqio_cgroup *bgrp; |
527 |
++ struct bfq_group *bfqg, *next, *prev = NULL; |
528 |
++ unsigned long flags; |
529 |
++ |
530 |
++ assert_spin_locked(bfqd->queue->queue_lock); |
531 |
++ |
532 |
++ for (; css != NULL && leaf != NULL; css = css->parent) { |
533 |
++ bgrp = css_to_bfqio(css); |
534 |
++ next = leaf->bfqd; |
535 |
++ |
536 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
537 |
++ BUG_ON(bfqg != NULL); |
538 |
++ |
539 |
++ spin_lock_irqsave(&bgrp->lock, flags); |
540 |
++ |
541 |
++ rcu_assign_pointer(leaf->bfqd, bfqd); |
542 |
++ hlist_add_head_rcu(&leaf->group_node, &bgrp->group_data); |
543 |
++ hlist_add_head(&leaf->bfqd_node, &bfqd->group_list); |
544 |
++ |
545 |
++ spin_unlock_irqrestore(&bgrp->lock, flags); |
546 |
++ |
547 |
++ prev = leaf; |
548 |
++ leaf = next; |
549 |
++ } |
550 |
++ |
551 |
++ BUG_ON(css == NULL && leaf != NULL); |
552 |
++ if (css != NULL && prev != NULL) { |
553 |
++ bgrp = css_to_bfqio(css); |
554 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
555 |
++ bfq_group_set_parent(prev, bfqg); |
556 |
++ } |
557 |
++} |
558 |
++ |
559 |
++/** |
560 |
++ * bfq_find_alloc_group - return the group associated to @bfqd in @cgroup. |
561 |
++ * @bfqd: queue descriptor. |
562 |
++ * @cgroup: cgroup being searched for. |
563 |
++ * |
564 |
++ * Return a group associated to @bfqd in @cgroup, allocating one if |
565 |
++ * necessary. When a group is returned all the cgroups in the path |
566 |
++ * to the root have a group associated to @bfqd. |
567 |
++ * |
568 |
++ * If the allocation fails, return the root group: this breaks guarantees |
569 |
++ * but is a safe fallbak. If this loss becames a problem it can be |
570 |
++ * mitigated using the equivalent weight (given by the product of the |
571 |
++ * weights of the groups in the path from @group to the root) in the |
572 |
++ * root scheduler. |
573 |
++ * |
574 |
++ * We allocate all the missing nodes in the path from the leaf cgroup |
575 |
++ * to the root and we connect the nodes only after all the allocations |
576 |
++ * have been successful. |
577 |
++ */ |
578 |
++static struct bfq_group *bfq_find_alloc_group(struct bfq_data *bfqd, |
579 |
++ struct cgroup_subsys_state *css) |
580 |
++{ |
581 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
582 |
++ struct bfq_group *bfqg; |
583 |
++ |
584 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
585 |
++ if (bfqg != NULL) |
586 |
++ return bfqg; |
587 |
++ |
588 |
++ bfqg = bfq_group_chain_alloc(bfqd, css); |
589 |
++ if (bfqg != NULL) |
590 |
++ bfq_group_chain_link(bfqd, css, bfqg); |
591 |
++ else |
592 |
++ bfqg = bfqd->root_group; |
593 |
++ |
594 |
++ return bfqg; |
595 |
++} |
596 |
++ |
597 |
++/** |
598 |
++ * bfq_bfqq_move - migrate @bfqq to @bfqg. |
599 |
++ * @bfqd: queue descriptor. |
600 |
++ * @bfqq: the queue to move. |
601 |
++ * @entity: @bfqq's entity. |
602 |
++ * @bfqg: the group to move to. |
603 |
++ * |
604 |
++ * Move @bfqq to @bfqg, deactivating it from its old group and reactivating |
605 |
++ * it on the new one. Avoid putting the entity on the old group idle tree. |
606 |
++ * |
607 |
++ * Must be called under the queue lock; the cgroup owning @bfqg must |
608 |
++ * not disappear (by now this just means that we are called under |
609 |
++ * rcu_read_lock()). |
610 |
++ */ |
611 |
++static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
612 |
++ struct bfq_entity *entity, struct bfq_group *bfqg) |
613 |
++{ |
614 |
++ int busy, resume; |
615 |
++ |
616 |
++ busy = bfq_bfqq_busy(bfqq); |
617 |
++ resume = !RB_EMPTY_ROOT(&bfqq->sort_list); |
618 |
++ |
619 |
++ BUG_ON(resume && !entity->on_st); |
620 |
++ BUG_ON(busy && !resume && entity->on_st && |
621 |
++ bfqq != bfqd->in_service_queue); |
622 |
++ |
623 |
++ if (busy) { |
624 |
++ BUG_ON(atomic_read(&bfqq->ref) < 2); |
625 |
++ |
626 |
++ if (!resume) |
627 |
++ bfq_del_bfqq_busy(bfqd, bfqq, 0); |
628 |
++ else |
629 |
++ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
630 |
++ } else if (entity->on_st) |
631 |
++ bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); |
632 |
++ |
633 |
++ /* |
634 |
++ * Here we use a reference to bfqg. We don't need a refcounter |
635 |
++ * as the cgroup reference will not be dropped, so that its |
636 |
++ * destroy() callback will not be invoked. |
637 |
++ */ |
638 |
++ entity->parent = bfqg->my_entity; |
639 |
++ entity->sched_data = &bfqg->sched_data; |
640 |
++ |
641 |
++ if (busy && resume) |
642 |
++ bfq_activate_bfqq(bfqd, bfqq); |
643 |
++ |
644 |
++ if (bfqd->in_service_queue == NULL && !bfqd->rq_in_driver) |
645 |
++ bfq_schedule_dispatch(bfqd); |
646 |
++} |
647 |
++ |
648 |
++/** |
649 |
++ * __bfq_bic_change_cgroup - move @bic to @cgroup. |
650 |
++ * @bfqd: the queue descriptor. |
651 |
++ * @bic: the bic to move. |
652 |
++ * @cgroup: the cgroup to move to. |
653 |
++ * |
654 |
++ * Move bic to cgroup, assuming that bfqd->queue is locked; the caller |
655 |
++ * has to make sure that the reference to cgroup is valid across the call. |
656 |
++ * |
657 |
++ * NOTE: an alternative approach might have been to store the current |
658 |
++ * cgroup in bfqq and getting a reference to it, reducing the lookup |
659 |
++ * time here, at the price of slightly more complex code. |
660 |
++ */ |
661 |
++static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
662 |
++ struct bfq_io_cq *bic, |
663 |
++ struct cgroup_subsys_state *css) |
664 |
++{ |
665 |
++ struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0); |
666 |
++ struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1); |
667 |
++ struct bfq_entity *entity; |
668 |
++ struct bfq_group *bfqg; |
669 |
++ struct bfqio_cgroup *bgrp; |
670 |
++ |
671 |
++ bgrp = css_to_bfqio(css); |
672 |
++ |
673 |
++ bfqg = bfq_find_alloc_group(bfqd, css); |
674 |
++ if (async_bfqq != NULL) { |
675 |
++ entity = &async_bfqq->entity; |
676 |
++ |
677 |
++ if (entity->sched_data != &bfqg->sched_data) { |
678 |
++ bic_set_bfqq(bic, NULL, 0); |
679 |
++ bfq_log_bfqq(bfqd, async_bfqq, |
680 |
++ "bic_change_group: %p %d", |
681 |
++ async_bfqq, atomic_read(&async_bfqq->ref)); |
682 |
++ bfq_put_queue(async_bfqq); |
683 |
++ } |
684 |
++ } |
685 |
++ |
686 |
++ if (sync_bfqq != NULL) { |
687 |
++ entity = &sync_bfqq->entity; |
688 |
++ if (entity->sched_data != &bfqg->sched_data) |
689 |
++ bfq_bfqq_move(bfqd, sync_bfqq, entity, bfqg); |
690 |
++ } |
691 |
++ |
692 |
++ return bfqg; |
693 |
++} |
694 |
++ |
695 |
++/** |
696 |
++ * bfq_bic_change_cgroup - move @bic to @cgroup. |
697 |
++ * @bic: the bic being migrated. |
698 |
++ * @cgroup: the destination cgroup. |
699 |
++ * |
700 |
++ * When the task owning @bic is moved to @cgroup, @bic is immediately |
701 |
++ * moved into its new parent group. |
702 |
++ */ |
703 |
++static void bfq_bic_change_cgroup(struct bfq_io_cq *bic, |
704 |
++ struct cgroup_subsys_state *css) |
705 |
++{ |
706 |
++ struct bfq_data *bfqd; |
707 |
++ unsigned long uninitialized_var(flags); |
708 |
++ |
709 |
++ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), |
710 |
++ &flags); |
711 |
++ if (bfqd != NULL) { |
712 |
++ __bfq_bic_change_cgroup(bfqd, bic, css); |
713 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
714 |
++ } |
715 |
++} |
716 |
++ |
717 |
++/** |
718 |
++ * bfq_bic_update_cgroup - update the cgroup of @bic. |
719 |
++ * @bic: the @bic to update. |
720 |
++ * |
721 |
++ * Make sure that @bic is enqueued in the cgroup of the current task. |
722 |
++ * We need this in addition to moving bics during the cgroup attach |
723 |
++ * phase because the task owning @bic could be at its first disk |
724 |
++ * access or we may end up in the root cgroup as the result of a |
725 |
++ * memory allocation failure and here we try to move to the right |
726 |
++ * group. |
727 |
++ * |
728 |
++ * Must be called under the queue lock. It is safe to use the returned |
729 |
++ * value even after the rcu_read_unlock() as the migration/destruction |
730 |
++ * paths act under the queue lock too. IOW it is impossible to race with |
731 |
++ * group migration/destruction and end up with an invalid group as: |
732 |
++ * a) here cgroup has not yet been destroyed, nor its destroy callback |
733 |
++ * has started execution, as current holds a reference to it, |
734 |
++ * b) if it is destroyed after rcu_read_unlock() [after current is |
735 |
++ * migrated to a different cgroup] its attach() callback will have |
736 |
++ * taken care of remove all the references to the old cgroup data. |
737 |
++ */ |
738 |
++static struct bfq_group *bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
739 |
++{ |
740 |
++ struct bfq_data *bfqd = bic_to_bfqd(bic); |
741 |
++ struct bfq_group *bfqg; |
742 |
++ struct cgroup_subsys_state *css; |
743 |
++ |
744 |
++ BUG_ON(bfqd == NULL); |
745 |
++ |
746 |
++ rcu_read_lock(); |
747 |
++ css = task_css(current, bfqio_subsys_id); |
748 |
++ bfqg = __bfq_bic_change_cgroup(bfqd, bic, css); |
749 |
++ rcu_read_unlock(); |
750 |
++ |
751 |
++ return bfqg; |
752 |
++} |
753 |
++ |
754 |
++/** |
755 |
++ * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st. |
756 |
++ * @st: the service tree being flushed. |
757 |
++ */ |
758 |
++static inline void bfq_flush_idle_tree(struct bfq_service_tree *st) |
759 |
++{ |
760 |
++ struct bfq_entity *entity = st->first_idle; |
761 |
++ |
762 |
++ for (; entity != NULL; entity = st->first_idle) |
763 |
++ __bfq_deactivate_entity(entity, 0); |
764 |
++} |
765 |
++ |
766 |
++/** |
767 |
++ * bfq_reparent_leaf_entity - move leaf entity to the root_group. |
768 |
++ * @bfqd: the device data structure with the root group. |
769 |
++ * @entity: the entity to move. |
770 |
++ */ |
771 |
++static inline void bfq_reparent_leaf_entity(struct bfq_data *bfqd, |
772 |
++ struct bfq_entity *entity) |
773 |
++{ |
774 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
775 |
++ |
776 |
++ BUG_ON(bfqq == NULL); |
777 |
++ bfq_bfqq_move(bfqd, bfqq, entity, bfqd->root_group); |
778 |
++ return; |
779 |
++} |
780 |
++ |
781 |
++/** |
782 |
++ * bfq_reparent_active_entities - move to the root group all active entities. |
783 |
++ * @bfqd: the device data structure with the root group. |
784 |
++ * @bfqg: the group to move from. |
785 |
++ * @st: the service tree with the entities. |
786 |
++ * |
787 |
++ * Needs queue_lock to be taken and reference to be valid over the call. |
788 |
++ */ |
789 |
++static inline void bfq_reparent_active_entities(struct bfq_data *bfqd, |
790 |
++ struct bfq_group *bfqg, |
791 |
++ struct bfq_service_tree *st) |
792 |
++{ |
793 |
++ struct rb_root *active = &st->active; |
794 |
++ struct bfq_entity *entity = NULL; |
795 |
++ |
796 |
++ if (!RB_EMPTY_ROOT(&st->active)) |
797 |
++ entity = bfq_entity_of(rb_first(active)); |
798 |
++ |
799 |
++ for (; entity != NULL; entity = bfq_entity_of(rb_first(active))) |
800 |
++ bfq_reparent_leaf_entity(bfqd, entity); |
801 |
++ |
802 |
++ if (bfqg->sched_data.active_entity != NULL) |
803 |
++ bfq_reparent_leaf_entity(bfqd, bfqg->sched_data.active_entity); |
804 |
++ |
805 |
++ return; |
806 |
++} |
807 |
++ |
808 |
++/** |
809 |
++ * bfq_destroy_group - destroy @bfqg. |
810 |
++ * @bgrp: the bfqio_cgroup containing @bfqg. |
811 |
++ * @bfqg: the group being destroyed. |
812 |
++ * |
813 |
++ * Destroy @bfqg, making sure that it is not referenced from its parent. |
814 |
++ */ |
815 |
++static void bfq_destroy_group(struct bfqio_cgroup *bgrp, struct bfq_group *bfqg) |
816 |
++{ |
817 |
++ struct bfq_data *bfqd; |
818 |
++ struct bfq_service_tree *st; |
819 |
++ struct bfq_entity *entity = bfqg->my_entity; |
820 |
++ unsigned long uninitialized_var(flags); |
821 |
++ int i; |
822 |
++ |
823 |
++ hlist_del(&bfqg->group_node); |
824 |
++ |
825 |
++ /* |
826 |
++ * Empty all service_trees belonging to this group before deactivating |
827 |
++ * the group itself. |
828 |
++ */ |
829 |
++ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) { |
830 |
++ st = bfqg->sched_data.service_tree + i; |
831 |
++ |
832 |
++ /* |
833 |
++ * The idle tree may still contain bfq_queues belonging |
834 |
++ * to exited task because they never migrated to a different |
835 |
++ * cgroup from the one being destroyed now. Noone else |
836 |
++ * can access them so it's safe to act without any lock. |
837 |
++ */ |
838 |
++ bfq_flush_idle_tree(st); |
839 |
++ |
840 |
++ /* |
841 |
++ * It may happen that some queues are still active |
842 |
++ * (busy) upon group destruction (if the corresponding |
843 |
++ * processes have been forced to terminate). We move |
844 |
++ * all the leaf entities corresponding to these queues |
845 |
++ * to the root_group. |
846 |
++ * Also, it may happen that the group has an entity |
847 |
++ * under service, which is disconnected from the active |
848 |
++ * tree: it must be moved, too. |
849 |
++ * There is no need to put the sync queues, as the |
850 |
++ * scheduler has taken no reference. |
851 |
++ */ |
852 |
++ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
853 |
++ if (bfqd != NULL) { |
854 |
++ bfq_reparent_active_entities(bfqd, bfqg, st); |
855 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
856 |
++ } |
857 |
++ BUG_ON(!RB_EMPTY_ROOT(&st->active)); |
858 |
++ BUG_ON(!RB_EMPTY_ROOT(&st->idle)); |
859 |
++ } |
860 |
++ BUG_ON(bfqg->sched_data.next_active != NULL); |
861 |
++ BUG_ON(bfqg->sched_data.active_entity != NULL); |
862 |
++ |
863 |
++ /* |
864 |
++ * We may race with device destruction, take extra care when |
865 |
++ * dereferencing bfqg->bfqd. |
866 |
++ */ |
867 |
++ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
868 |
++ if (bfqd != NULL) { |
869 |
++ hlist_del(&bfqg->bfqd_node); |
870 |
++ __bfq_deactivate_entity(entity, 0); |
871 |
++ bfq_put_async_queues(bfqd, bfqg); |
872 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
873 |
++ } |
874 |
++ BUG_ON(entity->tree != NULL); |
875 |
++ |
876 |
++ /* |
877 |
++ * No need to defer the kfree() to the end of the RCU grace |
878 |
++ * period: we are called from the destroy() callback of our |
879 |
++ * cgroup, so we can be sure that noone is a) still using |
880 |
++ * this cgroup or b) doing lookups in it. |
881 |
++ */ |
882 |
++ kfree(bfqg); |
883 |
++} |
884 |
++ |
885 |
++static void bfq_end_raising_async(struct bfq_data *bfqd) |
886 |
++{ |
887 |
++ struct hlist_node *tmp; |
888 |
++ struct bfq_group *bfqg; |
889 |
++ |
890 |
++ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) |
891 |
++ bfq_end_raising_async_queues(bfqd, bfqg); |
892 |
++ bfq_end_raising_async_queues(bfqd, bfqd->root_group); |
893 |
++} |
894 |
++ |
895 |
++/** |
896 |
++ * bfq_disconnect_groups - diconnect @bfqd from all its groups. |
897 |
++ * @bfqd: the device descriptor being exited. |
898 |
++ * |
899 |
++ * When the device exits we just make sure that no lookup can return |
900 |
++ * the now unused group structures. They will be deallocated on cgroup |
901 |
++ * destruction. |
902 |
++ */ |
903 |
++static void bfq_disconnect_groups(struct bfq_data *bfqd) |
904 |
++{ |
905 |
++ struct hlist_node *tmp; |
906 |
++ struct bfq_group *bfqg; |
907 |
++ |
908 |
++ bfq_log(bfqd, "disconnect_groups beginning"); |
909 |
++ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) { |
910 |
++ hlist_del(&bfqg->bfqd_node); |
911 |
++ |
912 |
++ __bfq_deactivate_entity(bfqg->my_entity, 0); |
913 |
++ |
914 |
++ /* |
915 |
++ * Don't remove from the group hash, just set an |
916 |
++ * invalid key. No lookups can race with the |
917 |
++ * assignment as bfqd is being destroyed; this |
918 |
++ * implies also that new elements cannot be added |
919 |
++ * to the list. |
920 |
++ */ |
921 |
++ rcu_assign_pointer(bfqg->bfqd, NULL); |
922 |
++ |
923 |
++ bfq_log(bfqd, "disconnect_groups: put async for group %p", |
924 |
++ bfqg); |
925 |
++ bfq_put_async_queues(bfqd, bfqg); |
926 |
++ } |
927 |
++} |
928 |
++ |
929 |
++static inline void bfq_free_root_group(struct bfq_data *bfqd) |
930 |
++{ |
931 |
++ struct bfqio_cgroup *bgrp = &bfqio_root_cgroup; |
932 |
++ struct bfq_group *bfqg = bfqd->root_group; |
933 |
++ |
934 |
++ bfq_put_async_queues(bfqd, bfqg); |
935 |
++ |
936 |
++ spin_lock_irq(&bgrp->lock); |
937 |
++ hlist_del_rcu(&bfqg->group_node); |
938 |
++ spin_unlock_irq(&bgrp->lock); |
939 |
++ |
940 |
++ /* |
941 |
++ * No need to synchronize_rcu() here: since the device is gone |
942 |
++ * there cannot be any read-side access to its root_group. |
943 |
++ */ |
944 |
++ kfree(bfqg); |
945 |
++} |
946 |
++ |
947 |
++static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
948 |
++{ |
949 |
++ struct bfq_group *bfqg; |
950 |
++ struct bfqio_cgroup *bgrp; |
951 |
++ int i; |
952 |
++ |
953 |
++ bfqg = kzalloc_node(sizeof(*bfqg), GFP_KERNEL, node); |
954 |
++ if (bfqg == NULL) |
955 |
++ return NULL; |
956 |
++ |
957 |
++ bfqg->entity.parent = NULL; |
958 |
++ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
959 |
++ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
960 |
++ |
961 |
++ bgrp = &bfqio_root_cgroup; |
962 |
++ spin_lock_irq(&bgrp->lock); |
963 |
++ rcu_assign_pointer(bfqg->bfqd, bfqd); |
964 |
++ hlist_add_head_rcu(&bfqg->group_node, &bgrp->group_data); |
965 |
++ spin_unlock_irq(&bgrp->lock); |
966 |
++ |
967 |
++ return bfqg; |
968 |
++} |
969 |
++ |
970 |
++#define SHOW_FUNCTION(__VAR) \ |
971 |
++static u64 bfqio_cgroup_##__VAR##_read(struct cgroup_subsys_state *css, \ |
972 |
++ struct cftype *cftype) \ |
973 |
++{ \ |
974 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
975 |
++ u64 ret = -ENODEV; \ |
976 |
++ \ |
977 |
++ mutex_lock(&bfqio_mutex); \ |
978 |
++ if (bfqio_is_removed(bgrp)) \ |
979 |
++ goto out_unlock; \ |
980 |
++ \ |
981 |
++ spin_lock_irq(&bgrp->lock); \ |
982 |
++ ret = bgrp->__VAR; \ |
983 |
++ spin_unlock_irq(&bgrp->lock); \ |
984 |
++ \ |
985 |
++out_unlock: \ |
986 |
++ mutex_unlock(&bfqio_mutex); \ |
987 |
++ return ret; \ |
988 |
++} |
989 |
++ |
990 |
++SHOW_FUNCTION(weight); |
991 |
++SHOW_FUNCTION(ioprio); |
992 |
++SHOW_FUNCTION(ioprio_class); |
993 |
++#undef SHOW_FUNCTION |
994 |
++ |
995 |
++#define STORE_FUNCTION(__VAR, __MIN, __MAX) \ |
996 |
++static int bfqio_cgroup_##__VAR##_write(struct cgroup_subsys_state *css,\ |
997 |
++ struct cftype *cftype, \ |
998 |
++ u64 val) \ |
999 |
++{ \ |
1000 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
1001 |
++ struct bfq_group *bfqg; \ |
1002 |
++ int ret = -EINVAL; \ |
1003 |
++ \ |
1004 |
++ if (val < (__MIN) || val > (__MAX)) \ |
1005 |
++ return ret; \ |
1006 |
++ \ |
1007 |
++ ret = -ENODEV; \ |
1008 |
++ mutex_lock(&bfqio_mutex); \ |
1009 |
++ if (bfqio_is_removed(bgrp)) \ |
1010 |
++ goto out_unlock; \ |
1011 |
++ ret = 0; \ |
1012 |
++ \ |
1013 |
++ spin_lock_irq(&bgrp->lock); \ |
1014 |
++ bgrp->__VAR = (unsigned short)val; \ |
1015 |
++ hlist_for_each_entry(bfqg, &bgrp->group_data, group_node) { \ |
1016 |
++ /* \ |
1017 |
++ * Setting the ioprio_changed flag of the entity \ |
1018 |
++ * to 1 with new_##__VAR == ##__VAR would re-set \ |
1019 |
++ * the value of the weight to its ioprio mapping. \ |
1020 |
++ * Set the flag only if necessary. \ |
1021 |
++ */ \ |
1022 |
++ if ((unsigned short)val != bfqg->entity.new_##__VAR) { \ |
1023 |
++ bfqg->entity.new_##__VAR = (unsigned short)val; \ |
1024 |
++ smp_wmb(); \ |
1025 |
++ bfqg->entity.ioprio_changed = 1; \ |
1026 |
++ } \ |
1027 |
++ } \ |
1028 |
++ spin_unlock_irq(&bgrp->lock); \ |
1029 |
++ \ |
1030 |
++out_unlock: \ |
1031 |
++ mutex_unlock(&bfqio_mutex); \ |
1032 |
++ return ret; \ |
1033 |
++} |
1034 |
++ |
1035 |
++STORE_FUNCTION(weight, BFQ_MIN_WEIGHT, BFQ_MAX_WEIGHT); |
1036 |
++STORE_FUNCTION(ioprio, 0, IOPRIO_BE_NR - 1); |
1037 |
++STORE_FUNCTION(ioprio_class, IOPRIO_CLASS_RT, IOPRIO_CLASS_IDLE); |
1038 |
++#undef STORE_FUNCTION |
1039 |
++ |
1040 |
++static struct cftype bfqio_files[] = { |
1041 |
++ { |
1042 |
++ .name = "weight", |
1043 |
++ .read_u64 = bfqio_cgroup_weight_read, |
1044 |
++ .write_u64 = bfqio_cgroup_weight_write, |
1045 |
++ }, |
1046 |
++ { |
1047 |
++ .name = "ioprio", |
1048 |
++ .read_u64 = bfqio_cgroup_ioprio_read, |
1049 |
++ .write_u64 = bfqio_cgroup_ioprio_write, |
1050 |
++ }, |
1051 |
++ { |
1052 |
++ .name = "ioprio_class", |
1053 |
++ .read_u64 = bfqio_cgroup_ioprio_class_read, |
1054 |
++ .write_u64 = bfqio_cgroup_ioprio_class_write, |
1055 |
++ }, |
1056 |
++ { }, /* terminate */ |
1057 |
++}; |
1058 |
++ |
1059 |
++static struct cgroup_subsys_state *bfqio_create(struct cgroup_subsys_state |
1060 |
++ *parent_css) |
1061 |
++{ |
1062 |
++ struct bfqio_cgroup *bgrp; |
1063 |
++ |
1064 |
++ if (parent_css != NULL) { |
1065 |
++ bgrp = kzalloc(sizeof(*bgrp), GFP_KERNEL); |
1066 |
++ if (bgrp == NULL) |
1067 |
++ return ERR_PTR(-ENOMEM); |
1068 |
++ } else |
1069 |
++ bgrp = &bfqio_root_cgroup; |
1070 |
++ |
1071 |
++ spin_lock_init(&bgrp->lock); |
1072 |
++ INIT_HLIST_HEAD(&bgrp->group_data); |
1073 |
++ bgrp->ioprio = BFQ_DEFAULT_GRP_IOPRIO; |
1074 |
++ bgrp->ioprio_class = BFQ_DEFAULT_GRP_CLASS; |
1075 |
++ |
1076 |
++ return &bgrp->css; |
1077 |
++} |
1078 |
++ |
1079 |
++/* |
1080 |
++ * We cannot support shared io contexts, as we have no means to support |
1081 |
++ * two tasks with the same ioc in two different groups without major rework |
1082 |
++ * of the main bic/bfqq data structures. By now we allow a task to change |
1083 |
++ * its cgroup only if it's the only owner of its ioc; the drawback of this |
1084 |
++ * behavior is that a group containing a task that forked using CLONE_IO |
1085 |
++ * will not be destroyed until the tasks sharing the ioc die. |
1086 |
++ */ |
1087 |
++static int bfqio_can_attach(struct cgroup_subsys_state *css, |
1088 |
++ struct cgroup_taskset *tset) |
1089 |
++{ |
1090 |
++ struct task_struct *task; |
1091 |
++ struct io_context *ioc; |
1092 |
++ int ret = 0; |
1093 |
++ |
1094 |
++ cgroup_taskset_for_each(task, css, tset) { |
1095 |
++ /* |
1096 |
++ * task_lock() is needed to avoid races with |
1097 |
++ * exit_io_context() |
1098 |
++ */ |
1099 |
++ task_lock(task); |
1100 |
++ ioc = task->io_context; |
1101 |
++ if (ioc != NULL && atomic_read(&ioc->nr_tasks) > 1) |
1102 |
++ /* |
1103 |
++ * ioc == NULL means that the task is either too young |
1104 |
++ * or exiting: if it has still no ioc the ioc can't be |
1105 |
++ * shared, if the task is exiting the attach will fail |
1106 |
++ * anyway, no matter what we return here. |
1107 |
++ */ |
1108 |
++ ret = -EINVAL; |
1109 |
++ task_unlock(task); |
1110 |
++ if (ret) |
1111 |
++ break; |
1112 |
++ } |
1113 |
++ |
1114 |
++ return ret; |
1115 |
++} |
1116 |
++ |
1117 |
++static void bfqio_attach(struct cgroup_subsys_state *css, |
1118 |
++ struct cgroup_taskset *tset) |
1119 |
++{ |
1120 |
++ struct task_struct *task; |
1121 |
++ struct io_context *ioc; |
1122 |
++ struct io_cq *icq; |
1123 |
++ |
1124 |
++ /* |
1125 |
++ * IMPORTANT NOTE: The move of more than one process at a time to a |
1126 |
++ * new group has not yet been tested. |
1127 |
++ */ |
1128 |
++ cgroup_taskset_for_each(task, css, tset) { |
1129 |
++ ioc = get_task_io_context(task, GFP_ATOMIC, NUMA_NO_NODE); |
1130 |
++ if (ioc) { |
1131 |
++ /* |
1132 |
++ * Handle cgroup change here. |
1133 |
++ */ |
1134 |
++ rcu_read_lock(); |
1135 |
++ hlist_for_each_entry_rcu(icq, &ioc->icq_list, ioc_node) |
1136 |
++ if (!strncmp( |
1137 |
++ icq->q->elevator->type->elevator_name, |
1138 |
++ "bfq", ELV_NAME_MAX)) |
1139 |
++ bfq_bic_change_cgroup(icq_to_bic(icq), |
1140 |
++ css); |
1141 |
++ rcu_read_unlock(); |
1142 |
++ put_io_context(ioc); |
1143 |
++ } |
1144 |
++ } |
1145 |
++} |
1146 |
++ |
1147 |
++static void bfqio_destroy(struct cgroup_subsys_state *css) |
1148 |
++{ |
1149 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
1150 |
++ struct hlist_node *tmp; |
1151 |
++ struct bfq_group *bfqg; |
1152 |
++ |
1153 |
++ /* |
1154 |
++ * Since we are destroying the cgroup, there are no more tasks |
1155 |
++ * referencing it, and all the RCU grace periods that may have |
1156 |
++ * referenced it are ended (as the destruction of the parent |
1157 |
++ * cgroup is RCU-safe); bgrp->group_data will not be accessed by |
1158 |
++ * anything else and we don't need any synchronization. |
1159 |
++ */ |
1160 |
++ hlist_for_each_entry_safe(bfqg, tmp, &bgrp->group_data, group_node) |
1161 |
++ bfq_destroy_group(bgrp, bfqg); |
1162 |
++ |
1163 |
++ BUG_ON(!hlist_empty(&bgrp->group_data)); |
1164 |
++ |
1165 |
++ kfree(bgrp); |
1166 |
++} |
1167 |
++ |
1168 |
++static int bfqio_css_online(struct cgroup_subsys_state *css) |
1169 |
++{ |
1170 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
1171 |
++ |
1172 |
++ mutex_lock(&bfqio_mutex); |
1173 |
++ bgrp->online = true; |
1174 |
++ mutex_unlock(&bfqio_mutex); |
1175 |
++ |
1176 |
++ return 0; |
1177 |
++} |
1178 |
++ |
1179 |
++static void bfqio_css_offline(struct cgroup_subsys_state *css) |
1180 |
++{ |
1181 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
1182 |
++ |
1183 |
++ mutex_lock(&bfqio_mutex); |
1184 |
++ bgrp->online = false; |
1185 |
++ mutex_unlock(&bfqio_mutex); |
1186 |
++} |
1187 |
++ |
1188 |
++struct cgroup_subsys bfqio_subsys = { |
1189 |
++ .name = "bfqio", |
1190 |
++ .css_alloc = bfqio_create, |
1191 |
++ .css_online = bfqio_css_online, |
1192 |
++ .css_offline = bfqio_css_offline, |
1193 |
++ .can_attach = bfqio_can_attach, |
1194 |
++ .attach = bfqio_attach, |
1195 |
++ .css_free = bfqio_destroy, |
1196 |
++ .subsys_id = bfqio_subsys_id, |
1197 |
++ .base_cftypes = bfqio_files, |
1198 |
++}; |
1199 |
++#else |
1200 |
++static inline void bfq_init_entity(struct bfq_entity *entity, |
1201 |
++ struct bfq_group *bfqg) |
1202 |
++{ |
1203 |
++ entity->weight = entity->new_weight; |
1204 |
++ entity->orig_weight = entity->new_weight; |
1205 |
++ entity->ioprio = entity->new_ioprio; |
1206 |
++ entity->ioprio_class = entity->new_ioprio_class; |
1207 |
++ entity->sched_data = &bfqg->sched_data; |
1208 |
++} |
1209 |
++ |
1210 |
++static inline struct bfq_group * |
1211 |
++bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
1212 |
++{ |
1213 |
++ struct bfq_data *bfqd = bic_to_bfqd(bic); |
1214 |
++ return bfqd->root_group; |
1215 |
++} |
1216 |
++ |
1217 |
++static inline void bfq_bfqq_move(struct bfq_data *bfqd, |
1218 |
++ struct bfq_queue *bfqq, |
1219 |
++ struct bfq_entity *entity, |
1220 |
++ struct bfq_group *bfqg) |
1221 |
++{ |
1222 |
++} |
1223 |
++ |
1224 |
++static void bfq_end_raising_async(struct bfq_data *bfqd) |
1225 |
++{ |
1226 |
++ bfq_end_raising_async_queues(bfqd, bfqd->root_group); |
1227 |
++} |
1228 |
++ |
1229 |
++static inline void bfq_disconnect_groups(struct bfq_data *bfqd) |
1230 |
++{ |
1231 |
++ bfq_put_async_queues(bfqd, bfqd->root_group); |
1232 |
++} |
1233 |
++ |
1234 |
++static inline void bfq_free_root_group(struct bfq_data *bfqd) |
1235 |
++{ |
1236 |
++ kfree(bfqd->root_group); |
1237 |
++} |
1238 |
++ |
1239 |
++static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
1240 |
++{ |
1241 |
++ struct bfq_group *bfqg; |
1242 |
++ int i; |
1243 |
++ |
1244 |
++ bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); |
1245 |
++ if (bfqg == NULL) |
1246 |
++ return NULL; |
1247 |
++ |
1248 |
++ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
1249 |
++ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
1250 |
++ |
1251 |
++ return bfqg; |
1252 |
++} |
1253 |
++#endif |
1254 |
+diff --git a/block/bfq-ioc.c b/block/bfq-ioc.c |
1255 |
+new file mode 100644 |
1256 |
+index 0000000..7f6b000 |
1257 |
+--- /dev/null |
1258 |
++++ b/block/bfq-ioc.c |
1259 |
+@@ -0,0 +1,36 @@ |
1260 |
++/* |
1261 |
++ * BFQ: I/O context handling. |
1262 |
++ * |
1263 |
++ * Based on ideas and code from CFQ: |
1264 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
1265 |
++ * |
1266 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
1267 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
1268 |
++ * |
1269 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
1270 |
++ */ |
1271 |
++ |
1272 |
++/** |
1273 |
++ * icq_to_bic - convert iocontext queue structure to bfq_io_cq. |
1274 |
++ * @icq: the iocontext queue. |
1275 |
++ */ |
1276 |
++static inline struct bfq_io_cq *icq_to_bic(struct io_cq *icq) |
1277 |
++{ |
1278 |
++ /* bic->icq is the first member, %NULL will convert to %NULL */ |
1279 |
++ return container_of(icq, struct bfq_io_cq, icq); |
1280 |
++} |
1281 |
++ |
1282 |
++/** |
1283 |
++ * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. |
1284 |
++ * @bfqd: the lookup key. |
1285 |
++ * @ioc: the io_context of the process doing I/O. |
1286 |
++ * |
1287 |
++ * Queue lock must be held. |
1288 |
++ */ |
1289 |
++static inline struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd, |
1290 |
++ struct io_context *ioc) |
1291 |
++{ |
1292 |
++ if (ioc) |
1293 |
++ return icq_to_bic(ioc_lookup_icq(ioc, bfqd->queue)); |
1294 |
++ return NULL; |
1295 |
++} |
1296 |
+diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c |
1297 |
+new file mode 100644 |
1298 |
+index 0000000..7670400 |
1299 |
+--- /dev/null |
1300 |
++++ b/block/bfq-iosched.c |
1301 |
+@@ -0,0 +1,3268 @@ |
1302 |
++/* |
1303 |
++ * BFQ, or Budget Fair Queueing, disk scheduler. |
1304 |
++ * |
1305 |
++ * Based on ideas and code from CFQ: |
1306 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
1307 |
++ * |
1308 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
1309 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
1310 |
++ * |
1311 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
1312 |
++ * |
1313 |
++ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
1314 |
++ * |
1315 |
++ * BFQ is a proportional share disk scheduling algorithm based on the |
1316 |
++ * slice-by-slice service scheme of CFQ. But BFQ assigns budgets, measured in |
1317 |
++ * number of sectors, to tasks instead of time slices. The disk is not granted |
1318 |
++ * to the in-service task for a given time slice, but until it has exahusted |
1319 |
++ * its assigned budget. This change from the time to the service domain allows |
1320 |
++ * BFQ to distribute the disk bandwidth among tasks as desired, without any |
1321 |
++ * distortion due to ZBR, workload fluctuations or other factors. BFQ uses an |
1322 |
++ * ad hoc internal scheduler, called B-WF2Q+, to schedule tasks according to |
1323 |
++ * their budgets (more precisely BFQ schedules queues associated to tasks). |
1324 |
++ * Thanks to this accurate scheduler, BFQ can afford to assign high budgets to |
1325 |
++ * disk-bound non-seeky tasks (to boost the throughput), and yet guarantee low |
1326 |
++ * latencies to interactive and soft real-time applications. |
1327 |
++ * |
1328 |
++ * BFQ is described in [1], where also a reference to the initial, more |
1329 |
++ * theoretical paper on BFQ can be found. The interested reader can find in |
1330 |
++ * the latter paper full details on the main algorithm as well as formulas of |
1331 |
++ * the guarantees, plus formal proofs of all the properties. With respect to |
1332 |
++ * the version of BFQ presented in these papers, this implementation adds a |
1333 |
++ * few more heuristics, such as the one that guarantees a low latency to soft |
1334 |
++ * real-time applications, and a hierarchical extension based on H-WF2Q+. |
1335 |
++ * |
1336 |
++ * B-WF2Q+ is based on WF2Q+, that is described in [2], together with |
1337 |
++ * H-WF2Q+, while the augmented tree used to implement B-WF2Q+ with O(log N) |
1338 |
++ * complexity derives from the one introduced with EEVDF in [3]. |
1339 |
++ * |
1340 |
++ * [1] P. Valente and M. Andreolini, ``Improving Application Responsiveness |
1341 |
++ * with the BFQ Disk I/O Scheduler'', |
1342 |
++ * Proceedings of the 5th Annual International Systems and Storage |
1343 |
++ * Conference (SYSTOR '12), June 2012. |
1344 |
++ * |
1345 |
++ * http://algogroup.unimo.it/people/paolo/disk_sched/bf1-v1-suite-results.pdf |
1346 |
++ * |
1347 |
++ * [2] Jon C.R. Bennett and H. Zhang, ``Hierarchical Packet Fair Queueing |
1348 |
++ * Algorithms,'' IEEE/ACM Transactions on Networking, 5(5):675-689, |
1349 |
++ * Oct 1997. |
1350 |
++ * |
1351 |
++ * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz |
1352 |
++ * |
1353 |
++ * [3] I. Stoica and H. Abdel-Wahab, ``Earliest Eligible Virtual Deadline |
1354 |
++ * First: A Flexible and Accurate Mechanism for Proportional Share |
1355 |
++ * Resource Allocation,'' technical report. |
1356 |
++ * |
1357 |
++ * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf |
1358 |
++ */ |
1359 |
++#include <linux/module.h> |
1360 |
++#include <linux/slab.h> |
1361 |
++#include <linux/blkdev.h> |
1362 |
++#include <linux/cgroup.h> |
1363 |
++#include <linux/elevator.h> |
1364 |
++#include <linux/jiffies.h> |
1365 |
++#include <linux/rbtree.h> |
1366 |
++#include <linux/ioprio.h> |
1367 |
++#include "bfq.h" |
1368 |
++#include "blk.h" |
1369 |
++ |
1370 |
++/* Max number of dispatches in one round of service. */ |
1371 |
++static const int bfq_quantum = 4; |
1372 |
++ |
1373 |
++/* Expiration time of sync (0) and async (1) requests, in jiffies. */ |
1374 |
++static const int bfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
1375 |
++ |
1376 |
++/* Maximum backwards seek, in KiB. */ |
1377 |
++static const int bfq_back_max = 16 * 1024; |
1378 |
++ |
1379 |
++/* Penalty of a backwards seek, in number of sectors. */ |
1380 |
++static const int bfq_back_penalty = 2; |
1381 |
++ |
1382 |
++/* Idling period duration, in jiffies. */ |
1383 |
++static int bfq_slice_idle = HZ / 125; |
1384 |
++ |
1385 |
++/* Default maximum budget values, in sectors and number of requests. */ |
1386 |
++static const int bfq_default_max_budget = 16 * 1024; |
1387 |
++static const int bfq_max_budget_async_rq = 4; |
1388 |
++ |
1389 |
++/* |
1390 |
++ * Async to sync throughput distribution is controlled as follows: |
1391 |
++ * when an async request is served, the entity is charged the number |
1392 |
++ * of sectors of the request, multipled by the factor below |
1393 |
++ */ |
1394 |
++static const int bfq_async_charge_factor = 10; |
1395 |
++ |
1396 |
++/* Default timeout values, in jiffies, approximating CFQ defaults. */ |
1397 |
++static const int bfq_timeout_sync = HZ / 8; |
1398 |
++static int bfq_timeout_async = HZ / 25; |
1399 |
++ |
1400 |
++struct kmem_cache *bfq_pool; |
1401 |
++ |
1402 |
++/* Below this threshold (in ms), we consider thinktime immediate. */ |
1403 |
++#define BFQ_MIN_TT 2 |
1404 |
++ |
1405 |
++/* hw_tag detection: parallel requests threshold and min samples needed. */ |
1406 |
++#define BFQ_HW_QUEUE_THRESHOLD 4 |
1407 |
++#define BFQ_HW_QUEUE_SAMPLES 32 |
1408 |
++ |
1409 |
++#define BFQQ_SEEK_THR (sector_t)(8 * 1024) |
1410 |
++#define BFQQ_SEEKY(bfqq) ((bfqq)->seek_mean > BFQQ_SEEK_THR) |
1411 |
++ |
1412 |
++/* Min samples used for peak rate estimation (for autotuning). */ |
1413 |
++#define BFQ_PEAK_RATE_SAMPLES 32 |
1414 |
++ |
1415 |
++/* Shift used for peak rate fixed precision calculations. */ |
1416 |
++#define BFQ_RATE_SHIFT 16 |
1417 |
++ |
1418 |
++/* |
1419 |
++ * The duration of the weight raising for interactive applications is |
1420 |
++ * computed automatically (as default behaviour), using the following |
1421 |
++ * formula: duration = (R / r) * T, where r is the peak rate of the |
1422 |
++ * disk, and R and T are two reference parameters. In particular, R is |
1423 |
++ * the peak rate of a reference disk, and T is about the maximum time |
1424 |
++ * for starting popular large applications on that disk, under BFQ and |
1425 |
++ * while reading two files in parallel. Finally, BFQ uses two |
1426 |
++ * different pairs (R, T) depending on whether the disk is rotational |
1427 |
++ * or non-rotational. |
1428 |
++ */ |
1429 |
++#define T_rot (msecs_to_jiffies(5500)) |
1430 |
++#define T_nonrot (msecs_to_jiffies(2000)) |
1431 |
++/* Next two quantities are in sectors/usec, left-shifted by BFQ_RATE_SHIFT */ |
1432 |
++#define R_rot 17415 |
1433 |
++#define R_nonrot 34791 |
1434 |
++ |
1435 |
++#define BFQ_SERVICE_TREE_INIT ((struct bfq_service_tree) \ |
1436 |
++ { RB_ROOT, RB_ROOT, NULL, NULL, 0, 0 }) |
1437 |
++ |
1438 |
++#define RQ_BIC(rq) ((struct bfq_io_cq *) (rq)->elv.priv[0]) |
1439 |
++#define RQ_BFQQ(rq) ((rq)->elv.priv[1]) |
1440 |
++ |
1441 |
++static inline void bfq_schedule_dispatch(struct bfq_data *bfqd); |
1442 |
++ |
1443 |
++#include "bfq-ioc.c" |
1444 |
++#include "bfq-sched.c" |
1445 |
++#include "bfq-cgroup.c" |
1446 |
++ |
1447 |
++#define bfq_class_idle(bfqq) ((bfqq)->entity.ioprio_class ==\ |
1448 |
++ IOPRIO_CLASS_IDLE) |
1449 |
++#define bfq_class_rt(bfqq) ((bfqq)->entity.ioprio_class ==\ |
1450 |
++ IOPRIO_CLASS_RT) |
1451 |
++ |
1452 |
++#define bfq_sample_valid(samples) ((samples) > 80) |
1453 |
++ |
1454 |
++/* |
1455 |
++ * We regard a request as SYNC, if either it's a read or has the SYNC bit |
1456 |
++ * set (in which case it could also be a direct WRITE). |
1457 |
++ */ |
1458 |
++static inline int bfq_bio_sync(struct bio *bio) |
1459 |
++{ |
1460 |
++ if (bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC)) |
1461 |
++ return 1; |
1462 |
++ |
1463 |
++ return 0; |
1464 |
++} |
1465 |
++ |
1466 |
++/* |
1467 |
++ * Scheduler run of queue, if there are requests pending and no one in the |
1468 |
++ * driver that will restart queueing. |
1469 |
++ */ |
1470 |
++static inline void bfq_schedule_dispatch(struct bfq_data *bfqd) |
1471 |
++{ |
1472 |
++ if (bfqd->queued != 0) { |
1473 |
++ bfq_log(bfqd, "schedule dispatch"); |
1474 |
++ kblockd_schedule_work(bfqd->queue, &bfqd->unplug_work); |
1475 |
++ } |
1476 |
++} |
1477 |
++ |
1478 |
++/* |
1479 |
++ * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1480 |
++ * We choose the request that is closesr to the head right now. Distance |
1481 |
++ * behind the head is penalized and only allowed to a certain extent. |
1482 |
++ */ |
1483 |
++static struct request *bfq_choose_req(struct bfq_data *bfqd, |
1484 |
++ struct request *rq1, |
1485 |
++ struct request *rq2, |
1486 |
++ sector_t last) |
1487 |
++{ |
1488 |
++ sector_t s1, s2, d1 = 0, d2 = 0; |
1489 |
++ unsigned long back_max; |
1490 |
++#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
1491 |
++#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ |
1492 |
++ unsigned wrap = 0; /* bit mask: requests behind the disk head? */ |
1493 |
++ |
1494 |
++ if (rq1 == NULL || rq1 == rq2) |
1495 |
++ return rq2; |
1496 |
++ if (rq2 == NULL) |
1497 |
++ return rq1; |
1498 |
++ |
1499 |
++ if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
1500 |
++ return rq1; |
1501 |
++ else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) |
1502 |
++ return rq2; |
1503 |
++ if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) |
1504 |
++ return rq1; |
1505 |
++ else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) |
1506 |
++ return rq2; |
1507 |
++ |
1508 |
++ s1 = blk_rq_pos(rq1); |
1509 |
++ s2 = blk_rq_pos(rq2); |
1510 |
++ |
1511 |
++ /* |
1512 |
++ * By definition, 1KiB is 2 sectors. |
1513 |
++ */ |
1514 |
++ back_max = bfqd->bfq_back_max * 2; |
1515 |
++ |
1516 |
++ /* |
1517 |
++ * Strict one way elevator _except_ in the case where we allow |
1518 |
++ * short backward seeks which are biased as twice the cost of a |
1519 |
++ * similar forward seek. |
1520 |
++ */ |
1521 |
++ if (s1 >= last) |
1522 |
++ d1 = s1 - last; |
1523 |
++ else if (s1 + back_max >= last) |
1524 |
++ d1 = (last - s1) * bfqd->bfq_back_penalty; |
1525 |
++ else |
1526 |
++ wrap |= BFQ_RQ1_WRAP; |
1527 |
++ |
1528 |
++ if (s2 >= last) |
1529 |
++ d2 = s2 - last; |
1530 |
++ else if (s2 + back_max >= last) |
1531 |
++ d2 = (last - s2) * bfqd->bfq_back_penalty; |
1532 |
++ else |
1533 |
++ wrap |= BFQ_RQ2_WRAP; |
1534 |
++ |
1535 |
++ /* Found required data */ |
1536 |
++ |
1537 |
++ /* |
1538 |
++ * By doing switch() on the bit mask "wrap" we avoid having to |
1539 |
++ * check two variables for all permutations: --> faster! |
1540 |
++ */ |
1541 |
++ switch (wrap) { |
1542 |
++ case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
1543 |
++ if (d1 < d2) |
1544 |
++ return rq1; |
1545 |
++ else if (d2 < d1) |
1546 |
++ return rq2; |
1547 |
++ else { |
1548 |
++ if (s1 >= s2) |
1549 |
++ return rq1; |
1550 |
++ else |
1551 |
++ return rq2; |
1552 |
++ } |
1553 |
++ |
1554 |
++ case BFQ_RQ2_WRAP: |
1555 |
++ return rq1; |
1556 |
++ case BFQ_RQ1_WRAP: |
1557 |
++ return rq2; |
1558 |
++ case (BFQ_RQ1_WRAP|BFQ_RQ2_WRAP): /* both rqs wrapped */ |
1559 |
++ default: |
1560 |
++ /* |
1561 |
++ * Since both rqs are wrapped, |
1562 |
++ * start with the one that's further behind head |
1563 |
++ * (--> only *one* back seek required), |
1564 |
++ * since back seek takes more time than forward. |
1565 |
++ */ |
1566 |
++ if (s1 <= s2) |
1567 |
++ return rq1; |
1568 |
++ else |
1569 |
++ return rq2; |
1570 |
++ } |
1571 |
++} |
1572 |
++ |
1573 |
++static struct bfq_queue * |
1574 |
++bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root, |
1575 |
++ sector_t sector, struct rb_node **ret_parent, |
1576 |
++ struct rb_node ***rb_link) |
1577 |
++{ |
1578 |
++ struct rb_node **p, *parent; |
1579 |
++ struct bfq_queue *bfqq = NULL; |
1580 |
++ |
1581 |
++ parent = NULL; |
1582 |
++ p = &root->rb_node; |
1583 |
++ while (*p) { |
1584 |
++ struct rb_node **n; |
1585 |
++ |
1586 |
++ parent = *p; |
1587 |
++ bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
1588 |
++ |
1589 |
++ /* |
1590 |
++ * Sort strictly based on sector. Smallest to the left, |
1591 |
++ * largest to the right. |
1592 |
++ */ |
1593 |
++ if (sector > blk_rq_pos(bfqq->next_rq)) |
1594 |
++ n = &(*p)->rb_right; |
1595 |
++ else if (sector < blk_rq_pos(bfqq->next_rq)) |
1596 |
++ n = &(*p)->rb_left; |
1597 |
++ else |
1598 |
++ break; |
1599 |
++ p = n; |
1600 |
++ bfqq = NULL; |
1601 |
++ } |
1602 |
++ |
1603 |
++ *ret_parent = parent; |
1604 |
++ if (rb_link) |
1605 |
++ *rb_link = p; |
1606 |
++ |
1607 |
++ bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d", |
1608 |
++ (long long unsigned)sector, |
1609 |
++ bfqq != NULL ? bfqq->pid : 0); |
1610 |
++ |
1611 |
++ return bfqq; |
1612 |
++} |
1613 |
++ |
1614 |
++static void bfq_rq_pos_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
1615 |
++{ |
1616 |
++ struct rb_node **p, *parent; |
1617 |
++ struct bfq_queue *__bfqq; |
1618 |
++ |
1619 |
++ if (bfqq->pos_root != NULL) { |
1620 |
++ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
1621 |
++ bfqq->pos_root = NULL; |
1622 |
++ } |
1623 |
++ |
1624 |
++ if (bfq_class_idle(bfqq)) |
1625 |
++ return; |
1626 |
++ if (!bfqq->next_rq) |
1627 |
++ return; |
1628 |
++ |
1629 |
++ bfqq->pos_root = &bfqd->rq_pos_tree; |
1630 |
++ __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root, |
1631 |
++ blk_rq_pos(bfqq->next_rq), &parent, &p); |
1632 |
++ if (__bfqq == NULL) { |
1633 |
++ rb_link_node(&bfqq->pos_node, parent, p); |
1634 |
++ rb_insert_color(&bfqq->pos_node, bfqq->pos_root); |
1635 |
++ } else |
1636 |
++ bfqq->pos_root = NULL; |
1637 |
++} |
1638 |
++ |
1639 |
++static struct request *bfq_find_next_rq(struct bfq_data *bfqd, |
1640 |
++ struct bfq_queue *bfqq, |
1641 |
++ struct request *last) |
1642 |
++{ |
1643 |
++ struct rb_node *rbnext = rb_next(&last->rb_node); |
1644 |
++ struct rb_node *rbprev = rb_prev(&last->rb_node); |
1645 |
++ struct request *next = NULL, *prev = NULL; |
1646 |
++ |
1647 |
++ BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1648 |
++ |
1649 |
++ if (rbprev != NULL) |
1650 |
++ prev = rb_entry_rq(rbprev); |
1651 |
++ |
1652 |
++ if (rbnext != NULL) |
1653 |
++ next = rb_entry_rq(rbnext); |
1654 |
++ else { |
1655 |
++ rbnext = rb_first(&bfqq->sort_list); |
1656 |
++ if (rbnext && rbnext != &last->rb_node) |
1657 |
++ next = rb_entry_rq(rbnext); |
1658 |
++ } |
1659 |
++ |
1660 |
++ return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); |
1661 |
++} |
1662 |
++ |
1663 |
++static void bfq_del_rq_rb(struct request *rq) |
1664 |
++{ |
1665 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
1666 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
1667 |
++ const int sync = rq_is_sync(rq); |
1668 |
++ |
1669 |
++ BUG_ON(bfqq->queued[sync] == 0); |
1670 |
++ bfqq->queued[sync]--; |
1671 |
++ bfqd->queued--; |
1672 |
++ |
1673 |
++ elv_rb_del(&bfqq->sort_list, rq); |
1674 |
++ |
1675 |
++ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
1676 |
++ if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) |
1677 |
++ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
1678 |
++ /* |
1679 |
++ * Remove queue from request-position tree as it is empty. |
1680 |
++ */ |
1681 |
++ if (bfqq->pos_root != NULL) { |
1682 |
++ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
1683 |
++ bfqq->pos_root = NULL; |
1684 |
++ } |
1685 |
++ } |
1686 |
++} |
1687 |
++ |
1688 |
++/* see the definition of bfq_async_charge_factor for details */ |
1689 |
++static inline unsigned long bfq_serv_to_charge(struct request *rq, |
1690 |
++ struct bfq_queue *bfqq) |
1691 |
++{ |
1692 |
++ return blk_rq_sectors(rq) * |
1693 |
++ (1 + ((!bfq_bfqq_sync(bfqq)) * (bfqq->raising_coeff == 1) * |
1694 |
++ bfq_async_charge_factor)); |
1695 |
++} |
1696 |
++ |
1697 |
++/** |
1698 |
++ * bfq_updated_next_req - update the queue after a new next_rq selection. |
1699 |
++ * @bfqd: the device data the queue belongs to. |
1700 |
++ * @bfqq: the queue to update. |
1701 |
++ * |
1702 |
++ * If the first request of a queue changes we make sure that the queue |
1703 |
++ * has enough budget to serve at least its first request (if the |
1704 |
++ * request has grown). We do this because if the queue has not enough |
1705 |
++ * budget for its first request, it has to go through two dispatch |
1706 |
++ * rounds to actually get it dispatched. |
1707 |
++ */ |
1708 |
++static void bfq_updated_next_req(struct bfq_data *bfqd, |
1709 |
++ struct bfq_queue *bfqq) |
1710 |
++{ |
1711 |
++ struct bfq_entity *entity = &bfqq->entity; |
1712 |
++ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
1713 |
++ struct request *next_rq = bfqq->next_rq; |
1714 |
++ unsigned long new_budget; |
1715 |
++ |
1716 |
++ if (next_rq == NULL) |
1717 |
++ return; |
1718 |
++ |
1719 |
++ if (bfqq == bfqd->in_service_queue) |
1720 |
++ /* |
1721 |
++ * In order not to break guarantees, budgets cannot be |
1722 |
++ * changed after an entity has been selected. |
1723 |
++ */ |
1724 |
++ return; |
1725 |
++ |
1726 |
++ BUG_ON(entity->tree != &st->active); |
1727 |
++ BUG_ON(entity == entity->sched_data->active_entity); |
1728 |
++ |
1729 |
++ new_budget = max_t(unsigned long, bfqq->max_budget, |
1730 |
++ bfq_serv_to_charge(next_rq, bfqq)); |
1731 |
++ entity->budget = new_budget; |
1732 |
++ bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", new_budget); |
1733 |
++ bfq_activate_bfqq(bfqd, bfqq); |
1734 |
++} |
1735 |
++ |
1736 |
++static inline unsigned int bfq_wrais_duration(struct bfq_data *bfqd) |
1737 |
++{ |
1738 |
++ u64 dur; |
1739 |
++ |
1740 |
++ if (bfqd->bfq_raising_max_time > 0) |
1741 |
++ return bfqd->bfq_raising_max_time; |
1742 |
++ |
1743 |
++ dur = bfqd->RT_prod; |
1744 |
++ do_div(dur, bfqd->peak_rate); |
1745 |
++ |
1746 |
++ return dur; |
1747 |
++} |
1748 |
++ |
1749 |
++static void bfq_add_rq_rb(struct request *rq) |
1750 |
++{ |
1751 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
1752 |
++ struct bfq_entity *entity = &bfqq->entity; |
1753 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
1754 |
++ struct request *next_rq, *prev; |
1755 |
++ unsigned long old_raising_coeff = bfqq->raising_coeff; |
1756 |
++ int idle_for_long_time = 0; |
1757 |
++ |
1758 |
++ bfq_log_bfqq(bfqd, bfqq, "add_rq_rb %d", rq_is_sync(rq)); |
1759 |
++ bfqq->queued[rq_is_sync(rq)]++; |
1760 |
++ bfqd->queued++; |
1761 |
++ |
1762 |
++ elv_rb_add(&bfqq->sort_list, rq); |
1763 |
++ |
1764 |
++ /* |
1765 |
++ * Check if this request is a better next-serve candidate. |
1766 |
++ */ |
1767 |
++ prev = bfqq->next_rq; |
1768 |
++ next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); |
1769 |
++ BUG_ON(next_rq == NULL); |
1770 |
++ bfqq->next_rq = next_rq; |
1771 |
++ |
1772 |
++ /* |
1773 |
++ * Adjust priority tree position, if next_rq changes. |
1774 |
++ */ |
1775 |
++ if (prev != bfqq->next_rq) |
1776 |
++ bfq_rq_pos_tree_add(bfqd, bfqq); |
1777 |
++ |
1778 |
++ if (!bfq_bfqq_busy(bfqq)) { |
1779 |
++ int soft_rt = bfqd->bfq_raising_max_softrt_rate > 0 && |
1780 |
++ time_is_before_jiffies(bfqq->soft_rt_next_start); |
1781 |
++ idle_for_long_time = time_is_before_jiffies( |
1782 |
++ bfqq->budget_timeout + |
1783 |
++ bfqd->bfq_raising_min_idle_time); |
1784 |
++ entity->budget = max_t(unsigned long, bfqq->max_budget, |
1785 |
++ bfq_serv_to_charge(next_rq, bfqq)); |
1786 |
++ |
1787 |
++ if (!bfqd->low_latency) |
1788 |
++ goto add_bfqq_busy; |
1789 |
++ |
1790 |
++ /* |
1791 |
++ * If the queue is not being boosted and has been idle |
1792 |
++ * for enough time, start a weight-raising period |
1793 |
++ */ |
1794 |
++ if (old_raising_coeff == 1 && |
1795 |
++ (idle_for_long_time || soft_rt)) { |
1796 |
++ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
1797 |
++ if (idle_for_long_time) |
1798 |
++ bfqq->raising_cur_max_time = |
1799 |
++ bfq_wrais_duration(bfqd); |
1800 |
++ else |
1801 |
++ bfqq->raising_cur_max_time = |
1802 |
++ bfqd->bfq_raising_rt_max_time; |
1803 |
++ bfq_log_bfqq(bfqd, bfqq, |
1804 |
++ "wrais starting at %llu msec," |
1805 |
++ "rais_max_time %u", |
1806 |
++ bfqq->last_rais_start_finish, |
1807 |
++ jiffies_to_msecs(bfqq-> |
1808 |
++ raising_cur_max_time)); |
1809 |
++ } else if (old_raising_coeff > 1) { |
1810 |
++ if (idle_for_long_time) |
1811 |
++ bfqq->raising_cur_max_time = |
1812 |
++ bfq_wrais_duration(bfqd); |
1813 |
++ else if (bfqq->raising_cur_max_time == |
1814 |
++ bfqd->bfq_raising_rt_max_time && |
1815 |
++ !soft_rt) { |
1816 |
++ bfqq->raising_coeff = 1; |
1817 |
++ bfq_log_bfqq(bfqd, bfqq, |
1818 |
++ "wrais ending at %llu msec," |
1819 |
++ "rais_max_time %u", |
1820 |
++ bfqq->last_rais_start_finish, |
1821 |
++ jiffies_to_msecs(bfqq-> |
1822 |
++ raising_cur_max_time)); |
1823 |
++ } else if ((bfqq->last_rais_start_finish + |
1824 |
++ bfqq->raising_cur_max_time < |
1825 |
++ jiffies + bfqd->bfq_raising_rt_max_time) && |
1826 |
++ soft_rt) { |
1827 |
++ /* |
1828 |
++ * |
1829 |
++ * The remaining weight-raising time is lower |
1830 |
++ * than bfqd->bfq_raising_rt_max_time, which |
1831 |
++ * means that the application is enjoying |
1832 |
++ * weight raising either because deemed soft rt |
1833 |
++ * in the near past, or because deemed |
1834 |
++ * interactive a long ago. In both cases, |
1835 |
++ * resetting now the current remaining weight- |
1836 |
++ * raising time for the application to the |
1837 |
++ * weight-raising duration for soft rt |
1838 |
++ * applications would not cause any latency |
1839 |
++ * increase for the application (as the new |
1840 |
++ * duration would be higher than the remaining |
1841 |
++ * time). |
1842 |
++ * |
1843 |
++ * In addition, the application is now meeting |
1844 |
++ * the requirements for being deemed soft rt. |
1845 |
++ * In the end we can correctly and safely |
1846 |
++ * (re)charge the weight-raising duration for |
1847 |
++ * the application with the weight-raising |
1848 |
++ * duration for soft rt applications. |
1849 |
++ * |
1850 |
++ * In particular, doing this recharge now, i.e., |
1851 |
++ * before the weight-raising period for the |
1852 |
++ * application finishes, reduces the probability |
1853 |
++ * of the following negative scenario: |
1854 |
++ * 1) the weight of a soft rt application is |
1855 |
++ * raised at startup (as for any newly |
1856 |
++ * created application), |
1857 |
++ * 2) since the application is not interactive, |
1858 |
++ * at a certain time weight-raising is |
1859 |
++ * stopped for the application, |
1860 |
++ * 3) at that time the application happens to |
1861 |
++ * still have pending requests, and hence |
1862 |
++ * is destined to not have a chance to be |
1863 |
++ * deemed soft rt before these requests are |
1864 |
++ * completed (see the comments to the |
1865 |
++ * function bfq_bfqq_softrt_next_start() |
1866 |
++ * for details on soft rt detection), |
1867 |
++ * 4) these pending requests experience a high |
1868 |
++ * latency because the application is not |
1869 |
++ * weight-raised while they are pending. |
1870 |
++ */ |
1871 |
++ bfqq->last_rais_start_finish = jiffies; |
1872 |
++ bfqq->raising_cur_max_time = |
1873 |
++ bfqd->bfq_raising_rt_max_time; |
1874 |
++ } |
1875 |
++ } |
1876 |
++ if (old_raising_coeff != bfqq->raising_coeff) |
1877 |
++ entity->ioprio_changed = 1; |
1878 |
++add_bfqq_busy: |
1879 |
++ bfqq->last_idle_bklogged = jiffies; |
1880 |
++ bfqq->service_from_backlogged = 0; |
1881 |
++ bfq_clear_bfqq_softrt_update(bfqq); |
1882 |
++ bfq_add_bfqq_busy(bfqd, bfqq); |
1883 |
++ } else { |
1884 |
++ if (bfqd->low_latency && old_raising_coeff == 1 && |
1885 |
++ !rq_is_sync(rq) && |
1886 |
++ bfqq->last_rais_start_finish + |
1887 |
++ time_is_before_jiffies( |
1888 |
++ bfqd->bfq_raising_min_inter_arr_async)) { |
1889 |
++ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
1890 |
++ bfqq->raising_cur_max_time = bfq_wrais_duration(bfqd); |
1891 |
++ |
1892 |
++ bfqd->raised_busy_queues++; |
1893 |
++ entity->ioprio_changed = 1; |
1894 |
++ bfq_log_bfqq(bfqd, bfqq, |
1895 |
++ "non-idle wrais starting at %llu msec," |
1896 |
++ "rais_max_time %u", |
1897 |
++ bfqq->last_rais_start_finish, |
1898 |
++ jiffies_to_msecs(bfqq-> |
1899 |
++ raising_cur_max_time)); |
1900 |
++ } |
1901 |
++ bfq_updated_next_req(bfqd, bfqq); |
1902 |
++ } |
1903 |
++ |
1904 |
++ if (bfqd->low_latency && |
1905 |
++ (old_raising_coeff == 1 || bfqq->raising_coeff == 1 || |
1906 |
++ idle_for_long_time)) |
1907 |
++ bfqq->last_rais_start_finish = jiffies; |
1908 |
++} |
1909 |
++ |
1910 |
++static void bfq_reposition_rq_rb(struct bfq_queue *bfqq, struct request *rq) |
1911 |
++{ |
1912 |
++ elv_rb_del(&bfqq->sort_list, rq); |
1913 |
++ bfqq->queued[rq_is_sync(rq)]--; |
1914 |
++ bfqq->bfqd->queued--; |
1915 |
++ bfq_add_rq_rb(rq); |
1916 |
++} |
1917 |
++ |
1918 |
++static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, |
1919 |
++ struct bio *bio) |
1920 |
++{ |
1921 |
++ struct task_struct *tsk = current; |
1922 |
++ struct bfq_io_cq *bic; |
1923 |
++ struct bfq_queue *bfqq; |
1924 |
++ |
1925 |
++ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
1926 |
++ if (bic == NULL) |
1927 |
++ return NULL; |
1928 |
++ |
1929 |
++ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
1930 |
++ if (bfqq != NULL) |
1931 |
++ return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio)); |
1932 |
++ |
1933 |
++ return NULL; |
1934 |
++} |
1935 |
++ |
1936 |
++static void bfq_activate_request(struct request_queue *q, struct request *rq) |
1937 |
++{ |
1938 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
1939 |
++ |
1940 |
++ bfqd->rq_in_driver++; |
1941 |
++ bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1942 |
++ bfq_log(bfqd, "activate_request: new bfqd->last_position %llu", |
1943 |
++ (long long unsigned)bfqd->last_position); |
1944 |
++} |
1945 |
++ |
1946 |
++static void bfq_deactivate_request(struct request_queue *q, struct request *rq) |
1947 |
++{ |
1948 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
1949 |
++ |
1950 |
++ WARN_ON(bfqd->rq_in_driver == 0); |
1951 |
++ bfqd->rq_in_driver--; |
1952 |
++} |
1953 |
++ |
1954 |
++static void bfq_remove_request(struct request *rq) |
1955 |
++{ |
1956 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
1957 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
1958 |
++ |
1959 |
++ if (bfqq->next_rq == rq) { |
1960 |
++ bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); |
1961 |
++ bfq_updated_next_req(bfqd, bfqq); |
1962 |
++ } |
1963 |
++ |
1964 |
++ list_del_init(&rq->queuelist); |
1965 |
++ bfq_del_rq_rb(rq); |
1966 |
++ |
1967 |
++ if (rq->cmd_flags & REQ_META) { |
1968 |
++ WARN_ON(bfqq->meta_pending == 0); |
1969 |
++ bfqq->meta_pending--; |
1970 |
++ } |
1971 |
++} |
1972 |
++ |
1973 |
++static int bfq_merge(struct request_queue *q, struct request **req, |
1974 |
++ struct bio *bio) |
1975 |
++{ |
1976 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
1977 |
++ struct request *__rq; |
1978 |
++ |
1979 |
++ __rq = bfq_find_rq_fmerge(bfqd, bio); |
1980 |
++ if (__rq != NULL && elv_rq_merge_ok(__rq, bio)) { |
1981 |
++ *req = __rq; |
1982 |
++ return ELEVATOR_FRONT_MERGE; |
1983 |
++ } |
1984 |
++ |
1985 |
++ return ELEVATOR_NO_MERGE; |
1986 |
++} |
1987 |
++ |
1988 |
++static void bfq_merged_request(struct request_queue *q, struct request *req, |
1989 |
++ int type) |
1990 |
++{ |
1991 |
++ if (type == ELEVATOR_FRONT_MERGE) { |
1992 |
++ struct bfq_queue *bfqq = RQ_BFQQ(req); |
1993 |
++ |
1994 |
++ bfq_reposition_rq_rb(bfqq, req); |
1995 |
++ } |
1996 |
++} |
1997 |
++ |
1998 |
++static void bfq_merged_requests(struct request_queue *q, struct request *rq, |
1999 |
++ struct request *next) |
2000 |
++{ |
2001 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
2002 |
++ |
2003 |
++ /* |
2004 |
++ * Reposition in fifo if next is older than rq. |
2005 |
++ */ |
2006 |
++ if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && |
2007 |
++ time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
2008 |
++ list_move(&rq->queuelist, &next->queuelist); |
2009 |
++ rq_set_fifo_time(rq, rq_fifo_time(next)); |
2010 |
++ } |
2011 |
++ |
2012 |
++ if (bfqq->next_rq == next) |
2013 |
++ bfqq->next_rq = rq; |
2014 |
++ |
2015 |
++ bfq_remove_request(next); |
2016 |
++} |
2017 |
++ |
2018 |
++/* Must be called with bfqq != NULL */ |
2019 |
++static inline void bfq_bfqq_end_raising(struct bfq_queue *bfqq) |
2020 |
++{ |
2021 |
++ BUG_ON(bfqq == NULL); |
2022 |
++ if (bfq_bfqq_busy(bfqq)) |
2023 |
++ bfqq->bfqd->raised_busy_queues--; |
2024 |
++ bfqq->raising_coeff = 1; |
2025 |
++ bfqq->raising_cur_max_time = 0; |
2026 |
++ /* Trigger a weight change on the next activation of the queue */ |
2027 |
++ bfqq->entity.ioprio_changed = 1; |
2028 |
++} |
2029 |
++ |
2030 |
++static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
2031 |
++ struct bfq_group *bfqg) |
2032 |
++{ |
2033 |
++ int i, j; |
2034 |
++ |
2035 |
++ for (i = 0; i < 2; i++) |
2036 |
++ for (j = 0; j < IOPRIO_BE_NR; j++) |
2037 |
++ if (bfqg->async_bfqq[i][j] != NULL) |
2038 |
++ bfq_bfqq_end_raising(bfqg->async_bfqq[i][j]); |
2039 |
++ if (bfqg->async_idle_bfqq != NULL) |
2040 |
++ bfq_bfqq_end_raising(bfqg->async_idle_bfqq); |
2041 |
++} |
2042 |
++ |
2043 |
++static void bfq_end_raising(struct bfq_data *bfqd) |
2044 |
++{ |
2045 |
++ struct bfq_queue *bfqq; |
2046 |
++ |
2047 |
++ spin_lock_irq(bfqd->queue->queue_lock); |
2048 |
++ |
2049 |
++ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) |
2050 |
++ bfq_bfqq_end_raising(bfqq); |
2051 |
++ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) |
2052 |
++ bfq_bfqq_end_raising(bfqq); |
2053 |
++ bfq_end_raising_async(bfqd); |
2054 |
++ |
2055 |
++ spin_unlock_irq(bfqd->queue->queue_lock); |
2056 |
++} |
2057 |
++ |
2058 |
++static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
2059 |
++ struct bio *bio) |
2060 |
++{ |
2061 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
2062 |
++ struct bfq_io_cq *bic; |
2063 |
++ struct bfq_queue *bfqq; |
2064 |
++ |
2065 |
++ /* |
2066 |
++ * Disallow merge of a sync bio into an async request. |
2067 |
++ */ |
2068 |
++ if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
2069 |
++ return 0; |
2070 |
++ |
2071 |
++ /* |
2072 |
++ * Lookup the bfqq that this bio will be queued with. Allow |
2073 |
++ * merge only if rq is queued there. |
2074 |
++ * Queue lock is held here. |
2075 |
++ */ |
2076 |
++ bic = bfq_bic_lookup(bfqd, current->io_context); |
2077 |
++ if (bic == NULL) |
2078 |
++ return 0; |
2079 |
++ |
2080 |
++ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
2081 |
++ return bfqq == RQ_BFQQ(rq); |
2082 |
++} |
2083 |
++ |
2084 |
++static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
2085 |
++ struct bfq_queue *bfqq) |
2086 |
++{ |
2087 |
++ if (bfqq != NULL) { |
2088 |
++ bfq_mark_bfqq_must_alloc(bfqq); |
2089 |
++ bfq_mark_bfqq_budget_new(bfqq); |
2090 |
++ bfq_clear_bfqq_fifo_expire(bfqq); |
2091 |
++ |
2092 |
++ bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
2093 |
++ |
2094 |
++ bfq_log_bfqq(bfqd, bfqq, |
2095 |
++ "set_in_service_queue, cur-budget = %lu", |
2096 |
++ bfqq->entity.budget); |
2097 |
++ } |
2098 |
++ |
2099 |
++ bfqd->in_service_queue = bfqq; |
2100 |
++} |
2101 |
++ |
2102 |
++/* |
2103 |
++ * Get and set a new queue for service. |
2104 |
++ */ |
2105 |
++static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd, |
2106 |
++ struct bfq_queue *bfqq) |
2107 |
++{ |
2108 |
++ if (!bfqq) |
2109 |
++ bfqq = bfq_get_next_queue(bfqd); |
2110 |
++ else |
2111 |
++ bfq_get_next_queue_forced(bfqd, bfqq); |
2112 |
++ |
2113 |
++ __bfq_set_in_service_queue(bfqd, bfqq); |
2114 |
++ return bfqq; |
2115 |
++} |
2116 |
++ |
2117 |
++static inline sector_t bfq_dist_from_last(struct bfq_data *bfqd, |
2118 |
++ struct request *rq) |
2119 |
++{ |
2120 |
++ if (blk_rq_pos(rq) >= bfqd->last_position) |
2121 |
++ return blk_rq_pos(rq) - bfqd->last_position; |
2122 |
++ else |
2123 |
++ return bfqd->last_position - blk_rq_pos(rq); |
2124 |
++} |
2125 |
++ |
2126 |
++/* |
2127 |
++ * Return true if bfqq has no request pending and rq is close enough to |
2128 |
++ * bfqd->last_position, or if rq is closer to bfqd->last_position than |
2129 |
++ * bfqq->next_rq |
2130 |
++ */ |
2131 |
++static inline int bfq_rq_close(struct bfq_data *bfqd, struct request *rq) |
2132 |
++{ |
2133 |
++ return bfq_dist_from_last(bfqd, rq) <= BFQQ_SEEK_THR; |
2134 |
++} |
2135 |
++ |
2136 |
++static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
2137 |
++{ |
2138 |
++ struct rb_root *root = &bfqd->rq_pos_tree; |
2139 |
++ struct rb_node *parent, *node; |
2140 |
++ struct bfq_queue *__bfqq; |
2141 |
++ sector_t sector = bfqd->last_position; |
2142 |
++ |
2143 |
++ if (RB_EMPTY_ROOT(root)) |
2144 |
++ return NULL; |
2145 |
++ |
2146 |
++ /* |
2147 |
++ * First, if we find a request starting at the end of the last |
2148 |
++ * request, choose it. |
2149 |
++ */ |
2150 |
++ __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL); |
2151 |
++ if (__bfqq != NULL) |
2152 |
++ return __bfqq; |
2153 |
++ |
2154 |
++ /* |
2155 |
++ * If the exact sector wasn't found, the parent of the NULL leaf |
2156 |
++ * will contain the closest sector (rq_pos_tree sorted by next_request |
2157 |
++ * position). |
2158 |
++ */ |
2159 |
++ __bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
2160 |
++ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
2161 |
++ return __bfqq; |
2162 |
++ |
2163 |
++ if (blk_rq_pos(__bfqq->next_rq) < sector) |
2164 |
++ node = rb_next(&__bfqq->pos_node); |
2165 |
++ else |
2166 |
++ node = rb_prev(&__bfqq->pos_node); |
2167 |
++ if (node == NULL) |
2168 |
++ return NULL; |
2169 |
++ |
2170 |
++ __bfqq = rb_entry(node, struct bfq_queue, pos_node); |
2171 |
++ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
2172 |
++ return __bfqq; |
2173 |
++ |
2174 |
++ return NULL; |
2175 |
++} |
2176 |
++ |
2177 |
++/* |
2178 |
++ * bfqd - obvious |
2179 |
++ * cur_bfqq - passed in so that we don't decide that the current queue |
2180 |
++ * is closely cooperating with itself. |
2181 |
++ * |
2182 |
++ * We are assuming that cur_bfqq has dispatched at least one request, |
2183 |
++ * and that bfqd->last_position reflects a position on the disk associated |
2184 |
++ * with the I/O issued by cur_bfqq. |
2185 |
++ */ |
2186 |
++static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
2187 |
++ struct bfq_queue *cur_bfqq) |
2188 |
++{ |
2189 |
++ struct bfq_queue *bfqq; |
2190 |
++ |
2191 |
++ if (bfq_class_idle(cur_bfqq)) |
2192 |
++ return NULL; |
2193 |
++ if (!bfq_bfqq_sync(cur_bfqq)) |
2194 |
++ return NULL; |
2195 |
++ if (BFQQ_SEEKY(cur_bfqq)) |
2196 |
++ return NULL; |
2197 |
++ |
2198 |
++ /* If device has only one backlogged bfq_queue, don't search. */ |
2199 |
++ if (bfqd->busy_queues == 1) |
2200 |
++ return NULL; |
2201 |
++ |
2202 |
++ /* |
2203 |
++ * We should notice if some of the queues are cooperating, e.g. |
2204 |
++ * working closely on the same area of the disk. In that case, |
2205 |
++ * we can group them together and don't waste time idling. |
2206 |
++ */ |
2207 |
++ bfqq = bfqq_close(bfqd); |
2208 |
++ if (bfqq == NULL || bfqq == cur_bfqq) |
2209 |
++ return NULL; |
2210 |
++ |
2211 |
++ /* |
2212 |
++ * Do not merge queues from different bfq_groups. |
2213 |
++ */ |
2214 |
++ if (bfqq->entity.parent != cur_bfqq->entity.parent) |
2215 |
++ return NULL; |
2216 |
++ |
2217 |
++ /* |
2218 |
++ * It only makes sense to merge sync queues. |
2219 |
++ */ |
2220 |
++ if (!bfq_bfqq_sync(bfqq)) |
2221 |
++ return NULL; |
2222 |
++ if (BFQQ_SEEKY(bfqq)) |
2223 |
++ return NULL; |
2224 |
++ |
2225 |
++ /* |
2226 |
++ * Do not merge queues of different priority classes. |
2227 |
++ */ |
2228 |
++ if (bfq_class_rt(bfqq) != bfq_class_rt(cur_bfqq)) |
2229 |
++ return NULL; |
2230 |
++ |
2231 |
++ return bfqq; |
2232 |
++} |
2233 |
++ |
2234 |
++/* |
2235 |
++ * If enough samples have been computed, return the current max budget |
2236 |
++ * stored in bfqd, which is dynamically updated according to the |
2237 |
++ * estimated disk peak rate; otherwise return the default max budget |
2238 |
++ */ |
2239 |
++static inline unsigned long bfq_max_budget(struct bfq_data *bfqd) |
2240 |
++{ |
2241 |
++ if (bfqd->budgets_assigned < 194) |
2242 |
++ return bfq_default_max_budget; |
2243 |
++ else |
2244 |
++ return bfqd->bfq_max_budget; |
2245 |
++} |
2246 |
++ |
2247 |
++/* |
2248 |
++ * Return min budget, which is a fraction of the current or default |
2249 |
++ * max budget (trying with 1/32) |
2250 |
++ */ |
2251 |
++static inline unsigned long bfq_min_budget(struct bfq_data *bfqd) |
2252 |
++{ |
2253 |
++ if (bfqd->budgets_assigned < 194) |
2254 |
++ return bfq_default_max_budget / 32; |
2255 |
++ else |
2256 |
++ return bfqd->bfq_max_budget / 32; |
2257 |
++} |
2258 |
++ |
2259 |
++/* |
2260 |
++ * Decides whether idling should be done for given device and |
2261 |
++ * given in-service queue. |
2262 |
++ */ |
2263 |
++static inline bool bfq_queue_nonrot_noidle(struct bfq_data *bfqd, |
2264 |
++ struct bfq_queue *in_service_bfqq) |
2265 |
++{ |
2266 |
++ if (in_service_bfqq == NULL) |
2267 |
++ return false; |
2268 |
++ /* |
2269 |
++ * If device is SSD it has no seek penalty, disable idling; but |
2270 |
++ * do so only if: |
2271 |
++ * - device does not support queuing, otherwise we still have |
2272 |
++ * a problem with sync vs async workloads; |
2273 |
++ * - the queue is not weight-raised, to preserve guarantees. |
2274 |
++ */ |
2275 |
++ return (blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag && |
2276 |
++ in_service_bfqq->raising_coeff == 1); |
2277 |
++} |
2278 |
++ |
2279 |
++static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
2280 |
++{ |
2281 |
++ struct bfq_queue *bfqq = bfqd->in_service_queue; |
2282 |
++ struct bfq_io_cq *bic; |
2283 |
++ unsigned long sl; |
2284 |
++ |
2285 |
++ WARN_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
2286 |
++ |
2287 |
++ /* Tasks have exited, don't wait. */ |
2288 |
++ bic = bfqd->in_service_bic; |
2289 |
++ if (bic == NULL || atomic_read(&bic->icq.ioc->active_ref) == 0) |
2290 |
++ return; |
2291 |
++ |
2292 |
++ bfq_mark_bfqq_wait_request(bfqq); |
2293 |
++ |
2294 |
++ /* |
2295 |
++ * We don't want to idle for seeks, but we do want to allow |
2296 |
++ * fair distribution of slice time for a process doing back-to-back |
2297 |
++ * seeks. So allow a little bit of time for him to submit a new rq. |
2298 |
++ * |
2299 |
++ * To prevent processes with (partly) seeky workloads from |
2300 |
++ * being too ill-treated, grant them a small fraction of the |
2301 |
++ * assigned budget before reducing the waiting time to |
2302 |
++ * BFQ_MIN_TT. This happened to help reduce latency. |
2303 |
++ */ |
2304 |
++ sl = bfqd->bfq_slice_idle; |
2305 |
++ if (bfq_sample_valid(bfqq->seek_samples) && BFQQ_SEEKY(bfqq) && |
2306 |
++ bfqq->entity.service > bfq_max_budget(bfqd) / 8 && |
2307 |
++ bfqq->raising_coeff == 1) |
2308 |
++ sl = min(sl, msecs_to_jiffies(BFQ_MIN_TT)); |
2309 |
++ else if (bfqq->raising_coeff > 1) |
2310 |
++ sl = sl * 3; |
2311 |
++ bfqd->last_idling_start = ktime_get(); |
2312 |
++ mod_timer(&bfqd->idle_slice_timer, jiffies + sl); |
2313 |
++ bfq_log(bfqd, "arm idle: %u/%u ms", |
2314 |
++ jiffies_to_msecs(sl), jiffies_to_msecs(bfqd->bfq_slice_idle)); |
2315 |
++} |
2316 |
++ |
2317 |
++/* |
2318 |
++ * Set the maximum time for the in-service queue to consume its |
2319 |
++ * budget. This prevents seeky processes from lowering the disk |
2320 |
++ * throughput (always guaranteed with a time slice scheme as in CFQ). |
2321 |
++ */ |
2322 |
++static void bfq_set_budget_timeout(struct bfq_data *bfqd) |
2323 |
++{ |
2324 |
++ struct bfq_queue *bfqq = bfqd->in_service_queue; |
2325 |
++ unsigned int timeout_coeff; |
2326 |
++ if (bfqq->raising_cur_max_time == bfqd->bfq_raising_rt_max_time) |
2327 |
++ timeout_coeff = 1; |
2328 |
++ else |
2329 |
++ timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; |
2330 |
++ |
2331 |
++ bfqd->last_budget_start = ktime_get(); |
2332 |
++ |
2333 |
++ bfq_clear_bfqq_budget_new(bfqq); |
2334 |
++ bfqq->budget_timeout = jiffies + |
2335 |
++ bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * timeout_coeff; |
2336 |
++ |
2337 |
++ bfq_log_bfqq(bfqd, bfqq, "set budget_timeout %u", |
2338 |
++ jiffies_to_msecs(bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * |
2339 |
++ timeout_coeff)); |
2340 |
++} |
2341 |
++ |
2342 |
++/* |
2343 |
++ * Move request from internal lists to the request queue dispatch list. |
2344 |
++ */ |
2345 |
++static void bfq_dispatch_insert(struct request_queue *q, struct request *rq) |
2346 |
++{ |
2347 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
2348 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
2349 |
++ |
2350 |
++ bfq_remove_request(rq); |
2351 |
++ bfqq->dispatched++; |
2352 |
++ elv_dispatch_sort(q, rq); |
2353 |
++ |
2354 |
++ if (bfq_bfqq_sync(bfqq)) |
2355 |
++ bfqd->sync_flight++; |
2356 |
++} |
2357 |
++ |
2358 |
++/* |
2359 |
