1 |
Author: mpagano |
2 |
Date: 2014-04-01 18:44:20 +0000 (Tue, 01 Apr 2014) |
3 |
New Revision: 2727 |
4 |
|
5 |
Added: |
6 |
genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch1 |
7 |
Removed: |
8 |
genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch |
9 |
Modified: |
10 |
genpatches-2.6/trunk/3.14/0000_README |
11 |
Log: |
12 |
Fix BFQ patch name to correct PATCH_DEPTH |
13 |
|
14 |
Modified: genpatches-2.6/trunk/3.14/0000_README |
15 |
=================================================================== |
16 |
--- genpatches-2.6/trunk/3.14/0000_README 2014-04-01 12:40:19 UTC (rev 2726) |
17 |
+++ genpatches-2.6/trunk/3.14/0000_README 2014-04-01 18:44:20 UTC (rev 2727) |
18 |
@@ -86,7 +86,7 @@ |
19 |
From: http://algo.ing.unimo.it/people/paolo/disk_sched/ |
20 |
Desc: BFQ v7r2 patch 1 for 3.14: Build, cgroups and kconfig bits |
21 |
|
22 |
-Patch: 5002_BFQ-2-block-introduce-the-v7r2-I-O-sched-for-3.14.patch |
23 |
+Patch: 5002_BFQ-2-block-introduce-the-v7r2-I-O-sched-for-3.14.patch1 |
24 |
From: http://algo.ing.unimo.it/people/paolo/disk_sched/ |
25 |
Desc: BFQ v7r2 patch 2 for 3.14: BFQ Scheduler |
26 |
|
27 |
|
28 |
Deleted: genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch |
29 |
=================================================================== |
30 |
--- genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch 2014-04-01 12:40:19 UTC (rev 2726) |
31 |
+++ genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch 2014-04-01 18:44:20 UTC (rev 2727) |
32 |
@@ -1,6065 +0,0 @@ |
33 |
-From 5055277df59d9280da6b60cf90bed8e5e57dc44d Mon Sep 17 00:00:00 2001 |
34 |
-From: Paolo Valente <paolo.valente@×××××××.it> |
35 |
-Date: Thu, 9 May 2013 19:10:02 +0200 |
36 |
-Subject: [PATCH 2/3] block: introduce the BFQ-v7r2 I/O sched for 3.14 |
37 |
- |
38 |
-Add the BFQ-v7r2 I/O scheduler to 3.14. |
39 |
-The general structure is borrowed from CFQ, as much of the code for |
40 |
-handling I/O contexts. Over time, several useful features have been |
41 |
-ported from CFQ as well (details in the changelog in README.BFQ). A |
42 |
-(bfq_)queue is associated to each task doing I/O on a device, and each |
43 |
-time a scheduling decision has to be made a queue is selected and served |
44 |
-until it expires. |
45 |
- |
46 |
- - Slices are given in the service domain: tasks are assigned |
47 |
- budgets, measured in number of sectors. Once got the disk, a task |
48 |
- must however consume its assigned budget within a configurable |
49 |
- maximum time (by default, the maximum possible value of the |
50 |
- budgets is automatically computed to comply with this timeout). |
51 |
- This allows the desired latency vs "throughput boosting" tradeoff |
52 |
- to be set. |
53 |
- |
54 |
- - Budgets are scheduled according to a variant of WF2Q+, implemented |
55 |
- using an augmented rb-tree to take eligibility into account while |
56 |
- preserving an O(log N) overall complexity. |
57 |
- |
58 |
- - A low-latency tunable is provided; if enabled, both interactive |
59 |
- and soft real-time applications are guaranteed a very low latency. |
60 |
- |
61 |
- - Latency guarantees are preserved also in the presence of NCQ. |
62 |
- |
63 |
- - Also with flash-based devices, a high throughput is achieved |
64 |
- while still preserving latency guarantees. |
65 |
- |
66 |
- - BFQ features Early Queue Merge (EQM), a sort of fusion of the |
67 |
- cooperating-queue-merging and the preemption mechanisms present |
68 |
- in CFQ. EQM is in fact a unified mechanism that tries to get a |
69 |
- sequential read pattern, and hence a high throughput, with any |
70 |
- set of processes performing interleaved I/O over a contiguous |
71 |
- sequence of sectors. |
72 |
- |
73 |
- - BFQ supports full hierarchical scheduling, exporting a cgroups |
74 |
- interface. Since each node has a full scheduler, each group can |
75 |
- be assigned its own weight. |
76 |
- |
77 |
- - If the cgroups interface is not used, only I/O priorities can be |
78 |
- assigned to processes, with ioprio values mapped to weights |
79 |
- with the relation weight = IOPRIO_BE_NR - ioprio. |
80 |
- |
81 |
- - ioprio classes are served in strict priority order, i.e., lower |
82 |
- priority queues are not served as long as there are higher |
83 |
- priority queues. Among queues in the same class the bandwidth is |
84 |
- distributed in proportion to the weight of each queue. A very |
85 |
- thin extra bandwidth is however guaranteed to the Idle class, to |
86 |
- prevent it from starving. |
87 |
- |
88 |
-Signed-off-by: Paolo Valente <paolo.valente@×××××××.it> |
89 |
-Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
90 |
---- |
91 |
- block/bfq-cgroup.c | 926 +++++++++++++++ |
92 |
- block/bfq-ioc.c | 36 + |
93 |
- block/bfq-iosched.c | 3300 +++++++++++++++++++++++++++++++++++++++++++++++++++ |
94 |
- block/bfq-sched.c | 1078 +++++++++++++++++ |
95 |
- block/bfq.h | 622 ++++++++++ |
96 |
- 5 files changed, 5962 insertions(+) |
97 |
- create mode 100644 block/bfq-cgroup.c |
98 |
- create mode 100644 block/bfq-ioc.c |
99 |
- create mode 100644 block/bfq-iosched.c |
100 |
- create mode 100644 block/bfq-sched.c |
101 |
- create mode 100644 block/bfq.h |
102 |
- |
103 |
-diff --git a/block/bfq-cgroup.c b/block/bfq-cgroup.c |
104 |
-new file mode 100644 |
105 |
-index 0000000..bcecdb4 |
106 |
---- /dev/null |
107 |
-+++ b/block/bfq-cgroup.c |
108 |
-@@ -0,0 +1,926 @@ |
109 |
-+/* |
110 |
-+ * BFQ: CGROUPS support. |
111 |
-+ * |
112 |
-+ * Based on ideas and code from CFQ: |
113 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
114 |
-+ * |
115 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
116 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
117 |
-+ * |
118 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
119 |
-+ * |
120 |
-+ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
121 |
-+ */ |
122 |
-+ |
123 |
-+#ifdef CONFIG_CGROUP_BFQIO |
124 |
-+ |
125 |
-+static DEFINE_MUTEX(bfqio_mutex); |
126 |
-+ |
127 |
-+static bool bfqio_is_removed(struct bfqio_cgroup *bgrp) |
128 |
-+{ |
129 |
-+ return bgrp ? !bgrp->online : false; |
130 |
-+} |
131 |
-+ |
132 |
-+static struct bfqio_cgroup bfqio_root_cgroup = { |
133 |
-+ .weight = BFQ_DEFAULT_GRP_WEIGHT, |
134 |
-+ .ioprio = BFQ_DEFAULT_GRP_IOPRIO, |
135 |
-+ .ioprio_class = BFQ_DEFAULT_GRP_CLASS, |
136 |
-+}; |
137 |
-+ |
138 |
-+static inline void bfq_init_entity(struct bfq_entity *entity, |
139 |
-+ struct bfq_group *bfqg) |
140 |
-+{ |
141 |
-+ entity->weight = entity->new_weight; |
142 |
-+ entity->orig_weight = entity->new_weight; |
143 |
-+ entity->ioprio = entity->new_ioprio; |
144 |
-+ entity->ioprio_class = entity->new_ioprio_class; |
145 |
-+ entity->parent = bfqg->my_entity; |
146 |
-+ entity->sched_data = &bfqg->sched_data; |
147 |
-+} |
148 |
-+ |
149 |
-+static struct bfqio_cgroup *css_to_bfqio(struct cgroup_subsys_state *css) |
150 |
-+{ |
151 |
-+ return css ? container_of(css, struct bfqio_cgroup, css) : NULL; |
152 |
-+} |
153 |
-+ |
154 |
-+/* |
155 |
-+ * Search the bfq_group for bfqd into the hash table (by now only a list) |
156 |
-+ * of bgrp. Must be called under rcu_read_lock(). |
157 |
-+ */ |
158 |
-+static struct bfq_group *bfqio_lookup_group(struct bfqio_cgroup *bgrp, |
159 |
-+ struct bfq_data *bfqd) |
160 |
-+{ |
161 |
-+ struct bfq_group *bfqg; |
162 |
-+ void *key; |
163 |
-+ |
164 |
-+ hlist_for_each_entry_rcu(bfqg, &bgrp->group_data, group_node) { |
165 |
-+ key = rcu_dereference(bfqg->bfqd); |
166 |
-+ if (key == bfqd) |
167 |
-+ return bfqg; |
168 |
-+ } |
169 |
-+ |
170 |
-+ return NULL; |
171 |
-+} |
172 |
-+ |
173 |
-+static inline void bfq_group_init_entity(struct bfqio_cgroup *bgrp, |
174 |
-+ struct bfq_group *bfqg) |
175 |
-+{ |
176 |
-+ struct bfq_entity *entity = &bfqg->entity; |
177 |
-+ |
178 |
-+ /* |
179 |
-+ * If the weight of the entity has never been set via the sysfs |
180 |
-+ * interface, then bgrp->weight == 0. In this case we initialize |
181 |
-+ * the weight from the current ioprio value. Otherwise, the group |
182 |
-+ * weight, if set, has priority over the ioprio value. |
183 |
-+ */ |
184 |
-+ if (bgrp->weight == 0) { |
185 |
-+ entity->new_weight = bfq_ioprio_to_weight(bgrp->ioprio); |
186 |
-+ entity->new_ioprio = bgrp->ioprio; |
187 |
-+ } else { |
188 |
-+ entity->new_weight = bgrp->weight; |
189 |
-+ entity->new_ioprio = bfq_weight_to_ioprio(bgrp->weight); |
190 |
-+ } |
191 |
-+ entity->orig_weight = entity->weight = entity->new_weight; |
192 |
-+ entity->ioprio = entity->new_ioprio; |
193 |
-+ entity->ioprio_class = entity->new_ioprio_class = bgrp->ioprio_class; |
194 |
-+ entity->my_sched_data = &bfqg->sched_data; |
195 |
-+} |
196 |
-+ |
197 |
-+static inline void bfq_group_set_parent(struct bfq_group *bfqg, |
198 |
-+ struct bfq_group *parent) |
199 |
-+{ |
200 |
-+ struct bfq_entity *entity; |
201 |
-+ |
202 |
-+ BUG_ON(parent == NULL); |
203 |
-+ BUG_ON(bfqg == NULL); |
204 |
-+ |
205 |
-+ entity = &bfqg->entity; |
206 |
-+ entity->parent = parent->my_entity; |
207 |
-+ entity->sched_data = &parent->sched_data; |
208 |
-+} |
209 |
-+ |
210 |
-+/** |
211 |
-+ * bfq_group_chain_alloc - allocate a chain of groups. |
212 |
-+ * @bfqd: queue descriptor. |
213 |
-+ * @css: the leaf cgroup_subsys_state this chain starts from. |
214 |
-+ * |
215 |
-+ * Allocate a chain of groups starting from the one belonging to |
216 |
-+ * @cgroup up to the root cgroup. Stop if a cgroup on the chain |
217 |
-+ * to the root has already an allocated group on @bfqd. |
218 |
-+ */ |
219 |
-+static struct bfq_group *bfq_group_chain_alloc(struct bfq_data *bfqd, |
220 |
-+ struct cgroup_subsys_state *css) |
221 |
-+{ |
222 |
-+ struct bfqio_cgroup *bgrp; |
223 |
-+ struct bfq_group *bfqg, *prev = NULL, *leaf = NULL; |
224 |
-+ |
225 |
-+ for (; css != NULL; css = css->parent) { |
226 |
-+ bgrp = css_to_bfqio(css); |
227 |
-+ |
228 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
229 |
-+ if (bfqg != NULL) { |
230 |
-+ /* |
231 |
-+ * All the cgroups in the path from there to the |
232 |
-+ * root must have a bfq_group for bfqd, so we don't |
233 |
-+ * need any more allocations. |
234 |
-+ */ |
235 |
-+ break; |
236 |
-+ } |
237 |
-+ |
238 |
-+ bfqg = kzalloc(sizeof(*bfqg), GFP_ATOMIC); |
239 |
-+ if (bfqg == NULL) |
240 |
-+ goto cleanup; |
241 |
-+ |
242 |
-+ bfq_group_init_entity(bgrp, bfqg); |
243 |
-+ bfqg->my_entity = &bfqg->entity; |
244 |
-+ |
245 |
-+ if (leaf == NULL) { |
246 |
-+ leaf = bfqg; |
247 |
-+ prev = leaf; |
248 |
-+ } else { |
249 |
-+ bfq_group_set_parent(prev, bfqg); |
250 |
-+ /* |
251 |
-+ * Build a list of allocated nodes using the bfqd |
252 |
-+ * filed, that is still unused and will be initialized |
253 |
-+ * only after the node will be connected. |
254 |
-+ */ |
255 |
-+ prev->bfqd = bfqg; |
256 |
-+ prev = bfqg; |
257 |
-+ } |
258 |
-+ } |
259 |
-+ |
260 |
-+ return leaf; |
261 |
-+ |
262 |
-+cleanup: |
263 |
-+ while (leaf != NULL) { |
264 |
-+ prev = leaf; |
265 |
-+ leaf = leaf->bfqd; |
266 |
-+ kfree(prev); |
267 |
-+ } |
268 |
-+ |
269 |
-+ return NULL; |
270 |
-+} |
271 |
-+ |
272 |
-+/** |
273 |
-+ * bfq_group_chain_link - link an allocated group chain to a cgroup hierarchy. |
274 |
-+ * @bfqd: the queue descriptor. |
275 |
-+ * @css: the leaf cgroup_subsys_state to start from. |
276 |
-+ * @leaf: the leaf group (to be associated to @cgroup). |
277 |
-+ * |
278 |
-+ * Try to link a chain of groups to a cgroup hierarchy, connecting the |
279 |
-+ * nodes bottom-up, so we can be sure that when we find a cgroup in the |
280 |
-+ * hierarchy that already as a group associated to @bfqd all the nodes |
281 |
-+ * in the path to the root cgroup have one too. |
282 |
-+ * |
283 |
-+ * On locking: the queue lock protects the hierarchy (there is a hierarchy |
284 |
-+ * per device) while the bfqio_cgroup lock protects the list of groups |
285 |
-+ * belonging to the same cgroup. |
286 |
-+ */ |
287 |
-+static void bfq_group_chain_link(struct bfq_data *bfqd, |
288 |
-+ struct cgroup_subsys_state *css, |
289 |
-+ struct bfq_group *leaf) |
290 |
-+{ |
291 |
-+ struct bfqio_cgroup *bgrp; |
292 |
-+ struct bfq_group *bfqg, *next, *prev = NULL; |
293 |
-+ unsigned long flags; |
294 |
-+ |
295 |
-+ assert_spin_locked(bfqd->queue->queue_lock); |
296 |
-+ |
297 |
-+ for (; css != NULL && leaf != NULL; css = css->parent) { |
298 |
-+ bgrp = css_to_bfqio(css); |
299 |
-+ next = leaf->bfqd; |
300 |
-+ |
301 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
302 |
-+ BUG_ON(bfqg != NULL); |
303 |
-+ |
304 |
-+ spin_lock_irqsave(&bgrp->lock, flags); |
305 |
-+ |
306 |
-+ rcu_assign_pointer(leaf->bfqd, bfqd); |
307 |
-+ hlist_add_head_rcu(&leaf->group_node, &bgrp->group_data); |
308 |
-+ hlist_add_head(&leaf->bfqd_node, &bfqd->group_list); |
309 |
-+ |
310 |
-+ spin_unlock_irqrestore(&bgrp->lock, flags); |
311 |
-+ |
312 |
-+ prev = leaf; |
313 |
-+ leaf = next; |
314 |
-+ } |
315 |
-+ |
316 |
-+ BUG_ON(css == NULL && leaf != NULL); |
317 |
-+ if (css != NULL && prev != NULL) { |
318 |
-+ bgrp = css_to_bfqio(css); |
319 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
320 |
-+ bfq_group_set_parent(prev, bfqg); |
321 |
-+ } |
322 |
-+} |
323 |
-+ |
324 |
-+/** |
325 |
-+ * bfq_find_alloc_group - return the group associated to @bfqd in @cgroup. |
326 |
-+ * @bfqd: queue descriptor. |
327 |
-+ * @cgroup: cgroup being searched for. |
328 |
-+ * |
329 |
-+ * Return a group associated to @bfqd in @cgroup, allocating one if |
330 |
-+ * necessary. When a group is returned all the cgroups in the path |
331 |
-+ * to the root have a group associated to @bfqd. |
332 |
-+ * |
333 |
-+ * If the allocation fails, return the root group: this breaks guarantees |
334 |
-+ * but is a safe fallback. If this loss becomes a problem it can be |
335 |
-+ * mitigated using the equivalent weight (given by the product of the |
336 |
-+ * weights of the groups in the path from @group to the root) in the |
337 |
-+ * root scheduler. |
338 |
-+ * |
339 |
-+ * We allocate all the missing nodes in the path from the leaf cgroup |