++ * Return expired entry, or NULL to just start from scratch in rbtree. |
2360 |
++ */ |
2361 |
++static struct request *bfq_check_fifo(struct bfq_queue *bfqq) |
2362 |
++{ |
2363 |
++ struct request *rq = NULL; |
2364 |
++ |
2365 |
++ if (bfq_bfqq_fifo_expire(bfqq)) |
2366 |
++ return NULL; |
2367 |
++ |
2368 |
++ bfq_mark_bfqq_fifo_expire(bfqq); |
2369 |
++ |
2370 |
++ if (list_empty(&bfqq->fifo)) |
2371 |
++ return NULL; |
2372 |
++ |
2373 |
++ rq = rq_entry_fifo(bfqq->fifo.next); |
2374 |
++ |
2375 |
++ if (time_before(jiffies, rq_fifo_time(rq))) |
2376 |
++ return NULL; |
2377 |
++ |
2378 |
++ return rq; |
2379 |
++} |
2380 |
++ |
2381 |
++/* |
2382 |
++ * Must be called with the queue_lock held. |
2383 |
++ */ |
2384 |
++static int bfqq_process_refs(struct bfq_queue *bfqq) |
2385 |
++{ |
2386 |
++ int process_refs, io_refs; |
2387 |
++ |
2388 |
++ io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
2389 |
++ process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
2390 |
++ BUG_ON(process_refs < 0); |
2391 |
++ return process_refs; |
2392 |
++} |
2393 |
++ |
2394 |
++static void bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
2395 |
++{ |
2396 |
++ int process_refs, new_process_refs; |
2397 |
++ struct bfq_queue *__bfqq; |
2398 |
++ |
2399 |
++ /* |
2400 |
++ * If there are no process references on the new_bfqq, then it is |
2401 |
++ * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
2402 |
++ * may have dropped their last reference (not just their last process |
2403 |
++ * reference). |
2404 |
++ */ |
2405 |
++ if (!bfqq_process_refs(new_bfqq)) |
2406 |
++ return; |
2407 |
++ |
2408 |
++ /* Avoid a circular list and skip interim queue merges. */ |
2409 |
++ while ((__bfqq = new_bfqq->new_bfqq)) { |
2410 |
++ if (__bfqq == bfqq) |
2411 |
++ return; |
2412 |
++ new_bfqq = __bfqq; |
2413 |
++ } |
2414 |
++ |
2415 |
++ process_refs = bfqq_process_refs(bfqq); |
2416 |
++ new_process_refs = bfqq_process_refs(new_bfqq); |
2417 |
++ /* |
2418 |
++ * If the process for the bfqq has gone away, there is no |
2419 |
++ * sense in merging the queues. |
2420 |
++ */ |
2421 |
++ if (process_refs == 0 || new_process_refs == 0) |
2422 |
++ return; |
2423 |
++ |
2424 |
++ /* |
2425 |
++ * Merge in the direction of the lesser amount of work. |
2426 |
++ */ |
2427 |
++ if (new_process_refs >= process_refs) { |
2428 |
++ bfqq->new_bfqq = new_bfqq; |
2429 |
++ atomic_add(process_refs, &new_bfqq->ref); |
2430 |
++ } else { |
2431 |
++ new_bfqq->new_bfqq = bfqq; |
2432 |
++ atomic_add(new_process_refs, &bfqq->ref); |
2433 |
++ } |
2434 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
2435 |
++ new_bfqq->pid); |
2436 |
++} |
2437 |
++ |
2438 |
++static inline unsigned long bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
2439 |
++{ |
2440 |
++ struct bfq_entity *entity = &bfqq->entity; |
2441 |
++ return entity->budget - entity->service; |
2442 |
++} |
2443 |
++ |
2444 |
++static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
2445 |
++{ |
2446 |
++ BUG_ON(bfqq != bfqd->in_service_queue); |
2447 |
++ |
2448 |
++ __bfq_bfqd_reset_in_service(bfqd); |
2449 |
++ |
2450 |
++ /* |
2451 |
++ * If this bfqq is shared between multiple processes, check |
2452 |
++ * to make sure that those processes are still issuing I/Os |
2453 |
++ * within the mean seek distance. If not, it may be time to |
2454 |
++ * break the queues apart again. |
2455 |
++ */ |
2456 |
++ if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq)) |
2457 |
++ bfq_mark_bfqq_split_coop(bfqq); |
2458 |
++ |
2459 |
++ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
2460 |
++ /* |
2461 |
++ * overloading budget_timeout field to store when |
2462 |
++ * the queue remains with no backlog, used by |
2463 |
++ * the weight-raising mechanism |
2464 |
++ */ |
2465 |
++ bfqq->budget_timeout = jiffies; |
2466 |
++ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
2467 |
++ } else { |
2468 |
++ bfq_activate_bfqq(bfqd, bfqq); |
2469 |
++ /* |
2470 |
++ * Resort priority tree of potential close cooperators. |
2471 |
++ */ |
2472 |
++ bfq_rq_pos_tree_add(bfqd, bfqq); |
2473 |
++ } |
2474 |
++} |
2475 |
++ |
2476 |
++/** |
2477 |
++ * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. |
2478 |
++ * @bfqd: device data. |
2479 |
++ * @bfqq: queue to update. |
2480 |
++ * @reason: reason for expiration. |
2481 |
++ * |
2482 |
++ * Handle the feedback on @bfqq budget. See the body for detailed |
2483 |
++ * comments. |
2484 |
++ */ |
2485 |
++static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, |
2486 |
++ struct bfq_queue *bfqq, |
2487 |
++ enum bfqq_expiration reason) |
2488 |
++{ |
2489 |
++ struct request *next_rq; |
2490 |
++ unsigned long budget, min_budget; |
2491 |
++ |
2492 |
++ budget = bfqq->max_budget; |
2493 |
++ min_budget = bfq_min_budget(bfqd); |
2494 |
++ |
2495 |
++ BUG_ON(bfqq != bfqd->in_service_queue); |
2496 |
++ |
2497 |
++ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %lu, budg left %lu", |
2498 |
++ bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); |
2499 |
++ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %lu, min budg %lu", |
2500 |
++ budget, bfq_min_budget(bfqd)); |
2501 |
++ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", |
2502 |
++ bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue)); |
2503 |
++ |
2504 |
++ if (bfq_bfqq_sync(bfqq)) { |
2505 |
++ switch (reason) { |
2506 |
++ /* |
2507 |
++ * Caveat: in all the following cases we trade latency |
2508 |
++ * for throughput. |
2509 |
++ */ |
2510 |
++ case BFQ_BFQQ_TOO_IDLE: |
2511 |
++ /* |
2512 |
++ * This is the only case where we may reduce |
2513 |
++ * the budget: if there is no requets of the |
2514 |
++ * process still waiting for completion, then |
2515 |
++ * we assume (tentatively) that the timer has |
2516 |
++ * expired because the batch of requests of |
2517 |
++ * the process could have been served with a |
2518 |
++ * smaller budget. Hence, betting that |
2519 |
++ * process will behave in the same way when it |
2520 |
++ * becomes backlogged again, we reduce its |
2521 |
++ * next budget. As long as we guess right, |
2522 |
++ * this budget cut reduces the latency |
2523 |
++ * experienced by the process. |
2524 |
++ * |
2525 |
++ * However, if there are still outstanding |
2526 |
++ * requests, then the process may have not yet |
2527 |
++ * issued its next request just because it is |
2528 |
++ * still waiting for the completion of some of |
2529 |
++ * the still oustanding ones. So in this |
2530 |
++ * subcase we do not reduce its budget, on the |
2531 |
++ * contrary we increase it to possibly boost |
2532 |
++ * the throughput, as discussed in the |
2533 |
++ * comments to the BUDGET_TIMEOUT case. |
2534 |
++ */ |
2535 |
++ if (bfqq->dispatched > 0) /* still oustanding reqs */ |
2536 |
++ budget = min(budget * 2, bfqd->bfq_max_budget); |
2537 |
++ else { |
2538 |
++ if (budget > 5 * min_budget) |
2539 |
++ budget -= 4 * min_budget; |
2540 |
++ else |
2541 |
++ budget = min_budget; |
2542 |
++ } |
2543 |
++ break; |
2544 |
++ case BFQ_BFQQ_BUDGET_TIMEOUT: |
2545 |
++ /* |
2546 |
++ * We double the budget here because: 1) it |
2547 |
++ * gives the chance to boost the throughput if |
2548 |
++ * this is not a seeky process (which may have |
2549 |
++ * bumped into this timeout because of, e.g., |
2550 |
++ * ZBR), 2) together with charge_full_budget |
2551 |
++ * it helps give seeky processes higher |
2552 |
++ * timestamps, and hence be served less |
2553 |
++ * frequently. |
2554 |
++ */ |
2555 |
++ budget = min(budget * 2, bfqd->bfq_max_budget); |
2556 |
++ break; |
2557 |
++ case BFQ_BFQQ_BUDGET_EXHAUSTED: |
2558 |
++ /* |
2559 |
++ * The process still has backlog, and did not |
2560 |
++ * let either the budget timeout or the disk |
2561 |
++ * idling timeout expire. Hence it is not |
2562 |
++ * seeky, has a short thinktime and may be |
2563 |
++ * happy with a higher budget too. So |
2564 |
++ * definitely increase the budget of this good |
2565 |
++ * candidate to boost the disk throughput. |
2566 |
++ */ |
2567 |
++ budget = min(budget * 4, bfqd->bfq_max_budget); |
2568 |
++ break; |
2569 |
++ case BFQ_BFQQ_NO_MORE_REQUESTS: |
2570 |
++ /* |
2571 |
++ * Leave the budget unchanged. |
2572 |
++ */ |
2573 |
++ default: |
2574 |
++ return; |
2575 |
++ } |
2576 |
++ } else /* async queue */ |
2577 |
++ /* async queues get always the maximum possible budget |
2578 |
++ * (their ability to dispatch is limited by |
2579 |
++ * @bfqd->bfq_max_budget_async_rq). |
2580 |
++ */ |
2581 |
++ budget = bfqd->bfq_max_budget; |
2582 |
++ |
2583 |
++ bfqq->max_budget = budget; |
2584 |
++ |
2585 |
++ if (bfqd->budgets_assigned >= 194 && bfqd->bfq_user_max_budget == 0 && |
2586 |
++ bfqq->max_budget > bfqd->bfq_max_budget) |
2587 |
++ bfqq->max_budget = bfqd->bfq_max_budget; |
2588 |
++ |
2589 |
++ /* |
2590 |
++ * Make sure that we have enough budget for the next request. |
2591 |
++ * Since the finish time of the bfqq must be kept in sync with |
2592 |
++ * the budget, be sure to call __bfq_bfqq_expire() after the |
2593 |
++ * update. |
2594 |
++ */ |
2595 |
++ next_rq = bfqq->next_rq; |
2596 |
++ if (next_rq != NULL) |
2597 |
++ bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, |
2598 |
++ bfq_serv_to_charge(next_rq, bfqq)); |
2599 |
++ else |
2600 |
++ bfqq->entity.budget = bfqq->max_budget; |
2601 |
++ |
2602 |
++ bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %lu", |
2603 |
++ next_rq != NULL ? blk_rq_sectors(next_rq) : 0, |
2604 |
++ bfqq->entity.budget); |
2605 |
++} |
2606 |
++ |
2607 |
++static unsigned long bfq_calc_max_budget(u64 peak_rate, u64 timeout) |
2608 |
++{ |
2609 |
++ unsigned long max_budget; |
2610 |
++ |
2611 |
++ /* |
2612 |
++ * The max_budget calculated when autotuning is equal to the |
2613 |
++ * amount of sectors transfered in timeout_sync at the |
2614 |
++ * estimated peak rate. |
2615 |
++ */ |
2616 |
++ max_budget = (unsigned long)(peak_rate * 1000 * |
2617 |
++ timeout >> BFQ_RATE_SHIFT); |
2618 |
++ |
2619 |
++ return max_budget; |
2620 |
++} |
2621 |
++ |
2622 |
++/* |
2623 |
++ * In addition to updating the peak rate, checks whether the process |
2624 |
++ * is "slow", and returns 1 if so. This slow flag is used, in addition |
2625 |
++ * to the budget timeout, to reduce the amount of service provided to |
2626 |
++ * seeky processes, and hence reduce their chances to lower the |
2627 |
++ * throughput. See the code for more details. |
2628 |
++ */ |
2629 |
++static int bfq_update_peak_rate(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
2630 |
++ int compensate, enum bfqq_expiration reason) |
2631 |
++{ |
2632 |
++ u64 bw, usecs, expected, timeout; |
2633 |
++ ktime_t delta; |
2634 |
++ int update = 0; |
2635 |
++ |
2636 |
++ if (!bfq_bfqq_sync(bfqq) || bfq_bfqq_budget_new(bfqq)) |
2637 |
++ return 0; |
2638 |
++ |
2639 |
++ if (compensate) |
2640 |
++ delta = bfqd->last_idling_start; |
2641 |
++ else |
2642 |
++ delta = ktime_get(); |
2643 |
++ delta = ktime_sub(delta, bfqd->last_budget_start); |
2644 |
++ usecs = ktime_to_us(delta); |
2645 |
++ |
2646 |
++ /* Don't trust short/unrealistic values. */ |
2647 |
++ if (usecs < 100 || usecs >= LONG_MAX) |
2648 |
++ return 0; |
2649 |
++ |
2650 |
++ /* |
2651 |
++ * Calculate the bandwidth for the last slice. We use a 64 bit |
2652 |
++ * value to store the peak rate, in sectors per usec in fixed |
2653 |
++ * point math. We do so to have enough precision in the estimate |
2654 |
++ * and to avoid overflows. |
2655 |
++ */ |
2656 |
++ bw = (u64)bfqq->entity.service << BFQ_RATE_SHIFT; |
2657 |
++ do_div(bw, (unsigned long)usecs); |
2658 |
++ |
2659 |
++ timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
2660 |
++ |
2661 |
++ /* |
2662 |
++ * Use only long (> 20ms) intervals to filter out spikes for |
2663 |
++ * the peak rate estimation. |
2664 |
++ */ |
2665 |
++ if (usecs > 20000) { |
2666 |
++ if (bw > bfqd->peak_rate || |
2667 |
++ (!BFQQ_SEEKY(bfqq) && |
2668 |
++ reason == BFQ_BFQQ_BUDGET_TIMEOUT)) { |
2669 |
++ bfq_log(bfqd, "measured bw =%llu", bw); |
2670 |
++ /* |
2671 |
++ * To smooth oscillations use a low-pass filter with |
2672 |
++ * alpha=7/8, i.e., |
2673 |
++ * new_rate = (7/8) * old_rate + (1/8) * bw |
2674 |
++ */ |
2675 |
++ do_div(bw, 8); |
2676 |
++ if (bw == 0) |
2677 |
++ return 0; |
2678 |
++ bfqd->peak_rate *= 7; |
2679 |
++ do_div(bfqd->peak_rate, 8); |
2680 |
++ bfqd->peak_rate += bw; |
2681 |
++ update = 1; |
2682 |
++ bfq_log(bfqd, "new peak_rate=%llu", bfqd->peak_rate); |
2683 |
++ } |
2684 |
++ |
2685 |
++ update |= bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES - 1; |
2686 |
++ |
2687 |
++ if (bfqd->peak_rate_samples < BFQ_PEAK_RATE_SAMPLES) |
2688 |
++ bfqd->peak_rate_samples++; |
2689 |
++ |
2690 |
++ if (bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES && |
2691 |
++ update && bfqd->bfq_user_max_budget == 0) { |
2692 |
++ bfqd->bfq_max_budget = |
2693 |
++ bfq_calc_max_budget(bfqd->peak_rate, timeout); |
2694 |
++ bfq_log(bfqd, "new max_budget=%lu", |
2695 |
++ bfqd->bfq_max_budget); |
2696 |
++ } |
2697 |
++ } |
2698 |
++ |
2699 |
++ /* |
2700 |
++ * If the process has been served for a too short time |
2701 |
++ * interval to let its possible sequential accesses prevail on |
2702 |
++ * the initial seek time needed to move the disk head on the |
2703 |
++ * first sector it requested, then give the process a chance |
2704 |
++ * and for the moment return false. |
2705 |
++ */ |
2706 |
++ if (bfqq->entity.budget <= bfq_max_budget(bfqd) / 8) |
2707 |
++ return 0; |
2708 |
++ |
2709 |
++ /* |
2710 |
++ * A process is considered ``slow'' (i.e., seeky, so that we |
2711 |
++ * cannot treat it fairly in the service domain, as it would |
2712 |
++ * slow down too much the other processes) if, when a slice |
2713 |
++ * ends for whatever reason, it has received service at a |
2714 |
++ * rate that would not be high enough to complete the budget |
2715 |
++ * before the budget timeout expiration. |
2716 |
++ */ |
2717 |
++ expected = bw * 1000 * timeout >> BFQ_RATE_SHIFT; |
2718 |
++ |
2719 |
++ /* |
2720 |
++ * Caveat: processes doing IO in the slower disk zones will |
2721 |
++ * tend to be slow(er) even if not seeky. And the estimated |
2722 |
++ * peak rate will actually be an average over the disk |
2723 |
++ * surface. Hence, to not be too harsh with unlucky processes, |
2724 |
++ * we keep a budget/3 margin of safety before declaring a |
2725 |
++ * process slow. |
2726 |
++ */ |
2727 |
++ return expected > (4 * bfqq->entity.budget) / 3; |
2728 |
++} |
2729 |
++ |
2730 |
++/* |
2731 |
++ * To be deemed as soft real-time, an application must meet two requirements. |
2732 |
++ * The first is that the application must not require an average bandwidth |
2733 |
++ * higher than the approximate bandwidth required to playback or record a |
2734 |
++ * compressed high-definition video. |
2735 |
++ * The next function is invoked on the completion of the last request of a |
2736 |
++ * batch, to compute the next-start time instant, soft_rt_next_start, such |
2737 |
++ * that, if the next request of the application does not arrive before |
2738 |
++ * soft_rt_next_start, then the above requirement on the bandwidth is met. |
2739 |
++ * |
2740 |
++ * The second requirement is that the request pattern of the application is |
2741 |
++ * isochronous, i.e., that, after issuing a request or a batch of requests, the |
2742 |
++ * application stops for a while, then issues a new batch, and so on. For this |
2743 |
++ * reason the next function is invoked to compute soft_rt_next_start only for |
2744 |
++ * applications that meet this requirement, whereas soft_rt_next_start is set |
2745 |
++ * to infinity for applications that do not. |
2746 |
++ * |
2747 |
++ * Unfortunately, even a greedy application may happen to behave in an |
2748 |
++ * isochronous way if several processes are competing for the CPUs. In fact, |
2749 |
++ * in this scenario the application stops issuing requests while the CPUs are |
2750 |
++ * busy serving other processes, then restarts, then stops again for a while, |
2751 |
++ * and so on. In addition, if the disk achieves a low enough throughput with |
2752 |
++ * the request pattern issued by the application, then the above bandwidth |
2753 |
++ * requirement may happen to be met too. To prevent such a greedy application |
2754 |
++ * to be deemed as soft real-time, a further rule is used in the computation |
2755 |
++ * of soft_rt_next_start: soft_rt_next_start must be higher than the current |
2756 |
++ * time plus the maximum time for which the arrival of a request is waited |
2757 |
++ * for when a sync queue becomes idle, namely bfqd->bfq_slice_idle. This |
2758 |
++ * filters out greedy applications, as the latter issue instead their next |
2759 |
++ * request as soon as possible after the last one has been completed (in |
2760 |
++ * contrast, when a batch of requests is completed, a soft real-time |
2761 |
++ * application spends some time processing data). |
2762 |
++ * |
2763 |
++ * Actually, the last filter may easily generate false positives if: only |
2764 |
++ * bfqd->bfq_slice_idle is used as a reference time interval, and one or |
2765 |
++ * both the following two cases occur: |
2766 |
++ * 1) HZ is so low that the duration of a jiffie is comparable to or higher |
2767 |
++ * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with |
2768 |
++ * HZ=100. |
2769 |
++ * 2) jiffies, instead of increasing at a constant rate, may stop increasing |
2770 |
++ * for a while, then suddenly 'jump' by several units to recover the lost |
2771 |
++ * increments. This seems to happen, e.g., inside virtual machines. |
2772 |
++ * To address this issue, we do not use as a reference time interval just |
2773 |
++ * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In |
2774 |
++ * particular we add the minimum number of jiffies for which the filter seems |
2775 |
++ * to be quite precise also in embedded systems and KVM/QEMU virtual machines. |
2776 |
++ */ |
2777 |
++static inline u64 bfq_bfqq_softrt_next_start(struct bfq_data *bfqd, |
2778 |
++ struct bfq_queue *bfqq) |
2779 |
++{ |
2780 |
++ return max(bfqq->last_idle_bklogged + |
2781 |
++ HZ * bfqq->service_from_backlogged / |
2782 |
++ bfqd->bfq_raising_max_softrt_rate, |
2783 |
++ (u64)jiffies + bfqq->bfqd->bfq_slice_idle + 4); |
2784 |
++} |
2785 |
++ |
2786 |
++/** |
2787 |
++ * bfq_bfqq_expire - expire a queue. |
2788 |
++ * @bfqd: device owning the queue. |
2789 |
++ * @bfqq: the queue to expire. |
2790 |
++ * @compensate: if true, compensate for the time spent idling. |
2791 |
++ * @reason: the reason causing the expiration. |
2792 |
++ * |
2793 |
++ * |
2794 |
++ * If the process associated to the queue is slow (i.e., seeky), or in |
2795 |
++ * case of budget timeout, or, finally, if it is async, we |
2796 |
++ * artificially charge it an entire budget (independently of the |
2797 |
++ * actual service it received). As a consequence, the queue will get |
2798 |
++ * higher timestamps than the correct ones upon reactivation, and |
2799 |
++ * hence it will be rescheduled as if it had received more service |
2800 |
++ * than what it actually received. In the end, this class of processes |
2801 |
++ * will receive less service in proportion to how slowly they consume |
2802 |
++ * their budgets (and hence how seriously they tend to lower the |
2803 |
++ * throughput). |
2804 |
++ * |
2805 |
++ * In contrast, when a queue expires because it has been idling for |
2806 |
++ * too much or because it exhausted its budget, we do not touch the |
2807 |
++ * amount of service it has received. Hence when the queue will be |
2808 |
++ * reactivated and its timestamps updated, the latter will be in sync |
2809 |
++ * with the actual service received by the queue until expiration. |
2810 |
++ * |
2811 |
++ * Charging a full budget to the first type of queues and the exact |
2812 |
++ * service to the others has the effect of using the WF2Q+ policy to |
2813 |
++ * schedule the former on a timeslice basis, without violating the |
2814 |
++ * service domain guarantees of the latter. |
2815 |
++ */ |
2816 |
++static void bfq_bfqq_expire(struct bfq_data *bfqd, |
2817 |
++ struct bfq_queue *bfqq, |
2818 |
++ int compensate, |
2819 |
++ enum bfqq_expiration reason) |
2820 |
++{ |
2821 |
++ int slow; |
2822 |
++ BUG_ON(bfqq != bfqd->in_service_queue); |
2823 |
++ |
2824 |
++ /* Update disk peak rate for autotuning and check whether the |
2825 |
++ * process is slow (see bfq_update_peak_rate). |
2826 |
++ */ |
2827 |
++ slow = bfq_update_peak_rate(bfqd, bfqq, compensate, reason); |
2828 |
++ |
2829 |
++ /* |
2830 |
++ * As above explained, 'punish' slow (i.e., seeky), timed-out |
2831 |
++ * and async queues, to favor sequential sync workloads. |
2832 |
++ * |
2833 |
++ * Processes doing IO in the slower disk zones will tend to be |
2834 |
++ * slow(er) even if not seeky. Hence, since the estimated peak |
2835 |
++ * rate is actually an average over the disk surface, these |
2836 |
++ * processes may timeout just for bad luck. To avoid punishing |
2837 |
++ * them we do not charge a full budget to a process that |
2838 |
++ * succeeded in consuming at least 2/3 of its budget. |
2839 |
++ */ |
2840 |
++ if (slow || (reason == BFQ_BFQQ_BUDGET_TIMEOUT && |
2841 |
++ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)) |
2842 |
++ bfq_bfqq_charge_full_budget(bfqq); |
2843 |
++ |
2844 |
++ bfqq->service_from_backlogged += bfqq->entity.service; |
2845 |
++ |
2846 |
++ if (bfqd->low_latency && bfqq->raising_coeff == 1) |
2847 |
++ bfqq->last_rais_start_finish = jiffies; |
2848 |
++ |
2849 |
++ if (bfqd->low_latency && bfqd->bfq_raising_max_softrt_rate > 0) { |
2850 |
++ if (reason != BFQ_BFQQ_BUDGET_TIMEOUT && |
2851 |
++ reason != BFQ_BFQQ_BUDGET_EXHAUSTED) { |
2852 |
++ /* |
2853 |
++ * If we get here, then the request pattern is |
2854 |
++ * isochronous (see the comments to the function |
2855 |
++ * bfq_bfqq_softrt_next_start()). However, if the |
2856 |
++ * queue still has in-flight requests, then it is |
2857 |
++ * better to postpone the computation of next_start |
2858 |
++ * to the next request completion. In fact, if we |
2859 |
++ * computed it now, then the application might pass |
2860 |
++ * the greedy-application filter improperly, because |
2861 |
++ * the arrival of its next request may happen to be |
2862 |
++ * higher than (jiffies + bfqq->bfqd->bfq_slice_idle) |
2863 |
++ * not because the application is truly soft real- |
2864 |
++ * time, but just because the application is currently |
2865 |
++ * waiting for the completion of some request before |
2866 |
++ * issuing, as quickly as possible, its next request. |
2867 |
++ */ |
2868 |
++ if (bfqq->dispatched > 0) { |
2869 |
++ bfqq->soft_rt_next_start = -1; |
2870 |
++ bfq_mark_bfqq_softrt_update(bfqq); |
2871 |
++ } else |
2872 |
++ bfqq->soft_rt_next_start = |
2873 |
++ bfq_bfqq_softrt_next_start(bfqd, bfqq); |
2874 |
++ } else |
2875 |
++ bfqq->soft_rt_next_start = -1; /* infinity */ |
2876 |
++ } |
2877 |
++ |
2878 |
++ bfq_log_bfqq(bfqd, bfqq, |
2879 |
++ "expire (%d, slow %d, num_disp %d, idle_win %d)", reason, slow, |
2880 |
++ bfqq->dispatched, bfq_bfqq_idle_window(bfqq)); |
2881 |
++ |
2882 |
++ /* Increase, decrease or leave budget unchanged according to reason */ |
2883 |
++ __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); |
2884 |
++ __bfq_bfqq_expire(bfqd, bfqq); |
2885 |
++} |
2886 |
++ |
2887 |
++/* |
2888 |
++ * Budget timeout is not implemented through a dedicated timer, but |
2889 |
++ * just checked on request arrivals and completions, as well as on |
2890 |
++ * idle timer expirations. |
2891 |
++ */ |
2892 |
++static int bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) |
2893 |
++{ |
2894 |
++ if (bfq_bfqq_budget_new(bfqq)) |
2895 |
++ return 0; |
2896 |
++ |
2897 |
++ if (time_before(jiffies, bfqq->budget_timeout)) |
2898 |
++ return 0; |
2899 |
++ |
2900 |
++ return 1; |
2901 |
++} |
2902 |
++ |
2903 |
++/* |
2904 |
++ * If we expire a queue that is waiting for the arrival of a new |
2905 |
++ * request, we may prevent the fictitious timestamp backshifting that |
2906 |
++ * allows the guarantees of the queue to be preserved (see [1] for |
2907 |
++ * this tricky aspect). Hence we return true only if this condition |
2908 |
++ * does not hold, or if the queue is slow enough to deserve only to be |
2909 |
++ * kicked off for preserving a high throughput. |
2910 |
++*/ |
2911 |
++static inline int bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) |
2912 |
++{ |
2913 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, |
2914 |
++ "may_budget_timeout: wr %d left %d timeout %d", |
2915 |
++ bfq_bfqq_wait_request(bfqq), |
2916 |
++ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, |
2917 |
++ bfq_bfqq_budget_timeout(bfqq)); |
2918 |
++ |
2919 |
++ return (!bfq_bfqq_wait_request(bfqq) || |
2920 |
++ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) |
2921 |
++ && |
2922 |
++ bfq_bfqq_budget_timeout(bfqq); |
2923 |
++} |
2924 |
++ |
2925 |
++/* |
2926 |
++ * For weight-raised queues issuing sync requests, idling is always performed, |
2927 |
++ * as this is instrumental in guaranteeing a high fraction of the throughput |
2928 |
++ * to these queues, and hence in guaranteeing a lower latency for their |
2929 |
++ * requests. See [1] for details. |
2930 |
++ * |
2931 |
++ * For non-weight-raised queues, idling is instead disabled if the device is |
2932 |
++ * NCQ-enabled and non-rotational, as this boosts the throughput on such |
2933 |
++ * devices. |
2934 |
++ */ |
2935 |
++static inline bool bfq_bfqq_must_not_expire(struct bfq_queue *bfqq) |
2936 |
++{ |
2937 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
2938 |
++ |
2939 |
++ return bfq_bfqq_sync(bfqq) && ( |
2940 |
++ bfqq->raising_coeff > 1 || |
2941 |
++ (bfq_bfqq_idle_window(bfqq) && |
2942 |
++ !(bfqd->hw_tag && |
2943 |
++ (blk_queue_nonrot(bfqd->queue) || |
2944 |
++ /* |
2945 |
++ * If there are weight-raised busy queues, then do not idle |
2946 |
++ * the disk for a sync non-weight-raised queue, and hence |
2947 |
++ * expire the queue immediately if empty. Combined with the |
2948 |
++ * timestamping rules of BFQ (see [1] for details), this |
2949 |
++ * causes sync non-weight-raised queues to get a lower |
2950 |
++ * fraction of the disk throughput, and hence reduces the rate |
2951 |
++ * at which the processes associated to these queues ask for |
2952 |
++ * requests from the request pool. |
2953 |
++ * |
2954 |
++ * This is beneficial for weight-raised processes, when the |
2955 |
++ * system operates in request-pool saturation conditions |
2956 |
++ * (e.g., in the presence of write hogs). In fact, if |
2957 |
++ * non-weight-raised processes ask for requests at a lower |
2958 |
++ * rate, then weight-raised processes have a higher |
2959 |
++ * probability to get a request from the pool immediately |
2960 |
++ * (or at least soon) when they need one. Hence they have a |
2961 |
++ * higher probability to actually get a fraction of the disk |
2962 |
++ * throughput proportional to their high weight. This is |
2963 |
++ * especially true with NCQ-enabled drives, which enqueue |
2964 |
++ * several requests in advance and further reorder |
2965 |
++ * internally-queued requests. |
2966 |
++ * |
2967 |
++ * Mistreating non-weight-raised queues in the above-described |
2968 |
++ * way, when there are busy weight-raised queues, seems to |
2969 |
++ * mitigate starvation problems in the presence of heavy write |
2970 |
++ * workloads and NCQ, and hence to guarantee a higher |
2971 |
++ * application and system responsiveness in these hostile |
2972 |
++ * scenarios. |
2973 |
++ */ |
2974 |
++ bfqd->raised_busy_queues > 0) |
2975 |
++ ) |
2976 |
++ ) |
2977 |
++ ); |
2978 |
++} |
2979 |
++ |
2980 |
++/* |
2981 |
++ * If the in-service queue is empty, but it is sync and either of the following |
2982 |
++ * conditions holds, then: 1) the queue must remain in service and cannot be |
2983 |
++ * expired, and 2) the disk must be idled to wait for the possible arrival |
2984 |
++ * of a new request for the queue. The conditions are: |
2985 |
++ * - the device is rotational and not performing NCQ, and the queue has its |
2986 |
++ * idle window set (in this case, waiting for a new request for the queue |
2987 |
++ * is likely to boost the disk throughput); |
2988 |
++ * - the queue is weight-raised (waiting for the request is necessary to |
2989 |
++ * provide the queue with fairness and latency guarantees, see [1] for |
2990 |
++ * details). |
2991 |
++ */ |
2992 |
++static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) |
2993 |
++{ |
2994 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
2995 |
++ |
2996 |
++ return (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqd->bfq_slice_idle != 0 && |
2997 |
++ bfq_bfqq_must_not_expire(bfqq) && |
2998 |
++ !bfq_queue_nonrot_noidle(bfqd, bfqq)); |
2999 |
++} |
3000 |
++ |
3001 |
++/* |
3002 |
++ * Select a queue for service. If we have a current queue in service, |
3003 |
++ * check whether to continue servicing it, or retrieve and set a new one. |
3004 |
++ */ |
3005 |
++static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
3006 |
++{ |
3007 |
++ struct bfq_queue *bfqq, *new_bfqq = NULL; |
3008 |
++ struct request *next_rq; |
3009 |
++ enum bfqq_expiration reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
3010 |
++ |
3011 |
++ bfqq = bfqd->in_service_queue; |
3012 |
++ if (bfqq == NULL) |
3013 |
++ goto new_queue; |
3014 |
++ |
3015 |
++ bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); |
3016 |
++ |
3017 |
++ /* |
3018 |
++ * If another queue has a request waiting within our mean seek |
3019 |
++ * distance, let it run. The expire code will check for close |
3020 |
++ * cooperators and put the close queue at the front of the |
3021 |
++ * service tree. If possible, merge the expiring queue with the |
3022 |
++ * new bfqq. |
3023 |
++ */ |
3024 |
++ new_bfqq = bfq_close_cooperator(bfqd, bfqq); |
3025 |
++ if (new_bfqq != NULL && bfqq->new_bfqq == NULL) |
3026 |
++ bfq_setup_merge(bfqq, new_bfqq); |
3027 |
++ |
3028 |
++ if (bfq_may_expire_for_budg_timeout(bfqq) && |
3029 |
++ !timer_pending(&bfqd->idle_slice_timer) && |
3030 |
++ !bfq_bfqq_must_idle(bfqq)) |
3031 |
++ goto expire; |
3032 |
++ |
3033 |
++ next_rq = bfqq->next_rq; |
3034 |
++ /* |
3035 |
++ * If bfqq has requests queued and it has enough budget left to |
3036 |
++ * serve them, keep the queue, otherwise expire it. |
3037 |
++ */ |
3038 |
++ if (next_rq != NULL) { |
3039 |
++ if (bfq_serv_to_charge(next_rq, bfqq) > |
3040 |
++ bfq_bfqq_budget_left(bfqq)) { |
3041 |
++ reason = BFQ_BFQQ_BUDGET_EXHAUSTED; |
3042 |
++ goto expire; |
3043 |
++ } else { |
3044 |
++ /* |
3045 |
++ * The idle timer may be pending because we may not |
3046 |
++ * disable disk idling even when a new request arrives |
3047 |
++ */ |
3048 |
++ if (timer_pending(&bfqd->idle_slice_timer)) { |
3049 |
++ /* |
3050 |
++ * If we get here: 1) at least a new request |
3051 |
++ * has arrived but we have not disabled the |
3052 |
++ * timer because the request was too small, |
3053 |
++ * 2) then the block layer has unplugged the |
3054 |
++ * device, causing the dispatch to be invoked. |
3055 |
++ * |
3056 |
++ * Since the device is unplugged, now the |
3057 |
++ * requests are probably large enough to |
3058 |
++ * provide a reasonable throughput. |
3059 |
++ * So we disable idling. |
3060 |
++ */ |
3061 |
++ bfq_clear_bfqq_wait_request(bfqq); |
3062 |
++ del_timer(&bfqd->idle_slice_timer); |
3063 |
++ } |
3064 |
++ if (new_bfqq == NULL) |
3065 |
++ goto keep_queue; |
3066 |
++ else |
3067 |
++ goto expire; |
3068 |
++ } |
3069 |
++ } |
3070 |
++ |
3071 |
++ /* |
3072 |
++ * No requests pending. If the in-service queue has no cooperator and |
3073 |
++ * still has requests in flight (possibly waiting for a completion) |
3074 |
++ * or is idling for a new request, then keep it. |
3075 |
++ */ |
3076 |
++ if (new_bfqq == NULL && (timer_pending(&bfqd->idle_slice_timer) || |
3077 |
++ (bfqq->dispatched != 0 && bfq_bfqq_must_not_expire(bfqq)))) { |
3078 |
++ bfqq = NULL; |
3079 |
++ goto keep_queue; |
3080 |
++ } else if (new_bfqq != NULL && timer_pending(&bfqd->idle_slice_timer)) { |
3081 |
++ /* |
3082 |
++ * Expiring the queue because there is a close cooperator, |
3083 |
++ * cancel timer. |
3084 |
++ */ |
3085 |
++ bfq_clear_bfqq_wait_request(bfqq); |
3086 |
++ del_timer(&bfqd->idle_slice_timer); |
3087 |
++ } |
3088 |
++ |
3089 |
++ reason = BFQ_BFQQ_NO_MORE_REQUESTS; |
3090 |
++expire: |
3091 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, reason); |
3092 |
++new_queue: |
3093 |
++ bfqq = bfq_set_in_service_queue(bfqd, new_bfqq); |
3094 |
++ bfq_log(bfqd, "select_queue: new queue %d returned", |
3095 |
++ bfqq != NULL ? bfqq->pid : 0); |
3096 |
++keep_queue: |
3097 |
++ return bfqq; |
3098 |
++} |
3099 |
++ |
3100 |
++static void bfq_update_raising_data(struct bfq_data *bfqd, |
3101 |
++ struct bfq_queue *bfqq) |
3102 |
++{ |
3103 |
++ if (bfqq->raising_coeff > 1) { /* queue is being boosted */ |
3104 |
++ struct bfq_entity *entity = &bfqq->entity; |
3105 |
++ |
3106 |
++ bfq_log_bfqq(bfqd, bfqq, |
3107 |
++ "raising period dur %u/%u msec, " |
3108 |
++ "old raising coeff %u, w %d(%d)", |
3109 |
++ jiffies_to_msecs(jiffies - |
3110 |
++ bfqq->last_rais_start_finish), |
3111 |
++ jiffies_to_msecs(bfqq->raising_cur_max_time), |
3112 |
++ bfqq->raising_coeff, |
3113 |
++ bfqq->entity.weight, bfqq->entity.orig_weight); |
3114 |
++ |
3115 |
++ BUG_ON(bfqq != bfqd->in_service_queue && entity->weight != |
3116 |
++ entity->orig_weight * bfqq->raising_coeff); |
3117 |
++ if (entity->ioprio_changed) |
3118 |
++ bfq_log_bfqq(bfqd, bfqq, |
3119 |
++ "WARN: pending prio change"); |
3120 |
++ /* |
3121 |
++ * If too much time has elapsed from the beginning |
3122 |
++ * of this weight-raising, stop it. |
3123 |
++ */ |
3124 |
++ if (jiffies - bfqq->last_rais_start_finish > |
3125 |
++ bfqq->raising_cur_max_time) { |
3126 |
++ bfqq->last_rais_start_finish = jiffies; |
3127 |
++ bfq_log_bfqq(bfqd, bfqq, |
3128 |
++ "wrais ending at %llu msec," |
3129 |
++ "rais_max_time %u", |
3130 |
++ bfqq->last_rais_start_finish, |
3131 |
++ jiffies_to_msecs(bfqq-> |
3132 |
++ raising_cur_max_time)); |
3133 |
++ bfq_bfqq_end_raising(bfqq); |
3134 |
++ __bfq_entity_update_weight_prio( |
3135 |
++ bfq_entity_service_tree(entity), |
3136 |
++ entity); |
3137 |
++ } |
3138 |
++ } |
3139 |
++} |
3140 |
++ |
3141 |
++/* |
3142 |
++ * Dispatch one request from bfqq, moving it to the request queue |
3143 |
++ * dispatch list. |
3144 |
++ */ |
3145 |
++static int bfq_dispatch_request(struct bfq_data *bfqd, |
3146 |
++ struct bfq_queue *bfqq) |
3147 |
++{ |
3148 |
++ int dispatched = 0; |
3149 |
++ struct request *rq; |
3150 |
++ unsigned long service_to_charge; |
3151 |
++ |
3152 |
++ BUG_ON(RB_EMPTY_ROOT(&bfqq->sort_list)); |
3153 |
++ |
3154 |
++ /* Follow expired path, else get first next available. */ |
3155 |
++ rq = bfq_check_fifo(bfqq); |
3156 |
++ if (rq == NULL) |
3157 |
++ rq = bfqq->next_rq; |
3158 |
++ service_to_charge = bfq_serv_to_charge(rq, bfqq); |
3159 |
++ |
3160 |
++ if (service_to_charge > bfq_bfqq_budget_left(bfqq)) { |
3161 |
++ /* |
3162 |
++ * This may happen if the next rq is chosen |
3163 |
++ * in fifo order instead of sector order. |
3164 |
++ * The budget is properly dimensioned |
3165 |
++ * to be always sufficient to serve the next request |
3166 |
++ * only if it is chosen in sector order. The reason is |
3167 |
++ * that it would be quite inefficient and little useful |
3168 |
++ * to always make sure that the budget is large enough |
3169 |
++ * to serve even the possible next rq in fifo order. |
3170 |
++ * In fact, requests are seldom served in fifo order. |
3171 |
++ * |
3172 |
++ * Expire the queue for budget exhaustion, and |
3173 |
++ * make sure that the next act_budget is enough |
3174 |
++ * to serve the next request, even if it comes |
3175 |
++ * from the fifo expired path. |
3176 |
++ */ |
3177 |
++ bfqq->next_rq = rq; |
3178 |
++ /* |
3179 |
++ * Since this dispatch is failed, make sure that |
3180 |
++ * a new one will be performed |
3181 |
++ */ |
3182 |
++ if (!bfqd->rq_in_driver) |
3183 |
++ bfq_schedule_dispatch(bfqd); |
3184 |
++ goto expire; |
3185 |
++ } |
3186 |
++ |
3187 |
++ /* Finally, insert request into driver dispatch list. */ |
3188 |
++ bfq_bfqq_served(bfqq, service_to_charge); |
3189 |
++ bfq_dispatch_insert(bfqd->queue, rq); |
3190 |
++ |
3191 |
++ bfq_update_raising_data(bfqd, bfqq); |
3192 |
++ |
3193 |
++ bfq_log_bfqq(bfqd, bfqq, |
3194 |
++ "dispatched %u sec req (%llu), budg left %lu", |
3195 |
++ blk_rq_sectors(rq), |
3196 |
++ (long long unsigned)blk_rq_pos(rq), |
3197 |
++ bfq_bfqq_budget_left(bfqq)); |
3198 |
++ |
3199 |
++ dispatched++; |
3200 |
++ |
3201 |
++ if (bfqd->in_service_bic == NULL) { |
3202 |
++ atomic_long_inc(&RQ_BIC(rq)->icq.ioc->refcount); |
3203 |
++ bfqd->in_service_bic = RQ_BIC(rq); |
3204 |
++ } |
3205 |
++ |
3206 |
++ if (bfqd->busy_queues > 1 && ((!bfq_bfqq_sync(bfqq) && |
3207 |
++ dispatched >= bfqd->bfq_max_budget_async_rq) || |
3208 |
++ bfq_class_idle(bfqq))) |
3209 |
++ goto expire; |
3210 |
++ |
3211 |
++ return dispatched; |
3212 |
++ |
3213 |
++expire: |
3214 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_EXHAUSTED); |
3215 |
++ return dispatched; |
3216 |
++} |
3217 |
++ |
3218 |
++static int __bfq_forced_dispatch_bfqq(struct bfq_queue *bfqq) |
3219 |
++{ |
3220 |
++ int dispatched = 0; |
3221 |
++ |
3222 |
++ while (bfqq->next_rq != NULL) { |
3223 |
++ bfq_dispatch_insert(bfqq->bfqd->queue, bfqq->next_rq); |
3224 |
++ dispatched++; |
3225 |
++ } |
3226 |
++ |
3227 |
++ BUG_ON(!list_empty(&bfqq->fifo)); |
3228 |
++ return dispatched; |
3229 |
++} |
3230 |
++ |
3231 |
++/* |
3232 |
++ * Drain our current requests. Used for barriers and when switching |
3233 |
++ * io schedulers on-the-fly. |
3234 |
++ */ |
3235 |
++static int bfq_forced_dispatch(struct bfq_data *bfqd) |
3236 |
++{ |
3237 |
++ struct bfq_queue *bfqq, *n; |
3238 |
++ struct bfq_service_tree *st; |
3239 |
++ int dispatched = 0; |
3240 |
++ |
3241 |
++ bfqq = bfqd->in_service_queue; |
3242 |
++ if (bfqq != NULL) |
3243 |
++ __bfq_bfqq_expire(bfqd, bfqq); |
3244 |
++ |
3245 |
++ /* |
3246 |
++ * Loop through classes, and be careful to leave the scheduler |
3247 |
++ * in a consistent state, as feedback mechanisms and vtime |
3248 |
++ * updates cannot be disabled during the process. |
3249 |
++ */ |
3250 |
++ list_for_each_entry_safe(bfqq, n, &bfqd->active_list, bfqq_list) { |
3251 |
++ st = bfq_entity_service_tree(&bfqq->entity); |
3252 |
++ |
3253 |
++ dispatched += __bfq_forced_dispatch_bfqq(bfqq); |
3254 |
++ bfqq->max_budget = bfq_max_budget(bfqd); |
3255 |
++ |
3256 |
++ bfq_forget_idle(st); |
3257 |
++ } |
3258 |
++ |
3259 |
++ BUG_ON(bfqd->busy_queues != 0); |
3260 |
++ |
3261 |
++ return dispatched; |
3262 |
++} |
3263 |
++ |
3264 |
++static int bfq_dispatch_requests(struct request_queue *q, int force) |
3265 |
++{ |
3266 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
3267 |
++ struct bfq_queue *bfqq; |
3268 |
++ int max_dispatch; |
3269 |
++ |
3270 |
++ bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues); |
3271 |
++ if (bfqd->busy_queues == 0) |
3272 |
++ return 0; |
3273 |
++ |
3274 |
++ if (unlikely(force)) |
3275 |
++ return bfq_forced_dispatch(bfqd); |
3276 |
++ |
3277 |
++ bfqq = bfq_select_queue(bfqd); |
3278 |
++ if (bfqq == NULL) |
3279 |
++ return 0; |
3280 |
++ |
3281 |
++ max_dispatch = bfqd->bfq_quantum; |
3282 |
++ if (bfq_class_idle(bfqq)) |
3283 |
++ max_dispatch = 1; |
3284 |
++ |
3285 |
++ if (!bfq_bfqq_sync(bfqq)) |
3286 |
++ max_dispatch = bfqd->bfq_max_budget_async_rq; |
3287 |
++ |
3288 |
++ if (bfqq->dispatched >= max_dispatch) { |
3289 |
++ if (bfqd->busy_queues > 1) |
3290 |
++ return 0; |
3291 |
++ if (bfqq->dispatched >= 4 * max_dispatch) |
3292 |
++ return 0; |
3293 |
++ } |
3294 |
++ |
3295 |
++ if (bfqd->sync_flight != 0 && !bfq_bfqq_sync(bfqq)) |
3296 |
++ return 0; |
3297 |
++ |
3298 |
++ bfq_clear_bfqq_wait_request(bfqq); |
3299 |
++ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
3300 |
++ |
3301 |
++ if (!bfq_dispatch_request(bfqd, bfqq)) |
3302 |
++ return 0; |
3303 |
++ |
3304 |
++ bfq_log_bfqq(bfqd, bfqq, "dispatched one request of %d (max_disp %d)", |
3305 |
++ bfqq->pid, max_dispatch); |
3306 |
++ |
3307 |
++ return 1; |
3308 |
++} |
3309 |
++ |
3310 |
++/* |
3311 |
++ * Task holds one reference to the queue, dropped when task exits. Each rq |
3312 |
++ * in-flight on this queue also holds a reference, dropped when rq is freed. |
3313 |
++ * |
3314 |
++ * Queue lock must be held here. |
3315 |
++ */ |
3316 |
++static void bfq_put_queue(struct bfq_queue *bfqq) |
3317 |
++{ |
3318 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
3319 |
++ |
3320 |
++ BUG_ON(atomic_read(&bfqq->ref) <= 0); |
3321 |
++ |
3322 |
++ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p %d", bfqq, |
3323 |
++ atomic_read(&bfqq->ref)); |
3324 |
++ if (!atomic_dec_and_test(&bfqq->ref)) |
3325 |
++ return; |
3326 |
++ |
3327 |
++ BUG_ON(rb_first(&bfqq->sort_list) != NULL); |
3328 |
++ BUG_ON(bfqq->allocated[READ] + bfqq->allocated[WRITE] != 0); |
3329 |
++ BUG_ON(bfqq->entity.tree != NULL); |
3330 |
++ BUG_ON(bfq_bfqq_busy(bfqq)); |
3331 |
++ BUG_ON(bfqd->in_service_queue == bfqq); |
3332 |
++ |
3333 |
++ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p freed", bfqq); |
3334 |
++ |
3335 |
++ kmem_cache_free(bfq_pool, bfqq); |
3336 |
++} |
3337 |
++ |
3338 |
++static void bfq_put_cooperator(struct bfq_queue *bfqq) |
3339 |
++{ |
3340 |
++ struct bfq_queue *__bfqq, *next; |
3341 |
++ |
3342 |
++ /* |
3343 |
++ * If this queue was scheduled to merge with another queue, be |
3344 |
++ * sure to drop the reference taken on that queue (and others in |
3345 |
++ * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs. |
3346 |
++ */ |
3347 |
++ __bfqq = bfqq->new_bfqq; |
3348 |
++ while (__bfqq) { |
3349 |
++ if (__bfqq == bfqq) { |
3350 |
++ WARN(1, "bfqq->new_bfqq loop detected.\n"); |
3351 |
++ break; |
3352 |
++ } |
3353 |
++ next = __bfqq->new_bfqq; |
3354 |
++ bfq_put_queue(__bfqq); |
3355 |
++ __bfqq = next; |
3356 |
++ } |
3357 |
++} |
3358 |
++ |
3359 |
++static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
3360 |
++{ |
3361 |
++ if (bfqq == bfqd->in_service_queue) { |
3362 |
++ __bfq_bfqq_expire(bfqd, bfqq); |
3363 |
++ bfq_schedule_dispatch(bfqd); |
3364 |
++ } |
3365 |
++ |
3366 |
++ bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, |
3367 |
++ atomic_read(&bfqq->ref)); |
3368 |
++ |
3369 |
++ bfq_put_cooperator(bfqq); |
3370 |
++ |
3371 |
++ bfq_put_queue(bfqq); |
3372 |
++} |
3373 |
++ |
3374 |
++static void bfq_init_icq(struct io_cq *icq) |
3375 |
++{ |
3376 |
++ struct bfq_io_cq *bic = icq_to_bic(icq); |
3377 |
++ |
3378 |
++ bic->ttime.last_end_request = jiffies; |
3379 |
++} |
3380 |
++ |
3381 |
++static void bfq_exit_icq(struct io_cq *icq) |
3382 |
++{ |
3383 |
++ struct bfq_io_cq *bic = icq_to_bic(icq); |
3384 |
++ struct bfq_data *bfqd = bic_to_bfqd(bic); |
3385 |
++ |
3386 |
++ if (bic->bfqq[BLK_RW_ASYNC]) { |
3387 |
++ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_ASYNC]); |
3388 |
++ bic->bfqq[BLK_RW_ASYNC] = NULL; |
3389 |
++ } |
3390 |
++ |
3391 |
++ if (bic->bfqq[BLK_RW_SYNC]) { |
3392 |
++ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_SYNC]); |
3393 |
++ bic->bfqq[BLK_RW_SYNC] = NULL; |
3394 |
++ } |
3395 |
++} |
3396 |
++ |
3397 |
++/* |
3398 |
++ * Update the entity prio values; note that the new values will not |
3399 |
++ * be used until the next (re)activation. |
3400 |
++ */ |
3401 |
++static void bfq_init_prio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) |
3402 |
++{ |
3403 |
++ struct task_struct *tsk = current; |
3404 |
++ int ioprio_class; |
3405 |
++ |
3406 |
++ if (!bfq_bfqq_prio_changed(bfqq)) |
3407 |
++ return; |
3408 |
++ |
3409 |
++ ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
3410 |
++ switch (ioprio_class) { |
3411 |
++ default: |
3412 |
++ dev_err(bfqq->bfqd->queue->backing_dev_info.dev, |
3413 |
++ "bfq: bad prio %x\n", ioprio_class); |
3414 |
++ case IOPRIO_CLASS_NONE: |
3415 |
++ /* |
3416 |
++ * No prio set, inherit CPU scheduling settings. |
3417 |
++ */ |
3418 |
++ bfqq->entity.new_ioprio = task_nice_ioprio(tsk); |
3419 |
++ bfqq->entity.new_ioprio_class = task_nice_ioclass(tsk); |
3420 |
++ break; |
3421 |
++ case IOPRIO_CLASS_RT: |
3422 |
++ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
3423 |
++ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_RT; |
3424 |
++ break; |
3425 |
++ case IOPRIO_CLASS_BE: |
3426 |
++ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
3427 |
++ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_BE; |
3428 |
++ break; |
3429 |
++ case IOPRIO_CLASS_IDLE: |
3430 |
++ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_IDLE; |
3431 |
++ bfqq->entity.new_ioprio = 7; |
3432 |
++ bfq_clear_bfqq_idle_window(bfqq); |
3433 |
++ break; |
3434 |
++ } |
3435 |
++ |
3436 |
++ bfqq->entity.ioprio_changed = 1; |
3437 |
++ |
3438 |
++ /* |
3439 |
++ * Keep track of original prio settings in case we have to temporarily |
3440 |
++ * elevate the priority of this queue. |
3441 |
++ */ |
3442 |
++ bfqq->org_ioprio = bfqq->entity.new_ioprio; |
3443 |
++ bfq_clear_bfqq_prio_changed(bfqq); |
3444 |
++} |
3445 |
++ |
3446 |
++static void bfq_changed_ioprio(struct bfq_io_cq *bic) |
3447 |
++{ |
3448 |
++ struct bfq_data *bfqd; |
3449 |
++ struct bfq_queue *bfqq, *new_bfqq; |
3450 |
++ struct bfq_group *bfqg; |
3451 |
++ unsigned long uninitialized_var(flags); |
3452 |
++ int ioprio = bic->icq.ioc->ioprio; |
3453 |
++ |
3454 |
++ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), |
3455 |
++ &flags); |
3456 |
++ /* |
3457 |
++ * This condition may trigger on a newly created bic, be sure to drop |
3458 |
++ * the lock before returning. |
3459 |
++ */ |
3460 |
++ if (unlikely(bfqd == NULL) || likely(bic->ioprio == ioprio)) |
3461 |
++ goto out; |
3462 |
++ |
3463 |
++ bfqq = bic->bfqq[BLK_RW_ASYNC]; |
3464 |
++ if (bfqq != NULL) { |
3465 |
++ bfqg = container_of(bfqq->entity.sched_data, struct bfq_group, |
3466 |
++ sched_data); |
3467 |
++ new_bfqq = bfq_get_queue(bfqd, bfqg, BLK_RW_ASYNC, bic, |
3468 |
++ GFP_ATOMIC); |
3469 |
++ if (new_bfqq != NULL) { |
3470 |
++ bic->bfqq[BLK_RW_ASYNC] = new_bfqq; |
3471 |
++ bfq_log_bfqq(bfqd, bfqq, |
3472 |
++ "changed_ioprio: bfqq %p %d", |
3473 |
++ bfqq, atomic_read(&bfqq->ref)); |
3474 |
++ bfq_put_queue(bfqq); |
3475 |
++ } |
3476 |
++ } |
3477 |
++ |
3478 |
++ bfqq = bic->bfqq[BLK_RW_SYNC]; |
3479 |
++ if (bfqq != NULL) |
3480 |
++ bfq_mark_bfqq_prio_changed(bfqq); |
3481 |
++ |
3482 |
++ bic->ioprio = ioprio; |
3483 |
++ |
3484 |
++out: |
3485 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
3486 |
++} |
3487 |
++ |
3488 |
++static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
3489 |
++ pid_t pid, int is_sync) |
3490 |
++{ |
3491 |
++ RB_CLEAR_NODE(&bfqq->entity.rb_node); |
3492 |
++ INIT_LIST_HEAD(&bfqq->fifo); |
3493 |
++ |
3494 |
++ atomic_set(&bfqq->ref, 0); |
3495 |
++ bfqq->bfqd = bfqd; |
3496 |
++ |
3497 |
++ bfq_mark_bfqq_prio_changed(bfqq); |
3498 |
++ |
3499 |
++ if (is_sync) { |
3500 |
++ if (!bfq_class_idle(bfqq)) |
3501 |
++ bfq_mark_bfqq_idle_window(bfqq); |
3502 |
++ bfq_mark_bfqq_sync(bfqq); |
3503 |
++ } |
3504 |
++ |
3505 |
++ /* Tentative initial value to trade off between thr and lat */ |
3506 |
++ bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
3507 |
++ bfqq->pid = pid; |
3508 |
++ |
3509 |
++ bfqq->raising_coeff = 1; |
3510 |
++ bfqq->last_rais_start_finish = 0; |
3511 |
++ bfqq->soft_rt_next_start = -1; |
3512 |
++} |
3513 |
++ |
3514 |
++static struct bfq_queue *bfq_find_alloc_queue(struct bfq_data *bfqd, |
3515 |
++ struct bfq_group *bfqg, |
3516 |
++ int is_sync, |
3517 |
++ struct bfq_io_cq *bic, |
3518 |
++ gfp_t gfp_mask) |
3519 |
++{ |
3520 |
++ struct bfq_queue *bfqq, *new_bfqq = NULL; |
3521 |
++ |
3522 |
++retry: |
3523 |
++ /* bic always exists here */ |
3524 |
++ bfqq = bic_to_bfqq(bic, is_sync); |
3525 |
++ |
3526 |
++ /* |
3527 |
++ * Always try a new alloc if we fall back to the OOM bfqq |
3528 |
++ * originally, since it should just be a temporary situation. |
3529 |
++ */ |
3530 |
++ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
3531 |
++ bfqq = NULL; |
3532 |
++ if (new_bfqq != NULL) { |
3533 |
++ bfqq = new_bfqq; |
3534 |
++ new_bfqq = NULL; |
3535 |
++ } else if (gfp_mask & __GFP_WAIT) { |
3536 |
++ spin_unlock_irq(bfqd->queue->queue_lock); |
3537 |
++ new_bfqq = kmem_cache_alloc_node(bfq_pool, |
3538 |
++ gfp_mask | __GFP_ZERO, |
3539 |
++ bfqd->queue->node); |
3540 |
++ spin_lock_irq(bfqd->queue->queue_lock); |
3541 |
++ if (new_bfqq != NULL) |
3542 |
++ goto retry; |
3543 |
++ } else { |
3544 |
++ bfqq = kmem_cache_alloc_node(bfq_pool, |
3545 |
++ gfp_mask | __GFP_ZERO, |
3546 |
++ bfqd->queue->node); |
3547 |
++ } |
3548 |
++ |
3549 |
++ if (bfqq != NULL) { |
3550 |
++ bfq_init_bfqq(bfqd, bfqq, current->pid, is_sync); |
3551 |
++ bfq_log_bfqq(bfqd, bfqq, "allocated"); |
3552 |
++ } else { |
3553 |
++ bfqq = &bfqd->oom_bfqq; |
3554 |
++ bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); |
3555 |
++ } |
3556 |
++ |
3557 |
++ bfq_init_prio_data(bfqq, bic); |
3558 |
++ bfq_init_entity(&bfqq->entity, bfqg); |
3559 |
++ } |
3560 |
++ |
3561 |
++ if (new_bfqq != NULL) |
3562 |
++ kmem_cache_free(bfq_pool, new_bfqq); |
3563 |
++ |
3564 |
++ return bfqq; |
3565 |
++} |
3566 |
++ |
3567 |
++static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, |
3568 |
++ struct bfq_group *bfqg, |
3569 |
++ int ioprio_class, int ioprio) |
3570 |
++{ |
3571 |
++ switch (ioprio_class) { |
3572 |
++ case IOPRIO_CLASS_RT: |
3573 |
++ return &bfqg->async_bfqq[0][ioprio]; |
3574 |
++ case IOPRIO_CLASS_NONE: |
3575 |
++ ioprio = IOPRIO_NORM; |
3576 |
++ /* fall through */ |
3577 |
++ case IOPRIO_CLASS_BE: |
3578 |
++ return &bfqg->async_bfqq[1][ioprio]; |
3579 |
++ case IOPRIO_CLASS_IDLE: |
3580 |
++ return &bfqg->async_idle_bfqq; |
3581 |
++ default: |
3582 |
++ BUG(); |
3583 |
++ } |
3584 |
++} |
3585 |
++ |
3586 |
++static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
3587 |
++ struct bfq_group *bfqg, int is_sync, |
3588 |
++ struct bfq_io_cq *bic, gfp_t gfp_mask) |
3589 |
++{ |
3590 |
++ const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
3591 |
++ const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
3592 |
++ struct bfq_queue **async_bfqq = NULL; |
3593 |
++ struct bfq_queue *bfqq = NULL; |
3594 |
++ |
3595 |
++ if (!is_sync) { |
3596 |
++ async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
3597 |
++ ioprio); |
3598 |
++ bfqq = *async_bfqq; |
3599 |
++ } |
3600 |
++ |
3601 |
++ if (bfqq == NULL) |
3602 |
++ bfqq = bfq_find_alloc_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
3603 |
++ |
3604 |
++ /* |
3605 |
++ * Pin the queue now that it's allocated, scheduler exit will prune it. |
3606 |
++ */ |
3607 |
++ if (!is_sync && *async_bfqq == NULL) { |
3608 |
++ atomic_inc(&bfqq->ref); |
3609 |
++ bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", |
3610 |
++ bfqq, atomic_read(&bfqq->ref)); |
3611 |
++ *async_bfqq = bfqq; |
3612 |
++ } |
3613 |
++ |
3614 |
++ atomic_inc(&bfqq->ref); |
3615 |
++ bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, |
3616 |
++ atomic_read(&bfqq->ref)); |
3617 |
++ return bfqq; |
3618 |
++} |
3619 |
++ |
3620 |
++static void bfq_update_io_thinktime(struct bfq_data *bfqd, |
3621 |
++ struct bfq_io_cq *bic) |
3622 |
++{ |
3623 |
++ unsigned long elapsed = jiffies - bic->ttime.last_end_request; |
3624 |
++ unsigned long ttime = min(elapsed, 2UL * bfqd->bfq_slice_idle); |
3625 |
++ |
3626 |
++ bic->ttime.ttime_samples = (7*bic->ttime.ttime_samples + 256) / 8; |
3627 |
++ bic->ttime.ttime_total = (7*bic->ttime.ttime_total + 256*ttime) / 8; |
3628 |
++ bic->ttime.ttime_mean = (bic->ttime.ttime_total + 128) / |
3629 |
++ bic->ttime.ttime_samples; |
3630 |
++} |
3631 |
++ |
3632 |
++static void bfq_update_io_seektime(struct bfq_data *bfqd, |
3633 |
++ struct bfq_queue *bfqq, |
3634 |
++ struct request *rq) |
3635 |
++{ |
3636 |
++ sector_t sdist; |
3637 |
++ u64 total; |
3638 |
++ |
3639 |
++ if (bfqq->last_request_pos < blk_rq_pos(rq)) |
3640 |
++ sdist = blk_rq_pos(rq) - bfqq->last_request_pos; |
3641 |
++ else |
3642 |
++ sdist = bfqq->last_request_pos - blk_rq_pos(rq); |
3643 |
++ |
3644 |
++ /* |
3645 |
++ * Don't allow the seek distance to get too large from the |
3646 |
++ * odd fragment, pagein, etc. |
3647 |
++ */ |
3648 |
++ if (bfqq->seek_samples == 0) /* first request, not really a seek */ |
3649 |
++ sdist = 0; |
3650 |
++ else if (bfqq->seek_samples <= 60) /* second & third seek */ |
3651 |
++ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*1024); |
3652 |
++ else |
3653 |
++ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*64); |
3654 |
++ |
3655 |
++ bfqq->seek_samples = (7*bfqq->seek_samples + 256) / 8; |
3656 |
++ bfqq->seek_total = (7*bfqq->seek_total + (u64)256*sdist) / 8; |
3657 |
++ total = bfqq->seek_total + (bfqq->seek_samples/2); |
3658 |
++ do_div(total, bfqq->seek_samples); |
3659 |
++ bfqq->seek_mean = (sector_t)total; |
3660 |
++ |
3661 |
++ bfq_log_bfqq(bfqd, bfqq, "dist=%llu mean=%llu", (u64)sdist, |
3662 |
++ (u64)bfqq->seek_mean); |
3663 |
++} |
3664 |
++ |
3665 |
++/* |
3666 |
++ * Disable idle window if the process thinks too long or seeks so much that |
3667 |
++ * it doesn't matter. |
3668 |
++ */ |
3669 |
++static void bfq_update_idle_window(struct bfq_data *bfqd, |
3670 |
++ struct bfq_queue *bfqq, |
3671 |
++ struct bfq_io_cq *bic) |
3672 |
++{ |
3673 |
++ int enable_idle; |
3674 |
++ |
3675 |
++ /* Don't idle for async or idle io prio class. */ |
3676 |
++ if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq)) |
3677 |
++ return; |
3678 |
++ |
3679 |
++ enable_idle = bfq_bfqq_idle_window(bfqq); |
3680 |
++ |
3681 |
++ if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
3682 |
++ bfqd->bfq_slice_idle == 0 || |
3683 |
++ (bfqd->hw_tag && BFQQ_SEEKY(bfqq) && |
3684 |
++ bfqq->raising_coeff == 1)) |
3685 |
++ enable_idle = 0; |
3686 |
++ else if (bfq_sample_valid(bic->ttime.ttime_samples)) { |
3687 |
++ if (bic->ttime.ttime_mean > bfqd->bfq_slice_idle && |
3688 |
++ bfqq->raising_coeff == 1) |
3689 |
++ enable_idle = 0; |
3690 |
++ else |
3691 |
++ enable_idle = 1; |
3692 |
++ } |
3693 |
++ bfq_log_bfqq(bfqd, bfqq, "update_idle_window: enable_idle %d", |
3694 |
++ enable_idle); |
3695 |
++ |
3696 |
++ if (enable_idle) |
3697 |
++ bfq_mark_bfqq_idle_window(bfqq); |
3698 |
++ else |
3699 |
++ bfq_clear_bfqq_idle_window(bfqq); |
3700 |
++} |
3701 |
++ |
3702 |
++/* |
3703 |
++ * Called when a new fs request (rq) is added to bfqq. Check if there's |
3704 |
++ * something we should do about it. |
3705 |
++ */ |
3706 |
++static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
3707 |
++ struct request *rq) |
3708 |
++{ |
3709 |
++ struct bfq_io_cq *bic = RQ_BIC(rq); |
3710 |
++ |
3711 |
++ if (rq->cmd_flags & REQ_META) |
3712 |
++ bfqq->meta_pending++; |
3713 |
++ |
3714 |
++ bfq_update_io_thinktime(bfqd, bic); |
3715 |
++ bfq_update_io_seektime(bfqd, bfqq, rq); |
3716 |
++ if (bfqq->entity.service > bfq_max_budget(bfqd) / 8 || |
3717 |
++ !BFQQ_SEEKY(bfqq)) |
3718 |
++ bfq_update_idle_window(bfqd, bfqq, bic); |
3719 |
++ |
3720 |
++ bfq_log_bfqq(bfqd, bfqq, |
3721 |
++ "rq_enqueued: idle_window=%d (seeky %d, mean %llu)", |
3722 |
++ bfq_bfqq_idle_window(bfqq), BFQQ_SEEKY(bfqq), |
3723 |
++ (long long unsigned)bfqq->seek_mean); |
3724 |
++ |
3725 |
++ bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
3726 |
++ |
3727 |
++ if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) { |
3728 |
++ int small_req = bfqq->queued[rq_is_sync(rq)] == 1 && |
3729 |
++ blk_rq_sectors(rq) < 32; |
3730 |
++ int budget_timeout = bfq_bfqq_budget_timeout(bfqq); |
3731 |
++ |
3732 |
++ /* |
3733 |
++ * There is just this request queued: if the request |
3734 |
++ * is small and the queue is not to be expired, then |
3735 |
++ * just exit. |
3736 |
++ * |
3737 |
++ * In this way, if the disk is being idled to wait for |
3738 |
++ * a new request from the in-service queue, we avoid |
3739 |
++ * unplugging the device and committing the disk to serve |
3740 |
++ * just a small request. On the contrary, we wait for |
3741 |
++ * the block layer to decide when to unplug the device: |
3742 |
++ * hopefully, new requests will be merged to this one |
3743 |
++ * quickly, then the device will be unplugged and |
3744 |
++ * larger requests will be dispatched. |
3745 |
++ */ |
3746 |
++ if (small_req && !budget_timeout) |
3747 |
++ return; |
3748 |
++ |
3749 |
++ /* |
3750 |
++ * A large enough request arrived, or the queue is to |
3751 |
++ * be expired: in both cases disk idling is to be |
3752 |
++ * stopped, so clear wait_request flag and reset |
3753 |
++ * timer. |
3754 |
++ */ |
3755 |
++ bfq_clear_bfqq_wait_request(bfqq); |
3756 |
++ del_timer(&bfqd->idle_slice_timer); |
3757 |
++ |
3758 |
++ /* |
3759 |
++ * The queue is not empty, because a new request just |
3760 |
++ * arrived. Hence we can safely expire the queue, in |
3761 |
++ * case of budget timeout, without risking that the |
3762 |
++ * timestamps of the queue are not updated correctly. |
3763 |
++ * See [1] for more details. |
3764 |
++ */ |
3765 |
++ if (budget_timeout) |
3766 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_TIMEOUT); |
3767 |
++ |
3768 |
++ /* |
3769 |
++ * Let the request rip immediately, or let a new queue be |
3770 |
++ * selected if bfqq has just been expired. |
3771 |
++ */ |
3772 |
++ __blk_run_queue(bfqd->queue); |
3773 |
++ } |
3774 |
++} |
3775 |
++ |
3776 |
++static void bfq_insert_request(struct request_queue *q, struct request *rq) |
3777 |
++{ |
3778 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
3779 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
3780 |
++ |
3781 |
++ assert_spin_locked(bfqd->queue->queue_lock); |
3782 |
++ bfq_init_prio_data(bfqq, RQ_BIC(rq)); |
3783 |
++ |
3784 |
++ bfq_add_rq_rb(rq); |
3785 |
++ |
3786 |
++ rq_set_fifo_time(rq, jiffies + bfqd->bfq_fifo_expire[rq_is_sync(rq)]); |
3787 |
++ list_add_tail(&rq->queuelist, &bfqq->fifo); |
3788 |
++ |
3789 |
++ bfq_rq_enqueued(bfqd, bfqq, rq); |
3790 |
++} |
3791 |
++ |
3792 |
++static void bfq_update_hw_tag(struct bfq_data *bfqd) |
3793 |
++{ |
3794 |
++ bfqd->max_rq_in_driver = max(bfqd->max_rq_in_driver, |
3795 |
++ bfqd->rq_in_driver); |
3796 |
++ |
3797 |
++ if (bfqd->hw_tag == 1) |
3798 |
++ return; |
3799 |
++ |
3800 |
++ /* |
3801 |
++ * This sample is valid if the number of outstanding requests |
3802 |
++ * is large enough to allow a queueing behavior. Note that the |
3803 |
++ * sum is not exact, as it's not taking into account deactivated |
3804 |
++ * requests. |
3805 |
++ */ |
3806 |
++ if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD) |
3807 |
++ return; |
3808 |
++ |
3809 |
++ if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) |
3810 |
++ return; |
3811 |
++ |
3812 |
++ bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; |
3813 |
++ bfqd->max_rq_in_driver = 0; |
3814 |
++ bfqd->hw_tag_samples = 0; |
3815 |
++} |
3816 |
++ |
3817 |
++static void bfq_completed_request(struct request_queue *q, struct request *rq) |
3818 |
++{ |
3819 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
3820 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
3821 |
++ const int sync = rq_is_sync(rq); |
3822 |
++ |
3823 |
++ bfq_log_bfqq(bfqd, bfqq, "completed %u sects req (%d)", |
3824 |
++ blk_rq_sectors(rq), sync); |
3825 |
++ |
3826 |
++ bfq_update_hw_tag(bfqd); |
3827 |
++ |
3828 |
++ WARN_ON(!bfqd->rq_in_driver); |
3829 |
++ WARN_ON(!bfqq->dispatched); |
3830 |
++ bfqd->rq_in_driver--; |
3831 |
++ bfqq->dispatched--; |
3832 |
++ |
3833 |
++ if (bfq_bfqq_sync(bfqq)) |
3834 |
++ bfqd->sync_flight--; |
3835 |
++ |
3836 |
++ if (sync) |
3837 |
++ RQ_BIC(rq)->ttime.last_end_request = jiffies; |
3838 |
++ |
3839 |
++ /* |
3840 |
++ * The computation of softrt_next_start was scheduled for the next |
3841 |
++ * request completion: it is now time to compute it. |
3842 |
++ */ |
3843 |
++ if (bfq_bfqq_softrt_update(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list)) |
3844 |
++ bfqq->soft_rt_next_start = |
3845 |
++ bfq_bfqq_softrt_next_start(bfqd, bfqq); |
3846 |
++ |
3847 |
++ /* |
3848 |
++ * If this is the in-service queue, check if it needs to be expired, |
3849 |
++ * or if we want to idle in case it has no pending requests. |
3850 |
++ */ |
3851 |
++ if (bfqd->in_service_queue == bfqq) { |
3852 |
++ if (bfq_bfqq_budget_new(bfqq)) |
3853 |
++ bfq_set_budget_timeout(bfqd); |
3854 |
++ |
3855 |
++ if (bfq_bfqq_must_idle(bfqq)) { |
3856 |
++ bfq_arm_slice_timer(bfqd); |
3857 |
++ goto out; |
3858 |
++ } else if (bfq_may_expire_for_budg_timeout(bfqq)) |
3859 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_TIMEOUT); |
3860 |
++ else if (RB_EMPTY_ROOT(&bfqq->sort_list) && |
3861 |
++ (bfqq->dispatched == 0 || |
3862 |
++ !bfq_bfqq_must_not_expire(bfqq))) |
3863 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, |
3864 |
++ BFQ_BFQQ_NO_MORE_REQUESTS); |
3865 |
++ } |
3866 |
++ |
3867 |
++ if (!bfqd->rq_in_driver) |
3868 |
++ bfq_schedule_dispatch(bfqd); |
3869 |
++ |
3870 |
++out: |
3871 |
++ return; |
3872 |
++} |
3873 |
++ |
3874 |
++static inline int __bfq_may_queue(struct bfq_queue *bfqq) |
3875 |
++{ |
3876 |
++ if (bfq_bfqq_wait_request(bfqq) && bfq_bfqq_must_alloc(bfqq)) { |
3877 |
++ bfq_clear_bfqq_must_alloc(bfqq); |
3878 |
++ return ELV_MQUEUE_MUST; |
3879 |
++ } |
3880 |
++ |
3881 |
++ return ELV_MQUEUE_MAY; |
3882 |
++} |
3883 |
++ |
3884 |
++static int bfq_may_queue(struct request_queue *q, int rw) |
3885 |
++{ |
3886 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
3887 |
++ struct task_struct *tsk = current; |
3888 |
++ struct bfq_io_cq *bic; |
3889 |
++ struct bfq_queue *bfqq; |
3890 |
++ |
3891 |
++ /* |
3892 |
++ * Don't force setup of a queue from here, as a call to may_queue |
3893 |
++ * does not necessarily imply that a request actually will be queued. |
3894 |
++ * So just lookup a possibly existing queue, or return 'may queue' |
3895 |
++ * if that fails. |
3896 |
++ */ |
3897 |
++ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
3898 |
++ if (bic == NULL) |
3899 |
++ return ELV_MQUEUE_MAY; |
3900 |
++ |
3901 |
++ bfqq = bic_to_bfqq(bic, rw_is_sync(rw)); |
3902 |
++ if (bfqq != NULL) { |
3903 |
++ bfq_init_prio_data(bfqq, bic); |
3904 |
++ |
3905 |
++ return __bfq_may_queue(bfqq); |
3906 |
++ } |
3907 |
++ |
3908 |
++ return ELV_MQUEUE_MAY; |
3909 |
++} |
3910 |
++ |
3911 |
++/* |
3912 |
++ * Queue lock held here. |
3913 |
++ */ |
3914 |
++static void bfq_put_request(struct request *rq) |
3915 |
++{ |
3916 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
3917 |
++ |
3918 |
++ if (bfqq != NULL) { |
3919 |
++ const int rw = rq_data_dir(rq); |
3920 |
++ |
3921 |
++ BUG_ON(!bfqq->allocated[rw]); |
3922 |
++ bfqq->allocated[rw]--; |
3923 |
++ |
3924 |
++ rq->elv.priv[0] = NULL; |
3925 |
++ rq->elv.priv[1] = NULL; |
3926 |
++ |
3927 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "put_request %p, %d", |
3928 |
++ bfqq, atomic_read(&bfqq->ref)); |
3929 |
++ bfq_put_queue(bfqq); |
3930 |
++ } |
3931 |
++} |
3932 |
++ |
3933 |
++static struct bfq_queue * |
3934 |
++bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
3935 |
++ struct bfq_queue *bfqq) |
3936 |
++{ |
3937 |
++ bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
3938 |
++ (long unsigned)bfqq->new_bfqq->pid); |
3939 |
++ bic_set_bfqq(bic, bfqq->new_bfqq, 1); |
3940 |
++ bfq_mark_bfqq_coop(bfqq->new_bfqq); |
3941 |
++ bfq_put_queue(bfqq); |
3942 |
++ return bic_to_bfqq(bic, 1); |
3943 |
++} |
3944 |
++ |
3945 |
++/* |
3946 |
++ * Returns NULL if a new bfqq should be allocated, or the old bfqq if this |
3947 |
++ * was the last process referring to said bfqq. |
3948 |
++ */ |
3949 |
++static struct bfq_queue * |
3950 |
++bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) |
3951 |
++{ |
3952 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); |
3953 |
++ if (bfqq_process_refs(bfqq) == 1) { |
3954 |
++ bfqq->pid = current->pid; |
3955 |
++ bfq_clear_bfqq_coop(bfqq); |
3956 |
++ bfq_clear_bfqq_split_coop(bfqq); |
3957 |
++ return bfqq; |
3958 |
++ } |
3959 |
++ |
3960 |
++ bic_set_bfqq(bic, NULL, 1); |
3961 |
++ |
3962 |
++ bfq_put_cooperator(bfqq); |
3963 |
++ |
3964 |
++ bfq_put_queue(bfqq); |
3965 |
++ return NULL; |
3966 |
++} |
3967 |
++ |
3968 |
++/* |
3969 |
++ * Allocate bfq data structures associated with this request. |
3970 |
++ */ |
3971 |
++static int bfq_set_request(struct request_queue *q, struct request *rq, |
3972 |
++ struct bio *bio, gfp_t gfp_mask) |
3973 |
++{ |
3974 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
3975 |
++ struct bfq_io_cq *bic = icq_to_bic(rq->elv.icq); |
3976 |
++ const int rw = rq_data_dir(rq); |
3977 |
++ const int is_sync = rq_is_sync(rq); |
3978 |
++ struct bfq_queue *bfqq; |
3979 |
++ struct bfq_group *bfqg; |
3980 |
++ unsigned long flags; |
3981 |
++ |
3982 |
++ might_sleep_if(gfp_mask & __GFP_WAIT); |
3983 |
++ |
3984 |
++ bfq_changed_ioprio(bic); |
3985 |
++ |
3986 |
++ spin_lock_irqsave(q->queue_lock, flags); |
3987 |
++ |
3988 |
++ if (bic == NULL) |
3989 |
++ goto queue_fail; |
3990 |
++ |
3991 |
++ bfqg = bfq_bic_update_cgroup(bic); |
3992 |
++ |
3993 |
++new_queue: |
3994 |
++ bfqq = bic_to_bfqq(bic, is_sync); |
3995 |
++ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
3996 |
++ bfqq = bfq_get_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
3997 |
++ bic_set_bfqq(bic, bfqq, is_sync); |
3998 |
++ } else { |
3999 |
++ /* |
4000 |
++ * If the queue was seeky for too long, break it apart. |
4001 |
++ */ |
4002 |
++ if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { |
4003 |
++ bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); |
4004 |
++ bfqq = bfq_split_bfqq(bic, bfqq); |
4005 |
++ if (!bfqq) |
4006 |
++ goto new_queue; |
4007 |
++ } |
4008 |
++ |
4009 |
++ /* |
4010 |
++ * Check to see if this queue is scheduled to merge with |
4011 |
++ * another closely cooperating queue. The merging of queues |
4012 |
++ * happens here as it must be done in process context. |
4013 |
++ * The reference on new_bfqq was taken in merge_bfqqs. |
4014 |
++ */ |
4015 |
++ if (bfqq->new_bfqq != NULL) |
4016 |
++ bfqq = bfq_merge_bfqqs(bfqd, bic, bfqq); |
4017 |
++ } |
4018 |
++ |
4019 |
++ bfqq->allocated[rw]++; |
4020 |
++ atomic_inc(&bfqq->ref); |
4021 |
++ bfq_log_bfqq(bfqd, bfqq, "set_request: bfqq %p, %d", bfqq, |
4022 |
++ atomic_read(&bfqq->ref)); |
4023 |
++ |
4024 |
++ rq->elv.priv[0] = bic; |
4025 |
++ rq->elv.priv[1] = bfqq; |
4026 |
++ |
4027 |
++ spin_unlock_irqrestore(q->queue_lock, flags); |
4028 |
++ |
4029 |
++ return 0; |
4030 |
++ |
4031 |
++queue_fail: |
4032 |
++ bfq_schedule_dispatch(bfqd); |
4033 |
++ spin_unlock_irqrestore(q->queue_lock, flags); |
4034 |
++ |
4035 |
++ return 1; |
4036 |
++} |
4037 |
++ |
4038 |
++static void bfq_kick_queue(struct work_struct *work) |
4039 |
++{ |
4040 |
++ struct bfq_data *bfqd = |
4041 |
++ container_of(work, struct bfq_data, unplug_work); |
4042 |
++ struct request_queue *q = bfqd->queue; |
4043 |
++ |
4044 |
++ spin_lock_irq(q->queue_lock); |
4045 |
++ __blk_run_queue(q); |
4046 |
++ spin_unlock_irq(q->queue_lock); |
4047 |
++} |
4048 |
++ |
4049 |
++/* |
4050 |
++ * Handler of the expiration of the timer running if the in-service queue |
4051 |
++ * is idling inside its time slice. |
4052 |
++ */ |
4053 |
++static void bfq_idle_slice_timer(unsigned long data) |
4054 |
++{ |
4055 |
++ struct bfq_data *bfqd = (struct bfq_data *)data; |
4056 |
++ struct bfq_queue *bfqq; |
4057 |
++ unsigned long flags; |
4058 |
++ enum bfqq_expiration reason; |
4059 |
++ |
4060 |
++ spin_lock_irqsave(bfqd->queue->queue_lock, flags); |
4061 |
++ |
4062 |
++ bfqq = bfqd->in_service_queue; |
4063 |
++ /* |
4064 |
++ * Theoretical race here: the in-service queue can be NULL or different |
4065 |
++ * from the queue that was idling if the timer handler spins on |
4066 |
++ * the queue_lock and a new request arrives for the current |
4067 |
++ * queue and there is a full dispatch cycle that changes the |
4068 |
++ * in-service queue. This can hardly happen, but in the worst case |
4069 |
++ * we just expire a queue too early. |
4070 |
++ */ |
4071 |
++ if (bfqq != NULL) { |
4072 |
++ bfq_log_bfqq(bfqd, bfqq, "slice_timer expired"); |
4073 |
++ if (bfq_bfqq_budget_timeout(bfqq)) |
4074 |
++ /* |
4075 |
++ * Also here the queue can be safely expired |
4076 |
++ * for budget timeout without wasting |
4077 |
++ * guarantees |
4078 |
++ */ |
4079 |
++ reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
4080 |
++ else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) |
4081 |
++ /* |
4082 |
++ * The queue may not be empty upon timer expiration, |
4083 |
++ * because we may not disable the timer when the first |
4084 |
++ * request of the in-service queue arrives during |
4085 |
++ * disk idling |
4086 |
++ */ |
4087 |
++ reason = BFQ_BFQQ_TOO_IDLE; |
4088 |
++ else |
4089 |
++ goto schedule_dispatch; |
4090 |
++ |
4091 |
++ bfq_bfqq_expire(bfqd, bfqq, 1, reason); |
4092 |
++ } |
4093 |
++ |
4094 |
++schedule_dispatch: |
4095 |
++ bfq_schedule_dispatch(bfqd); |
4096 |
++ |
4097 |
++ spin_unlock_irqrestore(bfqd->queue->queue_lock, flags); |
4098 |
++} |
4099 |
++ |
4100 |
++static void bfq_shutdown_timer_wq(struct bfq_data *bfqd) |
4101 |
++{ |
4102 |
++ del_timer_sync(&bfqd->idle_slice_timer); |
4103 |
++ cancel_work_sync(&bfqd->unplug_work); |
4104 |
++} |
4105 |
++ |
4106 |
++static inline void __bfq_put_async_bfqq(struct bfq_data *bfqd, |
4107 |
++ struct bfq_queue **bfqq_ptr) |
4108 |
++{ |
4109 |
++ struct bfq_group *root_group = bfqd->root_group; |
4110 |
++ struct bfq_queue *bfqq = *bfqq_ptr; |
4111 |
++ |
4112 |
++ bfq_log(bfqd, "put_async_bfqq: %p", bfqq); |
4113 |
++ if (bfqq != NULL) { |
4114 |
++ bfq_bfqq_move(bfqd, bfqq, &bfqq->entity, root_group); |
4115 |
++ bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
4116 |
++ bfqq, atomic_read(&bfqq->ref)); |
4117 |