340 |
-+ * to the root and we connect the nodes only after all the allocations |
341 |
-+ * have been successful. |
342 |
-+ */ |
343 |
-+static struct bfq_group *bfq_find_alloc_group(struct bfq_data *bfqd, |
344 |
-+ struct cgroup_subsys_state *css) |
345 |
-+{ |
346 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
347 |
-+ struct bfq_group *bfqg; |
348 |
-+ |
349 |
-+ bfqg = bfqio_lookup_group(bgrp, bfqd); |
350 |
-+ if (bfqg != NULL) |
351 |
-+ return bfqg; |
352 |
-+ |
353 |
-+ bfqg = bfq_group_chain_alloc(bfqd, css); |
354 |
-+ if (bfqg != NULL) |
355 |
-+ bfq_group_chain_link(bfqd, css, bfqg); |
356 |
-+ else |
357 |
-+ bfqg = bfqd->root_group; |
358 |
-+ |
359 |
-+ return bfqg; |
360 |
-+} |
361 |
-+ |
362 |
-+/** |
363 |
-+ * bfq_bfqq_move - migrate @bfqq to @bfqg. |
364 |
-+ * @bfqd: queue descriptor. |
365 |
-+ * @bfqq: the queue to move. |
366 |
-+ * @entity: @bfqq's entity. |
367 |
-+ * @bfqg: the group to move to. |
368 |
-+ * |
369 |
-+ * Move @bfqq to @bfqg, deactivating it from its old group and reactivating |
370 |
-+ * it on the new one. Avoid putting the entity on the old group idle tree. |
371 |
-+ * |
372 |
-+ * Must be called under the queue lock; the cgroup owning @bfqg must |
373 |
-+ * not disappear (by now this just means that we are called under |
374 |
-+ * rcu_read_lock()). |
375 |
-+ */ |
376 |
-+static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
377 |
-+ struct bfq_entity *entity, struct bfq_group *bfqg) |
378 |
-+{ |
379 |
-+ int busy, resume; |
380 |
-+ |
381 |
-+ busy = bfq_bfqq_busy(bfqq); |
382 |
-+ resume = !RB_EMPTY_ROOT(&bfqq->sort_list); |
383 |
-+ |
384 |
-+ BUG_ON(resume && !entity->on_st); |
385 |
-+ BUG_ON(busy && !resume && entity->on_st && |
386 |
-+ bfqq != bfqd->in_service_queue); |
387 |
-+ |
388 |
-+ if (busy) { |
389 |
-+ BUG_ON(atomic_read(&bfqq->ref) < 2); |
390 |
-+ |
391 |
-+ if (!resume) |
392 |
-+ bfq_del_bfqq_busy(bfqd, bfqq, 0); |
393 |
-+ else |
394 |
-+ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
395 |
-+ } else if (entity->on_st) |
396 |
-+ bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); |
397 |
-+ |
398 |
-+ /* |
399 |
-+ * Here we use a reference to bfqg. We don't need a refcounter |
400 |
-+ * as the cgroup reference will not be dropped, so that its |
401 |
-+ * destroy() callback will not be invoked. |
402 |
-+ */ |
403 |
-+ entity->parent = bfqg->my_entity; |
404 |
-+ entity->sched_data = &bfqg->sched_data; |
405 |
-+ |
406 |
-+ if (busy && resume) |
407 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
408 |
-+ |
409 |
-+ if (bfqd->in_service_queue == NULL && !bfqd->rq_in_driver) |
410 |
-+ bfq_schedule_dispatch(bfqd); |
411 |
-+} |
412 |
-+ |
413 |
-+/** |
414 |
-+ * __bfq_bic_change_cgroup - move @bic to @cgroup. |
415 |
-+ * @bfqd: the queue descriptor. |
416 |
-+ * @bic: the bic to move. |
417 |
-+ * @cgroup: the cgroup to move to. |
418 |
-+ * |
419 |
-+ * Move bic to cgroup, assuming that bfqd->queue is locked; the caller |
420 |
-+ * has to make sure that the reference to cgroup is valid across the call. |
421 |
-+ * |
422 |
-+ * NOTE: an alternative approach might have been to store the current |
423 |
-+ * cgroup in bfqq and getting a reference to it, reducing the lookup |
424 |
-+ * time here, at the price of slightly more complex code. |
425 |
-+ */ |
426 |
-+static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
427 |
-+ struct bfq_io_cq *bic, |
428 |
-+ struct cgroup_subsys_state *css) |
429 |
-+{ |
430 |
-+ struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0); |
431 |
-+ struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1); |
432 |
-+ struct bfq_entity *entity; |
433 |
-+ struct bfq_group *bfqg; |
434 |
-+ struct bfqio_cgroup *bgrp; |
435 |
-+ |
436 |
-+ bgrp = css_to_bfqio(css); |
437 |
-+ |
438 |
-+ bfqg = bfq_find_alloc_group(bfqd, css); |
439 |
-+ if (async_bfqq != NULL) { |
440 |
-+ entity = &async_bfqq->entity; |
441 |
-+ |
442 |
-+ if (entity->sched_data != &bfqg->sched_data) { |
443 |
-+ bic_set_bfqq(bic, NULL, 0); |
444 |
-+ bfq_log_bfqq(bfqd, async_bfqq, |
445 |
-+ "bic_change_group: %p %d", |
446 |
-+ async_bfqq, atomic_read(&async_bfqq->ref)); |
447 |
-+ bfq_put_queue(async_bfqq); |
448 |
-+ } |
449 |
-+ } |
450 |
-+ |
451 |
-+ if (sync_bfqq != NULL) { |
452 |
-+ entity = &sync_bfqq->entity; |
453 |
-+ if (entity->sched_data != &bfqg->sched_data) |
454 |
-+ bfq_bfqq_move(bfqd, sync_bfqq, entity, bfqg); |
455 |
-+ } |
456 |
-+ |
457 |
-+ return bfqg; |
458 |
-+} |
459 |
-+ |
460 |
-+/** |
461 |
-+ * bfq_bic_change_cgroup - move @bic to @cgroup. |
462 |
-+ * @bic: the bic being migrated. |
463 |
-+ * @cgroup: the destination cgroup. |
464 |
-+ * |
465 |
-+ * When the task owning @bic is moved to @cgroup, @bic is immediately |
466 |
-+ * moved into its new parent group. |
467 |
-+ */ |
468 |
-+static void bfq_bic_change_cgroup(struct bfq_io_cq *bic, |
469 |
-+ struct cgroup_subsys_state *css) |
470 |
-+{ |
471 |
-+ struct bfq_data *bfqd; |
472 |
-+ unsigned long uninitialized_var(flags); |
473 |
-+ |
474 |
-+ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), |
475 |
-+ &flags); |
476 |
-+ if (bfqd != NULL) { |
477 |
-+ __bfq_bic_change_cgroup(bfqd, bic, css); |
478 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
479 |
-+ } |
480 |
-+} |
481 |
-+ |
482 |
-+/** |
483 |
-+ * bfq_bic_update_cgroup - update the cgroup of @bic. |
484 |
-+ * @bic: the @bic to update. |
485 |
-+ * |
486 |
-+ * Make sure that @bic is enqueued in the cgroup of the current task. |
487 |
-+ * We need this in addition to moving bics during the cgroup attach |
488 |
-+ * phase because the task owning @bic could be at its first disk |
489 |
-+ * access or we may end up in the root cgroup as the result of a |
490 |
-+ * memory allocation failure and here we try to move to the right |
491 |
-+ * group. |
492 |
-+ * |
493 |
-+ * Must be called under the queue lock. It is safe to use the returned |
494 |
-+ * value even after the rcu_read_unlock() as the migration/destruction |
495 |
-+ * paths act under the queue lock too. IOW it is impossible to race with |
496 |
-+ * group migration/destruction and end up with an invalid group as: |
497 |
-+ * a) here cgroup has not yet been destroyed, nor its destroy callback |
498 |
-+ * has started execution, as current holds a reference to it, |
499 |
-+ * b) if it is destroyed after rcu_read_unlock() [after current is |
500 |
-+ * migrated to a different cgroup] its attach() callback will have |
501 |
-+ * taken care of remove all the references to the old cgroup data. |
502 |
-+ */ |
503 |
-+static struct bfq_group *bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
504 |
-+{ |
505 |
-+ struct bfq_data *bfqd = bic_to_bfqd(bic); |
506 |
-+ struct bfq_group *bfqg; |
507 |
-+ struct cgroup_subsys_state *css; |
508 |
-+ |
509 |
-+ BUG_ON(bfqd == NULL); |
510 |
-+ |
511 |
-+ rcu_read_lock(); |
512 |
-+ css = task_css(current, bfqio_subsys_id); |
513 |
-+ bfqg = __bfq_bic_change_cgroup(bfqd, bic, css); |
514 |
-+ rcu_read_unlock(); |
515 |
-+ |
516 |
-+ return bfqg; |
517 |
-+} |
518 |
-+ |
519 |
-+/** |
520 |
-+ * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st. |
521 |
-+ * @st: the service tree being flushed. |
522 |
-+ */ |
523 |
-+static inline void bfq_flush_idle_tree(struct bfq_service_tree *st) |
524 |
-+{ |
525 |
-+ struct bfq_entity *entity = st->first_idle; |
526 |
-+ |
527 |
-+ for (; entity != NULL; entity = st->first_idle) |
528 |
-+ __bfq_deactivate_entity(entity, 0); |
529 |
-+} |
530 |
-+ |
531 |
-+/** |
532 |
-+ * bfq_reparent_leaf_entity - move leaf entity to the root_group. |
533 |
-+ * @bfqd: the device data structure with the root group. |
534 |
-+ * @entity: the entity to move. |
535 |
-+ */ |
536 |
-+static inline void bfq_reparent_leaf_entity(struct bfq_data *bfqd, |
537 |
-+ struct bfq_entity *entity) |
538 |
-+{ |
539 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
540 |
-+ |
541 |
-+ BUG_ON(bfqq == NULL); |
542 |
-+ bfq_bfqq_move(bfqd, bfqq, entity, bfqd->root_group); |
543 |
-+ return; |
544 |
-+} |
545 |
-+ |
546 |
-+/** |
547 |
-+ * bfq_reparent_active_entities - move to the root group all active entities. |
548 |
-+ * @bfqd: the device data structure with the root group. |
549 |
-+ * @bfqg: the group to move from. |
550 |
-+ * @st: the service tree with the entities. |
551 |
-+ * |
552 |
-+ * Needs queue_lock to be taken and reference to be valid over the call. |
553 |
-+ */ |
554 |
-+static inline void bfq_reparent_active_entities(struct bfq_data *bfqd, |
555 |
-+ struct bfq_group *bfqg, |
556 |
-+ struct bfq_service_tree *st) |
557 |
-+{ |
558 |
-+ struct rb_root *active = &st->active; |
559 |
-+ struct bfq_entity *entity = NULL; |
560 |
-+ |
561 |
-+ if (!RB_EMPTY_ROOT(&st->active)) |
562 |
-+ entity = bfq_entity_of(rb_first(active)); |
563 |
-+ |
564 |
-+ for (; entity != NULL; entity = bfq_entity_of(rb_first(active))) |
565 |
-+ bfq_reparent_leaf_entity(bfqd, entity); |
566 |
-+ |
567 |
-+ if (bfqg->sched_data.in_service_entity != NULL) |
568 |
-+ bfq_reparent_leaf_entity(bfqd, |
569 |
-+ bfqg->sched_data.in_service_entity); |
570 |
-+ |
571 |
-+ return; |
572 |
-+} |
573 |
-+ |
574 |
-+/** |
575 |
-+ * bfq_destroy_group - destroy @bfqg. |
576 |
-+ * @bgrp: the bfqio_cgroup containing @bfqg. |
577 |
-+ * @bfqg: the group being destroyed. |
578 |
-+ * |
579 |
-+ * Destroy @bfqg, making sure that it is not referenced from its parent. |
580 |
-+ */ |
581 |
-+static void bfq_destroy_group(struct bfqio_cgroup *bgrp, struct bfq_group *bfqg) |
582 |
-+{ |
583 |
-+ struct bfq_data *bfqd; |
584 |
-+ struct bfq_service_tree *st; |
585 |
-+ struct bfq_entity *entity = bfqg->my_entity; |
586 |
-+ unsigned long uninitialized_var(flags); |
587 |
-+ int i; |
588 |
-+ |
589 |
-+ hlist_del(&bfqg->group_node); |
590 |
-+ |
591 |
-+ /* |
592 |
-+ * Empty all service_trees belonging to this group before deactivating |
593 |
-+ * the group itself. |
594 |
-+ */ |
595 |
-+ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) { |
596 |
-+ st = bfqg->sched_data.service_tree + i; |
597 |
-+ |
598 |
-+ /* |
599 |
-+ * The idle tree may still contain bfq_queues belonging |
600 |
-+ * to exited task because they never migrated to a different |
601 |
-+ * cgroup from the one being destroyed now. No one else |
602 |
-+ * can access them so it's safe to act without any lock. |
603 |
-+ */ |
604 |
-+ bfq_flush_idle_tree(st); |
605 |
-+ |
606 |
-+ /* |
607 |
-+ * It may happen that some queues are still active |
608 |
-+ * (busy) upon group destruction (if the corresponding |
609 |
-+ * processes have been forced to terminate). We move |
610 |
-+ * all the leaf entities corresponding to these queues |
611 |
-+ * to the root_group. |
612 |
-+ * Also, it may happen that the group has an entity |
613 |
-+ * under service, which is disconnected from the active |
614 |
-+ * tree: it must be moved, too. |
615 |
-+ * There is no need to put the sync queues, as the |
616 |
-+ * scheduler has taken no reference. |
617 |
-+ */ |
618 |
-+ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
619 |
-+ if (bfqd != NULL) { |
620 |
-+ bfq_reparent_active_entities(bfqd, bfqg, st); |
621 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
622 |
-+ } |
623 |
-+ BUG_ON(!RB_EMPTY_ROOT(&st->active)); |
624 |
-+ BUG_ON(!RB_EMPTY_ROOT(&st->idle)); |
625 |
-+ } |
626 |
-+ BUG_ON(bfqg->sched_data.next_in_service != NULL); |
627 |
-+ BUG_ON(bfqg->sched_data.in_service_entity != NULL); |
628 |
-+ |
629 |
-+ /* |
630 |
-+ * We may race with device destruction, take extra care when |
631 |
-+ * dereferencing bfqg->bfqd. |
632 |
-+ */ |
633 |
-+ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
634 |
-+ if (bfqd != NULL) { |
635 |
-+ hlist_del(&bfqg->bfqd_node); |
636 |
-+ __bfq_deactivate_entity(entity, 0); |
637 |
-+ bfq_put_async_queues(bfqd, bfqg); |
638 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
639 |
-+ } |
640 |
-+ BUG_ON(entity->tree != NULL); |
641 |
-+ |
642 |
-+ /* |
643 |
-+ * No need to defer the kfree() to the end of the RCU grace |
644 |
-+ * period: we are called from the destroy() callback of our |
645 |
-+ * cgroup, so we can be sure that no one is a) still using |
646 |
-+ * this cgroup or b) doing lookups in it. |
647 |
-+ */ |
648 |
-+ kfree(bfqg); |
649 |
-+} |
650 |
-+ |
651 |
-+static void bfq_end_raising_async(struct bfq_data *bfqd) |
652 |
-+{ |
653 |
-+ struct hlist_node *tmp; |
654 |
-+ struct bfq_group *bfqg; |
655 |
-+ |
656 |
-+ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) |
657 |
-+ bfq_end_raising_async_queues(bfqd, bfqg); |
658 |
-+ bfq_end_raising_async_queues(bfqd, bfqd->root_group); |
659 |
-+} |
660 |
-+ |
661 |
-+/** |
662 |
-+ * bfq_disconnect_groups - disconnect @bfqd from all its groups. |
663 |
-+ * @bfqd: the device descriptor being exited. |
664 |
-+ * |
665 |
-+ * When the device exits we just make sure that no lookup can return |
666 |
-+ * the now unused group structures. They will be deallocated on cgroup |
667 |
-+ * destruction. |
668 |
-+ */ |
669 |
-+static void bfq_disconnect_groups(struct bfq_data *bfqd) |
670 |
-+{ |
671 |
-+ struct hlist_node *tmp; |
672 |
-+ struct bfq_group *bfqg; |
673 |
-+ |
674 |
-+ bfq_log(bfqd, "disconnect_groups beginning"); |
675 |
-+ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) { |
676 |
-+ hlist_del(&bfqg->bfqd_node); |
677 |
-+ |
678 |
-+ __bfq_deactivate_entity(bfqg->my_entity, 0); |
679 |
-+ |
680 |
-+ /* |
681 |
-+ * Don't remove from the group hash, just set an |
682 |
-+ * invalid key. No lookups can race with the |
683 |
-+ * assignment as bfqd is being destroyed; this |
684 |
-+ * implies also that new elements cannot be added |
685 |
-+ * to the list. |
686 |
-+ */ |
687 |
-+ rcu_assign_pointer(bfqg->bfqd, NULL); |
688 |
-+ |
689 |
-+ bfq_log(bfqd, "disconnect_groups: put async for group %p", |
690 |
-+ bfqg); |
691 |
-+ bfq_put_async_queues(bfqd, bfqg); |
692 |
-+ } |
693 |
-+} |
694 |
-+ |
695 |
-+static inline void bfq_free_root_group(struct bfq_data *bfqd) |
696 |
-+{ |
697 |
-+ struct bfqio_cgroup *bgrp = &bfqio_root_cgroup; |