++ bfq_put_queue(bfqq); |
4118 |
++ *bfqq_ptr = NULL; |
4119 |
++ } |
4120 |
++} |
4121 |
++ |
4122 |
++/* |
4123 |
++ * Release all the bfqg references to its async queues. If we are |
4124 |
++ * deallocating the group these queues may still contain requests, so |
4125 |
++ * we reparent them to the root cgroup (i.e., the only one that will |
4126 |
++ * exist for sure untill all the requests on a device are gone). |
4127 |
++ */ |
4128 |
++static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
4129 |
++{ |
4130 |
++ int i, j; |
4131 |
++ |
4132 |
++ for (i = 0; i < 2; i++) |
4133 |
++ for (j = 0; j < IOPRIO_BE_NR; j++) |
4134 |
++ __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]); |
4135 |
++ |
4136 |
++ __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq); |
4137 |
++} |
4138 |
++ |
4139 |
++static void bfq_exit_queue(struct elevator_queue *e) |
4140 |
++{ |
4141 |
++ struct bfq_data *bfqd = e->elevator_data; |
4142 |
++ struct request_queue *q = bfqd->queue; |
4143 |
++ struct bfq_queue *bfqq, *n; |
4144 |
++ |
4145 |
++ bfq_shutdown_timer_wq(bfqd); |
4146 |
++ |
4147 |
++ spin_lock_irq(q->queue_lock); |
4148 |
++ |
4149 |
++ BUG_ON(bfqd->in_service_queue != NULL); |
4150 |
++ list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) |
4151 |
++ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
4152 |
++ |
4153 |
++ bfq_disconnect_groups(bfqd); |
4154 |
++ spin_unlock_irq(q->queue_lock); |
4155 |
++ |
4156 |
++ bfq_shutdown_timer_wq(bfqd); |
4157 |
++ |
4158 |
++ synchronize_rcu(); |
4159 |
++ |
4160 |
++ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
4161 |
++ |
4162 |
++ bfq_free_root_group(bfqd); |
4163 |
++ kfree(bfqd); |
4164 |
++} |
4165 |
++ |
4166 |
++static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
4167 |
++{ |
4168 |
++ struct bfq_group *bfqg; |
4169 |
++ struct bfq_data *bfqd; |
4170 |
++ struct elevator_queue *eq; |
4171 |
++ |
4172 |
++ eq = elevator_alloc(q, e); |
4173 |
++ if (eq == NULL) |
4174 |
++ return -ENOMEM; |
4175 |
++ |
4176 |
++ bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node); |
4177 |
++ if (bfqd == NULL) { |
4178 |
++ kobject_put(&eq->kobj); |
4179 |
++ return -ENOMEM; |
4180 |
++ } |
4181 |
++ eq->elevator_data = bfqd; |
4182 |
++ |
4183 |
++ /* |
4184 |
++ * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. |
4185 |
++ * Grab a permanent reference to it, so that the normal code flow |
4186 |
++ * will not attempt to free it. |
4187 |
++ */ |
4188 |
++ bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, 1, 0); |
4189 |
++ atomic_inc(&bfqd->oom_bfqq.ref); |
4190 |
++ |
4191 |
++ bfqd->queue = q; |
4192 |
++ |
4193 |
++ spin_lock_irq(q->queue_lock); |
4194 |
++ q->elevator = eq; |
4195 |
++ spin_unlock_irq(q->queue_lock); |
4196 |
++ |
4197 |
++ bfqg = bfq_alloc_root_group(bfqd, q->node); |
4198 |
++ if (bfqg == NULL) { |
4199 |
++ kfree(bfqd); |
4200 |
++ kobject_put(&eq->kobj); |
4201 |
++ return -ENOMEM; |
4202 |
++ } |
4203 |
++ |
4204 |
++ bfqd->root_group = bfqg; |
4205 |
++ |
4206 |
++ init_timer(&bfqd->idle_slice_timer); |
4207 |
++ bfqd->idle_slice_timer.function = bfq_idle_slice_timer; |
4208 |
++ bfqd->idle_slice_timer.data = (unsigned long)bfqd; |
4209 |
++ |
4210 |
++ bfqd->rq_pos_tree = RB_ROOT; |
4211 |
++ |
4212 |
++ INIT_WORK(&bfqd->unplug_work, bfq_kick_queue); |
4213 |
++ |
4214 |
++ INIT_LIST_HEAD(&bfqd->active_list); |
4215 |
++ INIT_LIST_HEAD(&bfqd->idle_list); |
4216 |
++ |
4217 |
++ bfqd->hw_tag = -1; |
4218 |
++ |
4219 |
++ bfqd->bfq_max_budget = bfq_default_max_budget; |
4220 |
++ |
4221 |
++ bfqd->bfq_quantum = bfq_quantum; |
4222 |
++ bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; |
4223 |
++ bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; |
4224 |
++ bfqd->bfq_back_max = bfq_back_max; |
4225 |
++ bfqd->bfq_back_penalty = bfq_back_penalty; |
4226 |
++ bfqd->bfq_slice_idle = bfq_slice_idle; |
4227 |
++ bfqd->bfq_class_idle_last_service = 0; |
4228 |
++ bfqd->bfq_max_budget_async_rq = bfq_max_budget_async_rq; |
4229 |
++ bfqd->bfq_timeout[BLK_RW_ASYNC] = bfq_timeout_async; |
4230 |
++ bfqd->bfq_timeout[BLK_RW_SYNC] = bfq_timeout_sync; |
4231 |
++ |
4232 |
++ bfqd->low_latency = true; |
4233 |
++ |
4234 |
++ bfqd->bfq_raising_coeff = 20; |
4235 |
++ bfqd->bfq_raising_rt_max_time = msecs_to_jiffies(300); |
4236 |
++ bfqd->bfq_raising_max_time = 0; |
4237 |
++ bfqd->bfq_raising_min_idle_time = msecs_to_jiffies(2000); |
4238 |
++ bfqd->bfq_raising_min_inter_arr_async = msecs_to_jiffies(500); |
4239 |
++ bfqd->bfq_raising_max_softrt_rate = 7000; /* |
4240 |
++ * Approximate rate required |
4241 |
++ * to playback or record a |
4242 |
++ * high-definition compressed |
4243 |
++ * video. |
4244 |
++ */ |
4245 |
++ bfqd->raised_busy_queues = 0; |
4246 |
++ |
4247 |
++ /* Initially estimate the device's peak rate as the reference rate */ |
4248 |
++ if (blk_queue_nonrot(bfqd->queue)) { |
4249 |
++ bfqd->RT_prod = R_nonrot * T_nonrot; |
4250 |
++ bfqd->peak_rate = R_nonrot; |
4251 |
++ } else { |
4252 |
++ bfqd->RT_prod = R_rot * T_rot; |
4253 |
++ bfqd->peak_rate = R_rot; |
4254 |
++ } |
4255 |
++ |
4256 |
++ return 0; |
4257 |
++} |
4258 |
++ |
4259 |
++static void bfq_slab_kill(void) |
4260 |
++{ |
4261 |
++ if (bfq_pool != NULL) |
4262 |
++ kmem_cache_destroy(bfq_pool); |
4263 |
++} |
4264 |
++ |
4265 |
++static int __init bfq_slab_setup(void) |
4266 |
++{ |
4267 |
++ bfq_pool = KMEM_CACHE(bfq_queue, 0); |
4268 |
++ if (bfq_pool == NULL) |
4269 |
++ return -ENOMEM; |
4270 |
++ return 0; |
4271 |
++} |
4272 |
++ |
4273 |
++static ssize_t bfq_var_show(unsigned int var, char *page) |
4274 |
++{ |
4275 |
++ return sprintf(page, "%d\n", var); |
4276 |
++} |
4277 |
++ |
4278 |
++static ssize_t bfq_var_store(unsigned long *var, const char *page, size_t count) |
4279 |
++{ |
4280 |
++ unsigned long new_val; |
4281 |
++ int ret = kstrtoul(page, 10, &new_val); |
4282 |
++ |
4283 |
++ if (ret == 0) |
4284 |
++ *var = new_val; |
4285 |
++ |
4286 |
++ return count; |
4287 |
++} |
4288 |
++ |
4289 |
++static ssize_t bfq_raising_max_time_show(struct elevator_queue *e, char *page) |
4290 |
++{ |
4291 |
++ struct bfq_data *bfqd = e->elevator_data; |
4292 |
++ return sprintf(page, "%d\n", bfqd->bfq_raising_max_time > 0 ? |
4293 |
++ jiffies_to_msecs(bfqd->bfq_raising_max_time) : |
4294 |
++ jiffies_to_msecs(bfq_wrais_duration(bfqd))); |
4295 |
++} |
4296 |
++ |
4297 |
++static ssize_t bfq_weights_show(struct elevator_queue *e, char *page) |
4298 |
++{ |
4299 |
++ struct bfq_queue *bfqq; |
4300 |
++ struct bfq_data *bfqd = e->elevator_data; |
4301 |
++ ssize_t num_char = 0; |
4302 |
++ |
4303 |
++ num_char += sprintf(page + num_char, "Tot reqs queued %d\n\n", |
4304 |
++ bfqd->queued); |
4305 |
++ |
4306 |
++ spin_lock_irq(bfqd->queue->queue_lock); |
4307 |
++ |
4308 |
++ num_char += sprintf(page + num_char, "Active:\n"); |
4309 |
++ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) { |
4310 |
++ num_char += sprintf(page + num_char, |
4311 |
++ "pid%d: weight %hu, nr_queued %d %d," |
4312 |
++ " dur %d/%u\n", |
4313 |
++ bfqq->pid, |
4314 |
++ bfqq->entity.weight, |
4315 |
++ bfqq->queued[0], |
4316 |
++ bfqq->queued[1], |
4317 |
++ jiffies_to_msecs(jiffies - |
4318 |
++ bfqq->last_rais_start_finish), |
4319 |
++ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
4320 |
++ } |
4321 |
++ |
4322 |
++ num_char += sprintf(page + num_char, "Idle:\n"); |
4323 |
++ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) { |
4324 |
++ num_char += sprintf(page + num_char, |
4325 |
++ "pid%d: weight %hu, dur %d/%u\n", |
4326 |
++ bfqq->pid, |
4327 |
++ bfqq->entity.weight, |
4328 |
++ jiffies_to_msecs(jiffies - |
4329 |
++ bfqq->last_rais_start_finish), |
4330 |
++ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
4331 |
++ } |
4332 |
++ |
4333 |
++ spin_unlock_irq(bfqd->queue->queue_lock); |
4334 |
++ |
4335 |
++ return num_char; |
4336 |
++} |
4337 |
++ |
4338 |
++#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
4339 |
++static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
4340 |
++{ \ |
4341 |
++ struct bfq_data *bfqd = e->elevator_data; \ |
4342 |
++ unsigned int __data = __VAR; \ |
4343 |
++ if (__CONV) \ |
4344 |
++ __data = jiffies_to_msecs(__data); \ |
4345 |
++ return bfq_var_show(__data, (page)); \ |
4346 |
++} |
4347 |
++SHOW_FUNCTION(bfq_quantum_show, bfqd->bfq_quantum, 0); |
4348 |
++SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 1); |
4349 |
++SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 1); |
4350 |
++SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); |
4351 |
++SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); |
4352 |
++SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 1); |
4353 |
++SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); |
4354 |
++SHOW_FUNCTION(bfq_max_budget_async_rq_show, bfqd->bfq_max_budget_async_rq, 0); |
4355 |
++SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout[BLK_RW_SYNC], 1); |
4356 |
++SHOW_FUNCTION(bfq_timeout_async_show, bfqd->bfq_timeout[BLK_RW_ASYNC], 1); |
4357 |
++SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
4358 |
++SHOW_FUNCTION(bfq_raising_coeff_show, bfqd->bfq_raising_coeff, 0); |
4359 |
++SHOW_FUNCTION(bfq_raising_rt_max_time_show, bfqd->bfq_raising_rt_max_time, 1); |
4360 |
++SHOW_FUNCTION(bfq_raising_min_idle_time_show, bfqd->bfq_raising_min_idle_time, |
4361 |
++ 1); |
4362 |
++SHOW_FUNCTION(bfq_raising_min_inter_arr_async_show, |
4363 |
++ bfqd->bfq_raising_min_inter_arr_async, |
4364 |
++ 1); |
4365 |
++SHOW_FUNCTION(bfq_raising_max_softrt_rate_show, |
4366 |
++ bfqd->bfq_raising_max_softrt_rate, 0); |
4367 |
++#undef SHOW_FUNCTION |
4368 |
++ |
4369 |
++#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ |
4370 |
++static ssize_t \ |
4371 |
++__FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
4372 |
++{ \ |
4373 |
++ struct bfq_data *bfqd = e->elevator_data; \ |
4374 |
++ unsigned long uninitialized_var(__data); \ |
4375 |
++ int ret = bfq_var_store(&__data, (page), count); \ |
4376 |
++ if (__data < (MIN)) \ |
4377 |
++ __data = (MIN); \ |
4378 |
++ else if (__data > (MAX)) \ |
4379 |
++ __data = (MAX); \ |
4380 |
++ if (__CONV) \ |
4381 |
++ *(__PTR) = msecs_to_jiffies(__data); \ |
4382 |
++ else \ |
4383 |
++ *(__PTR) = __data; \ |
4384 |
++ return ret; \ |
4385 |
++} |
4386 |
++STORE_FUNCTION(bfq_quantum_store, &bfqd->bfq_quantum, 1, INT_MAX, 0); |
4387 |
++STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, |
4388 |
++ INT_MAX, 1); |
4389 |
++STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, |
4390 |
++ INT_MAX, 1); |
4391 |
++STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); |
4392 |
++STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, |
4393 |
++ INT_MAX, 0); |
4394 |
++STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 1); |
4395 |
++STORE_FUNCTION(bfq_max_budget_async_rq_store, &bfqd->bfq_max_budget_async_rq, |
4396 |
++ 1, INT_MAX, 0); |
4397 |
++STORE_FUNCTION(bfq_timeout_async_store, &bfqd->bfq_timeout[BLK_RW_ASYNC], 0, |
4398 |
++ INT_MAX, 1); |
4399 |
++STORE_FUNCTION(bfq_raising_coeff_store, &bfqd->bfq_raising_coeff, 1, |
4400 |
++ INT_MAX, 0); |
4401 |
++STORE_FUNCTION(bfq_raising_max_time_store, &bfqd->bfq_raising_max_time, 0, |
4402 |
++ INT_MAX, 1); |
4403 |
++STORE_FUNCTION(bfq_raising_rt_max_time_store, &bfqd->bfq_raising_rt_max_time, 0, |
4404 |
++ INT_MAX, 1); |
4405 |
++STORE_FUNCTION(bfq_raising_min_idle_time_store, |
4406 |
++ &bfqd->bfq_raising_min_idle_time, 0, INT_MAX, 1); |
4407 |
++STORE_FUNCTION(bfq_raising_min_inter_arr_async_store, |
4408 |
++ &bfqd->bfq_raising_min_inter_arr_async, 0, INT_MAX, 1); |
4409 |
++STORE_FUNCTION(bfq_raising_max_softrt_rate_store, |
4410 |
++ &bfqd->bfq_raising_max_softrt_rate, 0, INT_MAX, 0); |
4411 |
++#undef STORE_FUNCTION |
4412 |
++ |
4413 |
++/* do nothing for the moment */ |
4414 |
++static ssize_t bfq_weights_store(struct elevator_queue *e, |
4415 |
++ const char *page, size_t count) |
4416 |
++{ |
4417 |
++ return count; |
4418 |
++} |
4419 |
++ |
4420 |
++static inline unsigned long bfq_estimated_max_budget(struct bfq_data *bfqd) |
4421 |
++{ |
4422 |
++ u64 timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
4423 |
++ |
4424 |
++ if (bfqd->peak_rate_samples >= BFQ_PEAK_RATE_SAMPLES) |
4425 |
++ return bfq_calc_max_budget(bfqd->peak_rate, timeout); |
4426 |
++ else |
4427 |
++ return bfq_default_max_budget; |
4428 |
++} |
4429 |
++ |
4430 |
++static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
4431 |
++ const char *page, size_t count) |
4432 |
++{ |
4433 |
++ struct bfq_data *bfqd = e->elevator_data; |
4434 |
++ unsigned long uninitialized_var(__data); |
4435 |
++ int ret = bfq_var_store(&__data, (page), count); |
4436 |
++ |
4437 |
++ if (__data == 0) |
4438 |
++ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
4439 |
++ else { |
4440 |
++ if (__data > INT_MAX) |
4441 |
++ __data = INT_MAX; |
4442 |
++ bfqd->bfq_max_budget = __data; |
4443 |
++ } |
4444 |
++ |
4445 |
++ bfqd->bfq_user_max_budget = __data; |
4446 |
++ |
4447 |
++ return ret; |
4448 |
++} |
4449 |
++ |
4450 |
++static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, |
4451 |
++ const char *page, size_t count) |
4452 |
++{ |
4453 |
++ struct bfq_data *bfqd = e->elevator_data; |
4454 |
++ unsigned long uninitialized_var(__data); |
4455 |
++ int ret = bfq_var_store(&__data, (page), count); |
4456 |
++ |
4457 |
++ if (__data < 1) |
4458 |
++ __data = 1; |
4459 |
++ else if (__data > INT_MAX) |
4460 |
++ __data = INT_MAX; |
4461 |
++ |
4462 |
++ bfqd->bfq_timeout[BLK_RW_SYNC] = msecs_to_jiffies(__data); |
4463 |
++ if (bfqd->bfq_user_max_budget == 0) |
4464 |
++ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
4465 |
++ |
4466 |
++ return ret; |
4467 |
++} |
4468 |
++ |
4469 |
++static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
4470 |
++ const char *page, size_t count) |
4471 |
++{ |
4472 |
++ struct bfq_data *bfqd = e->elevator_data; |
4473 |
++ unsigned long uninitialized_var(__data); |
4474 |
++ int ret = bfq_var_store(&__data, (page), count); |
4475 |
++ |
4476 |
++ if (__data > 1) |
4477 |
++ __data = 1; |
4478 |
++ if (__data == 0 && bfqd->low_latency != 0) |
4479 |
++ bfq_end_raising(bfqd); |
4480 |
++ bfqd->low_latency = __data; |
4481 |
++ |
4482 |
++ return ret; |
4483 |
++} |
4484 |
++ |
4485 |
++#define BFQ_ATTR(name) \ |
4486 |
++ __ATTR(name, S_IRUGO|S_IWUSR, bfq_##name##_show, bfq_##name##_store) |
4487 |
++ |
4488 |
++static struct elv_fs_entry bfq_attrs[] = { |
4489 |
++ BFQ_ATTR(quantum), |
4490 |
++ BFQ_ATTR(fifo_expire_sync), |
4491 |
++ BFQ_ATTR(fifo_expire_async), |
4492 |
++ BFQ_ATTR(back_seek_max), |
4493 |
++ BFQ_ATTR(back_seek_penalty), |
4494 |
++ BFQ_ATTR(slice_idle), |
4495 |
++ BFQ_ATTR(max_budget), |
4496 |
++ BFQ_ATTR(max_budget_async_rq), |
4497 |
++ BFQ_ATTR(timeout_sync), |
4498 |
++ BFQ_ATTR(timeout_async), |
4499 |
++ BFQ_ATTR(low_latency), |
4500 |
++ BFQ_ATTR(raising_coeff), |
4501 |
++ BFQ_ATTR(raising_max_time), |
4502 |
++ BFQ_ATTR(raising_rt_max_time), |
4503 |
++ BFQ_ATTR(raising_min_idle_time), |
4504 |
++ BFQ_ATTR(raising_min_inter_arr_async), |
4505 |
++ BFQ_ATTR(raising_max_softrt_rate), |
4506 |
++ BFQ_ATTR(weights), |
4507 |
++ __ATTR_NULL |
4508 |
++}; |
4509 |
++ |
4510 |
++static struct elevator_type iosched_bfq = { |
4511 |
++ .ops = { |
4512 |
++ .elevator_merge_fn = bfq_merge, |
4513 |
++ .elevator_merged_fn = bfq_merged_request, |
4514 |
++ .elevator_merge_req_fn = bfq_merged_requests, |
4515 |
++ .elevator_allow_merge_fn = bfq_allow_merge, |
4516 |
++ .elevator_dispatch_fn = bfq_dispatch_requests, |
4517 |
++ .elevator_add_req_fn = bfq_insert_request, |
4518 |
++ .elevator_activate_req_fn = bfq_activate_request, |
4519 |
++ .elevator_deactivate_req_fn = bfq_deactivate_request, |
4520 |
++ .elevator_completed_req_fn = bfq_completed_request, |
4521 |
++ .elevator_former_req_fn = elv_rb_former_request, |
4522 |
++ .elevator_latter_req_fn = elv_rb_latter_request, |
4523 |
++ .elevator_init_icq_fn = bfq_init_icq, |
4524 |
++ .elevator_exit_icq_fn = bfq_exit_icq, |
4525 |
++ .elevator_set_req_fn = bfq_set_request, |
4526 |
++ .elevator_put_req_fn = bfq_put_request, |
4527 |
++ .elevator_may_queue_fn = bfq_may_queue, |
4528 |
++ .elevator_init_fn = bfq_init_queue, |
4529 |
++ .elevator_exit_fn = bfq_exit_queue, |
4530 |
++ }, |
4531 |
++ .icq_size = sizeof(struct bfq_io_cq), |
4532 |
++ .icq_align = __alignof__(struct bfq_io_cq), |
4533 |
++ .elevator_attrs = bfq_attrs, |
4534 |
++ .elevator_name = "bfq", |
4535 |
++ .elevator_owner = THIS_MODULE, |
4536 |
++}; |
4537 |
++ |
4538 |
++static int __init bfq_init(void) |
4539 |
++{ |
4540 |
++ /* |
4541 |
++ * Can be 0 on HZ < 1000 setups. |
4542 |
++ */ |
4543 |
++ if (bfq_slice_idle == 0) |
4544 |
++ bfq_slice_idle = 1; |
4545 |
++ |
4546 |
++ if (bfq_timeout_async == 0) |
4547 |
++ bfq_timeout_async = 1; |
4548 |
++ |
4549 |
++ if (bfq_slab_setup()) |
4550 |
++ return -ENOMEM; |
4551 |
++ |
4552 |
++ elv_register(&iosched_bfq); |
4553 |
++ printk(KERN_INFO "BFQ I/O-scheduler version: v7"); |
4554 |
++ |
4555 |
++ return 0; |
4556 |
++} |
4557 |
++ |
4558 |
++static void __exit bfq_exit(void) |
4559 |
++{ |
4560 |
++ elv_unregister(&iosched_bfq); |
4561 |
++ bfq_slab_kill(); |
4562 |
++} |
4563 |
++ |
4564 |
++module_init(bfq_init); |
4565 |
++module_exit(bfq_exit); |
4566 |
++ |
4567 |
++MODULE_AUTHOR("Fabio Checconi, Paolo Valente"); |
4568 |
++MODULE_LICENSE("GPL"); |
4569 |
++MODULE_DESCRIPTION("Budget Fair Queueing IO scheduler"); |
4570 |
+diff --git a/block/bfq-sched.c b/block/bfq-sched.c |
4571 |
+new file mode 100644 |
4572 |
+index 0000000..30df81c |
4573 |
+--- /dev/null |
4574 |
++++ b/block/bfq-sched.c |
4575 |
+@@ -0,0 +1,1077 @@ |
4576 |
++/* |
4577 |
++ * BFQ: Hierarchical B-WF2Q+ scheduler. |
4578 |
++ * |
4579 |
++ * Based on ideas and code from CFQ: |
4580 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
4581 |
++ * |
4582 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
4583 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
4584 |
++ * |
4585 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
4586 |
++ */ |
4587 |
++ |
4588 |
++#ifdef CONFIG_CGROUP_BFQIO |
4589 |
++#define for_each_entity(entity) \ |
4590 |
++ for (; entity != NULL; entity = entity->parent) |
4591 |
++ |
4592 |
++#define for_each_entity_safe(entity, parent) \ |
4593 |
++ for (; entity && ({ parent = entity->parent; 1; }); entity = parent) |
4594 |
++ |
4595 |
++static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
4596 |
++ int extract, |
4597 |
++ struct bfq_data *bfqd); |
4598 |
++ |
4599 |
++static inline void bfq_update_budget(struct bfq_entity *next_active) |
4600 |
++{ |
4601 |
++ struct bfq_entity *bfqg_entity; |
4602 |
++ struct bfq_group *bfqg; |
4603 |
++ struct bfq_sched_data *group_sd; |
4604 |
++ |
4605 |
++ BUG_ON(next_active == NULL); |
4606 |
++ |
4607 |
++ group_sd = next_active->sched_data; |
4608 |
++ |
4609 |
++ bfqg = container_of(group_sd, struct bfq_group, sched_data); |
4610 |
++ /* |
4611 |
++ * bfq_group's my_entity field is not NULL only if the group |
4612 |
++ * is not the root group. We must not touch the root entity |
4613 |
++ * as it must never become an active entity. |
4614 |
++ */ |
4615 |
++ bfqg_entity = bfqg->my_entity; |
4616 |
++ if (bfqg_entity != NULL) |
4617 |
++ bfqg_entity->budget = next_active->budget; |
4618 |
++} |
4619 |
++ |
4620 |
++static int bfq_update_next_active(struct bfq_sched_data *sd) |
4621 |
++{ |
4622 |
++ struct bfq_entity *next_active; |
4623 |
++ |
4624 |
++ if (sd->active_entity != NULL) |
4625 |
++ /* will update/requeue at the end of service */ |
4626 |
++ return 0; |
4627 |
++ |
4628 |
++ /* |
4629 |
++ * NOTE: this can be improved in many ways, such as returning |
4630 |
++ * 1 (and thus propagating upwards the update) only when the |
4631 |
++ * budget changes, or caching the bfqq that will be scheduled |
4632 |
++ * next from this subtree. By now we worry more about |
4633 |
++ * correctness than about performance... |
4634 |
++ */ |
4635 |
++ next_active = bfq_lookup_next_entity(sd, 0, NULL); |
4636 |
++ sd->next_active = next_active; |
4637 |
++ |
4638 |
++ if (next_active != NULL) |
4639 |
++ bfq_update_budget(next_active); |
4640 |
++ |
4641 |
++ return 1; |
4642 |
++} |
4643 |
++ |
4644 |
++static inline void bfq_check_next_active(struct bfq_sched_data *sd, |
4645 |
++ struct bfq_entity *entity) |
4646 |
++{ |
4647 |
++ BUG_ON(sd->next_active != entity); |
4648 |
++} |
4649 |
++#else |
4650 |
++#define for_each_entity(entity) \ |
4651 |
++ for (; entity != NULL; entity = NULL) |
4652 |
++ |
4653 |
++#define for_each_entity_safe(entity, parent) \ |
4654 |
++ for (parent = NULL; entity != NULL; entity = parent) |
4655 |
++ |
4656 |
++static inline int bfq_update_next_active(struct bfq_sched_data *sd) |
4657 |
++{ |
4658 |
++ return 0; |
4659 |
++} |
4660 |
++ |
4661 |
++static inline void bfq_check_next_active(struct bfq_sched_data *sd, |
4662 |
++ struct bfq_entity *entity) |
4663 |
++{ |
4664 |
++} |
4665 |
++ |
4666 |
++static inline void bfq_update_budget(struct bfq_entity *next_active) |
4667 |
++{ |
4668 |
++} |
4669 |
++#endif |
4670 |
++ |
4671 |
++/* |
4672 |
++ * Shift for timestamp calculations. This actually limits the maximum |
4673 |
++ * service allowed in one timestamp delta (small shift values increase it), |
4674 |
++ * the maximum total weight that can be used for the queues in the system |
4675 |
++ * (big shift values increase it), and the period of virtual time wraparounds. |
4676 |
++ */ |
4677 |
++#define WFQ_SERVICE_SHIFT 22 |
4678 |
++ |
4679 |
++/** |
4680 |
++ * bfq_gt - compare two timestamps. |
4681 |
++ * @a: first ts. |
4682 |
++ * @b: second ts. |
4683 |
++ * |
4684 |
++ * Return @a > @b, dealing with wrapping correctly. |
4685 |
++ */ |
4686 |
++static inline int bfq_gt(u64 a, u64 b) |
4687 |
++{ |
4688 |
++ return (s64)(a - b) > 0; |
4689 |
++} |
4690 |
++ |
4691 |
++static inline struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity) |
4692 |
++{ |
4693 |
++ struct bfq_queue *bfqq = NULL; |
4694 |
++ |
4695 |
++ BUG_ON(entity == NULL); |
4696 |
++ |
4697 |
++ if (entity->my_sched_data == NULL) |
4698 |
++ bfqq = container_of(entity, struct bfq_queue, entity); |
4699 |
++ |
4700 |
++ return bfqq; |
4701 |
++} |
4702 |
++ |
4703 |
++ |
4704 |
++/** |
4705 |
++ * bfq_delta - map service into the virtual time domain. |
4706 |
++ * @service: amount of service. |
4707 |
++ * @weight: scale factor (weight of an entity or weight sum). |
4708 |
++ */ |
4709 |
++static inline u64 bfq_delta(unsigned long service, |
4710 |
++ unsigned long weight) |
4711 |
++{ |
4712 |
++ u64 d = (u64)service << WFQ_SERVICE_SHIFT; |
4713 |
++ |
4714 |
++ do_div(d, weight); |
4715 |
++ return d; |
4716 |
++} |
4717 |
++ |
4718 |
++/** |
4719 |
++ * bfq_calc_finish - assign the finish time to an entity. |
4720 |
++ * @entity: the entity to act upon. |
4721 |
++ * @service: the service to be charged to the entity. |
4722 |
++ */ |
4723 |
++static inline void bfq_calc_finish(struct bfq_entity *entity, |
4724 |
++ unsigned long service) |
4725 |
++{ |
4726 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4727 |
++ |
4728 |
++ BUG_ON(entity->weight == 0); |
4729 |
++ |
4730 |
++ entity->finish = entity->start + |
4731 |
++ bfq_delta(service, entity->weight); |
4732 |
++ |
4733 |
++ if (bfqq != NULL) { |
4734 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, |
4735 |
++ "calc_finish: serv %lu, w %d", |
4736 |
++ service, entity->weight); |
4737 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, |
4738 |
++ "calc_finish: start %llu, finish %llu, delta %llu", |
4739 |
++ entity->start, entity->finish, |
4740 |
++ bfq_delta(service, entity->weight)); |
4741 |
++ } |
4742 |
++} |
4743 |
++ |
4744 |
++/** |
4745 |
++ * bfq_entity_of - get an entity from a node. |
4746 |
++ * @node: the node field of the entity. |
4747 |
++ * |
4748 |
++ * Convert a node pointer to the relative entity. This is used only |
4749 |
++ * to simplify the logic of some functions and not as the generic |
4750 |
++ * conversion mechanism because, e.g., in the tree walking functions, |
4751 |
++ * the check for a %NULL value would be redundant. |
4752 |
++ */ |
4753 |
++static inline struct bfq_entity *bfq_entity_of(struct rb_node *node) |
4754 |
++{ |
4755 |
++ struct bfq_entity *entity = NULL; |
4756 |
++ |
4757 |
++ if (node != NULL) |
4758 |
++ entity = rb_entry(node, struct bfq_entity, rb_node); |
4759 |
++ |
4760 |
++ return entity; |
4761 |
++} |
4762 |
++ |
4763 |
++/** |
4764 |
++ * bfq_extract - remove an entity from a tree. |
4765 |
++ * @root: the tree root. |
4766 |
++ * @entity: the entity to remove. |
4767 |
++ */ |
4768 |
++static inline void bfq_extract(struct rb_root *root, |
4769 |
++ struct bfq_entity *entity) |
4770 |
++{ |
4771 |
++ BUG_ON(entity->tree != root); |
4772 |
++ |
4773 |
++ entity->tree = NULL; |
4774 |
++ rb_erase(&entity->rb_node, root); |
4775 |
++} |
4776 |
++ |
4777 |
++/** |
4778 |
++ * bfq_idle_extract - extract an entity from the idle tree. |
4779 |
++ * @st: the service tree of the owning @entity. |
4780 |
++ * @entity: the entity being removed. |
4781 |
++ */ |
4782 |
++static void bfq_idle_extract(struct bfq_service_tree *st, |
4783 |
++ struct bfq_entity *entity) |
4784 |
++{ |
4785 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4786 |
++ struct rb_node *next; |
4787 |
++ |
4788 |
++ BUG_ON(entity->tree != &st->idle); |
4789 |
++ |
4790 |
++ if (entity == st->first_idle) { |
4791 |
++ next = rb_next(&entity->rb_node); |
4792 |
++ st->first_idle = bfq_entity_of(next); |
4793 |
++ } |
4794 |
++ |
4795 |
++ if (entity == st->last_idle) { |
4796 |
++ next = rb_prev(&entity->rb_node); |
4797 |
++ st->last_idle = bfq_entity_of(next); |
4798 |
++ } |
4799 |
++ |
4800 |
++ bfq_extract(&st->idle, entity); |
4801 |
++ |
4802 |
++ if (bfqq != NULL) |
4803 |
++ list_del(&bfqq->bfqq_list); |
4804 |
++} |
4805 |
++ |
4806 |
++/** |
4807 |
++ * bfq_insert - generic tree insertion. |
4808 |
++ * @root: tree root. |
4809 |
++ * @entity: entity to insert. |
4810 |
++ * |
4811 |
++ * This is used for the idle and the active tree, since they are both |
4812 |
++ * ordered by finish time. |
4813 |
++ */ |
4814 |
++static void bfq_insert(struct rb_root *root, struct bfq_entity *entity) |
4815 |
++{ |
4816 |
++ struct bfq_entity *entry; |
4817 |
++ struct rb_node **node = &root->rb_node; |
4818 |
++ struct rb_node *parent = NULL; |
4819 |
++ |
4820 |
++ BUG_ON(entity->tree != NULL); |
4821 |
++ |
4822 |
++ while (*node != NULL) { |
4823 |
++ parent = *node; |
4824 |
++ entry = rb_entry(parent, struct bfq_entity, rb_node); |
4825 |
++ |
4826 |
++ if (bfq_gt(entry->finish, entity->finish)) |
4827 |
++ node = &parent->rb_left; |
4828 |
++ else |
4829 |
++ node = &parent->rb_right; |
4830 |
++ } |
4831 |
++ |
4832 |
++ rb_link_node(&entity->rb_node, parent, node); |
4833 |
++ rb_insert_color(&entity->rb_node, root); |
4834 |
++ |
4835 |
++ entity->tree = root; |
4836 |
++} |
4837 |
++ |
4838 |
++/** |
4839 |
++ * bfq_update_min - update the min_start field of a entity. |
4840 |
++ * @entity: the entity to update. |
4841 |
++ * @node: one of its children. |
4842 |
++ * |
4843 |
++ * This function is called when @entity may store an invalid value for |
4844 |
++ * min_start due to updates to the active tree. The function assumes |
4845 |
++ * that the subtree rooted at @node (which may be its left or its right |
4846 |
++ * child) has a valid min_start value. |
4847 |
++ */ |
4848 |
++static inline void bfq_update_min(struct bfq_entity *entity, |
4849 |
++ struct rb_node *node) |
4850 |
++{ |
4851 |
++ struct bfq_entity *child; |
4852 |
++ |
4853 |
++ if (node != NULL) { |
4854 |
++ child = rb_entry(node, struct bfq_entity, rb_node); |
4855 |
++ if (bfq_gt(entity->min_start, child->min_start)) |
4856 |
++ entity->min_start = child->min_start; |
4857 |
++ } |
4858 |
++} |
4859 |
++ |
4860 |
++/** |
4861 |
++ * bfq_update_active_node - recalculate min_start. |
4862 |
++ * @node: the node to update. |
4863 |
++ * |
4864 |
++ * @node may have changed position or one of its children may have moved, |
4865 |
++ * this function updates its min_start value. The left and right subtrees |
4866 |
++ * are assumed to hold a correct min_start value. |
4867 |
++ */ |
4868 |
++static inline void bfq_update_active_node(struct rb_node *node) |
4869 |
++{ |
4870 |
++ struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node); |
4871 |
++ |
4872 |
++ entity->min_start = entity->start; |
4873 |
++ bfq_update_min(entity, node->rb_right); |
4874 |
++ bfq_update_min(entity, node->rb_left); |
4875 |
++} |
4876 |
++ |
4877 |
++/** |
4878 |
++ * bfq_update_active_tree - update min_start for the whole active tree. |
4879 |
++ * @node: the starting node. |
4880 |
++ * |
4881 |
++ * @node must be the deepest modified node after an update. This function |
4882 |
++ * updates its min_start using the values held by its children, assuming |
4883 |
++ * that they did not change, and then updates all the nodes that may have |
4884 |
++ * changed in the path to the root. The only nodes that may have changed |
4885 |
++ * are the ones in the path or their siblings. |
4886 |
++ */ |
4887 |
++static void bfq_update_active_tree(struct rb_node *node) |
4888 |
++{ |
4889 |
++ struct rb_node *parent; |
4890 |
++ |
4891 |
++up: |
4892 |
++ bfq_update_active_node(node); |
4893 |
++ |
4894 |
++ parent = rb_parent(node); |
4895 |
++ if (parent == NULL) |
4896 |
++ return; |
4897 |
++ |
4898 |
++ if (node == parent->rb_left && parent->rb_right != NULL) |
4899 |
++ bfq_update_active_node(parent->rb_right); |
4900 |
++ else if (parent->rb_left != NULL) |
4901 |
++ bfq_update_active_node(parent->rb_left); |
4902 |
++ |
4903 |
++ node = parent; |
4904 |
++ goto up; |
4905 |
++} |
4906 |
++ |
4907 |
++/** |
4908 |
++ * bfq_active_insert - insert an entity in the active tree of its group/device. |
4909 |
++ * @st: the service tree of the entity. |
4910 |
++ * @entity: the entity being inserted. |
4911 |
++ * |
4912 |
++ * The active tree is ordered by finish time, but an extra key is kept |
4913 |
++ * per each node, containing the minimum value for the start times of |
4914 |
++ * its children (and the node itself), so it's possible to search for |
4915 |
++ * the eligible node with the lowest finish time in logarithmic time. |
4916 |
++ */ |
4917 |
++static void bfq_active_insert(struct bfq_service_tree *st, |
4918 |
++ struct bfq_entity *entity) |
4919 |
++{ |
4920 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4921 |
++ struct rb_node *node = &entity->rb_node; |
4922 |
++ |
4923 |
++ bfq_insert(&st->active, entity); |
4924 |
++ |
4925 |
++ if (node->rb_left != NULL) |
4926 |
++ node = node->rb_left; |
4927 |
++ else if (node->rb_right != NULL) |
4928 |
++ node = node->rb_right; |
4929 |
++ |
4930 |
++ bfq_update_active_tree(node); |
4931 |
++ |
4932 |
++ if (bfqq != NULL) |
4933 |
++ list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list); |
4934 |
++} |
4935 |
++ |
4936 |
++/** |
4937 |
++ * bfq_ioprio_to_weight - calc a weight from an ioprio. |
4938 |
++ * @ioprio: the ioprio value to convert. |
4939 |
++ */ |
4940 |
++static unsigned short bfq_ioprio_to_weight(int ioprio) |
4941 |
++{ |
4942 |
++ WARN_ON(ioprio < 0 || ioprio >= IOPRIO_BE_NR); |
4943 |
++ return IOPRIO_BE_NR - ioprio; |
4944 |
++} |
4945 |
++ |
4946 |
++/** |
4947 |
++ * bfq_weight_to_ioprio - calc an ioprio from a weight. |
4948 |
++ * @weight: the weight value to convert. |
4949 |
++ * |
4950 |
++ * To preserve as mush as possible the old only-ioprio user interface, |
4951 |
++ * 0 is used as an escape ioprio value for weights (numerically) equal or |
4952 |
++ * larger than IOPRIO_BE_NR |
4953 |
++ */ |
4954 |
++static unsigned short bfq_weight_to_ioprio(int weight) |
4955 |
++{ |
4956 |
++ WARN_ON(weight < BFQ_MIN_WEIGHT || weight > BFQ_MAX_WEIGHT); |
4957 |
++ return IOPRIO_BE_NR - weight < 0 ? 0 : IOPRIO_BE_NR - weight; |
4958 |
++} |
4959 |
++ |
4960 |
++static inline void bfq_get_entity(struct bfq_entity *entity) |
4961 |
++{ |
4962 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4963 |
++ struct bfq_sched_data *sd; |
4964 |
++ |
4965 |
++ if (bfqq != NULL) { |
4966 |
++ sd = entity->sched_data; |
4967 |
++ atomic_inc(&bfqq->ref); |
4968 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d", |
4969 |
++ bfqq, atomic_read(&bfqq->ref)); |
4970 |
++ } |
4971 |
++} |
4972 |
++ |
4973 |
++/** |
4974 |
++ * bfq_find_deepest - find the deepest node that an extraction can modify. |
4975 |
++ * @node: the node being removed. |
4976 |
++ * |
4977 |
++ * Do the first step of an extraction in an rb tree, looking for the |
4978 |
++ * node that will replace @node, and returning the deepest node that |
4979 |
++ * the following modifications to the tree can touch. If @node is the |
4980 |
++ * last node in the tree return %NULL. |
4981 |
++ */ |
4982 |
++static struct rb_node *bfq_find_deepest(struct rb_node *node) |
4983 |
++{ |
4984 |
++ struct rb_node *deepest; |
4985 |
++ |
4986 |
++ if (node->rb_right == NULL && node->rb_left == NULL) |
4987 |
++ deepest = rb_parent(node); |
4988 |
++ else if (node->rb_right == NULL) |
4989 |
++ deepest = node->rb_left; |
4990 |
++ else if (node->rb_left == NULL) |
4991 |
++ deepest = node->rb_right; |
4992 |
++ else { |
4993 |
++ deepest = rb_next(node); |
4994 |
++ if (deepest->rb_right != NULL) |
4995 |
++ deepest = deepest->rb_right; |
4996 |
++ else if (rb_parent(deepest) != node) |
4997 |
++ deepest = rb_parent(deepest); |
4998 |
++ } |
4999 |
++ |
5000 |
++ return deepest; |
5001 |
++} |
5002 |
++ |
5003 |
++/** |
5004 |
++ * bfq_active_extract - remove an entity from the active tree. |
5005 |
++ * @st: the service_tree containing the tree. |
5006 |
++ * @entity: the entity being removed. |
5007 |
++ */ |
5008 |
++static void bfq_active_extract(struct bfq_service_tree *st, |
5009 |
++ struct bfq_entity *entity) |
5010 |
++{ |
5011 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
5012 |
++ struct rb_node *node; |
5013 |
++ |
5014 |
++ node = bfq_find_deepest(&entity->rb_node); |
5015 |
++ bfq_extract(&st->active, entity); |
5016 |
++ |
5017 |
++ if (node != NULL) |
5018 |
++ bfq_update_active_tree(node); |
5019 |
++ |
5020 |
++ if (bfqq != NULL) |
5021 |
++ list_del(&bfqq->bfqq_list); |
5022 |
++} |
5023 |
++ |
5024 |
++/** |
5025 |
++ * bfq_idle_insert - insert an entity into the idle tree. |
5026 |
++ * @st: the service tree containing the tree. |
5027 |
++ * @entity: the entity to insert. |
5028 |
++ */ |
5029 |
++static void bfq_idle_insert(struct bfq_service_tree *st, |
5030 |
++ struct bfq_entity *entity) |
5031 |
++{ |
5032 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
5033 |
++ struct bfq_entity *first_idle = st->first_idle; |
5034 |
++ struct bfq_entity *last_idle = st->last_idle; |
5035 |
++ |
5036 |
++ if (first_idle == NULL || bfq_gt(first_idle->finish, entity->finish)) |
5037 |
++ st->first_idle = entity; |
5038 |
++ if (last_idle == NULL || bfq_gt(entity->finish, last_idle->finish)) |
5039 |
++ st->last_idle = entity; |
5040 |
++ |
5041 |
++ bfq_insert(&st->idle, entity); |
5042 |
++ |
5043 |
++ if (bfqq != NULL) |
5044 |
++ list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list); |
5045 |
++} |
5046 |
++ |
5047 |
++/** |
5048 |
++ * bfq_forget_entity - remove an entity from the wfq trees. |
5049 |
++ * @st: the service tree. |
5050 |
++ * @entity: the entity being removed. |
5051 |
++ * |
5052 |
++ * Update the device status and forget everything about @entity, putting |
5053 |
++ * the device reference to it, if it is a queue. Entities belonging to |
5054 |
++ * groups are not refcounted. |
5055 |
++ */ |
5056 |
++static void bfq_forget_entity(struct bfq_service_tree *st, |
5057 |
++ struct bfq_entity *entity) |
5058 |
++{ |
5059 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
5060 |
++ struct bfq_sched_data *sd; |
5061 |
++ |
5062 |
++ BUG_ON(!entity->on_st); |
5063 |
++ |
5064 |
++ entity->on_st = 0; |
5065 |
++ st->wsum -= entity->weight; |
5066 |
++ if (bfqq != NULL) { |
5067 |
++ sd = entity->sched_data; |
5068 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "forget_entity: %p %d", |
5069 |
++ bfqq, atomic_read(&bfqq->ref)); |
5070 |
++ bfq_put_queue(bfqq); |
5071 |
++ } |
5072 |
++} |
5073 |
++ |
5074 |
++/** |
5075 |
++ * bfq_put_idle_entity - release the idle tree ref of an entity. |
5076 |
++ * @st: service tree for the entity. |
5077 |
++ * @entity: the entity being released. |
5078 |
++ */ |
5079 |
++static void bfq_put_idle_entity(struct bfq_service_tree *st, |
5080 |
++ struct bfq_entity *entity) |
5081 |
++{ |
5082 |
++ bfq_idle_extract(st, entity); |
5083 |
++ bfq_forget_entity(st, entity); |
5084 |
++} |
5085 |
++ |
5086 |
++/** |
5087 |
++ * bfq_forget_idle - update the idle tree if necessary. |
5088 |
++ * @st: the service tree to act upon. |
5089 |
++ * |
5090 |
++ * To preserve the global O(log N) complexity we only remove one entry here; |
5091 |
++ * as the idle tree will not grow indefinitely this can be done safely. |
5092 |
++ */ |
5093 |
++static void bfq_forget_idle(struct bfq_service_tree *st) |
5094 |
++{ |
5095 |
++ struct bfq_entity *first_idle = st->first_idle; |
5096 |
++ struct bfq_entity *last_idle = st->last_idle; |
5097 |
++ |
5098 |
++ if (RB_EMPTY_ROOT(&st->active) && last_idle != NULL && |
5099 |
++ !bfq_gt(last_idle->finish, st->vtime)) { |
5100 |
++ /* |
5101 |
++ * Forget the whole idle tree, increasing the vtime past |
5102 |
++ * the last finish time of idle entities. |
5103 |
++ */ |
5104 |
++ st->vtime = last_idle->finish; |
5105 |
++ } |
5106 |
++ |
5107 |
++ if (first_idle != NULL && !bfq_gt(first_idle->finish, st->vtime)) |
5108 |
++ bfq_put_idle_entity(st, first_idle); |
5109 |
++} |
5110 |
++ |
5111 |
++static struct bfq_service_tree * |
5112 |
++__bfq_entity_update_weight_prio(struct bfq_service_tree *old_st, |
5113 |
++ struct bfq_entity *entity) |
5114 |
++{ |
5115 |
++ struct bfq_service_tree *new_st = old_st; |
5116 |
++ |
5117 |
++ if (entity->ioprio_changed) { |
5118 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
5119 |
++ |
5120 |
++ BUG_ON(old_st->wsum < entity->weight); |
5121 |
++ old_st->wsum -= entity->weight; |
5122 |
++ |
5123 |
++ if (entity->new_weight != entity->orig_weight) { |
5124 |
++ entity->orig_weight = entity->new_weight; |
5125 |
++ entity->ioprio = |
5126 |
++ bfq_weight_to_ioprio(entity->orig_weight); |
5127 |
++ } else if (entity->new_ioprio != entity->ioprio) { |
5128 |
++ entity->ioprio = entity->new_ioprio; |
5129 |
++ entity->orig_weight = |
5130 |
++ bfq_ioprio_to_weight(entity->ioprio); |
5131 |
++ } else |
5132 |
++ entity->new_weight = entity->orig_weight = |
5133 |
++ bfq_ioprio_to_weight(entity->ioprio); |
5134 |
++ |
5135 |
++ entity->ioprio_class = entity->new_ioprio_class; |
5136 |
++ entity->ioprio_changed = 0; |
5137 |
++ |
5138 |
++ /* |
5139 |
++ * NOTE: here we may be changing the weight too early, |
5140 |
++ * this will cause unfairness. The correct approach |
5141 |
++ * would have required additional complexity to defer |
5142 |
++ * weight changes to the proper time instants (i.e., |
5143 |
++ * when entity->finish <= old_st->vtime). |
5144 |
++ */ |
5145 |
++ new_st = bfq_entity_service_tree(entity); |
5146 |
++ entity->weight = entity->orig_weight * |
5147 |
++ (bfqq != NULL ? bfqq->raising_coeff : 1); |
5148 |
++ new_st->wsum += entity->weight; |
5149 |
++ |
5150 |
++ if (new_st != old_st) |
5151 |
++ entity->start = new_st->vtime; |
5152 |
++ } |
5153 |
++ |
5154 |
++ return new_st; |
5155 |
++} |
5156 |
++ |
5157 |
++/** |
5158 |
++ * bfq_bfqq_served - update the scheduler status after selection for service. |
5159 |
++ * @bfqq: the queue being served. |
5160 |
++ * @served: bytes to transfer. |
5161 |
++ * |
5162 |
++ * NOTE: this can be optimized, as the timestamps of upper level entities |
5163 |
++ * are synchronized every time a new bfqq is selected for service. By now, |
5164 |
++ * we keep it to better check consistency. |
5165 |
++ */ |
5166 |
++static void bfq_bfqq_served(struct bfq_queue *bfqq, unsigned long served) |
5167 |
++{ |
5168 |
++ struct bfq_entity *entity = &bfqq->entity; |
5169 |
++ struct bfq_service_tree *st; |
5170 |
++ |
5171 |
++ for_each_entity(entity) { |
5172 |
++ st = bfq_entity_service_tree(entity); |
5173 |
++ |
5174 |
++ entity->service += served; |
5175 |
++ BUG_ON(entity->service > entity->budget); |
5176 |
++ BUG_ON(st->wsum == 0); |
5177 |
++ |
5178 |
++ st->vtime += bfq_delta(served, st->wsum); |
5179 |
++ bfq_forget_idle(st); |
5180 |
++ } |
5181 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %lu secs", served); |
5182 |
++} |
5183 |
++ |
5184 |
++/** |
5185 |
++ * bfq_bfqq_charge_full_budget - set the service to the entity budget. |
5186 |
++ * @bfqq: the queue that needs a service update. |
5187 |
++ * |
5188 |
++ * When it's not possible to be fair in the service domain, because |
5189 |
++ * a queue is not consuming its budget fast enough (the meaning of |
5190 |
++ * fast depends on the timeout parameter), we charge it a full |
5191 |
++ * budget. In this way we should obtain a sort of time-domain |
5192 |
++ * fairness among all the seeky/slow queues. |
5193 |
++ */ |
5194 |
++static inline void bfq_bfqq_charge_full_budget(struct bfq_queue *bfqq) |
5195 |
++{ |
5196 |
++ struct bfq_entity *entity = &bfqq->entity; |
5197 |
++ |
5198 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "charge_full_budget"); |
5199 |
++ |
5200 |
++ bfq_bfqq_served(bfqq, entity->budget - entity->service); |
5201 |
++} |
5202 |
++ |
5203 |
++/** |
5204 |
++ * __bfq_activate_entity - activate an entity. |
5205 |
++ * @entity: the entity being activated. |
5206 |
++ * |
5207 |
++ * Called whenever an entity is activated, i.e., it is not active and one |
5208 |
++ * of its children receives a new request, or has to be reactivated due to |
5209 |
++ * budget exhaustion. It uses the current budget of the entity (and the |
5210 |
++ * service received if @entity is active) of the queue to calculate its |
5211 |
++ * timestamps. |
5212 |
++ */ |
5213 |
++static void __bfq_activate_entity(struct bfq_entity *entity) |
5214 |
++{ |
5215 |
++ struct bfq_sched_data *sd = entity->sched_data; |
5216 |
++ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
5217 |
++ |
5218 |
++ if (entity == sd->active_entity) { |
5219 |
++ BUG_ON(entity->tree != NULL); |
5220 |
++ /* |
5221 |
++ * If we are requeueing the current entity we have |
5222 |
++ * to take care of not charging to it service it has |
5223 |
++ * not received. |
5224 |
++ */ |
5225 |
++ bfq_calc_finish(entity, entity->service); |
5226 |
++ entity->start = entity->finish; |
5227 |
++ sd->active_entity = NULL; |
5228 |
++ } else if (entity->tree == &st->active) { |
5229 |
++ /* |
5230 |
++ * Requeueing an entity due to a change of some |
5231 |
++ * next_active entity below it. We reuse the old |
5232 |
++ * start time. |
5233 |
++ */ |
5234 |
++ bfq_active_extract(st, entity); |
5235 |
++ } else if (entity->tree == &st->idle) { |
5236 |
++ /* |
5237 |
++ * Must be on the idle tree, bfq_idle_extract() will |
5238 |
++ * check for that. |
5239 |
++ */ |
5240 |
++ bfq_idle_extract(st, entity); |
5241 |
++ entity->start = bfq_gt(st->vtime, entity->finish) ? |
5242 |
++ st->vtime : entity->finish; |
5243 |
++ } else { |
5244 |
++ /* |
5245 |
++ * The finish time of the entity may be invalid, and |
5246 |
++ * it is in the past for sure, otherwise the queue |
5247 |
++ * would have been on the idle tree. |
5248 |
++ */ |
5249 |
++ entity->start = st->vtime; |
5250 |
++ st->wsum += entity->weight; |
5251 |
++ bfq_get_entity(entity); |
5252 |
++ |
5253 |
++ BUG_ON(entity->on_st); |
5254 |
++ entity->on_st = 1; |
5255 |
++ } |
5256 |
++ |
5257 |
++ st = __bfq_entity_update_weight_prio(st, entity); |
5258 |
++ bfq_calc_finish(entity, entity->budget); |
5259 |
++ bfq_active_insert(st, entity); |
5260 |
++} |
5261 |
++ |
5262 |
++/** |
5263 |
++ * bfq_activate_entity - activate an entity and its ancestors if necessary. |
5264 |
++ * @entity: the entity to activate. |
5265 |
++ * |
5266 |
++ * Activate @entity and all the entities on the path from it to the root. |
5267 |
++ */ |
5268 |
++static void bfq_activate_entity(struct bfq_entity *entity) |
5269 |
++{ |
5270 |
++ struct bfq_sched_data *sd; |
5271 |
++ |
5272 |
++ for_each_entity(entity) { |
5273 |
++ __bfq_activate_entity(entity); |
5274 |
++ |
5275 |
++ sd = entity->sched_data; |
5276 |
++ if (!bfq_update_next_active(sd)) |
5277 |
++ /* |
5278 |
++ * No need to propagate the activation to the |
5279 |
++ * upper entities, as they will be updated when |
5280 |
++ * the active entity is rescheduled. |
5281 |
++ */ |
5282 |
++ break; |
5283 |
++ } |
5284 |
++} |
5285 |
++ |
5286 |
++/** |
5287 |
++ * __bfq_deactivate_entity - deactivate an entity from its service tree. |
5288 |
++ * @entity: the entity to deactivate. |
5289 |
++ * @requeue: if false, the entity will not be put into the idle tree. |
5290 |
++ * |
5291 |
++ * Deactivate an entity, independently from its previous state. If the |
5292 |
++ * entity was not on a service tree just return, otherwise if it is on |
5293 |
++ * any scheduler tree, extract it from that tree, and if necessary |
5294 |
++ * and if the caller did not specify @requeue, put it on the idle tree. |
5295 |
++ * |
5296 |
++ * Return %1 if the caller should update the entity hierarchy, i.e., |
5297 |
++ * if the entity was under service or if it was the next_active for |
5298 |
++ * its sched_data; return %0 otherwise. |
5299 |
++ */ |
5300 |
++static int __bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
5301 |
++{ |
5302 |
++ struct bfq_sched_data *sd = entity->sched_data; |
5303 |
++ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
5304 |
++ int was_active = entity == sd->active_entity; |
5305 |
++ int ret = 0; |
5306 |
++ |
5307 |
++ if (!entity->on_st) |
5308 |
++ return 0; |
5309 |
++ |
5310 |
++ BUG_ON(was_active && entity->tree != NULL); |
5311 |
++ |
5312 |
++ if (was_active) { |
5313 |
++ bfq_calc_finish(entity, entity->service); |
5314 |
++ sd->active_entity = NULL; |
5315 |
++ } else if (entity->tree == &st->active) |
5316 |
++ bfq_active_extract(st, entity); |
5317 |
++ else if (entity->tree == &st->idle) |
5318 |
++ bfq_idle_extract(st, entity); |
5319 |
++ else if (entity->tree != NULL) |
5320 |
++ BUG(); |
5321 |
++ |
5322 |
++ if (was_active || sd->next_active == entity) |
5323 |
++ ret = bfq_update_next_active(sd); |
5324 |
++ |
5325 |
++ if (!requeue || !bfq_gt(entity->finish, st->vtime)) |
5326 |
++ bfq_forget_entity(st, entity); |
5327 |
++ else |
5328 |
++ bfq_idle_insert(st, entity); |
5329 |
++ |
5330 |
++ BUG_ON(sd->active_entity == entity); |
5331 |
++ BUG_ON(sd->next_active == entity); |
5332 |
++ |
5333 |
++ return ret; |
5334 |
++} |
5335 |
++ |
5336 |
++/** |
5337 |
++ * bfq_deactivate_entity - deactivate an entity. |
5338 |
++ * @entity: the entity to deactivate. |
5339 |
++ * @requeue: true if the entity can be put on the idle tree |
5340 |
++ */ |
5341 |
++static void bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
5342 |
++{ |
5343 |
++ struct bfq_sched_data *sd; |
5344 |
++ struct bfq_entity *parent; |
5345 |
++ |
5346 |
++ for_each_entity_safe(entity, parent) { |
5347 |
++ sd = entity->sched_data; |
5348 |
++ |
5349 |
++ if (!__bfq_deactivate_entity(entity, requeue)) |
5350 |
++ /* |
5351 |
++ * The parent entity is still backlogged, and |
5352 |
++ * we don't need to update it as it is still |
5353 |
++ * under service. |
5354 |
++ */ |
5355 |
++ break; |
5356 |
++ |
5357 |
++ if (sd->next_active != NULL) |
5358 |
++ /* |
5359 |
++ * The parent entity is still backlogged and |
5360 |
++ * the budgets on the path towards the root |
5361 |
++ * need to be updated. |
5362 |
++ */ |
5363 |
++ goto update; |
5364 |
++ |
5365 |
++ /* |
5366 |
++ * If we reach there the parent is no more backlogged and |
5367 |
++ * we want to propagate the dequeue upwards. |
5368 |
++ */ |
5369 |
++ requeue = 1; |
5370 |
++ } |
5371 |
++ |
5372 |
++ return; |
5373 |
++ |
5374 |
++update: |
5375 |
++ entity = parent; |
5376 |
++ for_each_entity(entity) { |
5377 |
++ __bfq_activate_entity(entity); |
5378 |
++ |
5379 |
++ sd = entity->sched_data; |
5380 |
++ if (!bfq_update_next_active(sd)) |
5381 |
++ break; |
5382 |
++ } |
5383 |
++} |
5384 |
++ |
5385 |
++/** |
5386 |
++ * bfq_update_vtime - update vtime if necessary. |
5387 |
++ * @st: the service tree to act upon. |
5388 |
++ * |
5389 |
++ * If necessary update the service tree vtime to have at least one |
5390 |
++ * eligible entity, skipping to its start time. Assumes that the |
5391 |
++ * active tree of the device is not empty. |
5392 |
++ * |
5393 |
++ * NOTE: this hierarchical implementation updates vtimes quite often, |
5394 |
++ * we may end up with reactivated tasks getting timestamps after a |
5395 |
++ * vtime skip done because we needed a ->first_active entity on some |
5396 |
++ * intermediate node. |
5397 |
++ */ |
5398 |
++static void bfq_update_vtime(struct bfq_service_tree *st) |
5399 |
++{ |
5400 |
++ struct bfq_entity *entry; |
5401 |
++ struct rb_node *node = st->active.rb_node; |
5402 |
++ |
5403 |
++ entry = rb_entry(node, struct bfq_entity, rb_node); |
5404 |
++ if (bfq_gt(entry->min_start, st->vtime)) { |
5405 |
++ st->vtime = entry->min_start; |
5406 |
++ bfq_forget_idle(st); |
5407 |
++ } |
5408 |
++} |
5409 |
++ |
5410 |
++/** |
5411 |
++ * bfq_first_active - find the eligible entity with the smallest finish time |
5412 |
++ * @st: the service tree to select from. |
5413 |
++ * |
5414 |
++ * This function searches the first schedulable entity, starting from the |
5415 |
++ * root of the tree and going on the left every time on this side there is |
5416 |
++ * a subtree with at least one eligible (start >= vtime) entity. The path |
5417 |
++ * on the right is followed only if a) the left subtree contains no eligible |
5418 |
++ * entities and b) no eligible entity has been found yet. |
5419 |
++ */ |
5420 |
++static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st) |
5421 |
++{ |
5422 |
++ struct bfq_entity *entry, *first = NULL; |
5423 |
++ struct rb_node *node = st->active.rb_node; |
5424 |
++ |
5425 |
++ while (node != NULL) { |
5426 |
++ entry = rb_entry(node, struct bfq_entity, rb_node); |
5427 |
++left: |
5428 |
++ if (!bfq_gt(entry->start, st->vtime)) |
5429 |
++ first = entry; |
5430 |
++ |
5431 |
++ BUG_ON(bfq_gt(entry->min_start, st->vtime)); |
5432 |
++ |
5433 |
++ if (node->rb_left != NULL) { |
5434 |
++ entry = rb_entry(node->rb_left, |
5435 |
++ struct bfq_entity, rb_node); |
5436 |
++ if (!bfq_gt(entry->min_start, st->vtime)) { |
5437 |
++ node = node->rb_left; |
5438 |
++ goto left; |
5439 |
++ } |
5440 |
++ } |
5441 |
++ if (first != NULL) |
5442 |
++ break; |
5443 |
++ node = node->rb_right; |
5444 |
++ } |
5445 |
++ |
5446 |
++ BUG_ON(first == NULL && !RB_EMPTY_ROOT(&st->active)); |
5447 |
++ return first; |
5448 |
++} |
5449 |
++ |
5450 |
++/** |
5451 |
++ * __bfq_lookup_next_entity - return the first eligible entity in @st. |
5452 |
++ * @st: the service tree. |
5453 |
++ * |
5454 |
++ * Update the virtual time in @st and return the first eligible entity |
5455 |
++ * it contains. |
5456 |
++ */ |
5457 |
++static struct bfq_entity *__bfq_lookup_next_entity(struct bfq_service_tree *st, |
5458 |
++ bool force) |
5459 |
++{ |
5460 |
++ struct bfq_entity *entity, *new_next_active = NULL; |
5461 |
++ |
5462 |
++ if (RB_EMPTY_ROOT(&st->active)) |
5463 |
++ return NULL; |
5464 |
++ |
5465 |
++ bfq_update_vtime(st); |
5466 |
++ entity = bfq_first_active_entity(st); |
5467 |
++ BUG_ON(bfq_gt(entity->start, st->vtime)); |
5468 |
++ |
5469 |
++ /* |
5470 |
++ * If the chosen entity does not match with the sched_data's |
5471 |
++ * next_active and we are forcedly serving the IDLE priority |
5472 |
++ * class tree, bubble up budget update. |
5473 |
++ */ |
5474 |
++ if (unlikely(force && entity != entity->sched_data->next_active)) { |
5475 |
++ new_next_active = entity; |
5476 |
++ for_each_entity(new_next_active) |
5477 |
++ bfq_update_budget(new_next_active); |
5478 |
++ } |
5479 |
++ |
5480 |
++ return entity; |
5481 |
++} |
5482 |
++ |
5483 |
++/** |
5484 |
++ * bfq_lookup_next_entity - return the first eligible entity in @sd. |
5485 |
++ * @sd: the sched_data. |
5486 |
++ * @extract: if true the returned entity will be also extracted from @sd. |
5487 |
++ * |
5488 |
++ * NOTE: since we cache the next_active entity at each level of the |
5489 |
++ * hierarchy, the complexity of the lookup can be decreased with |
5490 |
++ * absolutely no effort just returning the cached next_active value; |
5491 |
++ * we prefer to do full lookups to test the consistency of * the data |
5492 |
++ * structures. |
5493 |
++ */ |
5494 |
++static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
5495 |
++ int extract, |
5496 |
++ struct bfq_data *bfqd) |
5497 |
++{ |
5498 |
++ struct bfq_service_tree *st = sd->service_tree; |
5499 |
++ struct bfq_entity *entity; |
5500 |
++ int i = 0; |
5501 |
++ |
5502 |
++ BUG_ON(sd->active_entity != NULL); |
5503 |
++ |
5504 |
++ if (bfqd != NULL && |
5505 |
++ jiffies - bfqd->bfq_class_idle_last_service > BFQ_CL_IDLE_TIMEOUT) { |
5506 |
++ entity = __bfq_lookup_next_entity(st + BFQ_IOPRIO_CLASSES - 1, |
5507 |
++ true); |
5508 |
++ if (entity != NULL) { |
5509 |
++ i = BFQ_IOPRIO_CLASSES - 1; |
5510 |
++ bfqd->bfq_class_idle_last_service = jiffies; |
5511 |
++ sd->next_active = entity; |
5512 |
++ } |
5513 |
++ } |
5514 |
++ for (; i < BFQ_IOPRIO_CLASSES; i++) { |
5515 |
++ entity = __bfq_lookup_next_entity(st + i, false); |
5516 |
++ if (entity != NULL) { |
5517 |
++ if (extract) { |
5518 |
++ bfq_check_next_active(sd, entity); |
5519 |
++ bfq_active_extract(st + i, entity); |
5520 |
++ sd->active_entity = entity; |
5521 |
++ sd->next_active = NULL; |
5522 |
++ } |
5523 |
++ break; |
5524 |
++ } |
5525 |
++ } |
5526 |
++ |
5527 |
++ return entity; |
5528 |
++} |
5529 |
++ |
5530 |
++/* |
5531 |
++ * Get next queue for service. |
5532 |
++ */ |
5533 |
++static struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd) |
5534 |
++{ |
5535 |
++ struct bfq_entity *entity = NULL; |
5536 |
++ struct bfq_sched_data *sd; |
5537 |
++ struct bfq_queue *bfqq; |
5538 |
++ |
5539 |
++ BUG_ON(bfqd->in_service_queue != NULL); |
5540 |
++ |
5541 |
++ if (bfqd->busy_queues == 0) |
5542 |
++ return NULL; |
5543 |
++ |
5544 |
++ sd = &bfqd->root_group->sched_data; |
5545 |
++ for (; sd != NULL; sd = entity->my_sched_data) { |
5546 |
++ entity = bfq_lookup_next_entity(sd, 1, bfqd); |
5547 |
++ BUG_ON(entity == NULL); |
5548 |
++ entity->service = 0; |
5549 |
++ } |
5550 |
++ |
5551 |
++ bfqq = bfq_entity_to_bfqq(entity); |
5552 |
++ BUG_ON(bfqq == NULL); |
5553 |
++ |
5554 |
++ return bfqq; |
5555 |
++} |
5556 |
++ |
5557 |
++/* |
5558 |
++ * Forced extraction of the given queue. |
5559 |
++ */ |
5560 |
++static void bfq_get_next_queue_forced(struct bfq_data *bfqd, |
5561 |
++ struct bfq_queue *bfqq) |
5562 |
++{ |
5563 |
++ struct bfq_entity *entity; |
5564 |
++ struct bfq_sched_data *sd; |
5565 |
++ |
5566 |
++ BUG_ON(bfqd->in_service_queue != NULL); |
5567 |
++ |
5568 |
++ entity = &bfqq->entity; |
5569 |
++ /* |
5570 |
++ * Bubble up extraction/update from the leaf to the root. |
5571 |
++ */ |
5572 |
++ for_each_entity(entity) { |
5573 |
++ sd = entity->sched_data; |
5574 |
++ bfq_update_budget(entity); |
5575 |
++ bfq_update_vtime(bfq_entity_service_tree(entity)); |
5576 |
++ bfq_active_extract(bfq_entity_service_tree(entity), entity); |
5577 |
++ sd->active_entity = entity; |
5578 |
++ sd->next_active = NULL; |
5579 |
++ entity->service = 0; |
5580 |
++ } |
5581 |
++ |
5582 |
++ return; |
5583 |
++} |
5584 |
++ |
5585 |
++static void __bfq_bfqd_reset_in_service(struct bfq_data *bfqd) |
5586 |
++{ |
5587 |
++ if (bfqd->in_service_bic != NULL) { |
5588 |
++ put_io_context(bfqd->in_service_bic->icq.ioc); |
5589 |
++ bfqd->in_service_bic = NULL; |
5590 |
++ } |
5591 |
++ |
5592 |
++ bfqd->in_service_queue = NULL; |
5593 |
++ del_timer(&bfqd->idle_slice_timer); |
5594 |
++} |
5595 |
++ |
5596 |
++static void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
5597 |
++ int requeue) |
5598 |
++{ |
5599 |
++ struct bfq_entity *entity = &bfqq->entity; |
5600 |
++ |
5601 |
++ if (bfqq == bfqd->in_service_queue) |
5602 |
++ __bfq_bfqd_reset_in_service(bfqd); |
5603 |
++ |
5604 |
++ bfq_deactivate_entity(entity, requeue); |
5605 |
++} |
5606 |
++ |
5607 |
++static void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
5608 |
++{ |
5609 |
++ struct bfq_entity *entity = &bfqq->entity; |
5610 |
++ |
5611 |
++ bfq_activate_entity(entity); |
5612 |
++} |
5613 |
++ |
5614 |
++/* |
5615 |
++ * Called when the bfqq no longer has requests pending, remove it from |
5616 |
++ * the service tree. |
5617 |
++ */ |
5618 |
++static void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
5619 |
++ int requeue) |
5620 |
++{ |
5621 |
++ BUG_ON(!bfq_bfqq_busy(bfqq)); |
5622 |
++ BUG_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
5623 |
++ |
5624 |
++ bfq_log_bfqq(bfqd, bfqq, "del from busy"); |
5625 |
++ |
5626 |
++ bfq_clear_bfqq_busy(bfqq); |
5627 |
++ |
5628 |
++ BUG_ON(bfqd->busy_queues == 0); |
5629 |
++ bfqd->busy_queues--; |
5630 |
++ if (bfqq->raising_coeff > 1) |
5631 |
++ bfqd->raised_busy_queues--; |
5632 |
++ |
5633 |
++ bfq_deactivate_bfqq(bfqd, bfqq, requeue); |
5634 |
++} |
5635 |
++ |
5636 |
++/* |
5637 |
++ * Called when an inactive queue receives a new request. |
5638 |
++ */ |
5639 |
++static void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
5640 |
++{ |
5641 |
++ BUG_ON(bfq_bfqq_busy(bfqq)); |
5642 |
++ BUG_ON(bfqq == bfqd->in_service_queue); |
5643 |
++ |
5644 |
++ bfq_log_bfqq(bfqd, bfqq, "add to busy"); |
5645 |
++ |
5646 |
++ bfq_activate_bfqq(bfqd, bfqq); |
5647 |
++ |
5648 |
++ bfq_mark_bfqq_busy(bfqq); |
5649 |
++ bfqd->busy_queues++; |
5650 |
++ if (bfqq->raising_coeff > 1) |
5651 |
++ bfqd->raised_busy_queues++; |
5652 |
++} |
5653 |
+diff --git a/block/bfq.h b/block/bfq.h |
5654 |
+new file mode 100644 |
5655 |
+index 0000000..68b28e3 |
5656 |
+--- /dev/null |
5657 |
++++ b/block/bfq.h |
5658 |
+@@ -0,0 +1,614 @@ |
5659 |
++/* |
5660 |
++ * BFQ-v7 for 3.13.0: data structures and common functions prototypes. |
5661 |
++ * |
5662 |
++ * Based on ideas and code from CFQ: |
5663 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
5664 |
++ * |
5665 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
5666 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
5667 |
++ * |
5668 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
5669 |
++ */ |
5670 |
++ |
5671 |
++#ifndef _BFQ_H |
5672 |
++#define _BFQ_H |
5673 |
++ |
5674 |
++#include <linux/blktrace_api.h> |
5675 |
++#include <linux/hrtimer.h> |
5676 |
++#include <linux/ioprio.h> |
5677 |
++#include <linux/rbtree.h> |
5678 |
++ |
5679 |
++#define BFQ_IOPRIO_CLASSES 3 |
5680 |
++#define BFQ_CL_IDLE_TIMEOUT (HZ/5) |
5681 |
++ |
5682 |
++#define BFQ_MIN_WEIGHT 1 |
5683 |
++#define BFQ_MAX_WEIGHT 1000 |
5684 |
++ |
5685 |
++#define BFQ_DEFAULT_GRP_WEIGHT 10 |
5686 |
++#define BFQ_DEFAULT_GRP_IOPRIO 0 |
5687 |
++#define BFQ_DEFAULT_GRP_CLASS IOPRIO_CLASS_BE |
5688 |
++ |
5689 |
++struct bfq_entity; |
5690 |
++ |
5691 |
++/** |
5692 |
++ * struct bfq_service_tree - per ioprio_class service tree. |
5693 |
++ * @active: tree for active entities (i.e., those backlogged). |
5694 |
++ * @idle: tree for idle entities (i.e., those not backlogged, with V <= F_i). |
5695 |
++ * @first_idle: idle entity with minimum F_i. |
5696 |
++ * @last_idle: idle entity with maximum F_i. |
5697 |
++ * @vtime: scheduler virtual time. |
5698 |
++ * @wsum: scheduler weight sum; active and idle entities contribute to it. |
5699 |
++ * |
5700 |
++ * Each service tree represents a B-WF2Q+ scheduler on its own. Each |
5701 |
++ * ioprio_class has its own independent scheduler, and so its own |
5702 |
++ * bfq_service_tree. All the fields are protected by the queue lock |
5703 |
++ * of the containing bfqd. |
5704 |
++ */ |
5705 |
++struct bfq_service_tree { |
5706 |
++ struct rb_root active; |
5707 |
++ struct rb_root idle; |
5708 |
++ |
5709 |
++ struct bfq_entity *first_idle; |
5710 |
++ struct bfq_entity *last_idle; |
5711 |
++ |
5712 |
++ u64 vtime; |
5713 |
++ unsigned long wsum; |
5714 |
++}; |
5715 |
++ |
5716 |
++/** |
5717 |
++ * struct bfq_sched_data - multi-class scheduler. |
5718 |
++ * @active_entity: entity under service. |
5719 |
++ * @next_active: head-of-the-line entity in the scheduler. |
5720 |
++ * @service_tree: array of service trees, one per ioprio_class. |
5721 |
++ * |
5722 |
++ * bfq_sched_data is the basic scheduler queue. It supports three |
5723 |
++ * ioprio_classes, and can be used either as a toplevel queue or as |
5724 |
++ * an intermediate queue on a hierarchical setup. |
5725 |
++ * @next_active points to the active entity of the sched_data service |
5726 |
++ * trees that will be scheduled next. |
5727 |
++ * |
5728 |
++ * The supported ioprio_classes are the same as in CFQ, in descending |
5729 |
++ * priority order, IOPRIO_CLASS_RT, IOPRIO_CLASS_BE, IOPRIO_CLASS_IDLE. |
5730 |
++ * Requests from higher priority queues are served before all the |
5731 |
++ * requests from lower priority queues; among requests of the same |
5732 |
++ * queue requests are served according to B-WF2Q+. |
5733 |
++ * All the fields are protected by the queue lock of the containing bfqd. |
5734 |
++ */ |
5735 |
++struct bfq_sched_data { |
5736 |
++ struct bfq_entity *active_entity; |
5737 |
++ struct bfq_entity *next_active; |
5738 |
++ struct bfq_service_tree service_tree[BFQ_IOPRIO_CLASSES]; |
5739 |
++}; |
5740 |
++ |
5741 |
++/** |
5742 |
++ * struct bfq_entity - schedulable entity. |
5743 |
++ * @rb_node: service_tree member. |
5744 |
++ * @on_st: flag, true if the entity is on a tree (either the active or |
5745 |
++ * the idle one of its service_tree). |
5746 |
++ * @finish: B-WF2Q+ finish timestamp (aka F_i). |
5747 |
++ * @start: B-WF2Q+ start timestamp (aka S_i). |
5748 |
++ * @tree: tree the entity is enqueued into; %NULL if not on a tree. |
5749 |
++ * @min_start: minimum start time of the (active) subtree rooted at |
5750 |
++ * this entity; used for O(log N) lookups into active trees. |
5751 |
++ * @service: service received during the last round of service. |
5752 |
++ * @budget: budget used to calculate F_i; F_i = S_i + @budget / @weight. |
5753 |
++ * @weight: weight of the queue |
5754 |
++ * @parent: parent entity, for hierarchical scheduling. |
5755 |
++ * @my_sched_data: for non-leaf nodes in the cgroup hierarchy, the |
5756 |
++ * associated scheduler queue, %NULL on leaf nodes. |
5757 |
++ * @sched_data: the scheduler queue this entity belongs to. |
5758 |
++ * @ioprio: the ioprio in use. |
5759 |
++ * @new_weight: when a weight change is requested, the new weight value. |
5760 |
++ * @orig_weight: original weight, used to implement weight boosting |
5761 |
++ * @new_ioprio: when an ioprio change is requested, the new ioprio value. |
5762 |
++ * @ioprio_class: the ioprio_class in use. |
5763 |
++ * @new_ioprio_class: when an ioprio_class change is requested, the new |
5764 |
++ * ioprio_class value. |
5765 |
++ * @ioprio_changed: flag, true when the user requested a weight, ioprio or |
5766 |
++ * ioprio_class change. |
5767 |
++ * |
5768 |
++ * A bfq_entity is used to represent either a bfq_queue (leaf node in the |
5769 |
++ * cgroup hierarchy) or a bfq_group into the upper level scheduler. Each |
5770 |
++ * entity belongs to the sched_data of the parent group in the cgroup |
5771 |
++ * hierarchy. Non-leaf entities have also their own sched_data, stored |
5772 |
++ * in @my_sched_data. |
5773 |
++ * |
5774 |
++ * Each entity stores independently its priority values; this would |
5775 |
++ * allow different weights on different devices, but this |
5776 |
++ * functionality is not exported to userspace by now. Priorities and |
5777 |
++ * weights are updated lazily, first storing the new values into the |
5778 |
++ * new_* fields, then setting the @ioprio_changed flag. As soon as |
5779 |
++ * there is a transition in the entity state that allows the priority |
5780 |
++ * update to take place the effective and the requested priority |
5781 |
++ * values are synchronized. |
5782 |
++ * |
5783 |
++ * Unless cgroups are used, the weight value is calculated from the |
5784 |
++ * ioprio to export the same interface as CFQ. When dealing with |
5785 |
++ * ``well-behaved'' queues (i.e., queues that do not spend too much |
5786 |
++ * time to consume their budget and have true sequential behavior, and |
5787 |
++ * when there are no external factors breaking anticipation) the |
5788 |
++ * relative weights at each level of the cgroups hierarchy should be |
5789 |
++ * guaranteed. All the fields are protected by the queue lock of the |
5790 |
++ * containing bfqd. |
5791 |
++ */ |
5792 |
++struct bfq_entity { |
5793 |
++ struct rb_node rb_node; |
5794 |
++ |
5795 |
++ int on_st; |
5796 |
++ |
5797 |
++ u64 finish; |
5798 |
++ u64 start; |
5799 |
++ |
5800 |
++ struct rb_root *tree; |
5801 |
++ |
5802 |
++ u64 min_start; |
5803 |
++ |
5804 |
++ unsigned long service, budget; |
5805 |
++ unsigned short weight, new_weight; |
5806 |
++ unsigned short orig_weight; |
5807 |
++ |
5808 |
++ struct bfq_entity *parent; |
5809 |
++ |
5810 |
++ struct bfq_sched_data *my_sched_data; |
5811 |
++ struct bfq_sched_data *sched_data; |
5812 |
++ |
5813 |
++ unsigned short ioprio, new_ioprio; |
5814 |
++ unsigned short ioprio_class, new_ioprio_class; |
5815 |
++ |
5816 |
++ int ioprio_changed; |
5817 |
++}; |
5818 |
++ |
5819 |
++struct bfq_group; |
5820 |
++ |
5821 |
++/** |
5822 |
++ * struct bfq_queue - leaf schedulable entity. |
5823 |
++ * @ref: reference counter. |
5824 |
++ * @bfqd: parent bfq_data. |
5825 |
++ * @new_bfqq: shared bfq_queue if queue is cooperating with |
5826 |
++ * one or more other queues. |
5827 |
++ * @pos_node: request-position tree member (see bfq_data's @rq_pos_tree). |
5828 |
++ * @pos_root: request-position tree root (see bfq_data's @rq_pos_tree). |
5829 |
++ * @sort_list: sorted list of pending requests. |
5830 |
++ * @next_rq: if fifo isn't expired, next request to serve. |
5831 |
++ * @queued: nr of requests queued in @sort_list. |
5832 |
++ * @allocated: currently allocated requests. |
5833 |
++ * @meta_pending: pending metadata requests. |
5834 |
++ * @fifo: fifo list of requests in sort_list. |
5835 |
++ * @entity: entity representing this queue in the scheduler. |
5836 |
++ * @max_budget: maximum budget allowed from the feedback mechanism. |
5837 |
++ * @budget_timeout: budget expiration (in jiffies). |
5838 |
++ * @dispatched: number of requests on the dispatch list or inside driver. |
5839 |
++ * @org_ioprio: saved ioprio during boosted periods. |
5840 |
++ * @flags: status flags. |
5841 |
++ * @bfqq_list: node for active/idle bfqq list inside our bfqd. |
5842 |
++ * @seek_samples: number of seeks sampled |
5843 |
++ * @seek_total: sum of the distances of the seeks sampled |
5844 |
++ * @seek_mean: mean seek distance |
5845 |
++ * @last_request_pos: position of the last request enqueued |
5846 |
++ * @pid: pid of the process owning the queue, used for logging purposes. |
5847 |
++ * @last_rais_start_time: last (idle -> weight-raised) transition attempt |
5848 |
++ * @raising_cur_max_time: current max raising time for this queue |
5849 |
++ * @last_idle_bklogged: time of the last transition of the @bfq_queue from |
5850 |
++ * idle to backlogged |
5851 |
++ * @service_from_backlogged: cumulative service received from the @bfq_queue |
5852 |
++ * since the last transition from idle to backlogged |
5853 |
++ * |
5854 |
++ * A bfq_queue is a leaf request queue; it can be associated to an io_context |
5855 |
++ * or more (if it is an async one). @cgroup holds a reference to the |
5856 |
++ * cgroup, to be sure that it does not disappear while a bfqq still |
5857 |
++ * references it (mostly to avoid races between request issuing and task |
5858 |
++ * migration followed by cgroup distruction). |
5859 |
++ * All the fields are protected by the queue lock of the containing bfqd. |
5860 |
++ */ |
5861 |
++struct bfq_queue { |
5862 |
++ atomic_t ref; |
5863 |
++ struct bfq_data *bfqd; |
5864 |
++ |
5865 |
++ /* fields for cooperating queues handling */ |
5866 |
++ struct bfq_queue *new_bfqq; |
5867 |
++ struct rb_node pos_node; |
5868 |
++ struct rb_root *pos_root; |
5869 |
++ |
5870 |
++ struct rb_root sort_list; |
5871 |
++ struct request *next_rq; |
5872 |
++ int queued[2]; |
5873 |
++ int allocated[2]; |
5874 |
++ int meta_pending; |
5875 |
++ struct list_head fifo; |
5876 |
++ |
5877 |
++ struct bfq_entity entity; |
5878 |
++ |
5879 |
++ unsigned long max_budget; |
5880 |
++ unsigned long budget_timeout; |
5881 |
++ |
5882 |
++ int dispatched; |
5883 |
++ |
5884 |
++ unsigned short org_ioprio; |
5885 |
++ |
5886 |
++ unsigned int flags; |
5887 |
++ |
5888 |
++ struct list_head bfqq_list; |
5889 |
++ |
5890 |
++ unsigned int seek_samples; |
5891 |
++ u64 seek_total; |
5892 |
++ sector_t seek_mean; |
5893 |
++ sector_t last_request_pos; |
5894 |
++ |
5895 |
++ pid_t pid; |
5896 |
++ |
5897 |
++ /* weight-raising fields */ |
5898 |
++ unsigned int raising_cur_max_time; |
5899 |
++ unsigned long soft_rt_next_start; |
5900 |
++ u64 last_rais_start_finish; |
5901 |
++ unsigned int raising_coeff; |
5902 |
++ u64 last_idle_bklogged; |
5903 |
++ unsigned long service_from_backlogged; |
5904 |
++}; |
5905 |
++ |
5906 |
++/** |
5907 |
++ * struct bfq_ttime - per process thinktime stats. |
5908 |
++ * @ttime_total: total process thinktime |
5909 |
++ * @ttime_samples: number of thinktime samples |
5910 |
++ * @ttime_mean: average process thinktime |
5911 |
++ */ |
5912 |
++struct bfq_ttime { |
5913 |
++ unsigned long last_end_request; |
5914 |
++ |
5915 |
++ unsigned long ttime_total; |
5916 |
++ unsigned long ttime_samples; |
5917 |
++ unsigned long ttime_mean; |
5918 |
++}; |
5919 |
++ |
5920 |
++/** |
5921 |
++ * struct bfq_io_cq - per (request_queue, io_context) structure. |
5922 |
++ * @icq: associated io_cq structure |
5923 |
++ * @bfqq: array of two process queues, the sync and the async |
5924 |