698 |
-+ struct bfq_group *bfqg = bfqd->root_group; |
699 |
-+ |
700 |
-+ bfq_put_async_queues(bfqd, bfqg); |
701 |
-+ |
702 |
-+ spin_lock_irq(&bgrp->lock); |
703 |
-+ hlist_del_rcu(&bfqg->group_node); |
704 |
-+ spin_unlock_irq(&bgrp->lock); |
705 |
-+ |
706 |
-+ /* |
707 |
-+ * No need to synchronize_rcu() here: since the device is gone |
708 |
-+ * there cannot be any read-side access to its root_group. |
709 |
-+ */ |
710 |
-+ kfree(bfqg); |
711 |
-+} |
712 |
-+ |
713 |
-+static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
714 |
-+{ |
715 |
-+ struct bfq_group *bfqg; |
716 |
-+ struct bfqio_cgroup *bgrp; |
717 |
-+ int i; |
718 |
-+ |
719 |
-+ bfqg = kzalloc_node(sizeof(*bfqg), GFP_KERNEL, node); |
720 |
-+ if (bfqg == NULL) |
721 |
-+ return NULL; |
722 |
-+ |
723 |
-+ bfqg->entity.parent = NULL; |
724 |
-+ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
725 |
-+ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
726 |
-+ |
727 |
-+ bgrp = &bfqio_root_cgroup; |
728 |
-+ spin_lock_irq(&bgrp->lock); |
729 |
-+ rcu_assign_pointer(bfqg->bfqd, bfqd); |
730 |
-+ hlist_add_head_rcu(&bfqg->group_node, &bgrp->group_data); |
731 |
-+ spin_unlock_irq(&bgrp->lock); |
732 |
-+ |
733 |
-+ return bfqg; |
734 |
-+} |
735 |
-+ |
736 |
-+#define SHOW_FUNCTION(__VAR) \ |
737 |
-+static u64 bfqio_cgroup_##__VAR##_read(struct cgroup_subsys_state *css, \ |
738 |
-+ struct cftype *cftype) \ |
739 |
-+{ \ |
740 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
741 |
-+ u64 ret = -ENODEV; \ |
742 |
-+ \ |
743 |
-+ mutex_lock(&bfqio_mutex); \ |
744 |
-+ if (bfqio_is_removed(bgrp)) \ |
745 |
-+ goto out_unlock; \ |
746 |
-+ \ |
747 |
-+ spin_lock_irq(&bgrp->lock); \ |
748 |
-+ ret = bgrp->__VAR; \ |
749 |
-+ spin_unlock_irq(&bgrp->lock); \ |
750 |
-+ \ |
751 |
-+out_unlock: \ |
752 |
-+ mutex_unlock(&bfqio_mutex); \ |
753 |
-+ return ret; \ |
754 |
-+} |
755 |
-+ |
756 |
-+SHOW_FUNCTION(weight); |
757 |
-+SHOW_FUNCTION(ioprio); |
758 |
-+SHOW_FUNCTION(ioprio_class); |
759 |
-+#undef SHOW_FUNCTION |
760 |
-+ |
761 |
-+#define STORE_FUNCTION(__VAR, __MIN, __MAX) \ |
762 |
-+static int bfqio_cgroup_##__VAR##_write(struct cgroup_subsys_state *css,\ |
763 |
-+ struct cftype *cftype, \ |
764 |
-+ u64 val) \ |
765 |
-+{ \ |
766 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
767 |
-+ struct bfq_group *bfqg; \ |
768 |
-+ int ret = -EINVAL; \ |
769 |
-+ \ |
770 |
-+ if (val < (__MIN) || val > (__MAX)) \ |
771 |
-+ return ret; \ |
772 |
-+ \ |
773 |
-+ ret = -ENODEV; \ |
774 |
-+ mutex_lock(&bfqio_mutex); \ |
775 |
-+ if (bfqio_is_removed(bgrp)) \ |
776 |
-+ goto out_unlock; \ |
777 |
-+ ret = 0; \ |
778 |
-+ \ |
779 |
-+ spin_lock_irq(&bgrp->lock); \ |
780 |
-+ bgrp->__VAR = (unsigned short)val; \ |
781 |
-+ hlist_for_each_entry(bfqg, &bgrp->group_data, group_node) { \ |
782 |
-+ /* \ |
783 |
-+ * Setting the ioprio_changed flag of the entity \ |
784 |
-+ * to 1 with new_##__VAR == ##__VAR would re-set \ |
785 |
-+ * the value of the weight to its ioprio mapping. \ |
786 |
-+ * Set the flag only if necessary. \ |
787 |
-+ */ \ |
788 |
-+ if ((unsigned short)val != bfqg->entity.new_##__VAR) { \ |
789 |
-+ bfqg->entity.new_##__VAR = (unsigned short)val; \ |
790 |
-+ /* \ |
791 |
-+ * Make sure that the above new value has been \ |
792 |
-+ * stored in bfqg->entity.new_##__VAR before \ |
793 |
-+ * setting the ioprio_changed flag. In fact, \ |
794 |
-+ * this flag may be read asynchronously (in \ |
795 |
-+ * critical sections protected by a different \ |
796 |
-+ * lock than that held here), and finding this \ |
797 |
-+ * flag set may cause the execution of the code \ |
798 |
-+ * for updating parameters whose value may \ |
799 |
-+ * depend also on bfqg->entity.new_##__VAR (in \ |
800 |
-+ * __bfq_entity_update_weight_prio). \ |
801 |
-+ * This barrier makes sure that the new value \ |
802 |
-+ * of bfqg->entity.new_##__VAR is correctly \ |
803 |
-+ * seen in that code. \ |
804 |
-+ */ \ |
805 |
-+ smp_wmb(); \ |
806 |
-+ bfqg->entity.ioprio_changed = 1; \ |
807 |
-+ } \ |
808 |
-+ } \ |
809 |
-+ spin_unlock_irq(&bgrp->lock); \ |
810 |
-+ \ |
811 |
-+out_unlock: \ |
812 |
-+ mutex_unlock(&bfqio_mutex); \ |
813 |
-+ return ret; \ |
814 |
-+} |
815 |
-+ |
816 |
-+STORE_FUNCTION(weight, BFQ_MIN_WEIGHT, BFQ_MAX_WEIGHT); |
817 |
-+STORE_FUNCTION(ioprio, 0, IOPRIO_BE_NR - 1); |
818 |
-+STORE_FUNCTION(ioprio_class, IOPRIO_CLASS_RT, IOPRIO_CLASS_IDLE); |
819 |
-+#undef STORE_FUNCTION |
820 |
-+ |
821 |
-+static struct cftype bfqio_files[] = { |
822 |
-+ { |
823 |
-+ .name = "weight", |
824 |
-+ .read_u64 = bfqio_cgroup_weight_read, |
825 |
-+ .write_u64 = bfqio_cgroup_weight_write, |
826 |
-+ }, |
827 |
-+ { |
828 |
-+ .name = "ioprio", |
829 |
-+ .read_u64 = bfqio_cgroup_ioprio_read, |
830 |
-+ .write_u64 = bfqio_cgroup_ioprio_write, |
831 |
-+ }, |
832 |
-+ { |
833 |
-+ .name = "ioprio_class", |
834 |
-+ .read_u64 = bfqio_cgroup_ioprio_class_read, |
835 |
-+ .write_u64 = bfqio_cgroup_ioprio_class_write, |
836 |
-+ }, |
837 |
-+ { }, /* terminate */ |
838 |
-+}; |
839 |
-+ |
840 |
-+static struct cgroup_subsys_state *bfqio_create(struct cgroup_subsys_state |
841 |
-+ *parent_css) |
842 |
-+{ |
843 |
-+ struct bfqio_cgroup *bgrp; |
844 |
-+ |
845 |
-+ if (parent_css != NULL) { |
846 |
-+ bgrp = kzalloc(sizeof(*bgrp), GFP_KERNEL); |
847 |
-+ if (bgrp == NULL) |
848 |
-+ return ERR_PTR(-ENOMEM); |
849 |
-+ } else |
850 |
-+ bgrp = &bfqio_root_cgroup; |
851 |
-+ |
852 |
-+ spin_lock_init(&bgrp->lock); |
853 |
-+ INIT_HLIST_HEAD(&bgrp->group_data); |
854 |
-+ bgrp->ioprio = BFQ_DEFAULT_GRP_IOPRIO; |
855 |
-+ bgrp->ioprio_class = BFQ_DEFAULT_GRP_CLASS; |
856 |
-+ |
857 |
-+ return &bgrp->css; |
858 |
-+} |
859 |
-+ |
860 |
-+/* |
861 |
-+ * We cannot support shared io contexts, as we have no means to support |
862 |
-+ * two tasks with the same ioc in two different groups without major rework |
863 |
-+ * of the main bic/bfqq data structures. By now we allow a task to change |
864 |
-+ * its cgroup only if it's the only owner of its ioc; the drawback of this |
865 |
-+ * behavior is that a group containing a task that forked using CLONE_IO |
866 |
-+ * will not be destroyed until the tasks sharing the ioc die. |
867 |
-+ */ |
868 |
-+static int bfqio_can_attach(struct cgroup_subsys_state *css, |
869 |
-+ struct cgroup_taskset *tset) |
870 |
-+{ |
871 |
-+ struct task_struct *task; |
872 |
-+ struct io_context *ioc; |
873 |
-+ int ret = 0; |
874 |
-+ |
875 |
-+ cgroup_taskset_for_each(task, css, tset) { |
876 |
-+ /* |
877 |
-+ * task_lock() is needed to avoid races with |
878 |
-+ * exit_io_context() |
879 |
-+ */ |
880 |
-+ task_lock(task); |
881 |
-+ ioc = task->io_context; |
882 |
-+ if (ioc != NULL && atomic_read(&ioc->nr_tasks) > 1) |
883 |
-+ /* |
884 |
-+ * ioc == NULL means that the task is either too young |
885 |
-+ * or exiting: if it has still no ioc the ioc can't be |
886 |
-+ * shared, if the task is exiting the attach will fail |
887 |
-+ * anyway, no matter what we return here. |
888 |
-+ */ |
889 |
-+ ret = -EINVAL; |
890 |
-+ task_unlock(task); |
891 |
-+ if (ret) |
892 |
-+ break; |
893 |
-+ } |
894 |
-+ |
895 |
-+ return ret; |
896 |
-+} |
897 |
-+ |
898 |
-+static void bfqio_attach(struct cgroup_subsys_state *css, |
899 |
-+ struct cgroup_taskset *tset) |
900 |
-+{ |
901 |
-+ struct task_struct *task; |
902 |
-+ struct io_context *ioc; |
903 |
-+ struct io_cq *icq; |
904 |
-+ |
905 |
-+ /* |
906 |
-+ * IMPORTANT NOTE: The move of more than one process at a time to a |
907 |
-+ * new group has not yet been tested. |
908 |
-+ */ |
909 |
-+ cgroup_taskset_for_each(task, css, tset) { |
910 |
-+ ioc = get_task_io_context(task, GFP_ATOMIC, NUMA_NO_NODE); |
911 |
-+ if (ioc) { |
912 |
-+ /* |
913 |
-+ * Handle cgroup change here. |
914 |
-+ */ |
915 |
-+ rcu_read_lock(); |
916 |
-+ hlist_for_each_entry_rcu(icq, &ioc->icq_list, ioc_node) |
917 |
-+ if (!strncmp( |
918 |
-+ icq->q->elevator->type->elevator_name, |
919 |
-+ "bfq", ELV_NAME_MAX)) |
920 |
-+ bfq_bic_change_cgroup(icq_to_bic(icq), |
921 |
-+ css); |
922 |
-+ rcu_read_unlock(); |
923 |
-+ put_io_context(ioc); |
924 |
-+ } |
925 |
-+ } |
926 |
-+} |
927 |
-+ |
928 |
-+static void bfqio_destroy(struct cgroup_subsys_state *css) |
929 |
-+{ |
930 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
931 |
-+ struct hlist_node *tmp; |
932 |
-+ struct bfq_group *bfqg; |
933 |
-+ |
934 |
-+ /* |
935 |
-+ * Since we are destroying the cgroup, there are no more tasks |
936 |
-+ * referencing it, and all the RCU grace periods that may have |
937 |
-+ * referenced it are ended (as the destruction of the parent |
938 |
-+ * cgroup is RCU-safe); bgrp->group_data will not be accessed by |
939 |
-+ * anything else and we don't need any synchronization. |
940 |
-+ */ |
941 |
-+ hlist_for_each_entry_safe(bfqg, tmp, &bgrp->group_data, group_node) |
942 |
-+ bfq_destroy_group(bgrp, bfqg); |
943 |
-+ |
944 |
-+ BUG_ON(!hlist_empty(&bgrp->group_data)); |
945 |
-+ |
946 |
-+ kfree(bgrp); |
947 |
-+} |
948 |
-+ |
949 |
-+static int bfqio_css_online(struct cgroup_subsys_state *css) |
950 |
-+{ |
951 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
952 |
-+ |
953 |
-+ mutex_lock(&bfqio_mutex); |
954 |
-+ bgrp->online = true; |
955 |
-+ mutex_unlock(&bfqio_mutex); |
956 |
-+ |
957 |
-+ return 0; |
958 |
-+} |
959 |
-+ |
960 |
-+static void bfqio_css_offline(struct cgroup_subsys_state *css) |
961 |
-+{ |
962 |
-+ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
963 |
-+ |
964 |
-+ mutex_lock(&bfqio_mutex); |
965 |
-+ bgrp->online = false; |
966 |
-+ mutex_unlock(&bfqio_mutex); |
967 |
-+} |
968 |
-+ |
969 |
-+struct cgroup_subsys bfqio_subsys = { |
970 |
-+ .name = "bfqio", |
971 |
-+ .css_alloc = bfqio_create, |
972 |
-+ .css_online = bfqio_css_online, |
973 |
-+ .css_offline = bfqio_css_offline, |
974 |
-+ .can_attach = bfqio_can_attach, |
975 |
-+ .attach = bfqio_attach, |
976 |
-+ .css_free = bfqio_destroy, |
977 |
-+ .subsys_id = bfqio_subsys_id, |
978 |
-+ .base_cftypes = bfqio_files, |
979 |
-+}; |
980 |
-+#else |
981 |
-+static inline void bfq_init_entity(struct bfq_entity *entity, |
982 |
-+ struct bfq_group *bfqg) |
983 |
-+{ |
984 |
-+ entity->weight = entity->new_weight; |
985 |
-+ entity->orig_weight = entity->new_weight; |
986 |
-+ entity->ioprio = entity->new_ioprio; |
987 |
-+ entity->ioprio_class = entity->new_ioprio_class; |
988 |
-+ entity->sched_data = &bfqg->sched_data; |
989 |
-+} |
990 |
-+ |
991 |
-+static inline struct bfq_group * |
992 |
-+bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
993 |
-+{ |
994 |
-+ struct bfq_data *bfqd = bic_to_bfqd(bic); |
995 |
-+ return bfqd->root_group; |
996 |
-+} |
997 |
-+ |
998 |
-+static inline void bfq_bfqq_move(struct bfq_data *bfqd, |
999 |
-+ struct bfq_queue *bfqq, |
1000 |
-+ struct bfq_entity *entity, |
1001 |
-+ struct bfq_group *bfqg) |
1002 |
-+{ |
1003 |
-+} |
1004 |
-+ |
1005 |
-+static void bfq_end_raising_async(struct bfq_data *bfqd) |
1006 |
-+{ |
1007 |
-+ bfq_end_raising_async_queues(bfqd, bfqd->root_group); |
1008 |
-+} |
1009 |
-+ |
1010 |
-+static inline void bfq_disconnect_groups(struct bfq_data *bfqd) |
1011 |
-+{ |
1012 |
-+ bfq_put_async_queues(bfqd, bfqd->root_group); |
1013 |
-+} |
1014 |
-+ |
1015 |
-+static inline void bfq_free_root_group(struct bfq_data *bfqd) |
1016 |
-+{ |
1017 |
-+ kfree(bfqd->root_group); |
1018 |
-+} |
1019 |
-+ |
1020 |
-+static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
1021 |
-+{ |
1022 |
-+ struct bfq_group *bfqg; |
1023 |
-+ int i; |
1024 |
-+ |
1025 |
-+ bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); |
1026 |
-+ if (bfqg == NULL) |
1027 |
-+ return NULL; |
1028 |
-+ |
1029 |
-+ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
1030 |
-+ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
1031 |
-+ |
1032 |
-+ return bfqg; |
1033 |
-+} |
1034 |
-+#endif |
1035 |
-diff --git a/block/bfq-ioc.c b/block/bfq-ioc.c |
1036 |
-new file mode 100644 |
1037 |
-index 0000000..7f6b000 |
1038 |
---- /dev/null |
1039 |
-+++ b/block/bfq-ioc.c |
1040 |
-@@ -0,0 +1,36 @@ |
1041 |
-+/* |
1042 |
-+ * BFQ: I/O context handling. |
1043 |
-+ * |
1044 |
-+ * Based on ideas and code from CFQ: |
1045 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
1046 |
-+ * |
1047 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
1048 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
1049 |
-+ * |
1050 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
1051 |
-+ */ |
1052 |
-+ |
1053 |
-+/** |
1054 |
-+ * icq_to_bic - convert iocontext queue structure to bfq_io_cq. |
1055 |
-+ * @icq: the iocontext queue. |
1056 |
-+ */ |
1057 |
-+static inline struct bfq_io_cq *icq_to_bic(struct io_cq *icq) |
1058 |
-+{ |
1059 |
-+ /* bic->icq is the first member, %NULL will convert to %NULL */ |
1060 |
-+ return container_of(icq, struct bfq_io_cq, icq); |
1061 |
-+} |
1062 |
-+ |
1063 |
-+/** |
1064 |
-+ * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. |
1065 |
-+ * @bfqd: the lookup key. |
1066 |
-+ * @ioc: the io_context of the process doing I/O. |
1067 |
-+ * |
1068 |
-+ * Queue lock must be held. |
1069 |
-+ */ |
1070 |
-+static inline struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd, |
1071 |
-+ struct io_context *ioc) |
1072 |
-+{ |
1073 |
-+ if (ioc) |
1074 |
-+ return icq_to_bic(ioc_lookup_icq(ioc, bfqd->queue)); |
1075 |
-+ return NULL; |
1076 |
-+} |
1077 |
-diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c |
1078 |
-new file mode 100644 |
1079 |
-index 0000000..f5f71e4 |
1080 |
---- /dev/null |
1081 |
-+++ b/block/bfq-iosched.c |
1082 |
-@@ -0,0 +1,3300 @@ |
1083 |
-+/* |
1084 |
-+ * Budget Fair Queueing (BFQ) disk scheduler. |
1085 |
-+ * |
1086 |
-+ * Based on ideas and code from CFQ: |
1087 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
1088 |
-+ * |
1089 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
1090 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
1091 |
-+ * |
1092 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
1093 |
-+ * |
1094 |
-+ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