++ * @ttime: associated @bfq_ttime struct |
5925 |
++ */ |
5926 |
++struct bfq_io_cq { |
5927 |
++ struct io_cq icq; /* must be the first member */ |
5928 |
++ struct bfq_queue *bfqq[2]; |
5929 |
++ struct bfq_ttime ttime; |
5930 |
++ int ioprio; |
5931 |
++}; |
5932 |
++ |
5933 |
++/** |
5934 |
++ * struct bfq_data - per device data structure. |
5935 |
++ * @queue: request queue for the managed device. |
5936 |
++ * @root_group: root bfq_group for the device. |
5937 |
++ * @rq_pos_tree: rbtree sorted by next_request position, |
5938 |
++ * used when determining if two or more queues |
5939 |
++ * have interleaving requests (see bfq_close_cooperator). |
5940 |
++ * @busy_queues: number of bfq_queues containing requests (including the |
5941 |
++ * queue under service, even if it is idling). |
5942 |
++ * @raised_busy_queues: number of weight-raised busy bfq_queues. |
5943 |
++ * @queued: number of queued requests. |
5944 |
++ * @rq_in_driver: number of requests dispatched and waiting for completion. |
5945 |
++ * @sync_flight: number of sync requests in the driver. |
5946 |
++ * @max_rq_in_driver: max number of reqs in driver in the last @hw_tag_samples |
5947 |
++ * completed requests . |
5948 |
++ * @hw_tag_samples: nr of samples used to calculate hw_tag. |
5949 |
++ * @hw_tag: flag set to one if the driver is showing a queueing behavior. |
5950 |
++ * @budgets_assigned: number of budgets assigned. |
5951 |
++ * @idle_slice_timer: timer set when idling for the next sequential request |
5952 |
++ * from the queue under service. |
5953 |
++ * @unplug_work: delayed work to restart dispatching on the request queue. |
5954 |
++ * @in_service_queue: bfq_queue under service. |
5955 |
++ * @in_service_bic: bfq_io_cq (bic) associated with the @in_service_queue. |
5956 |
++ * @last_position: on-disk position of the last served request. |
5957 |
++ * @last_budget_start: beginning of the last budget. |
5958 |
++ * @last_idling_start: beginning of the last idle slice. |
5959 |
++ * @peak_rate: peak transfer rate observed for a budget. |
5960 |
++ * @peak_rate_samples: number of samples used to calculate @peak_rate. |
5961 |
++ * @bfq_max_budget: maximum budget allotted to a bfq_queue before rescheduling. |
5962 |
++ * @group_list: list of all the bfq_groups active on the device. |
5963 |
++ * @active_list: list of all the bfq_queues active on the device. |
5964 |
++ * @idle_list: list of all the bfq_queues idle on the device. |
5965 |
++ * @bfq_quantum: max number of requests dispatched per dispatch round. |
5966 |
++ * @bfq_fifo_expire: timeout for async/sync requests; when it expires |
5967 |
++ * requests are served in fifo order. |
5968 |
++ * @bfq_back_penalty: weight of backward seeks wrt forward ones. |
5969 |
++ * @bfq_back_max: maximum allowed backward seek. |
5970 |
++ * @bfq_slice_idle: maximum idling time. |
5971 |
++ * @bfq_user_max_budget: user-configured max budget value (0 for auto-tuning). |
5972 |
++ * @bfq_max_budget_async_rq: maximum budget (in nr of requests) allotted to |
5973 |
++ * async queues. |
5974 |
++ * @bfq_timeout: timeout for bfq_queues to consume their budget; used to |
5975 |
++ * to prevent seeky queues to impose long latencies to well |
5976 |
++ * behaved ones (this also implies that seeky queues cannot |
5977 |
++ * receive guarantees in the service domain; after a timeout |
5978 |
++ * they are charged for the whole allocated budget, to try |
5979 |
++ * to preserve a behavior reasonably fair among them, but |
5980 |
++ * without service-domain guarantees). |
5981 |
++ * @bfq_raising_coeff: Maximum factor by which the weight of a boosted |
5982 |
++ * queue is multiplied |
5983 |
++ * @bfq_raising_max_time: maximum duration of a weight-raising period (jiffies) |
5984 |
++ * @bfq_raising_rt_max_time: maximum duration for soft real-time processes |
5985 |
++ * @bfq_raising_min_idle_time: minimum idle period after which weight-raising |
5986 |
++ * may be reactivated for a queue (in jiffies) |
5987 |
++ * @bfq_raising_min_inter_arr_async: minimum period between request arrivals |
5988 |
++ * after which weight-raising may be |
5989 |
++ * reactivated for an already busy queue |
5990 |
++ * (in jiffies) |
5991 |
++ * @bfq_raising_max_softrt_rate: max service-rate for a soft real-time queue, |
5992 |
++ * sectors per seconds |
5993 |
++ * @RT_prod: cached value of the product R*T used for computing the maximum |
5994 |
++ * duration of the weight raising automatically |
5995 |
++ * @oom_bfqq: fallback dummy bfqq for extreme OOM conditions |
5996 |
++ * |
5997 |
++ * All the fields are protected by the @queue lock. |
5998 |
++ */ |
5999 |
++struct bfq_data { |
6000 |
++ struct request_queue *queue; |
6001 |
++ |
6002 |
++ struct bfq_group *root_group; |
6003 |
++ |
6004 |
++ struct rb_root rq_pos_tree; |
6005 |
++ |
6006 |
++ int busy_queues; |
6007 |
++ int raised_busy_queues; |
6008 |
++ int queued; |
6009 |
++ int rq_in_driver; |
6010 |
++ int sync_flight; |
6011 |
++ |
6012 |
++ int max_rq_in_driver; |
6013 |
++ int hw_tag_samples; |
6014 |
++ int hw_tag; |
6015 |
++ |
6016 |
++ int budgets_assigned; |
6017 |
++ |
6018 |
++ struct timer_list idle_slice_timer; |
6019 |
++ struct work_struct unplug_work; |
6020 |
++ |
6021 |
++ struct bfq_queue *in_service_queue; |
6022 |
++ struct bfq_io_cq *in_service_bic; |
6023 |
++ |
6024 |
++ sector_t last_position; |
6025 |
++ |
6026 |
++ ktime_t last_budget_start; |
6027 |
++ ktime_t last_idling_start; |
6028 |
++ int peak_rate_samples; |
6029 |
++ u64 peak_rate; |
6030 |
++ unsigned long bfq_max_budget; |
6031 |
++ |
6032 |
++ struct hlist_head group_list; |
6033 |
++ struct list_head active_list; |
6034 |
++ struct list_head idle_list; |
6035 |
++ |
6036 |
++ unsigned int bfq_quantum; |
6037 |
++ unsigned int bfq_fifo_expire[2]; |
6038 |
++ unsigned int bfq_back_penalty; |
6039 |
++ unsigned int bfq_back_max; |
6040 |
++ unsigned int bfq_slice_idle; |
6041 |
++ u64 bfq_class_idle_last_service; |
6042 |
++ |
6043 |
++ unsigned int bfq_user_max_budget; |
6044 |
++ unsigned int bfq_max_budget_async_rq; |
6045 |
++ unsigned int bfq_timeout[2]; |
6046 |
++ |
6047 |
++ bool low_latency; |
6048 |
++ |
6049 |
++ /* parameters of the low_latency heuristics */ |
6050 |
++ unsigned int bfq_raising_coeff; |
6051 |
++ unsigned int bfq_raising_max_time; |
6052 |
++ unsigned int bfq_raising_rt_max_time; |
6053 |
++ unsigned int bfq_raising_min_idle_time; |
6054 |
++ unsigned long bfq_raising_min_inter_arr_async; |
6055 |
++ unsigned int bfq_raising_max_softrt_rate; |
6056 |
++ u64 RT_prod; |
6057 |
++ |
6058 |
++ struct bfq_queue oom_bfqq; |
6059 |
++}; |
6060 |
++ |
6061 |
++enum bfqq_state_flags { |
6062 |
++ BFQ_BFQQ_FLAG_busy = 0, /* has requests or is under service */ |
6063 |
++ BFQ_BFQQ_FLAG_wait_request, /* waiting for a request */ |
6064 |
++ BFQ_BFQQ_FLAG_must_alloc, /* must be allowed rq alloc */ |
6065 |
++ BFQ_BFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
6066 |
++ BFQ_BFQQ_FLAG_idle_window, /* slice idling enabled */ |
6067 |
++ BFQ_BFQQ_FLAG_prio_changed, /* task priority has changed */ |
6068 |
++ BFQ_BFQQ_FLAG_sync, /* synchronous queue */ |
6069 |
++ BFQ_BFQQ_FLAG_budget_new, /* no completion with this budget */ |
6070 |
++ BFQ_BFQQ_FLAG_coop, /* bfqq is shared */ |
6071 |
++ BFQ_BFQQ_FLAG_split_coop, /* shared bfqq will be splitted */ |
6072 |
++ BFQ_BFQQ_FLAG_softrt_update, /* needs softrt-next-start update */ |
6073 |
++}; |
6074 |
++ |
6075 |
++#define BFQ_BFQQ_FNS(name) \ |
6076 |
++static inline void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ |
6077 |
++{ \ |
6078 |
++ (bfqq)->flags |= (1 << BFQ_BFQQ_FLAG_##name); \ |
6079 |
++} \ |
6080 |
++static inline void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ |
6081 |
++{ \ |
6082 |
++ (bfqq)->flags &= ~(1 << BFQ_BFQQ_FLAG_##name); \ |
6083 |
++} \ |
6084 |
++static inline int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ |
6085 |
++{ \ |
6086 |
++ return ((bfqq)->flags & (1 << BFQ_BFQQ_FLAG_##name)) != 0; \ |
6087 |
++} |
6088 |
++ |
6089 |
++BFQ_BFQQ_FNS(busy); |
6090 |
++BFQ_BFQQ_FNS(wait_request); |
6091 |
++BFQ_BFQQ_FNS(must_alloc); |
6092 |
++BFQ_BFQQ_FNS(fifo_expire); |
6093 |
++BFQ_BFQQ_FNS(idle_window); |
6094 |
++BFQ_BFQQ_FNS(prio_changed); |
6095 |
++BFQ_BFQQ_FNS(sync); |
6096 |
++BFQ_BFQQ_FNS(budget_new); |
6097 |
++BFQ_BFQQ_FNS(coop); |
6098 |
++BFQ_BFQQ_FNS(split_coop); |
6099 |
++BFQ_BFQQ_FNS(softrt_update); |
6100 |
++#undef BFQ_BFQQ_FNS |
6101 |
++ |
6102 |
++/* Logging facilities. */ |
6103 |
++#define bfq_log_bfqq(bfqd, bfqq, fmt, args...) \ |
6104 |
++ blk_add_trace_msg((bfqd)->queue, "bfq%d " fmt, (bfqq)->pid, ##args) |
6105 |
++ |
6106 |
++#define bfq_log(bfqd, fmt, args...) \ |
6107 |
++ blk_add_trace_msg((bfqd)->queue, "bfq " fmt, ##args) |
6108 |
++ |
6109 |
++/* Expiration reasons. */ |
6110 |
++enum bfqq_expiration { |
6111 |
++ BFQ_BFQQ_TOO_IDLE = 0, /* queue has been idling for too long */ |
6112 |
++ BFQ_BFQQ_BUDGET_TIMEOUT, /* budget took too long to be used */ |
6113 |
++ BFQ_BFQQ_BUDGET_EXHAUSTED, /* budget consumed */ |
6114 |
++ BFQ_BFQQ_NO_MORE_REQUESTS, /* the queue has no more requests */ |
6115 |
++}; |
6116 |
++ |
6117 |
++#ifdef CONFIG_CGROUP_BFQIO |
6118 |
++/** |
6119 |
++ * struct bfq_group - per (device, cgroup) data structure. |
6120 |
++ * @entity: schedulable entity to insert into the parent group sched_data. |
6121 |
++ * @sched_data: own sched_data, to contain child entities (they may be |
6122 |
++ * both bfq_queues and bfq_groups). |
6123 |
++ * @group_node: node to be inserted into the bfqio_cgroup->group_data |
6124 |
++ * list of the containing cgroup's bfqio_cgroup. |
6125 |
++ * @bfqd_node: node to be inserted into the @bfqd->group_list list |
6126 |
++ * of the groups active on the same device; used for cleanup. |
6127 |
++ * @bfqd: the bfq_data for the device this group acts upon. |
6128 |
++ * @async_bfqq: array of async queues for all the tasks belonging to |
6129 |
++ * the group, one queue per ioprio value per ioprio_class, |
6130 |
++ * except for the idle class that has only one queue. |
6131 |
++ * @async_idle_bfqq: async queue for the idle class (ioprio is ignored). |
6132 |
++ * @my_entity: pointer to @entity, %NULL for the toplevel group; used |
6133 |
++ * to avoid too many special cases during group creation/migration. |
6134 |
++ * |
6135 |
++ * Each (device, cgroup) pair has its own bfq_group, i.e., for each cgroup |
6136 |
++ * there is a set of bfq_groups, each one collecting the lower-level |
6137 |
++ * entities belonging to the group that are acting on the same device. |
6138 |
++ * |
6139 |
++ * Locking works as follows: |
6140 |
++ * o @group_node is protected by the bfqio_cgroup lock, and is accessed |
6141 |
++ * via RCU from its readers. |
6142 |
++ * o @bfqd is protected by the queue lock, RCU is used to access it |
6143 |
++ * from the readers. |
6144 |
++ * o All the other fields are protected by the @bfqd queue lock. |
6145 |
++ */ |
6146 |
++struct bfq_group { |
6147 |
++ struct bfq_entity entity; |
6148 |
++ struct bfq_sched_data sched_data; |
6149 |
++ |
6150 |
++ struct hlist_node group_node; |
6151 |
++ struct hlist_node bfqd_node; |
6152 |
++ |
6153 |
++ void *bfqd; |
6154 |
++ |
6155 |
++ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
6156 |
++ struct bfq_queue *async_idle_bfqq; |
6157 |
++ |
6158 |
++ struct bfq_entity *my_entity; |
6159 |
++}; |
6160 |
++ |
6161 |
++/** |
6162 |
++ * struct bfqio_cgroup - bfq cgroup data structure. |
6163 |
++ * @css: subsystem state for bfq in the containing cgroup. |
6164 |
++ * @online: flag marked when the subsystem is inserted. |
6165 |
++ * @weight: cgroup weight. |
6166 |
++ * @ioprio: cgroup ioprio. |
6167 |
++ * @ioprio_class: cgroup ioprio_class. |
6168 |
++ * @lock: spinlock that protects @ioprio, @ioprio_class and @group_data. |
6169 |
++ * @group_data: list containing the bfq_group belonging to this cgroup. |
6170 |
++ * |
6171 |
++ * @group_data is accessed using RCU, with @lock protecting the updates, |
6172 |
++ * @ioprio and @ioprio_class are protected by @lock. |
6173 |
++ */ |
6174 |
++struct bfqio_cgroup { |
6175 |
++ struct cgroup_subsys_state css; |
6176 |
++ bool online; |
6177 |
++ |
6178 |
++ unsigned short weight, ioprio, ioprio_class; |
6179 |
++ |
6180 |
++ spinlock_t lock; |
6181 |
++ struct hlist_head group_data; |
6182 |
++}; |
6183 |
++#else |
6184 |
++struct bfq_group { |
6185 |
++ struct bfq_sched_data sched_data; |
6186 |
++ |
6187 |
++ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
6188 |
++ struct bfq_queue *async_idle_bfqq; |
6189 |
++}; |
6190 |
++#endif |
6191 |
++ |
6192 |
++static inline struct bfq_service_tree * |
6193 |
++bfq_entity_service_tree(struct bfq_entity *entity) |
6194 |
++{ |
6195 |
++ struct bfq_sched_data *sched_data = entity->sched_data; |
6196 |
++ unsigned int idx = entity->ioprio_class - 1; |
6197 |
++ |
6198 |
++ BUG_ON(idx >= BFQ_IOPRIO_CLASSES); |
6199 |
++ BUG_ON(sched_data == NULL); |
6200 |
++ |
6201 |
++ return sched_data->service_tree + idx; |
6202 |
++} |
6203 |
++ |
6204 |
++static inline struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, |
6205 |
++ int is_sync) |
6206 |
++{ |
6207 |
++ return bic->bfqq[!!is_sync]; |
6208 |
++} |
6209 |
++ |
6210 |
++static inline void bic_set_bfqq(struct bfq_io_cq *bic, |
6211 |
++ struct bfq_queue *bfqq, int is_sync) |
6212 |
++{ |
6213 |
++ bic->bfqq[!!is_sync] = bfqq; |
6214 |
++} |
6215 |
++ |
6216 |
++static inline struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) |
6217 |
++{ |
6218 |
++ return bic->icq.q->elevator->elevator_data; |
6219 |
++} |
6220 |
++ |
6221 |
++/** |
6222 |
++ * bfq_get_bfqd_locked - get a lock to a bfqd using a RCU protected pointer. |
6223 |
++ * @ptr: a pointer to a bfqd. |
6224 |
++ * @flags: storage for the flags to be saved. |
6225 |
++ * |
6226 |
++ * This function allows bfqg->bfqd to be protected by the |
6227 |
++ * queue lock of the bfqd they reference; the pointer is dereferenced |
6228 |
++ * under RCU, so the storage for bfqd is assured to be safe as long |
6229 |
++ * as the RCU read side critical section does not end. After the |
6230 |
++ * bfqd->queue->queue_lock is taken the pointer is rechecked, to be |
6231 |
++ * sure that no other writer accessed it. If we raced with a writer, |
6232 |
++ * the function returns NULL, with the queue unlocked, otherwise it |
6233 |
++ * returns the dereferenced pointer, with the queue locked. |
6234 |
++ */ |
6235 |
++static inline struct bfq_data *bfq_get_bfqd_locked(void **ptr, |
6236 |
++ unsigned long *flags) |
6237 |
++{ |
6238 |
++ struct bfq_data *bfqd; |
6239 |
++ |
6240 |
++ rcu_read_lock(); |
6241 |
++ bfqd = rcu_dereference(*(struct bfq_data **)ptr); |
6242 |
++ |
6243 |
++ if (bfqd != NULL) { |
6244 |
++ spin_lock_irqsave(bfqd->queue->queue_lock, *flags); |
6245 |
++ if (*ptr == bfqd) |
6246 |
++ goto out; |
6247 |
++ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
6248 |
++ } |
6249 |
++ |
6250 |
++ bfqd = NULL; |
6251 |
++out: |
6252 |
++ rcu_read_unlock(); |
6253 |
++ return bfqd; |
6254 |
++} |
6255 |
++ |
6256 |
++static inline void bfq_put_bfqd_unlock(struct bfq_data *bfqd, |
6257 |
++ unsigned long *flags) |
6258 |
++{ |
6259 |
++ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
6260 |
++} |
6261 |
++ |
6262 |
++static void bfq_changed_ioprio(struct bfq_io_cq *bic); |
6263 |
++static void bfq_put_queue(struct bfq_queue *bfqq); |
6264 |
++static void bfq_dispatch_insert(struct request_queue *q, struct request *rq); |
6265 |
++static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
6266 |
++ struct bfq_group *bfqg, int is_sync, |
6267 |
++ struct bfq_io_cq *bic, gfp_t gfp_mask); |
6268 |
++static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
6269 |
++ struct bfq_group *bfqg); |
6270 |
++static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg); |
6271 |
++static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq); |
6272 |
++#endif |
6273 |
+-- |
6274 |
+1.8.5.2 |
6275 |
+ |
6276 |
|
6277 |
Deleted: genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-v6r2-I-O-sched-for-3.11.patch1 |
6278 |
=================================================================== |
6279 |
--- genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-v6r2-I-O-sched-for-3.11.patch1 2014-01-29 14:41:45 UTC (rev 2660) |
6280 |
+++ genpatches-2.6/trunk/3.14/5000_BFQ-2-block-introduce-the-v6r2-I-O-sched-for-3.11.patch1 2014-01-30 16:49:47 UTC (rev 2661) |
6281 |
@@ -1,5773 +0,0 @@ |
6282 |
-From 009b78bafe1763f71e6bdbb4f536b564a73b7db5 Mon Sep 17 00:00:00 2001 |
6283 |
-From: Arianna Avanzini <avanzini.arianna@×××××.com> |
6284 |
-Date: Thu, 9 May 2013 19:10:02 +0200 |
6285 |
-Subject: [PATCH 2/3] block: introduce the BFQ-v6r2 I/O sched for 3.11 |
6286 |
- |
6287 |
-Add the BFQ-v6r2 I/O scheduler to 3.11. |
6288 |
-The general structure is borrowed from CFQ, as much code. A (bfq_)queue |
6289 |
-is associated to each task doing I/O on a device, and each time a |
6290 |
-scheduling decision has to be made a queue is selected and served until |
6291 |
-it expires. |
6292 |
- |
6293 |
- - Slices are given in the service domain: tasks are assigned |
6294 |
- budgets, measured in number of sectors. Once got the disk, a task |
6295 |
- must however consume its assigned budget within a configurable |
6296 |
- maximum time (by default, the maximum possible value of the |
6297 |
- budgets is automatically computed to comply with this timeout). |
6298 |
- This allows the desired latency vs "throughput boosting" tradeoff |
6299 |
- to be set. |
6300 |
- |
6301 |
- - Budgets are scheduled according to a variant of WF2Q+, implemented |
6302 |
- using an augmented rb-tree to take eligibility into account while |
6303 |
- preserving an O(log N) overall complexity. |
6304 |
- |
6305 |
- - A low-latency tunable is provided; if enabled, both interactive |
6306 |
- and soft real-time applications are guaranteed very low latency. |
6307 |
- |
6308 |
- - Latency guarantees are preserved also in presence of NCQ. |
6309 |
- |
6310 |
- - Also with flash-based devices, a high throughput is achieved while |
6311 |
- still preserving latency guarantees. |
6312 |
- |
6313 |
- - Useful features borrowed from CFQ: cooperating-queues merging (with |
6314 |
- some additional optimizations with respect to the original CFQ version), |
6315 |
- static fallback queue for OOM. |
6316 |
- |
6317 |
- - BFQ supports full hierarchical scheduling, exporting a cgroups |
6318 |
- interface. Each node has a full scheduler, so each group can |
6319 |
- be assigned its own ioprio (mapped to a weight, see next point) |
6320 |
- and an ioprio_class. |
6321 |
- |
6322 |
- - If the cgroups interface is used, weights can be explictly |
6323 |
- assigned, otherwise ioprio values are mapped to weights using the |
6324 |
- relation weight = IOPRIO_BE_NR - ioprio. |
6325 |
- |
6326 |
- - ioprio classes are served in strict priority order, i.e., lower |
6327 |
- priority queues are not served as long as there are higher |
6328 |
- priority queues. Among queues in the same class the bandwidth is |
6329 |
- distributed in proportion to the weight of each queue. A very |
6330 |
- thin extra bandwidth is however guaranteed to the Idle class, to |
6331 |
- prevent it from starving. |
6332 |
- |
6333 |
-Signed-off-by: Paolo Valente <paolo.valente@×××××××.it> |
6334 |
-Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
6335 |
---- |
6336 |
- block/bfq-cgroup.c | 881 +++++++++++++++ |
6337 |
- block/bfq-ioc.c | 36 + |
6338 |
- block/bfq-iosched.c | 3082 +++++++++++++++++++++++++++++++++++++++++++++++++++ |
6339 |
- block/bfq-sched.c | 1072 ++++++++++++++++++ |
6340 |
- block/bfq.h | 603 ++++++++++ |
6341 |
- 5 files changed, 5674 insertions(+) |
6342 |
- create mode 100644 block/bfq-cgroup.c |
6343 |
- create mode 100644 block/bfq-ioc.c |
6344 |
- create mode 100644 block/bfq-iosched.c |
6345 |
- create mode 100644 block/bfq-sched.c |
6346 |
- create mode 100644 block/bfq.h |
6347 |
- |
6348 |
-diff --git a/block/bfq-cgroup.c b/block/bfq-cgroup.c |
6349 |
-new file mode 100644 |
6350 |
-index 0000000..bb9b851 |
6351 |
---- /dev/null |
6352 |
-+++ b/block/bfq-cgroup.c |
6353 |
-@@ -0,0 +1,881 @@ |
6354 |
-+/* |
6355 |
-+ * BFQ: CGROUPS support. |
6356 |
-+ * |
6357 |
-+ * Based on ideas and code from CFQ: |
6358 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
6359 |
-+ * |
6360 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
6361 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
6362 |
-+ * |
6363 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
6364 |
-+ * |
6365 |
-+ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
6366 |
-+ */ |
6367 |
-+ |
6368 |
-+#ifdef CONFIG_CGROUP_BFQIO |
6369 |
-+ |
6370 |
-+static DEFINE_MUTEX(bfqio_mutex); |
6371 |
-+ |
6372 |
-+static bool bfqio_is_removed(struct cgroup *cgroup) |
6373 |
-+{ |
6374 |
-+ return test_bit(CGRP_DEAD, &cgroup->flags); |
6375 |
-+} |
6376 |
-+ |
6377 |
-+static struct bfqio_cgroup bfqio_root_cgroup = { |
6378 |
-+ .weight = BFQ_DEFAULT_GRP_WEIGHT, |
6379 |
-+ .ioprio = BFQ_DEFAULT_GRP_IOPRIO, |
6380 |
-+ .ioprio_class = BFQ_DEFAULT_GRP_CLASS, |
6381 |
-+}; |
6382 |
-+ |
6383 |
-+static inline void bfq_init_entity(struct bfq_entity *entity, |
6384 |
-+ struct bfq_group *bfqg) |
6385 |
-+{ |
6386 |
-+ entity->weight = entity->new_weight; |
6387 |
-+ entity->orig_weight = entity->new_weight; |
6388 |
-+ entity->ioprio = entity->new_ioprio; |
6389 |
-+ entity->ioprio_class = entity->new_ioprio_class; |
6390 |
-+ entity->parent = bfqg->my_entity; |
6391 |
-+ entity->sched_data = &bfqg->sched_data; |
6392 |
-+} |
6393 |
-+ |
6394 |
-+static struct bfqio_cgroup *cgroup_to_bfqio(struct cgroup *cgroup) |
6395 |
-+{ |
6396 |
-+ return container_of(cgroup_subsys_state(cgroup, bfqio_subsys_id), |
6397 |
-+ struct bfqio_cgroup, css); |
6398 |
-+} |
6399 |
-+ |
6400 |
-+/* |
6401 |
-+ * Search the bfq_group for bfqd into the hash table (by now only a list) |
6402 |
-+ * of bgrp. Must be called under rcu_read_lock(). |
6403 |
-+ */ |
6404 |
-+static struct bfq_group *bfqio_lookup_group(struct bfqio_cgroup *bgrp, |
6405 |
-+ struct bfq_data *bfqd) |
6406 |
-+{ |
6407 |
-+ struct bfq_group *bfqg; |
6408 |
-+ void *key; |
6409 |
-+ |
6410 |
-+ hlist_for_each_entry_rcu(bfqg, &bgrp->group_data, group_node) { |
6411 |
-+ key = rcu_dereference(bfqg->bfqd); |
6412 |
-+ if (key == bfqd) |
6413 |
-+ return bfqg; |
6414 |
-+ } |
6415 |
-+ |
6416 |
-+ return NULL; |
6417 |
-+} |
6418 |
-+ |
6419 |
-+static inline void bfq_group_init_entity(struct bfqio_cgroup *bgrp, |
6420 |
-+ struct bfq_group *bfqg) |
6421 |
-+{ |
6422 |
-+ struct bfq_entity *entity = &bfqg->entity; |
6423 |
-+ |
6424 |
-+ /* |
6425 |
-+ * If the weight of the entity has never been set via the sysfs |
6426 |
-+ * interface, then bgrp->weight == 0. In this case we initialize |
6427 |
-+ * the weight from the current ioprio value. Otherwise, the group |
6428 |
-+ * weight, if set, has priority over the ioprio value. |
6429 |
-+ */ |
6430 |
-+ if (bgrp->weight == 0) { |
6431 |
-+ entity->new_weight = bfq_ioprio_to_weight(bgrp->ioprio); |
6432 |
-+ entity->new_ioprio = bgrp->ioprio; |
6433 |
-+ } else { |
6434 |
-+ entity->new_weight = bgrp->weight; |
6435 |
-+ entity->new_ioprio = bfq_weight_to_ioprio(bgrp->weight); |
6436 |
-+ } |
6437 |
-+ entity->orig_weight = entity->weight = entity->new_weight; |
6438 |
-+ entity->ioprio = entity->new_ioprio; |
6439 |
-+ entity->ioprio_class = entity->new_ioprio_class = bgrp->ioprio_class; |
6440 |
-+ entity->my_sched_data = &bfqg->sched_data; |
6441 |
-+} |
6442 |
-+ |
6443 |
-+static inline void bfq_group_set_parent(struct bfq_group *bfqg, |
6444 |
-+ struct bfq_group *parent) |
6445 |
-+{ |
6446 |
-+ struct bfq_entity *entity; |
6447 |
-+ |
6448 |
-+ BUG_ON(parent == NULL); |
6449 |
-+ BUG_ON(bfqg == NULL); |
6450 |
-+ |
6451 |
-+ entity = &bfqg->entity; |
6452 |
-+ entity->parent = parent->my_entity; |
6453 |
-+ entity->sched_data = &parent->sched_data; |
6454 |
-+} |
6455 |
-+ |
6456 |
-+/** |
6457 |
-+ * bfq_group_chain_alloc - allocate a chain of groups. |
6458 |
-+ * @bfqd: queue descriptor. |
6459 |
-+ * @cgroup: the leaf cgroup this chain starts from. |
6460 |
-+ * |
6461 |
-+ * Allocate a chain of groups starting from the one belonging to |
6462 |
-+ * @cgroup up to the root cgroup. Stop if a cgroup on the chain |
6463 |
-+ * to the root has already an allocated group on @bfqd. |
6464 |
-+ */ |
6465 |
-+static struct bfq_group *bfq_group_chain_alloc(struct bfq_data *bfqd, |
6466 |
-+ struct cgroup *cgroup) |
6467 |
-+{ |
6468 |
-+ struct bfqio_cgroup *bgrp; |
6469 |
-+ struct bfq_group *bfqg, *prev = NULL, *leaf = NULL; |
6470 |
-+ |
6471 |
-+ for (; cgroup != NULL; cgroup = cgroup->parent) { |
6472 |
-+ bgrp = cgroup_to_bfqio(cgroup); |
6473 |
-+ |
6474 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6475 |
-+ if (bfqg != NULL) { |
6476 |
-+ /* |
6477 |
-+ * All the cgroups in the path from there to the |
6478 |
-+ * root must have a bfq_group for bfqd, so we don't |
6479 |
-+ * need any more allocations. |
6480 |
-+ */ |
6481 |
-+ break; |
6482 |
-+ } |
6483 |
-+ |
6484 |
-+ bfqg = kzalloc(sizeof(*bfqg), GFP_ATOMIC); |
6485 |
-+ if (bfqg == NULL) |
6486 |
-+ goto cleanup; |
6487 |
-+ |
6488 |
-+ bfq_group_init_entity(bgrp, bfqg); |
6489 |
-+ bfqg->my_entity = &bfqg->entity; |
6490 |
-+ |
6491 |
-+ if (leaf == NULL) { |
6492 |
-+ leaf = bfqg; |
6493 |
-+ prev = leaf; |
6494 |
-+ } else { |
6495 |
-+ bfq_group_set_parent(prev, bfqg); |
6496 |
-+ /* |
6497 |
-+ * Build a list of allocated nodes using the bfqd |
6498 |
-+ * filed, that is still unused and will be initialized |
6499 |
-+ * only after the node will be connected. |
6500 |
-+ */ |
6501 |
-+ prev->bfqd = bfqg; |
6502 |
-+ prev = bfqg; |
6503 |
-+ } |
6504 |
-+ } |
6505 |
-+ |
6506 |
-+ return leaf; |
6507 |
-+ |
6508 |
-+cleanup: |
6509 |
-+ while (leaf != NULL) { |
6510 |
-+ prev = leaf; |
6511 |
-+ leaf = leaf->bfqd; |
6512 |
-+ kfree(prev); |
6513 |
-+ } |
6514 |
-+ |
6515 |
-+ return NULL; |
6516 |
-+} |
6517 |
-+ |
6518 |
-+/** |
6519 |
-+ * bfq_group_chain_link - link an allocatd group chain to a cgroup hierarchy. |
6520 |
-+ * @bfqd: the queue descriptor. |
6521 |
-+ * @cgroup: the leaf cgroup to start from. |
6522 |
-+ * @leaf: the leaf group (to be associated to @cgroup). |
6523 |
-+ * |
6524 |
-+ * Try to link a chain of groups to a cgroup hierarchy, connecting the |
6525 |
-+ * nodes bottom-up, so we can be sure that when we find a cgroup in the |
6526 |
-+ * hierarchy that already as a group associated to @bfqd all the nodes |
6527 |
-+ * in the path to the root cgroup have one too. |
6528 |
-+ * |
6529 |
-+ * On locking: the queue lock protects the hierarchy (there is a hierarchy |
6530 |
-+ * per device) while the bfqio_cgroup lock protects the list of groups |
6531 |
-+ * belonging to the same cgroup. |
6532 |
-+ */ |
6533 |
-+static void bfq_group_chain_link(struct bfq_data *bfqd, struct cgroup *cgroup, |
6534 |
-+ struct bfq_group *leaf) |
6535 |
-+{ |
6536 |
-+ struct bfqio_cgroup *bgrp; |
6537 |
-+ struct bfq_group *bfqg, *next, *prev = NULL; |
6538 |
-+ unsigned long flags; |
6539 |
-+ |
6540 |
-+ assert_spin_locked(bfqd->queue->queue_lock); |
6541 |
-+ |
6542 |
-+ for (; cgroup != NULL && leaf != NULL; cgroup = cgroup->parent) { |
6543 |
-+ bgrp = cgroup_to_bfqio(cgroup); |
6544 |
-+ next = leaf->bfqd; |
6545 |
-+ |
6546 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6547 |
-+ BUG_ON(bfqg != NULL); |
6548 |
-+ |
6549 |
-+ spin_lock_irqsave(&bgrp->lock, flags); |
6550 |
-+ |
6551 |
-+ rcu_assign_pointer(leaf->bfqd, bfqd); |
6552 |
-+ hlist_add_head_rcu(&leaf->group_node, &bgrp->group_data); |
6553 |
-+ hlist_add_head(&leaf->bfqd_node, &bfqd->group_list); |
6554 |
-+ |
6555 |
-+ spin_unlock_irqrestore(&bgrp->lock, flags); |
6556 |
-+ |
6557 |
-+ prev = leaf; |
6558 |
-+ leaf = next; |
6559 |
-+ } |
6560 |
-+ |
6561 |
-+ BUG_ON(cgroup == NULL && leaf != NULL); |
6562 |
-+ if (cgroup != NULL && prev != NULL) { |
6563 |
-+ bgrp = cgroup_to_bfqio(cgroup); |
6564 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6565 |
-+ bfq_group_set_parent(prev, bfqg); |
6566 |
-+ } |
6567 |
-+} |
6568 |
-+ |
6569 |
-+/** |
6570 |
-+ * bfq_find_alloc_group - return the group associated to @bfqd in @cgroup. |
6571 |
-+ * @bfqd: queue descriptor. |
6572 |
-+ * @cgroup: cgroup being searched for. |
6573 |
-+ * |
6574 |
-+ * Return a group associated to @bfqd in @cgroup, allocating one if |
6575 |
-+ * necessary. When a group is returned all the cgroups in the path |
6576 |
-+ * to the root have a group associated to @bfqd. |
6577 |
-+ * |
6578 |
-+ * If the allocation fails, return the root group: this breaks guarantees |
6579 |
-+ * but is a safe fallbak. If this loss becames a problem it can be |
6580 |
-+ * mitigated using the equivalent weight (given by the product of the |
6581 |
-+ * weights of the groups in the path from @group to the root) in the |
6582 |
-+ * root scheduler. |
6583 |
-+ * |
6584 |
-+ * We allocate all the missing nodes in the path from the leaf cgroup |
6585 |
-+ * to the root and we connect the nodes only after all the allocations |
6586 |
-+ * have been successful. |
6587 |
-+ */ |
6588 |
-+static struct bfq_group *bfq_find_alloc_group(struct bfq_data *bfqd, |
6589 |
-+ struct cgroup *cgroup) |
6590 |
-+{ |
6591 |
-+ struct bfqio_cgroup *bgrp = cgroup_to_bfqio(cgroup); |
6592 |
-+ struct bfq_group *bfqg; |
6593 |
-+ |
6594 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6595 |
-+ if (bfqg != NULL) |
6596 |
-+ return bfqg; |
6597 |
-+ |
6598 |
-+ bfqg = bfq_group_chain_alloc(bfqd, cgroup); |
6599 |
-+ if (bfqg != NULL) |
6600 |
-+ bfq_group_chain_link(bfqd, cgroup, bfqg); |
6601 |
-+ else |
6602 |
-+ bfqg = bfqd->root_group; |
6603 |
-+ |
6604 |
-+ return bfqg; |
6605 |
-+} |
6606 |
-+ |
6607 |
-+/** |
6608 |
-+ * bfq_bfqq_move - migrate @bfqq to @bfqg. |
6609 |
-+ * @bfqd: queue descriptor. |
6610 |
-+ * @bfqq: the queue to move. |
6611 |
-+ * @entity: @bfqq's entity. |
6612 |
-+ * @bfqg: the group to move to. |
6613 |
-+ * |
6614 |
-+ * Move @bfqq to @bfqg, deactivating it from its old group and reactivating |
6615 |
-+ * it on the new one. Avoid putting the entity on the old group idle tree. |
6616 |
-+ * |
6617 |
-+ * Must be called under the queue lock; the cgroup owning @bfqg must |
6618 |
-+ * not disappear (by now this just means that we are called under |
6619 |
-+ * rcu_read_lock()). |
6620 |
-+ */ |
6621 |
-+static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
6622 |
-+ struct bfq_entity *entity, struct bfq_group *bfqg) |
6623 |
-+{ |
6624 |
-+ int busy, resume; |
6625 |
-+ |
6626 |
-+ busy = bfq_bfqq_busy(bfqq); |
6627 |
-+ resume = !RB_EMPTY_ROOT(&bfqq->sort_list); |
6628 |
-+ |
6629 |
-+ BUG_ON(resume && !entity->on_st); |
6630 |
-+ BUG_ON(busy && !resume && entity->on_st && bfqq != bfqd->active_queue); |
6631 |
-+ |
6632 |
-+ if (busy) { |
6633 |
-+ BUG_ON(atomic_read(&bfqq->ref) < 2); |
6634 |
-+ |
6635 |
-+ if (!resume) |
6636 |
-+ bfq_del_bfqq_busy(bfqd, bfqq, 0); |
6637 |
-+ else |
6638 |
-+ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
6639 |
-+ } else if (entity->on_st) |
6640 |
-+ bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); |
6641 |
-+ |
6642 |
-+ /* |
6643 |
-+ * Here we use a reference to bfqg. We don't need a refcounter |
6644 |
-+ * as the cgroup reference will not be dropped, so that its |
6645 |
-+ * destroy() callback will not be invoked. |
6646 |
-+ */ |
6647 |
-+ entity->parent = bfqg->my_entity; |
6648 |
-+ entity->sched_data = &bfqg->sched_data; |
6649 |
-+ |
6650 |
-+ if (busy && resume) |
6651 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
6652 |
-+ |
6653 |
-+ if (bfqd->active_queue == NULL && !bfqd->rq_in_driver) |
6654 |
-+ bfq_schedule_dispatch(bfqd); |
6655 |
-+} |
6656 |
-+ |
6657 |
-+/** |
6658 |
-+ * __bfq_bic_change_cgroup - move @bic to @cgroup. |
6659 |
-+ * @bfqd: the queue descriptor. |
6660 |
-+ * @bic: the bic to move. |
6661 |
-+ * @cgroup: the cgroup to move to. |
6662 |
-+ * |
6663 |
-+ * Move bic to cgroup, assuming that bfqd->queue is locked; the caller |
6664 |
-+ * has to make sure that the reference to cgroup is valid across the call. |
6665 |
-+ * |
6666 |
-+ * NOTE: an alternative approach might have been to store the current |
6667 |
-+ * cgroup in bfqq and getting a reference to it, reducing the lookup |
6668 |
-+ * time here, at the price of slightly more complex code. |
6669 |
-+ */ |
6670 |
-+static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
6671 |
-+ struct bfq_io_cq *bic, |
6672 |
-+ struct cgroup *cgroup) |
6673 |
-+{ |
6674 |
-+ struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0); |
6675 |
-+ struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1); |
6676 |
-+ struct bfq_entity *entity; |
6677 |
-+ struct bfq_group *bfqg; |
6678 |
-+ struct bfqio_cgroup *bgrp; |
6679 |
-+ |
6680 |
-+ bgrp = cgroup_to_bfqio(cgroup); |
6681 |
-+ |
6682 |
-+ bfqg = bfq_find_alloc_group(bfqd, cgroup); |
6683 |
-+ if (async_bfqq != NULL) { |
6684 |
-+ entity = &async_bfqq->entity; |
6685 |
-+ |
6686 |
-+ if (entity->sched_data != &bfqg->sched_data) { |
6687 |
-+ bic_set_bfqq(bic, NULL, 0); |
6688 |
-+ bfq_log_bfqq(bfqd, async_bfqq, |
6689 |
-+ "bic_change_group: %p %d", |
6690 |
-+ async_bfqq, atomic_read(&async_bfqq->ref)); |
6691 |
-+ bfq_put_queue(async_bfqq); |
6692 |
-+ } |
6693 |
-+ } |
6694 |
-+ |
6695 |
-+ if (sync_bfqq != NULL) { |
6696 |
-+ entity = &sync_bfqq->entity; |
6697 |
-+ if (entity->sched_data != &bfqg->sched_data) |
6698 |
-+ bfq_bfqq_move(bfqd, sync_bfqq, entity, bfqg); |
6699 |
-+ } |
6700 |
-+ |
6701 |
-+ return bfqg; |
6702 |
-+} |
6703 |
-+ |
6704 |
-+/** |
6705 |
-+ * bfq_bic_change_cgroup - move @bic to @cgroup. |
6706 |
-+ * @bic: the bic being migrated. |
6707 |
-+ * @cgroup: the destination cgroup. |
6708 |
-+ * |
6709 |
-+ * When the task owning @bic is moved to @cgroup, @bic is immediately |
6710 |
-+ * moved into its new parent group. |
6711 |
-+ */ |
6712 |
-+static void bfq_bic_change_cgroup(struct bfq_io_cq *bic, |
6713 |
-+ struct cgroup *cgroup) |
6714 |
-+{ |
6715 |
-+ struct bfq_data *bfqd; |
6716 |
-+ unsigned long uninitialized_var(flags); |
6717 |
-+ |
6718 |
-+ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), &flags); |
6719 |
-+ if (bfqd != NULL) { |
6720 |
-+ __bfq_bic_change_cgroup(bfqd, bic, cgroup); |
6721 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
6722 |
-+ } |
6723 |
-+} |
6724 |
-+ |
6725 |
-+/** |
6726 |
-+ * bfq_bic_update_cgroup - update the cgroup of @bic. |
6727 |
-+ * @bic: the @bic to update. |
6728 |
-+ * |
6729 |
-+ * Make sure that @bic is enqueued in the cgroup of the current task. |
6730 |
-+ * We need this in addition to moving bics during the cgroup attach |
6731 |
-+ * phase because the task owning @bic could be at its first disk |
6732 |
-+ * access or we may end up in the root cgroup as the result of a |
6733 |
-+ * memory allocation failure and here we try to move to the right |
6734 |
-+ * group. |
6735 |
-+ * |
6736 |
-+ * Must be called under the queue lock. It is safe to use the returned |
6737 |
-+ * value even after the rcu_read_unlock() as the migration/destruction |
6738 |
-+ * paths act under the queue lock too. IOW it is impossible to race with |
6739 |
-+ * group migration/destruction and end up with an invalid group as: |
6740 |
-+ * a) here cgroup has not yet been destroyed, nor its destroy callback |
6741 |
-+ * has started execution, as current holds a reference to it, |
6742 |
-+ * b) if it is destroyed after rcu_read_unlock() [after current is |
6743 |
-+ * migrated to a different cgroup] its attach() callback will have |
6744 |
-+ * taken care of remove all the references to the old cgroup data. |
6745 |
-+ */ |
6746 |
-+static struct bfq_group *bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
6747 |
-+{ |
6748 |
-+ struct bfq_data *bfqd = bic_to_bfqd(bic); |
6749 |
-+ struct bfq_group *bfqg; |
6750 |
-+ struct cgroup *cgroup; |
6751 |
-+ |
6752 |
-+ BUG_ON(bfqd == NULL); |
6753 |
-+ |
6754 |
-+ rcu_read_lock(); |
6755 |
-+ cgroup = task_cgroup(current, bfqio_subsys_id); |
6756 |
-+ bfqg = __bfq_bic_change_cgroup(bfqd, bic, cgroup); |
6757 |
-+ rcu_read_unlock(); |
6758 |
-+ |
6759 |
-+ return bfqg; |
6760 |
-+} |
6761 |
-+ |
6762 |
-+/** |
6763 |
-+ * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st. |
6764 |
-+ * @st: the service tree being flushed. |
6765 |
-+ */ |
6766 |
-+static inline void bfq_flush_idle_tree(struct bfq_service_tree *st) |
6767 |
-+{ |
6768 |
-+ struct bfq_entity *entity = st->first_idle; |
6769 |
-+ |
6770 |
-+ for (; entity != NULL; entity = st->first_idle) |
6771 |
-+ __bfq_deactivate_entity(entity, 0); |
6772 |
-+} |
6773 |
-+ |
6774 |
-+/** |
6775 |
-+ * bfq_reparent_leaf_entity - move leaf entity to the root_group. |
6776 |
-+ * @bfqd: the device data structure with the root group. |
6777 |
-+ * @entity: the entity to move. |
6778 |
-+ */ |
6779 |
-+static inline void bfq_reparent_leaf_entity(struct bfq_data *bfqd, |
6780 |
-+ struct bfq_entity *entity) |
6781 |
-+{ |
6782 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
6783 |
-+ |
6784 |
-+ BUG_ON(bfqq == NULL); |
6785 |
-+ bfq_bfqq_move(bfqd, bfqq, entity, bfqd->root_group); |
6786 |
-+ return; |
6787 |
-+} |
6788 |
-+ |
6789 |
-+/** |
6790 |
-+ * bfq_reparent_active_entities - move to the root group all active entities. |
6791 |
-+ * @bfqd: the device data structure with the root group. |
6792 |
-+ * @bfqg: the group to move from. |
6793 |
-+ * @st: the service tree with the entities. |
6794 |
-+ * |
6795 |
-+ * Needs queue_lock to be taken and reference to be valid over the call. |
6796 |
-+ */ |
6797 |
-+static inline void bfq_reparent_active_entities(struct bfq_data *bfqd, |
6798 |
-+ struct bfq_group *bfqg, |
6799 |
-+ struct bfq_service_tree *st) |
6800 |
-+{ |
6801 |
-+ struct rb_root *active = &st->active; |
6802 |
-+ struct bfq_entity *entity = NULL; |
6803 |
-+ |
6804 |
-+ if (!RB_EMPTY_ROOT(&st->active)) |
6805 |
-+ entity = bfq_entity_of(rb_first(active)); |
6806 |
-+ |
6807 |
-+ for (; entity != NULL ; entity = bfq_entity_of(rb_first(active))) |
6808 |
-+ bfq_reparent_leaf_entity(bfqd, entity); |
6809 |
-+ |
6810 |
-+ if (bfqg->sched_data.active_entity != NULL) |
6811 |
-+ bfq_reparent_leaf_entity(bfqd, bfqg->sched_data.active_entity); |
6812 |
-+ |
6813 |
-+ return; |
6814 |
-+} |
6815 |
-+ |
6816 |
-+/** |
6817 |
-+ * bfq_destroy_group - destroy @bfqg. |
6818 |
-+ * @bgrp: the bfqio_cgroup containing @bfqg. |
6819 |
-+ * @bfqg: the group being destroyed. |
6820 |
-+ * |
6821 |
-+ * Destroy @bfqg, making sure that it is not referenced from its parent. |
6822 |
-+ */ |
6823 |
-+static void bfq_destroy_group(struct bfqio_cgroup *bgrp, struct bfq_group *bfqg) |
6824 |
-+{ |
6825 |
-+ struct bfq_data *bfqd; |
6826 |
-+ struct bfq_service_tree *st; |
6827 |
-+ struct bfq_entity *entity = bfqg->my_entity; |
6828 |
-+ unsigned long uninitialized_var(flags); |
6829 |
-+ int i; |
6830 |
-+ |
6831 |
-+ hlist_del(&bfqg->group_node); |
6832 |
-+ |
6833 |
-+ /* |
6834 |
-+ * Empty all service_trees belonging to this group before deactivating |
6835 |
-+ * the group itself. |
6836 |
-+ */ |
6837 |
-+ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) { |
6838 |
-+ st = bfqg->sched_data.service_tree + i; |
6839 |
-+ |
6840 |
-+ /* |
6841 |
-+ * The idle tree may still contain bfq_queues belonging |
6842 |
-+ * to exited task because they never migrated to a different |
6843 |
-+ * cgroup from the one being destroyed now. Noone else |
6844 |
-+ * can access them so it's safe to act without any lock. |
6845 |
-+ */ |
6846 |
-+ bfq_flush_idle_tree(st); |
6847 |
-+ |
6848 |
-+ /* |
6849 |
-+ * It may happen that some queues are still active |
6850 |
-+ * (busy) upon group destruction (if the corresponding |
6851 |
-+ * processes have been forced to terminate). We move |
6852 |
-+ * all the leaf entities corresponding to these queues |
6853 |
-+ * to the root_group. |
6854 |
-+ * Also, it may happen that the group has an entity |
6855 |
-+ * under service, which is disconnected from the active |
6856 |
-+ * tree: it must be moved, too. |
6857 |
-+ * There is no need to put the sync queues, as the |
6858 |
-+ * scheduler has taken no reference. |
6859 |
-+ */ |
6860 |
-+ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
6861 |
-+ if (bfqd != NULL) { |
6862 |
-+ bfq_reparent_active_entities(bfqd, bfqg, st); |
6863 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
6864 |
-+ } |
6865 |
-+ BUG_ON(!RB_EMPTY_ROOT(&st->active)); |
6866 |
-+ BUG_ON(!RB_EMPTY_ROOT(&st->idle)); |
6867 |
-+ } |
6868 |
-+ BUG_ON(bfqg->sched_data.next_active != NULL); |
6869 |
-+ BUG_ON(bfqg->sched_data.active_entity != NULL); |
6870 |
-+ |
6871 |
-+ /* |
6872 |
-+ * We may race with device destruction, take extra care when |
6873 |
-+ * dereferencing bfqg->bfqd. |
6874 |
-+ */ |
6875 |
-+ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
6876 |
-+ if (bfqd != NULL) { |
6877 |
-+ hlist_del(&bfqg->bfqd_node); |
6878 |
-+ __bfq_deactivate_entity(entity, 0); |
6879 |
-+ bfq_put_async_queues(bfqd, bfqg); |
6880 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
6881 |
-+ } |
6882 |
-+ BUG_ON(entity->tree != NULL); |
6883 |
-+ |
6884 |
-+ /* |
6885 |
-+ * No need to defer the kfree() to the end of the RCU grace |
6886 |
-+ * period: we are called from the destroy() callback of our |
6887 |
-+ * cgroup, so we can be sure that noone is a) still using |
6888 |
-+ * this cgroup or b) doing lookups in it. |
6889 |
-+ */ |
6890 |
-+ kfree(bfqg); |
6891 |
-+} |
6892 |
-+ |
6893 |
-+static void bfq_end_raising_async(struct bfq_data *bfqd) |
6894 |
-+{ |
6895 |
-+ struct hlist_node *tmp; |
6896 |
-+ struct bfq_group *bfqg; |
6897 |
-+ |
6898 |
-+ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) |
6899 |
-+ bfq_end_raising_async_queues(bfqd, bfqg); |
6900 |
-+} |
6901 |
-+ |
6902 |
-+/** |
6903 |
-+ * bfq_disconnect_groups - diconnect @bfqd from all its groups. |
6904 |
-+ * @bfqd: the device descriptor being exited. |
6905 |
-+ * |
6906 |
-+ * When the device exits we just make sure that no lookup can return |
6907 |
-+ * the now unused group structures. They will be deallocated on cgroup |
6908 |
-+ * destruction. |
6909 |
-+ */ |
6910 |
-+static void bfq_disconnect_groups(struct bfq_data *bfqd) |
6911 |
-+{ |
6912 |
-+ struct hlist_node *tmp; |
6913 |
-+ struct bfq_group *bfqg; |
6914 |
-+ |
6915 |
-+ bfq_log(bfqd, "disconnect_groups beginning") ; |
6916 |
-+ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) { |
6917 |
-+ hlist_del(&bfqg->bfqd_node); |
6918 |
-+ |
6919 |
-+ __bfq_deactivate_entity(bfqg->my_entity, 0); |
6920 |
-+ |
6921 |
-+ /* |
6922 |
-+ * Don't remove from the group hash, just set an |
6923 |
-+ * invalid key. No lookups can race with the |
6924 |
-+ * assignment as bfqd is being destroyed; this |
6925 |
-+ * implies also that new elements cannot be added |
6926 |
-+ * to the list. |
6927 |
-+ */ |
6928 |
-+ rcu_assign_pointer(bfqg->bfqd, NULL); |
6929 |
-+ |
6930 |
-+ bfq_log(bfqd, "disconnect_groups: put async for group %p", |
6931 |
-+ bfqg) ; |
6932 |
-+ bfq_put_async_queues(bfqd, bfqg); |
6933 |
-+ } |
6934 |
-+} |
6935 |
-+ |
6936 |
-+static inline void bfq_free_root_group(struct bfq_data *bfqd) |
6937 |
-+{ |
6938 |
-+ struct bfqio_cgroup *bgrp = &bfqio_root_cgroup; |
6939 |
-+ struct bfq_group *bfqg = bfqd->root_group; |
6940 |
-+ |
6941 |
-+ bfq_put_async_queues(bfqd, bfqg); |
6942 |
-+ |
6943 |
-+ spin_lock_irq(&bgrp->lock); |
6944 |
-+ hlist_del_rcu(&bfqg->group_node); |
6945 |
-+ spin_unlock_irq(&bgrp->lock); |
6946 |
-+ |
6947 |
-+ /* |
6948 |
-+ * No need to synchronize_rcu() here: since the device is gone |
6949 |
-+ * there cannot be any read-side access to its root_group. |
6950 |
-+ */ |
6951 |
-+ kfree(bfqg); |
6952 |
-+} |
6953 |
-+ |
6954 |
-+static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
6955 |
-+{ |
6956 |
-+ struct bfq_group *bfqg; |
6957 |
-+ struct bfqio_cgroup *bgrp; |
6958 |
-+ int i; |
6959 |
-+ |
6960 |
-+ bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); |
6961 |
-+ if (bfqg == NULL) |
6962 |
-+ return NULL; |
6963 |
-+ |
6964 |
-+ bfqg->entity.parent = NULL; |
6965 |
-+ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
6966 |
-+ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
6967 |
-+ |
6968 |
-+ bgrp = &bfqio_root_cgroup; |
6969 |
-+ spin_lock_irq(&bgrp->lock); |
6970 |
-+ rcu_assign_pointer(bfqg->bfqd, bfqd); |
6971 |
-+ hlist_add_head_rcu(&bfqg->group_node, &bgrp->group_data); |
6972 |
-+ spin_unlock_irq(&bgrp->lock); |
6973 |
-+ |
6974 |
-+ return bfqg; |
6975 |
-+} |
6976 |
-+ |
6977 |
-+#define SHOW_FUNCTION(__VAR) \ |
6978 |
-+static u64 bfqio_cgroup_##__VAR##_read(struct cgroup *cgroup, \ |
6979 |
-+ struct cftype *cftype) \ |
6980 |
-+{ \ |
6981 |
-+ struct bfqio_cgroup *bgrp; \ |
6982 |
-+ u64 ret = -ENODEV; \ |
6983 |
-+ \ |
6984 |
-+ mutex_lock(&bfqio_mutex); \ |
6985 |
-+ if (bfqio_is_removed(cgroup)) \ |
6986 |
-+ goto out_unlock; \ |
6987 |
-+ \ |
6988 |
-+ bgrp = cgroup_to_bfqio(cgroup); \ |
6989 |
-+ spin_lock_irq(&bgrp->lock); \ |
6990 |
-+ ret = bgrp->__VAR; \ |
6991 |
-+ spin_unlock_irq(&bgrp->lock); \ |
6992 |
-+ \ |
6993 |
-+out_unlock: \ |
6994 |
-+ mutex_unlock(&bfqio_mutex); \ |
6995 |
-+ return ret; \ |
6996 |
-+} |
6997 |
-+ |
6998 |
-+SHOW_FUNCTION(weight); |
6999 |
-+SHOW_FUNCTION(ioprio); |
7000 |
-+SHOW_FUNCTION(ioprio_class); |
7001 |
-+#undef SHOW_FUNCTION |
7002 |
-+ |
7003 |
-+#define STORE_FUNCTION(__VAR, __MIN, __MAX) \ |
7004 |
-+static int bfqio_cgroup_##__VAR##_write(struct cgroup *cgroup, \ |
7005 |
-+ struct cftype *cftype, \ |
7006 |
-+ u64 val) \ |
7007 |
-+{ \ |
7008 |
-+ struct bfqio_cgroup *bgrp; \ |
7009 |
-+ struct bfq_group *bfqg; \ |
7010 |
-+ int ret = -EINVAL; \ |
7011 |
-+ \ |
7012 |
-+ if (val < (__MIN) || val > (__MAX)) \ |
7013 |
-+ return ret; \ |
7014 |
-+ \ |
7015 |
-+ ret = -ENODEV; \ |
7016 |
-+ mutex_lock(&bfqio_mutex); \ |
7017 |
-+ if (bfqio_is_removed(cgroup)) \ |
7018 |
-+ goto out_unlock; \ |
7019 |
-+ ret = 0; \ |
7020 |
-+ \ |
7021 |
-+ bgrp = cgroup_to_bfqio(cgroup); \ |
7022 |
-+ \ |
7023 |
-+ spin_lock_irq(&bgrp->lock); \ |
7024 |
-+ bgrp->__VAR = (unsigned short)val; \ |
7025 |
-+ hlist_for_each_entry(bfqg, &bgrp->group_data, group_node) { \ |
7026 |
-+ /* \ |
7027 |
-+ * Setting the ioprio_changed flag of the entity \ |
7028 |
-+ * to 1 with new_##__VAR == ##__VAR would re-set \ |
7029 |
-+ * the value of the weight to its ioprio mapping. \ |
7030 |
-+ * Set the flag only if necessary. \ |
7031 |
-+ */ \ |
7032 |
-+ if ((unsigned short)val != bfqg->entity.new_##__VAR) { \ |
7033 |
-+ bfqg->entity.new_##__VAR = (unsigned short)val; \ |
7034 |
-+ smp_wmb(); \ |
7035 |
-+ bfqg->entity.ioprio_changed = 1; \ |
7036 |
-+ } \ |
7037 |
-+ } \ |
7038 |
-+ spin_unlock_irq(&bgrp->lock); \ |
7039 |
-+ \ |
7040 |
-+out_unlock: \ |
7041 |
-+ mutex_unlock(&bfqio_mutex); \ |
7042 |
-+ return ret; \ |
7043 |
-+} |
7044 |
-+ |
7045 |
-+STORE_FUNCTION(weight, BFQ_MIN_WEIGHT, BFQ_MAX_WEIGHT); |
7046 |
-+STORE_FUNCTION(ioprio, 0, IOPRIO_BE_NR - 1); |
7047 |
-+STORE_FUNCTION(ioprio_class, IOPRIO_CLASS_RT, IOPRIO_CLASS_IDLE); |
7048 |
-+#undef STORE_FUNCTION |
7049 |
-+ |
7050 |
-+static struct cftype bfqio_files[] = { |
7051 |
-+ { |
7052 |
-+ .name = "weight", |
7053 |
-+ .read_u64 = bfqio_cgroup_weight_read, |
7054 |
-+ .write_u64 = bfqio_cgroup_weight_write, |
7055 |
-+ }, |
7056 |
-+ { |
7057 |
-+ .name = "ioprio", |
7058 |
-+ .read_u64 = bfqio_cgroup_ioprio_read, |
7059 |
-+ .write_u64 = bfqio_cgroup_ioprio_write, |
7060 |
-+ }, |
7061 |
-+ { |
7062 |
-+ .name = "ioprio_class", |
7063 |
-+ .read_u64 = bfqio_cgroup_ioprio_class_read, |
7064 |
-+ .write_u64 = bfqio_cgroup_ioprio_class_write, |
7065 |
-+ }, |
7066 |
-+ { }, /* terminate */ |
7067 |
-+}; |
7068 |
-+ |
7069 |
-+static struct cgroup_subsys_state *bfqio_create(struct cgroup *cgroup) |
7070 |
-+{ |
7071 |
-+ struct bfqio_cgroup *bgrp; |
7072 |
-+ |
7073 |
-+ if (cgroup->parent != NULL) { |
7074 |
-+ bgrp = kzalloc(sizeof(*bgrp), GFP_KERNEL); |
7075 |
-+ if (bgrp == NULL) |
7076 |
-+ return ERR_PTR(-ENOMEM); |
7077 |
-+ } else |
7078 |
-+ bgrp = &bfqio_root_cgroup; |
7079 |
-+ |
7080 |
-+ spin_lock_init(&bgrp->lock); |
7081 |
-+ INIT_HLIST_HEAD(&bgrp->group_data); |
7082 |
-+ bgrp->ioprio = BFQ_DEFAULT_GRP_IOPRIO; |
7083 |
-+ bgrp->ioprio_class = BFQ_DEFAULT_GRP_CLASS; |
7084 |
-+ |
7085 |
-+ return &bgrp->css; |
7086 |
-+} |
7087 |
-+ |
7088 |
-+/* |
7089 |
-+ * We cannot support shared io contexts, as we have no means to support |
7090 |
-+ * two tasks with the same ioc in two different groups without major rework |
7091 |
-+ * of the main bic/bfqq data structures. By now we allow a task to change |
7092 |
-+ * its cgroup only if it's the only owner of its ioc; the drawback of this |
7093 |
-+ * behavior is that a group containing a task that forked using CLONE_IO |
7094 |
-+ * will not be destroyed until the tasks sharing the ioc die. |
7095 |
-+ */ |
7096 |
-+static int bfqio_can_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
7097 |
-+{ |
7098 |
-+ struct task_struct *task; |
7099 |
-+ struct io_context *ioc; |
7100 |
-+ int ret = 0; |
7101 |
-+ |
7102 |
-+ cgroup_taskset_for_each(task, cgroup, tset) { |
7103 |
-+ /* task_lock() is needed to avoid races with exit_io_context() */ |
7104 |
-+ task_lock(task); |
7105 |
-+ ioc = task->io_context; |
7106 |
-+ if (ioc != NULL && atomic_read(&ioc->nr_tasks) > 1) |
7107 |
-+ /* |
7108 |
-+ * ioc == NULL means that the task is either too young or |
7109 |
-+ * exiting: if it has still no ioc the ioc can't be shared, |
7110 |
-+ * if the task is exiting the attach will fail anyway, no |
7111 |
-+ * matter what we return here. |
7112 |
-+ */ |
7113 |
-+ ret = -EINVAL; |
7114 |
-+ task_unlock(task); |
7115 |
-+ if (ret) |
7116 |
-+ break; |
7117 |
-+ } |
7118 |
-+ |
7119 |
-+ return ret; |
7120 |
-+} |
7121 |
-+ |
7122 |
-+static void bfqio_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
7123 |
-+{ |
7124 |
-+ struct task_struct *task; |
7125 |
-+ struct io_context *ioc; |
7126 |
-+ struct io_cq *icq; |
7127 |
-+ |
7128 |
-+ /* |
7129 |
-+ * IMPORTANT NOTE: The move of more than one process at a time to a |
7130 |
-+ * new group has not yet been tested. |
7131 |
-+ */ |
7132 |
-+ cgroup_taskset_for_each(task, cgroup, tset) { |
7133 |
-+ ioc = get_task_io_context(task, GFP_ATOMIC, NUMA_NO_NODE); |
7134 |
-+ if (ioc) { |
7135 |
-+ /* |
7136 |
-+ * Handle cgroup change here. |
7137 |
-+ */ |
7138 |
-+ rcu_read_lock(); |
7139 |
-+ hlist_for_each_entry_rcu(icq, &ioc->icq_list, ioc_node) |
7140 |
-+ if (!strncmp(icq->q->elevator->type->elevator_name, |
7141 |
-+ "bfq", ELV_NAME_MAX)) |
7142 |
-+ bfq_bic_change_cgroup(icq_to_bic(icq), |
7143 |
-+ cgroup); |
7144 |
-+ rcu_read_unlock(); |
7145 |
-+ put_io_context(ioc); |
7146 |
-+ } |
7147 |
-+ } |
7148 |
-+} |
7149 |
-+ |
7150 |
-+static void bfqio_destroy(struct cgroup *cgroup) |
7151 |
-+{ |
7152 |
-+ struct bfqio_cgroup *bgrp = cgroup_to_bfqio(cgroup); |
7153 |
-+ struct hlist_node *tmp; |
7154 |
-+ struct bfq_group *bfqg; |
7155 |
-+ |
7156 |
-+ /* |
7157 |
-+ * Since we are destroying the cgroup, there are no more tasks |
7158 |
-+ * referencing it, and all the RCU grace periods that may have |
7159 |
-+ * referenced it are ended (as the destruction of the parent |
7160 |
-+ * cgroup is RCU-safe); bgrp->group_data will not be accessed by |
7161 |
-+ * anything else and we don't need any synchronization. |
7162 |
-+ */ |
7163 |
-+ hlist_for_each_entry_safe(bfqg, tmp, &bgrp->group_data, group_node) |
7164 |
-+ bfq_destroy_group(bgrp, bfqg); |
7165 |
-+ |
7166 |
-+ BUG_ON(!hlist_empty(&bgrp->group_data)); |
7167 |
-+ |
7168 |
-+ kfree(bgrp); |
7169 |
-+} |
7170 |
-+ |
7171 |
-+struct cgroup_subsys bfqio_subsys = { |
7172 |
-+ .name = "bfqio", |
7173 |
-+ .css_alloc = bfqio_create, |
7174 |
-+ .can_attach = bfqio_can_attach, |
7175 |
-+ .attach = bfqio_attach, |
7176 |
-+ .css_free = bfqio_destroy, |
7177 |
-+ .subsys_id = bfqio_subsys_id, |
7178 |
-+ .base_cftypes = bfqio_files, |
7179 |
-+}; |
7180 |
-+#else |
7181 |
-+static inline void bfq_init_entity(struct bfq_entity *entity, |
7182 |
-+ struct bfq_group *bfqg) |
7183 |
-+{ |
7184 |
-+ entity->weight = entity->new_weight; |
7185 |
-+ entity->orig_weight = entity->new_weight; |
7186 |
-+ entity->ioprio = entity->new_ioprio; |
7187 |
-+ entity->ioprio_class = entity->new_ioprio_class; |
7188 |
-+ entity->sched_data = &bfqg->sched_data; |
7189 |
-+} |
7190 |
-+ |
7191 |
-+static inline struct bfq_group * |
7192 |
-+bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
7193 |
-+{ |
7194 |
-+ struct bfq_data *bfqd = bic_to_bfqd(bic); |
7195 |
-+ return bfqd->root_group; |
7196 |
-+} |
7197 |
-+ |
7198 |
-+static inline void bfq_bfqq_move(struct bfq_data *bfqd, |
7199 |
-+ struct bfq_queue *bfqq, |
7200 |
-+ struct bfq_entity *entity, |
7201 |
-+ struct bfq_group *bfqg) |
7202 |
-+{ |
7203 |
-+} |
7204 |
-+ |
7205 |
-+static void bfq_end_raising_async(struct bfq_data *bfqd) |
7206 |
-+{ |
7207 |
-+ bfq_end_raising_async_queues(bfqd, bfqd->root_group); |
7208 |
-+} |
7209 |
-+ |
7210 |
-+static inline void bfq_disconnect_groups(struct bfq_data *bfqd) |
7211 |
-+{ |
7212 |
-+ bfq_put_async_queues(bfqd, bfqd->root_group); |
7213 |
-+} |
7214 |
-+ |
7215 |
-+static inline void bfq_free_root_group(struct bfq_data *bfqd) |
7216 |
-+{ |
7217 |
-+ kfree(bfqd->root_group); |
7218 |
-+} |
7219 |
-+ |
7220 |
-+static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
7221 |
-+{ |
7222 |
-+ struct bfq_group *bfqg; |
7223 |
-+ int i; |
7224 |
-+ |
7225 |
-+ bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); |
7226 |
-+ if (bfqg == NULL) |
7227 |
-+ return NULL; |
7228 |
-+ |
7229 |
-+ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
7230 |
-+ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
7231 |
-+ |
7232 |
-+ return bfqg; |
7233 |
-+} |
7234 |
-+#endif |
7235 |
-diff --git a/block/bfq-ioc.c b/block/bfq-ioc.c |
7236 |
-new file mode 100644 |
7237 |
-index 0000000..326e3ec |
7238 |
---- /dev/null |
7239 |
-+++ b/block/bfq-ioc.c |
7240 |
-@@ -0,0 +1,36 @@ |
7241 |
-+/* |
7242 |
-+ * BFQ: I/O context handling. |
7243 |
-+ * |
7244 |
-+ * Based on ideas and code from CFQ: |
7245 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
7246 |
-+ * |
7247 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
7248 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
7249 |
-+ * |
7250 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
7251 |
-+ */ |
7252 |
-+ |
7253 |
-+/** |
7254 |
-+ * icq_to_bic - convert iocontext queue structure to bfq_io_cq. |
7255 |
-+ * @icq: the iocontext queue. |
7256 |
-+ */ |
7257 |
-+static inline struct bfq_io_cq *icq_to_bic(struct io_cq *icq) |
7258 |
-+{ |
7259 |
-+ /* bic->icq is the first member, %NULL will convert to %NULL */ |
7260 |
-+ return container_of(icq, struct bfq_io_cq, icq); |
7261 |
-+} |
7262 |
-+ |
7263 |
-+/** |
7264 |
-+ * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. |
7265 |
-+ * @bfqd: the lookup key. |
7266 |
-+ * @ioc: the io_context of the process doing I/O. |
7267 |
-+ * |
7268 |
-+ * Queue lock must be held. |
7269 |
-+ */ |
7270 |
-+static inline struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd, |
7271 |
-+ struct io_context *ioc) |
7272 |
-+{ |
7273 |
-+ if(ioc) |
7274 |
-+ return icq_to_bic(ioc_lookup_icq(ioc, bfqd->queue)); |
7275 |
-+ return NULL; |
7276 |
-+} |
7277 |
-diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c |
7278 |
-new file mode 100644 |
7279 |
-index 0000000..0ed2746 |
7280 |
---- /dev/null |
7281 |
-+++ b/block/bfq-iosched.c |
7282 |
-@@ -0,0 +1,3082 @@ |
7283 |
-+/* |
7284 |
-+ * BFQ, or Budget Fair Queueing, disk scheduler. |
7285 |
-+ * |
7286 |
-+ * Based on ideas and code from CFQ: |
7287 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
7288 |
-+ * |
7289 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
7290 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
7291 |
-+ * |
7292 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
7293 |
-+ * |
7294 |
-+ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
7295 |
-+ * |
7296 |
-+ * BFQ is a proportional share disk scheduling algorithm based on the |
7297 |
-+ * slice-by-slice service scheme of CFQ. But BFQ assigns budgets, |
7298 |
-+ * measured in number of sectors, to tasks instead of time slices. |
7299 |
-+ * The disk is not granted to the active task for a given time slice, |
7300 |
-+ * but until it has exahusted its assigned budget. This change from |
7301 |
-+ * the time to the service domain allows BFQ to distribute the disk |
7302 |
-+ * bandwidth among tasks as desired, without any distortion due to |
7303 |
-+ * ZBR, workload fluctuations or other factors. BFQ uses an ad hoc |
7304 |
-+ * internal scheduler, called B-WF2Q+, to schedule tasks according to |
7305 |
-+ * their budgets. Thanks to this accurate scheduler, BFQ can afford |
7306 |
-+ * to assign high budgets to disk-bound non-seeky tasks (to boost the |
7307 |
-+ * throughput), and yet guarantee low latencies to interactive and |
7308 |
-+ * soft real-time applications. |
7309 |
-+ * |
7310 |
-+ * BFQ has been introduced in [1], where the interested reader can |
7311 |
-+ * find an accurate description of the algorithm, the bandwidth |
7312 |
-+ * distribution and latency guarantees it provides, plus formal proofs |
7313 |
-+ * of all the properties. With respect to the algorithm presented in |
7314 |
-+ * the paper, this implementation adds several little heuristics, and |
7315 |
-+ * a hierarchical extension, based on H-WF2Q+. |
7316 |
-+ * |
7317 |
-+ * B-WF2Q+ is based on WF2Q+, that is described in [2], together with |
7318 |
-+ * H-WF2Q+, while the augmented tree used to implement B-WF2Q+ with O(log N) |
7319 |
-+ * complexity derives from the one introduced with EEVDF in [3]. |
7320 |
-+ * |
7321 |
-+ * [1] P. Valente and F. Checconi, ``High Throughput Disk Scheduling |
7322 |
-+ * with Deterministic Guarantees on Bandwidth Distribution,'', |
7323 |
-+ * IEEE Transactions on Computer, May 2010. |
7324 |
-+ * |
7325 |
-+ * http://algo.ing.unimo.it/people/paolo/disk_sched/bfq-techreport.pdf |
7326 |
-+ * |
7327 |
-+ * [2] Jon C.R. Bennett and H. Zhang, ``Hierarchical Packet Fair Queueing |
7328 |
-+ * Algorithms,'' IEEE/ACM Transactions on Networking, 5(5):675-689, |
7329 |
-+ * Oct 1997. |
7330 |
-+ * |
7331 |
-+ * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz |
7332 |
-+ * |
7333 |
-+ * [3] I. Stoica and H. Abdel-Wahab, ``Earliest Eligible Virtual Deadline |
7334 |
-+ * First: A Flexible and Accurate Mechanism for Proportional Share |
7335 |
-+ * Resource Allocation,'' technical report. |
7336 |
-+ * |
7337 |
-+ * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf |
7338 |
-+ */ |
7339 |
-+#include <linux/module.h> |
7340 |
-+#include <linux/slab.h> |
7341 |
-+#include <linux/blkdev.h> |
7342 |
-+#include <linux/cgroup.h> |
7343 |
-+#include <linux/elevator.h> |
7344 |
-+#include <linux/jiffies.h> |
7345 |
-+#include <linux/rbtree.h> |
7346 |
-+#include <linux/ioprio.h> |
7347 |
-+#include "bfq.h" |
7348 |
-+#include "blk.h" |
7349 |
-+ |
7350 |
-+/* Max number of dispatches in one round of service. */ |
7351 |
-+static const int bfq_quantum = 4; |
7352 |
-+ |
7353 |
-+/* Expiration time of sync (0) and async (1) requests, in jiffies. */ |
7354 |
-+static const int bfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
7355 |
-+ |
7356 |
-+/* Maximum backwards seek, in KiB. */ |
7357 |
-+static const int bfq_back_max = 16 * 1024; |
7358 |
-+ |
7359 |
-+/* Penalty of a backwards seek, in number of sectors. */ |
7360 |
-+static const int bfq_back_penalty = 2; |
7361 |
-+ |
7362 |
-+/* Idling period duration, in jiffies. */ |
7363 |
-+static int bfq_slice_idle = HZ / 125; |
7364 |
-+ |
7365 |
-+/* Default maximum budget values, in sectors and number of requests. */ |
7366 |
-+static const int bfq_default_max_budget = 16 * 1024; |
7367 |
-+static const int bfq_max_budget_async_rq = 4; |
7368 |
-+ |
7369 |
-+/* |
7370 |
-+ * Async to sync throughput distribution is controlled as follows: |
7371 |
-+ * when an async request is served, the entity is charged the number |
7372 |
-+ * of sectors of the request, multipled by the factor below |
7373 |
-+ */ |
7374 |
-+static const int bfq_async_charge_factor = 10; |
7375 |
-+ |
7376 |
-+/* Default timeout values, in jiffies, approximating CFQ defaults. */ |
7377 |
-+static const int bfq_timeout_sync = HZ / 8; |
7378 |
-+static int bfq_timeout_async = HZ / 25; |
7379 |
-+ |
7380 |
-+struct kmem_cache *bfq_pool; |
7381 |
-+ |
7382 |
-+/* Below this threshold (in ms), we consider thinktime immediate. */ |
7383 |
-+#define BFQ_MIN_TT 2 |
7384 |
-+ |
7385 |
-+/* hw_tag detection: parallel requests threshold and min samples needed. */ |
7386 |
-+#define BFQ_HW_QUEUE_THRESHOLD 4 |
7387 |
-+#define BFQ_HW_QUEUE_SAMPLES 32 |
7388 |
-+ |
7389 |
-+#define BFQQ_SEEK_THR (sector_t)(8 * 1024) |
7390 |
-+#define BFQQ_SEEKY(bfqq) ((bfqq)->seek_mean > BFQQ_SEEK_THR) |
7391 |
-+ |
7392 |
-+/* Min samples used for peak rate estimation (for autotuning). */ |
7393 |
-+#define BFQ_PEAK_RATE_SAMPLES 32 |
7394 |
-+ |
7395 |
-+/* Shift used for peak rate fixed precision calculations. */ |
7396 |
-+#define BFQ_RATE_SHIFT 16 |
7397 |
-+ |
7398 |
-+/* |
7399 |
-+ * The duration of the weight raising for interactive applications is |
7400 |
-+ * computed automatically (as default behaviour), using the following |
7401 |
-+ * formula: duration = (R / r) * T, where r is the peak rate of the |
7402 |
-+ * disk, and R and T are two reference parameters. In particular, R is |
7403 |
-+ * the peak rate of a reference disk, and T is about the maximum time |
7404 |
-+ * for starting popular large applications on that disk, under BFQ and |
7405 |
-+ * while reading two files in parallel. Finally, BFQ uses two |
7406 |
-+ * different pairs (R, T) depending on whether the disk is rotational |
7407 |
-+ * or non-rotational. |
7408 |
-+ */ |
7409 |
-+#define T_rot (msecs_to_jiffies(5500)) |
7410 |
-+#define T_nonrot (msecs_to_jiffies(2000)) |
7411 |
-+/* Next two quantities are in sectors/usec, left-shifted by BFQ_RATE_SHIFT */ |
7412 |
-+#define R_rot 17415 |
7413 |
-+#define R_nonrot 34791 |
7414 |
-+ |
7415 |
-+#define BFQ_SERVICE_TREE_INIT ((struct bfq_service_tree) \ |
7416 |
-+ { RB_ROOT, RB_ROOT, NULL, NULL, 0, 0 }) |
7417 |
-+ |
7418 |
-+#define RQ_BIC(rq) ((struct bfq_io_cq *) (rq)->elv.priv[0]) |
7419 |
-+#define RQ_BFQQ(rq) ((rq)->elv.priv[1]) |
7420 |
-+ |
7421 |
-+static inline void bfq_schedule_dispatch(struct bfq_data *bfqd); |
7422 |
-+ |
7423 |
-+#include "bfq-ioc.c" |
7424 |
-+#include "bfq-sched.c" |
7425 |
-+#include "bfq-cgroup.c" |
7426 |
-+ |
7427 |
-+#define bfq_class_idle(bfqq) ((bfqq)->entity.ioprio_class ==\ |
7428 |
-+ IOPRIO_CLASS_IDLE) |
7429 |
-+#define bfq_class_rt(bfqq) ((bfqq)->entity.ioprio_class ==\ |
7430 |
-+ IOPRIO_CLASS_RT) |
7431 |
-+ |
7432 |
-+#define bfq_sample_valid(samples) ((samples) > 80) |
7433 |
-+ |
7434 |
-+/* |
7435 |
-+ * We regard a request as SYNC, if either it's a read or has the SYNC bit |
7436 |
-+ * set (in which case it could also be a direct WRITE). |
7437 |
-+ */ |
7438 |
-+static inline int bfq_bio_sync(struct bio *bio) |
7439 |
-+{ |
7440 |
-+ if (bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC)) |
7441 |
-+ return 1; |
7442 |
-+ |
7443 |
-+ return 0; |
7444 |
-+} |
7445 |
-+ |
7446 |
-+/* |
7447 |
-+ * Scheduler run of queue, if there are requests pending and no one in the |
7448 |
-+ * driver that will restart queueing. |
7449 |
-+ */ |
7450 |
-+static inline void bfq_schedule_dispatch(struct bfq_data *bfqd) |
7451 |
-+{ |
7452 |
-+ if (bfqd->queued != 0) { |
7453 |
-+ bfq_log(bfqd, "schedule dispatch"); |
7454 |
-+ kblockd_schedule_work(bfqd->queue, &bfqd->unplug_work); |
7455 |
-+ } |
7456 |
-+} |
7457 |
-+ |
7458 |
-+/* |
7459 |
-+ * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
7460 |
-+ * We choose the request that is closesr to the head right now. Distance |
7461 |
-+ * behind the head is penalized and only allowed to a certain extent. |
7462 |
-+ */ |
7463 |
-+static struct request *bfq_choose_req(struct bfq_data *bfqd, |
7464 |
-+ struct request *rq1, |
7465 |
-+ struct request *rq2, |
7466 |
-+ sector_t last) |
7467 |
-+{ |
7468 |
-+ sector_t s1, s2, d1 = 0, d2 = 0; |
7469 |
-+ unsigned long back_max; |
7470 |
-+#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
7471 |
-+#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ |
7472 |
-+ unsigned wrap = 0; /* bit mask: requests behind the disk head? */ |
7473 |
-+ |
7474 |
-+ if (rq1 == NULL || rq1 == rq2) |
7475 |
-+ return rq2; |
7476 |
-+ if (rq2 == NULL) |
7477 |
-+ return rq1; |
7478 |
-+ |
7479 |
-+ if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
7480 |
-+ return rq1; |
7481 |
-+ else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) |
7482 |
-+ return rq2; |
7483 |
-+ if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) |
7484 |
-+ return rq1; |
7485 |
-+ else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) |
7486 |
-+ return rq2; |
7487 |
-+ |
7488 |
-+ s1 = blk_rq_pos(rq1); |
7489 |
-+ s2 = blk_rq_pos(rq2); |
7490 |
-+ |
7491 |
-+ /* |
7492 |
-+ * By definition, 1KiB is 2 sectors. |
7493 |
-+ */ |
7494 |
-+ back_max = bfqd->bfq_back_max * 2; |
7495 |
-+ |
7496 |
-+ /* |
7497 |
-+ * Strict one way elevator _except_ in the case where we allow |
7498 |
-+ * short backward seeks which are biased as twice the cost of a |
7499 |
-+ * similar forward seek. |
7500 |
-+ */ |
7501 |
-+ if (s1 >= last) |
7502 |
-+ d1 = s1 - last; |
7503 |
-+ else if (s1 + back_max >= last) |
7504 |
-+ d1 = (last - s1) * bfqd->bfq_back_penalty; |
7505 |
-+ else |
7506 |
-+ wrap |= BFQ_RQ1_WRAP; |
7507 |
-+ |
7508 |
-+ if (s2 >= last) |
7509 |
-+ d2 = s2 - last; |
7510 |
-+ else if (s2 + back_max >= last) |
7511 |
-+ d2 = (last - s2) * bfqd->bfq_back_penalty; |
7512 |
-+ else |
7513 |
-+ wrap |= BFQ_RQ2_WRAP; |
7514 |
-+ |
7515 |
-+ /* Found required data */ |
7516 |
-+ |
7517 |
-+ /* |
7518 |
-+ * By doing switch() on the bit mask "wrap" we avoid having to |
7519 |
-+ * check two variables for all permutations: --> faster! |
7520 |
-+ */ |
7521 |
-+ switch (wrap) { |
7522 |
-+ case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
7523 |
-+ if (d1 < d2) |
7524 |
-+ return rq1; |
7525 |
-+ else if (d2 < d1) |
7526 |
-+ return rq2; |
7527 |
-+ else { |
7528 |
-+ if (s1 >= s2) |
7529 |
-+ return rq1; |
7530 |
-+ else |
7531 |
-+ return rq2; |
7532 |
-+ } |
7533 |
-+ |
7534 |
-+ case BFQ_RQ2_WRAP: |
7535 |
-+ return rq1; |
7536 |
-+ case BFQ_RQ1_WRAP: |
7537 |
-+ return rq2; |
7538 |
-+ case (BFQ_RQ1_WRAP|BFQ_RQ2_WRAP): /* both rqs wrapped */ |
7539 |
-+ default: |
7540 |
-+ /* |
7541 |
-+ * Since both rqs are wrapped, |
7542 |
-+ * start with the one that's further behind head |
7543 |
-+ * (--> only *one* back seek required), |
7544 |
-+ * since back seek takes more time than forward. |
7545 |
-+ */ |
7546 |
-+ if (s1 <= s2) |
7547 |
-+ return rq1; |
7548 |
-+ else |
7549 |
-+ return rq2; |
7550 |
-+ } |
7551 |
-+} |
7552 |
-+ |
7553 |
-+static struct bfq_queue * |
7554 |
-+bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root, |
7555 |
-+ sector_t sector, struct rb_node **ret_parent, |
7556 |
-+ struct rb_node ***rb_link) |
7557 |
-+{ |
7558 |
-+ struct rb_node **p, *parent; |
7559 |
-+ struct bfq_queue *bfqq = NULL; |
7560 |
-+ |
7561 |
-+ parent = NULL; |
7562 |
-+ p = &root->rb_node; |
7563 |
-+ while (*p) { |
7564 |
-+ struct rb_node **n; |
7565 |
-+ |
7566 |
-+ parent = *p; |
7567 |
-+ bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
7568 |
-+ |
7569 |
-+ /* |
7570 |
-+ * Sort strictly based on sector. Smallest to the left, |
7571 |
-+ * largest to the right. |
7572 |
-+ */ |
7573 |
-+ if (sector > blk_rq_pos(bfqq->next_rq)) |
7574 |
-+ n = &(*p)->rb_right; |
7575 |
-+ else if (sector < blk_rq_pos(bfqq->next_rq)) |
7576 |
-+ n = &(*p)->rb_left; |
7577 |
-+ else |
7578 |
-+ break; |
7579 |
-+ p = n; |
7580 |
-+ bfqq = NULL; |
7581 |
-+ } |
7582 |
-+ |
7583 |
-+ *ret_parent = parent; |
7584 |
-+ if (rb_link) |
7585 |
-+ *rb_link = p; |
7586 |
-+ |
7587 |
-+ bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d", |
7588 |
-+ (long long unsigned)sector, |
7589 |
-+ bfqq != NULL ? bfqq->pid : 0); |
7590 |
-+ |
7591 |
-+ return bfqq; |
7592 |
-+} |
7593 |
-+ |
7594 |
-+static void bfq_rq_pos_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
7595 |
-+{ |
7596 |
-+ struct rb_node **p, *parent; |
7597 |
-+ struct bfq_queue *__bfqq; |
7598 |
-+ |
7599 |
-+ if (bfqq->pos_root != NULL) { |
7600 |
-+ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
7601 |
-+ bfqq->pos_root = NULL; |
7602 |
-+ } |
7603 |
-+ |
7604 |
-+ if (bfq_class_idle(bfqq)) |
7605 |
-+ return; |
7606 |
-+ if (!bfqq->next_rq) |
7607 |
-+ return; |
7608 |
-+ |
7609 |
-+ bfqq->pos_root = &bfqd->rq_pos_tree; |
7610 |
-+ __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root, |
7611 |
-+ blk_rq_pos(bfqq->next_rq), &parent, &p); |
7612 |
-+ if (__bfqq == NULL) { |
7613 |
-+ rb_link_node(&bfqq->pos_node, parent, p); |
7614 |
-+ rb_insert_color(&bfqq->pos_node, bfqq->pos_root); |
7615 |
-+ } else |
7616 |
-+ bfqq->pos_root = NULL; |
7617 |
-+} |
7618 |
-+ |
7619 |
-+static struct request *bfq_find_next_rq(struct bfq_data *bfqd, |
7620 |
-+ struct bfq_queue *bfqq, |
7621 |
-+ struct request *last) |
7622 |
-+{ |
7623 |
-+ struct rb_node *rbnext = rb_next(&last->rb_node); |
7624 |
-+ struct rb_node *rbprev = rb_prev(&last->rb_node); |
7625 |
-+ struct request *next = NULL, *prev = NULL; |
7626 |
-+ |
7627 |
-+ BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
7628 |
-+ |
7629 |
-+ if (rbprev != NULL) |
7630 |
-+ prev = rb_entry_rq(rbprev); |
7631 |
-+ |
7632 |
-+ if (rbnext != NULL) |
7633 |
-+ next = rb_entry_rq(rbnext); |
7634 |
-+ else { |
7635 |
-+ rbnext = rb_first(&bfqq->sort_list); |
7636 |
-+ if (rbnext && rbnext != &last->rb_node) |
7637 |
-+ next = rb_entry_rq(rbnext); |
7638 |
-+ } |
7639 |
-+ |
7640 |
-+ return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); |
7641 |
-+} |
7642 |
-+ |
7643 |
-+static void bfq_del_rq_rb(struct request *rq) |
7644 |
-+{ |
7645 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7646 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
7647 |
-+ const int sync = rq_is_sync(rq); |
7648 |
-+ |
7649 |
-+ BUG_ON(bfqq->queued[sync] == 0); |
7650 |
-+ bfqq->queued[sync]--; |
7651 |
-+ bfqd->queued--; |
7652 |
-+ |
7653 |
-+ elv_rb_del(&bfqq->sort_list, rq); |
7654 |
-+ |
7655 |
-+ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
7656 |
-+ if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->active_queue) |
7657 |
-+ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
7658 |
-+ /* |
7659 |
-+ * Remove queue from request-position tree as it is empty. |
7660 |
-+ */ |
7661 |
-+ if (bfqq->pos_root != NULL) { |
7662 |
-+ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
7663 |
-+ bfqq->pos_root = NULL; |
7664 |
-+ } |
7665 |
-+ } |
7666 |
-+} |
7667 |
-+ |
7668 |
-+/* see the definition of bfq_async_charge_factor for details */ |
7669 |
-+static inline unsigned long bfq_serv_to_charge(struct request *rq, |
7670 |
-+ struct bfq_queue *bfqq) |
7671 |
-+{ |
7672 |
-+ return blk_rq_sectors(rq) * |
7673 |
-+ (1 + ((!bfq_bfqq_sync(bfqq)) * (bfqq->raising_coeff == 1) * |
7674 |
-+ bfq_async_charge_factor)); |
7675 |
-+} |
7676 |
-+ |
7677 |
-+/** |
7678 |
-+ * bfq_updated_next_req - update the queue after a new next_rq selection. |
7679 |
-+ * @bfqd: the device data the queue belongs to. |
7680 |
-+ * @bfqq: the queue to update. |
7681 |
-+ * |
7682 |
-+ * If the first request of a queue changes we make sure that the queue |
7683 |
-+ * has enough budget to serve at least its first request (if the |
7684 |
-+ * request has grown). We do this because if the queue has not enough |
7685 |
-+ * budget for its first request, it has to go through two dispatch |
7686 |
-+ * rounds to actually get it dispatched. |
7687 |
-+ */ |
7688 |
-+static void bfq_updated_next_req(struct bfq_data *bfqd, |
7689 |
-+ struct bfq_queue *bfqq) |
7690 |
-+{ |
7691 |
-+ struct bfq_entity *entity = &bfqq->entity; |
7692 |
-+ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
7693 |
-+ struct request *next_rq = bfqq->next_rq; |
7694 |
-+ unsigned long new_budget; |
7695 |
-+ |
7696 |
-+ if (next_rq == NULL) |
7697 |
-+ return; |
7698 |
-+ |
7699 |
-+ if (bfqq == bfqd->active_queue) |
7700 |
-+ /* |
7701 |
-+ * In order not to break guarantees, budgets cannot be |
7702 |
-+ * changed after an entity has been selected. |
7703 |
-+ */ |
7704 |
-+ return; |
7705 |
-+ |
7706 |
-+ BUG_ON(entity->tree != &st->active); |
7707 |
-+ BUG_ON(entity == entity->sched_data->active_entity); |
7708 |
-+ |
7709 |
-+ new_budget = max_t(unsigned long, bfqq->max_budget, |
7710 |
-+ bfq_serv_to_charge(next_rq, bfqq)); |
7711 |
-+ entity->budget = new_budget; |
7712 |
-+ bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", new_budget); |
7713 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
7714 |
-+} |
7715 |
-+ |
7716 |
-+static inline unsigned int bfq_wrais_duration(struct bfq_data *bfqd) |
7717 |
-+{ |
7718 |
-+ u64 dur; |
7719 |
-+ |
7720 |
-+ if (bfqd->bfq_raising_max_time > 0) |
7721 |
-+ return bfqd->bfq_raising_max_time; |
7722 |
-+ |
7723 |
-+ dur = bfqd->RT_prod; |
7724 |
-+ do_div(dur, bfqd->peak_rate); |
7725 |
-+ |
7726 |
-+ return dur; |
7727 |
-+} |
7728 |
-+ |
7729 |
-+static void bfq_add_rq_rb(struct request *rq) |
7730 |
-+{ |
7731 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7732 |
-+ struct bfq_entity *entity = &bfqq->entity; |
7733 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
7734 |
-+ struct request *next_rq, *prev; |
7735 |
-+ unsigned long old_raising_coeff = bfqq->raising_coeff; |
7736 |
-+ int idle_for_long_time = bfqq->budget_timeout + |
7737 |
-+ bfqd->bfq_raising_min_idle_time < jiffies; |
7738 |
-+ |
7739 |
-+ bfq_log_bfqq(bfqd, bfqq, "add_rq_rb %d", rq_is_sync(rq)); |
7740 |
-+ bfqq->queued[rq_is_sync(rq)]++; |
7741 |
-+ bfqd->queued++; |
7742 |
-+ |
7743 |
-+ elv_rb_add(&bfqq->sort_list, rq); |
7744 |
-+ |
7745 |
-+ /* |
7746 |
-+ * Check if this request is a better next-serve candidate. |
7747 |
-+ */ |
7748 |
-+ prev = bfqq->next_rq; |
7749 |
-+ next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); |
7750 |
-+ BUG_ON(next_rq == NULL); |
7751 |
-+ bfqq->next_rq = next_rq; |
7752 |
-+ |
7753 |
-+ /* |
7754 |
-+ * Adjust priority tree position, if next_rq changes. |
7755 |
-+ */ |
7756 |
-+ if (prev != bfqq->next_rq) |
7757 |
-+ bfq_rq_pos_tree_add(bfqd, bfqq); |
7758 |
-+ |
7759 |
-+ if (!bfq_bfqq_busy(bfqq)) { |
7760 |
-+ int soft_rt = bfqd->bfq_raising_max_softrt_rate > 0 && |
7761 |
-+ bfqq->soft_rt_next_start < jiffies; |
7762 |
-+ entity->budget = max_t(unsigned long, bfqq->max_budget, |
7763 |
-+ bfq_serv_to_charge(next_rq, bfqq)); |
7764 |
-+ |
7765 |
-+ if (! bfqd->low_latency) |
7766 |
-+ goto add_bfqq_busy; |
7767 |
-+ |
7768 |
-+ /* |
7769 |
-+ * If the queue is not being boosted and has been idle |
7770 |
-+ * for enough time, start a weight-raising period |
7771 |
-+ */ |
7772 |
-+ if(old_raising_coeff == 1 && (idle_for_long_time || soft_rt)) { |
7773 |
-+ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
7774 |
-+ if (idle_for_long_time) |
7775 |
-+ bfqq->raising_cur_max_time = |
7776 |
-+ bfq_wrais_duration(bfqd); |
7777 |
-+ else |
7778 |
-+ bfqq->raising_cur_max_time = |
7779 |
-+ bfqd->bfq_raising_rt_max_time; |
7780 |
-+ bfq_log_bfqq(bfqd, bfqq, |
7781 |
-+ "wrais starting at %llu msec," |
7782 |
-+ "rais_max_time %u", |
7783 |
-+ bfqq->last_rais_start_finish, |
7784 |
-+ jiffies_to_msecs(bfqq-> |
7785 |
-+ raising_cur_max_time)); |
7786 |
-+ } else if (old_raising_coeff > 1) { |
7787 |
-+ if (idle_for_long_time) |
7788 |
-+ bfqq->raising_cur_max_time = |
7789 |
-+ bfq_wrais_duration(bfqd); |
7790 |
-+ else if (bfqq->raising_cur_max_time == |
7791 |
-+ bfqd->bfq_raising_rt_max_time && |
7792 |
-+ !soft_rt) { |
7793 |
-+ bfqq->raising_coeff = 1; |
7794 |
-+ bfq_log_bfqq(bfqd, bfqq, |
7795 |
-+ "wrais ending at %llu msec," |
7796 |
-+ "rais_max_time %u", |
7797 |
-+ bfqq->last_rais_start_finish, |
7798 |
-+ jiffies_to_msecs(bfqq-> |
7799 |
-+ raising_cur_max_time)); |
7800 |
-+ } |
7801 |
-+ } |
7802 |
-+ if (old_raising_coeff != bfqq->raising_coeff) |
7803 |
-+ entity->ioprio_changed = 1; |
7804 |
-+add_bfqq_busy: |
7805 |
-+ bfq_add_bfqq_busy(bfqd, bfqq); |
7806 |
-+ } else { |
7807 |
-+ if(bfqd->low_latency && old_raising_coeff == 1 && |
7808 |
-+ !rq_is_sync(rq) && |
7809 |
-+ bfqq->last_rais_start_finish + |
7810 |
-+ bfqd->bfq_raising_min_inter_arr_async < jiffies) { |
7811 |
-+ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
7812 |
-+ bfqq->raising_cur_max_time = bfq_wrais_duration(bfqd); |
7813 |
-+ |
7814 |
-+ entity->ioprio_changed = 1; |
7815 |
-+ bfq_log_bfqq(bfqd, bfqq, |
7816 |
-+ "non-idle wrais starting at %llu msec," |
7817 |
-+ "rais_max_time %u", |
7818 |
-+ bfqq->last_rais_start_finish, |
7819 |
-+ jiffies_to_msecs(bfqq-> |
7820 |
-+ raising_cur_max_time)); |
7821 |
-+ } |
7822 |
-+ bfq_updated_next_req(bfqd, bfqq); |
7823 |
-+ } |
7824 |
-+ |
7825 |
-+ if(bfqd->low_latency && |
7826 |
-+ (old_raising_coeff == 1 || bfqq->raising_coeff == 1 || |
7827 |
-+ idle_for_long_time)) |
7828 |
-+ bfqq->last_rais_start_finish = jiffies; |
7829 |
-+} |
7830 |
-+ |
7831 |
-+static void bfq_reposition_rq_rb(struct bfq_queue *bfqq, struct request *rq) |
7832 |
-+{ |
7833 |
-+ elv_rb_del(&bfqq->sort_list, rq); |
7834 |
-+ bfqq->queued[rq_is_sync(rq)]--; |
7835 |
-+ bfqq->bfqd->queued--; |
7836 |
-+ bfq_add_rq_rb(rq); |
7837 |
-+} |
7838 |
-+ |
7839 |
-+static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, |
7840 |
-+ struct bio *bio) |
7841 |
-+{ |
7842 |
-+ struct task_struct *tsk = current; |
7843 |
-+ struct bfq_io_cq *bic; |
7844 |
-+ struct bfq_queue *bfqq; |
7845 |
-+ |
7846 |
-+ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
7847 |
-+ if (bic == NULL) |
7848 |
-+ return NULL; |
7849 |
-+ |
7850 |
-+ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
7851 |
-+ if (bfqq != NULL) { |
7852 |
-+ sector_t sector = bio->bi_sector + bio_sectors(bio); |
7853 |
-+ |
7854 |
-+ return elv_rb_find(&bfqq->sort_list, sector); |
7855 |
-+ } |
7856 |
-+ |
7857 |
-+ return NULL; |
7858 |
-+} |
7859 |
-+ |
7860 |
-+static void bfq_activate_request(struct request_queue *q, struct request *rq) |
7861 |
-+{ |
7862 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
7863 |
-+ |
7864 |
-+ bfqd->rq_in_driver++; |
7865 |
-+ bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
7866 |
-+ bfq_log(bfqd, "activate_request: new bfqd->last_position %llu", |
7867 |
-+ (long long unsigned)bfqd->last_position); |
7868 |
-+} |
7869 |
-+ |
7870 |
-+static void bfq_deactivate_request(struct request_queue *q, struct request *rq) |
7871 |
-+{ |
7872 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
7873 |
-+ |
7874 |
-+ WARN_ON(bfqd->rq_in_driver == 0); |
7875 |
-+ bfqd->rq_in_driver--; |
7876 |
-+} |
7877 |
-+ |
7878 |
-+static void bfq_remove_request(struct request *rq) |
7879 |
-+{ |
7880 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7881 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
7882 |
-+ |
7883 |
-+ if (bfqq->next_rq == rq) { |
7884 |
-+ bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); |
7885 |
-+ bfq_updated_next_req(bfqd, bfqq); |
7886 |
-+ } |
7887 |
-+ |
7888 |
-+ list_del_init(&rq->queuelist); |
7889 |
-+ bfq_del_rq_rb(rq); |
7890 |
-+ |
7891 |
-+ if (rq->cmd_flags & REQ_META) { |
7892 |
-+ WARN_ON(bfqq->meta_pending == 0); |
7893 |
-+ bfqq->meta_pending--; |
7894 |
-+ } |
7895 |
-+} |
7896 |
-+ |
7897 |
-+static int bfq_merge(struct request_queue *q, struct request **req, |
7898 |
-+ struct bio *bio) |
7899 |
-+{ |
7900 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
7901 |
-+ struct request *__rq; |
7902 |
-+ |
7903 |
-+ __rq = bfq_find_rq_fmerge(bfqd, bio); |
7904 |
-+ if (__rq != NULL && elv_rq_merge_ok(__rq, bio)) { |
7905 |
-+ *req = __rq; |
7906 |
-+ return ELEVATOR_FRONT_MERGE; |
7907 |
-+ } |
7908 |
-+ |
7909 |
-+ return ELEVATOR_NO_MERGE; |
7910 |
-+} |
7911 |
-+ |
7912 |
-+static void bfq_merged_request(struct request_queue *q, struct request *req, |
7913 |
-+ int type) |
7914 |
-+{ |
7915 |
-+ if (type == ELEVATOR_FRONT_MERGE) { |
7916 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(req); |
7917 |
-+ |
7918 |
-+ bfq_reposition_rq_rb(bfqq, req); |
7919 |
-+ } |
7920 |
-+} |
7921 |
-+ |
7922 |
-+static void bfq_merged_requests(struct request_queue *q, struct request *rq, |
7923 |
-+ struct request *next) |
7924 |
-+{ |
7925 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7926 |
-+ |
7927 |
-+ /* |
7928 |
-+ * Reposition in fifo if next is older than rq. |
7929 |
-+ */ |
7930 |
-+ if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && |
7931 |
-+ time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
7932 |
-+ list_move(&rq->queuelist, &next->queuelist); |
7933 |
-+ rq_set_fifo_time(rq, rq_fifo_time(next)); |
7934 |
-+ } |
7935 |
-+ |
7936 |
-+ if (bfqq->next_rq == next) |
7937 |
-+ bfqq->next_rq = rq; |
7938 |
-+ |
7939 |
-+ bfq_remove_request(next); |
7940 |
-+} |
7941 |
-+ |
7942 |
-+/* Must be called with bfqq != NULL */ |
7943 |
-+static inline void bfq_bfqq_end_raising(struct bfq_queue *bfqq) |
7944 |
-+{ |
7945 |
-+ BUG_ON(bfqq == NULL); |
7946 |
-+ bfqq->raising_coeff = 1; |
7947 |
-+ bfqq->raising_cur_max_time = 0; |
7948 |
-+ /* Trigger a weight change on the next activation of the queue */ |
7949 |
-+ bfqq->entity.ioprio_changed = 1; |
7950 |
-+} |
7951 |
-+ |
7952 |
-+static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
7953 |
-+ struct bfq_group *bfqg) |
7954 |
-+{ |
7955 |
-+ int i, j; |
7956 |
-+ |
7957 |
-+ for (i = 0; i < 2; i++) |
7958 |
-+ for (j = 0; j < IOPRIO_BE_NR; j++) |
7959 |
-+ if (bfqg->async_bfqq[i][j] != NULL) |
7960 |
-+ bfq_bfqq_end_raising(bfqg->async_bfqq[i][j]); |
7961 |
-+ if (bfqg->async_idle_bfqq != NULL) |
7962 |
-+ bfq_bfqq_end_raising(bfqg->async_idle_bfqq); |
7963 |
-+} |
7964 |
-+ |
7965 |
-+static void bfq_end_raising(struct bfq_data *bfqd) |
7966 |
-+{ |
7967 |
-+ struct bfq_queue *bfqq; |
7968 |
-+ |
7969 |
-+ spin_lock_irq(bfqd->queue->queue_lock); |
7970 |
-+ |
7971 |
-+ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) |
7972 |
-+ bfq_bfqq_end_raising(bfqq); |
7973 |
-+ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) |
7974 |
-+ bfq_bfqq_end_raising(bfqq); |
7975 |
-+ bfq_end_raising_async(bfqd); |
7976 |
-+ |
7977 |
-+ spin_unlock_irq(bfqd->queue->queue_lock); |
7978 |
-+} |
7979 |
-+ |
7980 |
-+static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
7981 |
-+ struct bio *bio) |
7982 |
-+{ |
7983 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
7984 |
-+ struct bfq_io_cq *bic; |
7985 |
-+ struct bfq_queue *bfqq; |
7986 |
-+ |
7987 |
-+ /* |
7988 |
-+ * Disallow merge of a sync bio into an async request. |
7989 |
-+ */ |
7990 |
-+ if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
7991 |
-+ return 0; |
7992 |
-+ |
7993 |
-+ /* |
7994 |
-+ * Lookup the bfqq that this bio will be queued with. Allow |
7995 |
-+ * merge only if rq is queued there. |
7996 |
-+ * Queue lock is held here. |
7997 |
-+ */ |
7998 |
-+ bic = bfq_bic_lookup(bfqd, current->io_context); |
7999 |
-+ if (bic == NULL) |
8000 |
-+ return 0; |
8001 |
-+ |
8002 |
-+ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
8003 |
-+ return bfqq == RQ_BFQQ(rq); |
8004 |
-+} |
8005 |
-+ |
8006 |
-+static void __bfq_set_active_queue(struct bfq_data *bfqd, |
8007 |
-+ struct bfq_queue *bfqq) |
8008 |
-+{ |
8009 |
-+ if (bfqq != NULL) { |
8010 |
-+ bfq_mark_bfqq_must_alloc(bfqq); |
8011 |
-+ bfq_mark_bfqq_budget_new(bfqq); |
8012 |
-+ bfq_clear_bfqq_fifo_expire(bfqq); |
8013 |
-+ |
8014 |
-+ bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
8015 |
-+ |
8016 |
-+ bfq_log_bfqq(bfqd, bfqq, "set_active_queue, cur-budget = %lu", |
8017 |
-+ bfqq->entity.budget); |
8018 |
-+ } |
8019 |
-+ |
8020 |
-+ bfqd->active_queue = bfqq; |
8021 |
-+} |
8022 |
-+ |
8023 |
-+/* |
8024 |
-+ * Get and set a new active queue for service. |
8025 |
-+ */ |
8026 |
-+static struct bfq_queue *bfq_set_active_queue(struct bfq_data *bfqd, |
8027 |
-+ struct bfq_queue *bfqq) |
8028 |
-+{ |
8029 |
-+ if (!bfqq) |
8030 |
-+ bfqq = bfq_get_next_queue(bfqd); |
8031 |
-+ else |
8032 |
-+ bfq_get_next_queue_forced(bfqd, bfqq); |
8033 |
-+ |
8034 |
-+ __bfq_set_active_queue(bfqd, bfqq); |
8035 |
-+ return bfqq; |
8036 |
-+} |
8037 |
-+ |
8038 |
-+static inline sector_t bfq_dist_from_last(struct bfq_data *bfqd, |
8039 |
-+ struct request *rq) |
8040 |
-+{ |
8041 |
-+ if (blk_rq_pos(rq) >= bfqd->last_position) |
8042 |
-+ return blk_rq_pos(rq) - bfqd->last_position; |
8043 |
-+ else |
8044 |
-+ return bfqd->last_position - blk_rq_pos(rq); |
8045 |
-+} |
8046 |
-+ |
8047 |
-+/* |
8048 |
-+ * Return true if bfqq has no request pending and rq is close enough to |
8049 |
-+ * bfqd->last_position, or if rq is closer to bfqd->last_position than |
8050 |
-+ * bfqq->next_rq |
8051 |
-+ */ |
8052 |
-+static inline int bfq_rq_close(struct bfq_data *bfqd, struct request *rq) |
8053 |
-+{ |
8054 |
-+ return bfq_dist_from_last(bfqd, rq) <= BFQQ_SEEK_THR; |
8055 |
-+} |
8056 |
-+ |
8057 |
-+static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
8058 |
-+{ |
8059 |
-+ struct rb_root *root = &bfqd->rq_pos_tree; |
8060 |
-+ struct rb_node *parent, *node; |
8061 |
-+ struct bfq_queue *__bfqq; |
8062 |
-+ sector_t sector = bfqd->last_position; |
8063 |
-+ |
8064 |
-+ if (RB_EMPTY_ROOT(root)) |
8065 |
-+ return NULL; |
8066 |
-+ |
8067 |
-+ /* |
8068 |
-+ * First, if we find a request starting at the end of the last |
8069 |
-+ * request, choose it. |
8070 |
-+ */ |
8071 |
-+ __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL); |
8072 |
-+ if (__bfqq != NULL) |
8073 |
-+ return __bfqq; |
8074 |
-+ |
8075 |
-+ /* |
8076 |
-+ * If the exact sector wasn't found, the parent of the NULL leaf |
8077 |
-+ * will contain the closest sector (rq_pos_tree sorted by next_request |
8078 |
-+ * position). |
8079 |
-+ */ |
8080 |
-+ __bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
8081 |
-+ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
8082 |
-+ return __bfqq; |
8083 |
-+ |
8084 |
-+ if (blk_rq_pos(__bfqq->next_rq) < sector) |
8085 |
-+ node = rb_next(&__bfqq->pos_node); |
8086 |
-+ else |
8087 |
-+ node = rb_prev(&__bfqq->pos_node); |
8088 |
-+ if (node == NULL) |
8089 |
-+ return NULL; |
8090 |
-+ |
8091 |
-+ __bfqq = rb_entry(node, struct bfq_queue, pos_node); |
8092 |
-+ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
8093 |
-+ return __bfqq; |
8094 |
-+ |
8095 |
-+ return NULL; |
8096 |
-+} |
8097 |
-+ |
8098 |
-+/* |
8099 |
-+ * bfqd - obvious |
8100 |
-+ * cur_bfqq - passed in so that we don't decide that the current queue |
8101 |
-+ * is closely cooperating with itself. |
8102 |
-+ * |
8103 |
-+ * We are assuming that cur_bfqq has dispatched at least one request, |
8104 |
-+ * and that bfqd->last_position reflects a position on the disk associated |
8105 |
-+ * with the I/O issued by cur_bfqq. |
8106 |
-+ */ |
8107 |
-+static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
8108 |
-+ struct bfq_queue *cur_bfqq) |
8109 |
-+{ |
8110 |
-+ struct bfq_queue *bfqq; |
8111 |
-+ |
8112 |
-+ if (bfq_class_idle(cur_bfqq)) |
8113 |
-+ return NULL; |
8114 |
-+ if (!bfq_bfqq_sync(cur_bfqq)) |
8115 |
-+ return NULL; |
8116 |
-+ if (BFQQ_SEEKY(cur_bfqq)) |
8117 |
-+ return NULL; |
8118 |
-+ |
8119 |
-+ /* If device has only one backlogged bfq_queue, don't search. */ |
8120 |
-+ if (bfqd->busy_queues == 1) |
8121 |
-+ return NULL; |
8122 |
-+ |
8123 |
-+ /* |
8124 |
-+ * We should notice if some of the queues are cooperating, e.g. |
8125 |
-+ * working closely on the same area of the disk. In that case, |
8126 |
-+ * we can group them together and don't waste time idling. |
8127 |
-+ */ |
8128 |
-+ bfqq = bfqq_close(bfqd); |
8129 |
-+ if (bfqq == NULL || bfqq == cur_bfqq) |
8130 |
-+ return NULL; |
8131 |
-+ |
8132 |
-+ /* |
8133 |
-+ * Do not merge queues from different bfq_groups. |
8134 |
-+ */ |
8135 |
-+ if (bfqq->entity.parent != cur_bfqq->entity.parent) |
8136 |
-+ return NULL; |
8137 |
-+ |
8138 |
-+ /* |
8139 |
-+ * It only makes sense to merge sync queues. |
8140 |
-+ */ |
8141 |
-+ if (!bfq_bfqq_sync(bfqq)) |
8142 |
-+ return NULL; |
8143 |
-+ if (BFQQ_SEEKY(bfqq)) |
8144 |
-+ return NULL; |
8145 |
-+ |
8146 |
-+ /* |
8147 |
-+ * Do not merge queues of different priority classes. |
8148 |
-+ */ |
8149 |
-+ if (bfq_class_rt(bfqq) != bfq_class_rt(cur_bfqq)) |
8150 |
-+ return NULL; |
8151 |
-+ |
8152 |
-+ return bfqq; |
8153 |
-+} |
8154 |
-+ |
8155 |
-+/* |
8156 |
-+ * If enough samples have been computed, return the current max budget |
8157 |
-+ * stored in bfqd, which is dynamically updated according to the |
8158 |
-+ * estimated disk peak rate; otherwise return the default max budget |
8159 |
-+ */ |
8160 |
-+static inline unsigned long bfq_max_budget(struct bfq_data *bfqd) |
8161 |
-+{ |
8162 |
-+ if (bfqd->budgets_assigned < 194) |
8163 |
-+ return bfq_default_max_budget; |
8164 |
-+ else |
8165 |
-+ return bfqd->bfq_max_budget; |
8166 |
-+} |
8167 |
-+ |
8168 |
-+/* |
8169 |
-+ * Return min budget, which is a fraction of the current or default |
8170 |
-+ * max budget (trying with 1/32) |
8171 |
-+ */ |
8172 |
-+static inline unsigned long bfq_min_budget(struct bfq_data *bfqd) |
8173 |
-+{ |
8174 |
-+ if (bfqd->budgets_assigned < 194) |
8175 |
-+ return bfq_default_max_budget / 32; |
8176 |
-+ else |
8177 |
-+ return bfqd->bfq_max_budget / 32; |
8178 |
-+} |
8179 |
-+ |
8180 |
-+/* |
8181 |
-+ * Decides whether idling should be done for given device and |
8182 |
-+ * given active queue. |
8183 |
-+ */ |
8184 |
-+static inline bool bfq_queue_nonrot_noidle(struct bfq_data *bfqd, |
8185 |
-+ struct bfq_queue *active_bfqq) |
8186 |
-+{ |
8187 |
-+ if (active_bfqq == NULL) |
8188 |
-+ return false; |
8189 |
-+ /* |
8190 |
-+ * If device is SSD it has no seek penalty, disable idling; but |
8191 |
-+ * do so only if: |
8192 |
-+ * - device does not support queuing, otherwise we still have |
8193 |
-+ * a problem with sync vs async workloads; |
8194 |
-+ * - the queue is not weight-raised, to preserve guarantees. |
8195 |
-+ */ |
8196 |
-+ return (blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag && |
8197 |
-+ active_bfqq->raising_coeff == 1); |
8198 |
-+} |
8199 |
-+ |
8200 |
-+static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
8201 |
-+{ |
8202 |
-+ struct bfq_queue *bfqq = bfqd->active_queue; |
8203 |
-+ struct bfq_io_cq *bic; |
8204 |
-+ unsigned long sl; |
8205 |
-+ |
8206 |
-+ WARN_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
8207 |
-+ |
8208 |
-+ /* Tasks have exited, don't wait. */ |
8209 |
-+ bic = bfqd->active_bic; |
8210 |
-+ if (bic == NULL || atomic_read(&bic->icq.ioc->active_ref) == 0) |
8211 |
-+ return; |
8212 |
-+ |
8213 |
-+ bfq_mark_bfqq_wait_request(bfqq); |
8214 |
-+ |
8215 |
-+ /* |
8216 |
-+ * We don't want to idle for seeks, but we do want to allow |
8217 |
-+ * fair distribution of slice time for a process doing back-to-back |
8218 |
-+ * seeks. So allow a little bit of time for him to submit a new rq. |
8219 |
-+ * |
8220 |
-+ * To prevent processes with (partly) seeky workloads from |
8221 |
-+ * being too ill-treated, grant them a small fraction of the |
8222 |
-+ * assigned budget before reducing the waiting time to |
8223 |
-+ * BFQ_MIN_TT. This happened to help reduce latency. |
8224 |
-+ */ |
8225 |
-+ sl = bfqd->bfq_slice_idle; |
8226 |
-+ if (bfq_sample_valid(bfqq->seek_samples) && BFQQ_SEEKY(bfqq) && |
8227 |
-+ bfqq->entity.service > bfq_max_budget(bfqd) / 8 && |
8228 |
-+ bfqq->raising_coeff == 1) |
8229 |
-+ sl = min(sl, msecs_to_jiffies(BFQ_MIN_TT)); |
8230 |
-+ else if (bfqq->raising_coeff > 1) |
8231 |
-+ sl = sl * 3; |
8232 |
-+ bfqd->last_idling_start = ktime_get(); |
8233 |
-+ mod_timer(&bfqd->idle_slice_timer, jiffies + sl); |
8234 |
-+ bfq_log(bfqd, "arm idle: %u/%u ms", |
8235 |
-+ jiffies_to_msecs(sl), jiffies_to_msecs(bfqd->bfq_slice_idle)); |
8236 |
-+} |
8237 |
-+ |
8238 |
-+/* |
8239 |
-+ * Set the maximum time for the active queue to consume its |
8240 |
-+ * budget. This prevents seeky processes from lowering the disk |
8241 |
-+ * throughput (always guaranteed with a time slice scheme as in CFQ). |
8242 |
-+ */ |
8243 |
-+static void bfq_set_budget_timeout(struct bfq_data *bfqd) |
8244 |
-+{ |
8245 |
-+ struct bfq_queue *bfqq = bfqd->active_queue; |
8246 |
-+ unsigned int timeout_coeff; |
8247 |
-+ if (bfqq->raising_cur_max_time == bfqd->bfq_raising_rt_max_time) |
8248 |
-+ timeout_coeff = 1; |
8249 |
-+ else |
8250 |
-+ timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; |
8251 |
-+ |
8252 |
-+ bfqd->last_budget_start = ktime_get(); |
8253 |
-+ |
8254 |
-+ bfq_clear_bfqq_budget_new(bfqq); |
8255 |
-+ bfqq->budget_timeout = jiffies + |
8256 |
-+ bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * timeout_coeff; |
8257 |
-+ |
8258 |
-+ bfq_log_bfqq(bfqd, bfqq, "set budget_timeout %u", |
8259 |
-+ jiffies_to_msecs(bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * |
8260 |
-+ timeout_coeff)); |
8261 |
-+} |
8262 |
-+ |
8263 |
-+/* |
8264 |
-+ * Move request from internal lists to the request queue dispatch list. |
8265 |
-+ */ |
8266 |
-+static void bfq_dispatch_insert(struct request_queue *q, struct request *rq) |
8267 |
-+{ |
8268 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
8269 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
8270 |
-+ |
8271 |
-+ bfq_remove_request(rq); |
8272 |
-+ bfqq->dispatched++; |
8273 |
-+ elv_dispatch_sort(q, rq); |
8274 |
-+ |
8275 |
-+ if (bfq_bfqq_sync(bfqq)) |
8276 |
-+ bfqd->sync_flight++; |
8277 |
-+} |
8278 |
-+ |
8279 |
-+/* |
8280 |
-+ * Return expired entry, or NULL to just start from scratch in rbtree. |
8281 |
-+ */ |
8282 |
-+static struct request *bfq_check_fifo(struct bfq_queue *bfqq) |
8283 |
-+{ |
8284 |
-+ struct request *rq = NULL; |
8285 |
-+ |
8286 |
-+ if (bfq_bfqq_fifo_expire(bfqq)) |
8287 |
-+ return NULL; |
8288 |
-+ |
8289 |
-+ bfq_mark_bfqq_fifo_expire(bfqq); |
8290 |
-+ |
8291 |
-+ if (list_empty(&bfqq->fifo)) |
8292 |
-+ return NULL; |
8293 |
-+ |
8294 |
-+ rq = rq_entry_fifo(bfqq->fifo.next); |
8295 |
-+ |
8296 |
-+ if (time_before(jiffies, rq_fifo_time(rq))) |
8297 |
-+ return NULL; |
8298 |
-+ |
8299 |
-+ return rq; |
8300 |
-+} |
8301 |
-+ |
8302 |
-+/* |
8303 |
-+ * Must be called with the queue_lock held. |
8304 |
-+ */ |
8305 |
-+static int bfqq_process_refs(struct bfq_queue *bfqq) |
8306 |
-+{ |
8307 |
-+ int process_refs, io_refs; |
8308 |
-+ |
8309 |
-+ io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
8310 |
-+ process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
8311 |
-+ BUG_ON(process_refs < 0); |
8312 |
-+ return process_refs; |
8313 |
-+} |
8314 |
-+ |
8315 |
-+static void bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
8316 |
-+{ |
8317 |
-+ int process_refs, new_process_refs; |
8318 |
-+ struct bfq_queue *__bfqq; |
8319 |
-+ |
8320 |
-+ /* |
8321 |
-+ * If there are no process references on the new_bfqq, then it is |
8322 |
-+ * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
8323 |
-+ * may have dropped their last reference (not just their last process |
8324 |
-+ * reference). |
8325 |
-+ */ |
8326 |
-+ if (!bfqq_process_refs(new_bfqq)) |
8327 |
-+ return; |
8328 |
-+ |
8329 |
-+ /* Avoid a circular list and skip interim queue merges. */ |
8330 |
-+ while ((__bfqq = new_bfqq->new_bfqq)) { |
8331 |
-+ if (__bfqq == bfqq) |
8332 |
-+ return; |
8333 |
-+ new_bfqq = __bfqq; |
8334 |
-+ } |
8335 |
-+ |
8336 |
-+ process_refs = bfqq_process_refs(bfqq); |
8337 |
-+ new_process_refs = bfqq_process_refs(new_bfqq); |
8338 |
-+ /* |
8339 |
-+ * If the process for the bfqq has gone away, there is no |
8340 |
-+ * sense in merging the queues. |
8341 |
-+ */ |
8342 |
-+ if (process_refs == 0 || new_process_refs == 0) |
8343 |
-+ return; |
8344 |
-+ |
8345 |
-+ /* |
8346 |
-+ * Merge in the direction of the lesser amount of work. |
8347 |
-+ */ |
8348 |
-+ if (new_process_refs >= process_refs) { |
8349 |
-+ bfqq->new_bfqq = new_bfqq; |
8350 |
-+ atomic_add(process_refs, &new_bfqq->ref); |
8351 |
-+ } else { |
8352 |
-+ new_bfqq->new_bfqq = bfqq; |
8353 |
-+ atomic_add(new_process_refs, &bfqq->ref); |
8354 |
-+ } |
8355 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
8356 |
-+ new_bfqq->pid); |
8357 |
-+} |
8358 |
-+ |
8359 |
-+static inline unsigned long bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
8360 |
-+{ |
8361 |
-+ struct bfq_entity *entity = &bfqq->entity; |
8362 |
-+ return entity->budget - entity->service; |
8363 |
-+} |
8364 |
-+ |
8365 |
-+static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
8366 |
-+{ |
8367 |
-+ BUG_ON(bfqq != bfqd->active_queue); |
8368 |
-+ |
8369 |
-+ __bfq_bfqd_reset_active(bfqd); |
8370 |
-+ |
8371 |
-+ /* |
8372 |
-+ * If this bfqq is shared between multiple processes, check |
8373 |
-+ * to make sure that those processes are still issuing I/Os |
8374 |
-+ * within the mean seek distance. If not, it may be time to |
8375 |
-+ * break the queues apart again. |
8376 |
-+ */ |
8377 |
-+ if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq)) |
8378 |
-+ bfq_mark_bfqq_split_coop(bfqq); |
8379 |
-+ |
8380 |
-+ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
8381 |
-+ /* |
8382 |
-+ * overloading budget_timeout field to store when |
8383 |
-+ * the queue remains with no backlog, used by |
8384 |
-+ * the weight-raising mechanism |
8385 |
-+ */ |
8386 |
-+ bfqq->budget_timeout = jiffies ; |
8387 |
-+ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
8388 |
-+ } else { |
8389 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
8390 |
-+ /* |
8391 |
-+ * Resort priority tree of potential close cooperators. |
8392 |
-+ */ |
8393 |
-+ bfq_rq_pos_tree_add(bfqd, bfqq); |
8394 |
-+ } |
8395 |
-+} |
8396 |
-+ |
8397 |
-+/** |
8398 |
-+ * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. |
8399 |
-+ * @bfqd: device data. |
8400 |
-+ * @bfqq: queue to update. |
8401 |
-+ * @reason: reason for expiration. |
8402 |
-+ * |
8403 |
-+ * Handle the feedback on @bfqq budget. See the body for detailed |
8404 |
-+ * comments. |
8405 |
-+ */ |
8406 |
-+static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, |
8407 |
-+ struct bfq_queue *bfqq, |
8408 |
-+ enum bfqq_expiration reason) |
8409 |
-+{ |
8410 |
-+ struct request *next_rq; |
8411 |
-+ unsigned long budget, min_budget; |
8412 |
-+ |
8413 |
-+ budget = bfqq->max_budget; |
8414 |
-+ min_budget = bfq_min_budget(bfqd); |
8415 |
-+ |
8416 |
-+ BUG_ON(bfqq != bfqd->active_queue); |
8417 |
-+ |
8418 |
-+ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %lu, budg left %lu", |
8419 |
-+ bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); |
8420 |
-+ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %lu, min budg %lu", |
8421 |
-+ budget, bfq_min_budget(bfqd)); |
8422 |
-+ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", |
8423 |
-+ bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->active_queue)); |
8424 |
-+ |
8425 |
-+ if (bfq_bfqq_sync(bfqq)) { |
8426 |
-+ switch (reason) { |
8427 |
-+ /* |
8428 |
-+ * Caveat: in all the following cases we trade latency |
8429 |
-+ * for throughput. |
8430 |
-+ */ |
8431 |
-+ case BFQ_BFQQ_TOO_IDLE: |
8432 |
-+ /* |
8433 |
-+ * This is the only case where we may reduce |
8434 |
-+ * the budget: if there is no requets of the |
8435 |
-+ * process still waiting for completion, then |
8436 |
-+ * we assume (tentatively) that the timer has |
8437 |
-+ * expired because the batch of requests of |
8438 |
-+ * the process could have been served with a |
8439 |
-+ * smaller budget. Hence, betting that |
8440 |
-+ * process will behave in the same way when it |
8441 |
-+ * becomes backlogged again, we reduce its |
8442 |
-+ * next budget. As long as we guess right, |
8443 |
-+ * this budget cut reduces the latency |
8444 |
-+ * experienced by the process. |
8445 |
-+ * |
8446 |
-+ * However, if there are still outstanding |
8447 |
-+ * requests, then the process may have not yet |
8448 |
-+ * issued its next request just because it is |
8449 |
-+ * still waiting for the completion of some of |
8450 |
-+ * the still oustanding ones. So in this |
8451 |
-+ * subcase we do not reduce its budget, on the |
8452 |
-+ * contrary we increase it to possibly boost |
8453 |
-+ * the throughput, as discussed in the |
8454 |
-+ * comments to the BUDGET_TIMEOUT case. |
8455 |
-+ */ |
8456 |
-+ if (bfqq->dispatched > 0) /* still oustanding reqs */ |
8457 |
-+ budget = min(budget * 2, bfqd->bfq_max_budget); |
8458 |
-+ else { |
8459 |
-+ if (budget > 5 * min_budget) |
8460 |
-+ budget -= 4 * min_budget; |
8461 |
-+ else |
8462 |
-+ budget = min_budget; |
8463 |
-+ } |
8464 |
-+ break; |
8465 |
-+ case BFQ_BFQQ_BUDGET_TIMEOUT: |
8466 |
-+ /* |
8467 |
-+ * We double the budget here because: 1) it |
8468 |
-+ * gives the chance to boost the throughput if |
8469 |
-+ * this is not a seeky process (which may have |
8470 |
-+ * bumped into this timeout because of, e.g., |
8471 |
-+ * ZBR), 2) together with charge_full_budget |
8472 |
-+ * it helps give seeky processes higher |
8473 |
-+ * timestamps, and hence be served less |
8474 |
-+ * frequently. |
8475 |
-+ */ |
8476 |
-+ budget = min(budget * 2, bfqd->bfq_max_budget); |
8477 |
-+ break; |
8478 |
-+ case BFQ_BFQQ_BUDGET_EXHAUSTED: |
8479 |
-+ /* |
8480 |
-+ * The process still has backlog, and did not |
8481 |
-+ * let either the budget timeout or the disk |
8482 |
-+ * idling timeout expire. Hence it is not |
8483 |
-+ * seeky, has a short thinktime and may be |
8484 |
-+ * happy with a higher budget too. So |
8485 |
-+ * definitely increase the budget of this good |
8486 |
-+ * candidate to boost the disk throughput. |
8487 |
-+ */ |
8488 |
-+ budget = min(budget * 4, bfqd->bfq_max_budget); |
8489 |
-+ break; |
8490 |
-+ case BFQ_BFQQ_NO_MORE_REQUESTS: |
8491 |
-+ /* |
8492 |
-+ * Leave the budget unchanged. |
8493 |
-+ */ |
8494 |
-+ default: |
8495 |
-+ return; |
8496 |
-+ } |
8497 |
-+ } else /* async queue */ |
8498 |
-+ /* async queues get always the maximum possible budget |
8499 |
-+ * (their ability to dispatch is limited by |
8500 |
-+ * @bfqd->bfq_max_budget_async_rq). |
8501 |
-+ */ |
8502 |
-+ budget = bfqd->bfq_max_budget; |
8503 |
-+ |
8504 |
-+ bfqq->max_budget = budget; |
8505 |
-+ |
8506 |
-+ if (bfqd->budgets_assigned >= 194 && bfqd->bfq_user_max_budget == 0 && |
8507 |
-+ bfqq->max_budget > bfqd->bfq_max_budget) |
8508 |
-+ bfqq->max_budget = bfqd->bfq_max_budget; |
8509 |
-+ |
8510 |
-+ /* |
8511 |
-+ * Make sure that we have enough budget for the next request. |
8512 |
-+ * Since the finish time of the bfqq must be kept in sync with |
8513 |
-+ * the budget, be sure to call __bfq_bfqq_expire() after the |
8514 |
-+ * update. |
8515 |
-+ */ |
8516 |
-+ next_rq = bfqq->next_rq; |
8517 |
-+ if (next_rq != NULL) |
8518 |
-+ bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, |
8519 |
-+ bfq_serv_to_charge(next_rq, bfqq)); |
8520 |
-+ else |
8521 |
-+ bfqq->entity.budget = bfqq->max_budget; |
8522 |
-+ |
8523 |
-+ bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %lu", |
8524 |
-+ next_rq != NULL ? blk_rq_sectors(next_rq) : 0, |
8525 |
-+ bfqq->entity.budget); |
8526 |
-+} |
8527 |
-+ |
8528 |
-+static unsigned long bfq_calc_max_budget(u64 peak_rate, u64 timeout) |
8529 |
-+{ |
8530 |
-+ unsigned long max_budget; |
8531 |
-+ |
8532 |
-+ /* |
8533 |
-+ * The max_budget calculated when autotuning is equal to the |
8534 |
-+ * amount of sectors transfered in timeout_sync at the |
8535 |
-+ * estimated peak rate. |
8536 |
-+ */ |
8537 |
-+ max_budget = (unsigned long)(peak_rate * 1000 * |
8538 |
-+ timeout >> BFQ_RATE_SHIFT); |
8539 |
-+ |
8540 |
-+ return max_budget; |
8541 |
-+} |
8542 |
-+ |
8543 |
-+/* |
8544 |
-+ * In addition to updating the peak rate, checks whether the process |
8545 |
-+ * is "slow", and returns 1 if so. This slow flag is used, in addition |
8546 |
-+ * to the budget timeout, to reduce the amount of service provided to |
8547 |
-+ * seeky processes, and hence reduce their chances to lower the |
8548 |
-+ * throughput. See the code for more details. |
8549 |
-+ */ |
8550 |
-+static int bfq_update_peak_rate(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
8551 |
-+ int compensate, enum bfqq_expiration reason) |
8552 |
-+{ |
8553 |
-+ u64 bw, usecs, expected, timeout; |
8554 |
-+ ktime_t delta; |
8555 |
-+ int update = 0; |
8556 |
-+ |
8557 |
-+ if (!bfq_bfqq_sync(bfqq) || bfq_bfqq_budget_new(bfqq)) |
8558 |
-+ return 0; |
8559 |
-+ |
8560 |
-+ if (compensate) |
8561 |
-+ delta = bfqd->last_idling_start; |
8562 |
-+ else |
8563 |
-+ delta = ktime_get(); |
8564 |
-+ delta = ktime_sub(delta, bfqd->last_budget_start); |
8565 |
-+ usecs = ktime_to_us(delta); |
8566 |
-+ |
8567 |
-+ /* Don't trust short/unrealistic values. */ |
8568 |
-+ if (usecs < 100 || usecs >= LONG_MAX) |
8569 |
-+ return 0; |
8570 |
-+ |
8571 |
-+ /* |
8572 |
-+ * Calculate the bandwidth for the last slice. We use a 64 bit |
8573 |
-+ * value to store the peak rate, in sectors per usec in fixed |
8574 |
-+ * point math. We do so to have enough precision in the estimate |
8575 |
-+ * and to avoid overflows. |
8576 |
-+ */ |
8577 |
-+ bw = (u64)bfqq->entity.service << BFQ_RATE_SHIFT; |
8578 |
-+ do_div(bw, (unsigned long)usecs); |
8579 |
-+ |
8580 |
-+ timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
8581 |
-+ |
8582 |
-+ /* |
8583 |
-+ * Use only long (> 20ms) intervals to filter out spikes for |
8584 |
-+ * the peak rate estimation. |
8585 |
-+ */ |
8586 |
-+ if (usecs > 20000) { |
8587 |
-+ if (bw > bfqd->peak_rate || |
8588 |
-+ (!BFQQ_SEEKY(bfqq) && |
8589 |
-+ reason == BFQ_BFQQ_BUDGET_TIMEOUT)) { |
8590 |
-+ bfq_log(bfqd, "measured bw =%llu", bw); |
8591 |
-+ /* |
8592 |
-+ * To smooth oscillations use a low-pass filter with |
8593 |
-+ * alpha=7/8, i.e., |
8594 |
-+ * new_rate = (7/8) * old_rate + (1/8) * bw |
8595 |
-+ */ |
8596 |
-+ do_div(bw, 8); |
8597 |
-+ if (bw == 0) |
8598 |
-+ return 0; |
8599 |
-+ bfqd->peak_rate *= 7; |
8600 |
-+ do_div(bfqd->peak_rate, 8); |
8601 |
-+ bfqd->peak_rate += bw; |
8602 |
-+ update = 1; |
8603 |
-+ bfq_log(bfqd, "new peak_rate=%llu", bfqd->peak_rate); |
8604 |
-+ } |
8605 |
-+ |
8606 |
-+ update |= bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES - 1; |
8607 |
-+ |
8608 |
-+ if (bfqd->peak_rate_samples < BFQ_PEAK_RATE_SAMPLES) |
8609 |
-+ bfqd->peak_rate_samples++; |
8610 |
-+ |
8611 |
-+ if (bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES && |
8612 |
-+ update && bfqd->bfq_user_max_budget == 0) { |
8613 |
-+ bfqd->bfq_max_budget = |
8614 |
-+ bfq_calc_max_budget(bfqd->peak_rate, timeout); |
8615 |
-+ bfq_log(bfqd, "new max_budget=%lu", |
8616 |
-+ bfqd->bfq_max_budget); |
8617 |
-+ } |
8618 |
-+ } |
8619 |
-+ |
8620 |
-+ /* |
8621 |
-+ * If the process has been served for a too short time |
8622 |
-+ * interval to let its possible sequential accesses prevail on |
8623 |
-+ * the initial seek time needed to move the disk head on the |
8624 |
-+ * first sector it requested, then give the process a chance |
8625 |
-+ * and for the moment return false. |
8626 |
-+ */ |
8627 |
-+ if (bfqq->entity.budget <= bfq_max_budget(bfqd) / 8) |
8628 |
-+ return 0; |
8629 |
-+ |
8630 |
-+ /* |
8631 |
-+ * A process is considered ``slow'' (i.e., seeky, so that we |
8632 |
-+ * cannot treat it fairly in the service domain, as it would |
8633 |
-+ * slow down too much the other processes) if, when a slice |
8634 |
-+ * ends for whatever reason, it has received service at a |
8635 |
-+ * rate that would not be high enough to complete the budget |
8636 |
-+ * before the budget timeout expiration. |
8637 |
-+ */ |
8638 |
-+ expected = bw * 1000 * timeout >> BFQ_RATE_SHIFT; |
8639 |
-+ |
8640 |
-+ /* |
8641 |
-+ * Caveat: processes doing IO in the slower disk zones will |
8642 |
-+ * tend to be slow(er) even if not seeky. And the estimated |
8643 |
-+ * peak rate will actually be an average over the disk |
8644 |
-+ * surface. Hence, to not be too harsh with unlucky processes, |
8645 |
-+ * we keep a budget/3 margin of safety before declaring a |
8646 |
-+ * process slow. |
8647 |
-+ */ |
8648 |
-+ return expected > (4 * bfqq->entity.budget) / 3; |
8649 |
-+} |
8650 |
-+ |
8651 |
-+/** |
8652 |
-+ * bfq_bfqq_expire - expire a queue. |
8653 |
-+ * @bfqd: device owning the queue. |
8654 |
-+ * @bfqq: the queue to expire. |
8655 |
-+ * @compensate: if true, compensate for the time spent idling. |
8656 |
-+ * @reason: the reason causing the expiration. |
8657 |
-+ * |
8658 |
-+ * |
8659 |
-+ * If the process associated to the queue is slow (i.e., seeky), or in |
8660 |
-+ * case of budget timeout, or, finally, if it is async, we |
8661 |
-+ * artificially charge it an entire budget (independently of the |
8662 |
-+ * actual service it received). As a consequence, the queue will get |
8663 |
-+ * higher timestamps than the correct ones upon reactivation, and |
8664 |
-+ * hence it will be rescheduled as if it had received more service |
8665 |
-+ * than what it actually received. In the end, this class of processes |
8666 |
-+ * will receive less service in proportion to how slowly they consume |
8667 |
-+ * their budgets (and hence how seriously they tend to lower the |
8668 |
-+ * throughput). |
8669 |
-+ * |
8670 |
-+ * In contrast, when a queue expires because it has been idling for |
8671 |
-+ * too much or because it exhausted its budget, we do not touch the |
8672 |
-+ * amount of service it has received. Hence when the queue will be |
8673 |
-+ * reactivated and its timestamps updated, the latter will be in sync |
8674 |
-+ * with the actual service received by the queue until expiration. |
8675 |
-+ * |
8676 |
-+ * Charging a full budget to the first type of queues and the exact |
8677 |
-+ * service to the others has the effect of using the WF2Q+ policy to |
8678 |
-+ * schedule the former on a timeslice basis, without violating the |
8679 |
-+ * service domain guarantees of the latter. |
8680 |
-+ */ |
8681 |
-+static void bfq_bfqq_expire(struct bfq_data *bfqd, |
8682 |
-+ struct bfq_queue *bfqq, |
8683 |
-+ int compensate, |
8684 |
-+ enum bfqq_expiration reason) |
8685 |
-+{ |
8686 |
-+ int slow; |
8687 |
-+ BUG_ON(bfqq != bfqd->active_queue); |
8688 |
-+ |
8689 |
-+ /* Update disk peak rate for autotuning and check whether the |
8690 |
-+ * process is slow (see bfq_update_peak_rate). |
8691 |
-+ */ |
8692 |
-+ slow = bfq_update_peak_rate(bfqd, bfqq, compensate, reason); |
8693 |
-+ |
8694 |
-+ /* |
8695 |
-+ * As above explained, 'punish' slow (i.e., seeky), timed-out |
8696 |
-+ * and async queues, to favor sequential sync workloads. |
8697 |
-+ * |
8698 |
-+ * Processes doing IO in the slower disk zones will tend to be |
8699 |
-+ * slow(er) even if not seeky. Hence, since the estimated peak |
8700 |
-+ * rate is actually an average over the disk surface, these |
8701 |
-+ * processes may timeout just for bad luck. To avoid punishing |
8702 |
-+ * them we do not charge a full budget to a process that |
8703 |
-+ * succeeded in consuming at least 2/3 of its budget. |
8704 |
-+ */ |
8705 |
-+ if (slow || (reason == BFQ_BFQQ_BUDGET_TIMEOUT && |
8706 |
-+ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)) |
8707 |
-+ bfq_bfqq_charge_full_budget(bfqq); |
8708 |
-+ |
8709 |
-+ if (bfqd->low_latency && bfqq->raising_coeff == 1) |
8710 |
-+ bfqq->last_rais_start_finish = jiffies; |
8711 |
-+ |
8712 |
-+ if (bfqd->low_latency && bfqd->bfq_raising_max_softrt_rate > 0) { |
8713 |
-+ if(reason != BFQ_BFQQ_BUDGET_TIMEOUT) |
8714 |
-+ bfqq->soft_rt_next_start = |
8715 |
-+ jiffies + |
8716 |
-+ HZ * bfqq->entity.service / |
8717 |
-+ bfqd->bfq_raising_max_softrt_rate; |
8718 |
-+ else |
8719 |
-+ bfqq->soft_rt_next_start = -1; /* infinity */ |
8720 |
-+ } |
8721 |
-+ bfq_log_bfqq(bfqd, bfqq, |
8722 |
-+ "expire (%d, slow %d, num_disp %d, idle_win %d)", reason, slow, |
8723 |
-+ bfqq->dispatched, bfq_bfqq_idle_window(bfqq)); |
8724 |
-+ |
8725 |
-+ /* Increase, decrease or leave budget unchanged according to reason */ |
8726 |
-+ __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); |
8727 |
-+ __bfq_bfqq_expire(bfqd, bfqq); |
8728 |
-+} |
8729 |
-+ |
8730 |
-+/* |
8731 |
-+ * Budget timeout is not implemented through a dedicated timer, but |
8732 |
-+ * just checked on request arrivals and completions, as well as on |
8733 |
-+ * idle timer expirations. |
8734 |
-+ */ |
8735 |
-+static int bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) |
8736 |
-+{ |
8737 |
-+ if (bfq_bfqq_budget_new(bfqq)) |
8738 |
-+ return 0; |
8739 |
-+ |
8740 |
-+ if (time_before(jiffies, bfqq->budget_timeout)) |
8741 |
-+ return 0; |
8742 |
-+ |
8743 |
-+ return 1; |
8744 |
-+} |
8745 |
-+ |
8746 |
-+/* |
8747 |
-+ * If we expire a queue that is waiting for the arrival of a new |
8748 |
-+ * request, we may prevent the fictitious timestamp backshifting that |
8749 |
-+ * allows the guarantees of the queue to be preserved (see [1] for |
8750 |
-+ * this tricky aspect). Hence we return true only if this condition |
8751 |
-+ * does not hold, or if the queue is slow enough to deserve only to be |
8752 |
-+ * kicked off for preserving a high throughput. |
8753 |
-+*/ |
8754 |
-+static inline int bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) |
8755 |
-+{ |
8756 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, |
8757 |
-+ "may_budget_timeout: wr %d left %d timeout %d", |
8758 |
-+ bfq_bfqq_wait_request(bfqq), |
8759 |
-+ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, |
8760 |
-+ bfq_bfqq_budget_timeout(bfqq)); |
8761 |
-+ |
8762 |
-+ return (!bfq_bfqq_wait_request(bfqq) || |
8763 |
-+ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) |
8764 |
-+ && |
8765 |
-+ bfq_bfqq_budget_timeout(bfqq); |
8766 |
-+} |
8767 |
-+ |
8768 |
-+/* |
8769 |
-+ * If the active queue is empty, but it is sync and either of the following |
8770 |
-+ * conditions holds, then: 1) the queue must remain active and cannot be |
8771 |
-+ * expired, and 2) the disk must be idled to wait for the possible arrival |
8772 |
-+ * of a new request for the queue. The conditions are: |
8773 |
-+ * - the device is rotational and not performing NCQ, and the queue has its |
8774 |
-+ * idle window set (in this case, waiting for a new request for the queue |
8775 |
-+ * is likely to boost the disk throughput); |
8776 |
-+ * - the queue is weight-raised (waiting for the request is necessary for |
8777 |
-+ * providing the queue with fairness and latency guarantees). |
8778 |
-+ */ |
8779 |
-+static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq, |
8780 |
-+ int budg_timeout) |
8781 |
-+{ |
8782 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
8783 |
-+ |
8784 |
-+ return (bfq_bfqq_sync(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) && |
8785 |
-+ bfqd->bfq_slice_idle != 0 && |
8786 |
-+ ((bfq_bfqq_idle_window(bfqq) && !bfqd->hw_tag && |
8787 |
-+ !blk_queue_nonrot(bfqd->queue)) |
8788 |
-+ || bfqq->raising_coeff > 1) && |
8789 |
-+ (bfqd->rq_in_driver == 0 || |
8790 |
-+ budg_timeout || |
8791 |
-+ bfqq->raising_coeff > 1) && |
8792 |
-+ !bfq_close_cooperator(bfqd, bfqq) && |
8793 |
-+ (!bfq_bfqq_coop(bfqq) || |
8794 |
-+ !bfq_bfqq_some_coop_idle(bfqq)) && |
8795 |
-+ !bfq_queue_nonrot_noidle(bfqd, bfqq)); |
8796 |
-+} |
8797 |
-+ |
8798 |
-+/* |
8799 |
-+ * Select a queue for service. If we have a current active queue, |
8800 |
-+ * check whether to continue servicing it, or retrieve and set a new one. |
8801 |
-+ */ |
8802 |
-+static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
8803 |
-+{ |
8804 |
-+ struct bfq_queue *bfqq, *new_bfqq = NULL; |
8805 |
-+ struct request *next_rq; |
8806 |
-+ enum bfqq_expiration reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
8807 |
-+ int budg_timeout; |
8808 |
-+ |
8809 |
-+ bfqq = bfqd->active_queue; |
8810 |
-+ if (bfqq == NULL) |
8811 |
-+ goto new_queue; |
8812 |
-+ |
8813 |
-+ bfq_log_bfqq(bfqd, bfqq, "select_queue: already active queue"); |
8814 |
-+ |
8815 |
-+ /* |
8816 |
-+ * If another queue has a request waiting within our mean seek |
8817 |
-+ * distance, let it run. The expire code will check for close |
8818 |
-+ * cooperators and put the close queue at the front of the |
8819 |
-+ * service tree. If possible, merge the expiring queue with the |
8820 |
-+ * new bfqq. |
8821 |
-+ */ |
8822 |
-+ new_bfqq = bfq_close_cooperator(bfqd, bfqq); |
8823 |
-+ if (new_bfqq != NULL && bfqq->new_bfqq == NULL) |
8824 |
-+ bfq_setup_merge(bfqq, new_bfqq); |
8825 |
-+ |
8826 |
-+ budg_timeout = bfq_may_expire_for_budg_timeout(bfqq); |
8827 |
-+ if (budg_timeout && |
8828 |
-+ !bfq_bfqq_must_idle(bfqq, budg_timeout)) |
8829 |
-+ goto expire; |
8830 |
-+ |
8831 |
-+ next_rq = bfqq->next_rq; |
8832 |
-+ /* |
8833 |
-+ * If bfqq has requests queued and it has enough budget left to |
8834 |
-+ * serve them, keep the queue, otherwise expire it. |
8835 |
-+ */ |
8836 |
-+ if (next_rq != NULL) { |
8837 |
-+ if (bfq_serv_to_charge(next_rq, bfqq) > |
8838 |
-+ bfq_bfqq_budget_left(bfqq)) { |
8839 |
-+ reason = BFQ_BFQQ_BUDGET_EXHAUSTED; |
8840 |
-+ goto expire; |
8841 |
-+ } else { |
8842 |
-+ /* |
8843 |
-+ * The idle timer may be pending because we may not |
8844 |
-+ * disable disk idling even when a new request arrives |
8845 |
-+ */ |
8846 |
-+ if (timer_pending(&bfqd->idle_slice_timer)) { |
8847 |
-+ /* |
8848 |
-+ * If we get here: 1) at least a new request |
8849 |
-+ * has arrived but we have not disabled the |
8850 |
-+ * timer because the request was too small, |
8851 |
-+ * 2) then the block layer has unplugged the |
8852 |
-+ * device, causing the dispatch to be invoked. |
8853 |
-+ * |
8854 |
-+ * Since the device is unplugged, now the |
8855 |
-+ * requests are probably large enough to |
8856 |
-+ * provide a reasonable throughput. |
8857 |
-+ * So we disable idling. |
8858 |
-+ */ |
8859 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
8860 |
-+ del_timer(&bfqd->idle_slice_timer); |
8861 |
-+ } |
8862 |
-+ if (new_bfqq == NULL) |
8863 |
-+ goto keep_queue; |
8864 |
-+ else |
8865 |
-+ goto expire; |
8866 |
-+ } |
8867 |
-+ } |
8868 |
-+ |
8869 |
-+ /* |
8870 |
-+ * No requests pending. If there is no cooperator, and the active |
8871 |
-+ * queue still has requests in flight or is idling for a new request, |
8872 |
-+ * then keep it. |
8873 |
-+ */ |
8874 |
-+ if (new_bfqq == NULL && (timer_pending(&bfqd->idle_slice_timer) || |
8875 |
-+ (bfqq->dispatched != 0 && |
8876 |
-+ (bfq_bfqq_idle_window(bfqq) || bfqq->raising_coeff > 1) && |
8877 |
-+ !bfq_queue_nonrot_noidle(bfqd, bfqq)))) { |
8878 |
-+ bfqq = NULL; |
8879 |
-+ goto keep_queue; |
8880 |
-+ } else if (new_bfqq != NULL && timer_pending(&bfqd->idle_slice_timer)) { |
8881 |
-+ /* |
8882 |
-+ * Expiring the queue because there is a close cooperator, |
8883 |
-+ * cancel timer. |
8884 |
-+ */ |
8885 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
8886 |
-+ del_timer(&bfqd->idle_slice_timer); |
8887 |
-+ } |
8888 |
-+ |
8889 |
-+ reason = BFQ_BFQQ_NO_MORE_REQUESTS; |
8890 |
-+expire: |
8891 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, reason); |
8892 |
-+new_queue: |
8893 |
-+ bfqq = bfq_set_active_queue(bfqd, new_bfqq); |
8894 |
-+ bfq_log(bfqd, "select_queue: new queue %d returned", |
8895 |
-+ bfqq != NULL ? bfqq->pid : 0); |
8896 |
-+keep_queue: |
8897 |
-+ return bfqq; |
8898 |
-+} |
8899 |
-+ |
8900 |
-+static void update_raising_data(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
8901 |
-+{ |
8902 |
-+ if (bfqq->raising_coeff > 1) { /* queue is being boosted */ |
8903 |
-+ struct bfq_entity *entity = &bfqq->entity; |
8904 |
-+ |
8905 |
-+ bfq_log_bfqq(bfqd, bfqq, |
8906 |
-+ "raising period dur %u/%u msec, " |
8907 |
-+ "old raising coeff %u, w %d(%d)", |
8908 |
-+ jiffies_to_msecs(jiffies - |
8909 |
-+ bfqq->last_rais_start_finish), |
8910 |
-+ jiffies_to_msecs(bfqq->raising_cur_max_time), |
8911 |
-+ bfqq->raising_coeff, |
8912 |
-+ bfqq->entity.weight, bfqq->entity.orig_weight); |
8913 |
-+ |
8914 |
-+ BUG_ON(bfqq != bfqd->active_queue && entity->weight != |
8915 |
-+ entity->orig_weight * bfqq->raising_coeff); |
8916 |
-+ if(entity->ioprio_changed) |
8917 |
-+ bfq_log_bfqq(bfqd, bfqq, |
8918 |
-+ "WARN: pending prio change"); |
8919 |
-+ /* |
8920 |
-+ * If too much time has elapsed from the beginning |
8921 |
-+ * of this weight-raising period and process is not soft |
8922 |
-+ * real-time, stop it |
8923 |
-+ */ |
8924 |
-+ if (jiffies - bfqq->last_rais_start_finish > |
8925 |
-+ bfqq->raising_cur_max_time) { |
8926 |
-+ int soft_rt = bfqd->bfq_raising_max_softrt_rate > 0 && |
8927 |
-+ bfqq->soft_rt_next_start < jiffies; |
8928 |
-+ |
8929 |
-+ bfqq->last_rais_start_finish = jiffies; |
8930 |
-+ if (soft_rt) |
8931 |
-+ bfqq->raising_cur_max_time = |
8932 |
-+ bfqd->bfq_raising_rt_max_time; |
8933 |
-+ else { |
8934 |
-+ bfq_log_bfqq(bfqd, bfqq, |
8935 |
-+ "wrais ending at %llu msec," |
8936 |
-+ "rais_max_time %u", |
8937 |
-+ bfqq->last_rais_start_finish, |
8938 |
-+ jiffies_to_msecs(bfqq-> |
8939 |
-+ raising_cur_max_time)); |
8940 |
-+ bfq_bfqq_end_raising(bfqq); |
8941 |
-+ __bfq_entity_update_weight_prio( |
8942 |
-+ bfq_entity_service_tree(entity), |
8943 |
-+ entity); |
8944 |
-+ } |
8945 |
-+ } |
8946 |
-+ } |
8947 |
-+} |
8948 |
-+ |
8949 |
-+/* |
8950 |
-+ * Dispatch one request from bfqq, moving it to the request queue |
8951 |
-+ * dispatch list. |
8952 |
-+ */ |
8953 |
-+static int bfq_dispatch_request(struct bfq_data *bfqd, |
8954 |
-+ struct bfq_queue *bfqq) |
8955 |
-+{ |
8956 |
-+ int dispatched = 0; |
8957 |
-+ struct request *rq; |
8958 |
-+ unsigned long service_to_charge; |
8959 |
-+ |
8960 |
-+ BUG_ON(RB_EMPTY_ROOT(&bfqq->sort_list)); |
8961 |
-+ |
8962 |
-+ /* Follow expired path, else get first next available. */ |
8963 |
-+ rq = bfq_check_fifo(bfqq); |
8964 |
-+ if (rq == NULL) |
8965 |
-+ rq = bfqq->next_rq; |
8966 |
-+ service_to_charge = bfq_serv_to_charge(rq, bfqq); |
8967 |
-+ |
8968 |
-+ if (service_to_charge > bfq_bfqq_budget_left(bfqq)) { |
8969 |
-+ /* |
8970 |
-+ * This may happen if the next rq is chosen |
8971 |
-+ * in fifo order instead of sector order. |
8972 |
-+ * The budget is properly dimensioned |
8973 |
-+ * to be always sufficient to serve the next request |
8974 |
-+ * only if it is chosen in sector order. The reason is |
8975 |
-+ * that it would be quite inefficient and little useful |
8976 |
-+ * to always make sure that the budget is large enough |
8977 |
-+ * to serve even the possible next rq in fifo order. |
8978 |
-+ * In fact, requests are seldom served in fifo order. |
8979 |
-+ * |
8980 |
-+ * Expire the queue for budget exhaustion, and |
8981 |
-+ * make sure that the next act_budget is enough |
8982 |
-+ * to serve the next request, even if it comes |
8983 |
-+ * from the fifo expired path. |
8984 |
-+ */ |
8985 |
-+ bfqq->next_rq = rq; |
8986 |
-+ /* |
8987 |
-+ * Since this dispatch is failed, make sure that |
8988 |
-+ * a new one will be performed |
8989 |
-+ */ |
8990 |
-+ if (!bfqd->rq_in_driver) |
8991 |
-+ bfq_schedule_dispatch(bfqd); |
8992 |
-+ goto expire; |
8993 |
-+ } |
8994 |
-+ |
8995 |
-+ /* Finally, insert request into driver dispatch list. */ |
8996 |
-+ bfq_bfqq_served(bfqq, service_to_charge); |
8997 |
-+ bfq_dispatch_insert(bfqd->queue, rq); |
8998 |
-+ |
8999 |
-+ update_raising_data(bfqd, bfqq); |
9000 |
-+ |
9001 |
-+ bfq_log_bfqq(bfqd, bfqq, "dispatched %u sec req (%llu), " |
9002 |
-+ "budg left %lu", |
9003 |
-+ blk_rq_sectors(rq), |
9004 |
-+ (long long unsigned)blk_rq_pos(rq), |
9005 |
-+ bfq_bfqq_budget_left(bfqq)); |
9006 |
-+ |
9007 |
-+ dispatched++; |
9008 |
-+ |
9009 |
-+ if (bfqd->active_bic == NULL) { |
9010 |
-+ atomic_long_inc(&RQ_BIC(rq)->icq.ioc->refcount); |
9011 |
-+ bfqd->active_bic = RQ_BIC(rq); |
9012 |
-+ } |
9013 |
-+ |
9014 |
-+ if (bfqd->busy_queues > 1 && ((!bfq_bfqq_sync(bfqq) && |
9015 |
-+ dispatched >= bfqd->bfq_max_budget_async_rq) || |
9016 |
-+ bfq_class_idle(bfqq))) |
9017 |
-+ goto expire; |
9018 |
-+ |
9019 |
-+ return dispatched; |
9020 |
-+ |
9021 |
-+expire: |
9022 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_EXHAUSTED); |
9023 |
-+ return dispatched; |
9024 |
-+} |
9025 |
-+ |
9026 |
-+static int __bfq_forced_dispatch_bfqq(struct bfq_queue *bfqq) |
9027 |
-+{ |
9028 |
-+ int dispatched = 0; |
9029 |
-+ |
9030 |
-+ while (bfqq->next_rq != NULL) { |
9031 |
-+ bfq_dispatch_insert(bfqq->bfqd->queue, bfqq->next_rq); |
9032 |
-+ dispatched++; |
9033 |
-+ } |
9034 |
-+ |
9035 |
-+ BUG_ON(!list_empty(&bfqq->fifo)); |
9036 |
-+ return dispatched; |
9037 |
-+} |
9038 |
-+ |
9039 |
-+/* |
9040 |
-+ * Drain our current requests. Used for barriers and when switching |
9041 |
-+ * io schedulers on-the-fly. |
9042 |
-+ */ |
9043 |
-+static int bfq_forced_dispatch(struct bfq_data *bfqd) |
9044 |
-+{ |
9045 |
-+ struct bfq_queue *bfqq, *n; |
9046 |
-+ struct bfq_service_tree *st; |
9047 |
-+ int dispatched = 0; |
9048 |
-+ |
9049 |
-+ bfqq = bfqd->active_queue; |
9050 |
-+ if (bfqq != NULL) |
9051 |
-+ __bfq_bfqq_expire(bfqd, bfqq); |
9052 |
-+ |
9053 |
-+ /* |
9054 |
-+ * Loop through classes, and be careful to leave the scheduler |
9055 |
-+ * in a consistent state, as feedback mechanisms and vtime |
9056 |
-+ * updates cannot be disabled during the process. |
9057 |
-+ */ |
9058 |
-+ list_for_each_entry_safe(bfqq, n, &bfqd->active_list, bfqq_list) { |
9059 |
-+ st = bfq_entity_service_tree(&bfqq->entity); |
9060 |
-+ |
9061 |
-+ dispatched += __bfq_forced_dispatch_bfqq(bfqq); |
9062 |
-+ bfqq->max_budget = bfq_max_budget(bfqd); |
9063 |
-+ |
9064 |
-+ bfq_forget_idle(st); |
9065 |
-+ } |
9066 |
-+ |
9067 |
-+ BUG_ON(bfqd->busy_queues != 0); |
9068 |
-+ |
9069 |
-+ return dispatched; |
9070 |
-+} |
9071 |
-+ |
9072 |
-+static int bfq_dispatch_requests(struct request_queue *q, int force) |
9073 |
-+{ |
9074 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
9075 |
-+ struct bfq_queue *bfqq; |
9076 |
-+ int max_dispatch; |
9077 |
-+ |
9078 |
-+ bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues); |
9079 |
-+ if (bfqd->busy_queues == 0) |
9080 |
-+ return 0; |
9081 |
-+ |
9082 |
-+ if (unlikely(force)) |
9083 |
-+ return bfq_forced_dispatch(bfqd); |
9084 |
-+ |
9085 |
-+ if((bfqq = bfq_select_queue(bfqd)) == NULL) |
9086 |
-+ return 0; |
9087 |
-+ |
9088 |
-+ max_dispatch = bfqd->bfq_quantum; |
9089 |
-+ if (bfq_class_idle(bfqq)) |
9090 |
-+ max_dispatch = 1; |
9091 |
-+ |
9092 |
-+ if (!bfq_bfqq_sync(bfqq)) |
9093 |
-+ max_dispatch = bfqd->bfq_max_budget_async_rq; |
9094 |
-+ |
9095 |
-+ if (bfqq->dispatched >= max_dispatch) { |
9096 |
-+ if (bfqd->busy_queues > 1) |
9097 |
-+ return 0; |
9098 |
-+ if (bfqq->dispatched >= 4 * max_dispatch) |
9099 |
-+ return 0; |
9100 |
-+ } |
9101 |
-+ |
9102 |
-+ if (bfqd->sync_flight != 0 && !bfq_bfqq_sync(bfqq)) |
9103 |
-+ return 0; |
9104 |
-+ |
9105 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
9106 |
-+ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
9107 |
-+ |
9108 |
-+ if (! bfq_dispatch_request(bfqd, bfqq)) |
9109 |
-+ return 0; |
9110 |
-+ |
9111 |
-+ bfq_log_bfqq(bfqd, bfqq, "dispatched one request of %d" |
9112 |
-+ "(max_disp %d)", bfqq->pid, max_dispatch); |
9113 |
-+ |
9114 |
-+ return 1; |
9115 |
-+} |
9116 |
-+ |
9117 |
-+/* |
9118 |
-+ * Task holds one reference to the queue, dropped when task exits. Each rq |
9119 |
-+ * in-flight on this queue also holds a reference, dropped when rq is freed. |
9120 |
-+ * |
9121 |
-+ * Queue lock must be held here. |
9122 |
-+ */ |
9123 |
-+static void bfq_put_queue(struct bfq_queue *bfqq) |
9124 |
-+{ |
9125 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
9126 |
-+ |
9127 |
-+ BUG_ON(atomic_read(&bfqq->ref) <= 0); |
9128 |
-+ |
9129 |
-+ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p %d", bfqq, |
9130 |
-+ atomic_read(&bfqq->ref)); |
9131 |
-+ if (!atomic_dec_and_test(&bfqq->ref)) |
9132 |
-+ return; |
9133 |
-+ |
9134 |
-+ BUG_ON(rb_first(&bfqq->sort_list) != NULL); |
9135 |
-+ BUG_ON(bfqq->allocated[READ] + bfqq->allocated[WRITE] != 0); |
9136 |
-+ BUG_ON(bfqq->entity.tree != NULL); |
9137 |
-+ BUG_ON(bfq_bfqq_busy(bfqq)); |
9138 |
-+ BUG_ON(bfqd->active_queue == bfqq); |
9139 |
-+ |
9140 |
-+ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p freed", bfqq); |
9141 |
-+ |
9142 |
-+ kmem_cache_free(bfq_pool, bfqq); |
9143 |
-+} |
9144 |
-+ |
9145 |
-+static void bfq_put_cooperator(struct bfq_queue *bfqq) |
9146 |
-+{ |
9147 |
-+ struct bfq_queue *__bfqq, *next; |
9148 |
-+ |
9149 |
-+ /* |
9150 |
-+ * If this queue was scheduled to merge with another queue, be |
9151 |
-+ * sure to drop the reference taken on that queue (and others in |
9152 |
-+ * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs. |
9153 |
-+ */ |
9154 |
-+ __bfqq = bfqq->new_bfqq; |
9155 |
-+ while (__bfqq) { |
9156 |
-+ if (__bfqq == bfqq) { |
9157 |
-+ WARN(1, "bfqq->new_bfqq loop detected.\n"); |
9158 |
-+ break; |
9159 |
-+ } |
9160 |
-+ next = __bfqq->new_bfqq; |
9161 |
-+ bfq_put_queue(__bfqq); |
9162 |
-+ __bfqq = next; |
9163 |
-+ } |
9164 |
-+} |
9165 |
-+ |
9166 |
-+static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
9167 |
-+{ |
9168 |
-+ if (bfqq == bfqd->active_queue) { |
9169 |
-+ __bfq_bfqq_expire(bfqd, bfqq); |
9170 |
-+ bfq_schedule_dispatch(bfqd); |
9171 |
-+ } |
9172 |
-+ |
9173 |
-+ bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, |
9174 |
-+ atomic_read(&bfqq->ref)); |
9175 |
-+ |
9176 |
-+ bfq_put_cooperator(bfqq); |
9177 |
-+ |
9178 |
-+ bfq_put_queue(bfqq); |
9179 |
-+} |
9180 |
-+ |
9181 |
-+static void bfq_init_icq(struct io_cq *icq) |
9182 |
-+{ |
9183 |
-+ struct bfq_io_cq *bic = icq_to_bic(icq); |
9184 |
-+ |
9185 |
-+ bic->ttime.last_end_request = jiffies; |
9186 |
-+} |
9187 |
-+ |
9188 |
-+static void bfq_exit_icq(struct io_cq *icq) |
9189 |
-+{ |
9190 |
-+ struct bfq_io_cq *bic = icq_to_bic(icq); |
9191 |
-+ struct bfq_data *bfqd = bic_to_bfqd(bic); |
9192 |
-+ |
9193 |
-+ if (bic->bfqq[BLK_RW_ASYNC]) { |
9194 |
-+ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_ASYNC]); |
9195 |
-+ bic->bfqq[BLK_RW_ASYNC] = NULL; |
9196 |
-+ } |
9197 |
-+ |
9198 |
-+ if (bic->bfqq[BLK_RW_SYNC]) { |
9199 |
-+ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_SYNC]); |
9200 |
-+ bic->bfqq[BLK_RW_SYNC] = NULL; |
9201 |
-+ } |
9202 |
-+} |
9203 |
-+ |
9204 |
-+/* |
9205 |
-+ * Update the entity prio values; note that the new values will not |
9206 |
-+ * be used until the next (re)activation. |
9207 |
-+ */ |
9208 |
-+static void bfq_init_prio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) |
9209 |
-+{ |
9210 |
-+ struct task_struct *tsk = current; |
9211 |
-+ int ioprio_class; |
9212 |
-+ |
9213 |
-+ if (!bfq_bfqq_prio_changed(bfqq)) |
9214 |
-+ return; |
9215 |
-+ |
9216 |
-+ ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
9217 |
-+ switch (ioprio_class) { |
9218 |
-+ default: |
9219 |
-+ printk(KERN_ERR "bfq: bad prio %x\n", ioprio_class); |
9220 |
-+ case IOPRIO_CLASS_NONE: |
9221 |
-+ /* |
9222 |
-+ * No prio set, inherit CPU scheduling settings. |
9223 |
-+ */ |
9224 |
-+ bfqq->entity.new_ioprio = task_nice_ioprio(tsk); |
9225 |
-+ bfqq->entity.new_ioprio_class = task_nice_ioclass(tsk); |
9226 |
-+ break; |
9227 |
-+ case IOPRIO_CLASS_RT: |
9228 |
-+ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
9229 |
-+ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_RT; |
9230 |
-+ break; |
9231 |
-+ case IOPRIO_CLASS_BE: |
9232 |
-+ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
9233 |
-+ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_BE; |
9234 |
-+ break; |
9235 |
-+ case IOPRIO_CLASS_IDLE: |
9236 |
-+ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_IDLE; |
9237 |
-+ bfqq->entity.new_ioprio = 7; |
9238 |
-+ bfq_clear_bfqq_idle_window(bfqq); |
9239 |
-+ break; |
9240 |
-+ } |
9241 |
-+ |
9242 |
-+ bfqq->entity.ioprio_changed = 1; |
9243 |
-+ |
9244 |
-+ /* |
9245 |
-+ * Keep track of original prio settings in case we have to temporarily |
9246 |
-+ * elevate the priority of this queue. |
9247 |
-+ */ |
9248 |
-+ bfqq->org_ioprio = bfqq->entity.new_ioprio; |
9249 |
-+ bfq_clear_bfqq_prio_changed(bfqq); |
9250 |
-+} |
9251 |
-+ |
9252 |
-+static void bfq_changed_ioprio(struct bfq_io_cq *bic) |
9253 |
-+{ |
9254 |
-+ struct bfq_data *bfqd; |
9255 |
-+ struct bfq_queue *bfqq, *new_bfqq; |
9256 |
-+ struct bfq_group *bfqg; |
9257 |
-+ unsigned long uninitialized_var(flags); |
9258 |
-+ int ioprio = bic->icq.ioc->ioprio; |
9259 |
-+ |
9260 |
-+ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), &flags); |
9261 |
-+ /* |
9262 |
-+ * This condition may trigger on a newly created bic, be sure to drop the |
9263 |
-+ * lock before returning. |
9264 |
-+ */ |
9265 |
-+ if (unlikely(bfqd == NULL) || likely(bic->ioprio == ioprio)) |
9266 |
-+ goto out; |
9267 |
-+ |
9268 |
-+ bfqq = bic->bfqq[BLK_RW_ASYNC]; |
9269 |
-+ if (bfqq != NULL) { |
9270 |
-+ bfqg = container_of(bfqq->entity.sched_data, struct bfq_group, |
9271 |
-+ sched_data); |
9272 |
-+ new_bfqq = bfq_get_queue(bfqd, bfqg, BLK_RW_ASYNC, bic, |
9273 |
-+ GFP_ATOMIC); |
9274 |
-+ if (new_bfqq != NULL) { |
9275 |
-+ bic->bfqq[BLK_RW_ASYNC] = new_bfqq; |
9276 |
-+ bfq_log_bfqq(bfqd, bfqq, |
9277 |
-+ "changed_ioprio: bfqq %p %d", |
9278 |
-+ bfqq, atomic_read(&bfqq->ref)); |
9279 |
-+ bfq_put_queue(bfqq); |
9280 |
-+ } |
9281 |
-+ } |
9282 |
-+ |
9283 |
-+ bfqq = bic->bfqq[BLK_RW_SYNC]; |
9284 |
-+ if (bfqq != NULL) |
9285 |
-+ bfq_mark_bfqq_prio_changed(bfqq); |
9286 |
-+ |
9287 |
-+ bic->ioprio = ioprio; |
9288 |
-+ |
9289 |
-+out: |
9290 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
9291 |
-+} |
9292 |
-+ |
9293 |
-+static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
9294 |
-+ pid_t pid, int is_sync) |
9295 |
-+{ |
9296 |
-+ RB_CLEAR_NODE(&bfqq->entity.rb_node); |
9297 |
-+ INIT_LIST_HEAD(&bfqq->fifo); |
9298 |
-+ |
9299 |
-+ atomic_set(&bfqq->ref, 0); |
9300 |
-+ bfqq->bfqd = bfqd; |
9301 |
-+ |
9302 |
-+ bfq_mark_bfqq_prio_changed(bfqq); |
9303 |
-+ |
9304 |
-+ if (is_sync) { |
9305 |
-+ if (!bfq_class_idle(bfqq)) |
9306 |
-+ bfq_mark_bfqq_idle_window(bfqq); |
9307 |
-+ bfq_mark_bfqq_sync(bfqq); |
9308 |
-+ } |
9309 |
-+ |
9310 |
-+ /* Tentative initial value to trade off between thr and lat */ |
9311 |
-+ bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
9312 |
-+ bfqq->pid = pid; |
9313 |
-+ |
9314 |
-+ bfqq->raising_coeff = 1; |
9315 |
-+ bfqq->last_rais_start_finish = 0; |
9316 |
-+ bfqq->soft_rt_next_start = -1; |
9317 |
-+} |
9318 |
-+ |
9319 |
-+static struct bfq_queue *bfq_find_alloc_queue(struct bfq_data *bfqd, |
9320 |
-+ struct bfq_group *bfqg, |
9321 |
-+ int is_sync, |
9322 |
-+ struct bfq_io_cq *bic, |
9323 |
-+ gfp_t gfp_mask) |
9324 |
-+{ |
9325 |
-+ struct bfq_queue *bfqq, *new_bfqq = NULL; |
9326 |
-+ |
9327 |
-+retry: |
9328 |
-+ /* bic always exists here */ |
9329 |
-+ bfqq = bic_to_bfqq(bic, is_sync); |
9330 |
-+ |
9331 |
-+ /* |
9332 |
-+ * Always try a new alloc if we fall back to the OOM bfqq |
9333 |
-+ * originally, since it should just be a temporary situation. |
9334 |
-+ */ |
9335 |
-+ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
9336 |
-+ bfqq = NULL; |
9337 |
-+ if (new_bfqq != NULL) { |
9338 |
-+ bfqq = new_bfqq; |
9339 |
-+ new_bfqq = NULL; |
9340 |
-+ } else if (gfp_mask & __GFP_WAIT) { |
9341 |
-+ spin_unlock_irq(bfqd->queue->queue_lock); |
9342 |
-+ new_bfqq = kmem_cache_alloc_node(bfq_pool, |
9343 |
-+ gfp_mask | __GFP_ZERO, |
9344 |
-+ bfqd->queue->node); |
9345 |
-+ spin_lock_irq(bfqd->queue->queue_lock); |
9346 |
-+ if (new_bfqq != NULL) |
9347 |
-+ goto retry; |
9348 |
-+ } else { |
9349 |
-+ bfqq = kmem_cache_alloc_node(bfq_pool, |
9350 |
-+ gfp_mask | __GFP_ZERO, |
9351 |
-+ bfqd->queue->node); |
9352 |
-+ } |
9353 |
-+ |
9354 |
-+ if (bfqq != NULL) { |
9355 |
-+ bfq_init_bfqq(bfqd, bfqq, current->pid, is_sync); |
9356 |
-+ bfq_log_bfqq(bfqd, bfqq, "allocated"); |
9357 |
-+ } else { |
9358 |
-+ bfqq = &bfqd->oom_bfqq; |
9359 |
-+ bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); |
9360 |
-+ } |
9361 |
-+ |
9362 |
-+ bfq_init_prio_data(bfqq, bic); |
9363 |
-+ bfq_init_entity(&bfqq->entity, bfqg); |
9364 |
-+ } |
9365 |
-+ |
9366 |
-+ if (new_bfqq != NULL) |
9367 |
-+ kmem_cache_free(bfq_pool, new_bfqq); |
9368 |
-+ |
9369 |
-+ return bfqq; |
9370 |
-+} |
9371 |
-+ |
9372 |
-+static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, |
9373 |
-+ struct bfq_group *bfqg, |
9374 |
-+ int ioprio_class, int ioprio) |
9375 |
-+{ |
9376 |
-+ switch (ioprio_class) { |
9377 |
-+ case IOPRIO_CLASS_RT: |
9378 |
-+ return &bfqg->async_bfqq[0][ioprio]; |
9379 |
-+ case IOPRIO_CLASS_NONE: |
9380 |
-+ ioprio = IOPRIO_NORM; |
9381 |
-+ /* fall through */ |
9382 |
-+ case IOPRIO_CLASS_BE: |
9383 |
-+ return &bfqg->async_bfqq[1][ioprio]; |
9384 |
-+ case IOPRIO_CLASS_IDLE: |
9385 |
-+ return &bfqg->async_idle_bfqq; |
9386 |
-+ default: |
9387 |
-+ BUG(); |
9388 |
-+ } |
9389 |
-+} |
9390 |
-+ |
9391 |
-+static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
9392 |
-+ struct bfq_group *bfqg, int is_sync, |
9393 |
-+ struct bfq_io_cq *bic, gfp_t gfp_mask) |
9394 |
-+{ |
9395 |
-+ const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
9396 |
-+ const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
9397 |
-+ struct bfq_queue **async_bfqq = NULL; |
9398 |
-+ struct bfq_queue *bfqq = NULL; |
9399 |
-+ |
9400 |
-+ if (!is_sync) { |
9401 |
-+ async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
9402 |
-+ ioprio); |
9403 |
-+ bfqq = *async_bfqq; |
9404 |
-+ } |
9405 |
-+ |
9406 |
-+ if (bfqq == NULL) |
9407 |
-+ bfqq = bfq_find_alloc_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
9408 |
-+ |
9409 |
-+ /* |
9410 |
-+ * Pin the queue now that it's allocated, scheduler exit will prune it. |
9411 |
-+ */ |
9412 |
-+ if (!is_sync && *async_bfqq == NULL) { |
9413 |
-+ atomic_inc(&bfqq->ref); |
9414 |
-+ bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", |
9415 |
-+ bfqq, atomic_read(&bfqq->ref)); |
9416 |
-+ *async_bfqq = bfqq; |
9417 |
-+ } |
9418 |
-+ |
9419 |
-+ atomic_inc(&bfqq->ref); |
9420 |
-+ bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, |
9421 |
-+ atomic_read(&bfqq->ref)); |
9422 |
-+ return bfqq; |
9423 |
-+} |
9424 |
-+ |
9425 |
-+static void bfq_update_io_thinktime(struct bfq_data *bfqd, |
9426 |
-+ struct bfq_io_cq *bic) |
9427 |
-+{ |
9428 |
-+ unsigned long elapsed = jiffies - bic->ttime.last_end_request; |
9429 |
-+ unsigned long ttime = min(elapsed, 2UL * bfqd->bfq_slice_idle); |
9430 |
-+ |
9431 |
-+ bic->ttime.ttime_samples = (7*bic->ttime.ttime_samples + 256) / 8; |
9432 |
-+ bic->ttime.ttime_total = (7*bic->ttime.ttime_total + 256*ttime) / 8; |
9433 |
-+ bic->ttime.ttime_mean = (bic->ttime.ttime_total + 128) / bic->ttime.ttime_samples; |
9434 |
-+} |
9435 |
-+ |
9436 |
-+static void bfq_update_io_seektime(struct bfq_data *bfqd, |
9437 |
-+ struct bfq_queue *bfqq, |
9438 |
-+ struct request *rq) |
9439 |
-+{ |
9440 |
-+ sector_t sdist; |
9441 |
-+ u64 total; |
9442 |
-+ |
9443 |
-+ if (bfqq->last_request_pos < blk_rq_pos(rq)) |
9444 |
-+ sdist = blk_rq_pos(rq) - bfqq->last_request_pos; |
9445 |
-+ else |
9446 |
-+ sdist = bfqq->last_request_pos - blk_rq_pos(rq); |
9447 |
-+ |
9448 |
-+ /* |
9449 |
-+ * Don't allow the seek distance to get too large from the |
9450 |
-+ * odd fragment, pagein, etc. |
9451 |
-+ */ |
9452 |
-+ if (bfqq->seek_samples == 0) /* first request, not really a seek */ |
9453 |
-+ sdist = 0; |
9454 |
-+ else if (bfqq->seek_samples <= 60) /* second & third seek */ |
9455 |
-+ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*1024); |
9456 |
-+ else |
9457 |
-+ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*64); |
9458 |
-+ |
9459 |
-+ bfqq->seek_samples = (7*bfqq->seek_samples + 256) / 8; |
9460 |
-+ bfqq->seek_total = (7*bfqq->seek_total + (u64)256*sdist) / 8; |
9461 |
-+ total = bfqq->seek_total + (bfqq->seek_samples/2); |
9462 |
-+ do_div(total, bfqq->seek_samples); |
9463 |
-+ if (bfq_bfqq_coop(bfqq)) { |
9464 |
-+ /* |
9465 |
-+ * If the mean seektime increases for a (non-seeky) shared |
9466 |
-+ * queue, some cooperator is likely to be idling too much. |
9467 |
-+ * On the contrary, if it decreases, some cooperator has |
9468 |
-+ * probably waked up. |
9469 |
-+ * |
9470 |
-+ */ |
9471 |
-+ if ((sector_t)total < bfqq->seek_mean) |
9472 |