1095 |
-+ * |
1096 |
-+ * BFQ is a proportional share disk scheduling algorithm based on the |
1097 |
-+ * slice-by-slice service scheme of CFQ. But BFQ assigns budgets, measured in |
1098 |
-+ * number of sectors, to tasks instead of time slices. The disk is not granted |
1099 |
-+ * to the in-service task for a given time slice, but until it has exhausted |
1100 |
-+ * its assigned budget. This change from the time to the service domain allows |
1101 |
-+ * BFQ to distribute the disk bandwidth among tasks as desired, without any |
1102 |
-+ * distortion due to ZBR, workload fluctuations or other factors. BFQ uses an |
1103 |
-+ * ad hoc internal scheduler, called B-WF2Q+, to schedule tasks according to |
1104 |
-+ * their budgets (more precisely BFQ schedules queues associated to tasks). |
1105 |
-+ * Thanks to this accurate scheduler, BFQ can afford to assign high budgets to |
1106 |
-+ * disk-bound non-seeky tasks (to boost the throughput), and yet guarantee low |
1107 |
-+ * latencies to interactive and soft real-time applications. |
1108 |
-+ * |
1109 |
-+ * BFQ is described in [1], where also a reference to the initial, more |
1110 |
-+ * theoretical paper on BFQ can be found. The interested reader can find in |
1111 |
-+ * the latter paper full details on the main algorithm as well as formulas of |
1112 |
-+ * the guarantees, plus formal proofs of all the properties. With respect to |
1113 |
-+ * the version of BFQ presented in these papers, this implementation adds a |
1114 |
-+ * few more heuristics, such as the one that guarantees a low latency to soft |
1115 |
-+ * real-time applications, and a hierarchical extension based on H-WF2Q+. |
1116 |
-+ * |
1117 |
-+ * B-WF2Q+ is based on WF2Q+, that is described in [2], together with |
1118 |
-+ * H-WF2Q+, while the augmented tree used to implement B-WF2Q+ with O(log N) |
1119 |
-+ * complexity derives from the one introduced with EEVDF in [3]. |
1120 |
-+ * |
1121 |
-+ * [1] P. Valente and M. Andreolini, ``Improving Application Responsiveness |
1122 |
-+ * with the BFQ Disk I/O Scheduler'', |
1123 |
-+ * Proceedings of the 5th Annual International Systems and Storage |
1124 |
-+ * Conference (SYSTOR '12), June 2012. |
1125 |
-+ * |
1126 |
-+ * http://algogroup.unimo.it/people/paolo/disk_sched/bf1-v1-suite-results.pdf |
1127 |
-+ * |
1128 |
-+ * [2] Jon C.R. Bennett and H. Zhang, ``Hierarchical Packet Fair Queueing |
1129 |
-+ * Algorithms,'' IEEE/ACM Transactions on Networking, 5(5):675-689, |
1130 |
-+ * Oct 1997. |
1131 |
-+ * |
1132 |
-+ * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz |
1133 |
-+ * |
1134 |
-+ * [3] I. Stoica and H. Abdel-Wahab, ``Earliest Eligible Virtual Deadline |
1135 |
-+ * First: A Flexible and Accurate Mechanism for Proportional Share |
1136 |
-+ * Resource Allocation,'' technical report. |
1137 |
-+ * |
1138 |
-+ * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf |
1139 |
-+ */ |
1140 |
-+#include <linux/module.h> |
1141 |
-+#include <linux/slab.h> |
1142 |
-+#include <linux/blkdev.h> |
1143 |
-+#include <linux/cgroup.h> |
1144 |
-+#include <linux/elevator.h> |
1145 |
-+#include <linux/jiffies.h> |
1146 |
-+#include <linux/rbtree.h> |
1147 |
-+#include <linux/ioprio.h> |
1148 |
-+#include "bfq.h" |
1149 |
-+#include "blk.h" |
1150 |
-+ |
1151 |
-+/* Max number of dispatches in one round of service. */ |
1152 |
-+static const int bfq_quantum = 4; |
1153 |
-+ |
1154 |
-+/* Expiration time of sync (0) and async (1) requests, in jiffies. */ |
1155 |
-+static const int bfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
1156 |
-+ |
1157 |
-+/* Maximum backwards seek, in KiB. */ |
1158 |
-+static const int bfq_back_max = 16 * 1024; |
1159 |
-+ |
1160 |
-+/* Penalty of a backwards seek, in number of sectors. */ |
1161 |
-+static const int bfq_back_penalty = 2; |
1162 |
-+ |
1163 |
-+/* Idling period duration, in jiffies. */ |
1164 |
-+static int bfq_slice_idle = HZ / 125; |
1165 |
-+ |
1166 |
-+/* Default maximum budget values, in sectors and number of requests. */ |
1167 |
-+static const int bfq_default_max_budget = 16 * 1024; |
1168 |
-+static const int bfq_max_budget_async_rq = 4; |
1169 |
-+ |
1170 |
-+/* |
1171 |
-+ * Async to sync throughput distribution is controlled as follows: |
1172 |
-+ * when an async request is served, the entity is charged the number |
1173 |
-+ * of sectors of the request, multiplied by the factor below |
1174 |
-+ */ |
1175 |
-+static const int bfq_async_charge_factor = 10; |
1176 |
-+ |
1177 |
-+/* Default timeout values, in jiffies, approximating CFQ defaults. */ |
1178 |
-+static const int bfq_timeout_sync = HZ / 8; |
1179 |
-+static int bfq_timeout_async = HZ / 25; |
1180 |
-+ |
1181 |
-+struct kmem_cache *bfq_pool; |
1182 |
-+ |
1183 |
-+/* Below this threshold (in ms), we consider thinktime immediate. */ |
1184 |
-+#define BFQ_MIN_TT 2 |
1185 |
-+ |
1186 |
-+/* hw_tag detection: parallel requests threshold and min samples needed. */ |
1187 |
-+#define BFQ_HW_QUEUE_THRESHOLD 4 |
1188 |
-+#define BFQ_HW_QUEUE_SAMPLES 32 |
1189 |
-+ |
1190 |
-+#define BFQQ_SEEK_THR (sector_t)(8 * 1024) |
1191 |
-+#define BFQQ_SEEKY(bfqq) ((bfqq)->seek_mean > BFQQ_SEEK_THR) |
1192 |
-+ |
1193 |
-+/* Min samples used for peak rate estimation (for autotuning). */ |
1194 |
-+#define BFQ_PEAK_RATE_SAMPLES 32 |
1195 |
-+ |
1196 |
-+/* Shift used for peak rate fixed precision calculations. */ |
1197 |
-+#define BFQ_RATE_SHIFT 16 |
1198 |
-+ |
1199 |
-+/* |
1200 |
-+ * The duration of the weight raising for interactive applications is |
1201 |
-+ * computed automatically (as default behaviour), using the following |
1202 |
-+ * formula: duration = (R / r) * T, where r is the peak rate of the |
1203 |
-+ * disk, and R and T are two reference parameters. In particular, R is |
1204 |
-+ * the peak rate of a reference disk, and T is about the maximum time |
1205 |
-+ * for starting popular large applications on that disk, under BFQ and |
1206 |
-+ * while reading two files in parallel. Finally, BFQ uses two |
1207 |
-+ * different pairs (R, T) depending on whether the disk is rotational |
1208 |
-+ * or non-rotational. |
1209 |
-+ */ |
1210 |
-+#define T_rot (msecs_to_jiffies(5500)) |
1211 |
-+#define T_nonrot (msecs_to_jiffies(2000)) |
1212 |
-+/* Next two quantities are in sectors/usec, left-shifted by BFQ_RATE_SHIFT */ |
1213 |
-+#define R_rot 17415 |
1214 |
-+#define R_nonrot 34791 |
1215 |
-+ |
1216 |
-+#define BFQ_SERVICE_TREE_INIT ((struct bfq_service_tree) \ |
1217 |
-+ { RB_ROOT, RB_ROOT, NULL, NULL, 0, 0 }) |
1218 |
-+ |
1219 |
-+#define RQ_BIC(rq) ((struct bfq_io_cq *) (rq)->elv.priv[0]) |
1220 |
-+#define RQ_BFQQ(rq) ((rq)->elv.priv[1]) |
1221 |
-+ |
1222 |
-+static inline void bfq_schedule_dispatch(struct bfq_data *bfqd); |
1223 |
-+ |
1224 |
-+#include "bfq-ioc.c" |
1225 |
-+#include "bfq-sched.c" |
1226 |
-+#include "bfq-cgroup.c" |
1227 |
-+ |
1228 |
-+#define bfq_class_idle(bfqq) ((bfqq)->entity.ioprio_class ==\ |
1229 |
-+ IOPRIO_CLASS_IDLE) |
1230 |
-+#define bfq_class_rt(bfqq) ((bfqq)->entity.ioprio_class ==\ |
1231 |
-+ IOPRIO_CLASS_RT) |
1232 |
-+ |
1233 |
-+#define bfq_sample_valid(samples) ((samples) > 80) |
1234 |
-+ |
1235 |
-+/* |
1236 |
-+ * We regard a request as SYNC, if either it's a read or has the SYNC bit |
1237 |
-+ * set (in which case it could also be a direct WRITE). |
1238 |
-+ */ |
1239 |
-+static inline int bfq_bio_sync(struct bio *bio) |
1240 |
-+{ |
1241 |
-+ if (bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC)) |
1242 |
-+ return 1; |
1243 |
-+ |
1244 |
-+ return 0; |
1245 |
-+} |
1246 |
-+ |
1247 |
-+/* |
1248 |
-+ * Scheduler run of queue, if there are requests pending and no one in the |
1249 |
-+ * driver that will restart queueing. |
1250 |
-+ */ |
1251 |
-+static inline void bfq_schedule_dispatch(struct bfq_data *bfqd) |
1252 |
-+{ |
1253 |
-+ if (bfqd->queued != 0) { |
1254 |
-+ bfq_log(bfqd, "schedule dispatch"); |
1255 |
-+ kblockd_schedule_work(bfqd->queue, &bfqd->unplug_work); |
1256 |
-+ } |
1257 |
-+} |
1258 |
-+ |
1259 |
-+/* |
1260 |
-+ * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
1261 |
-+ * We choose the request that is closesr to the head right now. Distance |
1262 |
-+ * behind the head is penalized and only allowed to a certain extent. |
1263 |
-+ */ |
1264 |
-+static struct request *bfq_choose_req(struct bfq_data *bfqd, |
1265 |
-+ struct request *rq1, |
1266 |
-+ struct request *rq2, |
1267 |
-+ sector_t last) |
1268 |
-+{ |
1269 |
-+ sector_t s1, s2, d1 = 0, d2 = 0; |
1270 |
-+ unsigned long back_max; |
1271 |
-+#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
1272 |
-+#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ |
1273 |
-+ unsigned wrap = 0; /* bit mask: requests behind the disk head? */ |
1274 |
-+ |
1275 |
-+ if (rq1 == NULL || rq1 == rq2) |
1276 |
-+ return rq2; |
1277 |
-+ if (rq2 == NULL) |
1278 |
-+ return rq1; |
1279 |
-+ |
1280 |
-+ if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
1281 |
-+ return rq1; |
1282 |
-+ else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) |
1283 |
-+ return rq2; |
1284 |
-+ if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) |
1285 |
-+ return rq1; |
1286 |
-+ else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) |
1287 |
-+ return rq2; |
1288 |
-+ |
1289 |
-+ s1 = blk_rq_pos(rq1); |
1290 |
-+ s2 = blk_rq_pos(rq2); |
1291 |
-+ |
1292 |
-+ /* |
1293 |
-+ * By definition, 1KiB is 2 sectors. |
1294 |
-+ */ |
1295 |
-+ back_max = bfqd->bfq_back_max * 2; |
1296 |
-+ |
1297 |
-+ /* |
1298 |
-+ * Strict one way elevator _except_ in the case where we allow |
1299 |
-+ * short backward seeks which are biased as twice the cost of a |
1300 |
-+ * similar forward seek. |
1301 |
-+ */ |
1302 |
-+ if (s1 >= last) |
1303 |
-+ d1 = s1 - last; |
1304 |
-+ else if (s1 + back_max >= last) |
1305 |
-+ d1 = (last - s1) * bfqd->bfq_back_penalty; |
1306 |
-+ else |
1307 |
-+ wrap |= BFQ_RQ1_WRAP; |
1308 |
-+ |
1309 |
-+ if (s2 >= last) |
1310 |
-+ d2 = s2 - last; |
1311 |
-+ else if (s2 + back_max >= last) |
1312 |
-+ d2 = (last - s2) * bfqd->bfq_back_penalty; |
1313 |
-+ else |
1314 |
-+ wrap |= BFQ_RQ2_WRAP; |
1315 |
-+ |
1316 |
-+ /* Found required data */ |
1317 |
-+ |
1318 |
-+ /* |
1319 |
-+ * By doing switch() on the bit mask "wrap" we avoid having to |
1320 |
-+ * check two variables for all permutations: --> faster! |
1321 |
-+ */ |
1322 |
-+ switch (wrap) { |
1323 |
-+ case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
1324 |
-+ if (d1 < d2) |
1325 |
-+ return rq1; |
1326 |
-+ else if (d2 < d1) |
1327 |
-+ return rq2; |
1328 |
-+ else { |
1329 |
-+ if (s1 >= s2) |
1330 |
-+ return rq1; |
1331 |
-+ else |
1332 |
-+ return rq2; |
1333 |
-+ } |
1334 |
-+ |
1335 |
-+ case BFQ_RQ2_WRAP: |
1336 |
-+ return rq1; |
1337 |
-+ case BFQ_RQ1_WRAP: |
1338 |
-+ return rq2; |
1339 |
-+ case (BFQ_RQ1_WRAP|BFQ_RQ2_WRAP): /* both rqs wrapped */ |
1340 |
-+ default: |
1341 |
-+ /* |
1342 |
-+ * Since both rqs are wrapped, |
1343 |
-+ * start with the one that's further behind head |
1344 |
-+ * (--> only *one* back seek required), |
1345 |
-+ * since back seek takes more time than forward. |
1346 |
-+ */ |
1347 |
-+ if (s1 <= s2) |
1348 |
-+ return rq1; |
1349 |
-+ else |
1350 |
-+ return rq2; |
1351 |
-+ } |
1352 |
-+} |
1353 |
-+ |
1354 |
-+static struct bfq_queue * |
1355 |
-+bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root, |
1356 |
-+ sector_t sector, struct rb_node **ret_parent, |
1357 |
-+ struct rb_node ***rb_link) |
1358 |
-+{ |
1359 |
-+ struct rb_node **p, *parent; |
1360 |
-+ struct bfq_queue *bfqq = NULL; |
1361 |
-+ |
1362 |
-+ parent = NULL; |
1363 |
-+ p = &root->rb_node; |
1364 |
-+ while (*p) { |
1365 |
-+ struct rb_node **n; |
1366 |
-+ |
1367 |
-+ parent = *p; |
1368 |
-+ bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
1369 |
-+ |
1370 |
-+ /* |
1371 |
-+ * Sort strictly based on sector. Smallest to the left, |
1372 |
-+ * largest to the right. |
1373 |
-+ */ |
1374 |
-+ if (sector > blk_rq_pos(bfqq->next_rq)) |
1375 |
-+ n = &(*p)->rb_right; |
1376 |
-+ else if (sector < blk_rq_pos(bfqq->next_rq)) |
1377 |
-+ n = &(*p)->rb_left; |
1378 |
-+ else |
1379 |
-+ break; |
1380 |
-+ p = n; |
1381 |
-+ bfqq = NULL; |
1382 |
-+ } |
1383 |
-+ |
1384 |
-+ *ret_parent = parent; |
1385 |
-+ if (rb_link) |
1386 |
-+ *rb_link = p; |
1387 |
-+ |
1388 |
-+ bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d", |
1389 |
-+ (long long unsigned)sector, |
1390 |
-+ bfqq != NULL ? bfqq->pid : 0); |
1391 |
-+ |
1392 |
-+ return bfqq; |
1393 |
-+} |
1394 |
-+ |
1395 |
-+static void bfq_rq_pos_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
1396 |
-+{ |
1397 |
-+ struct rb_node **p, *parent; |
1398 |
-+ struct bfq_queue *__bfqq; |
1399 |
-+ |
1400 |
-+ if (bfqq->pos_root != NULL) { |
1401 |
-+ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
1402 |
-+ bfqq->pos_root = NULL; |
1403 |
-+ } |
1404 |
-+ |
1405 |
-+ if (bfq_class_idle(bfqq)) |
1406 |
-+ return; |
1407 |
-+ if (!bfqq->next_rq) |
1408 |
-+ return; |
1409 |
-+ |
1410 |
-+ bfqq->pos_root = &bfqd->rq_pos_tree; |
1411 |
-+ __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root, |
1412 |
-+ blk_rq_pos(bfqq->next_rq), &parent, &p); |
1413 |
-+ if (__bfqq == NULL) { |
1414 |
-+ rb_link_node(&bfqq->pos_node, parent, p); |
1415 |
-+ rb_insert_color(&bfqq->pos_node, bfqq->pos_root); |
1416 |
-+ } else |
1417 |
-+ bfqq->pos_root = NULL; |
1418 |
-+} |
1419 |
-+ |
1420 |
-+static struct request *bfq_find_next_rq(struct bfq_data *bfqd, |
1421 |
-+ struct bfq_queue *bfqq, |
1422 |
-+ struct request *last) |
1423 |
-+{ |
1424 |
-+ struct rb_node *rbnext = rb_next(&last->rb_node); |
1425 |
-+ struct rb_node *rbprev = rb_prev(&last->rb_node); |
1426 |
-+ struct request *next = NULL, *prev = NULL; |
1427 |
-+ |
1428 |
-+ BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
1429 |
-+ |
1430 |
-+ if (rbprev != NULL) |
1431 |
-+ prev = rb_entry_rq(rbprev); |
1432 |
-+ |
1433 |
-+ if (rbnext != NULL) |
1434 |
-+ next = rb_entry_rq(rbnext); |
1435 |
-+ else { |