-+ bfq_mark_bfqq_some_coop_idle(bfqq) ; |
9473 |
-+ else if ((sector_t)total > bfqq->seek_mean) |
9474 |
-+ bfq_clear_bfqq_some_coop_idle(bfqq) ; |
9475 |
-+ } |
9476 |
-+ bfqq->seek_mean = (sector_t)total; |
9477 |
-+ |
9478 |
-+ bfq_log_bfqq(bfqd, bfqq, "dist=%llu mean=%llu", (u64)sdist, |
9479 |
-+ (u64)bfqq->seek_mean); |
9480 |
-+} |
9481 |
-+ |
9482 |
-+/* |
9483 |
-+ * Disable idle window if the process thinks too long or seeks so much that |
9484 |
-+ * it doesn't matter. |
9485 |
-+ */ |
9486 |
-+static void bfq_update_idle_window(struct bfq_data *bfqd, |
9487 |
-+ struct bfq_queue *bfqq, |
9488 |
-+ struct bfq_io_cq *bic) |
9489 |
-+{ |
9490 |
-+ int enable_idle; |
9491 |
-+ |
9492 |
-+ /* Don't idle for async or idle io prio class. */ |
9493 |
-+ if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq)) |
9494 |
-+ return; |
9495 |
-+ |
9496 |
-+ enable_idle = bfq_bfqq_idle_window(bfqq); |
9497 |
-+ |
9498 |
-+ if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
9499 |
-+ bfqd->bfq_slice_idle == 0 || |
9500 |
-+ (bfqd->hw_tag && BFQQ_SEEKY(bfqq) && |
9501 |
-+ bfqq->raising_coeff == 1)) |
9502 |
-+ enable_idle = 0; |
9503 |
-+ else if (bfq_sample_valid(bic->ttime.ttime_samples)) { |
9504 |
-+ if (bic->ttime.ttime_mean > bfqd->bfq_slice_idle && |
9505 |
-+ bfqq->raising_coeff == 1) |
9506 |
-+ enable_idle = 0; |
9507 |
-+ else |
9508 |
-+ enable_idle = 1; |
9509 |
-+ } |
9510 |
-+ bfq_log_bfqq(bfqd, bfqq, "update_idle_window: enable_idle %d", |
9511 |
-+ enable_idle); |
9512 |
-+ |
9513 |
-+ if (enable_idle) |
9514 |
-+ bfq_mark_bfqq_idle_window(bfqq); |
9515 |
-+ else |
9516 |
-+ bfq_clear_bfqq_idle_window(bfqq); |
9517 |
-+} |
9518 |
-+ |
9519 |
-+/* |
9520 |
-+ * Called when a new fs request (rq) is added to bfqq. Check if there's |
9521 |
-+ * something we should do about it. |
9522 |
-+ */ |
9523 |
-+static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
9524 |
-+ struct request *rq) |
9525 |
-+{ |
9526 |
-+ struct bfq_io_cq *bic = RQ_BIC(rq); |
9527 |
-+ |
9528 |
-+ if (rq->cmd_flags & REQ_META) |
9529 |
-+ bfqq->meta_pending++; |
9530 |
-+ |
9531 |
-+ bfq_update_io_thinktime(bfqd, bic); |
9532 |
-+ bfq_update_io_seektime(bfqd, bfqq, rq); |
9533 |
-+ if (bfqq->entity.service > bfq_max_budget(bfqd) / 8 || |
9534 |
-+ !BFQQ_SEEKY(bfqq)) |
9535 |
-+ bfq_update_idle_window(bfqd, bfqq, bic); |
9536 |
-+ |
9537 |
-+ bfq_log_bfqq(bfqd, bfqq, |
9538 |
-+ "rq_enqueued: idle_window=%d (seeky %d, mean %llu)", |
9539 |
-+ bfq_bfqq_idle_window(bfqq), BFQQ_SEEKY(bfqq), |
9540 |
-+ (long long unsigned)bfqq->seek_mean); |
9541 |
-+ |
9542 |
-+ bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
9543 |
-+ |
9544 |
-+ if (bfqq == bfqd->active_queue) { |
9545 |
-+ /* |
9546 |
-+ * If there is just this request queued and the request |
9547 |
-+ * is small, just exit. |
9548 |
-+ * In this way, if the disk is being idled to wait for a new |
9549 |
-+ * request from the active queue, we avoid unplugging the |
9550 |
-+ * device now. |
9551 |
-+ * |
9552 |
-+ * By doing so, we spare the disk to be committed |
9553 |
-+ * to serve just a small request. On the contrary, we wait for |
9554 |
-+ * the block layer to decide when to unplug the device: |
9555 |
-+ * hopefully, new requests will be merged to this |
9556 |
-+ * one quickly, then the device will be unplugged |
9557 |
-+ * and larger requests will be dispatched. |
9558 |
-+ */ |
9559 |
-+ if (bfqq->queued[rq_is_sync(rq)] == 1 && |
9560 |
-+ blk_rq_sectors(rq) < 32) { |
9561 |
-+ return; |
9562 |
-+ } |
9563 |
-+ if (bfq_bfqq_wait_request(bfqq)) { |
9564 |
-+ /* |
9565 |
-+ * If we are waiting for a request for this queue, let |
9566 |
-+ * it rip immediately and flag that we must not expire |
9567 |
-+ * this queue just now. |
9568 |
-+ */ |
9569 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
9570 |
-+ del_timer(&bfqd->idle_slice_timer); |
9571 |
-+ /* |
9572 |
-+ * Here we can safely expire the queue, in |
9573 |
-+ * case of budget timeout, without wasting |
9574 |
-+ * guarantees |
9575 |
-+ */ |
9576 |
-+ if (bfq_bfqq_budget_timeout(bfqq)) |
9577 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, |
9578 |
-+ BFQ_BFQQ_BUDGET_TIMEOUT); |
9579 |
-+ __blk_run_queue(bfqd->queue); |
9580 |
-+ } |
9581 |
-+ } |
9582 |
-+} |
9583 |
-+ |
9584 |
-+static void bfq_insert_request(struct request_queue *q, struct request *rq) |
9585 |
-+{ |
9586 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
9587 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
9588 |
-+ |
9589 |
-+ assert_spin_locked(bfqd->queue->queue_lock); |
9590 |
-+ bfq_init_prio_data(bfqq, RQ_BIC(rq)); |
9591 |
-+ |
9592 |
-+ bfq_add_rq_rb(rq); |
9593 |
-+ |
9594 |
-+ rq_set_fifo_time(rq, jiffies + bfqd->bfq_fifo_expire[rq_is_sync(rq)]); |
9595 |
-+ list_add_tail(&rq->queuelist, &bfqq->fifo); |
9596 |
-+ |
9597 |
-+ bfq_rq_enqueued(bfqd, bfqq, rq); |
9598 |
-+} |
9599 |
-+ |
9600 |
-+static void bfq_update_hw_tag(struct bfq_data *bfqd) |
9601 |
-+{ |
9602 |
-+ bfqd->max_rq_in_driver = max(bfqd->max_rq_in_driver, |
9603 |
-+ bfqd->rq_in_driver); |
9604 |
-+ |
9605 |
-+ if (bfqd->hw_tag == 1) |
9606 |
-+ return; |
9607 |
-+ |
9608 |
-+ /* |
9609 |
-+ * This sample is valid if the number of outstanding requests |
9610 |
-+ * is large enough to allow a queueing behavior. Note that the |
9611 |
-+ * sum is not exact, as it's not taking into account deactivated |
9612 |
-+ * requests. |
9613 |
-+ */ |
9614 |
-+ if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD) |
9615 |
-+ return; |
9616 |
-+ |
9617 |
-+ if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) |
9618 |
-+ return; |
9619 |
-+ |
9620 |
-+ bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; |
9621 |
-+ bfqd->max_rq_in_driver = 0; |
9622 |
-+ bfqd->hw_tag_samples = 0; |
9623 |
-+} |
9624 |
-+ |
9625 |
-+static void bfq_completed_request(struct request_queue *q, struct request *rq) |
9626 |
-+{ |
9627 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
9628 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
9629 |
-+ const int sync = rq_is_sync(rq); |
9630 |
-+ |
9631 |
-+ bfq_log_bfqq(bfqd, bfqq, "completed %u sects req (%d)", |
9632 |
-+ blk_rq_sectors(rq), sync); |
9633 |
-+ |
9634 |
-+ bfq_update_hw_tag(bfqd); |
9635 |
-+ |
9636 |
-+ WARN_ON(!bfqd->rq_in_driver); |
9637 |
-+ WARN_ON(!bfqq->dispatched); |
9638 |
-+ bfqd->rq_in_driver--; |
9639 |
-+ bfqq->dispatched--; |
9640 |
-+ |
9641 |
-+ if (bfq_bfqq_sync(bfqq)) |
9642 |
-+ bfqd->sync_flight--; |
9643 |
-+ |
9644 |
-+ if (sync) |
9645 |
-+ RQ_BIC(rq)->ttime.last_end_request = jiffies; |
9646 |
-+ |
9647 |
-+ /* |
9648 |
-+ * If this is the active queue, check if it needs to be expired, |
9649 |
-+ * or if we want to idle in case it has no pending requests. |
9650 |
-+ */ |
9651 |
-+ if (bfqd->active_queue == bfqq) { |
9652 |
-+ int budg_timeout = bfq_may_expire_for_budg_timeout(bfqq); |
9653 |
-+ if (bfq_bfqq_budget_new(bfqq)) |
9654 |
-+ bfq_set_budget_timeout(bfqd); |
9655 |
-+ |
9656 |
-+ /* Idling is disabled also for cooperation issues: |
9657 |
-+ * 1) there is a close cooperator for the queue, or |
9658 |
-+ * 2) the queue is shared and some cooperator is likely |
9659 |
-+ * to be idle (in this case, by not arming the idle timer, |
9660 |
-+ * we try to slow down the queue, to prevent the zones |
9661 |
-+ * of the disk accessed by the active cooperators to become |
9662 |
-+ * too distant from the zone that will be accessed by the |
9663 |
-+ * currently idle cooperators) |
9664 |
-+ */ |
9665 |
-+ if (bfq_bfqq_must_idle(bfqq, budg_timeout)) |
9666 |
-+ bfq_arm_slice_timer(bfqd); |
9667 |
-+ else if (budg_timeout) |
9668 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_TIMEOUT); |
9669 |
-+ } |
9670 |
-+ |
9671 |
-+ if (!bfqd->rq_in_driver) |
9672 |
-+ bfq_schedule_dispatch(bfqd); |
9673 |
-+} |
9674 |
-+ |
9675 |
-+static inline int __bfq_may_queue(struct bfq_queue *bfqq) |
9676 |
-+{ |
9677 |
-+ if (bfq_bfqq_wait_request(bfqq) && bfq_bfqq_must_alloc(bfqq)) { |
9678 |
-+ bfq_clear_bfqq_must_alloc(bfqq); |
9679 |
-+ return ELV_MQUEUE_MUST; |
9680 |
-+ } |
9681 |
-+ |
9682 |
-+ return ELV_MQUEUE_MAY; |
9683 |
-+} |
9684 |
-+ |
9685 |
-+static int bfq_may_queue(struct request_queue *q, int rw) |
9686 |
-+{ |
9687 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
9688 |
-+ struct task_struct *tsk = current; |
9689 |
-+ struct bfq_io_cq *bic; |
9690 |
-+ struct bfq_queue *bfqq; |
9691 |
-+ |
9692 |
-+ /* |
9693 |
-+ * Don't force setup of a queue from here, as a call to may_queue |
9694 |
-+ * does not necessarily imply that a request actually will be queued. |
9695 |
-+ * So just lookup a possibly existing queue, or return 'may queue' |
9696 |
-+ * if that fails. |
9697 |
-+ */ |
9698 |
-+ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
9699 |
-+ if (bic == NULL) |
9700 |
-+ return ELV_MQUEUE_MAY; |
9701 |
-+ |
9702 |
-+ bfqq = bic_to_bfqq(bic, rw_is_sync(rw)); |
9703 |
-+ if (bfqq != NULL) { |
9704 |
-+ bfq_init_prio_data(bfqq, bic); |
9705 |
-+ |
9706 |
-+ return __bfq_may_queue(bfqq); |
9707 |
-+ } |
9708 |
-+ |
9709 |
-+ return ELV_MQUEUE_MAY; |
9710 |
-+} |
9711 |
-+ |
9712 |
-+/* |
9713 |
-+ * Queue lock held here. |
9714 |
-+ */ |
9715 |
-+static void bfq_put_request(struct request *rq) |
9716 |
-+{ |
9717 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
9718 |
-+ |
9719 |
-+ if (bfqq != NULL) { |
9720 |
-+ const int rw = rq_data_dir(rq); |
9721 |
-+ |
9722 |
-+ BUG_ON(!bfqq->allocated[rw]); |
9723 |
-+ bfqq->allocated[rw]--; |
9724 |
-+ |
9725 |
-+ rq->elv.priv[0] = NULL; |
9726 |
-+ rq->elv.priv[1] = NULL; |
9727 |
-+ |
9728 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "put_request %p, %d", |
9729 |
-+ bfqq, atomic_read(&bfqq->ref)); |
9730 |
-+ bfq_put_queue(bfqq); |
9731 |
-+ } |
9732 |
-+} |
9733 |
-+ |
9734 |
-+static struct bfq_queue * |
9735 |
-+bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
9736 |
-+ struct bfq_queue *bfqq) |
9737 |
-+{ |
9738 |
-+ bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
9739 |
-+ (long unsigned)bfqq->new_bfqq->pid); |
9740 |
-+ bic_set_bfqq(bic, bfqq->new_bfqq, 1); |
9741 |
-+ bfq_mark_bfqq_coop(bfqq->new_bfqq); |
9742 |
-+ bfq_put_queue(bfqq); |
9743 |
-+ return bic_to_bfqq(bic, 1); |
9744 |
-+} |
9745 |
-+ |
9746 |
-+/* |
9747 |
-+ * Returns NULL if a new bfqq should be allocated, or the old bfqq if this |
9748 |
-+ * was the last process referring to said bfqq. |
9749 |
-+ */ |
9750 |
-+static struct bfq_queue * |
9751 |
-+bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) |
9752 |
-+{ |
9753 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); |
9754 |
-+ if (bfqq_process_refs(bfqq) == 1) { |
9755 |
-+ bfqq->pid = current->pid; |
9756 |
-+ bfq_clear_bfqq_some_coop_idle(bfqq); |
9757 |
-+ bfq_clear_bfqq_coop(bfqq); |
9758 |
-+ bfq_clear_bfqq_split_coop(bfqq); |
9759 |
-+ return bfqq; |
9760 |
-+ } |
9761 |
-+ |
9762 |
-+ bic_set_bfqq(bic, NULL, 1); |
9763 |
-+ |
9764 |
-+ bfq_put_cooperator(bfqq); |
9765 |
-+ |
9766 |
-+ bfq_put_queue(bfqq); |
9767 |
-+ return NULL; |
9768 |
-+} |
9769 |
-+ |
9770 |
-+/* |
9771 |
-+ * Allocate bfq data structures associated with this request. |
9772 |
-+ */ |
9773 |
-+static int bfq_set_request(struct request_queue *q, struct request *rq, |
9774 |
-+ struct bio *bio, gfp_t gfp_mask) |
9775 |
-+{ |
9776 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
9777 |
-+ struct bfq_io_cq *bic = icq_to_bic(rq->elv.icq); |
9778 |
-+ const int rw = rq_data_dir(rq); |
9779 |
-+ const int is_sync = rq_is_sync(rq); |
9780 |
-+ struct bfq_queue *bfqq; |
9781 |
-+ struct bfq_group *bfqg; |
9782 |
-+ unsigned long flags; |
9783 |
-+ |
9784 |
-+ might_sleep_if(gfp_mask & __GFP_WAIT); |
9785 |
-+ |
9786 |
-+ bfq_changed_ioprio(bic); |
9787 |
-+ |
9788 |
-+ spin_lock_irqsave(q->queue_lock, flags); |
9789 |
-+ |
9790 |
-+ if (bic == NULL) |
9791 |
-+ goto queue_fail; |
9792 |
-+ |
9793 |
-+ bfqg = bfq_bic_update_cgroup(bic); |
9794 |
-+ |
9795 |
-+new_queue: |
9796 |
-+ bfqq = bic_to_bfqq(bic, is_sync); |
9797 |
-+ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
9798 |
-+ bfqq = bfq_get_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
9799 |
-+ bic_set_bfqq(bic, bfqq, is_sync); |
9800 |
-+ } else { |
9801 |
-+ /* |
9802 |
-+ * If the queue was seeky for too long, break it apart. |
9803 |
-+ */ |
9804 |
-+ if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { |
9805 |
-+ bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); |
9806 |
-+ bfqq = bfq_split_bfqq(bic, bfqq); |
9807 |
-+ if (!bfqq) |
9808 |
-+ goto new_queue; |
9809 |
-+ } |
9810 |
-+ |
9811 |
-+ /* |
9812 |
-+ * Check to see if this queue is scheduled to merge with |
9813 |
-+ * another closely cooperating queue. The merging of queues |
9814 |
-+ * happens here as it must be done in process context. |
9815 |
-+ * The reference on new_bfqq was taken in merge_bfqqs. |
9816 |
-+ */ |
9817 |
-+ if (bfqq->new_bfqq != NULL) |
9818 |
-+ bfqq = bfq_merge_bfqqs(bfqd, bic, bfqq); |
9819 |
-+ } |
9820 |
-+ |
9821 |
-+ bfqq->allocated[rw]++; |
9822 |
-+ atomic_inc(&bfqq->ref); |
9823 |
-+ bfq_log_bfqq(bfqd, bfqq, "set_request: bfqq %p, %d", bfqq, |
9824 |
-+ atomic_read(&bfqq->ref)); |
9825 |
-+ |
9826 |
-+ rq->elv.priv[0] = bic; |
9827 |
-+ rq->elv.priv[1] = bfqq; |
9828 |
-+ |
9829 |
-+ spin_unlock_irqrestore(q->queue_lock, flags); |
9830 |
-+ |
9831 |
-+ return 0; |
9832 |
-+ |
9833 |
-+queue_fail: |
9834 |
-+ bfq_schedule_dispatch(bfqd); |
9835 |
-+ spin_unlock_irqrestore(q->queue_lock, flags); |
9836 |
-+ |
9837 |
-+ return 1; |
9838 |
-+} |
9839 |
-+ |
9840 |
-+static void bfq_kick_queue(struct work_struct *work) |
9841 |
-+{ |
9842 |
-+ struct bfq_data *bfqd = |
9843 |
-+ container_of(work, struct bfq_data, unplug_work); |
9844 |
-+ struct request_queue *q = bfqd->queue; |
9845 |
-+ |
9846 |
-+ spin_lock_irq(q->queue_lock); |
9847 |
-+ __blk_run_queue(q); |
9848 |
-+ spin_unlock_irq(q->queue_lock); |
9849 |
-+} |
9850 |
-+ |
9851 |
-+/* |
9852 |
-+ * Handler of the expiration of the timer running if the active_queue |
9853 |
-+ * is idling inside its time slice. |
9854 |
-+ */ |
9855 |
-+static void bfq_idle_slice_timer(unsigned long data) |
9856 |
-+{ |
9857 |
-+ struct bfq_data *bfqd = (struct bfq_data *)data; |
9858 |
-+ struct bfq_queue *bfqq; |
9859 |
-+ unsigned long flags; |
9860 |
-+ enum bfqq_expiration reason; |
9861 |
-+ |
9862 |
-+ spin_lock_irqsave(bfqd->queue->queue_lock, flags); |
9863 |
-+ |
9864 |
-+ bfqq = bfqd->active_queue; |
9865 |
-+ /* |
9866 |
-+ * Theoretical race here: active_queue can be NULL or different |
9867 |
-+ * from the queue that was idling if the timer handler spins on |
9868 |
-+ * the queue_lock and a new request arrives for the current |
9869 |
-+ * queue and there is a full dispatch cycle that changes the |
9870 |
-+ * active_queue. This can hardly happen, but in the worst case |
9871 |
-+ * we just expire a queue too early. |
9872 |
-+ */ |
9873 |
-+ if (bfqq != NULL) { |
9874 |
-+ bfq_log_bfqq(bfqd, bfqq, "slice_timer expired"); |
9875 |
-+ if (bfq_bfqq_budget_timeout(bfqq)) |
9876 |
-+ /* |
9877 |
-+ * Also here the queue can be safely expired |
9878 |
-+ * for budget timeout without wasting |
9879 |
-+ * guarantees |
9880 |
-+ */ |
9881 |
-+ reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
9882 |
-+ else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) |
9883 |
-+ /* |
9884 |
-+ * The queue may not be empty upon timer expiration, |
9885 |
-+ * because we may not disable the timer when the first |
9886 |
-+ * request of the active queue arrives during |
9887 |
-+ * disk idling |
9888 |
-+ */ |
9889 |
-+ reason = BFQ_BFQQ_TOO_IDLE; |
9890 |
-+ else |
9891 |
-+ goto schedule_dispatch; |
9892 |
-+ |
9893 |
-+ bfq_bfqq_expire(bfqd, bfqq, 1, reason); |
9894 |
-+ } |
9895 |
-+ |
9896 |
-+schedule_dispatch: |
9897 |
-+ bfq_schedule_dispatch(bfqd); |
9898 |
-+ |
9899 |
-+ spin_unlock_irqrestore(bfqd->queue->queue_lock, flags); |
9900 |
-+} |
9901 |
-+ |
9902 |
-+static void bfq_shutdown_timer_wq(struct bfq_data *bfqd) |
9903 |
-+{ |
9904 |
-+ del_timer_sync(&bfqd->idle_slice_timer); |
9905 |
-+ cancel_work_sync(&bfqd->unplug_work); |
9906 |
-+} |
9907 |
-+ |
9908 |
-+static inline void __bfq_put_async_bfqq(struct bfq_data *bfqd, |
9909 |
-+ struct bfq_queue **bfqq_ptr) |
9910 |
-+{ |
9911 |
-+ struct bfq_group *root_group = bfqd->root_group; |
9912 |
-+ struct bfq_queue *bfqq = *bfqq_ptr; |
9913 |
-+ |
9914 |
-+ bfq_log(bfqd, "put_async_bfqq: %p", bfqq); |
9915 |
-+ if (bfqq != NULL) { |
9916 |
-+ bfq_bfqq_move(bfqd, bfqq, &bfqq->entity, root_group); |
9917 |
-+ bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
9918 |
-+ bfqq, atomic_read(&bfqq->ref)); |
9919 |
-+ bfq_put_queue(bfqq); |
9920 |
-+ *bfqq_ptr = NULL; |
9921 |
-+ } |
9922 |
-+} |
9923 |
-+ |
9924 |
-+/* |
9925 |
-+ * Release all the bfqg references to its async queues. If we are |
9926 |
-+ * deallocating the group these queues may still contain requests, so |
9927 |
-+ * we reparent them to the root cgroup (i.e., the only one that will |
9928 |
-+ * exist for sure untill all the requests on a device are gone). |
9929 |
-+ */ |
9930 |
-+static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
9931 |
-+{ |
9932 |
-+ int i, j; |
9933 |
-+ |
9934 |
-+ for (i = 0; i < 2; i++) |
9935 |
-+ for (j = 0; j < IOPRIO_BE_NR; j++) |
9936 |
-+ __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]); |
9937 |
-+ |
9938 |
-+ __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq); |
9939 |
-+} |
9940 |
-+ |
9941 |
-+static void bfq_exit_queue(struct elevator_queue *e) |
9942 |
-+{ |
9943 |
-+ struct bfq_data *bfqd = e->elevator_data; |
9944 |
-+ struct request_queue *q = bfqd->queue; |
9945 |
-+ struct bfq_queue *bfqq, *n; |
9946 |
-+ |
9947 |
-+ bfq_shutdown_timer_wq(bfqd); |
9948 |
-+ |
9949 |
-+ spin_lock_irq(q->queue_lock); |
9950 |
-+ |
9951 |
-+ BUG_ON(bfqd->active_queue != NULL); |
9952 |
-+ list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) |
9953 |
-+ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
9954 |
-+ |
9955 |
-+ bfq_disconnect_groups(bfqd); |
9956 |
-+ spin_unlock_irq(q->queue_lock); |
9957 |
-+ |
9958 |
-+ bfq_shutdown_timer_wq(bfqd); |
9959 |
-+ |
9960 |
-+ synchronize_rcu(); |
9961 |
-+ |
9962 |
-+ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
9963 |
-+ |
9964 |
-+ bfq_free_root_group(bfqd); |
9965 |
-+ kfree(bfqd); |
9966 |
-+} |
9967 |
-+ |
9968 |
-+static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
9969 |
-+{ |
9970 |
-+ struct bfq_group *bfqg; |
9971 |
-+ struct bfq_data *bfqd; |
9972 |
-+ struct elevator_queue *eq; |
9973 |
-+ |
9974 |
-+ eq = elevator_alloc(q, e); |
9975 |
-+ if (eq == NULL) |
9976 |
-+ return -ENOMEM; |
9977 |
-+ |
9978 |
-+ bfqd = kmalloc_node(sizeof(*bfqd), GFP_KERNEL | __GFP_ZERO, q->node); |
9979 |
-+ if (bfqd == NULL) { |
9980 |
-+ kobject_put(&eq->kobj); |
9981 |
-+ return -ENOMEM; |
9982 |
-+ } |
9983 |
-+ eq->elevator_data = bfqd; |
9984 |
-+ |
9985 |
-+ /* |
9986 |
-+ * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. |
9987 |
-+ * Grab a permanent reference to it, so that the normal code flow |
9988 |
-+ * will not attempt to free it. |
9989 |
-+ */ |
9990 |
-+ bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, 1, 0); |
9991 |
-+ atomic_inc(&bfqd->oom_bfqq.ref); |
9992 |
-+ |
9993 |
-+ bfqd->queue = q; |
9994 |
-+ |
9995 |
-+ spin_lock_irq(q->queue_lock); |
9996 |
-+ q->elevator = eq; |
9997 |
-+ spin_unlock_irq(q->queue_lock); |
9998 |
-+ |
9999 |
-+ bfqg = bfq_alloc_root_group(bfqd, q->node); |
10000 |
-+ if (bfqg == NULL) { |
10001 |
-+ kfree(bfqd); |
10002 |
-+ kobject_put(&eq->kobj); |
10003 |
-+ return -ENOMEM; |
10004 |
-+ } |
10005 |
-+ |
10006 |
-+ bfqd->root_group = bfqg; |
10007 |
-+ |
10008 |
-+ init_timer(&bfqd->idle_slice_timer); |
10009 |
-+ bfqd->idle_slice_timer.function = bfq_idle_slice_timer; |
10010 |
-+ bfqd->idle_slice_timer.data = (unsigned long)bfqd; |
10011 |
-+ |
10012 |
-+ bfqd->rq_pos_tree = RB_ROOT; |
10013 |
-+ |
10014 |
-+ INIT_WORK(&bfqd->unplug_work, bfq_kick_queue); |
10015 |
-+ |
10016 |
-+ INIT_LIST_HEAD(&bfqd->active_list); |
10017 |
-+ INIT_LIST_HEAD(&bfqd->idle_list); |
10018 |
-+ |
10019 |
-+ bfqd->hw_tag = -1; |
10020 |
-+ |
10021 |
-+ bfqd->bfq_max_budget = bfq_default_max_budget; |
10022 |
-+ |
10023 |
-+ bfqd->bfq_quantum = bfq_quantum; |
10024 |
-+ bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; |
10025 |
-+ bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; |
10026 |
-+ bfqd->bfq_back_max = bfq_back_max; |
10027 |
-+ bfqd->bfq_back_penalty = bfq_back_penalty; |
10028 |
-+ bfqd->bfq_slice_idle = bfq_slice_idle; |
10029 |
-+ bfqd->bfq_class_idle_last_service = 0; |
10030 |
-+ bfqd->bfq_max_budget_async_rq = bfq_max_budget_async_rq; |
10031 |
-+ bfqd->bfq_timeout[BLK_RW_ASYNC] = bfq_timeout_async; |
10032 |
-+ bfqd->bfq_timeout[BLK_RW_SYNC] = bfq_timeout_sync; |
10033 |
-+ |
10034 |
-+ bfqd->low_latency = true; |
10035 |
-+ |
10036 |
-+ bfqd->bfq_raising_coeff = 20; |
10037 |
-+ bfqd->bfq_raising_rt_max_time = msecs_to_jiffies(300); |
10038 |
-+ bfqd->bfq_raising_max_time = 0; |
10039 |
-+ bfqd->bfq_raising_min_idle_time = msecs_to_jiffies(2000); |
10040 |
-+ bfqd->bfq_raising_min_inter_arr_async = msecs_to_jiffies(500); |
10041 |
-+ bfqd->bfq_raising_max_softrt_rate = 7000; |
10042 |
-+ |
10043 |
-+ /* Initially estimate the device's peak rate as the reference rate */ |
10044 |
-+ if (blk_queue_nonrot(bfqd->queue)) { |
10045 |
-+ bfqd->RT_prod = R_nonrot * T_nonrot; |
10046 |
-+ bfqd->peak_rate = R_nonrot; |
10047 |
-+ } else { |
10048 |
-+ bfqd->RT_prod = R_rot * T_rot; |
10049 |
-+ bfqd->peak_rate = R_rot; |
10050 |
-+ } |
10051 |
-+ |
10052 |
-+ return 0; |
10053 |
-+} |
10054 |
-+ |
10055 |
-+static void bfq_slab_kill(void) |
10056 |
-+{ |
10057 |
-+ if (bfq_pool != NULL) |
10058 |
-+ kmem_cache_destroy(bfq_pool); |
10059 |
-+} |
10060 |
-+ |
10061 |
-+static int __init bfq_slab_setup(void) |
10062 |
-+{ |
10063 |
-+ bfq_pool = KMEM_CACHE(bfq_queue, 0); |
10064 |
-+ if (bfq_pool == NULL) |
10065 |
-+ return -ENOMEM; |
10066 |
-+ return 0; |
10067 |
-+} |
10068 |
-+ |
10069 |
-+static ssize_t bfq_var_show(unsigned int var, char *page) |
10070 |
-+{ |
10071 |
-+ return sprintf(page, "%d\n", var); |
10072 |
-+} |
10073 |
-+ |
10074 |
-+static ssize_t bfq_var_store(unsigned long *var, const char *page, size_t count) |
10075 |
-+{ |
10076 |
-+ unsigned long new_val; |
10077 |
-+ int ret = strict_strtoul(page, 10, &new_val); |
10078 |
-+ |
10079 |
-+ if (ret == 0) |
10080 |
-+ *var = new_val; |
10081 |
-+ |
10082 |
-+ return count; |
10083 |
-+} |
10084 |
-+ |
10085 |
-+static ssize_t bfq_raising_max_time_show(struct elevator_queue *e, char *page) |
10086 |
-+{ |
10087 |
-+ struct bfq_data *bfqd = e->elevator_data; |
10088 |
-+ return sprintf(page, "%d\n", bfqd->bfq_raising_max_time > 0 ? |
10089 |
-+ jiffies_to_msecs(bfqd->bfq_raising_max_time) : |
10090 |
-+ jiffies_to_msecs(bfq_wrais_duration(bfqd))); |
10091 |
-+} |
10092 |
-+ |
10093 |
-+static ssize_t bfq_weights_show(struct elevator_queue *e, char *page) |
10094 |
-+{ |
10095 |
-+ struct bfq_queue *bfqq; |
10096 |
-+ struct bfq_data *bfqd = e->elevator_data; |
10097 |
-+ ssize_t num_char = 0; |
10098 |
-+ |
10099 |
-+ num_char += sprintf(page + num_char, "Tot reqs queued %d\n\n", |
10100 |
-+ bfqd->queued); |
10101 |
-+ |
10102 |
-+ spin_lock_irq(bfqd->queue->queue_lock); |
10103 |
-+ |
10104 |
-+ num_char += sprintf(page + num_char, "Active:\n"); |
10105 |
-+ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) { |
10106 |
-+ num_char += sprintf(page + num_char, |
10107 |
-+ "pid%d: weight %hu, nr_queued %d %d," |
10108 |
-+ " dur %d/%u\n", |
10109 |
-+ bfqq->pid, |
10110 |
-+ bfqq->entity.weight, |
10111 |
-+ bfqq->queued[0], |
10112 |
-+ bfqq->queued[1], |
10113 |
-+ jiffies_to_msecs(jiffies - |
10114 |
-+ bfqq->last_rais_start_finish), |
10115 |
-+ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
10116 |
-+ } |
10117 |
-+ |
10118 |
-+ num_char += sprintf(page + num_char, "Idle:\n"); |
10119 |
-+ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) { |
10120 |
-+ num_char += sprintf(page + num_char, |
10121 |
-+ "pid%d: weight %hu, dur %d/%u\n", |
10122 |
-+ bfqq->pid, |
10123 |
-+ bfqq->entity.weight, |
10124 |
-+ jiffies_to_msecs(jiffies - |
10125 |
-+ bfqq->last_rais_start_finish), |
10126 |
-+ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
10127 |
-+ } |
10128 |
-+ |
10129 |
-+ spin_unlock_irq(bfqd->queue->queue_lock); |
10130 |
-+ |
10131 |
-+ return num_char; |
10132 |
-+} |
10133 |
-+ |
10134 |
-+#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
10135 |
-+static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
10136 |
-+{ \ |
10137 |
-+ struct bfq_data *bfqd = e->elevator_data; \ |
10138 |
-+ unsigned int __data = __VAR; \ |
10139 |
-+ if (__CONV) \ |
10140 |
-+ __data = jiffies_to_msecs(__data); \ |
10141 |
-+ return bfq_var_show(__data, (page)); \ |
10142 |
-+} |
10143 |
-+SHOW_FUNCTION(bfq_quantum_show, bfqd->bfq_quantum, 0); |
10144 |
-+SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 1); |
10145 |
-+SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 1); |
10146 |
-+SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); |
10147 |
-+SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); |
10148 |
-+SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 1); |
10149 |
-+SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); |
10150 |
-+SHOW_FUNCTION(bfq_max_budget_async_rq_show, bfqd->bfq_max_budget_async_rq, 0); |
10151 |
-+SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout[BLK_RW_SYNC], 1); |
10152 |
-+SHOW_FUNCTION(bfq_timeout_async_show, bfqd->bfq_timeout[BLK_RW_ASYNC], 1); |
10153 |
-+SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
10154 |
-+SHOW_FUNCTION(bfq_raising_coeff_show, bfqd->bfq_raising_coeff, 0); |
10155 |
-+SHOW_FUNCTION(bfq_raising_rt_max_time_show, bfqd->bfq_raising_rt_max_time, 1); |
10156 |
-+SHOW_FUNCTION(bfq_raising_min_idle_time_show, bfqd->bfq_raising_min_idle_time, |
10157 |
-+ 1); |
10158 |
-+SHOW_FUNCTION(bfq_raising_min_inter_arr_async_show, |
10159 |
-+ bfqd->bfq_raising_min_inter_arr_async, |
10160 |
-+ 1); |
10161 |
-+SHOW_FUNCTION(bfq_raising_max_softrt_rate_show, |
10162 |
-+ bfqd->bfq_raising_max_softrt_rate, 0); |
10163 |
-+#undef SHOW_FUNCTION |
10164 |
-+ |
10165 |
-+#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ |
10166 |
-+static ssize_t \ |
10167 |
-+__FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
10168 |
-+{ \ |
10169 |
-+ struct bfq_data *bfqd = e->elevator_data; \ |
10170 |
-+ unsigned long uninitialized_var(__data); \ |
10171 |
-+ int ret = bfq_var_store(&__data, (page), count); \ |
10172 |
-+ if (__data < (MIN)) \ |
10173 |
-+ __data = (MIN); \ |
10174 |
-+ else if (__data > (MAX)) \ |
10175 |
-+ __data = (MAX); \ |
10176 |
-+ if (__CONV) \ |
10177 |
-+ *(__PTR) = msecs_to_jiffies(__data); \ |
10178 |
-+ else \ |
10179 |
-+ *(__PTR) = __data; \ |
10180 |
-+ return ret; \ |
10181 |
-+} |
10182 |
-+STORE_FUNCTION(bfq_quantum_store, &bfqd->bfq_quantum, 1, INT_MAX, 0); |
10183 |
-+STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, |
10184 |
-+ INT_MAX, 1); |
10185 |
-+STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, |
10186 |
-+ INT_MAX, 1); |
10187 |
-+STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); |
10188 |
-+STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, |
10189 |
-+ INT_MAX, 0); |
10190 |
-+STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 1); |
10191 |
-+STORE_FUNCTION(bfq_max_budget_async_rq_store, &bfqd->bfq_max_budget_async_rq, |
10192 |
-+ 1, INT_MAX, 0); |
10193 |
-+STORE_FUNCTION(bfq_timeout_async_store, &bfqd->bfq_timeout[BLK_RW_ASYNC], 0, |
10194 |
-+ INT_MAX, 1); |
10195 |
-+STORE_FUNCTION(bfq_raising_coeff_store, &bfqd->bfq_raising_coeff, 1, |
10196 |
-+ INT_MAX, 0); |
10197 |
-+STORE_FUNCTION(bfq_raising_max_time_store, &bfqd->bfq_raising_max_time, 0, |
10198 |
-+ INT_MAX, 1); |
10199 |
-+STORE_FUNCTION(bfq_raising_rt_max_time_store, &bfqd->bfq_raising_rt_max_time, 0, |
10200 |
-+ INT_MAX, 1); |
10201 |
-+STORE_FUNCTION(bfq_raising_min_idle_time_store, |
10202 |
-+ &bfqd->bfq_raising_min_idle_time, 0, INT_MAX, 1); |
10203 |
-+STORE_FUNCTION(bfq_raising_min_inter_arr_async_store, |
10204 |
-+ &bfqd->bfq_raising_min_inter_arr_async, 0, INT_MAX, 1); |
10205 |
-+STORE_FUNCTION(bfq_raising_max_softrt_rate_store, |
10206 |
-+ &bfqd->bfq_raising_max_softrt_rate, 0, INT_MAX, 0); |
10207 |
-+#undef STORE_FUNCTION |
10208 |
-+ |
10209 |
-+/* do nothing for the moment */ |
10210 |
-+static ssize_t bfq_weights_store(struct elevator_queue *e, |
10211 |
-+ const char *page, size_t count) |
10212 |
-+{ |
10213 |
-+ return count; |
10214 |
-+} |
10215 |
-+ |
10216 |
-+static inline unsigned long bfq_estimated_max_budget(struct bfq_data *bfqd) |
10217 |
-+{ |
10218 |
-+ u64 timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
10219 |
-+ |
10220 |
-+ if (bfqd->peak_rate_samples >= BFQ_PEAK_RATE_SAMPLES) |
10221 |
-+ return bfq_calc_max_budget(bfqd->peak_rate, timeout); |
10222 |
-+ else |
10223 |
-+ return bfq_default_max_budget; |
10224 |
-+} |
10225 |
-+ |
10226 |
-+static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
10227 |
-+ const char *page, size_t count) |
10228 |
-+{ |
10229 |
-+ struct bfq_data *bfqd = e->elevator_data; |
10230 |
-+ unsigned long uninitialized_var(__data); |
10231 |
-+ int ret = bfq_var_store(&__data, (page), count); |
10232 |
-+ |
10233 |
-+ if (__data == 0) |
10234 |
-+ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
10235 |
-+ else { |
10236 |
-+ if (__data > INT_MAX) |
10237 |
-+ __data = INT_MAX; |
10238 |
-+ bfqd->bfq_max_budget = __data; |
10239 |
-+ } |
10240 |
-+ |
10241 |
-+ bfqd->bfq_user_max_budget = __data; |
10242 |
-+ |
10243 |
-+ return ret; |
10244 |
-+} |
10245 |
-+ |
10246 |
-+static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, |
10247 |
-+ const char *page, size_t count) |
10248 |
-+{ |
10249 |
-+ struct bfq_data *bfqd = e->elevator_data; |
10250 |
-+ unsigned long uninitialized_var(__data); |
10251 |
-+ int ret = bfq_var_store(&__data, (page), count); |
10252 |
-+ |
10253 |
-+ if (__data < 1) |
10254 |
-+ __data = 1; |
10255 |
-+ else if (__data > INT_MAX) |
10256 |
-+ __data = INT_MAX; |
10257 |
-+ |
10258 |
-+ bfqd->bfq_timeout[BLK_RW_SYNC] = msecs_to_jiffies(__data); |
10259 |
-+ if (bfqd->bfq_user_max_budget == 0) |
10260 |
-+ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
10261 |
-+ |
10262 |
-+ return ret; |
10263 |
-+} |
10264 |
-+ |
10265 |
-+static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
10266 |
-+ const char *page, size_t count) |
10267 |
-+{ |
10268 |
-+ struct bfq_data *bfqd = e->elevator_data; |
10269 |
-+ unsigned long uninitialized_var(__data); |
10270 |
-+ int ret = bfq_var_store(&__data, (page), count); |
10271 |
-+ |
10272 |
-+ if (__data > 1) |
10273 |
-+ __data = 1; |
10274 |
-+ if (__data == 0 && bfqd->low_latency != 0) |
10275 |
-+ bfq_end_raising(bfqd); |
10276 |
-+ bfqd->low_latency = __data; |
10277 |
-+ |
10278 |
-+ return ret; |
10279 |
-+} |
10280 |
-+ |
10281 |
-+#define BFQ_ATTR(name) \ |
10282 |
-+ __ATTR(name, S_IRUGO|S_IWUSR, bfq_##name##_show, bfq_##name##_store) |
10283 |
-+ |
10284 |
-+static struct elv_fs_entry bfq_attrs[] = { |
10285 |
-+ BFQ_ATTR(quantum), |
10286 |
-+ BFQ_ATTR(fifo_expire_sync), |
10287 |
-+ BFQ_ATTR(fifo_expire_async), |
10288 |
-+ BFQ_ATTR(back_seek_max), |
10289 |
-+ BFQ_ATTR(back_seek_penalty), |
10290 |
-+ BFQ_ATTR(slice_idle), |
10291 |
-+ BFQ_ATTR(max_budget), |
10292 |
-+ BFQ_ATTR(max_budget_async_rq), |
10293 |
-+ BFQ_ATTR(timeout_sync), |
10294 |
-+ BFQ_ATTR(timeout_async), |
10295 |
-+ BFQ_ATTR(low_latency), |
10296 |
-+ BFQ_ATTR(raising_coeff), |
10297 |
-+ BFQ_ATTR(raising_max_time), |
10298 |
-+ BFQ_ATTR(raising_rt_max_time), |
10299 |
-+ BFQ_ATTR(raising_min_idle_time), |
10300 |
-+ BFQ_ATTR(raising_min_inter_arr_async), |
10301 |
-+ BFQ_ATTR(raising_max_softrt_rate), |
10302 |
-+ BFQ_ATTR(weights), |
10303 |
-+ __ATTR_NULL |
10304 |
-+}; |
10305 |
-+ |
10306 |
-+static struct elevator_type iosched_bfq = { |
10307 |
-+ .ops = { |
10308 |
-+ .elevator_merge_fn = bfq_merge, |
10309 |
-+ .elevator_merged_fn = bfq_merged_request, |
10310 |
-+ .elevator_merge_req_fn = bfq_merged_requests, |
10311 |
-+ .elevator_allow_merge_fn = bfq_allow_merge, |
10312 |
-+ .elevator_dispatch_fn = bfq_dispatch_requests, |
10313 |
-+ .elevator_add_req_fn = bfq_insert_request, |
10314 |
-+ .elevator_activate_req_fn = bfq_activate_request, |
10315 |
-+ .elevator_deactivate_req_fn = bfq_deactivate_request, |
10316 |
-+ .elevator_completed_req_fn = bfq_completed_request, |
10317 |
-+ .elevator_former_req_fn = elv_rb_former_request, |
10318 |
-+ .elevator_latter_req_fn = elv_rb_latter_request, |
10319 |
-+ .elevator_init_icq_fn = bfq_init_icq, |
10320 |
-+ .elevator_exit_icq_fn = bfq_exit_icq, |
10321 |
-+ .elevator_set_req_fn = bfq_set_request, |
10322 |
-+ .elevator_put_req_fn = bfq_put_request, |
10323 |
-+ .elevator_may_queue_fn = bfq_may_queue, |
10324 |
-+ .elevator_init_fn = bfq_init_queue, |
10325 |
-+ .elevator_exit_fn = bfq_exit_queue, |
10326 |
-+ }, |
10327 |
-+ .icq_size = sizeof(struct bfq_io_cq), |
10328 |
-+ .icq_align = __alignof__(struct bfq_io_cq), |
10329 |
-+ .elevator_attrs = bfq_attrs, |
10330 |
-+ .elevator_name = "bfq", |
10331 |
-+ .elevator_owner = THIS_MODULE, |
10332 |
-+}; |
10333 |
-+ |
10334 |
-+static int __init bfq_init(void) |
10335 |
-+{ |
10336 |
-+ /* |
10337 |
-+ * Can be 0 on HZ < 1000 setups. |
10338 |
-+ */ |
10339 |
-+ if (bfq_slice_idle == 0) |
10340 |
-+ bfq_slice_idle = 1; |
10341 |
-+ |
10342 |
-+ if (bfq_timeout_async == 0) |
10343 |
-+ bfq_timeout_async = 1; |
10344 |
-+ |
10345 |
-+ if (bfq_slab_setup()) |
10346 |
-+ return -ENOMEM; |
10347 |
-+ |
10348 |
-+ elv_register(&iosched_bfq); |
10349 |
-+ |
10350 |
-+ return 0; |
10351 |
-+} |
10352 |
-+ |
10353 |
-+static void __exit bfq_exit(void) |
10354 |
-+{ |
10355 |
-+ elv_unregister(&iosched_bfq); |
10356 |
-+ bfq_slab_kill(); |
10357 |
-+} |
10358 |
-+ |
10359 |
-+module_init(bfq_init); |
10360 |
-+module_exit(bfq_exit); |
10361 |
-+ |
10362 |
-+MODULE_AUTHOR("Fabio Checconi, Paolo Valente"); |
10363 |
-+MODULE_LICENSE("GPL"); |
10364 |
-+MODULE_DESCRIPTION("Budget Fair Queueing IO scheduler"); |
10365 |
-diff --git a/block/bfq-sched.c b/block/bfq-sched.c |
10366 |
-new file mode 100644 |
10367 |
-index 0000000..03f8061 |
10368 |
---- /dev/null |
10369 |
-+++ b/block/bfq-sched.c |
10370 |
-@@ -0,0 +1,1072 @@ |
10371 |
-+/* |
10372 |
-+ * BFQ: Hierarchical B-WF2Q+ scheduler. |
10373 |
-+ * |
10374 |
-+ * Based on ideas and code from CFQ: |
10375 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
10376 |
-+ * |
10377 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
10378 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
10379 |
-+ * |
10380 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
10381 |
-+ */ |
10382 |
-+ |
10383 |
-+#ifdef CONFIG_CGROUP_BFQIO |
10384 |
-+#define for_each_entity(entity) \ |
10385 |
-+ for (; entity != NULL; entity = entity->parent) |
10386 |
-+ |
10387 |
-+#define for_each_entity_safe(entity, parent) \ |
10388 |
-+ for (; entity && ({ parent = entity->parent; 1; }); entity = parent) |
10389 |
-+ |
10390 |
-+static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
10391 |
-+ int extract, |
10392 |
-+ struct bfq_data *bfqd); |
10393 |
-+ |
10394 |
-+static inline void bfq_update_budget(struct bfq_entity *next_active) |
10395 |
-+{ |
10396 |
-+ struct bfq_entity *bfqg_entity; |
10397 |
-+ struct bfq_group *bfqg; |
10398 |
-+ struct bfq_sched_data *group_sd; |
10399 |
-+ |
10400 |
-+ BUG_ON(next_active == NULL); |
10401 |
-+ |
10402 |
-+ group_sd = next_active->sched_data; |
10403 |
-+ |
10404 |
-+ bfqg = container_of(group_sd, struct bfq_group, sched_data); |
10405 |
-+ /* |
10406 |
-+ * bfq_group's my_entity field is not NULL only if the group |
10407 |
-+ * is not the root group. We must not touch the root entity |
10408 |
-+ * as it must never become an active entity. |
10409 |
-+ */ |
10410 |
-+ bfqg_entity = bfqg->my_entity; |
10411 |
-+ if (bfqg_entity != NULL) |
10412 |
-+ bfqg_entity->budget = next_active->budget; |
10413 |
-+} |
10414 |
-+ |
10415 |
-+static int bfq_update_next_active(struct bfq_sched_data *sd) |
10416 |
-+{ |
10417 |
-+ struct bfq_entity *next_active; |
10418 |
-+ |
10419 |
-+ if (sd->active_entity != NULL) |
10420 |
-+ /* will update/requeue at the end of service */ |
10421 |
-+ return 0; |
10422 |
-+ |
10423 |
-+ /* |
10424 |
-+ * NOTE: this can be improved in many ways, such as returning |
10425 |
-+ * 1 (and thus propagating upwards the update) only when the |
10426 |
-+ * budget changes, or caching the bfqq that will be scheduled |
10427 |
-+ * next from this subtree. By now we worry more about |
10428 |
-+ * correctness than about performance... |
10429 |
-+ */ |
10430 |
-+ next_active = bfq_lookup_next_entity(sd, 0, NULL); |
10431 |
-+ sd->next_active = next_active; |
10432 |
-+ |
10433 |
-+ if (next_active != NULL) |
10434 |
-+ bfq_update_budget(next_active); |
10435 |
-+ |
10436 |
-+ return 1; |
10437 |
-+} |
10438 |
-+ |
10439 |
-+static inline void bfq_check_next_active(struct bfq_sched_data *sd, |
10440 |
-+ struct bfq_entity *entity) |
10441 |
-+{ |
10442 |
-+ BUG_ON(sd->next_active != entity); |
10443 |
-+} |
10444 |
-+#else |
10445 |
-+#define for_each_entity(entity) \ |
10446 |
-+ for (; entity != NULL; entity = NULL) |
10447 |
-+ |
10448 |
-+#define for_each_entity_safe(entity, parent) \ |
10449 |
-+ for (parent = NULL; entity != NULL; entity = parent) |
10450 |
-+ |
10451 |
-+static inline int bfq_update_next_active(struct bfq_sched_data *sd) |
10452 |
-+{ |
10453 |
-+ return 0; |
10454 |
-+} |
10455 |
-+ |
10456 |
-+static inline void bfq_check_next_active(struct bfq_sched_data *sd, |
10457 |
-+ struct bfq_entity *entity) |
10458 |
-+{ |
10459 |
-+} |
10460 |
-+ |
10461 |
-+static inline void bfq_update_budget(struct bfq_entity *next_active) |
10462 |
-+{ |
10463 |
-+} |
10464 |
-+#endif |
10465 |
-+ |
10466 |
-+/* |
10467 |
-+ * Shift for timestamp calculations. This actually limits the maximum |
10468 |
-+ * service allowed in one timestamp delta (small shift values increase it), |
10469 |
-+ * the maximum total weight that can be used for the queues in the system |
10470 |
-+ * (big shift values increase it), and the period of virtual time wraparounds. |
10471 |
-+ */ |
10472 |
-+#define WFQ_SERVICE_SHIFT 22 |
10473 |
-+ |
10474 |
-+/** |
10475 |
-+ * bfq_gt - compare two timestamps. |
10476 |
-+ * @a: first ts. |
10477 |
-+ * @b: second ts. |
10478 |
-+ * |
10479 |
-+ * Return @a > @b, dealing with wrapping correctly. |
10480 |
-+ */ |
10481 |
-+static inline int bfq_gt(u64 a, u64 b) |
10482 |
-+{ |
10483 |
-+ return (s64)(a - b) > 0; |
10484 |
-+} |
10485 |
-+ |
10486 |
-+static inline struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity) |
10487 |
-+{ |
10488 |
-+ struct bfq_queue *bfqq = NULL; |
10489 |
-+ |
10490 |
-+ BUG_ON(entity == NULL); |
10491 |
-+ |
10492 |
-+ if (entity->my_sched_data == NULL) |
10493 |
-+ bfqq = container_of(entity, struct bfq_queue, entity); |
10494 |
-+ |
10495 |
-+ return bfqq; |
10496 |
-+} |
10497 |
-+ |
10498 |
-+ |
10499 |
-+/** |
10500 |
-+ * bfq_delta - map service into the virtual time domain. |
10501 |
-+ * @service: amount of service. |
10502 |
-+ * @weight: scale factor (weight of an entity or weight sum). |
10503 |
-+ */ |
10504 |
-+static inline u64 bfq_delta(unsigned long service, |
10505 |
-+ unsigned long weight) |
10506 |
-+{ |
10507 |
-+ u64 d = (u64)service << WFQ_SERVICE_SHIFT; |
10508 |
-+ |
10509 |
-+ do_div(d, weight); |
10510 |
-+ return d; |
10511 |
-+} |
10512 |
-+ |
10513 |
-+/** |
10514 |
-+ * bfq_calc_finish - assign the finish time to an entity. |
10515 |
-+ * @entity: the entity to act upon. |
10516 |
-+ * @service: the service to be charged to the entity. |
10517 |
-+ */ |
10518 |
-+static inline void bfq_calc_finish(struct bfq_entity *entity, |
10519 |
-+ unsigned long service) |
10520 |
-+{ |
10521 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10522 |
-+ |
10523 |
-+ BUG_ON(entity->weight == 0); |
10524 |
-+ |
10525 |
-+ entity->finish = entity->start + |
10526 |
-+ bfq_delta(service, entity->weight); |
10527 |
-+ |
10528 |
-+ if (bfqq != NULL) { |
10529 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, |
10530 |
-+ "calc_finish: serv %lu, w %d", |
10531 |
-+ service, entity->weight); |
10532 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, |
10533 |
-+ "calc_finish: start %llu, finish %llu, delta %llu", |
10534 |
-+ entity->start, entity->finish, |
10535 |
-+ bfq_delta(service, entity->weight)); |
10536 |
-+ } |
10537 |
-+} |
10538 |
-+ |
10539 |
-+/** |
10540 |
-+ * bfq_entity_of - get an entity from a node. |
10541 |
-+ * @node: the node field of the entity. |
10542 |
-+ * |
10543 |
-+ * Convert a node pointer to the relative entity. This is used only |
10544 |
-+ * to simplify the logic of some functions and not as the generic |
10545 |
-+ * conversion mechanism because, e.g., in the tree walking functions, |
10546 |
-+ * the check for a %NULL value would be redundant. |
10547 |
-+ */ |
10548 |
-+static inline struct bfq_entity *bfq_entity_of(struct rb_node *node) |
10549 |
-+{ |
10550 |
-+ struct bfq_entity *entity = NULL; |
10551 |
-+ |
10552 |
-+ if (node != NULL) |
10553 |
-+ entity = rb_entry(node, struct bfq_entity, rb_node); |
10554 |
-+ |
10555 |
-+ return entity; |
10556 |
-+} |
10557 |
-+ |
10558 |
-+/** |
10559 |
-+ * bfq_extract - remove an entity from a tree. |
10560 |
-+ * @root: the tree root. |
10561 |
-+ * @entity: the entity to remove. |
10562 |
-+ */ |
10563 |
-+static inline void bfq_extract(struct rb_root *root, |
10564 |
-+ struct bfq_entity *entity) |
10565 |
-+{ |
10566 |
-+ BUG_ON(entity->tree != root); |
10567 |
-+ |
10568 |
-+ entity->tree = NULL; |
10569 |
-+ rb_erase(&entity->rb_node, root); |
10570 |
-+} |
10571 |
-+ |
10572 |
-+/** |
10573 |
-+ * bfq_idle_extract - extract an entity from the idle tree. |
10574 |
-+ * @st: the service tree of the owning @entity. |
10575 |
-+ * @entity: the entity being removed. |
10576 |
-+ */ |
10577 |
-+static void bfq_idle_extract(struct bfq_service_tree *st, |
10578 |
-+ struct bfq_entity *entity) |
10579 |
-+{ |
10580 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10581 |
-+ struct rb_node *next; |
10582 |
-+ |
10583 |
-+ BUG_ON(entity->tree != &st->idle); |
10584 |
-+ |
10585 |
-+ if (entity == st->first_idle) { |
10586 |
-+ next = rb_next(&entity->rb_node); |
10587 |
-+ st->first_idle = bfq_entity_of(next); |
10588 |
-+ } |
10589 |
-+ |
10590 |
-+ if (entity == st->last_idle) { |
10591 |
-+ next = rb_prev(&entity->rb_node); |
10592 |
-+ st->last_idle = bfq_entity_of(next); |
10593 |
-+ } |
10594 |
-+ |
10595 |
-+ bfq_extract(&st->idle, entity); |
10596 |
-+ |
10597 |
-+ if (bfqq != NULL) |
10598 |
-+ list_del(&bfqq->bfqq_list); |
10599 |
-+} |
10600 |
-+ |
10601 |
-+/** |
10602 |
-+ * bfq_insert - generic tree insertion. |
10603 |
-+ * @root: tree root. |
10604 |
-+ * @entity: entity to insert. |
10605 |
-+ * |
10606 |
-+ * This is used for the idle and the active tree, since they are both |
10607 |
-+ * ordered by finish time. |
10608 |
-+ */ |
10609 |
-+static void bfq_insert(struct rb_root *root, struct bfq_entity *entity) |
10610 |
-+{ |
10611 |
-+ struct bfq_entity *entry; |
10612 |
-+ struct rb_node **node = &root->rb_node; |
10613 |
-+ struct rb_node *parent = NULL; |
10614 |
-+ |
10615 |
-+ BUG_ON(entity->tree != NULL); |
10616 |
-+ |
10617 |
-+ while (*node != NULL) { |
10618 |
-+ parent = *node; |
10619 |
-+ entry = rb_entry(parent, struct bfq_entity, rb_node); |
10620 |
-+ |
10621 |
-+ if (bfq_gt(entry->finish, entity->finish)) |
10622 |
-+ node = &parent->rb_left; |
10623 |
-+ else |
10624 |
-+ node = &parent->rb_right; |
10625 |
-+ } |
10626 |
-+ |
10627 |
-+ rb_link_node(&entity->rb_node, parent, node); |
10628 |
-+ rb_insert_color(&entity->rb_node, root); |
10629 |
-+ |
10630 |
-+ entity->tree = root; |
10631 |
-+} |
10632 |
-+ |
10633 |
-+/** |
10634 |
-+ * bfq_update_min - update the min_start field of a entity. |
10635 |
-+ * @entity: the entity to update. |
10636 |
-+ * @node: one of its children. |
10637 |
-+ * |
10638 |
-+ * This function is called when @entity may store an invalid value for |
10639 |
-+ * min_start due to updates to the active tree. The function assumes |
10640 |
-+ * that the subtree rooted at @node (which may be its left or its right |
10641 |
-+ * child) has a valid min_start value. |
10642 |
-+ */ |
10643 |
-+static inline void bfq_update_min(struct bfq_entity *entity, |
10644 |
-+ struct rb_node *node) |
10645 |
-+{ |
10646 |
-+ struct bfq_entity *child; |
10647 |
-+ |
10648 |
-+ if (node != NULL) { |
10649 |
-+ child = rb_entry(node, struct bfq_entity, rb_node); |
10650 |
-+ if (bfq_gt(entity->min_start, child->min_start)) |
10651 |
-+ entity->min_start = child->min_start; |
10652 |
-+ } |
10653 |
-+} |
10654 |
-+ |
10655 |
-+/** |
10656 |
-+ * bfq_update_active_node - recalculate min_start. |
10657 |
-+ * @node: the node to update. |
10658 |
-+ * |
10659 |
-+ * @node may have changed position or one of its children may have moved, |
10660 |
-+ * this function updates its min_start value. The left and right subtrees |
10661 |
-+ * are assumed to hold a correct min_start value. |
10662 |
-+ */ |
10663 |
-+static inline void bfq_update_active_node(struct rb_node *node) |
10664 |
-+{ |
10665 |
-+ struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node); |
10666 |
-+ |
10667 |
-+ entity->min_start = entity->start; |
10668 |
-+ bfq_update_min(entity, node->rb_right); |
10669 |
-+ bfq_update_min(entity, node->rb_left); |
10670 |
-+} |
10671 |
-+ |
10672 |
-+/** |
10673 |
-+ * bfq_update_active_tree - update min_start for the whole active tree. |
10674 |
-+ * @node: the starting node. |
10675 |
-+ * |
10676 |
-+ * @node must be the deepest modified node after an update. This function |
10677 |
-+ * updates its min_start using the values held by its children, assuming |
10678 |
-+ * that they did not change, and then updates all the nodes that may have |
10679 |
-+ * changed in the path to the root. The only nodes that may have changed |
10680 |
-+ * are the ones in the path or their siblings. |
10681 |
-+ */ |
10682 |
-+static void bfq_update_active_tree(struct rb_node *node) |
10683 |
-+{ |
10684 |
-+ struct rb_node *parent; |
10685 |
-+ |
10686 |
-+up: |
10687 |
-+ bfq_update_active_node(node); |
10688 |
-+ |
10689 |
-+ parent = rb_parent(node); |
10690 |
-+ if (parent == NULL) |
10691 |
-+ return; |
10692 |
-+ |
10693 |
-+ if (node == parent->rb_left && parent->rb_right != NULL) |
10694 |
-+ bfq_update_active_node(parent->rb_right); |
10695 |
-+ else if (parent->rb_left != NULL) |
10696 |
-+ bfq_update_active_node(parent->rb_left); |
10697 |
-+ |
10698 |
-+ node = parent; |
10699 |
-+ goto up; |
10700 |
-+} |
10701 |
-+ |
10702 |
-+/** |
10703 |
-+ * bfq_active_insert - insert an entity in the active tree of its group/device. |
10704 |
-+ * @st: the service tree of the entity. |
10705 |
-+ * @entity: the entity being inserted. |
10706 |
-+ * |
10707 |
-+ * The active tree is ordered by finish time, but an extra key is kept |
10708 |
-+ * per each node, containing the minimum value for the start times of |
10709 |
-+ * its children (and the node itself), so it's possible to search for |
10710 |
-+ * the eligible node with the lowest finish time in logarithmic time. |
10711 |
-+ */ |
10712 |
-+static void bfq_active_insert(struct bfq_service_tree *st, |
10713 |
-+ struct bfq_entity *entity) |
10714 |
-+{ |
10715 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10716 |
-+ struct rb_node *node = &entity->rb_node; |
10717 |
-+ |
10718 |
-+ bfq_insert(&st->active, entity); |
10719 |
-+ |
10720 |
-+ if (node->rb_left != NULL) |
10721 |
-+ node = node->rb_left; |
10722 |
-+ else if (node->rb_right != NULL) |
10723 |
-+ node = node->rb_right; |
10724 |
-+ |
10725 |
-+ bfq_update_active_tree(node); |
10726 |
-+ |
10727 |
-+ if (bfqq != NULL) |
10728 |
-+ list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list); |
10729 |
-+} |
10730 |
-+ |
10731 |
-+/** |
10732 |
-+ * bfq_ioprio_to_weight - calc a weight from an ioprio. |
10733 |
-+ * @ioprio: the ioprio value to convert. |
10734 |
-+ */ |
10735 |
-+static unsigned short bfq_ioprio_to_weight(int ioprio) |
10736 |
-+{ |
10737 |
-+ WARN_ON(ioprio < 0 || ioprio >= IOPRIO_BE_NR); |
10738 |
-+ return IOPRIO_BE_NR - ioprio; |
10739 |
-+} |
10740 |
-+ |
10741 |
-+/** |
10742 |
-+ * bfq_weight_to_ioprio - calc an ioprio from a weight. |
10743 |
-+ * @weight: the weight value to convert. |
10744 |
-+ * |
10745 |
-+ * To preserve as mush as possible the old only-ioprio user interface, |
10746 |
-+ * 0 is used as an escape ioprio value for weights (numerically) equal or |
10747 |
-+ * larger than IOPRIO_BE_NR |
10748 |
-+ */ |
10749 |
-+static unsigned short bfq_weight_to_ioprio(int weight) |
10750 |
-+{ |
10751 |
-+ WARN_ON(weight < BFQ_MIN_WEIGHT || weight > BFQ_MAX_WEIGHT); |
10752 |
-+ return IOPRIO_BE_NR - weight < 0 ? 0 : IOPRIO_BE_NR - weight; |
10753 |
-+} |
10754 |
-+ |
10755 |
-+static inline void bfq_get_entity(struct bfq_entity *entity) |
10756 |
-+{ |
10757 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10758 |
-+ struct bfq_sched_data *sd; |
10759 |
-+ |
10760 |
-+ if (bfqq != NULL) { |
10761 |
-+ sd = entity->sched_data; |
10762 |
-+ atomic_inc(&bfqq->ref); |
10763 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d", |
10764 |
-+ bfqq, atomic_read(&bfqq->ref)); |
10765 |
-+ } |
10766 |
-+} |
10767 |
-+ |
10768 |
-+/** |
10769 |
-+ * bfq_find_deepest - find the deepest node that an extraction can modify. |
10770 |
-+ * @node: the node being removed. |
10771 |
-+ * |
10772 |
-+ * Do the first step of an extraction in an rb tree, looking for the |
10773 |
-+ * node that will replace @node, and returning the deepest node that |
10774 |
-+ * the following modifications to the tree can touch. If @node is the |
10775 |
-+ * last node in the tree return %NULL. |
10776 |
-+ */ |
10777 |
-+static struct rb_node *bfq_find_deepest(struct rb_node *node) |
10778 |
-+{ |
10779 |
-+ struct rb_node *deepest; |
10780 |
-+ |
10781 |
-+ if (node->rb_right == NULL && node->rb_left == NULL) |
10782 |
-+ deepest = rb_parent(node); |
10783 |
-+ else if (node->rb_right == NULL) |
10784 |
-+ deepest = node->rb_left; |
10785 |
-+ else if (node->rb_left == NULL) |
10786 |
-+ deepest = node->rb_right; |
10787 |
-+ else { |
10788 |
-+ deepest = rb_next(node); |
10789 |
-+ if (deepest->rb_right != NULL) |
10790 |
-+ deepest = deepest->rb_right; |
10791 |
-+ else if (rb_parent(deepest) != node) |
10792 |
-+ deepest = rb_parent(deepest); |
10793 |
-+ } |
10794 |
-+ |
10795 |
-+ return deepest; |
10796 |
-+} |
10797 |
-+ |
10798 |
-+/** |
10799 |
-+ * bfq_active_extract - remove an entity from the active tree. |
10800 |
-+ * @st: the service_tree containing the tree. |
10801 |
-+ * @entity: the entity being removed. |
10802 |
-+ */ |
10803 |
-+static void bfq_active_extract(struct bfq_service_tree *st, |
10804 |
-+ struct bfq_entity *entity) |
10805 |
-+{ |
10806 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10807 |
-+ struct rb_node *node; |
10808 |
-+ |
10809 |
-+ node = bfq_find_deepest(&entity->rb_node); |
10810 |
-+ bfq_extract(&st->active, entity); |
10811 |
-+ |
10812 |
-+ if (node != NULL) |
10813 |
-+ bfq_update_active_tree(node); |
10814 |
-+ |
10815 |
-+ if (bfqq != NULL) |
10816 |
-+ list_del(&bfqq->bfqq_list); |
10817 |
-+} |
10818 |
-+ |
10819 |
-+/** |
10820 |
-+ * bfq_idle_insert - insert an entity into the idle tree. |
10821 |
-+ * @st: the service tree containing the tree. |
10822 |
-+ * @entity: the entity to insert. |
10823 |
-+ */ |
10824 |
-+static void bfq_idle_insert(struct bfq_service_tree *st, |
10825 |
-+ struct bfq_entity *entity) |
10826 |
-+{ |
10827 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10828 |
-+ struct bfq_entity *first_idle = st->first_idle; |
10829 |
-+ struct bfq_entity *last_idle = st->last_idle; |
10830 |
-+ |
10831 |
-+ if (first_idle == NULL || bfq_gt(first_idle->finish, entity->finish)) |
10832 |
-+ st->first_idle = entity; |
10833 |
-+ if (last_idle == NULL || bfq_gt(entity->finish, last_idle->finish)) |
10834 |
-+ st->last_idle = entity; |
10835 |
-+ |
10836 |
-+ bfq_insert(&st->idle, entity); |
10837 |
-+ |
10838 |
-+ if (bfqq != NULL) |
10839 |
-+ list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list); |
10840 |
-+} |
10841 |
-+ |
10842 |
-+/** |
10843 |
-+ * bfq_forget_entity - remove an entity from the wfq trees. |
10844 |
-+ * @st: the service tree. |
10845 |
-+ * @entity: the entity being removed. |
10846 |
-+ * |
10847 |
-+ * Update the device status and forget everything about @entity, putting |
10848 |
-+ * the device reference to it, if it is a queue. Entities belonging to |
10849 |
-+ * groups are not refcounted. |
10850 |
-+ */ |
10851 |
-+static void bfq_forget_entity(struct bfq_service_tree *st, |
10852 |
-+ struct bfq_entity *entity) |
10853 |
-+{ |
10854 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10855 |
-+ struct bfq_sched_data *sd; |
10856 |
-+ |
10857 |
-+ BUG_ON(!entity->on_st); |
10858 |
-+ |
10859 |
-+ entity->on_st = 0; |
10860 |
-+ st->wsum -= entity->weight; |
10861 |
-+ if (bfqq != NULL) { |
10862 |
-+ sd = entity->sched_data; |
10863 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "forget_entity: %p %d", |
10864 |
-+ bfqq, atomic_read(&bfqq->ref)); |
10865 |
-+ bfq_put_queue(bfqq); |
10866 |
-+ } |
10867 |
-+} |
10868 |
-+ |
10869 |
-+/** |
10870 |
-+ * bfq_put_idle_entity - release the idle tree ref of an entity. |
10871 |
-+ * @st: service tree for the entity. |
10872 |
-+ * @entity: the entity being released. |
10873 |
-+ */ |
10874 |
-+static void bfq_put_idle_entity(struct bfq_service_tree *st, |
10875 |
-+ struct bfq_entity *entity) |
10876 |
-+{ |
10877 |
-+ bfq_idle_extract(st, entity); |
10878 |
-+ bfq_forget_entity(st, entity); |
10879 |
-+} |
10880 |
-+ |
10881 |
-+/** |
10882 |
-+ * bfq_forget_idle - update the idle tree if necessary. |
10883 |
-+ * @st: the service tree to act upon. |
10884 |
-+ * |
10885 |
-+ * To preserve the global O(log N) complexity we only remove one entry here; |
10886 |
-+ * as the idle tree will not grow indefinitely this can be done safely. |
10887 |
-+ */ |
10888 |
-+static void bfq_forget_idle(struct bfq_service_tree *st) |
10889 |
-+{ |
10890 |
-+ struct bfq_entity *first_idle = st->first_idle; |
10891 |
-+ struct bfq_entity *last_idle = st->last_idle; |
10892 |
-+ |
10893 |
-+ if (RB_EMPTY_ROOT(&st->active) && last_idle != NULL && |
10894 |
-+ !bfq_gt(last_idle->finish, st->vtime)) { |
10895 |
-+ /* |
10896 |
-+ * Forget the whole idle tree, increasing the vtime past |
10897 |
-+ * the last finish time of idle entities. |
10898 |
-+ */ |
10899 |
-+ st->vtime = last_idle->finish; |
10900 |
-+ } |
10901 |
-+ |
10902 |
-+ if (first_idle != NULL && !bfq_gt(first_idle->finish, st->vtime)) |
10903 |
-+ bfq_put_idle_entity(st, first_idle); |
10904 |
-+} |
10905 |
-+ |
10906 |
-+static struct bfq_service_tree * |
10907 |
-+__bfq_entity_update_weight_prio(struct bfq_service_tree *old_st, |
10908 |
-+ struct bfq_entity *entity) |
10909 |
-+{ |
10910 |
-+ struct bfq_service_tree *new_st = old_st; |
10911 |
-+ |
10912 |
-+ if (entity->ioprio_changed) { |
10913 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10914 |
-+ |
10915 |
-+ BUG_ON(old_st->wsum < entity->weight); |
10916 |
-+ old_st->wsum -= entity->weight; |
10917 |
-+ |
10918 |
-+ if (entity->new_weight != entity->orig_weight) { |
10919 |
-+ entity->orig_weight = entity->new_weight; |
10920 |
-+ entity->ioprio = |
10921 |
-+ bfq_weight_to_ioprio(entity->orig_weight); |
10922 |
-+ } else if (entity->new_ioprio != entity->ioprio) { |
10923 |
-+ entity->ioprio = entity->new_ioprio; |
10924 |
-+ entity->orig_weight = |
10925 |
-+ bfq_ioprio_to_weight(entity->ioprio); |
10926 |
-+ } else |
10927 |
-+ entity->new_weight = entity->orig_weight = |
10928 |
-+ bfq_ioprio_to_weight(entity->ioprio); |
10929 |
-+ |
10930 |
-+ entity->ioprio_class = entity->new_ioprio_class; |
10931 |
-+ entity->ioprio_changed = 0; |
10932 |
-+ |
10933 |
-+ /* |
10934 |
-+ * NOTE: here we may be changing the weight too early, |
10935 |
-+ * this will cause unfairness. The correct approach |
10936 |
-+ * would have required additional complexity to defer |
10937 |
-+ * weight changes to the proper time instants (i.e., |
10938 |
-+ * when entity->finish <= old_st->vtime). |
10939 |
-+ */ |
10940 |
-+ new_st = bfq_entity_service_tree(entity); |
10941 |
-+ entity->weight = entity->orig_weight * |
10942 |
-+ (bfqq != NULL ? bfqq->raising_coeff : 1); |
10943 |
-+ new_st->wsum += entity->weight; |
10944 |
-+ |
10945 |
-+ if (new_st != old_st) |
10946 |
-+ entity->start = new_st->vtime; |
10947 |
-+ } |
10948 |
-+ |
10949 |
-+ return new_st; |
10950 |
-+} |
10951 |
-+ |
10952 |
-+/** |
10953 |
-+ * bfq_bfqq_served - update the scheduler status after selection for service. |
10954 |
-+ * @bfqq: the queue being served. |
10955 |
-+ * @served: bytes to transfer. |
10956 |
-+ * |
10957 |
-+ * NOTE: this can be optimized, as the timestamps of upper level entities |
10958 |
-+ * are synchronized every time a new bfqq is selected for service. By now, |
10959 |
-+ * we keep it to better check consistency. |
10960 |
-+ */ |
10961 |
-+static void bfq_bfqq_served(struct bfq_queue *bfqq, unsigned long served) |
10962 |
-+{ |
10963 |
-+ struct bfq_entity *entity = &bfqq->entity; |
10964 |
-+ struct bfq_service_tree *st; |
10965 |
-+ |
10966 |
-+ for_each_entity(entity) { |
10967 |
-+ st = bfq_entity_service_tree(entity); |
10968 |
-+ |
10969 |
-+ entity->service += served; |
10970 |
-+ BUG_ON(entity->service > entity->budget); |
10971 |
-+ BUG_ON(st->wsum == 0); |
10972 |
-+ |
10973 |
-+ st->vtime += bfq_delta(served, st->wsum); |
10974 |
-+ bfq_forget_idle(st); |
10975 |
-+ } |
10976 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %lu secs", served); |
10977 |
-+} |
10978 |
-+ |
10979 |
-+/** |
10980 |
-+ * bfq_bfqq_charge_full_budget - set the service to the entity budget. |
10981 |
-+ * @bfqq: the queue that needs a service update. |
10982 |
-+ * |
10983 |
-+ * When it's not possible to be fair in the service domain, because |
10984 |
-+ * a queue is not consuming its budget fast enough (the meaning of |
10985 |
-+ * fast depends on the timeout parameter), we charge it a full |
10986 |
-+ * budget. In this way we should obtain a sort of time-domain |
10987 |
-+ * fairness among all the seeky/slow queues. |
10988 |
-+ */ |
10989 |
-+static inline void bfq_bfqq_charge_full_budget(struct bfq_queue *bfqq) |
10990 |
-+{ |
10991 |
-+ struct bfq_entity *entity = &bfqq->entity; |
10992 |
-+ |
10993 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "charge_full_budget"); |
10994 |
-+ |
10995 |
-+ bfq_bfqq_served(bfqq, entity->budget - entity->service); |
10996 |
-+} |
10997 |
-+ |
10998 |
-+/** |
10999 |
-+ * __bfq_activate_entity - activate an entity. |
11000 |
-+ * @entity: the entity being activated. |
11001 |
-+ * |
11002 |
-+ * Called whenever an entity is activated, i.e., it is not active and one |
11003 |
-+ * of its children receives a new request, or has to be reactivated due to |
11004 |
-+ * budget exhaustion. It uses the current budget of the entity (and the |
11005 |
-+ * service received if @entity is active) of the queue to calculate its |
11006 |
-+ * timestamps. |
11007 |
-+ */ |
11008 |
-+static void __bfq_activate_entity(struct bfq_entity *entity) |
11009 |
-+{ |
11010 |
-+ struct bfq_sched_data *sd = entity->sched_data; |
11011 |
-+ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
11012 |
-+ |
11013 |
-+ if (entity == sd->active_entity) { |
11014 |
-+ BUG_ON(entity->tree != NULL); |
11015 |
-+ /* |
11016 |
-+ * If we are requeueing the current entity we have |
11017 |
-+ * to take care of not charging to it service it has |
11018 |
-+ * not received. |
11019 |
-+ */ |
11020 |
-+ bfq_calc_finish(entity, entity->service); |
11021 |
-+ entity->start = entity->finish; |
11022 |
-+ sd->active_entity = NULL; |
11023 |
-+ } else if (entity->tree == &st->active) { |
11024 |
-+ /* |
11025 |
-+ * Requeueing an entity due to a change of some |
11026 |
-+ * next_active entity below it. We reuse the old |
11027 |
-+ * start time. |
11028 |
-+ */ |
11029 |
-+ bfq_active_extract(st, entity); |
11030 |
-+ } else if (entity->tree == &st->idle) { |
11031 |
-+ /* |
11032 |
-+ * Must be on the idle tree, bfq_idle_extract() will |
11033 |
-+ * check for that. |
11034 |
-+ */ |
11035 |
-+ bfq_idle_extract(st, entity); |
11036 |
-+ entity->start = bfq_gt(st->vtime, entity->finish) ? |
11037 |
-+ st->vtime : entity->finish; |
11038 |
-+ } else { |
11039 |
-+ /* |
11040 |
-+ * The finish time of the entity may be invalid, and |
11041 |
-+ * it is in the past for sure, otherwise the queue |
11042 |
-+ * would have been on the idle tree. |
11043 |
-+ */ |
11044 |
-+ entity->start = st->vtime; |
11045 |
-+ st->wsum += entity->weight; |
11046 |
-+ bfq_get_entity(entity); |
11047 |
-+ |
11048 |
-+ BUG_ON(entity->on_st); |
11049 |
-+ entity->on_st = 1; |
11050 |
-+ } |
11051 |
-+ |
11052 |
-+ st = __bfq_entity_update_weight_prio(st, entity); |
11053 |
-+ bfq_calc_finish(entity, entity->budget); |
11054 |
-+ bfq_active_insert(st, entity); |
11055 |
-+} |
11056 |
-+ |
11057 |
-+/** |
11058 |
-+ * bfq_activate_entity - activate an entity and its ancestors if necessary. |
11059 |
-+ * @entity: the entity to activate. |
11060 |
-+ * |
11061 |
-+ * Activate @entity and all the entities on the path from it to the root. |
11062 |
-+ */ |
11063 |
-+static void bfq_activate_entity(struct bfq_entity *entity) |
11064 |
-+{ |
11065 |
-+ struct bfq_sched_data *sd; |
11066 |
-+ |
11067 |
-+ for_each_entity(entity) { |
11068 |
-+ __bfq_activate_entity(entity); |
11069 |
-+ |
11070 |
-+ sd = entity->sched_data; |
11071 |
-+ if (!bfq_update_next_active(sd)) |
11072 |
-+ /* |
11073 |
-+ * No need to propagate the activation to the |
11074 |
-+ * upper entities, as they will be updated when |
11075 |
-+ * the active entity is rescheduled. |
11076 |
-+ */ |
11077 |
-+ break; |
11078 |
-+ } |
11079 |
-+} |
11080 |
-+ |
11081 |
-+/** |
11082 |
-+ * __bfq_deactivate_entity - deactivate an entity from its service tree. |
11083 |
-+ * @entity: the entity to deactivate. |
11084 |
-+ * @requeue: if false, the entity will not be put into the idle tree. |
11085 |
-+ * |
11086 |
-+ * Deactivate an entity, independently from its previous state. If the |
11087 |
-+ * entity was not on a service tree just return, otherwise if it is on |
11088 |
-+ * any scheduler tree, extract it from that tree, and if necessary |
11089 |
-+ * and if the caller did not specify @requeue, put it on the idle tree. |
11090 |
-+ * |
11091 |
-+ * Return %1 if the caller should update the entity hierarchy, i.e., |
11092 |
-+ * if the entity was under service or if it was the next_active for |
11093 |
-+ * its sched_data; return %0 otherwise. |
11094 |
-+ */ |
11095 |
-+static int __bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
11096 |
-+{ |
11097 |
-+ struct bfq_sched_data *sd = entity->sched_data; |
11098 |
-+ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
11099 |
-+ int was_active = entity == sd->active_entity; |
11100 |
-+ int ret = 0; |
11101 |
-+ |
11102 |
-+ if (!entity->on_st) |
11103 |
-+ return 0; |
11104 |
-+ |
11105 |
-+ BUG_ON(was_active && entity->tree != NULL); |
11106 |
-+ |
11107 |
-+ if (was_active) { |
11108 |
-+ bfq_calc_finish(entity, entity->service); |
11109 |
-+ sd->active_entity = NULL; |
11110 |
-+ } else if (entity->tree == &st->active) |
11111 |
-+ bfq_active_extract(st, entity); |
11112 |
-+ else if (entity->tree == &st->idle) |
11113 |
-+ bfq_idle_extract(st, entity); |
11114 |
-+ else if (entity->tree != NULL) |
11115 |
-+ BUG(); |
11116 |
-+ |
11117 |
-+ if (was_active || sd->next_active == entity) |
11118 |
-+ ret = bfq_update_next_active(sd); |
11119 |
-+ |
11120 |
-+ if (!requeue || !bfq_gt(entity->finish, st->vtime)) |
11121 |
-+ bfq_forget_entity(st, entity); |
11122 |
-+ else |
11123 |
-+ bfq_idle_insert(st, entity); |
11124 |
-+ |
11125 |
-+ BUG_ON(sd->active_entity == entity); |
11126 |
-+ BUG_ON(sd->next_active == entity); |
11127 |
-+ |
11128 |
-+ return ret; |
11129 |
-+} |
11130 |
-+ |
11131 |
-+/** |
11132 |
-+ * bfq_deactivate_entity - deactivate an entity. |
11133 |
-+ * @entity: the entity to deactivate. |
11134 |
-+ * @requeue: true if the entity can be put on the idle tree |
11135 |
-+ */ |
11136 |
-+static void bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
11137 |
-+{ |
11138 |
-+ struct bfq_sched_data *sd; |
11139 |
-+ struct bfq_entity *parent; |
11140 |
-+ |
11141 |
-+ for_each_entity_safe(entity, parent) { |
11142 |
-+ sd = entity->sched_data; |
11143 |
-+ |
11144 |
-+ if (!__bfq_deactivate_entity(entity, requeue)) |
11145 |
-+ /* |
11146 |
-+ * The parent entity is still backlogged, and |
11147 |
-+ * we don't need to update it as it is still |
11148 |
-+ * under service. |
11149 |
-+ */ |
11150 |
-+ break; |
11151 |
-+ |
11152 |
-+ if (sd->next_active != NULL) |
11153 |
-+ /* |
11154 |
-+ * The parent entity is still backlogged and |
11155 |
-+ * the budgets on the path towards the root |
11156 |
-+ * need to be updated. |
11157 |
-+ */ |
11158 |
-+ goto update; |
11159 |
-+ |
11160 |
-+ /* |
11161 |
-+ * If we reach there the parent is no more backlogged and |
11162 |
-+ * we want to propagate the dequeue upwards. |
11163 |
-+ */ |
11164 |
-+ requeue = 1; |
11165 |
-+ } |
11166 |
-+ |
11167 |
-+ return; |
11168 |
-+ |
11169 |
-+update: |
11170 |
-+ entity = parent; |
11171 |
-+ for_each_entity(entity) { |
11172 |
-+ __bfq_activate_entity(entity); |
11173 |
-+ |
11174 |
-+ sd = entity->sched_data; |
11175 |
-+ if (!bfq_update_next_active(sd)) |
11176 |
-+ break; |
11177 |
-+ } |
11178 |
-+} |
11179 |
-+ |
11180 |
-+/** |
11181 |
-+ * bfq_update_vtime - update vtime if necessary. |
11182 |
-+ * @st: the service tree to act upon. |
11183 |
-+ * |
11184 |
-+ * If necessary update the service tree vtime to have at least one |
11185 |
-+ * eligible entity, skipping to its start time. Assumes that the |
11186 |
-+ * active tree of the device is not empty. |
11187 |
-+ * |
11188 |
-+ * NOTE: this hierarchical implementation updates vtimes quite often, |
11189 |
-+ * we may end up with reactivated tasks getting timestamps after a |
11190 |
-+ * vtime skip done because we needed a ->first_active entity on some |
11191 |
-+ * intermediate node. |
11192 |
-+ */ |
11193 |
-+static void bfq_update_vtime(struct bfq_service_tree *st) |
11194 |
-+{ |
11195 |
-+ struct bfq_entity *entry; |
11196 |
-+ struct rb_node *node = st->active.rb_node; |
11197 |
-+ |
11198 |
-+ entry = rb_entry(node, struct bfq_entity, rb_node); |
11199 |
-+ if (bfq_gt(entry->min_start, st->vtime)) { |
11200 |
-+ st->vtime = entry->min_start; |
11201 |
-+ bfq_forget_idle(st); |
11202 |
-+ } |
11203 |
-+} |
11204 |
-+ |
11205 |
-+/** |
11206 |
-+ * bfq_first_active - find the eligible entity with the smallest finish time |
11207 |
-+ * @st: the service tree to select from. |
11208 |
-+ * |
11209 |
-+ * This function searches the first schedulable entity, starting from the |
11210 |
-+ * root of the tree and going on the left every time on this side there is |
11211 |
-+ * a subtree with at least one eligible (start >= vtime) entity. The path |
11212 |
-+ * on the right is followed only if a) the left subtree contains no eligible |
11213 |
-+ * entities and b) no eligible entity has been found yet. |
11214 |
-+ */ |
11215 |
-+static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st) |
11216 |
-+{ |
11217 |
-+ struct bfq_entity *entry, *first = NULL; |
11218 |
-+ struct rb_node *node = st->active.rb_node; |
11219 |
-+ |
11220 |
-+ while (node != NULL) { |
11221 |
-+ entry = rb_entry(node, struct bfq_entity, rb_node); |
11222 |
-+left: |
11223 |
-+ if (!bfq_gt(entry->start, st->vtime)) |
11224 |
-+ first = entry; |
11225 |
-+ |
11226 |
-+ BUG_ON(bfq_gt(entry->min_start, st->vtime)); |
11227 |
-+ |
11228 |
-+ if (node->rb_left != NULL) { |
11229 |
-+ entry = rb_entry(node->rb_left, |
11230 |
-+ struct bfq_entity, rb_node); |
11231 |
-+ if (!bfq_gt(entry->min_start, st->vtime)) { |
11232 |
-+ node = node->rb_left; |
11233 |
-+ goto left; |
11234 |
-+ } |
11235 |
-+ } |
11236 |
-+ if (first != NULL) |
11237 |
-+ break; |
11238 |
-+ node = node->rb_right; |
11239 |
-+ } |
11240 |
-+ |
11241 |
-+ BUG_ON(first == NULL && !RB_EMPTY_ROOT(&st->active)); |
11242 |
-+ return first; |
11243 |
-+} |
11244 |
-+ |
11245 |
-+/** |
11246 |
-+ * __bfq_lookup_next_entity - return the first eligible entity in @st. |
11247 |
-+ * @st: the service tree. |
11248 |
-+ * |
11249 |
-+ * Update the virtual time in @st and return the first eligible entity |
11250 |
-+ * it contains. |
11251 |
-+ */ |
11252 |
-+static struct bfq_entity *__bfq_lookup_next_entity(struct bfq_service_tree *st, |
11253 |
-+ bool force) |
11254 |
-+{ |
11255 |
-+ struct bfq_entity *entity, *new_next_active = NULL; |
11256 |
-+ |
11257 |
-+ if (RB_EMPTY_ROOT(&st->active)) |
11258 |
-+ return NULL; |
11259 |
-+ |
11260 |
-+ bfq_update_vtime(st); |
11261 |
-+ entity = bfq_first_active_entity(st); |
11262 |
-+ BUG_ON(bfq_gt(entity->start, st->vtime)); |
11263 |
-+ |
11264 |
-+ /* |
11265 |
-+ * If the chosen entity does not match with the sched_data's |
11266 |
-+ * next_active and we are forcedly serving the IDLE priority |
11267 |
-+ * class tree, bubble up budget update. |
11268 |
-+ */ |
11269 |
-+ if (unlikely(force && entity != entity->sched_data->next_active)) { |
11270 |
-+ new_next_active = entity; |
11271 |
-+ for_each_entity(new_next_active) |
11272 |
-+ bfq_update_budget(new_next_active); |
11273 |
-+ } |
11274 |
-+ |
11275 |
-+ return entity; |
11276 |
-+} |
11277 |
-+ |
11278 |
-+/** |
11279 |
-+ * bfq_lookup_next_entity - return the first eligible entity in @sd. |
11280 |
-+ * @sd: the sched_data. |
11281 |
-+ * @extract: if true the returned entity will be also extracted from @sd. |
11282 |
-+ * |
11283 |
-+ * NOTE: since we cache the next_active entity at each level of the |
11284 |
-+ * hierarchy, the complexity of the lookup can be decreased with |
11285 |
-+ * absolutely no effort just returning the cached next_active value; |
11286 |
-+ * we prefer to do full lookups to test the consistency of * the data |
11287 |
-+ * structures. |
11288 |
-+ */ |
11289 |
-+static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
11290 |
-+ int extract, |
11291 |
-+ struct bfq_data *bfqd) |
11292 |
-+{ |
11293 |
-+ struct bfq_service_tree *st = sd->service_tree; |
11294 |
-+ struct bfq_entity *entity; |
11295 |
-+ int i=0; |
11296 |
-+ |
11297 |
-+ BUG_ON(sd->active_entity != NULL); |
11298 |
-+ |
11299 |
-+ if (bfqd != NULL && |
11300 |
-+ jiffies - bfqd->bfq_class_idle_last_service > BFQ_CL_IDLE_TIMEOUT) { |
11301 |
-+ entity = __bfq_lookup_next_entity(st + BFQ_IOPRIO_CLASSES - 1, true); |
11302 |
-+ if (entity != NULL) { |
11303 |
-+ i = BFQ_IOPRIO_CLASSES - 1; |
11304 |
-+ bfqd->bfq_class_idle_last_service = jiffies; |
11305 |
-+ sd->next_active = entity; |
11306 |
-+ } |
11307 |
-+ } |
11308 |
-+ for (; i < BFQ_IOPRIO_CLASSES; i++) { |
11309 |
-+ entity = __bfq_lookup_next_entity(st + i, false); |
11310 |
-+ if (entity != NULL) { |
11311 |
-+ if (extract) { |
11312 |
-+ bfq_check_next_active(sd, entity); |
11313 |
-+ bfq_active_extract(st + i, entity); |
11314 |
-+ sd->active_entity = entity; |
11315 |
-+ sd->next_active = NULL; |
11316 |
-+ } |
11317 |
-+ break; |
11318 |
-+ } |
11319 |
-+ } |
11320 |
-+ |
11321 |
-+ return entity; |
11322 |
-+} |
11323 |
-+ |
11324 |
-+/* |
11325 |
-+ * Get next queue for service. |
11326 |
-+ */ |
11327 |
-+static struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd) |
11328 |
-+{ |
11329 |
-+ struct bfq_entity *entity = NULL; |
11330 |
-+ struct bfq_sched_data *sd; |
11331 |
-+ struct bfq_queue *bfqq; |
11332 |
-+ |
11333 |
-+ BUG_ON(bfqd->active_queue != NULL); |
11334 |
-+ |
11335 |
-+ if (bfqd->busy_queues == 0) |
11336 |
-+ return NULL; |
11337 |
-+ |
11338 |
-+ sd = &bfqd->root_group->sched_data; |
11339 |
-+ for (; sd != NULL; sd = entity->my_sched_data) { |
11340 |
-+ entity = bfq_lookup_next_entity(sd, 1, bfqd); |
11341 |
-+ BUG_ON(entity == NULL); |
11342 |
-+ entity->service = 0; |
11343 |
-+ } |
11344 |
-+ |
11345 |
-+ bfqq = bfq_entity_to_bfqq(entity); |
11346 |
-+ BUG_ON(bfqq == NULL); |
11347 |
-+ |
11348 |
-+ return bfqq; |
11349 |
-+} |
11350 |
-+ |
11351 |
-+/* |
11352 |
-+ * Forced extraction of the given queue. |
11353 |
-+ */ |
11354 |
-+static void bfq_get_next_queue_forced(struct bfq_data *bfqd, |
11355 |
-+ struct bfq_queue *bfqq) |
11356 |
-+{ |
11357 |
-+ struct bfq_entity *entity; |
11358 |
-+ struct bfq_sched_data *sd; |
11359 |
-+ |
11360 |
-+ BUG_ON(bfqd->active_queue != NULL); |
11361 |
-+ |
11362 |
-+ entity = &bfqq->entity; |
11363 |
-+ /* |
11364 |
-+ * Bubble up extraction/update from the leaf to the root. |
11365 |
-+ */ |
11366 |
-+ for_each_entity(entity) { |
11367 |
-+ sd = entity->sched_data; |
11368 |
-+ bfq_update_budget(entity); |
11369 |
-+ bfq_update_vtime(bfq_entity_service_tree(entity)); |
11370 |
-+ bfq_active_extract(bfq_entity_service_tree(entity), entity); |
11371 |
-+ sd->active_entity = entity; |
11372 |
-+ sd->next_active = NULL; |
11373 |
-+ entity->service = 0; |
11374 |
-+ } |
11375 |
-+ |
11376 |
-+ return; |
11377 |
-+} |
11378 |
-+ |
11379 |
-+static void __bfq_bfqd_reset_active(struct bfq_data *bfqd) |
11380 |
-+{ |
11381 |
-+ if (bfqd->active_bic != NULL) { |
11382 |
-+ put_io_context(bfqd->active_bic->icq.ioc); |
11383 |
-+ bfqd->active_bic = NULL; |
11384 |
-+ } |
11385 |
-+ |
11386 |
-+ bfqd->active_queue = NULL; |
11387 |
-+ del_timer(&bfqd->idle_slice_timer); |
11388 |
-+} |
11389 |
-+ |
11390 |
-+static void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
11391 |
-+ int requeue) |
11392 |
-+{ |
11393 |
-+ struct bfq_entity *entity = &bfqq->entity; |
11394 |
-+ |
11395 |
-+ if (bfqq == bfqd->active_queue) |
11396 |
-+ __bfq_bfqd_reset_active(bfqd); |
11397 |
-+ |
11398 |
-+ bfq_deactivate_entity(entity, requeue); |
11399 |
-+} |
11400 |
-+ |
11401 |
-+static void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
11402 |
-+{ |
11403 |
-+ struct bfq_entity *entity = &bfqq->entity; |
11404 |
-+ |
11405 |
-+ bfq_activate_entity(entity); |
11406 |
-+} |
11407 |
-+ |
11408 |
-+/* |
11409 |
-+ * Called when the bfqq no longer has requests pending, remove it from |
11410 |
-+ * the service tree. |
11411 |
-+ */ |
11412 |
-+static void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
11413 |
-+ int requeue) |
11414 |
-+{ |
11415 |
-+ BUG_ON(!bfq_bfqq_busy(bfqq)); |
11416 |
-+ BUG_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
11417 |
-+ |
11418 |
-+ bfq_log_bfqq(bfqd, bfqq, "del from busy"); |
11419 |
-+ |
11420 |
-+ bfq_clear_bfqq_busy(bfqq); |
11421 |
-+ |
11422 |
-+ BUG_ON(bfqd->busy_queues == 0); |
11423 |
-+ bfqd->busy_queues--; |
11424 |
-+ |
11425 |
-+ bfq_deactivate_bfqq(bfqd, bfqq, requeue); |
11426 |
-+} |
11427 |
-+ |
11428 |
-+/* |
11429 |
-+ * Called when an inactive queue receives a new request. |
11430 |
-+ */ |
11431 |
-+static void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
11432 |
-+{ |
11433 |
-+ BUG_ON(bfq_bfqq_busy(bfqq)); |
11434 |
-+ BUG_ON(bfqq == bfqd->active_queue); |
11435 |
-+ |
11436 |
-+ bfq_log_bfqq(bfqd, bfqq, "add to busy"); |
11437 |
-+ |
11438 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
11439 |
-+ |
11440 |
-+ bfq_mark_bfqq_busy(bfqq); |
11441 |
-+ bfqd->busy_queues++; |
11442 |
-+} |
11443 |
-diff --git a/block/bfq.h b/block/bfq.h |
11444 |
-new file mode 100644 |
11445 |
-index 0000000..48ecde9 |
11446 |
---- /dev/null |
11447 |
-+++ b/block/bfq.h |
11448 |
-@@ -0,0 +1,603 @@ |
11449 |
-+/* |
11450 |
-+ * BFQ-v6r2 for 3.10.0: data structures and common functions prototypes. |
11451 |
-+ * |
11452 |
-+ * Based on ideas and code from CFQ: |
11453 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
11454 |
-+ * |
11455 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
11456 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
11457 |
-+ * |
11458 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
11459 |
-+ */ |
11460 |
-+ |
11461 |
-+#ifndef _BFQ_H |
11462 |
-+#define _BFQ_H |
11463 |
-+ |
11464 |
-+#include <linux/blktrace_api.h> |
11465 |
-+#include <linux/hrtimer.h> |
11466 |
-+#include <linux/ioprio.h> |
11467 |
-+#include <linux/rbtree.h> |
11468 |
-+ |
11469 |
-+#define BFQ_IOPRIO_CLASSES 3 |
11470 |
-+#define BFQ_CL_IDLE_TIMEOUT HZ/5 |
11471 |
-+ |
11472 |
-+#define BFQ_MIN_WEIGHT 1 |
11473 |
-+#define BFQ_MAX_WEIGHT 1000 |
11474 |
-+ |
11475 |
-+#define BFQ_DEFAULT_GRP_WEIGHT 10 |
11476 |
-+#define BFQ_DEFAULT_GRP_IOPRIO 0 |
11477 |
-+#define BFQ_DEFAULT_GRP_CLASS IOPRIO_CLASS_BE |
11478 |
-+ |
11479 |
-+struct bfq_entity; |
11480 |
-+ |
11481 |
-+/** |
11482 |
-+ * struct bfq_service_tree - per ioprio_class service tree. |
11483 |
-+ * @active: tree for active entities (i.e., those backlogged). |
11484 |
-+ * @idle: tree for idle entities (i.e., those not backlogged, with V <= F_i). |
11485 |
-+ * @first_idle: idle entity with minimum F_i. |
11486 |
-+ * @last_idle: idle entity with maximum F_i. |
11487 |
-+ * @vtime: scheduler virtual time. |
11488 |
-+ * @wsum: scheduler weight sum; active and idle entities contribute to it. |
11489 |
-+ * |
11490 |
-+ * Each service tree represents a B-WF2Q+ scheduler on its own. Each |
11491 |
-+ * ioprio_class has its own independent scheduler, and so its own |
11492 |
-+ * bfq_service_tree. All the fields are protected by the queue lock |
11493 |
-+ * of the containing bfqd. |
11494 |
-+ */ |
11495 |
-+struct bfq_service_tree { |
11496 |
-+ struct rb_root active; |
11497 |
-+ struct rb_root idle; |
11498 |
-+ |
11499 |
-+ struct bfq_entity *first_idle; |
11500 |
-+ struct bfq_entity *last_idle; |
11501 |
-+ |
11502 |
-+ u64 vtime; |
11503 |
-+ unsigned long wsum; |
11504 |
-+}; |
11505 |
-+ |
11506 |
-+/** |
11507 |
-+ * struct bfq_sched_data - multi-class scheduler. |
11508 |
-+ * @active_entity: entity under service. |
11509 |
-+ * @next_active: head-of-the-line entity in the scheduler. |
11510 |
-+ * @service_tree: array of service trees, one per ioprio_class. |
11511 |
-+ * |
11512 |
-+ * bfq_sched_data is the basic scheduler queue. It supports three |
11513 |
-+ * ioprio_classes, and can be used either as a toplevel queue or as |
11514 |
-+ * an intermediate queue on a hierarchical setup. |
11515 |
-+ * @next_active points to the active entity of the sched_data service |
11516 |
-+ * trees that will be scheduled next. |
11517 |
-+ * |
11518 |
-+ * The supported ioprio_classes are the same as in CFQ, in descending |
11519 |
-+ * priority order, IOPRIO_CLASS_RT, IOPRIO_CLASS_BE, IOPRIO_CLASS_IDLE. |
11520 |
-+ * Requests from higher priority queues are served before all the |
11521 |
-+ * requests from lower priority queues; among requests of the same |
11522 |
-+ * queue requests are served according to B-WF2Q+. |
11523 |
-+ * All the fields are protected by the queue lock of the containing bfqd. |
11524 |
-+ */ |
11525 |
-+struct bfq_sched_data { |
11526 |
-+ struct bfq_entity *active_entity; |
11527 |
-+ struct bfq_entity *next_active; |
11528 |
-+ struct bfq_service_tree service_tree[BFQ_IOPRIO_CLASSES]; |
11529 |
-+}; |
11530 |
-+ |
11531 |
-+/** |
11532 |
-+ * struct bfq_entity - schedulable entity. |
11533 |
-+ * @rb_node: service_tree member. |
11534 |
-+ * @on_st: flag, true if the entity is on a tree (either the active or |
11535 |
-+ * the idle one of its service_tree). |
11536 |
-+ * @finish: B-WF2Q+ finish timestamp (aka F_i). |
11537 |
-+ * @start: B-WF2Q+ start timestamp (aka S_i). |
11538 |
-+ * @tree: tree the entity is enqueued into; %NULL if not on a tree. |
11539 |
-+ * @min_start: minimum start time of the (active) subtree rooted at |
11540 |
-+ * this entity; used for O(log N) lookups into active trees. |
11541 |
-+ * @service: service received during the last round of service. |
11542 |
-+ * @budget: budget used to calculate F_i; F_i = S_i + @budget / @weight. |
11543 |
-+ * @weight: weight of the queue |
11544 |
-+ * @parent: parent entity, for hierarchical scheduling. |
11545 |
-+ * @my_sched_data: for non-leaf nodes in the cgroup hierarchy, the |
11546 |
-+ * associated scheduler queue, %NULL on leaf nodes. |
11547 |
-+ * @sched_data: the scheduler queue this entity belongs to. |
11548 |
-+ * @ioprio: the ioprio in use. |
11549 |
-+ * @new_weight: when a weight change is requested, the new weight value. |
11550 |
-+ * @orig_weight: original weight, used to implement weight boosting |
11551 |
-+ * @new_ioprio: when an ioprio change is requested, the new ioprio value. |
11552 |
-+ * @ioprio_class: the ioprio_class in use. |
11553 |
-+ * @new_ioprio_class: when an ioprio_class change is requested, the new |
11554 |
-+ * ioprio_class value. |
11555 |
-+ * @ioprio_changed: flag, true when the user requested a weight, ioprio or |
11556 |
-+ * ioprio_class change. |
11557 |
-+ * |
11558 |
-+ * A bfq_entity is used to represent either a bfq_queue (leaf node in the |
11559 |
-+ * cgroup hierarchy) or a bfq_group into the upper level scheduler. Each |
11560 |
-+ * entity belongs to the sched_data of the parent group in the cgroup |
11561 |
-+ * hierarchy. Non-leaf entities have also their own sched_data, stored |
11562 |
-+ * in @my_sched_data. |
11563 |
-+ * |
11564 |
-+ * Each entity stores independently its priority values; this would |
11565 |
-+ * allow different weights on different devices, but this |
11566 |
-+ * functionality is not exported to userspace by now. Priorities and |
11567 |
-+ * weights are updated lazily, first storing the new values into the |
11568 |
-+ * new_* fields, then setting the @ioprio_changed flag. As soon as |
11569 |
-+ * there is a transition in the entity state that allows the priority |
11570 |
-+ * update to take place the effective and the requested priority |
11571 |
-+ * values are synchronized. |
11572 |
-+ * |
11573 |
-+ * Unless cgroups are used, the weight value is calculated from the |
11574 |
-+ * ioprio to export the same interface as CFQ. When dealing with |
11575 |
-+ * ``well-behaved'' queues (i.e., queues that do not spend too much |
11576 |
-+ * time to consume their budget and have true sequential behavior, and |
11577 |
-+ * when there are no external factors breaking anticipation) the |
11578 |
-+ * relative weights at each level of the cgroups hierarchy should be |
11579 |
-+ * guaranteed. All the fields are protected by the queue lock of the |
11580 |
-+ * containing bfqd. |
11581 |
-+ */ |
11582 |
-+struct bfq_entity { |
11583 |
-+ struct rb_node rb_node; |
11584 |
-+ |
11585 |
-+ int on_st; |
11586 |
-+ |
11587 |
-+ u64 finish; |
11588 |
-+ u64 start; |
11589 |
-+ |
11590 |
-+ struct rb_root *tree; |
11591 |
-+ |
11592 |
-+ u64 min_start; |
11593 |
-+ |
11594 |
-+ unsigned long service, budget; |
11595 |
-+ unsigned short weight, new_weight; |
11596 |
-+ unsigned short orig_weight; |
11597 |
-+ |
11598 |
-+ struct bfq_entity *parent; |
11599 |
-+ |
11600 |
-+ struct bfq_sched_data *my_sched_data; |
11601 |
-+ struct bfq_sched_data *sched_data; |
11602 |
-+ |
11603 |
-+ unsigned short ioprio, new_ioprio; |
11604 |
-+ unsigned short ioprio_class, new_ioprio_class; |
11605 |
-+ |
11606 |
-+ int ioprio_changed; |
11607 |
-+}; |
11608 |
-+ |
11609 |
-+struct bfq_group; |
11610 |
-+ |
11611 |
-+/** |
11612 |
-+ * struct bfq_queue - leaf schedulable entity. |
11613 |
-+ * @ref: reference counter. |
11614 |
-+ * @bfqd: parent bfq_data. |
11615 |
-+ * @new_bfqq: shared bfq_queue if queue is cooperating with |
11616 |
-+ * one or more other queues. |
11617 |
-+ * @pos_node: request-position tree member (see bfq_data's @rq_pos_tree). |
11618 |
-+ * @pos_root: request-position tree root (see bfq_data's @rq_pos_tree). |
11619 |
-+ * @sort_list: sorted list of pending requests. |
11620 |
-+ * @next_rq: if fifo isn't expired, next request to serve. |
11621 |
-+ * @queued: nr of requests queued in @sort_list. |
11622 |
-+ * @allocated: currently allocated requests. |
11623 |
-+ * @meta_pending: pending metadata requests. |
11624 |
-+ * @fifo: fifo list of requests in sort_list. |
11625 |
-+ * @entity: entity representing this queue in the scheduler. |
11626 |
-+ * @max_budget: maximum budget allowed from the feedback mechanism. |
11627 |
-+ * @budget_timeout: budget expiration (in jiffies). |
11628 |
-+ * @dispatched: number of requests on the dispatch list or inside driver. |
11629 |
-+ * @org_ioprio: saved ioprio during boosted periods. |
11630 |
-+ * @flags: status flags. |
11631 |
-+ * @bfqq_list: node for active/idle bfqq list inside our bfqd. |
11632 |
-+ * @seek_samples: number of seeks sampled |
11633 |
-+ * @seek_total: sum of the distances of the seeks sampled |
11634 |
-+ * @seek_mean: mean seek distance |
11635 |
-+ * @last_request_pos: position of the last request enqueued |
11636 |
-+ * @pid: pid of the process owning the queue, used for logging purposes. |
11637 |
-+ * @last_rais_start_time: last (idle -> weight-raised) transition attempt |
11638 |
-+ * @raising_cur_max_time: current max raising time for this queue |
11639 |
-+ * |
11640 |
-+ * A bfq_queue is a leaf request queue; it can be associated to an io_context |
11641 |
-+ * or more (if it is an async one). @cgroup holds a reference to the |
11642 |
-+ * cgroup, to be sure that it does not disappear while a bfqq still |
11643 |
-+ * references it (mostly to avoid races between request issuing and task |
11644 |
-+ * migration followed by cgroup distruction). |
11645 |
-+ * All the fields are protected by the queue lock of the containing bfqd. |
11646 |
-+ */ |
11647 |
-+struct bfq_queue { |
11648 |
-+ atomic_t ref; |
11649 |
-+ struct bfq_data *bfqd; |
11650 |
-+ |
11651 |
-+ /* fields for cooperating queues handling */ |
11652 |
-+ struct bfq_queue *new_bfqq; |
11653 |
-+ struct rb_node pos_node; |
11654 |
-+ struct rb_root *pos_root; |
11655 |
-+ |
11656 |
-+ struct rb_root sort_list; |
11657 |
-+ struct request *next_rq; |
11658 |
-+ int queued[2]; |
11659 |
-+ int allocated[2]; |
11660 |
-+ int meta_pending; |
11661 |
-+ struct list_head fifo; |
11662 |
-+ |
11663 |
-+ struct bfq_entity entity; |
11664 |
-+ |
11665 |
-+ unsigned long max_budget; |
11666 |
-+ unsigned long budget_timeout; |
11667 |
-+ |
11668 |
-+ int dispatched; |
11669 |
-+ |
11670 |
-+ unsigned short org_ioprio; |
11671 |
-+ |
11672 |
-+ unsigned int flags; |
11673 |
-+ |
11674 |
-+ struct list_head bfqq_list; |
11675 |
-+ |
11676 |
-+ unsigned int seek_samples; |
11677 |
-+ u64 seek_total; |
11678 |
-+ sector_t seek_mean; |
11679 |
-+ sector_t last_request_pos; |
11680 |
-+ |
11681 |
-+ pid_t pid; |
11682 |
-+ |
11683 |
-+ /* weight-raising fields */ |
11684 |
-+ unsigned int raising_cur_max_time; |
11685 |
-+ u64 last_rais_start_finish, soft_rt_next_start; |
11686 |
-+ unsigned int raising_coeff; |
11687 |
-+}; |
11688 |
-+ |
11689 |
-+/** |
11690 |
-+ * struct bfq_ttime - per process thinktime stats. |
11691 |
-+ * @ttime_total: total process thinktime |
11692 |
-+ * @ttime_samples: number of thinktime samples |
11693 |
-+ * @ttime_mean: average process thinktime |
11694 |
-+ */ |
11695 |
-+struct bfq_ttime { |
11696 |
-+ unsigned long last_end_request; |
11697 |
-+ |
11698 |
-+ unsigned long ttime_total; |
11699 |
-+ unsigned long ttime_samples; |
11700 |
-+ unsigned long ttime_mean; |
11701 |
-+}; |
11702 |
-+ |
11703 |
-+/** |
11704 |
-+ * struct bfq_io_cq - per (request_queue, io_context) structure. |
11705 |
-+ * @icq: associated io_cq structure |
11706 |
-+ * @bfqq: array of two process queues, the sync and the async |
11707 |
-+ * @ttime: associated @bfq_ttime struct |
11708 |
-+ */ |
11709 |
-+struct bfq_io_cq { |
11710 |
-+ struct io_cq icq; /* must be the first member */ |
11711 |
-+ struct bfq_queue *bfqq[2]; |
11712 |
-+ struct bfq_ttime ttime; |
11713 |
-+ int ioprio; |
11714 |
-+}; |
11715 |
-+ |
11716 |
-+/** |
11717 |
-+ * struct bfq_data - per device data structure. |
11718 |
-+ * @queue: request queue for the managed device. |
11719 |
-+ * @root_group: root bfq_group for the device. |
11720 |
-+ * @rq_pos_tree: rbtree sorted by next_request position, |
11721 |
-+ * used when determining if two or more queues |
11722 |
-+ * have interleaving requests (see bfq_close_cooperator). |
11723 |
-+ * @busy_queues: number of bfq_queues containing requests (including the |
11724 |
-+ * queue under service, even if it is idling). |
11725 |
-+ * @queued: number of queued requests. |
11726 |
-+ * @rq_in_driver: number of requests dispatched and waiting for completion. |
11727 |
-+ * @sync_flight: number of sync requests in the driver. |
11728 |
-+ * @max_rq_in_driver: max number of reqs in driver in the last @hw_tag_samples |
11729 |
-+ * completed requests . |
11730 |
-+ * @hw_tag_samples: nr of samples used to calculate hw_tag. |
11731 |
-+ * @hw_tag: flag set to one if the driver is showing a queueing behavior. |
11732 |
-+ * @budgets_assigned: number of budgets assigned. |
11733 |
-+ * @idle_slice_timer: timer set when idling for the next sequential request |
11734 |
-+ * from the queue under service. |
11735 |
-+ * @unplug_work: delayed work to restart dispatching on the request queue. |
11736 |
-+ * @active_queue: bfq_queue under service. |
11737 |
-+ * @active_bic: bfq_io_cq (bic) associated with the @active_queue. |
11738 |
-+ * @last_position: on-disk position of the last served request. |
11739 |
-+ * @last_budget_start: beginning of the last budget. |
11740 |
-+ * @last_idling_start: beginning of the last idle slice. |
11741 |
-+ * @peak_rate: peak transfer rate observed for a budget. |
11742 |
-+ * @peak_rate_samples: number of samples used to calculate @peak_rate. |
11743 |
-+ * @bfq_max_budget: maximum budget allotted to a bfq_queue before rescheduling. |
11744 |
-+ * @group_list: list of all the bfq_groups active on the device. |
11745 |
-+ * @active_list: list of all the bfq_queues active on the device. |
11746 |
-+ * @idle_list: list of all the bfq_queues idle on the device. |
11747 |
-+ * @bfq_quantum: max number of requests dispatched per dispatch round. |
11748 |
-+ * @bfq_fifo_expire: timeout for async/sync requests; when it expires |
11749 |
-+ * requests are served in fifo order. |
11750 |
-+ * @bfq_back_penalty: weight of backward seeks wrt forward ones. |
11751 |
-+ * @bfq_back_max: maximum allowed backward seek. |
11752 |
-+ * @bfq_slice_idle: maximum idling time. |
11753 |
-+ * @bfq_user_max_budget: user-configured max budget value (0 for auto-tuning). |
11754 |
-+ * @bfq_max_budget_async_rq: maximum budget (in nr of requests) allotted to |
11755 |
-+ * async queues. |
11756 |
-+ * @bfq_timeout: timeout for bfq_queues to consume their budget; used to |
11757 |
-+ * to prevent seeky queues to impose long latencies to well |
11758 |
-+ * behaved ones (this also implies that seeky queues cannot |
11759 |
-+ * receive guarantees in the service domain; after a timeout |
11760 |
-+ * they are charged for the whole allocated budget, to try |
11761 |
-+ * to preserve a behavior reasonably fair among them, but |
11762 |
-+ * without service-domain guarantees). |
11763 |
-+ * @bfq_raising_coeff: Maximum factor by which the weight of a boosted |
11764 |
-+ * queue is multiplied |
11765 |
-+ * @bfq_raising_max_time: maximum duration of a weight-raising period (jiffies) |
11766 |
-+ * @bfq_raising_rt_max_time: maximum duration for soft real-time processes |
11767 |
-+ * @bfq_raising_min_idle_time: minimum idle period after which weight-raising |
11768 |
-+ * may be reactivated for a queue (in jiffies) |
11769 |
-+ * @bfq_raising_min_inter_arr_async: minimum period between request arrivals |
11770 |
-+ * after which weight-raising may be |
11771 |
-+ * reactivated for an already busy queue |
11772 |
-+ * (in jiffies) |
11773 |
-+ * @bfq_raising_max_softrt_rate: max service-rate for a soft real-time queue, |
11774 |
-+ * sectors per seconds |
11775 |
-+ * @RT_prod: cached value of the product R*T used for computing the maximum |
11776 |
-+ * duration of the weight raising automatically |
11777 |
-+ * @oom_bfqq: fallback dummy bfqq for extreme OOM conditions |
11778 |
-+ * |
11779 |
-+ * All the fields are protected by the @queue lock. |
11780 |
-+ */ |
11781 |
-+struct bfq_data { |
11782 |
-+ struct request_queue *queue; |
11783 |
-+ |
11784 |
-+ struct bfq_group *root_group; |
11785 |
-+ |
11786 |
-+ struct rb_root rq_pos_tree; |
11787 |
-+ |
11788 |
-+ int busy_queues; |
11789 |
-+ int queued; |
11790 |
-+ int rq_in_driver; |
11791 |
-+ int sync_flight; |
11792 |
-+ |
11793 |
-+ int max_rq_in_driver; |
11794 |
-+ int hw_tag_samples; |
11795 |
-+ int hw_tag; |
11796 |
-+ |
11797 |
-+ int budgets_assigned; |
11798 |
-+ |
11799 |
-+ struct timer_list idle_slice_timer; |
11800 |
-+ struct work_struct unplug_work; |
11801 |
-+ |
11802 |
-+ struct bfq_queue *active_queue; |
11803 |
-+ struct bfq_io_cq *active_bic; |
11804 |
-+ |
11805 |
-+ sector_t last_position; |
11806 |
-+ |
11807 |
-+ ktime_t last_budget_start; |
11808 |
-+ ktime_t last_idling_start; |
11809 |
-+ int peak_rate_samples; |
11810 |
-+ u64 peak_rate; |
11811 |
-+ unsigned long bfq_max_budget; |
11812 |
-+ |
11813 |
-+ struct hlist_head group_list; |
11814 |
-+ struct list_head active_list; |
11815 |
-+ struct list_head idle_list; |
11816 |
-+ |
11817 |
-+ unsigned int bfq_quantum; |
11818 |
-+ unsigned int bfq_fifo_expire[2]; |
11819 |
-+ unsigned int bfq_back_penalty; |
11820 |
-+ unsigned int bfq_back_max; |
11821 |
-+ unsigned int bfq_slice_idle; |
11822 |
-+ u64 bfq_class_idle_last_service; |
11823 |
-+ |
11824 |
-+ unsigned int bfq_user_max_budget; |
11825 |
-+ unsigned int bfq_max_budget_async_rq; |
11826 |
-+ unsigned int bfq_timeout[2]; |
11827 |
-+ |
11828 |
-+ bool low_latency; |
11829 |
-+ |
11830 |
-+ /* parameters of the low_latency heuristics */ |
11831 |
-+ unsigned int bfq_raising_coeff; |
11832 |
-+ unsigned int bfq_raising_max_time; |
11833 |
-+ unsigned int bfq_raising_rt_max_time; |
11834 |
-+ unsigned int bfq_raising_min_idle_time; |
11835 |
-+ unsigned int bfq_raising_min_inter_arr_async; |
11836 |
-+ unsigned int bfq_raising_max_softrt_rate; |
11837 |
-+ u64 RT_prod; |
11838 |
-+ |
11839 |
-+ struct bfq_queue oom_bfqq; |
11840 |
-+}; |
11841 |
-+ |
11842 |
-+enum bfqq_state_flags { |
11843 |
-+ BFQ_BFQQ_FLAG_busy = 0, /* has requests or is under service */ |
11844 |
-+ BFQ_BFQQ_FLAG_wait_request, /* waiting for a request */ |
11845 |
-+ BFQ_BFQQ_FLAG_must_alloc, /* must be allowed rq alloc */ |
11846 |
-+ BFQ_BFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
11847 |
-+ BFQ_BFQQ_FLAG_idle_window, /* slice idling enabled */ |
11848 |
-+ BFQ_BFQQ_FLAG_prio_changed, /* task priority has changed */ |
11849 |
-+ BFQ_BFQQ_FLAG_sync, /* synchronous queue */ |
11850 |
-+ BFQ_BFQQ_FLAG_budget_new, /* no completion with this budget */ |
11851 |
-+ BFQ_BFQQ_FLAG_coop, /* bfqq is shared */ |
11852 |
-+ BFQ_BFQQ_FLAG_split_coop, /* shared bfqq will be splitted */ |
11853 |
-+ BFQ_BFQQ_FLAG_some_coop_idle, /* some cooperator is inactive */ |
11854 |
-+}; |
11855 |
-+ |
11856 |
-+#define BFQ_BFQQ_FNS(name) \ |
11857 |
-+static inline void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ |
11858 |
-+{ \ |
11859 |
-+ (bfqq)->flags |= (1 << BFQ_BFQQ_FLAG_##name); \ |
11860 |
-+} \ |
11861 |
-+static inline void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ |
11862 |
-+{ \ |
11863 |
-+ (bfqq)->flags &= ~(1 << BFQ_BFQQ_FLAG_##name); \ |
11864 |
-+} \ |
11865 |
-+static inline int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ |
11866 |
-+{ \ |
11867 |
-+ return ((bfqq)->flags & (1 << BFQ_BFQQ_FLAG_##name)) != 0; \ |
11868 |
-+} |
11869 |
-+ |
11870 |
-+BFQ_BFQQ_FNS(busy); |
11871 |
-+BFQ_BFQQ_FNS(wait_request); |
11872 |
-+BFQ_BFQQ_FNS(must_alloc); |
11873 |
-+BFQ_BFQQ_FNS(fifo_expire); |
11874 |
-+BFQ_BFQQ_FNS(idle_window); |
11875 |
-+BFQ_BFQQ_FNS(prio_changed); |
11876 |
-+BFQ_BFQQ_FNS(sync); |
11877 |
-+BFQ_BFQQ_FNS(budget_new); |
11878 |
-+BFQ_BFQQ_FNS(coop); |
11879 |
-+BFQ_BFQQ_FNS(split_coop); |
11880 |
-+BFQ_BFQQ_FNS(some_coop_idle); |
11881 |
-+#undef BFQ_BFQQ_FNS |
11882 |
-+ |
11883 |
-+/* Logging facilities. */ |
11884 |
-+#define bfq_log_bfqq(bfqd, bfqq, fmt, args...) \ |
11885 |
-+ blk_add_trace_msg((bfqd)->queue, "bfq%d " fmt, (bfqq)->pid, ##args) |
11886 |
-+ |
11887 |
-+#define bfq_log(bfqd, fmt, args...) \ |
11888 |
-+ blk_add_trace_msg((bfqd)->queue, "bfq " fmt, ##args) |
11889 |
-+ |
11890 |
-+/* Expiration reasons. */ |
11891 |
-+enum bfqq_expiration { |
11892 |
-+ BFQ_BFQQ_TOO_IDLE = 0, /* queue has been idling for too long */ |
11893 |
-+ BFQ_BFQQ_BUDGET_TIMEOUT, /* budget took too long to be used */ |
11894 |
-+ BFQ_BFQQ_BUDGET_EXHAUSTED, /* budget consumed */ |
11895 |
-+ BFQ_BFQQ_NO_MORE_REQUESTS, /* the queue has no more requests */ |
11896 |
-+}; |
11897 |
-+ |
11898 |
-+#ifdef CONFIG_CGROUP_BFQIO |
11899 |
-+/** |
11900 |
-+ * struct bfq_group - per (device, cgroup) data structure. |
11901 |
-+ * @entity: schedulable entity to insert into the parent group sched_data. |
11902 |
-+ * @sched_data: own sched_data, to contain child entities (they may be |
11903 |
-+ * both bfq_queues and bfq_groups). |
11904 |
-+ * @group_node: node to be inserted into the bfqio_cgroup->group_data |
11905 |
-+ * list of the containing cgroup's bfqio_cgroup. |
11906 |
-+ * @bfqd_node: node to be inserted into the @bfqd->group_list list |
11907 |
-+ * of the groups active on the same device; used for cleanup. |
11908 |
-+ * @bfqd: the bfq_data for the device this group acts upon. |
11909 |
-+ * @async_bfqq: array of async queues for all the tasks belonging to |
11910 |
-+ * the group, one queue per ioprio value per ioprio_class, |
11911 |
-+ * except for the idle class that has only one queue. |
11912 |
-+ * @async_idle_bfqq: async queue for the idle class (ioprio is ignored). |
11913 |
-+ * @my_entity: pointer to @entity, %NULL for the toplevel group; used |
11914 |
-+ * to avoid too many special cases during group creation/migration. |
11915 |
-+ * |
11916 |
-+ * Each (device, cgroup) pair has its own bfq_group, i.e., for each cgroup |
11917 |
-+ * there is a set of bfq_groups, each one collecting the lower-level |
11918 |
-+ * entities belonging to the group that are acting on the same device. |
11919 |
-+ * |
11920 |
-+ * Locking works as follows: |
11921 |
-+ * o @group_node is protected by the bfqio_cgroup lock, and is accessed |
11922 |
-+ * via RCU from its readers. |
11923 |
-+ * o @bfqd is protected by the queue lock, RCU is used to access it |
11924 |
-+ * from the readers. |
11925 |
-+ * o All the other fields are protected by the @bfqd queue lock. |
11926 |
-+ */ |
11927 |
-+struct bfq_group { |
11928 |
-+ struct bfq_entity entity; |
11929 |
-+ struct bfq_sched_data sched_data; |
11930 |
-+ |
11931 |
-+ struct hlist_node group_node; |
11932 |
-+ struct hlist_node bfqd_node; |
11933 |
-+ |
11934 |
-+ void *bfqd; |
11935 |
-+ |
11936 |
-+ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
11937 |
-+ struct bfq_queue *async_idle_bfqq; |
11938 |
-+ |
11939 |
-+ struct bfq_entity *my_entity; |
11940 |
-+}; |
11941 |
-+ |
11942 |
-+/** |
11943 |
-+ * struct bfqio_cgroup - bfq cgroup data structure. |
11944 |
-+ * @css: subsystem state for bfq in the containing cgroup. |
11945 |
-+ * @weight: cgroup weight. |
11946 |
-+ * @ioprio: cgroup ioprio. |
11947 |
-+ * @ioprio_class: cgroup ioprio_class. |
11948 |
-+ * @lock: spinlock that protects @ioprio, @ioprio_class and @group_data. |
11949 |
-+ * @group_data: list containing the bfq_group belonging to this cgroup. |
11950 |
-+ * |
11951 |
-+ * @group_data is accessed using RCU, with @lock protecting the updates, |
11952 |
-+ * @ioprio and @ioprio_class are protected by @lock. |
11953 |
-+ */ |
11954 |
-+struct bfqio_cgroup { |
11955 |
-+ struct cgroup_subsys_state css; |
11956 |
-+ |
11957 |
-+ unsigned short weight, ioprio, ioprio_class; |
11958 |
-+ |
11959 |
-+ spinlock_t lock; |
11960 |
-+ struct hlist_head group_data; |
11961 |
-+}; |
11962 |
-+#else |
11963 |
-+struct bfq_group { |
11964 |
-+ struct bfq_sched_data sched_data; |
11965 |
-+ |
11966 |
-+ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
11967 |
-+ struct bfq_queue *async_idle_bfqq; |
11968 |
-+}; |
11969 |
-+#endif |
11970 |
-+ |
11971 |
-+static inline struct bfq_service_tree * |
11972 |
-+bfq_entity_service_tree(struct bfq_entity *entity) |
11973 |
-+{ |
11974 |
-+ struct bfq_sched_data *sched_data = entity->sched_data; |
11975 |
-+ unsigned int idx = entity->ioprio_class - 1; |
11976 |
-+ |
11977 |
-+ BUG_ON(idx >= BFQ_IOPRIO_CLASSES); |
11978 |
-+ BUG_ON(sched_data == NULL); |
11979 |
-+ |
11980 |
-+ return sched_data->service_tree + idx; |
11981 |
-+} |
11982 |
-+ |
11983 |
-+static inline struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, |
11984 |
-+ int is_sync) |
11985 |
-+{ |
11986 |
-+ return bic->bfqq[!!is_sync]; |
11987 |
-+} |
11988 |
-+ |
11989 |
-+static inline void bic_set_bfqq(struct bfq_io_cq *bic, |
11990 |
-+ struct bfq_queue *bfqq, int is_sync) |
11991 |
-+{ |
11992 |
-+ bic->bfqq[!!is_sync] = bfqq; |
11993 |
-+} |
11994 |
-+ |
11995 |
-+static inline struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) |
11996 |
-+{ |
11997 |
-+ return bic->icq.q->elevator->elevator_data; |
11998 |
-+} |
11999 |
-+ |
12000 |
-+/** |
12001 |
-+ * bfq_get_bfqd_locked - get a lock to a bfqd using a RCU protected pointer. |
12002 |
-+ * @ptr: a pointer to a bfqd. |
12003 |
-+ * @flags: storage for the flags to be saved. |
12004 |
-+ * |
12005 |
-+ * This function allows bfqg->bfqd to be protected by the |
12006 |
-+ * queue lock of the bfqd they reference; the pointer is dereferenced |
12007 |
-+ * under RCU, so the storage for bfqd is assured to be safe as long |
12008 |
-+ * as the RCU read side critical section does not end. After the |
12009 |
-+ * bfqd->queue->queue_lock is taken the pointer is rechecked, to be |
12010 |
-+ * sure that no other writer accessed it. If we raced with a writer, |
12011 |
-+ * the function returns NULL, with the queue unlocked, otherwise it |
12012 |
-+ * returns the dereferenced pointer, with the queue locked. |
12013 |
-+ */ |
12014 |
-+static inline struct bfq_data *bfq_get_bfqd_locked(void **ptr, |
12015 |
-+ unsigned long *flags) |
12016 |
-+{ |
12017 |
-+ struct bfq_data *bfqd; |
12018 |
-+ |
12019 |
-+ rcu_read_lock(); |
12020 |
-+ bfqd = rcu_dereference(*(struct bfq_data **)ptr); |
12021 |
-+ |
12022 |
-+ if (bfqd != NULL) { |
12023 |
-+ spin_lock_irqsave(bfqd->queue->queue_lock, *flags); |
12024 |
-+ if (*ptr == bfqd) |
12025 |
-+ goto out; |
12026 |
-+ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
12027 |
-+ } |
12028 |
-+ |
12029 |
-+ bfqd = NULL; |
12030 |
-+out: |
12031 |
-+ rcu_read_unlock(); |
12032 |
-+ return bfqd; |
12033 |
-+} |
12034 |
-+ |
12035 |
-+static inline void bfq_put_bfqd_unlock(struct bfq_data *bfqd, |
12036 |
-+ unsigned long *flags) |
12037 |
-+{ |
12038 |
-+ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
12039 |
-+} |
12040 |
-+ |
12041 |
-+static void bfq_changed_ioprio(struct bfq_io_cq *bic); |
12042 |
-+static void bfq_put_queue(struct bfq_queue *bfqq); |
12043 |
-+static void bfq_dispatch_insert(struct request_queue *q, struct request *rq); |
12044 |
-+static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
12045 |
-+ struct bfq_group *bfqg, int is_sync, |
12046 |
-+ struct bfq_io_cq *bic, gfp_t gfp_mask); |
12047 |
-+static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
12048 |
-+ struct bfq_group *bfqg); |
12049 |
-+static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg); |
12050 |
-+static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq); |
12051 |
-+#endif |
12052 |
--- |
12053 |
-1.8.1.4 |
12054 |
- |
12055 |
|
12056 |
Deleted: genpatches-2.6/trunk/3.14/5000_BFQ-3-block-add-Early-Queue-Merge-EQM-v6r2-for-3.11.0.patch1 |
12057 |
=================================================================== |
12058 |
--- genpatches-2.6/trunk/3.14/5000_BFQ-3-block-add-Early-Queue-Merge-EQM-v6r2-for-3.11.0.patch1 2014-01-29 14:41:45 UTC (rev 2660) |
12059 |
+++ genpatches-2.6/trunk/3.14/5000_BFQ-3-block-add-Early-Queue-Merge-EQM-v6r2-for-3.11.0.patch1 2014-01-30 16:49:47 UTC (rev 2661) |
12060 |
@@ -1,1049 +0,0 @@ |
12061 |
-From 9acaa783ecab69925d38c6aca7252ff565a093d0 Mon Sep 17 00:00:00 2001 |
12062 |
-From: Mauro Andreolini <mauro.andreolini@×××××××.it> |
12063 |
-Date: Fri, 14 Jun 2013 13:46:47 +0200 |
12064 |
-Subject: [PATCH 3/3] block, bfq: add Early Queue Merge (EQM) to BFQ-v6r2 for |
12065 |
- 3.11.0 |
12066 |
- |
12067 |
-A set of processes may happen to perform interleaved reads, i.e., requests |
12068 |
-whose union would give rise to a sequential read pattern. There are two |
12069 |
-typical cases: in the first case, processes read fixed-size chunks of |
12070 |
-data at a fixed distance from each other, while in the second case processes |
12071 |
-may read variable-size chunks at variable distances. The latter case occurs |
12072 |
-for example with KVM, which splits the I/O generated by the guest into |
12073 |
-multiple chunks, and lets these chunks be served by a pool of cooperating |
12074 |
-processes, iteratively assigning the next chunk of I/O to the first |
12075 |
-available process. CFQ uses actual queue merging for the first type of |
12076 |
-processes, whereas it uses preemption to get a sequential read pattern out |
12077 |
-of the read requests performed by the second type of processes. In the end |
12078 |
-it uses two different mechanisms to achieve the same goal: boosting the |
12079 |
-throughput with interleaved I/O. |
12080 |
- |
12081 |
-This patch introduces Early Queue Merge (EQM), a unified mechanism to get a |
12082 |
-sequential read pattern with both types of processes. The main idea is |
12083 |
-checking newly arrived requests against the next request of the active queue |
12084 |
-both in case of actual request insert and in case of request merge. By doing |
12085 |
-so, both the types of processes can be handled by just merging their queues. |
12086 |
-EQM is then simpler and more compact than the pair of mechanisms used in |
12087 |
-CFQ. |
12088 |
- |
12089 |
-Finally, EQM also preserves the typical low-latency properties of BFQ, by |
12090 |
-properly restoring the weight-raising state of a queue when it gets back to |
12091 |
-a non-merged state. |
12092 |
- |
12093 |
-Signed-off-by: Mauro Andreolini <mauro.andreolini@×××××××.it> |
12094 |
-Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
12095 |
-Reviewed-by: Paolo Valente <paolo.valente@×××××××.it> |
12096 |
---- |
12097 |
- block/bfq-iosched.c | 653 ++++++++++++++++++++++++++++++++++++---------------- |
12098 |
- block/bfq-sched.c | 28 --- |
12099 |
- block/bfq.h | 16 ++ |
12100 |
- 3 files changed, 466 insertions(+), 231 deletions(-) |
12101 |
- |
12102 |
-diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c |
12103 |
-index 0ed2746..bbe79fb 100644 |
12104 |
---- a/block/bfq-iosched.c |
12105 |
-+++ b/block/bfq-iosched.c |
12106 |
-@@ -444,6 +444,43 @@ static inline unsigned int bfq_wrais_duration(struct bfq_data *bfqd) |
12107 |
- return dur; |
12108 |
- } |
12109 |
- |
12110 |
-+static inline void |
12111 |
-+bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_io_cq *bic) |
12112 |
-+{ |
12113 |
-+ if (bic->saved_idle_window) |
12114 |
-+ bfq_mark_bfqq_idle_window(bfqq); |
12115 |
-+ else |
12116 |
-+ bfq_clear_bfqq_idle_window(bfqq); |
12117 |
-+ if (bic->raising_time_left && bfqq->bfqd->low_latency) { |
12118 |
-+ /* |
12119 |
-+ * Start a weight raising period with the duration given by |
12120 |
-+ * the raising_time_left snapshot. |
12121 |
-+ */ |
12122 |
-+ bfqq->raising_coeff = bfqq->bfqd->bfq_raising_coeff; |
12123 |
-+ bfqq->raising_cur_max_time = bic->raising_time_left; |
12124 |
-+ bfqq->last_rais_start_finish = jiffies; |
12125 |
-+ } |
12126 |
-+ /* |
12127 |
-+ * Clear raising_time_left to prevent bfq_bfqq_save_state() from |
12128 |
-+ * getting confused about the queue's need of a weight-raising |
12129 |
-+ * period. |
12130 |
-+ */ |
12131 |
-+ bic->raising_time_left = 0; |
12132 |
-+} |
12133 |
-+ |
12134 |
-+/* |
12135 |
-+ * Must be called with the queue_lock held. |
12136 |
-+ */ |
12137 |
-+static int bfqq_process_refs(struct bfq_queue *bfqq) |
12138 |
-+{ |
12139 |
-+ int process_refs, io_refs; |
12140 |
-+ |
12141 |
-+ io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
12142 |
-+ process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
12143 |
-+ BUG_ON(process_refs < 0); |
12144 |
-+ return process_refs; |
12145 |
-+} |
12146 |
-+ |
12147 |
- static void bfq_add_rq_rb(struct request *rq) |
12148 |
- { |
12149 |
- struct bfq_queue *bfqq = RQ_BFQQ(rq); |
12150 |
-@@ -483,11 +520,20 @@ static void bfq_add_rq_rb(struct request *rq) |
12151 |
- if (! bfqd->low_latency) |
12152 |
- goto add_bfqq_busy; |
12153 |
- |
12154 |
-+ if (bfq_bfqq_just_split(bfqq)) |
12155 |
-+ goto set_ioprio_changed; |
12156 |
-+ |
12157 |
- /* |
12158 |
-- * If the queue is not being boosted and has been idle |
12159 |
-- * for enough time, start a weight-raising period |
12160 |
-+ * If the queue: |
12161 |
-+ * - is not being boosted, |
12162 |
-+ * - has been idle for enough time, |
12163 |
-+ * - is not a sync queue or is linked to a bfq_io_cq (it is |
12164 |
-+ * shared "for its nature" or it is not shared and its |
12165 |
-+ * requests have not been redirected to a shared queue) |
12166 |
-+ * start a weight-raising period. |
12167 |
- */ |
12168 |
-- if(old_raising_coeff == 1 && (idle_for_long_time || soft_rt)) { |
12169 |
-+ if(old_raising_coeff == 1 && (idle_for_long_time || soft_rt) && |
12170 |
-+ (!bfq_bfqq_sync(bfqq) || bfqq->bic != NULL)) { |
12171 |
- bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
12172 |
- if (idle_for_long_time) |
12173 |
- bfqq->raising_cur_max_time = |
12174 |
-@@ -517,6 +563,7 @@ static void bfq_add_rq_rb(struct request *rq) |
12175 |
- raising_cur_max_time)); |
12176 |
- } |
12177 |
- } |
12178 |
-+set_ioprio_changed: |
12179 |
- if (old_raising_coeff != bfqq->raising_coeff) |
12180 |
- entity->ioprio_changed = 1; |
12181 |
- add_bfqq_busy: |
12182 |
-@@ -695,89 +742,35 @@ static void bfq_end_raising(struct bfq_data *bfqd) |
12183 |
- spin_unlock_irq(bfqd->queue->queue_lock); |
12184 |
- } |
12185 |
- |
12186 |
--static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
12187 |
-- struct bio *bio) |
12188 |
-+static inline sector_t bfq_io_struct_pos(void *io_struct, bool request) |
12189 |
- { |
12190 |
-- struct bfq_data *bfqd = q->elevator->elevator_data; |
12191 |
-- struct bfq_io_cq *bic; |
12192 |
-- struct bfq_queue *bfqq; |
12193 |
-- |
12194 |
-- /* |
12195 |
-- * Disallow merge of a sync bio into an async request. |
12196 |
-- */ |
12197 |
-- if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
12198 |
-- return 0; |
12199 |
-- |
12200 |
-- /* |
12201 |
-- * Lookup the bfqq that this bio will be queued with. Allow |
12202 |
-- * merge only if rq is queued there. |
12203 |
-- * Queue lock is held here. |
12204 |
-- */ |
12205 |
-- bic = bfq_bic_lookup(bfqd, current->io_context); |
12206 |
-- if (bic == NULL) |
12207 |
-- return 0; |
12208 |
-- |
12209 |
-- bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
12210 |
-- return bfqq == RQ_BFQQ(rq); |
12211 |
--} |
12212 |
-- |
12213 |
--static void __bfq_set_active_queue(struct bfq_data *bfqd, |
12214 |
-- struct bfq_queue *bfqq) |
12215 |
--{ |
12216 |
-- if (bfqq != NULL) { |
12217 |
-- bfq_mark_bfqq_must_alloc(bfqq); |
12218 |
-- bfq_mark_bfqq_budget_new(bfqq); |
12219 |
-- bfq_clear_bfqq_fifo_expire(bfqq); |
12220 |
-- |
12221 |
-- bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
12222 |
-- |
12223 |
-- bfq_log_bfqq(bfqd, bfqq, "set_active_queue, cur-budget = %lu", |
12224 |
-- bfqq->entity.budget); |
12225 |
-- } |
12226 |
-- |
12227 |
-- bfqd->active_queue = bfqq; |
12228 |
--} |
12229 |
-- |
12230 |
--/* |
12231 |
-- * Get and set a new active queue for service. |
12232 |
-- */ |
12233 |
--static struct bfq_queue *bfq_set_active_queue(struct bfq_data *bfqd, |
12234 |
-- struct bfq_queue *bfqq) |
12235 |
--{ |
12236 |
-- if (!bfqq) |
12237 |
-- bfqq = bfq_get_next_queue(bfqd); |
12238 |
-+ if (request) |
12239 |
-+ return blk_rq_pos(io_struct); |
12240 |
- else |
12241 |
-- bfq_get_next_queue_forced(bfqd, bfqq); |
12242 |
-- |
12243 |
-- __bfq_set_active_queue(bfqd, bfqq); |
12244 |
-- return bfqq; |
12245 |
-+ return ((struct bio *)io_struct)->bi_sector; |
12246 |
- } |
12247 |
- |
12248 |
--static inline sector_t bfq_dist_from_last(struct bfq_data *bfqd, |
12249 |
-- struct request *rq) |
12250 |
-+static inline sector_t bfq_dist_from(sector_t pos1, |
12251 |
-+ sector_t pos2) |
12252 |
- { |
12253 |
-- if (blk_rq_pos(rq) >= bfqd->last_position) |
12254 |
-- return blk_rq_pos(rq) - bfqd->last_position; |
12255 |
-+ if (pos1 >= pos2) |
12256 |
-+ return pos1 - pos2; |
12257 |
- else |
12258 |
-- return bfqd->last_position - blk_rq_pos(rq); |
12259 |
-+ return pos2 - pos1; |
12260 |
- } |
12261 |
- |
12262 |
--/* |
12263 |
-- * Return true if bfqq has no request pending and rq is close enough to |
12264 |
-- * bfqd->last_position, or if rq is closer to bfqd->last_position than |
12265 |
-- * bfqq->next_rq |
12266 |
-- */ |
12267 |
--static inline int bfq_rq_close(struct bfq_data *bfqd, struct request *rq) |
12268 |
-+static inline int bfq_rq_close_to_sector(void *io_struct, bool request, |
12269 |
-+ sector_t sector) |
12270 |
- { |
12271 |
-- return bfq_dist_from_last(bfqd, rq) <= BFQQ_SEEK_THR; |
12272 |
-+ return bfq_dist_from(bfq_io_struct_pos(io_struct, request), sector) <= |
12273 |
-+ BFQQ_SEEK_THR; |
12274 |
- } |
12275 |
- |
12276 |
--static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
12277 |
-+static struct bfq_queue *bfqq_close(struct bfq_data *bfqd, sector_t sector) |
12278 |
- { |
12279 |
- struct rb_root *root = &bfqd->rq_pos_tree; |
12280 |
- struct rb_node *parent, *node; |
12281 |
- struct bfq_queue *__bfqq; |
12282 |
-- sector_t sector = bfqd->last_position; |
12283 |
- |
12284 |
- if (RB_EMPTY_ROOT(root)) |
12285 |
- return NULL; |
12286 |
-@@ -796,7 +789,7 @@ static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
12287 |
- * position). |
12288 |
- */ |
12289 |
- __bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
12290 |
-- if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
12291 |
-+ if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) |
12292 |
- return __bfqq; |
12293 |
- |
12294 |
- if (blk_rq_pos(__bfqq->next_rq) < sector) |
12295 |
-@@ -807,7 +800,7 @@ static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
12296 |
- return NULL; |
12297 |
- |
12298 |
- __bfqq = rb_entry(node, struct bfq_queue, pos_node); |
12299 |
-- if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
12300 |
-+ if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) |
12301 |
- return __bfqq; |
12302 |
- |
12303 |
- return NULL; |
12304 |
-@@ -816,14 +809,12 @@ static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
12305 |
- /* |
12306 |
- * bfqd - obvious |
12307 |
- * cur_bfqq - passed in so that we don't decide that the current queue |
12308 |
-- * is closely cooperating with itself. |
12309 |
-- * |
12310 |
-- * We are assuming that cur_bfqq has dispatched at least one request, |
12311 |
-- * and that bfqd->last_position reflects a position on the disk associated |
12312 |
-- * with the I/O issued by cur_bfqq. |
12313 |
-+ * is closely cooperating with itself |
12314 |
-+ * sector - used as a reference point to search for a close queue |
12315 |
- */ |
12316 |
- static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
12317 |
-- struct bfq_queue *cur_bfqq) |
12318 |
-+ struct bfq_queue *cur_bfqq, |
12319 |
-+ sector_t sector) |
12320 |
- { |
12321 |
- struct bfq_queue *bfqq; |
12322 |
- |
12323 |
-@@ -843,7 +834,7 @@ static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
12324 |
- * working closely on the same area of the disk. In that case, |
12325 |
- * we can group them together and don't waste time idling. |
12326 |
- */ |
12327 |
-- bfqq = bfqq_close(bfqd); |
12328 |
-+ bfqq = bfqq_close(bfqd, sector); |
12329 |
- if (bfqq == NULL || bfqq == cur_bfqq) |
12330 |
- return NULL; |
12331 |
- |
12332 |
-@@ -870,6 +861,275 @@ static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
12333 |
- return bfqq; |
12334 |
- } |
12335 |
- |
12336 |
-+static struct bfq_queue * |
12337 |
-+bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
12338 |
-+{ |
12339 |
-+ int process_refs, new_process_refs; |
12340 |
-+ struct bfq_queue *__bfqq; |
12341 |
-+ |
12342 |
-+ /* |
12343 |
-+ * If there are no process references on the new_bfqq, then it is |
12344 |
-+ * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
12345 |
-+ * may have dropped their last reference (not just their last process |
12346 |
-+ * reference). |
12347 |
-+ */ |
12348 |
-+ if (!bfqq_process_refs(new_bfqq)) |
12349 |
-+ return NULL; |
12350 |
-+ |
12351 |
-+ /* Avoid a circular list and skip interim queue merges. */ |
12352 |
-+ while ((__bfqq = new_bfqq->new_bfqq)) { |
12353 |
-+ if (__bfqq == bfqq) |
12354 |
-+ return NULL; |
12355 |
-+ new_bfqq = __bfqq; |
12356 |
-+ } |
12357 |
-+ |
12358 |
-+ process_refs = bfqq_process_refs(bfqq); |
12359 |
-+ new_process_refs = bfqq_process_refs(new_bfqq); |
12360 |
-+ /* |
12361 |
-+ * If the process for the bfqq has gone away, there is no |
12362 |
-+ * sense in merging the queues. |
12363 |
-+ */ |
12364 |
-+ if (process_refs == 0 || new_process_refs == 0) |
12365 |
-+ return NULL; |
12366 |
-+ |
12367 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
12368 |
-+ new_bfqq->pid); |
12369 |
-+ |
12370 |
-+ /* |
12371 |
-+ * Merging is just a redirection: the requests of the process owning |
12372 |
-+ * one of the two queues are redirected to the other queue. The latter |
12373 |
-+ * queue, in its turn, is set as shared if this is the first time that |
12374 |
-+ * the requests of some process are redirected to it. |
12375 |
-+ * |
12376 |
-+ * We redirect bfqq to new_bfqq and not the opposite, because we |
12377 |
-+ * are in the context of the process owning bfqq, hence we have the |
12378 |
-+ * io_cq of this process. So we can immediately configure this io_cq |
12379 |
-+ * to redirect the requests of the process to new_bfqq. |
12380 |
-+ * |
12381 |
-+ * NOTE, even if new_bfqq coincides with the active queue, the io_cq of |
12382 |
-+ * new_bfqq is not available, because, if the active queue is shared, |
12383 |
-+ * bfqd->active_bic may not point to the io_cq of the active queue. |
12384 |
-+ * Redirecting the requests of the process owning bfqq to the currently |
12385 |
-+ * active queue is in any case the best option, as we feed the active queue |
12386 |
-+ * with new requests close to the last request served and, by doing so, |
12387 |
-+ * hopefully increase the throughput. |
12388 |
-+ */ |
12389 |
-+ bfqq->new_bfqq = new_bfqq; |
12390 |
-+ atomic_add(process_refs, &new_bfqq->ref); |
12391 |
-+ return new_bfqq; |
12392 |
-+} |
12393 |
-+ |
12394 |
-+/* |
12395 |
-+ * Attempt to schedule a merge of bfqq with the currently active queue or |
12396 |
-+ * with a close queue among the scheduled queues. |
12397 |
-+ * Return NULL if no merge was scheduled, a pointer to the shared bfq_queue |
12398 |
-+ * structure otherwise. |
12399 |
-+ */ |
12400 |
-+static struct bfq_queue * |
12401 |
-+bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
12402 |
-+ void *io_struct, bool request) |
12403 |
-+{ |
12404 |
-+ struct bfq_queue *active_bfqq, *new_bfqq; |
12405 |
-+ |
12406 |
-+ if (bfqq->new_bfqq) |
12407 |
-+ return bfqq->new_bfqq; |
12408 |
-+ |
12409 |
-+ if (!io_struct) |
12410 |
-+ return NULL; |
12411 |
-+ |
12412 |
-+ active_bfqq = bfqd->active_queue; |
12413 |
-+ |
12414 |
-+ if (active_bfqq == NULL || active_bfqq == bfqq || !bfqd->active_bic) |
12415 |
-+ goto check_scheduled; |
12416 |
-+ |
12417 |
-+ if (bfq_class_idle(active_bfqq) || bfq_class_idle(bfqq)) |
12418 |
-+ goto check_scheduled; |
12419 |
-+ |
12420 |
-+ if (bfq_class_rt(active_bfqq) != bfq_class_rt(bfqq)) |
12421 |
-+ goto check_scheduled; |
12422 |
-+ |
12423 |
-+ if (active_bfqq->entity.parent != bfqq->entity.parent) |
12424 |
-+ goto check_scheduled; |
12425 |
-+ |
12426 |
-+ if (bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) && |
12427 |
-+ bfq_bfqq_sync(active_bfqq) && bfq_bfqq_sync(bfqq)) |
12428 |
-+ if ((new_bfqq = bfq_setup_merge(bfqq, active_bfqq))) |
12429 |
-+ return new_bfqq; /* Merge with the active queue */ |
12430 |
-+ |
12431 |
-+ /* |
12432 |
-+ * Check whether there is a cooperator among currently scheduled |
12433 |
-+ * queues. The only thing we need is that the bio/request is not |
12434 |
-+ * NULL, as we need it to establish whether a cooperator exists. |
12435 |
-+ */ |
12436 |
-+check_scheduled: |
12437 |
-+ new_bfqq = bfq_close_cooperator(bfqd, bfqq, |
12438 |
-+ bfq_io_struct_pos(io_struct, request)); |
12439 |
-+ if (new_bfqq) |
12440 |
-+ return bfq_setup_merge(bfqq, new_bfqq); |
12441 |
-+ |
12442 |
-+ return NULL; |
12443 |
-+} |
12444 |
-+ |
12445 |
-+static inline void |
12446 |
-+bfq_bfqq_save_state(struct bfq_queue *bfqq) |
12447 |
-+{ |
12448 |
-+ /* |
12449 |
-+ * If bfqq->bic == NULL, the queue is already shared or its requests |
12450 |
-+ * have already been redirected to a shared queue; both idle window |
12451 |
-+ * and weight raising state have already been saved. Do nothing. |
12452 |
-+ */ |
12453 |
-+ if (bfqq->bic == NULL) |
12454 |
-+ return; |
12455 |
-+ if (bfqq->bic->raising_time_left) |
12456 |
-+ /* |
12457 |
-+ * This is the queue of a just-started process, and would |
12458 |
-+ * deserve weight raising: we set raising_time_left to the full |
12459 |
-+ * weight-raising duration to trigger weight-raising when and |
12460 |
-+ * if the queue is split and the first request of the queue |
12461 |
-+ * is enqueued. |
12462 |
-+ */ |
12463 |
-+ bfqq->bic->raising_time_left = bfq_wrais_duration(bfqq->bfqd); |
12464 |
-+ else if (bfqq->raising_coeff > 1) { |
12465 |
-+ unsigned long wrais_duration = |
12466 |
-+ jiffies - bfqq->last_rais_start_finish; |
12467 |
-+ /* |
12468 |
-+ * It may happen that a queue's weight raising period lasts |
12469 |
-+ * longer than its raising_cur_max_time, as weight raising is |
12470 |
-+ * handled only when a request is enqueued or dispatched (it |
12471 |
-+ * does not use any timer). If the weight raising period is |
12472 |
-+ * about to end, don't save it. |
12473 |
-+ */ |
12474 |
-+ if (bfqq->raising_cur_max_time <= wrais_duration) |
12475 |
-+ bfqq->bic->raising_time_left = 0; |
12476 |
-+ else |
12477 |
-+ bfqq->bic->raising_time_left = |
12478 |
-+ bfqq->raising_cur_max_time - wrais_duration; |
12479 |
-+ /* |
12480 |
-+ * The bfq_queue is becoming shared or the requests of the |
12481 |
-+ * process owning the queue are being redirected to a shared |
12482 |
-+ * queue. Stop the weight raising period of the queue, as in |
12483 |
-+ * both cases it should not be owned by an interactive or soft |
12484 |
-+ * real-time application. |
12485 |
-+ */ |
12486 |
-+ bfq_bfqq_end_raising(bfqq); |
12487 |
-+ } else |
12488 |
-+ bfqq->bic->raising_time_left = 0; |
12489 |
-+ bfqq->bic->saved_idle_window = bfq_bfqq_idle_window(bfqq); |
12490 |
-+} |
12491 |
-+ |
12492 |
-+static inline void |
12493 |
-+bfq_get_bic_reference(struct bfq_queue *bfqq) |
12494 |
-+{ |
12495 |
-+ /* |
12496 |
-+ * If bfqq->bic has a non-NULL value, the bic to which it belongs |
12497 |
-+ * is about to begin using a shared bfq_queue. |
12498 |
-+ */ |
12499 |
-+ if (bfqq->bic) |
12500 |
-+ atomic_long_inc(&bfqq->bic->icq.ioc->refcount); |
12501 |
-+} |
12502 |
-+ |
12503 |
-+static void |
12504 |
-+bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
12505 |
-+ struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
12506 |
-+{ |
12507 |
-+ bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
12508 |
-+ (long unsigned)new_bfqq->pid); |
12509 |
-+ /* Save weight raising and idle window of the merged queues */ |
12510 |
-+ bfq_bfqq_save_state(bfqq); |
12511 |
-+ bfq_bfqq_save_state(new_bfqq); |
12512 |
-+ /* |
12513 |
-+ * Grab a reference to the bic, to prevent it from being destroyed |
12514 |
-+ * before being possibly touched by a bfq_split_bfqq(). |
12515 |
-+ */ |
12516 |
-+ bfq_get_bic_reference(bfqq); |
12517 |
-+ bfq_get_bic_reference(new_bfqq); |
12518 |
-+ /* Merge queues (that is, let bic redirect its requests to new_bfqq) */ |
12519 |
-+ bic_set_bfqq(bic, new_bfqq, 1); |
12520 |
-+ bfq_mark_bfqq_coop(new_bfqq); |
12521 |
-+ /* |
12522 |
-+ * new_bfqq now belongs to at least two bics (it is a shared queue): set |
12523 |
-+ * new_bfqq->bic to NULL. bfqq either: |
12524 |
-+ * - does not belong to any bic any more, and hence bfqq->bic must |
12525 |
-+ * be set to NULL, or |
12526 |
-+ * - is a queue whose owning bics have already been redirected to a |
12527 |
-+ * different queue, hence the queue is destined to not belong to any |
12528 |
-+ * bic soon and bfqq->bic is already NULL (therefore the next |
12529 |
-+ * assignment causes no harm). |
12530 |
-+ */ |
12531 |
-+ new_bfqq->bic = NULL; |
12532 |
-+ bfqq->bic = NULL; |
12533 |
-+ bfq_put_queue(bfqq); |
12534 |
-+} |
12535 |
-+ |
12536 |
-+static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
12537 |
-+ struct bio *bio) |
12538 |
-+{ |
12539 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
12540 |
-+ struct bfq_io_cq *bic; |
12541 |
-+ struct bfq_queue *bfqq, *new_bfqq; |
12542 |
-+ |
12543 |
-+ /* |
12544 |
-+ * Disallow merge of a sync bio into an async request. |
12545 |
-+ */ |
12546 |
-+ if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
12547 |
-+ return 0; |
12548 |
-+ |
12549 |
-+ /* |
12550 |
-+ * Lookup the bfqq that this bio will be queued with. Allow |
12551 |
-+ * merge only if rq is queued there. |
12552 |
-+ * Queue lock is held here. |
12553 |
-+ */ |
12554 |
-+ bic = bfq_bic_lookup(bfqd, current->io_context); |
12555 |
-+ if (bic == NULL) |
12556 |
-+ return 0; |
12557 |
-+ |
12558 |
-+ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
12559 |
-+ /* |
12560 |
-+ * We take advantage of this function to perform an early merge |
12561 |
-+ * of the queues of possible cooperating processes. |
12562 |
-+ */ |
12563 |
-+ if (bfqq != NULL && |
12564 |
-+ (new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false))) { |
12565 |
-+ bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq); |
12566 |
-+ /* |
12567 |
-+ * If we get here, the bio will be queued in the shared queue, |
12568 |
-+ * i.e., new_bfqq, so use new_bfqq to decide whether bio and |
12569 |
-+ * rq can be merged. |
12570 |
-+ */ |
12571 |
-+ bfqq = new_bfqq; |
12572 |
-+ } |
12573 |
-+ |
12574 |
-+ return bfqq == RQ_BFQQ(rq); |
12575 |
-+} |
12576 |
-+ |
12577 |
-+static void __bfq_set_active_queue(struct bfq_data *bfqd, |
12578 |
-+ struct bfq_queue *bfqq) |
12579 |
-+{ |
12580 |
-+ if (bfqq != NULL) { |
12581 |
-+ bfq_mark_bfqq_must_alloc(bfqq); |
12582 |
-+ bfq_mark_bfqq_budget_new(bfqq); |
12583 |
-+ bfq_clear_bfqq_fifo_expire(bfqq); |
12584 |
-+ |
12585 |
-+ bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
12586 |
-+ |
12587 |
-+ bfq_log_bfqq(bfqd, bfqq, "set_active_queue, cur-budget = %lu", |
12588 |
-+ bfqq->entity.budget); |
12589 |
-+ } |
12590 |
-+ |
12591 |
-+ bfqd->active_queue = bfqq; |
12592 |
-+} |
12593 |
-+ |
12594 |
-+/* |
12595 |
-+ * Get and set a new active queue for service. |
12596 |
-+ */ |
12597 |
-+static struct bfq_queue *bfq_set_active_queue(struct bfq_data *bfqd) |
12598 |
-+{ |
12599 |
-+ struct bfq_queue *bfqq = bfq_get_next_queue(bfqd); |
12600 |
-+ |
12601 |
-+ __bfq_set_active_queue(bfqd, bfqq); |
12602 |
-+ return bfqq; |
12603 |
-+} |
12604 |
-+ |
12605 |
- /* |
12606 |
- * If enough samples have been computed, return the current max budget |
12607 |
- * stored in bfqd, which is dynamically updated according to the |
12608 |
-@@ -1017,63 +1277,6 @@ static struct request *bfq_check_fifo(struct bfq_queue *bfqq) |
12609 |
- return rq; |
12610 |
- } |
12611 |
- |
12612 |
--/* |
12613 |
-- * Must be called with the queue_lock held. |
12614 |
-- */ |
12615 |
--static int bfqq_process_refs(struct bfq_queue *bfqq) |
12616 |
--{ |
12617 |
-- int process_refs, io_refs; |
12618 |
-- |
12619 |
-- io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
12620 |
-- process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
12621 |
-- BUG_ON(process_refs < 0); |
12622 |
-- return process_refs; |
12623 |
--} |
12624 |
-- |
12625 |
--static void bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
12626 |
--{ |
12627 |
-- int process_refs, new_process_refs; |
12628 |
-- struct bfq_queue *__bfqq; |
12629 |
-- |
12630 |
-- /* |
12631 |
-- * If there are no process references on the new_bfqq, then it is |
12632 |
-- * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
12633 |
-- * may have dropped their last reference (not just their last process |
12634 |
-- * reference). |
12635 |
-- */ |
12636 |
-- if (!bfqq_process_refs(new_bfqq)) |
12637 |
-- return; |
12638 |
-- |
12639 |
-- /* Avoid a circular list and skip interim queue merges. */ |
12640 |
-- while ((__bfqq = new_bfqq->new_bfqq)) { |
12641 |
-- if (__bfqq == bfqq) |
12642 |
-- return; |
12643 |
-- new_bfqq = __bfqq; |
12644 |
-- } |
12645 |
-- |
12646 |
-- process_refs = bfqq_process_refs(bfqq); |
12647 |
-- new_process_refs = bfqq_process_refs(new_bfqq); |
12648 |
-- /* |
12649 |
-- * If the process for the bfqq has gone away, there is no |
12650 |
-- * sense in merging the queues. |
12651 |
-- */ |
12652 |
-- if (process_refs == 0 || new_process_refs == 0) |
12653 |
-- return; |
12654 |
-- |
12655 |
-- /* |
12656 |
-- * Merge in the direction of the lesser amount of work. |
12657 |
-- */ |
12658 |
-- if (new_process_refs >= process_refs) { |
12659 |
-- bfqq->new_bfqq = new_bfqq; |
12660 |
-- atomic_add(process_refs, &new_bfqq->ref); |
12661 |
-- } else { |
12662 |
-- new_bfqq->new_bfqq = bfqq; |
12663 |
-- atomic_add(new_process_refs, &bfqq->ref); |
12664 |
-- } |
12665 |
-- bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
12666 |
-- new_bfqq->pid); |
12667 |
--} |
12668 |
-- |
12669 |
- static inline unsigned long bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
12670 |
- { |
12671 |
- struct bfq_entity *entity = &bfqq->entity; |
12672 |
-@@ -1493,6 +1696,14 @@ static inline int bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) |
12673 |
- * is likely to boost the disk throughput); |
12674 |
- * - the queue is weight-raised (waiting for the request is necessary for |
12675 |
- * providing the queue with fairness and latency guarantees). |
12676 |
-+ * |
12677 |
-+ * In any case, idling can be disabled for cooperation issues, if |
12678 |
-+ * 1) there is a close cooperator for the queue, or |
12679 |
-+ * 2) the queue is shared and some cooperator is likely to be idle (in this |
12680 |
-+ * case, by not arming the idle timer, we try to slow down the queue, to |
12681 |
-+ * prevent the zones of the disk accessed by the active cooperators to |
12682 |
-+ * become too distant from the zone that will be accessed by the currently |
12683 |
-+ * idle cooperators). |
12684 |
- */ |
12685 |
- static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq, |
12686 |
- int budg_timeout) |
12687 |
-@@ -1507,7 +1718,7 @@ static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq, |
12688 |
- (bfqd->rq_in_driver == 0 || |
12689 |
- budg_timeout || |
12690 |
- bfqq->raising_coeff > 1) && |
12691 |
-- !bfq_close_cooperator(bfqd, bfqq) && |
12692 |
-+ !bfq_close_cooperator(bfqd, bfqq, bfqd->last_position) && |
12693 |
- (!bfq_bfqq_coop(bfqq) || |
12694 |
- !bfq_bfqq_some_coop_idle(bfqq)) && |
12695 |
- !bfq_queue_nonrot_noidle(bfqd, bfqq)); |
12696 |
-@@ -1519,7 +1730,7 @@ static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq, |
12697 |
- */ |
12698 |
- static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
12699 |
- { |
12700 |
-- struct bfq_queue *bfqq, *new_bfqq = NULL; |
12701 |
-+ struct bfq_queue *bfqq; |
12702 |
- struct request *next_rq; |
12703 |
- enum bfqq_expiration reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
12704 |
- int budg_timeout; |
12705 |
-@@ -1530,17 +1741,6 @@ static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
12706 |
- |
12707 |
- bfq_log_bfqq(bfqd, bfqq, "select_queue: already active queue"); |
12708 |
- |
12709 |
-- /* |
12710 |
-- * If another queue has a request waiting within our mean seek |
12711 |
-- * distance, let it run. The expire code will check for close |
12712 |
-- * cooperators and put the close queue at the front of the |
12713 |
-- * service tree. If possible, merge the expiring queue with the |
12714 |
-- * new bfqq. |
12715 |
-- */ |
12716 |
-- new_bfqq = bfq_close_cooperator(bfqd, bfqq); |
12717 |
-- if (new_bfqq != NULL && bfqq->new_bfqq == NULL) |
12718 |
-- bfq_setup_merge(bfqq, new_bfqq); |
12719 |
-- |
12720 |
- budg_timeout = bfq_may_expire_for_budg_timeout(bfqq); |
12721 |
- if (budg_timeout && |
12722 |
- !bfq_bfqq_must_idle(bfqq, budg_timeout)) |
12723 |
-@@ -1577,10 +1777,7 @@ static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
12724 |
- bfq_clear_bfqq_wait_request(bfqq); |
12725 |
- del_timer(&bfqd->idle_slice_timer); |
12726 |
- } |
12727 |
-- if (new_bfqq == NULL) |
12728 |
-- goto keep_queue; |
12729 |
-- else |
12730 |
-- goto expire; |
12731 |
-+ goto keep_queue; |
12732 |
- } |
12733 |
- } |
12734 |
- |
12735 |
-@@ -1589,26 +1786,19 @@ static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
12736 |
- * queue still has requests in flight or is idling for a new request, |
12737 |
- * then keep it. |
12738 |
- */ |
12739 |
-- if (new_bfqq == NULL && (timer_pending(&bfqd->idle_slice_timer) || |
12740 |
-+ if (timer_pending(&bfqd->idle_slice_timer) || |
12741 |
- (bfqq->dispatched != 0 && |
12742 |
- (bfq_bfqq_idle_window(bfqq) || bfqq->raising_coeff > 1) && |
12743 |
-- !bfq_queue_nonrot_noidle(bfqd, bfqq)))) { |
12744 |
-+ !bfq_queue_nonrot_noidle(bfqd, bfqq))) { |
12745 |
- bfqq = NULL; |
12746 |
- goto keep_queue; |
12747 |
-- } else if (new_bfqq != NULL && timer_pending(&bfqd->idle_slice_timer)) { |
12748 |
-- /* |
12749 |
-- * Expiring the queue because there is a close cooperator, |
12750 |
-- * cancel timer. |
12751 |
-- */ |
12752 |
-- bfq_clear_bfqq_wait_request(bfqq); |
12753 |
-- del_timer(&bfqd->idle_slice_timer); |
12754 |
- } |
12755 |
- |
12756 |
- reason = BFQ_BFQQ_NO_MORE_REQUESTS; |
12757 |
- expire: |
12758 |
- bfq_bfqq_expire(bfqd, bfqq, 0, reason); |
12759 |
- new_queue: |
12760 |
-- bfqq = bfq_set_active_queue(bfqd, new_bfqq); |
12761 |
-+ bfqq = bfq_set_active_queue(bfqd); |
12762 |
- bfq_log(bfqd, "select_queue: new queue %d returned", |
12763 |
- bfqq != NULL ? bfqq->pid : 0); |
12764 |
- keep_queue: |
12765 |
-@@ -1617,9 +1807,8 @@ keep_queue: |
12766 |
- |
12767 |
- static void update_raising_data(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
12768 |
- { |
12769 |
-+ struct bfq_entity *entity = &bfqq->entity; |
12770 |
- if (bfqq->raising_coeff > 1) { /* queue is being boosted */ |
12771 |
-- struct bfq_entity *entity = &bfqq->entity; |
12772 |
-- |
12773 |
- bfq_log_bfqq(bfqd, bfqq, |
12774 |
- "raising period dur %u/%u msec, " |
12775 |
- "old raising coeff %u, w %d(%d)", |
12776 |
-@@ -1656,12 +1845,14 @@ static void update_raising_data(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
12777 |
- jiffies_to_msecs(bfqq-> |
12778 |
- raising_cur_max_time)); |
12779 |
- bfq_bfqq_end_raising(bfqq); |
12780 |
-- __bfq_entity_update_weight_prio( |
12781 |
-- bfq_entity_service_tree(entity), |
12782 |
-- entity); |
12783 |
- } |
12784 |
- } |
12785 |
- } |
12786 |
-+ /* Update weight both if it must be raised and if it must be lowered */ |
12787 |
-+ if ((entity->weight > entity->orig_weight) != (bfqq->raising_coeff > 1)) |
12788 |
-+ __bfq_entity_update_weight_prio( |
12789 |
-+ bfq_entity_service_tree(entity), |
12790 |
-+ entity); |
12791 |
- } |
12792 |
- |
12793 |
- /* |
12794 |
-@@ -1901,6 +2092,25 @@ static void bfq_init_icq(struct io_cq *icq) |
12795 |
- struct bfq_io_cq *bic = icq_to_bic(icq); |
12796 |
- |
12797 |
- bic->ttime.last_end_request = jiffies; |
12798 |
-+ /* |
12799 |
-+ * A newly created bic indicates that the process has just |
12800 |
-+ * started doing I/O, and is probably mapping into memory its |
12801 |
-+ * executable and libraries: it definitely needs weight raising. |
12802 |
-+ * There is however the possibility that the process performs, |
12803 |
-+ * for a while, I/O close to some other process. EQM intercepts |
12804 |
-+ * this behavior and may merge the queue corresponding to the |
12805 |
-+ * process with some other queue, BEFORE the weight of the queue |
12806 |
-+ * is raised. Merged queues are not weight-raised (they are assumed |
12807 |
-+ * to belong to processes that benefit only from high throughput). |
12808 |
-+ * If the merge is basically the consequence of an accident, then |
12809 |
-+ * the queue will be split soon and will get back its old weight. |
12810 |
-+ * It is then important to write down somewhere that this queue |
12811 |
-+ * does need weight raising, even if it did not make it to get its |
12812 |
-+ * weight raised before being merged. To this purpose, we overload |
12813 |
-+ * the field raising_time_left and assign 1 to it, to mark the queue |
12814 |
-+ * as needing weight raising. |
12815 |
-+ */ |
12816 |
-+ bic->raising_time_left = 1; |
12817 |
- } |
12818 |
- |
12819 |
- static void bfq_exit_icq(struct io_cq *icq) |
12820 |
-@@ -1914,6 +2124,13 @@ static void bfq_exit_icq(struct io_cq *icq) |
12821 |
- } |
12822 |
- |
12823 |
- if (bic->bfqq[BLK_RW_SYNC]) { |
12824 |
-+ /* |
12825 |
-+ * If the bic is using a shared queue, put the reference |
12826 |
-+ * taken on the io_context when the bic started using a |
12827 |
-+ * shared bfq_queue. |
12828 |
-+ */ |
12829 |
-+ if (bfq_bfqq_coop(bic->bfqq[BLK_RW_SYNC])) |
12830 |
-+ put_io_context(icq->ioc); |
12831 |
- bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_SYNC]); |
12832 |
- bic->bfqq[BLK_RW_SYNC] = NULL; |
12833 |
- } |
12834 |
-@@ -2211,6 +2428,10 @@ static void bfq_update_idle_window(struct bfq_data *bfqd, |
12835 |
- if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq)) |
12836 |
- return; |
12837 |
- |
12838 |
-+ /* Idle window just restored, statistics are meaningless. */ |
12839 |
-+ if (bfq_bfqq_just_split(bfqq)) |
12840 |
-+ return; |
12841 |
-+ |
12842 |
- enable_idle = bfq_bfqq_idle_window(bfqq); |
12843 |
- |
12844 |
- if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
12845 |
-@@ -2251,6 +2472,7 @@ static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
12846 |
- if (bfqq->entity.service > bfq_max_budget(bfqd) / 8 || |
12847 |
- !BFQQ_SEEKY(bfqq)) |
12848 |
- bfq_update_idle_window(bfqd, bfqq, bic); |
12849 |
-+ bfq_clear_bfqq_just_split(bfqq); |
12850 |
- |
12851 |
- bfq_log_bfqq(bfqd, bfqq, |
12852 |
- "rq_enqueued: idle_window=%d (seeky %d, mean %llu)", |
12853 |
-@@ -2302,13 +2524,45 @@ static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
12854 |
- static void bfq_insert_request(struct request_queue *q, struct request *rq) |
12855 |
- { |
12856 |
- struct bfq_data *bfqd = q->elevator->elevator_data; |
12857 |
-- struct bfq_queue *bfqq = RQ_BFQQ(rq); |
12858 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq), *new_bfqq; |
12859 |
- |
12860 |
- assert_spin_locked(bfqd->queue->queue_lock); |
12861 |
-+ |
12862 |
-+ /* |
12863 |
-+ * An unplug may trigger a requeue of a request from the device |
12864 |
-+ * driver: make sure we are in process context while trying to |
12865 |
-+ * merge two bfq_queues. |
12866 |
-+ */ |
12867 |
-+ if (!in_interrupt() && |
12868 |
-+ (new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true))) { |
12869 |
-+ if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq) |
12870 |
-+ new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1); |
12871 |
-+ /* |
12872 |
-+ * Release the request's reference to the old bfqq |
12873 |
-+ * and make sure one is taken to the shared queue. |
12874 |
-+ */ |
12875 |
-+ new_bfqq->allocated[rq_data_dir(rq)]++; |
12876 |
-+ bfqq->allocated[rq_data_dir(rq)]--; |
12877 |
-+ atomic_inc(&new_bfqq->ref); |
12878 |
-+ bfq_put_queue(bfqq); |
12879 |
-+ if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq) |
12880 |
-+ bfq_merge_bfqqs(bfqd, RQ_BIC(rq), bfqq, new_bfqq); |