1436 |
-+ rbnext = rb_first(&bfqq->sort_list); |
1437 |
-+ if (rbnext && rbnext != &last->rb_node) |
1438 |
-+ next = rb_entry_rq(rbnext); |
1439 |
-+ } |
1440 |
-+ |
1441 |
-+ return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); |
1442 |
-+} |
1443 |
-+ |
1444 |
-+static void bfq_del_rq_rb(struct request *rq) |
1445 |
-+{ |
1446 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
1447 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
1448 |
-+ const int sync = rq_is_sync(rq); |
1449 |
-+ |
1450 |
-+ BUG_ON(bfqq->queued[sync] == 0); |
1451 |
-+ bfqq->queued[sync]--; |
1452 |
-+ bfqd->queued--; |
1453 |
-+ |
1454 |
-+ elv_rb_del(&bfqq->sort_list, rq); |
1455 |
-+ |
1456 |
-+ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
1457 |
-+ if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) |
1458 |
-+ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
1459 |
-+ /* |
1460 |
-+ * Remove queue from request-position tree as it is empty. |
1461 |
-+ */ |
1462 |
-+ if (bfqq->pos_root != NULL) { |
1463 |
-+ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
1464 |
-+ bfqq->pos_root = NULL; |
1465 |
-+ } |
1466 |
-+ } |
1467 |
-+} |
1468 |
-+ |
1469 |
-+/* see the definition of bfq_async_charge_factor for details */ |
1470 |
-+static inline unsigned long bfq_serv_to_charge(struct request *rq, |
1471 |
-+ struct bfq_queue *bfqq) |
1472 |
-+{ |
1473 |
-+ return blk_rq_sectors(rq) * |
1474 |
-+ (1 + ((!bfq_bfqq_sync(bfqq)) * (bfqq->raising_coeff == 1) * |
1475 |
-+ bfq_async_charge_factor)); |
1476 |
-+} |
1477 |
-+ |
1478 |
-+/** |
1479 |
-+ * bfq_updated_next_req - update the queue after a new next_rq selection. |
1480 |
-+ * @bfqd: the device data the queue belongs to. |
1481 |
-+ * @bfqq: the queue to update. |
1482 |
-+ * |
1483 |
-+ * If the first request of a queue changes we make sure that the queue |
1484 |
-+ * has enough budget to serve at least its first request (if the |
1485 |
-+ * request has grown). We do this because if the queue has not enough |
1486 |
-+ * budget for its first request, it has to go through two dispatch |
1487 |
-+ * rounds to actually get it dispatched. |
1488 |
-+ */ |
1489 |
-+static void bfq_updated_next_req(struct bfq_data *bfqd, |
1490 |
-+ struct bfq_queue *bfqq) |
1491 |
-+{ |
1492 |
-+ struct bfq_entity *entity = &bfqq->entity; |
1493 |
-+ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
1494 |
-+ struct request *next_rq = bfqq->next_rq; |
1495 |
-+ unsigned long new_budget; |
1496 |
-+ |
1497 |
-+ if (next_rq == NULL) |
1498 |
-+ return; |
1499 |
-+ |
1500 |
-+ if (bfqq == bfqd->in_service_queue) |
1501 |
-+ /* |
1502 |
-+ * In order not to break guarantees, budgets cannot be |
1503 |
-+ * changed after an entity has been selected. |
1504 |
-+ */ |
1505 |
-+ return; |
1506 |
-+ |
1507 |
-+ BUG_ON(entity->tree != &st->active); |
1508 |
-+ BUG_ON(entity == entity->sched_data->in_service_entity); |
1509 |
-+ |
1510 |
-+ new_budget = max_t(unsigned long, bfqq->max_budget, |
1511 |
-+ bfq_serv_to_charge(next_rq, bfqq)); |
1512 |
-+ entity->budget = new_budget; |
1513 |
-+ bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", new_budget); |
1514 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
1515 |
-+} |
1516 |
-+ |
1517 |
-+static inline unsigned int bfq_wrais_duration(struct bfq_data *bfqd) |
1518 |
-+{ |
1519 |
-+ u64 dur; |
1520 |
-+ |
1521 |
-+ if (bfqd->bfq_raising_max_time > 0) |
1522 |
-+ return bfqd->bfq_raising_max_time; |
1523 |
-+ |
1524 |
-+ dur = bfqd->RT_prod; |
1525 |
-+ do_div(dur, bfqd->peak_rate); |
1526 |
-+ |
1527 |
-+ return dur; |
1528 |
-+} |
1529 |
-+ |
1530 |
-+static void bfq_add_rq_rb(struct request *rq) |
1531 |
-+{ |
1532 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
1533 |
-+ struct bfq_entity *entity = &bfqq->entity; |
1534 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
1535 |
-+ struct request *next_rq, *prev; |
1536 |
-+ unsigned long old_raising_coeff = bfqq->raising_coeff; |
1537 |
-+ int idle_for_long_time = 0; |
1538 |
-+ |
1539 |
-+ bfq_log_bfqq(bfqd, bfqq, "add_rq_rb %d", rq_is_sync(rq)); |
1540 |
-+ bfqq->queued[rq_is_sync(rq)]++; |
1541 |
-+ bfqd->queued++; |
1542 |
-+ |
1543 |
-+ elv_rb_add(&bfqq->sort_list, rq); |
1544 |
-+ |
1545 |
-+ /* |
1546 |
-+ * Check if this request is a better next-serve candidate. |
1547 |
-+ */ |
1548 |
-+ prev = bfqq->next_rq; |
1549 |
-+ next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); |
1550 |
-+ BUG_ON(next_rq == NULL); |
1551 |
-+ bfqq->next_rq = next_rq; |
1552 |
-+ |
1553 |
-+ /* |
1554 |
-+ * Adjust priority tree position, if next_rq changes. |
1555 |
-+ */ |
1556 |
-+ if (prev != bfqq->next_rq) |
1557 |
-+ bfq_rq_pos_tree_add(bfqd, bfqq); |
1558 |
-+ |
1559 |
-+ if (!bfq_bfqq_busy(bfqq)) { |
1560 |
-+ int soft_rt = bfqd->bfq_raising_max_softrt_rate > 0 && |
1561 |
-+ time_is_before_jiffies(bfqq->soft_rt_next_start); |
1562 |
-+ idle_for_long_time = time_is_before_jiffies( |
1563 |
-+ bfqq->budget_timeout + |
1564 |
-+ bfqd->bfq_raising_min_idle_time); |
1565 |
-+ entity->budget = max_t(unsigned long, bfqq->max_budget, |
1566 |
-+ bfq_serv_to_charge(next_rq, bfqq)); |
1567 |
-+ |
1568 |
-+ if (!bfqd->low_latency) |
1569 |
-+ goto add_bfqq_busy; |
1570 |
-+ |
1571 |
-+ /* |
1572 |
-+ * If the queue is not being boosted and has been idle |
1573 |
-+ * for enough time, start a weight-raising period |
1574 |
-+ */ |
1575 |
-+ if (old_raising_coeff == 1 && |
1576 |
-+ (idle_for_long_time || soft_rt)) { |
1577 |
-+ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
1578 |
-+ if (idle_for_long_time) |
1579 |
-+ bfqq->raising_cur_max_time = |
1580 |
-+ bfq_wrais_duration(bfqd); |
1581 |
-+ else |
1582 |
-+ bfqq->raising_cur_max_time = |
1583 |
-+ bfqd->bfq_raising_rt_max_time; |
1584 |
-+ bfq_log_bfqq(bfqd, bfqq, |
1585 |
-+ "wrais starting at %lu, " |
1586 |
-+ "rais_max_time %u", |
1587 |
-+ jiffies, |
1588 |
-+ jiffies_to_msecs(bfqq-> |
1589 |
-+ raising_cur_max_time)); |
1590 |
-+ } else if (old_raising_coeff > 1) { |
1591 |
-+ if (idle_for_long_time) |
1592 |
-+ bfqq->raising_cur_max_time = |
1593 |
-+ bfq_wrais_duration(bfqd); |
1594 |
-+ else if (bfqq->raising_cur_max_time == |
1595 |
-+ bfqd->bfq_raising_rt_max_time && |
1596 |
-+ !soft_rt) { |
1597 |
-+ bfqq->raising_coeff = 1; |
1598 |
-+ bfq_log_bfqq(bfqd, bfqq, |
1599 |
-+ "wrais ending at %lu, " |
1600 |
-+ "rais_max_time %u", |
1601 |
-+ jiffies, |
1602 |
-+ jiffies_to_msecs(bfqq-> |
1603 |
-+ raising_cur_max_time)); |
1604 |
-+ } else if (time_before( |
1605 |
-+ bfqq->last_rais_start_finish + |
1606 |
-+ bfqq->raising_cur_max_time, |
1607 |
-+ jiffies + |
1608 |
-+ bfqd->bfq_raising_rt_max_time) && |
1609 |
-+ soft_rt) { |
1610 |
-+ /* |
1611 |
-+ * |
1612 |
-+ * The remaining weight-raising time is lower |
1613 |
-+ * than bfqd->bfq_raising_rt_max_time, which |
1614 |
-+ * means that the application is enjoying |
1615 |
-+ * weight raising either because deemed soft- |
1616 |
-+ * rt in the near past, or because deemed |
1617 |
-+ * interactive a long ago. In both cases, |
1618 |
-+ * resetting now the current remaining weight- |
1619 |
-+ * raising time for the application to the |
1620 |
-+ * weight-raising duration for soft rt |
1621 |
-+ * applications would not cause any latency |
1622 |
-+ * increase for the application (as the new |
1623 |
-+ * duration would be higher than the remaining |
1624 |
-+ * time). |
1625 |
-+ * |
1626 |
-+ * In addition, the application is now meeting |
1627 |
-+ * the requirements for being deemed soft rt. |
1628 |
-+ * In the end we can correctly and safely |
1629 |
-+ * (re)charge the weight-raising duration for |
1630 |
-+ * the application with the weight-raising |
1631 |
-+ * duration for soft rt applications. |
1632 |
-+ * |
1633 |
-+ * In particular, doing this recharge now, i.e., |
1634 |
-+ * before the weight-raising period for the |
1635 |
-+ * application finishes, reduces the probability |
1636 |
-+ * of the following negative scenario: |
1637 |
-+ * 1) the weight of a soft rt application is |
1638 |
-+ * raised at startup (as for any newly |
1639 |
-+ * created application), |
1640 |
-+ * 2) since the application is not interactive, |
1641 |
-+ * at a certain time weight-raising is |
1642 |
-+ * stopped for the application, |
1643 |
-+ * 3) at that time the application happens to |
1644 |
-+ * still have pending requests, and hence |
1645 |
-+ * is destined to not have a chance to be |
1646 |
-+ * deemed soft rt before these requests are |
1647 |
-+ * completed (see the comments to the |
1648 |
-+ * function bfq_bfqq_softrt_next_start() |
1649 |
-+ * for details on soft rt detection), |
1650 |
-+ * 4) these pending requests experience a high |
1651 |
-+ * latency because the application is not |
1652 |
-+ * weight-raised while they are pending. |
1653 |
-+ */ |
1654 |
-+ bfqq->last_rais_start_finish = jiffies; |
1655 |
-+ bfqq->raising_cur_max_time = |
1656 |
-+ bfqd->bfq_raising_rt_max_time; |
1657 |
-+ } |
1658 |
-+ } |
1659 |
-+ if (old_raising_coeff != bfqq->raising_coeff) |
1660 |
-+ entity->ioprio_changed = 1; |
1661 |
-+add_bfqq_busy: |
1662 |
-+ bfqq->last_idle_bklogged = jiffies; |
1663 |
-+ bfqq->service_from_backlogged = 0; |
1664 |
-+ bfq_clear_bfqq_softrt_update(bfqq); |
1665 |
-+ bfq_add_bfqq_busy(bfqd, bfqq); |
1666 |
-+ } else { |
1667 |
-+ if (bfqd->low_latency && old_raising_coeff == 1 && |
1668 |
-+ !rq_is_sync(rq) && |
1669 |
-+ time_is_before_jiffies( |
1670 |
-+ bfqq->last_rais_start_finish + |
1671 |
-+ bfqd->bfq_raising_min_inter_arr_async)) { |
1672 |
-+ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
1673 |
-+ bfqq->raising_cur_max_time = bfq_wrais_duration(bfqd); |
1674 |
-+ |
1675 |
-+ bfqd->raised_busy_queues++; |
1676 |
-+ entity->ioprio_changed = 1; |
1677 |
-+ bfq_log_bfqq(bfqd, bfqq, |
1678 |
-+ "non-idle wrais starting at %lu, " |
1679 |
-+ "rais_max_time %u", |
1680 |
-+ jiffies, |
1681 |
-+ jiffies_to_msecs(bfqq-> |
1682 |
-+ raising_cur_max_time)); |
1683 |
-+ } |
1684 |
-+ bfq_updated_next_req(bfqd, bfqq); |
1685 |
-+ } |
1686 |
-+ |
1687 |
-+ if (bfqd->low_latency && |
1688 |
-+ (old_raising_coeff == 1 || bfqq->raising_coeff == 1 || |
1689 |
-+ idle_for_long_time)) |
1690 |
-+ bfqq->last_rais_start_finish = jiffies; |
1691 |
-+} |
1692 |
-+ |
1693 |
-+static void bfq_reposition_rq_rb(struct bfq_queue *bfqq, struct request *rq) |
1694 |
-+{ |
1695 |
-+ elv_rb_del(&bfqq->sort_list, rq); |
1696 |
-+ bfqq->queued[rq_is_sync(rq)]--; |
1697 |
-+ bfqq->bfqd->queued--; |
1698 |
-+ bfq_add_rq_rb(rq); |
1699 |
-+} |
1700 |
-+ |
1701 |
-+static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, |
1702 |
-+ struct bio *bio) |
1703 |
-+{ |
1704 |
-+ struct task_struct *tsk = current; |
1705 |
-+ struct bfq_io_cq *bic; |
1706 |
-+ struct bfq_queue *bfqq; |
1707 |
-+ |
1708 |
-+ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
1709 |
-+ if (bic == NULL) |
1710 |
-+ return NULL; |
1711 |
-+ |
1712 |
-+ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
1713 |
-+ if (bfqq != NULL) |
1714 |
-+ return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio)); |
1715 |
-+ |
1716 |
-+ return NULL; |
1717 |
-+} |
1718 |
-+ |
1719 |
-+static void bfq_activate_request(struct request_queue *q, struct request *rq) |
1720 |
-+{ |
1721 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
1722 |
-+ |
1723 |
-+ bfqd->rq_in_driver++; |
1724 |
-+ bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
1725 |
-+ bfq_log(bfqd, "activate_request: new bfqd->last_position %llu", |
1726 |
-+ (long long unsigned)bfqd->last_position); |
1727 |
-+} |
1728 |
-+ |
1729 |
-+static void bfq_deactivate_request(struct request_queue *q, struct request *rq) |
1730 |
-+{ |
1731 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
1732 |
-+ |
1733 |
-+ WARN_ON(bfqd->rq_in_driver == 0); |
1734 |
-+ bfqd->rq_in_driver--; |
1735 |
-+} |
1736 |
-+ |
1737 |
-+static void bfq_remove_request(struct request *rq) |
1738 |
-+{ |
1739 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
1740 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
1741 |
-+ |
1742 |
-+ if (bfqq->next_rq == rq) { |
1743 |
-+ bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); |
1744 |
-+ bfq_updated_next_req(bfqd, bfqq); |
1745 |
-+ } |
1746 |
-+ |
1747 |
-+ list_del_init(&rq->queuelist); |
1748 |
-+ bfq_del_rq_rb(rq); |
1749 |
-+ |
1750 |
-+ if (rq->cmd_flags & REQ_META) { |
1751 |
-+ WARN_ON(bfqq->meta_pending == 0); |
1752 |
-+ bfqq->meta_pending--; |
1753 |
-+ } |
1754 |
-+} |
1755 |
-+ |
1756 |
-+static int bfq_merge(struct request_queue *q, struct request **req, |
1757 |
-+ struct bio *bio) |
1758 |
-+{ |
1759 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
1760 |
-+ struct request *__rq; |
1761 |
-+ |
1762 |
-+ __rq = bfq_find_rq_fmerge(bfqd, bio); |
1763 |
-+ if (__rq != NULL && elv_rq_merge_ok(__rq, bio)) { |
1764 |
-+ *req = __rq; |
1765 |
-+ return ELEVATOR_FRONT_MERGE; |
1766 |
-+ } |
1767 |
-+ |
1768 |
-+ return ELEVATOR_NO_MERGE; |
1769 |
-+} |
1770 |
-+ |
1771 |
-+static void bfq_merged_request(struct request_queue *q, struct request *req, |
1772 |
-+ int type) |
1773 |
-+{ |
1774 |
-+ if (type == ELEVATOR_FRONT_MERGE) { |
1775 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(req); |
1776 |
-+ |
1777 |
-+ bfq_reposition_rq_rb(bfqq, req); |
1778 |
-+ } |
1779 |
-+} |
1780 |
-+ |
1781 |
-+static void bfq_merged_requests(struct request_queue *q, struct request *rq, |
1782 |
-+ struct request *next) |
1783 |
-+{ |
1784 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
1785 |
-+ |
1786 |
-+ /* |
1787 |
-+ * Reposition in fifo if next is older than rq. |
1788 |
-+ */ |
1789 |
-+ if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && |
1790 |
-+ time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
1791 |
-+ list_move(&rq->queuelist, &next->queuelist); |
1792 |
-+ rq_set_fifo_time(rq, rq_fifo_time(next)); |
1793 |
-+ } |
1794 |
-+ |
1795 |
-+ if (bfqq->next_rq == next) |
1796 |
-+ bfqq->next_rq = rq; |
1797 |
-+ |
1798 |
-+ bfq_remove_request(next); |
1799 |
-+} |
1800 |
-+ |
1801 |
-+/* Must be called with bfqq != NULL */ |