12881 |
-+ rq->elv.priv[1] = new_bfqq; |
12882 |
-+ bfqq = new_bfqq; |
12883 |
-+ } |
12884 |
-+ |
12885 |
- bfq_init_prio_data(bfqq, RQ_BIC(rq)); |
12886 |
- |
12887 |
- bfq_add_rq_rb(rq); |
12888 |
- |
12889 |
-+ /* |
12890 |
-+ * Here a newly-created bfq_queue has already started a weight-raising |
12891 |
-+ * period: clear raising_time_left to prevent bfq_bfqq_save_state() |
12892 |
-+ * from assigning it a full weight-raising period. See the detailed |
12893 |
-+ * comments about this field in bfq_init_icq(). |
12894 |
-+ */ |
12895 |
-+ if (bfqq->bic != NULL) |
12896 |
-+ bfqq->bic->raising_time_left = 0; |
12897 |
- rq_set_fifo_time(rq, jiffies + bfqd->bfq_fifo_expire[rq_is_sync(rq)]); |
12898 |
- list_add_tail(&rq->queuelist, &bfqq->fifo); |
12899 |
- |
12900 |
-@@ -2371,15 +2625,6 @@ static void bfq_completed_request(struct request_queue *q, struct request *rq) |
12901 |
- if (bfq_bfqq_budget_new(bfqq)) |
12902 |
- bfq_set_budget_timeout(bfqd); |
12903 |
- |
12904 |
-- /* Idling is disabled also for cooperation issues: |
12905 |
-- * 1) there is a close cooperator for the queue, or |
12906 |
-- * 2) the queue is shared and some cooperator is likely |
12907 |
-- * to be idle (in this case, by not arming the idle timer, |
12908 |
-- * we try to slow down the queue, to prevent the zones |
12909 |
-- * of the disk accessed by the active cooperators to become |
12910 |
-- * too distant from the zone that will be accessed by the |
12911 |
-- * currently idle cooperators) |
12912 |
-- */ |
12913 |
- if (bfq_bfqq_must_idle(bfqq, budg_timeout)) |
12914 |
- bfq_arm_slice_timer(bfqd); |
12915 |
- else if (budg_timeout) |
12916 |
-@@ -2449,18 +2694,6 @@ static void bfq_put_request(struct request *rq) |
12917 |
- } |
12918 |
- } |
12919 |
- |
12920 |
--static struct bfq_queue * |
12921 |
--bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
12922 |
-- struct bfq_queue *bfqq) |
12923 |
--{ |
12924 |
-- bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
12925 |
-- (long unsigned)bfqq->new_bfqq->pid); |
12926 |
-- bic_set_bfqq(bic, bfqq->new_bfqq, 1); |
12927 |
-- bfq_mark_bfqq_coop(bfqq->new_bfqq); |
12928 |
-- bfq_put_queue(bfqq); |
12929 |
-- return bic_to_bfqq(bic, 1); |
12930 |
--} |
12931 |
-- |
12932 |
- /* |
12933 |
- * Returns NULL if a new bfqq should be allocated, or the old bfqq if this |
12934 |
- * was the last process referring to said bfqq. |
12935 |
-@@ -2469,6 +2702,9 @@ static struct bfq_queue * |
12936 |
- bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) |
12937 |
- { |
12938 |
- bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); |
12939 |
-+ |
12940 |
-+ put_io_context(bic->icq.ioc); |
12941 |
-+ |
12942 |
- if (bfqq_process_refs(bfqq) == 1) { |
12943 |
- bfqq->pid = current->pid; |
12944 |
- bfq_clear_bfqq_some_coop_idle(bfqq); |
12945 |
-@@ -2498,6 +2734,7 @@ static int bfq_set_request(struct request_queue *q, struct request *rq, |
12946 |
- struct bfq_queue *bfqq; |
12947 |
- struct bfq_group *bfqg; |
12948 |
- unsigned long flags; |
12949 |
-+ bool split = false; |
12950 |
- |
12951 |
- might_sleep_if(gfp_mask & __GFP_WAIT); |
12952 |
- |
12953 |
-@@ -2516,24 +2753,14 @@ new_queue: |
12954 |
- bfqq = bfq_get_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
12955 |
- bic_set_bfqq(bic, bfqq, is_sync); |
12956 |
- } else { |
12957 |
-- /* |
12958 |
-- * If the queue was seeky for too long, break it apart. |
12959 |
-- */ |
12960 |
-+ /* If the queue was seeky for too long, break it apart. */ |
12961 |
- if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { |
12962 |
- bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); |
12963 |
- bfqq = bfq_split_bfqq(bic, bfqq); |
12964 |
-+ split = true; |
12965 |
- if (!bfqq) |
12966 |
- goto new_queue; |
12967 |
- } |
12968 |
-- |
12969 |
-- /* |
12970 |
-- * Check to see if this queue is scheduled to merge with |
12971 |
-- * another closely cooperating queue. The merging of queues |
12972 |
-- * happens here as it must be done in process context. |
12973 |
-- * The reference on new_bfqq was taken in merge_bfqqs. |
12974 |
-- */ |
12975 |
-- if (bfqq->new_bfqq != NULL) |
12976 |
-- bfqq = bfq_merge_bfqqs(bfqd, bic, bfqq); |
12977 |
- } |
12978 |
- |
12979 |
- bfqq->allocated[rw]++; |
12980 |
-@@ -2544,6 +2771,26 @@ new_queue: |
12981 |
- rq->elv.priv[0] = bic; |
12982 |
- rq->elv.priv[1] = bfqq; |
12983 |
- |
12984 |
-+ /* |
12985 |
-+ * If a bfq_queue has only one process reference, it is owned |
12986 |
-+ * by only one bfq_io_cq: we can set the bic field of the |
12987 |
-+ * bfq_queue to the address of that structure. Also, if the |
12988 |
-+ * queue has just been split, mark a flag so that the |
12989 |
-+ * information is available to the other scheduler hooks. |
12990 |
-+ */ |
12991 |
-+ if (bfqq_process_refs(bfqq) == 1) { |
12992 |
-+ bfqq->bic = bic; |
12993 |
-+ if (split) { |
12994 |
-+ bfq_mark_bfqq_just_split(bfqq); |
12995 |
-+ /* |
12996 |
-+ * If the queue has just been split from a shared queue, |
12997 |
-+ * restore the idle window and the possible weight |
12998 |
-+ * raising period. |
12999 |
-+ */ |
13000 |
-+ bfq_bfqq_resume_state(bfqq, bic); |
13001 |
-+ } |
13002 |
-+ } |
13003 |
-+ |
13004 |
- spin_unlock_irqrestore(q->queue_lock, flags); |
13005 |
- |
13006 |
- return 0; |
13007 |
-diff --git a/block/bfq-sched.c b/block/bfq-sched.c |
13008 |
-index 03f8061..a0edaa2 100644 |
13009 |
---- a/block/bfq-sched.c |
13010 |
-+++ b/block/bfq-sched.c |
13011 |
-@@ -978,34 +978,6 @@ static struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd) |
13012 |
- return bfqq; |
13013 |
- } |
13014 |
- |
13015 |
--/* |
13016 |
-- * Forced extraction of the given queue. |
13017 |
-- */ |
13018 |
--static void bfq_get_next_queue_forced(struct bfq_data *bfqd, |
13019 |
-- struct bfq_queue *bfqq) |
13020 |
--{ |
13021 |
-- struct bfq_entity *entity; |
13022 |
-- struct bfq_sched_data *sd; |
13023 |
-- |
13024 |
-- BUG_ON(bfqd->active_queue != NULL); |
13025 |
-- |
13026 |
-- entity = &bfqq->entity; |
13027 |
-- /* |
13028 |
-- * Bubble up extraction/update from the leaf to the root. |
13029 |
-- */ |
13030 |
-- for_each_entity(entity) { |
13031 |
-- sd = entity->sched_data; |
13032 |
-- bfq_update_budget(entity); |
13033 |
-- bfq_update_vtime(bfq_entity_service_tree(entity)); |
13034 |
-- bfq_active_extract(bfq_entity_service_tree(entity), entity); |
13035 |
-- sd->active_entity = entity; |
13036 |
-- sd->next_active = NULL; |
13037 |
-- entity->service = 0; |
13038 |
-- } |
13039 |
-- |
13040 |
-- return; |
13041 |
--} |
13042 |
-- |
13043 |
- static void __bfq_bfqd_reset_active(struct bfq_data *bfqd) |
13044 |
- { |
13045 |
- if (bfqd->active_bic != NULL) { |
13046 |
-diff --git a/block/bfq.h b/block/bfq.h |
13047 |
-index 48ecde9..bb52975 100644 |
13048 |
---- a/block/bfq.h |
13049 |
-+++ b/block/bfq.h |
13050 |
-@@ -188,6 +188,8 @@ struct bfq_group; |
13051 |
- * @pid: pid of the process owning the queue, used for logging purposes. |
13052 |
- * @last_rais_start_time: last (idle -> weight-raised) transition attempt |
13053 |
- * @raising_cur_max_time: current max raising time for this queue |
13054 |
-+ * @bic: pointer to the bfq_io_cq owning the bfq_queue, set to %NULL if the |
13055 |
-+ * queue is shared |
13056 |
- * |
13057 |
- * A bfq_queue is a leaf request queue; it can be associated to an io_context |
13058 |
- * or more (if it is an async one). @cgroup holds a reference to the |
13059 |
-@@ -231,6 +233,7 @@ struct bfq_queue { |
13060 |
- sector_t last_request_pos; |
13061 |
- |
13062 |
- pid_t pid; |
13063 |
-+ struct bfq_io_cq *bic; |
13064 |
- |
13065 |
- /* weight-raising fields */ |
13066 |
- unsigned int raising_cur_max_time; |
13067 |
-@@ -257,12 +260,23 @@ struct bfq_ttime { |
13068 |
- * @icq: associated io_cq structure |
13069 |
- * @bfqq: array of two process queues, the sync and the async |
13070 |
- * @ttime: associated @bfq_ttime struct |
13071 |
-+ * @raising_time_left: snapshot of the time left before weight raising ends |
13072 |
-+ * for the sync queue associated to this process; this |
13073 |
-+ * snapshot is taken to remember this value while the weight |
13074 |
-+ * raising is suspended because the queue is merged with a |
13075 |
-+ * shared queue, and is used to set @raising_cur_max_time |
13076 |
-+ * when the queue is split from the shared queue and its |
13077 |
-+ * weight is raised again |
13078 |
-+ * @saved_idle_window: same purpose as the previous field for the idle window |
13079 |
- */ |
13080 |
- struct bfq_io_cq { |
13081 |
- struct io_cq icq; /* must be the first member */ |
13082 |
- struct bfq_queue *bfqq[2]; |
13083 |
- struct bfq_ttime ttime; |
13084 |
- int ioprio; |
13085 |
-+ |
13086 |
-+ unsigned int raising_time_left; |
13087 |
-+ unsigned int saved_idle_window; |
13088 |
- }; |
13089 |
- |
13090 |
- /** |
13091 |
-@@ -403,6 +417,7 @@ enum bfqq_state_flags { |
13092 |
- BFQ_BFQQ_FLAG_coop, /* bfqq is shared */ |
13093 |
- BFQ_BFQQ_FLAG_split_coop, /* shared bfqq will be splitted */ |
13094 |
- BFQ_BFQQ_FLAG_some_coop_idle, /* some cooperator is inactive */ |
13095 |
-+ BFQ_BFQQ_FLAG_just_split, /* queue has just been split */ |
13096 |
- }; |
13097 |
- |
13098 |
- #define BFQ_BFQQ_FNS(name) \ |
13099 |
-@@ -430,6 +445,7 @@ BFQ_BFQQ_FNS(budget_new); |
13100 |
- BFQ_BFQQ_FNS(coop); |
13101 |
- BFQ_BFQQ_FNS(split_coop); |
13102 |
- BFQ_BFQQ_FNS(some_coop_idle); |
13103 |
-+BFQ_BFQQ_FNS(just_split); |
13104 |
- #undef BFQ_BFQQ_FNS |
13105 |
- |
13106 |
- /* Logging facilities. */ |
13107 |
--- |
13108 |
-1.8.1.4 |
13109 |
- |
13110 |
|
13111 |
Added: genpatches-2.6/trunk/3.14/5000_BFQ-3-block-bfq-add-Early-Queue-Merge-EQM-to-BFQ-v7-for-3.13.0.patch |
13112 |
=================================================================== |
13113 |
--- genpatches-2.6/trunk/3.14/5000_BFQ-3-block-bfq-add-Early-Queue-Merge-EQM-to-BFQ-v7-for-3.13.0.patch (rev 0) |
13114 |
+++ genpatches-2.6/trunk/3.14/5000_BFQ-3-block-bfq-add-Early-Queue-Merge-EQM-to-BFQ-v7-for-3.13.0.patch 2014-01-30 16:49:47 UTC (rev 2661) |
13115 |
@@ -0,0 +1,1034 @@ |
13116 |
+From 3cd9e2ea29c3ba9e420556e8ecf161d166186b63 Mon Sep 17 00:00:00 2001 |
13117 |
+From: Mauro Andreolini <mauro.andreolini@×××××××.it> |
13118 |
+Date: Thu, 23 Jan 2014 16:54:44 +0100 |
13119 |
+Subject: [PATCH 3/3] block, bfq: add Early Queue Merge (EQM) to BFQ-v7 for |
13120 |
+ 3.13.0 |
13121 |
+ |
13122 |
+A set of processes may happen to perform interleaved reads, i.e., requests |
13123 |
+whose union would give rise to a sequential read pattern. There are two |
13124 |
+typical cases: in the first case, processes read fixed-size chunks of |
13125 |
+data at a fixed distance from each other, while in the second case processes |
13126 |
+may read variable-size chunks at variable distances. The latter case occurs |
13127 |
+for example with KVM, which splits the I/O generated by the guest into |
13128 |
+multiple chunks, and lets these chunks be served by a pool of cooperating |
13129 |
+processes, iteratively assigning the next chunk of I/O to the first |
13130 |
+available process. CFQ uses actual queue merging for the first type of |
13131 |
+rocesses, whereas it uses preemption to get a sequential read pattern out |
13132 |
+of the read requests performed by the second type of processes. In the end |
13133 |
+it uses two different mechanisms to achieve the same goal: boosting the |
13134 |
+throughput with interleaved I/O. |
13135 |
+ |
13136 |
+This patch introduces Early Queue Merge (EQM), a unified mechanism to get a |
13137 |
+sequential read pattern with both types of processes. The main idea is |
13138 |
+checking newly arrived requests against the next request of the active queue |
13139 |
+both in case of actual request insert and in case of request merge. By doing |
13140 |
+so, both the types of processes can be handled by just merging their queues. |
13141 |
+EQM is then simpler and more compact than the pair of mechanisms used in |
13142 |
+CFQ. |
13143 |
+ |
13144 |
+Finally, EQM also preserves the typical low-latency properties of BFQ, by |
13145 |
+properly restoring the weight-raising state of a queue when it gets back to |
13146 |
+a non-merged state. |
13147 |
+ |
13148 |
+Signed-off-by: Mauro Andreolini <mauro.andreolini@×××××××.it> |
13149 |
+Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
13150 |
+Reviewed-by: Paolo Valente <paolo.valente@×××××××.it> |
13151 |
+--- |
13152 |
+ block/bfq-iosched.c | 657 ++++++++++++++++++++++++++++++++++++---------------- |
13153 |
+ block/bfq-sched.c | 28 --- |
13154 |
+ block/bfq.h | 16 ++ |
13155 |
+ 3 files changed, 474 insertions(+), 227 deletions(-) |
13156 |
+ |
13157 |
+diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c |
13158 |
+index 7670400..295236e 100644 |
13159 |
+--- a/block/bfq-iosched.c |
13160 |
++++ b/block/bfq-iosched.c |
13161 |
+@@ -445,6 +445,46 @@ static inline unsigned int bfq_wrais_duration(struct bfq_data *bfqd) |
13162 |
+ return dur; |
13163 |
+ } |
13164 |
+ |
13165 |
++static inline void |
13166 |
++bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_io_cq *bic) |
13167 |
++{ |
13168 |
++ if (bic->saved_idle_window) |
13169 |
++ bfq_mark_bfqq_idle_window(bfqq); |
13170 |
++ else |
13171 |
++ bfq_clear_bfqq_idle_window(bfqq); |
13172 |
++ if (bic->raising_time_left && bfqq->bfqd->low_latency) { |
13173 |
++ /* |
13174 |
++ * Start a weight raising period with the duration given by |
13175 |
++ * the raising_time_left snapshot. |
13176 |
++ */ |
13177 |
++ if (bfq_bfqq_busy(bfqq)) |
13178 |
++ bfqq->bfqd->raised_busy_queues++; |
13179 |
++ bfqq->raising_coeff = bfqq->bfqd->bfq_raising_coeff; |
13180 |
++ bfqq->raising_cur_max_time = bic->raising_time_left; |
13181 |
++ bfqq->last_rais_start_finish = jiffies; |
13182 |
++ bfqq->entity.ioprio_changed = 1; |
13183 |
++ } |
13184 |
++ /* |
13185 |
++ * Clear raising_time_left to prevent bfq_bfqq_save_state() from |
13186 |
++ * getting confused about the queue's need of a weight-raising |
13187 |
++ * period. |
13188 |
++ */ |
13189 |
++ bic->raising_time_left = 0; |
13190 |
++} |
13191 |
++ |
13192 |
++/* |
13193 |
++ * Must be called with the queue_lock held. |
13194 |
++ */ |
13195 |
++static int bfqq_process_refs(struct bfq_queue *bfqq) |
13196 |
++{ |
13197 |
++ int process_refs, io_refs; |
13198 |
++ |
13199 |
++ io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
13200 |
++ process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
13201 |
++ BUG_ON(process_refs < 0); |
13202 |
++ return process_refs; |
13203 |
++} |
13204 |
++ |
13205 |
+ static void bfq_add_rq_rb(struct request *rq) |
13206 |
+ { |
13207 |
+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
13208 |
+@@ -486,12 +526,20 @@ static void bfq_add_rq_rb(struct request *rq) |
13209 |
+ if (!bfqd->low_latency) |
13210 |
+ goto add_bfqq_busy; |
13211 |
+ |
13212 |
++ if (bfq_bfqq_just_split(bfqq)) |
13213 |
++ goto set_ioprio_changed; |
13214 |
++ |
13215 |
+ /* |
13216 |
+- * If the queue is not being boosted and has been idle |
13217 |
+- * for enough time, start a weight-raising period |
13218 |
++ * If the queue: |
13219 |
++ * - is not being boosted, |
13220 |
++ * - has been idle for enough time, |
13221 |
++ * - is not a sync queue or is linked to a bfq_io_cq (it is |
13222 |
++ * shared "for its nature" or it is not shared and its |
13223 |
++ * requests have not been redirected to a shared queue) |
13224 |
++ * start a weight-raising period. |
13225 |
+ */ |
13226 |
+- if (old_raising_coeff == 1 && |
13227 |
+- (idle_for_long_time || soft_rt)) { |
13228 |
++ if (old_raising_coeff == 1 && (idle_for_long_time || soft_rt) && |
13229 |
++ (!bfq_bfqq_sync(bfqq) || bfqq->bic != NULL)) { |
13230 |
+ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
13231 |
+ if (idle_for_long_time) |
13232 |
+ bfqq->raising_cur_max_time = |
13233 |
+@@ -572,6 +620,7 @@ static void bfq_add_rq_rb(struct request *rq) |
13234 |
+ bfqd->bfq_raising_rt_max_time; |
13235 |
+ } |
13236 |
+ } |
13237 |
++set_ioprio_changed: |
13238 |
+ if (old_raising_coeff != bfqq->raising_coeff) |
13239 |
+ entity->ioprio_changed = 1; |
13240 |
+ add_bfqq_busy: |
13241 |
+@@ -754,90 +803,35 @@ static void bfq_end_raising(struct bfq_data *bfqd) |
13242 |
+ spin_unlock_irq(bfqd->queue->queue_lock); |
13243 |
+ } |
13244 |
+ |
13245 |
+-static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
13246 |
+- struct bio *bio) |
13247 |
+-{ |
13248 |
+- struct bfq_data *bfqd = q->elevator->elevator_data; |
13249 |
+- struct bfq_io_cq *bic; |
13250 |
+- struct bfq_queue *bfqq; |
13251 |
+- |
13252 |
+- /* |
13253 |
+- * Disallow merge of a sync bio into an async request. |
13254 |
+- */ |
13255 |
+- if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
13256 |
+- return 0; |
13257 |
+- |
13258 |
+- /* |
13259 |
+- * Lookup the bfqq that this bio will be queued with. Allow |
13260 |
+- * merge only if rq is queued there. |
13261 |
+- * Queue lock is held here. |
13262 |
+- */ |
13263 |
+- bic = bfq_bic_lookup(bfqd, current->io_context); |
13264 |
+- if (bic == NULL) |
13265 |
+- return 0; |
13266 |
+- |
13267 |
+- bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
13268 |
+- return bfqq == RQ_BFQQ(rq); |
13269 |
+-} |
13270 |
+- |
13271 |
+-static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
13272 |
+- struct bfq_queue *bfqq) |
13273 |
+-{ |
13274 |
+- if (bfqq != NULL) { |
13275 |
+- bfq_mark_bfqq_must_alloc(bfqq); |
13276 |
+- bfq_mark_bfqq_budget_new(bfqq); |
13277 |
+- bfq_clear_bfqq_fifo_expire(bfqq); |
13278 |
+- |
13279 |
+- bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
13280 |
+- |
13281 |
+- bfq_log_bfqq(bfqd, bfqq, |
13282 |
+- "set_in_service_queue, cur-budget = %lu", |
13283 |
+- bfqq->entity.budget); |
13284 |
+- } |
13285 |
+- |
13286 |
+- bfqd->in_service_queue = bfqq; |
13287 |
+-} |
13288 |
+- |
13289 |
+-/* |
13290 |
+- * Get and set a new queue for service. |
13291 |
+- */ |
13292 |
+-static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd, |
13293 |
+- struct bfq_queue *bfqq) |
13294 |
++static inline sector_t bfq_io_struct_pos(void *io_struct, bool request) |
13295 |
+ { |
13296 |
+- if (!bfqq) |
13297 |
+- bfqq = bfq_get_next_queue(bfqd); |
13298 |
++ if (request) |
13299 |
++ return blk_rq_pos(io_struct); |
13300 |
+ else |
13301 |
+- bfq_get_next_queue_forced(bfqd, bfqq); |
13302 |
+- |
13303 |
+- __bfq_set_in_service_queue(bfqd, bfqq); |
13304 |
+- return bfqq; |
13305 |
++ return ((struct bio *)io_struct)->bi_sector; |
13306 |
+ } |
13307 |
+ |
13308 |
+-static inline sector_t bfq_dist_from_last(struct bfq_data *bfqd, |
13309 |
+- struct request *rq) |
13310 |
++static inline sector_t bfq_dist_from(sector_t pos1, |
13311 |
++ sector_t pos2) |
13312 |
+ { |
13313 |
+- if (blk_rq_pos(rq) >= bfqd->last_position) |
13314 |
+- return blk_rq_pos(rq) - bfqd->last_position; |
13315 |
++ if (pos1 >= pos2) |
13316 |
++ return pos1 - pos2; |
13317 |
+ else |
13318 |
+- return bfqd->last_position - blk_rq_pos(rq); |
13319 |
++ return pos2 - pos1; |
13320 |
+ } |
13321 |
+ |
13322 |
+-/* |
13323 |
+- * Return true if bfqq has no request pending and rq is close enough to |
13324 |
+- * bfqd->last_position, or if rq is closer to bfqd->last_position than |
13325 |
+- * bfqq->next_rq |
13326 |
+- */ |
13327 |
+-static inline int bfq_rq_close(struct bfq_data *bfqd, struct request *rq) |
13328 |
++static inline int bfq_rq_close_to_sector(void *io_struct, bool request, |
13329 |
++ sector_t sector) |
13330 |
+ { |
13331 |
+- return bfq_dist_from_last(bfqd, rq) <= BFQQ_SEEK_THR; |
13332 |
++ return bfq_dist_from(bfq_io_struct_pos(io_struct, request), sector) <= |
13333 |
++ BFQQ_SEEK_THR; |
13334 |
+ } |
13335 |
+ |
13336 |
+-static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
13337 |
++static struct bfq_queue *bfqq_close(struct bfq_data *bfqd, sector_t sector) |
13338 |
+ { |
13339 |
+ struct rb_root *root = &bfqd->rq_pos_tree; |
13340 |
+ struct rb_node *parent, *node; |
13341 |
+ struct bfq_queue *__bfqq; |
13342 |
+- sector_t sector = bfqd->last_position; |
13343 |
+ |
13344 |
+ if (RB_EMPTY_ROOT(root)) |
13345 |
+ return NULL; |
13346 |
+@@ -856,7 +850,7 @@ static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
13347 |
+ * position). |
13348 |
+ */ |
13349 |
+ __bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
13350 |
+- if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
13351 |
++ if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) |
13352 |
+ return __bfqq; |
13353 |
+ |
13354 |
+ if (blk_rq_pos(__bfqq->next_rq) < sector) |
13355 |
+@@ -867,7 +861,7 @@ static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
13356 |
+ return NULL; |
13357 |
+ |
13358 |
+ __bfqq = rb_entry(node, struct bfq_queue, pos_node); |
13359 |
+- if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
13360 |
++ if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector)) |
13361 |
+ return __bfqq; |
13362 |
+ |
13363 |
+ return NULL; |
13364 |
+@@ -876,14 +870,12 @@ static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
13365 |
+ /* |
13366 |
+ * bfqd - obvious |
13367 |
+ * cur_bfqq - passed in so that we don't decide that the current queue |
13368 |
+- * is closely cooperating with itself. |
13369 |
+- * |
13370 |
+- * We are assuming that cur_bfqq has dispatched at least one request, |
13371 |
+- * and that bfqd->last_position reflects a position on the disk associated |
13372 |
+- * with the I/O issued by cur_bfqq. |
13373 |
++ * is closely cooperating with itself |
13374 |
++ * sector - used as a reference point to search for a close queue |
13375 |
+ */ |
13376 |
+ static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
13377 |
+- struct bfq_queue *cur_bfqq) |
13378 |
++ struct bfq_queue *cur_bfqq, |
13379 |
++ sector_t sector) |
13380 |
+ { |
13381 |
+ struct bfq_queue *bfqq; |
13382 |
+ |
13383 |
+@@ -903,7 +895,7 @@ static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
13384 |
+ * working closely on the same area of the disk. In that case, |
13385 |
+ * we can group them together and don't waste time idling. |
13386 |
+ */ |
13387 |
+- bfqq = bfqq_close(bfqd); |
13388 |
++ bfqq = bfqq_close(bfqd, sector); |
13389 |
+ if (bfqq == NULL || bfqq == cur_bfqq) |
13390 |
+ return NULL; |
13391 |
+ |
13392 |
+@@ -930,6 +922,282 @@ static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
13393 |
+ return bfqq; |
13394 |
+ } |
13395 |
+ |
13396 |
++static struct bfq_queue * |
13397 |
++bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
13398 |
++{ |
13399 |
++ int process_refs, new_process_refs; |
13400 |
++ struct bfq_queue *__bfqq; |
13401 |
++ |
13402 |
++ /* |
13403 |
++ * If there are no process references on the new_bfqq, then it is |
13404 |
++ * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
13405 |
++ * may have dropped their last reference (not just their last process |
13406 |
++ * reference). |
13407 |
++ */ |
13408 |
++ if (!bfqq_process_refs(new_bfqq)) |
13409 |
++ return NULL; |
13410 |
++ |
13411 |
++ /* Avoid a circular list and skip interim queue merges. */ |
13412 |
++ while ((__bfqq = new_bfqq->new_bfqq)) { |
13413 |
++ if (__bfqq == bfqq) |
13414 |
++ return NULL; |
13415 |
++ new_bfqq = __bfqq; |
13416 |
++ } |
13417 |
++ |
13418 |
++ process_refs = bfqq_process_refs(bfqq); |
13419 |
++ new_process_refs = bfqq_process_refs(new_bfqq); |
13420 |
++ /* |
13421 |
++ * If the process for the bfqq has gone away, there is no |
13422 |
++ * sense in merging the queues. |
13423 |
++ */ |
13424 |
++ if (process_refs == 0 || new_process_refs == 0) |
13425 |
++ return NULL; |
13426 |
++ |
13427 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
13428 |
++ new_bfqq->pid); |
13429 |
++ |
13430 |
++ /* |
13431 |
++ * Merging is just a redirection: the requests of the process owning |
13432 |
++ * one of the two queues are redirected to the other queue. The latter |
13433 |
++ * queue, in its turn, is set as shared if this is the first time that |
13434 |
++ * the requests of some process are redirected to it. |
13435 |
++ * |
13436 |
++ * We redirect bfqq to new_bfqq and not the opposite, because we |
13437 |
++ * are in the context of the process owning bfqq, hence we have the |
13438 |
++ * io_cq of this process. So we can immediately configure this io_cq |
13439 |
++ * to redirect the requests of the process to new_bfqq. |
13440 |
++ * |
13441 |
++ * NOTE, even if new_bfqq coincides with the in-service queue, the |
13442 |
++ * io_cq of new_bfqq is not available, because, if the in-service queue |
13443 |
++ * is shared, bfqd->in_service_bic may not point to the io_cq of the |
13444 |
++ * in-service queue. |
13445 |
++ * Redirecting the requests of the process owning bfqq to the currently |
13446 |
++ * in-service queue is in any case the best option, as we feed the |
13447 |
++ * in-service queue with new requests close to the last request served |
13448 |
++ * and, by doing so, hopefully increase the throughput. |
13449 |
++ */ |
13450 |
++ bfqq->new_bfqq = new_bfqq; |
13451 |
++ atomic_add(process_refs, &new_bfqq->ref); |
13452 |
++ return new_bfqq; |
13453 |
++} |
13454 |
++ |
13455 |
++/* |
13456 |
++ * Attempt to schedule a merge of bfqq with the currently in-service queue or |
13457 |
++ * with a close queue among the scheduled queues. |
13458 |
++ * Return NULL if no merge was scheduled, a pointer to the shared bfq_queue |
13459 |
++ * structure otherwise. |
13460 |
++ */ |
13461 |
++static struct bfq_queue * |
13462 |
++bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
13463 |
++ void *io_struct, bool request) |
13464 |
++{ |
13465 |
++ struct bfq_queue *in_service_bfqq, *new_bfqq; |
13466 |
++ |
13467 |
++ if (bfqq->new_bfqq) |
13468 |
++ return bfqq->new_bfqq; |
13469 |
++ |
13470 |
++ if (!io_struct) |
13471 |
++ return NULL; |
13472 |
++ |
13473 |
++ in_service_bfqq = bfqd->in_service_queue; |
13474 |
++ |
13475 |
++ if (in_service_bfqq == NULL || in_service_bfqq == bfqq || |
13476 |
++ !bfqd->in_service_bic) |
13477 |
++ goto check_scheduled; |
13478 |
++ |
13479 |
++ if (bfq_class_idle(in_service_bfqq) || bfq_class_idle(bfqq)) |
13480 |
++ goto check_scheduled; |
13481 |
++ |
13482 |
++ if (bfq_class_rt(in_service_bfqq) != bfq_class_rt(bfqq)) |
13483 |
++ goto check_scheduled; |
13484 |
++ |
13485 |
++ if (in_service_bfqq->entity.parent != bfqq->entity.parent) |
13486 |
++ goto check_scheduled; |
13487 |
++ |
13488 |
++ if (bfq_rq_close_to_sector(io_struct, request, bfqd->last_position) && |
13489 |
++ bfq_bfqq_sync(in_service_bfqq) && bfq_bfqq_sync(bfqq)) { |
13490 |
++ new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq); |
13491 |
++ if (new_bfqq != NULL) |
13492 |
++ return new_bfqq; /* Merge with the in-service queue */ |
13493 |
++ } |
13494 |
++ |
13495 |
++ /* |
13496 |
++ * Check whether there is a cooperator among currently scheduled |
13497 |
++ * queues. The only thing we need is that the bio/request is not |
13498 |
++ * NULL, as we need it to establish whether a cooperator exists. |
13499 |
++ */ |
13500 |
++check_scheduled: |
13501 |
++ new_bfqq = bfq_close_cooperator(bfqd, bfqq, |
13502 |
++ bfq_io_struct_pos(io_struct, request)); |
13503 |
++ if (new_bfqq) |
13504 |
++ return bfq_setup_merge(bfqq, new_bfqq); |
13505 |
++ |
13506 |
++ return NULL; |
13507 |
++} |
13508 |
++ |
13509 |
++static inline void |
13510 |
++bfq_bfqq_save_state(struct bfq_queue *bfqq) |
13511 |
++{ |
13512 |
++ /* |
13513 |
++ * If bfqq->bic == NULL, the queue is already shared or its requests |
13514 |
++ * have already been redirected to a shared queue; both idle window |
13515 |
++ * and weight raising state have already been saved. Do nothing. |
13516 |
++ */ |
13517 |
++ if (bfqq->bic == NULL) |
13518 |
++ return; |
13519 |
++ if (bfqq->bic->raising_time_left) |
13520 |
++ /* |
13521 |
++ * This is the queue of a just-started process, and would |
13522 |
++ * deserve weight raising: we set raising_time_left to the full |
13523 |
++ * weight-raising duration to trigger weight-raising when and |
13524 |
++ * if the queue is split and the first request of the queue |
13525 |
++ * is enqueued. |
13526 |
++ */ |
13527 |
++ bfqq->bic->raising_time_left = bfq_wrais_duration(bfqq->bfqd); |
13528 |
++ else if (bfqq->raising_coeff > 1) { |
13529 |
++ unsigned long wrais_duration = |
13530 |
++ jiffies - bfqq->last_rais_start_finish; |
13531 |
++ /* |
13532 |
++ * It may happen that a queue's weight raising period lasts |
13533 |
++ * longer than its raising_cur_max_time, as weight raising is |
13534 |
++ * handled only when a request is enqueued or dispatched (it |
13535 |
++ * does not use any timer). If the weight raising period is |
13536 |
++ * about to end, don't save it. |
13537 |
++ */ |
13538 |
++ if (bfqq->raising_cur_max_time <= wrais_duration) |
13539 |
++ bfqq->bic->raising_time_left = 0; |
13540 |
++ else |
13541 |
++ bfqq->bic->raising_time_left = |
13542 |
++ bfqq->raising_cur_max_time - wrais_duration; |
13543 |
++ /* |
13544 |
++ * The bfq_queue is becoming shared or the requests of the |
13545 |
++ * process owning the queue are being redirected to a shared |
13546 |
++ * queue. Stop the weight raising period of the queue, as in |
13547 |
++ * both cases it should not be owned by an interactive or soft |
13548 |
++ * real-time application. |
13549 |
++ */ |
13550 |
++ bfq_bfqq_end_raising(bfqq); |
13551 |
++ } else |
13552 |
++ bfqq->bic->raising_time_left = 0; |
13553 |
++ bfqq->bic->saved_idle_window = bfq_bfqq_idle_window(bfqq); |
13554 |
++} |
13555 |
++ |
13556 |
++static inline void |
13557 |
++bfq_get_bic_reference(struct bfq_queue *bfqq) |
13558 |
++{ |
13559 |
++ /* |
13560 |
++ * If bfqq->bic has a non-NULL value, the bic to which it belongs |
13561 |
++ * is about to begin using a shared bfq_queue. |
13562 |
++ */ |
13563 |
++ if (bfqq->bic) |
13564 |
++ atomic_long_inc(&bfqq->bic->icq.ioc->refcount); |
13565 |
++} |
13566 |
++ |
13567 |
++static void |
13568 |
++bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
13569 |
++ struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
13570 |
++{ |
13571 |
++ bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
13572 |
++ (long unsigned)new_bfqq->pid); |
13573 |
++ /* Save weight raising and idle window of the merged queues */ |
13574 |
++ bfq_bfqq_save_state(bfqq); |
13575 |
++ bfq_bfqq_save_state(new_bfqq); |
13576 |
++ /* |
13577 |
++ * Grab a reference to the bic, to prevent it from being destroyed |
13578 |
++ * before being possibly touched by a bfq_split_bfqq(). |
13579 |
++ */ |
13580 |
++ bfq_get_bic_reference(bfqq); |
13581 |
++ bfq_get_bic_reference(new_bfqq); |
13582 |
++ /* Merge queues (that is, let bic redirect its requests to new_bfqq) */ |
13583 |
++ bic_set_bfqq(bic, new_bfqq, 1); |
13584 |
++ bfq_mark_bfqq_coop(new_bfqq); |
13585 |
++ /* |
13586 |
++ * new_bfqq now belongs to at least two bics (it is a shared queue): set |
13587 |
++ * new_bfqq->bic to NULL. bfqq either: |
13588 |
++ * - does not belong to any bic any more, and hence bfqq->bic must |
13589 |
++ * be set to NULL, or |
13590 |
++ * - is a queue whose owning bics have already been redirected to a |
13591 |
++ * different queue, hence the queue is destined to not belong to any |
13592 |
++ * bic soon and bfqq->bic is already NULL (therefore the next |
13593 |
++ * assignment causes no harm). |
13594 |
++ */ |
13595 |
++ new_bfqq->bic = NULL; |
13596 |
++ bfqq->bic = NULL; |
13597 |
++ bfq_put_queue(bfqq); |
13598 |
++} |
13599 |
++ |
13600 |
++static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
13601 |
++ struct bio *bio) |
13602 |
++{ |
13603 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
13604 |
++ struct bfq_io_cq *bic; |
13605 |
++ struct bfq_queue *bfqq, *new_bfqq; |
13606 |
++ |
13607 |
++ /* |
13608 |
++ * Disallow merge of a sync bio into an async request. |
13609 |
++ */ |
13610 |
++ if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
13611 |
++ return 0; |
13612 |
++ |
13613 |
++ /* |
13614 |
++ * Lookup the bfqq that this bio will be queued with. Allow |
13615 |
++ * merge only if rq is queued there. |
13616 |
++ * Queue lock is held here. |
13617 |
++ */ |
13618 |
++ bic = bfq_bic_lookup(bfqd, current->io_context); |
13619 |
++ if (bic == NULL) |
13620 |
++ return 0; |
13621 |
++ |
13622 |
++ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
13623 |
++ /* |
13624 |
++ * We take advantage of this function to perform an early merge |
13625 |
++ * of the queues of possible cooperating processes. |
13626 |
++ */ |
13627 |
++ if (bfqq != NULL) { |
13628 |
++ new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false); |
13629 |
++ if (new_bfqq != NULL) { |
13630 |
++ bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq); |
13631 |
++ /* |
13632 |
++ * If we get here, the bio will be queued in the shared queue, |
13633 |
++ * i.e., new_bfqq, so use new_bfqq to decide whether bio and |
13634 |
++ * rq can be merged. |
13635 |
++ */ |
13636 |
++ bfqq = new_bfqq; |
13637 |
++ } |
13638 |
++ } |
13639 |
++ |
13640 |
++ return bfqq == RQ_BFQQ(rq); |
13641 |
++} |
13642 |
++ |
13643 |
++static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
13644 |
++ struct bfq_queue *bfqq) |
13645 |
++{ |
13646 |
++ if (bfqq != NULL) { |
13647 |
++ bfq_mark_bfqq_must_alloc(bfqq); |
13648 |
++ bfq_mark_bfqq_budget_new(bfqq); |
13649 |
++ bfq_clear_bfqq_fifo_expire(bfqq); |
13650 |
++ |
13651 |
++ bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
13652 |
++ |
13653 |
++ bfq_log_bfqq(bfqd, bfqq, |
13654 |
++ "set_in_service_queue, cur-budget = %lu", |
13655 |
++ bfqq->entity.budget); |
13656 |
++ } |
13657 |
++ |
13658 |
++ bfqd->in_service_queue = bfqq; |
13659 |
++} |
13660 |
++ |
13661 |
++/* |
13662 |
++ * Get and set a new queue for service. |
13663 |
++ */ |
13664 |
++static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd) |
13665 |
++{ |
13666 |
++ struct bfq_queue *bfqq = bfq_get_next_queue(bfqd); |
13667 |
++ |
13668 |
++ __bfq_set_in_service_queue(bfqd, bfqq); |
13669 |
++ return bfqq; |
13670 |
++} |
13671 |
++ |
13672 |
+ /* |
13673 |
+ * If enough samples have been computed, return the current max budget |
13674 |
+ * stored in bfqd, which is dynamically updated according to the |
13675 |
+@@ -1077,63 +1345,6 @@ static struct request *bfq_check_fifo(struct bfq_queue *bfqq) |
13676 |
+ return rq; |
13677 |
+ } |
13678 |
+ |
13679 |
+-/* |
13680 |
+- * Must be called with the queue_lock held. |
13681 |
+- */ |
13682 |
+-static int bfqq_process_refs(struct bfq_queue *bfqq) |
13683 |
+-{ |
13684 |
+- int process_refs, io_refs; |
13685 |
+- |
13686 |
+- io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
13687 |
+- process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
13688 |
+- BUG_ON(process_refs < 0); |
13689 |
+- return process_refs; |
13690 |
+-} |
13691 |
+- |
13692 |
+-static void bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
13693 |
+-{ |
13694 |
+- int process_refs, new_process_refs; |
13695 |
+- struct bfq_queue *__bfqq; |
13696 |
+- |
13697 |
+- /* |
13698 |
+- * If there are no process references on the new_bfqq, then it is |
13699 |
+- * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
13700 |
+- * may have dropped their last reference (not just their last process |
13701 |
+- * reference). |
13702 |
+- */ |
13703 |
+- if (!bfqq_process_refs(new_bfqq)) |
13704 |
+- return; |
13705 |
+- |
13706 |
+- /* Avoid a circular list and skip interim queue merges. */ |
13707 |
+- while ((__bfqq = new_bfqq->new_bfqq)) { |
13708 |
+- if (__bfqq == bfqq) |
13709 |
+- return; |
13710 |
+- new_bfqq = __bfqq; |
13711 |
+- } |
13712 |
+- |
13713 |
+- process_refs = bfqq_process_refs(bfqq); |
13714 |
+- new_process_refs = bfqq_process_refs(new_bfqq); |
13715 |
+- /* |
13716 |
+- * If the process for the bfqq has gone away, there is no |
13717 |
+- * sense in merging the queues. |
13718 |
+- */ |
13719 |
+- if (process_refs == 0 || new_process_refs == 0) |
13720 |
+- return; |
13721 |
+- |
13722 |
+- /* |
13723 |
+- * Merge in the direction of the lesser amount of work. |
13724 |
+- */ |
13725 |
+- if (new_process_refs >= process_refs) { |
13726 |
+- bfqq->new_bfqq = new_bfqq; |
13727 |
+- atomic_add(process_refs, &new_bfqq->ref); |
13728 |
+- } else { |
13729 |
+- new_bfqq->new_bfqq = bfqq; |
13730 |
+- atomic_add(new_process_refs, &bfqq->ref); |
13731 |
+- } |
13732 |
+- bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
13733 |
+- new_bfqq->pid); |
13734 |
+-} |
13735 |
+- |
13736 |
+ static inline unsigned long bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
13737 |
+ { |
13738 |
+ struct bfq_entity *entity = &bfqq->entity; |
13739 |
+@@ -1703,7 +1914,7 @@ static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) |
13740 |
+ */ |
13741 |
+ static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
13742 |
+ { |
13743 |
+- struct bfq_queue *bfqq, *new_bfqq = NULL; |
13744 |
++ struct bfq_queue *bfqq; |
13745 |
+ struct request *next_rq; |
13746 |
+ enum bfqq_expiration reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
13747 |
+ |
13748 |
+@@ -1713,17 +1924,6 @@ static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
13749 |
+ |
13750 |
+ bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); |
13751 |
+ |
13752 |
+- /* |
13753 |
+- * If another queue has a request waiting within our mean seek |
13754 |
+- * distance, let it run. The expire code will check for close |
13755 |
+- * cooperators and put the close queue at the front of the |
13756 |
+- * service tree. If possible, merge the expiring queue with the |
13757 |
+- * new bfqq. |
13758 |
+- */ |
13759 |
+- new_bfqq = bfq_close_cooperator(bfqd, bfqq); |
13760 |
+- if (new_bfqq != NULL && bfqq->new_bfqq == NULL) |
13761 |
+- bfq_setup_merge(bfqq, new_bfqq); |
13762 |
+- |
13763 |
+ if (bfq_may_expire_for_budg_timeout(bfqq) && |
13764 |
+ !timer_pending(&bfqd->idle_slice_timer) && |
13765 |
+ !bfq_bfqq_must_idle(bfqq)) |
13766 |
+@@ -1760,36 +1960,26 @@ static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
13767 |
+ bfq_clear_bfqq_wait_request(bfqq); |
13768 |
+ del_timer(&bfqd->idle_slice_timer); |
13769 |
+ } |
13770 |
+- if (new_bfqq == NULL) |
13771 |
+- goto keep_queue; |
13772 |
+- else |
13773 |
+- goto expire; |
13774 |
++ goto keep_queue; |
13775 |
+ } |
13776 |
+ } |
13777 |
+ |
13778 |
+ /* |
13779 |
+- * No requests pending. If the in-service queue has no cooperator and |
13780 |
+- * still has requests in flight (possibly waiting for a completion) |
13781 |
+- * or is idling for a new request, then keep it. |
13782 |
++ * No requests pending. If the in-service queue still has requests in |
13783 |
++ * flight (possibly waiting for a completion) or is idling for a new |
13784 |
++ * request, then keep it. |
13785 |
+ */ |
13786 |
+- if (new_bfqq == NULL && (timer_pending(&bfqd->idle_slice_timer) || |
13787 |
+- (bfqq->dispatched != 0 && bfq_bfqq_must_not_expire(bfqq)))) { |
13788 |
++ if (timer_pending(&bfqd->idle_slice_timer) || |
13789 |
++ (bfqq->dispatched != 0 && bfq_bfqq_must_not_expire(bfqq))) { |
13790 |
+ bfqq = NULL; |
13791 |
+ goto keep_queue; |
13792 |
+- } else if (new_bfqq != NULL && timer_pending(&bfqd->idle_slice_timer)) { |
13793 |
+- /* |
13794 |
+- * Expiring the queue because there is a close cooperator, |
13795 |
+- * cancel timer. |
13796 |
+- */ |
13797 |
+- bfq_clear_bfqq_wait_request(bfqq); |
13798 |
+- del_timer(&bfqd->idle_slice_timer); |
13799 |
+ } |
13800 |
+ |
13801 |
+ reason = BFQ_BFQQ_NO_MORE_REQUESTS; |
13802 |
+ expire: |
13803 |
+ bfq_bfqq_expire(bfqd, bfqq, 0, reason); |
13804 |
+ new_queue: |
13805 |
+- bfqq = bfq_set_in_service_queue(bfqd, new_bfqq); |
13806 |
++ bfqq = bfq_set_in_service_queue(bfqd); |
13807 |
+ bfq_log(bfqd, "select_queue: new queue %d returned", |
13808 |
+ bfqq != NULL ? bfqq->pid : 0); |
13809 |
+ keep_queue: |
13810 |
+@@ -1799,9 +1989,8 @@ keep_queue: |
13811 |
+ static void bfq_update_raising_data(struct bfq_data *bfqd, |
13812 |
+ struct bfq_queue *bfqq) |
13813 |
+ { |
13814 |
++ struct bfq_entity *entity = &bfqq->entity; |
13815 |
+ if (bfqq->raising_coeff > 1) { /* queue is being boosted */ |
13816 |
+- struct bfq_entity *entity = &bfqq->entity; |
13817 |
+- |
13818 |
+ bfq_log_bfqq(bfqd, bfqq, |
13819 |
+ "raising period dur %u/%u msec, " |
13820 |
+ "old raising coeff %u, w %d(%d)", |
13821 |
+@@ -1818,7 +2007,7 @@ static void bfq_update_raising_data(struct bfq_data *bfqd, |
13822 |
+ "WARN: pending prio change"); |
13823 |
+ /* |
13824 |
+ * If too much time has elapsed from the beginning |
13825 |
+- * of this weight-raising, stop it. |
13826 |
++ * of this weight-raising period, stop it. |
13827 |
+ */ |
13828 |
+ if (jiffies - bfqq->last_rais_start_finish > |
13829 |
+ bfqq->raising_cur_max_time) { |
13830 |
+@@ -1830,11 +2019,13 @@ static void bfq_update_raising_data(struct bfq_data *bfqd, |
13831 |
+ jiffies_to_msecs(bfqq-> |
13832 |
+ raising_cur_max_time)); |
13833 |
+ bfq_bfqq_end_raising(bfqq); |
13834 |
+- __bfq_entity_update_weight_prio( |
13835 |
+- bfq_entity_service_tree(entity), |
13836 |
+- entity); |
13837 |
+ } |
13838 |
+ } |
13839 |
++ /* Update weight both if it must be raised and if it must be lowered */ |
13840 |
++ if ((entity->weight > entity->orig_weight) != (bfqq->raising_coeff > 1)) |
13841 |
++ __bfq_entity_update_weight_prio( |
13842 |
++ bfq_entity_service_tree(entity), |
13843 |
++ entity); |
13844 |
+ } |
13845 |
+ |
13846 |
+ /* |
13847 |
+@@ -2075,6 +2266,25 @@ static void bfq_init_icq(struct io_cq *icq) |
13848 |
+ struct bfq_io_cq *bic = icq_to_bic(icq); |
13849 |
+ |
13850 |
+ bic->ttime.last_end_request = jiffies; |
13851 |
++ /* |
13852 |
++ * A newly created bic indicates that the process has just |
13853 |
++ * started doing I/O, and is probably mapping into memory its |
13854 |
++ * executable and libraries: it definitely needs weight raising. |
13855 |
++ * There is however the possibility that the process performs, |
13856 |
++ * for a while, I/O close to some other process. EQM intercepts |
13857 |
++ * this behavior and may merge the queue corresponding to the |
13858 |
++ * process with some other queue, BEFORE the weight of the queue |
13859 |
++ * is raised. Merged queues are not weight-raised (they are assumed |
13860 |
++ * to belong to processes that benefit only from high throughput). |
13861 |
++ * If the merge is basically the consequence of an accident, then |
13862 |
++ * the queue will be split soon and will get back its old weight. |
13863 |
++ * It is then important to write down somewhere that this queue |
13864 |
++ * does need weight raising, even if it did not make it to get its |
13865 |
++ * weight raised before being merged. To this purpose, we overload |
13866 |
++ * the field raising_time_left and assign 1 to it, to mark the queue |
13867 |
++ * as needing weight raising. |
13868 |
++ */ |
13869 |
++ bic->raising_time_left = 1; |
13870 |
+ } |
13871 |
+ |
13872 |
+ static void bfq_exit_icq(struct io_cq *icq) |
13873 |
+@@ -2088,6 +2298,13 @@ static void bfq_exit_icq(struct io_cq *icq) |
13874 |
+ } |
13875 |
+ |
13876 |
+ if (bic->bfqq[BLK_RW_SYNC]) { |
13877 |
++ /* |
13878 |
++ * If the bic is using a shared queue, put the reference |
13879 |
++ * taken on the io_context when the bic started using a |
13880 |
++ * shared bfq_queue. |
13881 |
++ */ |
13882 |
++ if (bfq_bfqq_coop(bic->bfqq[BLK_RW_SYNC])) |
13883 |
++ put_io_context(icq->ioc); |
13884 |
+ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_SYNC]); |
13885 |
+ bic->bfqq[BLK_RW_SYNC] = NULL; |
13886 |
+ } |
13887 |
+@@ -2375,6 +2592,10 @@ static void bfq_update_idle_window(struct bfq_data *bfqd, |
13888 |
+ if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq)) |
13889 |
+ return; |
13890 |
+ |
13891 |
++ /* Idle window just restored, statistics are meaningless. */ |
13892 |
++ if (bfq_bfqq_just_split(bfqq)) |
13893 |
++ return; |
13894 |
++ |
13895 |
+ enable_idle = bfq_bfqq_idle_window(bfqq); |
13896 |
+ |
13897 |
+ if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
13898 |
+@@ -2415,6 +2636,7 @@ static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
13899 |
+ if (bfqq->entity.service > bfq_max_budget(bfqd) / 8 || |
13900 |
+ !BFQQ_SEEKY(bfqq)) |
13901 |
+ bfq_update_idle_window(bfqd, bfqq, bic); |
13902 |
++ bfq_clear_bfqq_just_split(bfqq); |
13903 |
+ |
13904 |
+ bfq_log_bfqq(bfqd, bfqq, |
13905 |
+ "rq_enqueued: idle_window=%d (seeky %d, mean %llu)", |
13906 |
+@@ -2475,13 +2697,48 @@ static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
13907 |
+ static void bfq_insert_request(struct request_queue *q, struct request *rq) |
13908 |
+ { |
13909 |
+ struct bfq_data *bfqd = q->elevator->elevator_data; |
13910 |
+- struct bfq_queue *bfqq = RQ_BFQQ(rq); |
13911 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq), *new_bfqq; |
13912 |
+ |
13913 |
+ assert_spin_locked(bfqd->queue->queue_lock); |
13914 |
++ |
13915 |
++ /* |
13916 |
++ * An unplug may trigger a requeue of a request from the device |
13917 |
++ * driver: make sure we are in process context while trying to |
13918 |
++ * merge two bfq_queues. |
13919 |
++ */ |
13920 |
++ if (!in_interrupt()) { |
13921 |
++ new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true); |
13922 |
++ if (new_bfqq != NULL) { |
13923 |
++ if (bic_to_bfqq(RQ_BIC(rq), 1) != bfqq) |
13924 |
++ new_bfqq = bic_to_bfqq(RQ_BIC(rq), 1); |
13925 |
++ /* |
13926 |
++ * Release the request's reference to the old bfqq |
13927 |
++ * and make sure one is taken to the shared queue. |
13928 |
++ */ |
13929 |
++ new_bfqq->allocated[rq_data_dir(rq)]++; |
13930 |
++ bfqq->allocated[rq_data_dir(rq)]--; |
13931 |
++ atomic_inc(&new_bfqq->ref); |
13932 |
++ bfq_put_queue(bfqq); |
13933 |
++ if (bic_to_bfqq(RQ_BIC(rq), 1) == bfqq) |
13934 |
++ bfq_merge_bfqqs(bfqd, RQ_BIC(rq), |
13935 |
++ bfqq, new_bfqq); |
13936 |
++ rq->elv.priv[1] = new_bfqq; |
13937 |
++ bfqq = new_bfqq; |
13938 |
++ } |
13939 |
++ } |
13940 |
++ |
13941 |
+ bfq_init_prio_data(bfqq, RQ_BIC(rq)); |
13942 |
+ |
13943 |
+ bfq_add_rq_rb(rq); |
13944 |
+ |
13945 |
++ /* |
13946 |
++ * Here a newly-created bfq_queue has already started a weight-raising |
13947 |
++ * period: clear raising_time_left to prevent bfq_bfqq_save_state() |
13948 |
++ * from assigning it a full weight-raising period. See the detailed |
13949 |
++ * comments about this field in bfq_init_icq(). |
13950 |
++ */ |
13951 |
++ if (bfqq->bic != NULL) |
13952 |
++ bfqq->bic->raising_time_left = 0; |
13953 |
+ rq_set_fifo_time(rq, jiffies + bfqd->bfq_fifo_expire[rq_is_sync(rq)]); |
13954 |
+ list_add_tail(&rq->queuelist, &bfqq->fifo); |
13955 |
+ |
13956 |
+@@ -2629,18 +2886,6 @@ static void bfq_put_request(struct request *rq) |
13957 |
+ } |
13958 |
+ } |
13959 |
+ |
13960 |
+-static struct bfq_queue * |
13961 |
+-bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
13962 |
+- struct bfq_queue *bfqq) |
13963 |
+-{ |
13964 |
+- bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
13965 |
+- (long unsigned)bfqq->new_bfqq->pid); |
13966 |
+- bic_set_bfqq(bic, bfqq->new_bfqq, 1); |
13967 |
+- bfq_mark_bfqq_coop(bfqq->new_bfqq); |
13968 |
+- bfq_put_queue(bfqq); |
13969 |
+- return bic_to_bfqq(bic, 1); |
13970 |
+-} |
13971 |
+- |
13972 |
+ /* |
13973 |
+ * Returns NULL if a new bfqq should be allocated, or the old bfqq if this |
13974 |
+ * was the last process referring to said bfqq. |
13975 |
+@@ -2649,6 +2894,9 @@ static struct bfq_queue * |
13976 |
+ bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) |
13977 |
+ { |
13978 |
+ bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); |
13979 |
++ |
13980 |
++ put_io_context(bic->icq.ioc); |
13981 |
++ |
13982 |
+ if (bfqq_process_refs(bfqq) == 1) { |
13983 |
+ bfqq->pid = current->pid; |
13984 |
+ bfq_clear_bfqq_coop(bfqq); |
13985 |
+@@ -2677,6 +2925,7 @@ static int bfq_set_request(struct request_queue *q, struct request *rq, |
13986 |
+ struct bfq_queue *bfqq; |
13987 |
+ struct bfq_group *bfqg; |
13988 |
+ unsigned long flags; |
13989 |
++ bool split = false; |
13990 |
+ |
13991 |
+ might_sleep_if(gfp_mask & __GFP_WAIT); |
13992 |
+ |
13993 |
+@@ -2695,24 +2944,14 @@ new_queue: |
13994 |
+ bfqq = bfq_get_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
13995 |
+ bic_set_bfqq(bic, bfqq, is_sync); |
13996 |
+ } else { |
13997 |
+- /* |
13998 |
+- * If the queue was seeky for too long, break it apart. |
13999 |
+- */ |
14000 |
++ /* If the queue was seeky for too long, break it apart. */ |
14001 |
+ if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { |
14002 |
+ bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); |
14003 |
+ bfqq = bfq_split_bfqq(bic, bfqq); |
14004 |
++ split = true; |
14005 |
+ if (!bfqq) |
14006 |
+ goto new_queue; |
14007 |
+ } |
14008 |
+- |
14009 |
+- /* |
14010 |
+- * Check to see if this queue is scheduled to merge with |
14011 |
+- * another closely cooperating queue. The merging of queues |
14012 |
+- * happens here as it must be done in process context. |
14013 |
+- * The reference on new_bfqq was taken in merge_bfqqs. |
14014 |
+- */ |
14015 |
+- if (bfqq->new_bfqq != NULL) |
14016 |
+- bfqq = bfq_merge_bfqqs(bfqd, bic, bfqq); |
14017 |
+ } |
14018 |
+ |
14019 |
+ bfqq->allocated[rw]++; |
14020 |
+@@ -2723,6 +2962,26 @@ new_queue: |
14021 |
+ rq->elv.priv[0] = bic; |
14022 |
+ rq->elv.priv[1] = bfqq; |
14023 |
+ |
14024 |
++ /* |
14025 |
++ * If a bfq_queue has only one process reference, it is owned |
14026 |
++ * by only one bfq_io_cq: we can set the bic field of the |
14027 |
++ * bfq_queue to the address of that structure. Also, if the |
14028 |
++ * queue has just been split, mark a flag so that the |
14029 |
++ * information is available to the other scheduler hooks. |
14030 |
++ */ |
14031 |
++ if (bfqq_process_refs(bfqq) == 1) { |
14032 |
++ bfqq->bic = bic; |
14033 |
++ if (split) { |
14034 |
++ bfq_mark_bfqq_just_split(bfqq); |
14035 |
++ /* |
14036 |
++ * If the queue has just been split from a shared queue, |
14037 |
++ * restore the idle window and the possible weight |
14038 |
++ * raising period. |
14039 |
++ */ |
14040 |
++ bfq_bfqq_resume_state(bfqq, bic); |
14041 |
++ } |
14042 |
++ } |
14043 |
++ |
14044 |
+ spin_unlock_irqrestore(q->queue_lock, flags); |
14045 |
+ |
14046 |
+ return 0; |
14047 |
+diff --git a/block/bfq-sched.c b/block/bfq-sched.c |
14048 |
+index 30df81c..47e66a8 100644 |
14049 |
+--- a/block/bfq-sched.c |
14050 |
++++ b/block/bfq-sched.c |
14051 |
+@@ -979,34 +979,6 @@ static struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd) |
14052 |
+ return bfqq; |
14053 |
+ } |
14054 |
+ |
14055 |
+-/* |
14056 |
+- * Forced extraction of the given queue. |
14057 |
+- */ |
14058 |
+-static void bfq_get_next_queue_forced(struct bfq_data *bfqd, |
14059 |
+- struct bfq_queue *bfqq) |
14060 |
+-{ |
14061 |
+- struct bfq_entity *entity; |
14062 |
+- struct bfq_sched_data *sd; |
14063 |
+- |
14064 |
+- BUG_ON(bfqd->in_service_queue != NULL); |
14065 |
+- |
14066 |
+- entity = &bfqq->entity; |
14067 |
+- /* |
14068 |
+- * Bubble up extraction/update from the leaf to the root. |
14069 |
+- */ |
14070 |
+- for_each_entity(entity) { |
14071 |
+- sd = entity->sched_data; |
14072 |
+- bfq_update_budget(entity); |
14073 |
+- bfq_update_vtime(bfq_entity_service_tree(entity)); |
14074 |
+- bfq_active_extract(bfq_entity_service_tree(entity), entity); |
14075 |
+- sd->active_entity = entity; |
14076 |
+- sd->next_active = NULL; |
14077 |
+- entity->service = 0; |
14078 |
+- } |
14079 |
+- |
14080 |
+- return; |
14081 |
+-} |
14082 |
+- |
14083 |
+ static void __bfq_bfqd_reset_in_service(struct bfq_data *bfqd) |
14084 |
+ { |
14085 |
+ if (bfqd->in_service_bic != NULL) { |
14086 |
+diff --git a/block/bfq.h b/block/bfq.h |
14087 |
+index 68b28e3..438f560 100644 |
14088 |
+--- a/block/bfq.h |
14089 |
++++ b/block/bfq.h |
14090 |
+@@ -192,6 +192,8 @@ struct bfq_group; |
14091 |
+ * idle to backlogged |
14092 |
+ * @service_from_backlogged: cumulative service received from the @bfq_queue |
14093 |
+ * since the last transition from idle to backlogged |
14094 |
++ * @bic: pointer to the bfq_io_cq owning the bfq_queue, set to %NULL if the |
14095 |
++ * queue is shared |
14096 |
+ * |
14097 |
+ * A bfq_queue is a leaf request queue; it can be associated to an io_context |
14098 |
+ * or more (if it is an async one). @cgroup holds a reference to the |
14099 |
+@@ -235,6 +237,7 @@ struct bfq_queue { |
14100 |
+ sector_t last_request_pos; |
14101 |
+ |
14102 |
+ pid_t pid; |
14103 |
++ struct bfq_io_cq *bic; |
14104 |
+ |
14105 |
+ /* weight-raising fields */ |
14106 |
+ unsigned int raising_cur_max_time; |
14107 |
+@@ -264,12 +267,23 @@ struct bfq_ttime { |
14108 |
+ * @icq: associated io_cq structure |
14109 |
+ * @bfqq: array of two process queues, the sync and the async |
14110 |
+ * @ttime: associated @bfq_ttime struct |
14111 |
++ * @raising_time_left: snapshot of the time left before weight raising ends |
14112 |
++ * for the sync queue associated to this process; this |
14113 |
++ * snapshot is taken to remember this value while the weight |
14114 |
++ * raising is suspended because the queue is merged with a |
14115 |
++ * shared queue, and is used to set @raising_cur_max_time |
14116 |
++ * when the queue is split from the shared queue and its |
14117 |
++ * weight is raised again |
14118 |
++ * @saved_idle_window: same purpose as the previous field for the idle window |
14119 |
+ */ |
14120 |
+ struct bfq_io_cq { |
14121 |
+ struct io_cq icq; /* must be the first member */ |
14122 |
+ struct bfq_queue *bfqq[2]; |
14123 |
+ struct bfq_ttime ttime; |
14124 |
+ int ioprio; |
14125 |
++ |
14126 |
++ unsigned int raising_time_left; |
14127 |
++ unsigned int saved_idle_window; |
14128 |
+ }; |
14129 |
+ |
14130 |
+ /** |
14131 |
+@@ -411,6 +425,7 @@ enum bfqq_state_flags { |
14132 |
+ BFQ_BFQQ_FLAG_budget_new, /* no completion with this budget */ |
14133 |
+ BFQ_BFQQ_FLAG_coop, /* bfqq is shared */ |
14134 |
+ BFQ_BFQQ_FLAG_split_coop, /* shared bfqq will be splitted */ |
14135 |
++ BFQ_BFQQ_FLAG_just_split, /* queue has just been split */ |
14136 |
+ BFQ_BFQQ_FLAG_softrt_update, /* needs softrt-next-start update */ |
14137 |
+ }; |
14138 |
+ |
14139 |
+@@ -438,6 +453,7 @@ BFQ_BFQQ_FNS(sync); |
14140 |
+ BFQ_BFQQ_FNS(budget_new); |
14141 |
+ BFQ_BFQQ_FNS(coop); |
14142 |
+ BFQ_BFQQ_FNS(split_coop); |
14143 |
++BFQ_BFQQ_FNS(just_split); |
14144 |
+ BFQ_BFQQ_FNS(softrt_update); |
14145 |
+ #undef BFQ_BFQQ_FNS |
14146 |
+ |
14147 |
+-- |
14148 |
+1.8.5.2 |
14149 |
+ |
14150 |
|
14151 |
Deleted: genpatches-2.6/trunk/3.14/5000_BFQ-4-block-Switch-from-BFQ-v6r2-for-3.11.0-to-BFQ-v6r2-fo.patch |
14152 |
=================================================================== |
14153 |
--- genpatches-2.6/trunk/3.14/5000_BFQ-4-block-Switch-from-BFQ-v6r2-for-3.11.0-to-BFQ-v6r2-fo.patch 2014-01-29 14:41:45 UTC (rev 2660) |
14154 |
+++ genpatches-2.6/trunk/3.14/5000_BFQ-4-block-Switch-from-BFQ-v6r2-for-3.11.0-to-BFQ-v6r2-fo.patch 2014-01-30 16:49:47 UTC (rev 2661) |
14155 |
@@ -1,362 +0,0 @@ |
14156 |
-From 2e1646d06515b7dd1344db547dfcf9a4640dee8e Mon Sep 17 00:00:00 2001 |
14157 |
-From: Arianna Avanzini <avanzini.arianna@×××××.com> |
14158 |
-Date: Wed, 11 Sep 2013 22:26:47 +0200 |
14159 |
-Subject: [PATCH] block: Switch from BFQ-v6r2 for 3.11.0 to BFQ-v6r2 for |
14160 |
- 3.12.0-rc1 |
14161 |
- |
14162 |
---- |
14163 |
- block/bfq-cgroup.c | 115 +++++++++++++++++++++++++++++++---------------------- |
14164 |
- block/bfq.h | 2 + |
14165 |
- 2 files changed, 70 insertions(+), 47 deletions(-) |
14166 |
- |
14167 |
-diff --git a/block/bfq-cgroup.c b/block/bfq-cgroup.c |
14168 |
-index bb9b851..afae4ca 100644 |
14169 |
---- a/block/bfq-cgroup.c |
14170 |
-+++ b/block/bfq-cgroup.c |
14171 |
-@@ -16,9 +16,9 @@ |
14172 |
- |
14173 |
- static DEFINE_MUTEX(bfqio_mutex); |
14174 |
- |
14175 |
--static bool bfqio_is_removed(struct cgroup *cgroup) |
14176 |
-+static bool bfqio_is_removed(struct bfqio_cgroup *bgrp) |
14177 |
- { |
14178 |
-- return test_bit(CGRP_DEAD, &cgroup->flags); |
14179 |
-+ return bgrp ? !bgrp->online : false; |
14180 |
- } |
14181 |
- |
14182 |
- static struct bfqio_cgroup bfqio_root_cgroup = { |
14183 |
-@@ -38,10 +38,9 @@ static inline void bfq_init_entity(struct bfq_entity *entity, |
14184 |
- entity->sched_data = &bfqg->sched_data; |
14185 |
- } |
14186 |
- |
14187 |
--static struct bfqio_cgroup *cgroup_to_bfqio(struct cgroup *cgroup) |
14188 |
-+static struct bfqio_cgroup *css_to_bfqio(struct cgroup_subsys_state *css) |
14189 |
- { |
14190 |
-- return container_of(cgroup_subsys_state(cgroup, bfqio_subsys_id), |
14191 |
-- struct bfqio_cgroup, css); |
14192 |
-+ return css ? container_of(css, struct bfqio_cgroup, css) : NULL; |
14193 |
- } |
14194 |
- |
14195 |
- /* |
14196 |
-@@ -103,20 +102,20 @@ static inline void bfq_group_set_parent(struct bfq_group *bfqg, |
14197 |
- /** |
14198 |
- * bfq_group_chain_alloc - allocate a chain of groups. |
14199 |
- * @bfqd: queue descriptor. |
14200 |
-- * @cgroup: the leaf cgroup this chain starts from. |
14201 |
-+ * @css: the leaf cgroup_subsys_state this chain starts from. |
14202 |
- * |
14203 |
- * Allocate a chain of groups starting from the one belonging to |
14204 |
- * @cgroup up to the root cgroup. Stop if a cgroup on the chain |
14205 |
- * to the root has already an allocated group on @bfqd. |
14206 |
- */ |
14207 |
- static struct bfq_group *bfq_group_chain_alloc(struct bfq_data *bfqd, |
14208 |
-- struct cgroup *cgroup) |
14209 |
-+ struct cgroup_subsys_state *css) |
14210 |
- { |
14211 |
- struct bfqio_cgroup *bgrp; |
14212 |
- struct bfq_group *bfqg, *prev = NULL, *leaf = NULL; |
14213 |
- |
14214 |
-- for (; cgroup != NULL; cgroup = cgroup->parent) { |
14215 |
-- bgrp = cgroup_to_bfqio(cgroup); |
14216 |
-+ for (; css != NULL; css = css->parent) { |
14217 |
-+ bgrp = css_to_bfqio(css); |
14218 |
- |
14219 |
- bfqg = bfqio_lookup_group(bgrp, bfqd); |
14220 |
- if (bfqg != NULL) { |
14221 |
-@@ -165,7 +164,7 @@ cleanup: |
14222 |
- /** |
14223 |
- * bfq_group_chain_link - link an allocatd group chain to a cgroup hierarchy. |
14224 |
- * @bfqd: the queue descriptor. |
14225 |
-- * @cgroup: the leaf cgroup to start from. |
14226 |
-+ * @css: the leaf cgroup_subsys_state to start from. |
14227 |
- * @leaf: the leaf group (to be associated to @cgroup). |
14228 |
- * |
14229 |
- * Try to link a chain of groups to a cgroup hierarchy, connecting the |
14230 |
-@@ -177,7 +176,8 @@ cleanup: |
14231 |
- * per device) while the bfqio_cgroup lock protects the list of groups |
14232 |
- * belonging to the same cgroup. |
14233 |
- */ |
14234 |
--static void bfq_group_chain_link(struct bfq_data *bfqd, struct cgroup *cgroup, |
14235 |
-+static void bfq_group_chain_link(struct bfq_data *bfqd, |
14236 |
-+ struct cgroup_subsys_state *css, |
14237 |
- struct bfq_group *leaf) |
14238 |
- { |
14239 |
- struct bfqio_cgroup *bgrp; |
14240 |
-@@ -186,8 +186,8 @@ static void bfq_group_chain_link(struct bfq_data *bfqd, struct cgroup *cgroup, |
14241 |
- |
14242 |
- assert_spin_locked(bfqd->queue->queue_lock); |
14243 |
- |
14244 |
-- for (; cgroup != NULL && leaf != NULL; cgroup = cgroup->parent) { |
14245 |
-- bgrp = cgroup_to_bfqio(cgroup); |
14246 |
-+ for (; css != NULL && leaf != NULL; css = css->parent) { |
14247 |
-+ bgrp = css_to_bfqio(css); |
14248 |
- next = leaf->bfqd; |
14249 |
- |
14250 |
- bfqg = bfqio_lookup_group(bgrp, bfqd); |
14251 |
-@@ -205,9 +205,9 @@ static void bfq_group_chain_link(struct bfq_data *bfqd, struct cgroup *cgroup, |
14252 |
- leaf = next; |
14253 |
- } |
14254 |
- |
14255 |
-- BUG_ON(cgroup == NULL && leaf != NULL); |
14256 |
-- if (cgroup != NULL && prev != NULL) { |
14257 |
-- bgrp = cgroup_to_bfqio(cgroup); |
14258 |
-+ BUG_ON(css == NULL && leaf != NULL); |
14259 |
-+ if (css != NULL && prev != NULL) { |
14260 |
-+ bgrp = css_to_bfqio(css); |
14261 |
- bfqg = bfqio_lookup_group(bgrp, bfqd); |
14262 |
- bfq_group_set_parent(prev, bfqg); |
14263 |
- } |
14264 |
-@@ -233,18 +233,18 @@ static void bfq_group_chain_link(struct bfq_data *bfqd, struct cgroup *cgroup, |
14265 |
- * have been successful. |
14266 |
- */ |
14267 |
- static struct bfq_group *bfq_find_alloc_group(struct bfq_data *bfqd, |
14268 |
-- struct cgroup *cgroup) |
14269 |
-+ struct cgroup_subsys_state *css) |
14270 |
- { |
14271 |
-- struct bfqio_cgroup *bgrp = cgroup_to_bfqio(cgroup); |
14272 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
14273 |
- struct bfq_group *bfqg; |
14274 |
- |
14275 |
- bfqg = bfqio_lookup_group(bgrp, bfqd); |
14276 |
- if (bfqg != NULL) |
14277 |
- return bfqg; |
14278 |
- |
14279 |
-- bfqg = bfq_group_chain_alloc(bfqd, cgroup); |
14280 |
-+ bfqg = bfq_group_chain_alloc(bfqd, css); |
14281 |
- if (bfqg != NULL) |
14282 |
-- bfq_group_chain_link(bfqd, cgroup, bfqg); |
14283 |
-+ bfq_group_chain_link(bfqd, css, bfqg); |
14284 |
- else |
14285 |
- bfqg = bfqd->root_group; |
14286 |
- |
14287 |
-@@ -315,8 +315,8 @@ static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
14288 |
- * time here, at the price of slightly more complex code. |
14289 |
- */ |
14290 |
- static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
14291 |
-- struct bfq_io_cq *bic, |
14292 |
-- struct cgroup *cgroup) |
14293 |
-+ struct bfq_io_cq *bic, |
14294 |
-+ struct cgroup_subsys_state *css) |
14295 |
- { |
14296 |
- struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0); |
14297 |
- struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1); |
14298 |
-@@ -324,9 +324,9 @@ static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
14299 |
- struct bfq_group *bfqg; |
14300 |
- struct bfqio_cgroup *bgrp; |
14301 |
- |
14302 |
-- bgrp = cgroup_to_bfqio(cgroup); |
14303 |
-+ bgrp = css_to_bfqio(css); |
14304 |
- |
14305 |
-- bfqg = bfq_find_alloc_group(bfqd, cgroup); |
14306 |
-+ bfqg = bfq_find_alloc_group(bfqd, css); |
14307 |
- if (async_bfqq != NULL) { |
14308 |
- entity = &async_bfqq->entity; |
14309 |
- |
14310 |
-@@ -357,14 +357,14 @@ static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
14311 |
- * moved into its new parent group. |
14312 |
- */ |
14313 |
- static void bfq_bic_change_cgroup(struct bfq_io_cq *bic, |
14314 |
-- struct cgroup *cgroup) |
14315 |
-+ struct cgroup_subsys_state *css) |
14316 |
- { |
14317 |
- struct bfq_data *bfqd; |
14318 |
- unsigned long uninitialized_var(flags); |
14319 |
- |
14320 |
- bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), &flags); |
14321 |
- if (bfqd != NULL) { |
14322 |
-- __bfq_bic_change_cgroup(bfqd, bic, cgroup); |
14323 |
-+ __bfq_bic_change_cgroup(bfqd, bic, css); |
14324 |
- bfq_put_bfqd_unlock(bfqd, &flags); |
14325 |
- } |
14326 |
- } |
14327 |
-@@ -394,13 +394,13 @@ static struct bfq_group *bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
14328 |
- { |
14329 |
- struct bfq_data *bfqd = bic_to_bfqd(bic); |
14330 |
- struct bfq_group *bfqg; |
14331 |
-- struct cgroup *cgroup; |
14332 |
-+ struct cgroup_subsys_state *css; |
14333 |
- |
14334 |
- BUG_ON(bfqd == NULL); |
14335 |
- |
14336 |
- rcu_read_lock(); |
14337 |
-- cgroup = task_cgroup(current, bfqio_subsys_id); |
14338 |
-- bfqg = __bfq_bic_change_cgroup(bfqd, bic, cgroup); |
14339 |
-+ css = task_css(current, bfqio_subsys_id); |
14340 |
-+ bfqg = __bfq_bic_change_cgroup(bfqd, bic, css); |
14341 |
- rcu_read_unlock(); |
14342 |
- |
14343 |
- return bfqg; |
14344 |
-@@ -622,17 +622,16 @@ static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
14345 |
- } |
14346 |
- |
14347 |
- #define SHOW_FUNCTION(__VAR) \ |
14348 |
--static u64 bfqio_cgroup_##__VAR##_read(struct cgroup *cgroup, \ |
14349 |
-+static u64 bfqio_cgroup_##__VAR##_read(struct cgroup_subsys_state *css, \ |
14350 |
- struct cftype *cftype) \ |
14351 |
- { \ |
14352 |
-- struct bfqio_cgroup *bgrp; \ |
14353 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
14354 |
- u64 ret = -ENODEV; \ |
14355 |
- \ |
14356 |
- mutex_lock(&bfqio_mutex); \ |
14357 |
-- if (bfqio_is_removed(cgroup)) \ |
14358 |
-+ if (bfqio_is_removed(bgrp)) \ |
14359 |
- goto out_unlock; \ |
14360 |
- \ |
14361 |
-- bgrp = cgroup_to_bfqio(cgroup); \ |
14362 |
- spin_lock_irq(&bgrp->lock); \ |
14363 |
- ret = bgrp->__VAR; \ |
14364 |
- spin_unlock_irq(&bgrp->lock); \ |
14365 |
-@@ -648,11 +647,11 @@ SHOW_FUNCTION(ioprio_class); |
14366 |
- #undef SHOW_FUNCTION |
14367 |
- |
14368 |
- #define STORE_FUNCTION(__VAR, __MIN, __MAX) \ |
14369 |
--static int bfqio_cgroup_##__VAR##_write(struct cgroup *cgroup, \ |
14370 |
-+static int bfqio_cgroup_##__VAR##_write(struct cgroup_subsys_state *css,\ |
14371 |
- struct cftype *cftype, \ |
14372 |
- u64 val) \ |
14373 |
- { \ |
14374 |
-- struct bfqio_cgroup *bgrp; \ |
14375 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
14376 |
- struct bfq_group *bfqg; \ |
14377 |
- int ret = -EINVAL; \ |
14378 |
- \ |
14379 |
-@@ -661,12 +660,10 @@ static int bfqio_cgroup_##__VAR##_write(struct cgroup *cgroup, \ |
14380 |
- \ |
14381 |
- ret = -ENODEV; \ |
14382 |
- mutex_lock(&bfqio_mutex); \ |
14383 |
-- if (bfqio_is_removed(cgroup)) \ |
14384 |
-+ if (bfqio_is_removed(bgrp)) \ |
14385 |
- goto out_unlock; \ |
14386 |
- ret = 0; \ |
14387 |
- \ |
14388 |
-- bgrp = cgroup_to_bfqio(cgroup); \ |
14389 |
-- \ |
14390 |
- spin_lock_irq(&bgrp->lock); \ |
14391 |
- bgrp->__VAR = (unsigned short)val; \ |
14392 |
- hlist_for_each_entry(bfqg, &bgrp->group_data, group_node) { \ |
14393 |
-@@ -713,11 +710,11 @@ static struct cftype bfqio_files[] = { |
14394 |
- { }, /* terminate */ |
14395 |
- }; |
14396 |
- |
14397 |
--static struct cgroup_subsys_state *bfqio_create(struct cgroup *cgroup) |
14398 |
-+static struct cgroup_subsys_state *bfqio_create(struct cgroup_subsys_state *parent_css) |
14399 |
- { |
14400 |
- struct bfqio_cgroup *bgrp; |
14401 |
- |
14402 |
-- if (cgroup->parent != NULL) { |
14403 |
-+ if (parent_css != NULL) { |
14404 |
- bgrp = kzalloc(sizeof(*bgrp), GFP_KERNEL); |
14405 |
- if (bgrp == NULL) |
14406 |
- return ERR_PTR(-ENOMEM); |
14407 |
-@@ -740,13 +737,14 @@ static struct cgroup_subsys_state *bfqio_create(struct cgroup *cgroup) |
14408 |
- * behavior is that a group containing a task that forked using CLONE_IO |
14409 |
- * will not be destroyed until the tasks sharing the ioc die. |
14410 |
- */ |
14411 |
--static int bfqio_can_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
14412 |
-+static int bfqio_can_attach(struct cgroup_subsys_state *css, |
14413 |
-+ struct cgroup_taskset *tset) |
14414 |
- { |
14415 |
- struct task_struct *task; |
14416 |
- struct io_context *ioc; |
14417 |
- int ret = 0; |
14418 |
- |
14419 |
-- cgroup_taskset_for_each(task, cgroup, tset) { |
14420 |
-+ cgroup_taskset_for_each(task, css, tset) { |
14421 |
- /* task_lock() is needed to avoid races with exit_io_context() */ |
14422 |
- task_lock(task); |
14423 |
- ioc = task->io_context; |
14424 |
-@@ -766,7 +764,8 @@ static int bfqio_can_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
14425 |
- return ret; |
14426 |
- } |
14427 |
- |
14428 |
--static void bfqio_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
14429 |
-+static void bfqio_attach(struct cgroup_subsys_state *css, |
14430 |
-+ struct cgroup_taskset *tset) |
14431 |
- { |
14432 |
- struct task_struct *task; |
14433 |
- struct io_context *ioc; |
14434 |
-@@ -776,7 +775,7 @@ static void bfqio_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
14435 |
- * IMPORTANT NOTE: The move of more than one process at a time to a |
14436 |
- * new group has not yet been tested. |
14437 |
- */ |
14438 |
-- cgroup_taskset_for_each(task, cgroup, tset) { |
14439 |
-+ cgroup_taskset_for_each(task, css, tset) { |
14440 |
- ioc = get_task_io_context(task, GFP_ATOMIC, NUMA_NO_NODE); |
14441 |
- if (ioc) { |
14442 |
- /* |
14443 |
-@@ -787,16 +786,16 @@ static void bfqio_attach(struct cgroup *cgroup, struct cgroup_taskset *tset) |
14444 |
- if (!strncmp(icq->q->elevator->type->elevator_name, |
14445 |
- "bfq", ELV_NAME_MAX)) |
14446 |
- bfq_bic_change_cgroup(icq_to_bic(icq), |
14447 |
-- cgroup); |
14448 |
-+ css); |
14449 |
- rcu_read_unlock(); |
14450 |
- put_io_context(ioc); |
14451 |
- } |
14452 |
- } |
14453 |
- } |
14454 |
- |
14455 |
--static void bfqio_destroy(struct cgroup *cgroup) |
14456 |
-+static void bfqio_destroy(struct cgroup_subsys_state *css) |
14457 |
- { |
14458 |
-- struct bfqio_cgroup *bgrp = cgroup_to_bfqio(cgroup); |
14459 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
14460 |
- struct hlist_node *tmp; |
14461 |
- struct bfq_group *bfqg; |
14462 |
- |
14463 |
-@@ -815,9 +814,31 @@ static void bfqio_destroy(struct cgroup *cgroup) |
14464 |
- kfree(bgrp); |
14465 |
- } |
14466 |
- |
14467 |
-+static int bfqio_css_online(struct cgroup_subsys_state *css) |
14468 |
-+{ |
14469 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
14470 |
-+ |
14471 |
-+ mutex_lock(&bfqio_mutex); |
14472 |
-+ bgrp->online = true; |
14473 |
-+ mutex_unlock(&bfqio_mutex); |
14474 |
-+ |
14475 |
-+ return 0; |
14476 |
-+} |
14477 |
-+ |
14478 |
-+static void bfqio_css_offline(struct cgroup_subsys_state *css) |
14479 |
-+{ |
14480 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
14481 |
-+ |
14482 |
-+ mutex_lock(&bfqio_mutex); |
14483 |
-+ bgrp->online = false; |
14484 |
-+ mutex_unlock(&bfqio_mutex); |
14485 |
-+} |
14486 |
-+ |
14487 |
- struct cgroup_subsys bfqio_subsys = { |
14488 |
- .name = "bfqio", |
14489 |
- .css_alloc = bfqio_create, |
14490 |
-+ .css_online = bfqio_css_online, |
14491 |
-+ .css_offline = bfqio_css_offline, |
14492 |
- .can_attach = bfqio_can_attach, |
14493 |
- .attach = bfqio_attach, |
14494 |
- .css_free = bfqio_destroy, |
14495 |
-diff --git a/block/bfq.h b/block/bfq.h |
14496 |
-index bb52975..885e62c 100644 |
14497 |
---- a/block/bfq.h |
14498 |
-+++ b/block/bfq.h |
14499 |
-@@ -510,6 +510,7 @@ struct bfq_group { |
14500 |
- /** |
14501 |
- * struct bfqio_cgroup - bfq cgroup data structure. |
14502 |
- * @css: subsystem state for bfq in the containing cgroup. |
14503 |
-+ * @online: flag marked when the subsystem is inserted. |
14504 |
- * @weight: cgroup weight. |
14505 |
- * @ioprio: cgroup ioprio. |
14506 |
- * @ioprio_class: cgroup ioprio_class. |
14507 |
-@@ -521,6 +522,7 @@ struct bfq_group { |
14508 |
- */ |
14509 |
- struct bfqio_cgroup { |
14510 |
- struct cgroup_subsys_state css; |
14511 |
-+ bool online; |
14512 |
- |
14513 |
- unsigned short weight, ioprio, ioprio_class; |
14514 |
- |
14515 |
--- |
14516 |
-1.8.1.4 |
14517 |
- |