1802 |
-+static inline void bfq_bfqq_end_raising(struct bfq_queue *bfqq) |
1803 |
-+{ |
1804 |
-+ BUG_ON(bfqq == NULL); |
1805 |
-+ if (bfq_bfqq_busy(bfqq)) |
1806 |
-+ bfqq->bfqd->raised_busy_queues--; |
1807 |
-+ bfqq->raising_coeff = 1; |
1808 |
-+ bfqq->raising_cur_max_time = 0; |
1809 |
-+ /* Trigger a weight change on the next activation of the queue */ |
1810 |
-+ bfqq->entity.ioprio_changed = 1; |
1811 |
-+} |
1812 |
-+ |
1813 |
-+static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
1814 |
-+ struct bfq_group *bfqg) |
1815 |
-+{ |
1816 |
-+ int i, j; |
1817 |
-+ |
1818 |
-+ for (i = 0; i < 2; i++) |
1819 |
-+ for (j = 0; j < IOPRIO_BE_NR; j++) |
1820 |
-+ if (bfqg->async_bfqq[i][j] != NULL) |
1821 |
-+ bfq_bfqq_end_raising(bfqg->async_bfqq[i][j]); |
1822 |
-+ if (bfqg->async_idle_bfqq != NULL) |
1823 |
-+ bfq_bfqq_end_raising(bfqg->async_idle_bfqq); |
1824 |
-+} |
1825 |
-+ |
1826 |
-+static void bfq_end_raising(struct bfq_data *bfqd) |
1827 |
-+{ |
1828 |
-+ struct bfq_queue *bfqq; |
1829 |
-+ |
1830 |
-+ spin_lock_irq(bfqd->queue->queue_lock); |
1831 |
-+ |
1832 |
-+ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) |
1833 |
-+ bfq_bfqq_end_raising(bfqq); |
1834 |
-+ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) |
1835 |
-+ bfq_bfqq_end_raising(bfqq); |
1836 |
-+ bfq_end_raising_async(bfqd); |
1837 |
-+ |
1838 |
-+ spin_unlock_irq(bfqd->queue->queue_lock); |
1839 |
-+} |
1840 |
-+ |
1841 |
-+static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
1842 |
-+ struct bio *bio) |
1843 |
-+{ |
1844 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
1845 |
-+ struct bfq_io_cq *bic; |
1846 |
-+ struct bfq_queue *bfqq; |
1847 |
-+ |
1848 |
-+ /* |
1849 |
-+ * Disallow merge of a sync bio into an async request. |
1850 |
-+ */ |
1851 |
-+ if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
1852 |
-+ return 0; |
1853 |
-+ |
1854 |
-+ /* |
1855 |
-+ * Lookup the bfqq that this bio will be queued with. Allow |
1856 |
-+ * merge only if rq is queued there. |
1857 |
-+ * Queue lock is held here. |
1858 |
-+ */ |
1859 |
-+ bic = bfq_bic_lookup(bfqd, current->io_context); |
1860 |
-+ if (bic == NULL) |
1861 |
-+ return 0; |
1862 |
-+ |
1863 |
-+ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
1864 |
-+ return bfqq == RQ_BFQQ(rq); |
1865 |
-+} |
1866 |
-+ |
1867 |
-+static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
1868 |
-+ struct bfq_queue *bfqq) |
1869 |
-+{ |
1870 |
-+ if (bfqq != NULL) { |
1871 |
-+ bfq_mark_bfqq_must_alloc(bfqq); |
1872 |
-+ bfq_mark_bfqq_budget_new(bfqq); |
1873 |
-+ bfq_clear_bfqq_fifo_expire(bfqq); |
1874 |
-+ |
1875 |
-+ bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
1876 |
-+ |
1877 |
-+ bfq_log_bfqq(bfqd, bfqq, |
1878 |
-+ "set_in_service_queue, cur-budget = %lu", |
1879 |
-+ bfqq->entity.budget); |
1880 |
-+ } |
1881 |
-+ |
1882 |
-+ bfqd->in_service_queue = bfqq; |
1883 |
-+} |
1884 |
-+ |
1885 |
-+/* |
1886 |
-+ * Get and set a new queue for service. |
1887 |
-+ */ |
1888 |
-+static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd, |
1889 |
-+ struct bfq_queue *bfqq) |
1890 |
-+{ |
1891 |
-+ if (!bfqq) |
1892 |
-+ bfqq = bfq_get_next_queue(bfqd); |
1893 |
-+ else |
1894 |
-+ bfq_get_next_queue_forced(bfqd, bfqq); |
1895 |
-+ |
1896 |
-+ __bfq_set_in_service_queue(bfqd, bfqq); |
1897 |
-+ return bfqq; |
1898 |
-+} |
1899 |
-+ |
1900 |
-+static inline sector_t bfq_dist_from_last(struct bfq_data *bfqd, |
1901 |
-+ struct request *rq) |
1902 |
-+{ |
1903 |
-+ if (blk_rq_pos(rq) >= bfqd->last_position) |
1904 |
-+ return blk_rq_pos(rq) - bfqd->last_position; |
1905 |
-+ else |
1906 |
-+ return bfqd->last_position - blk_rq_pos(rq); |
1907 |
-+} |
1908 |
-+ |
1909 |
-+/* |
1910 |
-+ * Return true if bfqq has no request pending and rq is close enough to |
1911 |
-+ * bfqd->last_position, or if rq is closer to bfqd->last_position than |
1912 |
-+ * bfqq->next_rq |
1913 |
-+ */ |
1914 |
-+static inline int bfq_rq_close(struct bfq_data *bfqd, struct request *rq) |
1915 |
-+{ |
1916 |
-+ return bfq_dist_from_last(bfqd, rq) <= BFQQ_SEEK_THR; |
1917 |
-+} |
1918 |
-+ |
1919 |
-+static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
1920 |
-+{ |
1921 |
-+ struct rb_root *root = &bfqd->rq_pos_tree; |
1922 |
-+ struct rb_node *parent, *node; |
1923 |
-+ struct bfq_queue *__bfqq; |
1924 |
-+ sector_t sector = bfqd->last_position; |
1925 |
-+ |
1926 |
-+ if (RB_EMPTY_ROOT(root)) |
1927 |
-+ return NULL; |
1928 |
-+ |
1929 |
-+ /* |
1930 |
-+ * First, if we find a request starting at the end of the last |
1931 |
-+ * request, choose it. |
1932 |
-+ */ |
1933 |
-+ __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL); |
1934 |
-+ if (__bfqq != NULL) |
1935 |
-+ return __bfqq; |
1936 |
-+ |
1937 |
-+ /* |
1938 |
-+ * If the exact sector wasn't found, the parent of the NULL leaf |
1939 |
-+ * will contain the closest sector (rq_pos_tree sorted by next_request |
1940 |
-+ * position). |
1941 |
-+ */ |
1942 |
-+ __bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
1943 |
-+ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
1944 |
-+ return __bfqq; |
1945 |
-+ |
1946 |
-+ if (blk_rq_pos(__bfqq->next_rq) < sector) |
1947 |
-+ node = rb_next(&__bfqq->pos_node); |
1948 |
-+ else |
1949 |
-+ node = rb_prev(&__bfqq->pos_node); |
1950 |
-+ if (node == NULL) |
1951 |
-+ return NULL; |
1952 |
-+ |
1953 |
-+ __bfqq = rb_entry(node, struct bfq_queue, pos_node); |
1954 |
-+ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
1955 |
-+ return __bfqq; |
1956 |
-+ |
1957 |
-+ return NULL; |
1958 |
-+} |
1959 |
-+ |
1960 |
-+/* |
1961 |
-+ * bfqd - obvious |
1962 |
-+ * cur_bfqq - passed in so that we don't decide that the current queue |
1963 |
-+ * is closely cooperating with itself. |
1964 |
-+ * |
1965 |
-+ * We are assuming that cur_bfqq has dispatched at least one request, |
1966 |
-+ * and that bfqd->last_position reflects a position on the disk associated |
1967 |
-+ * with the I/O issued by cur_bfqq. |
1968 |
-+ */ |
1969 |
-+static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
1970 |
-+ struct bfq_queue *cur_bfqq) |
1971 |
-+{ |
1972 |
-+ struct bfq_queue *bfqq; |
1973 |
-+ |
1974 |
-+ if (bfq_class_idle(cur_bfqq)) |
1975 |
-+ return NULL; |
1976 |
-+ if (!bfq_bfqq_sync(cur_bfqq)) |
1977 |
-+ return NULL; |
1978 |
-+ if (BFQQ_SEEKY(cur_bfqq)) |
1979 |
-+ return NULL; |
1980 |
-+ |
1981 |
-+ /* If device has only one backlogged bfq_queue, don't search. */ |
1982 |
-+ if (bfqd->busy_queues == 1) |
1983 |
-+ return NULL; |
1984 |
-+ |
1985 |
-+ /* |
1986 |
-+ * We should notice if some of the queues are cooperating, e.g. |
1987 |
-+ * working closely on the same area of the disk. In that case, |
1988 |
-+ * we can group them together and don't waste time idling. |
1989 |
-+ */ |
1990 |
-+ bfqq = bfqq_close(bfqd); |
1991 |
-+ if (bfqq == NULL || bfqq == cur_bfqq) |
1992 |
-+ return NULL; |
1993 |
-+ |
1994 |
-+ /* |
1995 |
-+ * Do not merge queues from different bfq_groups. |
1996 |
-+ */ |
1997 |
-+ if (bfqq->entity.parent != cur_bfqq->entity.parent) |
1998 |
-+ return NULL; |
1999 |
-+ |
2000 |
-+ /* |
2001 |
-+ * It only makes sense to merge sync queues. |
2002 |
-+ */ |
2003 |
-+ if (!bfq_bfqq_sync(bfqq)) |
2004 |
-+ return NULL; |
2005 |
-+ if (BFQQ_SEEKY(bfqq)) |
2006 |
-+ return NULL; |
2007 |
-+ |
2008 |
-+ /* |
2009 |
-+ * Do not merge queues of different priority classes. |
2010 |
-+ */ |
2011 |
-+ if (bfq_class_rt(bfqq) != bfq_class_rt(cur_bfqq)) |
2012 |
-+ return NULL; |
2013 |
-+ |
2014 |
-+ return bfqq; |
2015 |
-+} |
2016 |
-+ |
2017 |
-+/* |
2018 |
-+ * If enough samples have been computed, return the current max budget |
2019 |
-+ * stored in bfqd, which is dynamically updated according to the |
2020 |
-+ * estimated disk peak rate; otherwise return the default max budget |
2021 |
-+ */ |
2022 |
-+static inline unsigned long bfq_max_budget(struct bfq_data *bfqd) |
2023 |
-+{ |
2024 |
-+ if (bfqd->budgets_assigned < 194) |
2025 |
-+ return bfq_default_max_budget; |
2026 |
-+ else |
2027 |
-+ return bfqd->bfq_max_budget; |
2028 |
-+} |
2029 |
-+ |
2030 |
-+/* |
2031 |
-+ * Return min budget, which is a fraction of the current or default |
2032 |
-+ * max budget (trying with 1/32) |
2033 |
-+ */ |
2034 |
-+static inline unsigned long bfq_min_budget(struct bfq_data *bfqd) |
2035 |
-+{ |
2036 |
-+ if (bfqd->budgets_assigned < 194) |
2037 |
-+ return bfq_default_max_budget / 32; |
2038 |
-+ else |
2039 |
-+ return bfqd->bfq_max_budget / 32; |
2040 |
-+} |
2041 |
-+ |
2042 |
-+/* |
2043 |
-+ * Decides whether idling should be done for given device and |
2044 |
-+ * given in-service queue. |
2045 |
-+ */ |
2046 |
-+static inline bool bfq_queue_nonrot_noidle(struct bfq_data *bfqd, |
2047 |
-+ struct bfq_queue *in_service_bfqq) |
2048 |
-+{ |
2049 |
-+ if (in_service_bfqq == NULL) |
2050 |
-+ return false; |
2051 |
-+ /* |
2052 |
-+ * If the device is non-rotational, and hence has no seek penalty, |
2053 |
-+ * disable idling; but do so only if: |
2054 |
-+ * - device does not support queuing, otherwise we still have |
2055 |
-+ * a problem with sync vs async workloads; |
2056 |
-+ * - the queue is not weight-raised, to preserve guarantees. |
2057 |
-+ */ |
2058 |
-+ return blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag && |
2059 |
-+ (in_service_bfqq->raising_coeff == 1); |
2060 |
-+} |
2061 |
-+ |
2062 |
-+static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
2063 |
-+{ |
2064 |
-+ struct bfq_queue *bfqq = bfqd->in_service_queue; |
2065 |
-+ struct bfq_io_cq *bic; |
2066 |
-+ unsigned long sl; |
2067 |
-+ |
2068 |
-+ WARN_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
2069 |
-+ |
2070 |
-+ /* Tasks have exited, don't wait. */ |
2071 |
-+ bic = bfqd->in_service_bic; |
2072 |
-+ if (bic == NULL || atomic_read(&bic->icq.ioc->active_ref) == 0) |
2073 |
-+ return; |
2074 |
-+ |
2075 |
-+ bfq_mark_bfqq_wait_request(bfqq); |
2076 |
-+ |
2077 |
-+ /* |
2078 |
-+ * We don't want to idle for seeks, but we do want to allow |
2079 |
-+ * fair distribution of slice time for a process doing back-to-back |
2080 |
-+ * seeks. So allow a little bit of time for him to submit a new rq. |
2081 |
-+ * |
2082 |
-+ * To prevent processes with (partly) seeky workloads from |
2083 |
-+ * being too ill-treated, grant them a small fraction of the |
2084 |
-+ * assigned budget before reducing the waiting time to |
2085 |
-+ * BFQ_MIN_TT. This happened to help reduce latency. |
2086 |
-+ */ |
2087 |
-+ sl = bfqd->bfq_slice_idle; |
2088 |
-+ if (bfq_sample_valid(bfqq->seek_samples) && BFQQ_SEEKY(bfqq) && |
2089 |
-+ bfqq->entity.service > bfq_max_budget(bfqd) / 8 && |
2090 |
-+ bfqq->raising_coeff == 1) |
2091 |
-+ sl = min(sl, msecs_to_jiffies(BFQ_MIN_TT)); |
2092 |
-+ else if (bfqq->raising_coeff > 1) |
2093 |
-+ sl = sl * 3; |
2094 |
-+ bfqd->last_idling_start = ktime_get(); |
2095 |
-+ mod_timer(&bfqd->idle_slice_timer, jiffies + sl); |
2096 |
-+ bfq_log(bfqd, "arm idle: %u/%u ms", |
2097 |
-+ jiffies_to_msecs(sl), jiffies_to_msecs(bfqd->bfq_slice_idle)); |
2098 |
-+} |
2099 |
-+ |
2100 |
-+/* |
2101 |
-+ * Set the maximum time for the in-service queue to consume its |
2102 |
-+ * budget. This prevents seeky processes from lowering the disk |
2103 |
-+ * throughput (always guaranteed with a time slice scheme as in CFQ). |
2104 |
-+ */ |
2105 |
-+static void bfq_set_budget_timeout(struct bfq_data *bfqd) |
2106 |
-+{ |
2107 |
-+ struct bfq_queue *bfqq = bfqd->in_service_queue; |
2108 |
-+ unsigned int timeout_coeff; |
2109 |
-+ if (bfqq->raising_cur_max_time == bfqd->bfq_raising_rt_max_time) |
2110 |
-+ timeout_coeff = 1; |
2111 |
-+ else |
2112 |
-+ timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; |
2113 |
-+ |
2114 |
-+ bfqd->last_budget_start = ktime_get(); |
2115 |
-+ |
2116 |
-+ bfq_clear_bfqq_budget_new(bfqq); |
2117 |
-+ bfqq->budget_timeout = jiffies + |
2118 |
-+ bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * timeout_coeff; |
2119 |
-+ |
2120 |
-+ bfq_log_bfqq(bfqd, bfqq, "set budget_timeout %u", |
2121 |
-+ jiffies_to_msecs(bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * |
2122 |
-+ timeout_coeff)); |
2123 |
-+} |
2124 |
-+ |
2125 |
-+/* |
2126 |
-+ * Move request from internal lists to the request queue dispatch list. |
2127 |
-+ */ |
2128 |
-+static void bfq_dispatch_insert(struct request_queue *q, struct request *rq) |
2129 |
-+{ |
2130 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
2131 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
2132 |
-+ |
2133 |
-+ bfq_remove_request(rq); |
2134 |
-+ bfqq->dispatched++; |
2135 |
-+ elv_dispatch_sort(q, rq); |
2136 |
-+ |
2137 |
-+ if (bfq_bfqq_sync(bfqq)) |
2138 |
-+ bfqd->sync_flight++; |
2139 |
-+} |
2140 |
-+ |
2141 |
-+/* |
2142 |
-+ * Return expired entry, or NULL to just start from scratch in rbtree. |
2143 |
-+ */ |
2144 |
-+static struct request *bfq_check_fifo(struct bfq_queue *bfqq) |
2145 |
-+{ |
2146 |
-+ struct request *rq = NULL; |
2147 |
-+ |
2148 |
-+ if (bfq_bfqq_fifo_expire(bfqq)) |
2149 |
-+ return NULL; |
2150 |
-+ |
2151 |
-+ bfq_mark_bfqq_fifo_expire(bfqq); |
2152 |
-+ |
2153 |
-+ if (list_empty(&bfqq->fifo)) |
2154 |
-+ return NULL; |
2155 |
-+ |
2156 |
-+ rq = rq_entry_fifo(bfqq->fifo.next); |
2157 |
-+ |
2158 |
-+ if (time_before(jiffies, rq_fifo_time(rq))) |
2159 |
-+ return NULL; |
2160 |
-+ |
2161 |
-+ return rq; |
2162 |
-+} |
2163 |
-+ |
2164 |
-+/* |
2165 |
-+ * Must be called with the queue_lock held. |
2166 |
-+ */ |
2167 |
-+static int bfqq_process_refs(struct bfq_queue *bfqq) |
2168 |
-+{ |
2169 |
-+ int process_refs, io_refs; |
2170 |
-+ |
2171 |
-+ io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
2172 |
-+ process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
2173 |
-+ BUG_ON(process_refs < 0); |
2174 |
-+ return process_refs; |
2175 |
-+} |
2176 |
-+ |
2177 |
-+static void bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
2178 |
-+{ |
2179 |
-+ int process_refs, new_process_refs; |
2180 |
-+ struct bfq_queue *__bfqq; |
2181 |
-+ |
2182 |
-+ /* |
2183 |
-+ * If there are no process references on the new_bfqq, then it is |
2184 |
-+ * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
2185 |
-+ * may have dropped their last reference (not just their last process |
2186 |
-+ * reference). |
2187 |
-+ */ |
2188 |
-+ if (!bfqq_process_refs(new_bfqq)) |
2189 |
-+ return; |
2190 |
-+ |
2191 |
-+ /* Avoid a circular list and skip interim queue merges. */ |
2192 |
-+ while ((__bfqq = new_bfqq->new_bfqq)) { |
2193 |
-+ if (__bfqq == bfqq) |
2194 |
-+ return; |
2195 |
-+ new_bfqq = __bfqq; |
2196 |
-+ } |
2197 |
-+ |
2198 |
-+ process_refs = bfqq_process_refs(bfqq); |
2199 |
-+ new_process_refs = bfqq_process_refs(new_bfqq); |
2200 |
-+ /* |
2201 |
-+ * If the process for the bfqq has gone away, there is no |
2202 |
-+ * sense in merging the queues. |
2203 |
-+ */ |
2204 |
-+ if (process_refs == 0 || new_process_refs == 0) |
2205 |
-+ return; |
2206 |
-+ |
2207 |
-+ /* |
2208 |
-+ * Merge in the direction of the lesser amount of work. |
2209 |
-+ */ |
2210 |
-+ if (new_process_refs >= process_refs) { |
2211 |
-+ bfqq->new_bfqq = new_bfqq; |
2212 |
-+ atomic_add(process_refs, &new_bfqq->ref); |
2213 |
-+ } else { |
2214 |
-+ new_bfqq->new_bfqq = bfqq; |
2215 |
-+ atomic_add(new_process_refs, &bfqq->ref); |
2216 |
-+ } |
2217 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
2218 |
-+ new_bfqq->pid); |
2219 |
-+} |
2220 |
-+ |
2221 |
-+static inline unsigned long bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
2222 |
-+{ |
2223 |
-+ struct bfq_entity *entity = &bfqq->entity; |
2224 |
-+ return entity->budget - entity->service; |
2225 |
-+} |
2226 |
-+ |
2227 |
-+static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
2228 |
-+{ |
2229 |
-+ BUG_ON(bfqq != bfqd->in_service_queue); |
2230 |
-+ |
2231 |
-+ __bfq_bfqd_reset_in_service(bfqd); |
2232 |
-+ |
2233 |
-+ /* |
2234 |
-+ * If this bfqq is shared between multiple processes, check |
2235 |
-+ * to make sure that those processes are still issuing I/Os |
2236 |
-+ * within the mean seek distance. If not, it may be time to |
2237 |
-+ * break the queues apart again. |
2238 |
-+ */ |
2239 |
-+ if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq)) |
2240 |
-+ bfq_mark_bfqq_split_coop(bfqq); |
2241 |
-+ |
2242 |
-+ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
2243 |
-+ /* |
2244 |
-+ * overloading budget_timeout field to store when |
2245 |
-+ * the queue remains with no backlog, used by |
2246 |
-+ * the weight-raising mechanism |
2247 |
-+ */ |
2248 |
-+ bfqq->budget_timeout = jiffies; |
2249 |
-+ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
2250 |
-+ } else { |
2251 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
2252 |
-+ /* |
2253 |
-+ * Resort priority tree of potential close cooperators. |
2254 |
-+ */ |
2255 |
-+ bfq_rq_pos_tree_add(bfqd, bfqq); |
2256 |
-+ } |
2257 |
-+} |
2258 |
-+ |
2259 |
-+/** |
2260 |
-+ * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. |
2261 |
-+ * @bfqd: device data. |
2262 |
-+ * @bfqq: queue to update. |
2263 |
-+ * @reason: reason for expiration. |
2264 |
-+ * |
2265 |
-+ * Handle the feedback on @bfqq budget. See the body for detailed |
2266 |
-+ * comments. |
2267 |
-+ */ |
2268 |
-+static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, |
2269 |
-+ struct bfq_queue *bfqq, |
2270 |
-+ enum bfqq_expiration reason) |
2271 |
-+{ |
2272 |
-+ struct request *next_rq; |
2273 |
-+ unsigned long budget, min_budget; |
2274 |
-+ |
2275 |
-+ budget = bfqq->max_budget; |
2276 |
-+ min_budget = bfq_min_budget(bfqd); |
2277 |
-+ |
2278 |
-+ BUG_ON(bfqq != bfqd->in_service_queue); |
2279 |
-+ |
2280 |
-+ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %lu, budg left %lu", |
2281 |
-+ bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); |
2282 |
-+ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %lu, min budg %lu", |
2283 |
-+ budget, bfq_min_budget(bfqd)); |
2284 |
-+ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", |
2285 |
-+ bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue)); |
2286 |
-+ |
2287 |
-+ if (bfq_bfqq_sync(bfqq)) { |
2288 |
-+ switch (reason) { |
2289 |
-+ /* |
2290 |
-+ * Caveat: in all the following cases we trade latency |
2291 |
-+ * for throughput. |
2292 |
-+ */ |
2293 |
-+ case BFQ_BFQQ_TOO_IDLE: |
2294 |
-+ /* |
2295 |
-+ * This is the only case where we may reduce |
2296 |
-+ * the budget: if there is no request of the |
2297 |
-+ * process still waiting for completion, then |
2298 |
-+ * we assume (tentatively) that the timer has |
2299 |
-+ * expired because the batch of requests of |
2300 |
-+ * the process could have been served with a |
2301 |
-+ * smaller budget. Hence, betting that |
2302 |
-+ * process will behave in the same way when it |
2303 |
-+ * becomes backlogged again, we reduce its |
2304 |
-+ * next budget. As long as we guess right, |
2305 |
-+ * this budget cut reduces the latency |
2306 |
-+ * experienced by the process. |
2307 |
-+ * |
2308 |
-+ * However, if there are still outstanding |
2309 |
-+ * requests, then the process may have not yet |
2310 |
-+ * issued its next request just because it is |
2311 |
-+ * still waiting for the completion of some of |
2312 |
-+ * the still outstanding ones. So in this |
2313 |
-+ * subcase we do not reduce its budget, on the |
2314 |
-+ * contrary we increase it to possibly boost |
2315 |
-+ * the throughput, as discussed in the |
2316 |
-+ * comments to the BUDGET_TIMEOUT case. |
2317 |
-+ */ |
2318 |
-+ if (bfqq->dispatched > 0) /* still outstanding reqs */ |
2319 |
-+ budget = min(budget * 2, bfqd->bfq_max_budget); |
2320 |
-+ else { |
2321 |
-+ if (budget > 5 * min_budget) |
2322 |
-+ budget -= 4 * min_budget; |
2323 |
-+ else |
2324 |
-+ budget = min_budget; |
2325 |
-+ } |
2326 |
-+ break; |
2327 |
-+ case BFQ_BFQQ_BUDGET_TIMEOUT: |
2328 |
-+ /* |
2329 |
-+ * We double the budget here because: 1) it |
2330 |
-+ * gives the chance to boost the throughput if |
2331 |
-+ * this is not a seeky process (which may have |
2332 |
-+ * bumped into this timeout because of, e.g., |
2333 |
-+ * ZBR), 2) together with charge_full_budget |
2334 |
-+ * it helps give seeky processes higher |
2335 |
-+ * timestamps, and hence be served less |
2336 |
-+ * frequently. |
2337 |
-+ */ |
2338 |
-+ budget = min(budget * 2, bfqd->bfq_max_budget); |
2339 |
-+ break; |
2340 |
-+ case BFQ_BFQQ_BUDGET_EXHAUSTED: |
2341 |
-+ /* |
2342 |
-+ * The process still has backlog, and did not |
2343 |
-+ * let either the budget timeout or the disk |
2344 |
-+ * idling timeout expire. Hence it is not |
2345 |
-+ * seeky, has a short thinktime and may be |
2346 |
-+ * happy with a higher budget too. So |
2347 |
-+ * definitely increase the budget of this good |
2348 |
-+ * candidate to boost the disk throughput. |
2349 |
-+ */ |
2350 |
-+ budget = min(budget * 4, bfqd->bfq_max_budget); |
2351 |
-+ break; |
2352 |
-+ case BFQ_BFQQ_NO_MORE_REQUESTS: |
2353 |
-+ /* |
2354 |
-+ * Leave the budget unchanged. |
2355 |
-+ */ |
2356 |
-+ default: |
2357 |
-+ return; |
2358 |
-+ } |
2359 |
-+ } else /* async queue */ |
2360 |
-+ /* async queues get always the maximum possible budget |
2361 |
-+ * (their ability to dispatch is limited by |
2362 |
-+ * @bfqd->bfq_max_budget_async_rq). |
2363 |
-+ */ |
2364 |
-+ budget = bfqd->bfq_max_budget; |
2365 |
-+ |
2366 |
-+ bfqq->max_budget = budget; |
2367 |
-+ |
2368 |
-+ if (bfqd->budgets_assigned >= 194 && bfqd->bfq_user_max_budget == 0 && |
2369 |
-+ bfqq->max_budget > bfqd->bfq_max_budget) |
2370 |
-+ bfqq->max_budget = bfqd->bfq_max_budget; |
2371 |
-+ |
2372 |
-+ /* |
2373 |
-+ * Make sure that we have enough budget for the next request. |
2374 |
-+ * Since the finish time of the bfqq must be kept in sync with |
2375 |
-+ * the budget, be sure to call __bfq_bfqq_expire() after the |
2376 |
-+ * update. |
2377 |
-+ */ |
2378 |
-+ next_rq = bfqq->next_rq; |
2379 |
-+ if (next_rq != NULL) |
2380 |
-+ bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, |
2381 |
-+ bfq_serv_to_charge(next_rq, bfqq)); |
2382 |
-+ else |
2383 |
-+ bfqq->entity.budget = bfqq->max_budget; |
2384 |
-+ |
2385 |
-+ bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %lu", |
2386 |
-+ next_rq != NULL ? blk_rq_sectors(next_rq) : 0, |
2387 |
-+ bfqq->entity.budget); |
2388 |
-+} |
2389 |
-+ |
2390 |
-+static unsigned long bfq_calc_max_budget(u64 peak_rate, u64 timeout) |
2391 |
-+{ |
2392 |
-+ unsigned long max_budget; |
2393 |
-+ |
2394 |
-+ /* |
2395 |
-+ * The max_budget calculated when autotuning is equal to the |
2396 |
-+ * amount of sectors transfered in timeout_sync at the |
2397 |
-+ * estimated peak rate. |
2398 |
-+ */ |
2399 |
-+ max_budget = (unsigned long)(peak_rate * 1000 * |
2400 |
-+ timeout >> BFQ_RATE_SHIFT); |
2401 |
-+ |
2402 |
-+ return max_budget; |
2403 |
-+} |
2404 |
-+ |
2405 |
-+/* |
2406 |
-+ * In addition to updating the peak rate, checks whether the process |
2407 |
-+ * is "slow", and returns 1 if so. This slow flag is used, in addition |
2408 |
-+ * to the budget timeout, to reduce the amount of service provided to |
2409 |
-+ * seeky processes, and hence reduce their chances to lower the |
2410 |
-+ * throughput. See the code for more details. |
2411 |
-+ */ |
2412 |
-+static int bfq_update_peak_rate(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
2413 |
-+ int compensate, enum bfqq_expiration reason) |
2414 |
-+{ |
2415 |
-+ u64 bw, usecs, expected, timeout; |
2416 |
-+ ktime_t delta; |
2417 |
-+ int update = 0; |
2418 |
-+ |
2419 |
-+ if (!bfq_bfqq_sync(bfqq) || bfq_bfqq_budget_new(bfqq)) |
2420 |
-+ return 0; |
2421 |
-+ |
2422 |
-+ if (compensate) |
2423 |
-+ delta = bfqd->last_idling_start; |
2424 |
-+ else |
2425 |
-+ delta = ktime_get(); |
2426 |
-+ delta = ktime_sub(delta, bfqd->last_budget_start); |
2427 |
-+ usecs = ktime_to_us(delta); |
2428 |
-+ |
2429 |
-+ /* Don't trust short/unrealistic values. */ |
2430 |
-+ if (usecs < 100 || usecs >= LONG_MAX) |
2431 |
-+ return 0; |
2432 |
-+ |
2433 |
-+ /* |
2434 |
-+ * Calculate the bandwidth for the last slice. We use a 64 bit |
2435 |
-+ * value to store the peak rate, in sectors per usec in fixed |
2436 |
-+ * point math. We do so to have enough precision in the estimate |
2437 |
-+ * and to avoid overflows. |
2438 |
-+ */ |
2439 |
-+ bw = (u64)bfqq->entity.service << BFQ_RATE_SHIFT; |
2440 |
-+ do_div(bw, (unsigned long)usecs); |
2441 |
-+ |
2442 |
-+ timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
2443 |
-+ |
2444 |
-+ /* |
2445 |
-+ * Use only long (> 20ms) intervals to filter out spikes for |
2446 |
-+ * the peak rate estimation. |
2447 |
-+ */ |
2448 |
-+ if (usecs > 20000) { |
2449 |
-+ if (bw > bfqd->peak_rate || |
2450 |
-+ (!BFQQ_SEEKY(bfqq) && |
2451 |
-+ reason == BFQ_BFQQ_BUDGET_TIMEOUT)) { |
2452 |
-+ bfq_log(bfqd, "measured bw =%llu", bw); |
2453 |
-+ /* |
2454 |
-+ * To smooth oscillations use a low-pass filter with |
2455 |
-+ * alpha=7/8, i.e., |
2456 |
-+ * new_rate = (7/8) * old_rate + (1/8) * bw |
2457 |
-+ */ |
2458 |
-+ do_div(bw, 8); |
2459 |
-+ if (bw == 0) |
2460 |
-+ return 0; |
2461 |
-+ bfqd->peak_rate *= 7; |
2462 |
-+ do_div(bfqd->peak_rate, 8); |
2463 |
-+ bfqd->peak_rate += bw; |
2464 |
-+ update = 1; |
2465 |
-+ bfq_log(bfqd, "new peak_rate=%llu", bfqd->peak_rate); |
2466 |
-+ } |
2467 |
-+ |
2468 |
-+ update |= bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES - 1; |
2469 |
-+ |
2470 |
-+ if (bfqd->peak_rate_samples < BFQ_PEAK_RATE_SAMPLES) |
2471 |
-+ bfqd->peak_rate_samples++; |
2472 |
-+ |
2473 |
-+ if (bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES && |
2474 |
-+ update && bfqd->bfq_user_max_budget == 0) { |
2475 |
-+ bfqd->bfq_max_budget = |
2476 |
-+ bfq_calc_max_budget(bfqd->peak_rate, timeout); |
2477 |
-+ bfq_log(bfqd, "new max_budget=%lu", |
2478 |
-+ bfqd->bfq_max_budget); |
2479 |
-+ } |
2480 |
-+ } |
2481 |
-+ |
2482 |
-+ /* |
2483 |
-+ * If the process has been served for a too short time |
2484 |
-+ * interval to let its possible sequential accesses prevail on |
2485 |
-+ * the initial seek time needed to move the disk head on the |
2486 |
-+ * first sector it requested, then give the process a chance |
2487 |
-+ * and for the moment return false. |
2488 |
-+ */ |
2489 |
-+ if (bfqq->entity.budget <= bfq_max_budget(bfqd) / 8) |
2490 |
-+ return 0; |
2491 |
-+ |
2492 |
-+ /* |
2493 |
-+ * A process is considered ``slow'' (i.e., seeky, so that we |
2494 |
-+ * cannot treat it fairly in the service domain, as it would |
2495 |
-+ * slow down too much the other processes) if, when a slice |
2496 |
-+ * ends for whatever reason, it has received service at a |
2497 |
-+ * rate that would not be high enough to complete the budget |
2498 |
-+ * before the budget timeout expiration. |
2499 |
-+ */ |
2500 |
-+ expected = bw * 1000 * timeout >> BFQ_RATE_SHIFT; |
2501 |
-+ |
2502 |
-+ /* |
2503 |
-+ * Caveat: processes doing IO in the slower disk zones will |
2504 |
-+ * tend to be slow(er) even if not seeky. And the estimated |
2505 |
-+ * peak rate will actually be an average over the disk |
2506 |
-+ * surface. Hence, to not be too harsh with unlucky processes, |
2507 |
-+ * we keep a budget/3 margin of safety before declaring a |
2508 |
-+ * process slow. |
2509 |
-+ */ |
2510 |
-+ return expected > (4 * bfqq->entity.budget) / 3; |
2511 |
-+} |
2512 |
-+ |
2513 |
-+/* |
2514 |
-+ * To be deemed as soft real-time, an application must meet two requirements. |
2515 |
-+ * First, the application must not require an average bandwidth higher than |
2516 |
-+ * the approximate bandwidth required to playback or record a compressed high- |
2517 |
-+ * definition video. |
2518 |
-+ * The next function is invoked on the completion of the last request of a |
2519 |
-+ * batch, to compute the next-start time instant, soft_rt_next_start, such |
2520 |
-+ * that, if the next request of the application does not arrive before |
2521 |
-+ * soft_rt_next_start, then the above requirement on the bandwidth is met. |
2522 |
-+ * |
2523 |
-+ * The second requirement is that the request pattern of the application is |
2524 |
-+ * isochronous, i.e., that, after issuing a request or a batch of requests, |
2525 |
-+ * the application stops issuing new requests until all its pending requests |
2526 |
-+ * have been completed. After that, the application may issue a new batch, |
2527 |
-+ * and so on. |
2528 |
-+ * For this reason the next function is invoked to compute soft_rt_next_start |
2529 |
-+ * only for applications that meet this requirement, whereas soft_rt_next_start |
2530 |
-+ * is set to infinity for applications that do not. |
2531 |
-+ * |
2532 |
-+ * Unfortunately, even a greedy application may happen to behave in an |
2533 |
-+ * isochronous way if the CPU load is high. In fact, the application may stop |
2534 |
-+ * issuing requests while the CPUs are busy serving other processes, then |
2535 |
-+ * restart, then stop again for a while, and so on. In addition, if the disk |
2536 |
-+ * achieves a low enough throughput with the request pattern issued by the |
2537 |
-+ * application (e.g., because the request pattern is random and/or the device |
2538 |
-+ * is slow), then the application may meet the above bandwidth requirement too. |
2539 |
-+ * To prevent such a greedy application to be deemed as soft real-time, a |
2540 |
-+ * further rule is used in the computation of soft_rt_next_start: |
2541 |
-+ * soft_rt_next_start must be higher than the current time plus the maximum |
2542 |
-+ * time for which the arrival of a request is waited for when a sync queue |
2543 |
-+ * becomes idle, namely bfqd->bfq_slice_idle. |
2544 |
-+ * This filters out greedy applications, as the latter issue instead their next |
2545 |
-+ * request as soon as possible after the last one has been completed (in |
2546 |
-+ * contrast, when a batch of requests is completed, a soft real-time application |
2547 |
-+ * spends some time processing data). |
2548 |
-+ * |
2549 |
-+ * Unfortunately, the last filter may easily generate false positives if only |
2550 |
-+ * bfqd->bfq_slice_idle is used as a reference time interval and one or both |
2551 |
-+ * the following cases occur: |
2552 |
-+ * 1) HZ is so low that the duration of a jiffy is comparable to or higher |
2553 |
-+ * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with |
2554 |
-+ * HZ=100. |
2555 |
-+ * 2) jiffies, instead of increasing at a constant rate, may stop increasing |
2556 |
-+ * for a while, then suddenly 'jump' by several units to recover the lost |
2557 |
-+ * increments. This seems to happen, e.g., inside virtual machines. |
2558 |
-+ * To address this issue, we do not use as a reference time interval just |
2559 |
-+ * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In |
2560 |
-+ * particular we add the minimum number of jiffies for which the filter seems |
2561 |
-+ * to be quite precise also in embedded systems and KVM/QEMU virtual machines. |
2562 |
-+ */ |
2563 |
-+static inline unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd, |
2564 |
-+ struct bfq_queue *bfqq) |
2565 |
-+{ |
2566 |
-+ return max(bfqq->last_idle_bklogged + |
2567 |
-+ HZ * bfqq->service_from_backlogged / |
2568 |
-+ bfqd->bfq_raising_max_softrt_rate, |
2569 |
-+ jiffies + bfqq->bfqd->bfq_slice_idle + 4); |
2570 |
-+} |
2571 |
-+ |
2572 |
-+/* |
2573 |
-+ * Return the largest-possible time instant such that, for as long as possible, |
2574 |
-+ * the current time will be lower than this time instant according to the macro |
2575 |
-+ * time_is_before_jiffies(). |
2576 |
-+ */ |
2577 |
-+static inline unsigned long bfq_infinity_from_now(unsigned long now) |
2578 |
-+{ |
2579 |
-+ return now + ULONG_MAX / 2; |
2580 |
-+} |
2581 |
-+ |
2582 |
-+/** |
2583 |
-+ * bfq_bfqq_expire - expire a queue. |
2584 |
-+ * @bfqd: device owning the queue. |
2585 |
-+ * @bfqq: the queue to expire. |
2586 |
-+ * @compensate: if true, compensate for the time spent idling. |
2587 |
-+ * @reason: the reason causing the expiration. |
2588 |
-+ * |
2589 |
-+ * |
2590 |
-+ * If the process associated to the queue is slow (i.e., seeky), or in |
2591 |
-+ * case of budget timeout, or, finally, if it is async, we |
2592 |
-+ * artificially charge it an entire budget (independently of the |
2593 |
-+ * actual service it received). As a consequence, the queue will get |
2594 |
-+ * higher timestamps than the correct ones upon reactivation, and |
2595 |
-+ * hence it will be rescheduled as if it had received more service |
2596 |
-+ * than what it actually received. In the end, this class of processes |
2597 |
-+ * will receive less service in proportion to how slowly they consume |
2598 |
-+ * their budgets (and hence how seriously they tend to lower the |
2599 |
-+ * throughput). |
2600 |
-+ * |
2601 |
-+ * In contrast, when a queue expires because it has been idling for |
2602 |
-+ * too much or because it exhausted its budget, we do not touch the |
2603 |
-+ * amount of service it has received. Hence when the queue will be |
2604 |
-+ * reactivated and its timestamps updated, the latter will be in sync |
2605 |
-+ * with the actual service received by the queue until expiration. |
2606 |
-+ * |
2607 |
-+ * Charging a full budget to the first type of queues and the exact |
2608 |
-+ * service to the others has the effect of using the WF2Q+ policy to |
2609 |
-+ * schedule the former on a timeslice basis, without violating the |
2610 |
-+ * service domain guarantees of the latter. |
2611 |
-+ */ |
2612 |
-+static void bfq_bfqq_expire(struct bfq_data *bfqd, |
2613 |
-+ struct bfq_queue *bfqq, |
2614 |
-+ int compensate, |
2615 |
-+ enum bfqq_expiration reason) |
2616 |
-+{ |
2617 |
-+ int slow; |
2618 |
-+ BUG_ON(bfqq != bfqd->in_service_queue); |
2619 |
-+ |
2620 |
-+ /* Update disk peak rate for autotuning and check whether the |
2621 |
-+ * process is slow (see bfq_update_peak_rate). |
2622 |
-+ */ |
2623 |
-+ slow = bfq_update_peak_rate(bfqd, bfqq, compensate, reason); |
2624 |
-+ |
2625 |
-+ /* |
2626 |
-+ * As above explained, 'punish' slow (i.e., seeky), timed-out |
2627 |
-+ * and async queues, to favor sequential sync workloads. |
2628 |
-+ * |
2629 |
-+ * Processes doing IO in the slower disk zones will tend to be |
2630 |
-+ * slow(er) even if not seeky. Hence, since the estimated peak |
2631 |
-+ * rate is actually an average over the disk surface, these |
2632 |
-+ * processes may timeout just for bad luck. To avoid punishing |
2633 |
-+ * them we do not charge a full budget to a process that |
2634 |
-+ * succeeded in consuming at least 2/3 of its budget. |
2635 |
-+ */ |
2636 |
-+ if (slow || (reason == BFQ_BFQQ_BUDGET_TIMEOUT && |
2637 |
-+ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)) |
2638 |
-+ bfq_bfqq_charge_full_budget(bfqq); |
2639 |
-+ |
2640 |
-+ bfqq->service_from_backlogged += bfqq->entity.service; |
2641 |
-+ |
2642 |
-+ if (bfqd->low_latency && bfqq->raising_coeff == 1) |
2643 |
-+ bfqq->last_rais_start_finish = jiffies; |
2644 |
-+ |
2645 |
-+ if (bfqd->low_latency && bfqd->bfq_raising_max_softrt_rate > 0 && |
2646 |
-+ RB_EMPTY_ROOT(&bfqq->sort_list)) { |
2647 |
-+ /* |
2648 |
-+ * If we get here, and there are no outstanding requests, |
2649 |
-+ * then the request pattern is isochronous (see the comments |
2650 |
-+ * to the function bfq_bfqq_softrt_next_start()). Hence we can |
2651 |
-+ * compute soft_rt_next_start. If, instead, the queue still |
2652 |
-+ * has outstanding requests, then we have to wait for the |
2653 |
-+ * completion of all the outstanding requests to discover |
2654 |
-+ * whether the request pattern is actually isochronous. |
2655 |
-+ */ |
2656 |
-+ if (bfqq->dispatched == 0) |
2657 |
-+ bfqq->soft_rt_next_start = |
2658 |
-+ bfq_bfqq_softrt_next_start(bfqd, bfqq); |
2659 |
-+ else { |
2660 |
-+ /* |
2661 |
-+ * The application is still waiting for the |
2662 |
-+ * completion of one or more requests: |
2663 |
-+ * prevent it from possibly being incorrectly |
2664 |
-+ * deemed as soft real-time by setting its |
2665 |
-+ * soft_rt_next_start to infinity. In fact, |
2666 |
-+ * without this assignment, the application |
2667 |
-+ * would be incorrectly deemed as soft |
2668 |
-+ * real-time if: |
2669 |
-+ * 1) it issued a new request before the |
2670 |
-+ * completion of all its in-flight |
2671 |
-+ * requests, and |
2672 |
-+ * 2) at that time, its soft_rt_next_start |
2673 |
-+ * happened to be in the past. |
2674 |
-+ */ |
2675 |
-+ bfqq->soft_rt_next_start = |
2676 |
-+ bfq_infinity_from_now(jiffies); |
2677 |
-+ /* |
2678 |
-+ * Schedule an update of soft_rt_next_start to when |
2679 |
-+ * the task may be discovered to be isochronous. |
2680 |
-+ */ |
2681 |
-+ bfq_mark_bfqq_softrt_update(bfqq); |
2682 |
-+ } |
2683 |
-+ } |
2684 |
-+ |
2685 |
-+ bfq_log_bfqq(bfqd, bfqq, |
2686 |
-+ "expire (%d, slow %d, num_disp %d, idle_win %d)", reason, slow, |
2687 |
-+ bfqq->dispatched, bfq_bfqq_idle_window(bfqq)); |
2688 |
-+ |
2689 |
-+ /* Increase, decrease or leave budget unchanged according to reason */ |
2690 |
-+ __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); |
2691 |
-+ __bfq_bfqq_expire(bfqd, bfqq); |
2692 |
-+} |
2693 |
-+ |
2694 |
-+/* |
2695 |
-+ * Budget timeout is not implemented through a dedicated timer, but |
2696 |
-+ * just checked on request arrivals and completions, as well as on |
2697 |
-+ * idle timer expirations. |
2698 |
-+ */ |
2699 |
-+static int bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) |
2700 |
-+{ |
2701 |
-+ if (bfq_bfqq_budget_new(bfqq)) |
2702 |
-+ return 0; |
2703 |
-+ |
2704 |
-+ if (time_before(jiffies, bfqq->budget_timeout)) |
2705 |
-+ return 0; |
2706 |
-+ |
2707 |
-+ return 1; |
2708 |
-+} |
2709 |
-+ |
2710 |
-+/* |
2711 |
-+ * If we expire a queue that is waiting for the arrival of a new |
2712 |
-+ * request, we may prevent the fictitious timestamp back-shifting that |
2713 |
-+ * allows the guarantees of the queue to be preserved (see [1] for |
2714 |
-+ * this tricky aspect). Hence we return true only if this condition |
2715 |
-+ * does not hold, or if the queue is slow enough to deserve only to be |
2716 |
-+ * kicked off for preserving a high throughput. |
2717 |
-+*/ |
2718 |
-+static inline int bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) |
2719 |
-+{ |
2720 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, |
2721 |
-+ "may_budget_timeout: wr %d left %d timeout %d", |
2722 |
-+ bfq_bfqq_wait_request(bfqq), |
2723 |
-+ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, |
2724 |
-+ bfq_bfqq_budget_timeout(bfqq)); |
2725 |
-+ |
2726 |
-+ return (!bfq_bfqq_wait_request(bfqq) || |
2727 |
-+ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) |
2728 |
-+ && |
2729 |
-+ bfq_bfqq_budget_timeout(bfqq); |
2730 |
-+} |
2731 |
-+ |
2732 |
-+/* |
2733 |
-+ * For weight-raised queues issuing sync requests, idling is always performed, |
2734 |
-+ * as this is instrumental in guaranteeing a high fraction of the throughput |
2735 |
-+ * to these queues, and hence in guaranteeing a lower latency for their |
2736 |
-+ * requests. See [1] for details. |
2737 |
-+ * |
2738 |
-+ * For non-weight-raised queues, idling is instead disabled if the device is |
2739 |
-+ * NCQ-enabled and non-rotational, as this boosts the throughput on such |
2740 |
-+ * devices. |
2741 |
-+ */ |
2742 |
-+static inline bool bfq_bfqq_must_not_expire(struct bfq_queue *bfqq) |
2743 |
-+{ |
2744 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
2745 |
-+ |
2746 |
-+ return bfq_bfqq_sync(bfqq) && ( |
2747 |
-+ bfqq->raising_coeff > 1 || |
2748 |
-+ (bfq_bfqq_idle_window(bfqq) && |
2749 |
-+ !(bfqd->hw_tag && |
2750 |
-+ (blk_queue_nonrot(bfqd->queue) || |
2751 |
-+ /* |
2752 |
-+ * If there are weight-raised busy queues, then do not idle |
2753 |
-+ * the disk for a sync non-weight-raised queue, and hence |
2754 |
-+ * expire the queue immediately if empty. Combined with the |
2755 |
-+ * timestamping rules of BFQ (see [1] for details), this |
2756 |
-+ * causes sync non-weight-raised queues to get a lower |
2757 |
-+ * fraction of the disk throughput, and hence reduces the rate |
2758 |
-+ * at which the processes associated to these queues ask for |
2759 |
-+ * requests from the request pool. |
2760 |
-+ * |
2761 |
-+ * This is beneficial for weight-raised processes, when the |
2762 |
-+ * system operates in request-pool saturation conditions |
2763 |
-+ * (e.g., in the presence of write hogs). In fact, if |
2764 |
-+ * non-weight-raised processes ask for requests at a lower |
2765 |
-+ * rate, then weight-raised processes have a higher |
2766 |
-+ * probability to get a request from the pool immediately |
2767 |
-+ * (or at least soon) when they need one. Hence they have a |
2768 |
-+ * higher probability to actually get a fraction of the disk |
2769 |
-+ * throughput proportional to their high weight. This is |
2770 |
-+ * especially true with NCQ-enabled drives, which enqueue |
2771 |
-+ * several requests in advance and further reorder |
2772 |
-+ * internally-queued requests. |
2773 |
-+ * |
2774 |
-+ * Mistreating non-weight-raised queues in the above-described |
2775 |
-+ * way, when there are busy weight-raised queues, seems to |
2776 |
-+ * mitigate starvation problems in the presence of heavy write |
2777 |
-+ * workloads and NCQ, and hence to guarantee a higher |
2778 |
-+ * application and system responsiveness in these hostile |
2779 |
-+ * scenarios. |
2780 |
-+ */ |
2781 |
-+ bfqd->raised_busy_queues > 0) |
2782 |
-+ ) |
2783 |
-+ ) |
2784 |
-+ ); |
2785 |
-+} |
2786 |
-+ |
2787 |
-+/* |
2788 |
-+ * If the in-service queue is empty, but it is sync and either of the following |
2789 |
-+ * conditions holds, then: 1) the queue must remain in service and cannot be |
2790 |
-+ * expired, and 2) the disk must be idled to wait for the possible arrival |
2791 |
-+ * of a new request for the queue. The conditions are: |
2792 |
-+ * - the device is rotational and not performing NCQ, and the queue has its |
2793 |
-+ * idle window set (in this case, waiting for a new request for the queue |
2794 |
-+ * is likely to boost the disk throughput); |
2795 |
-+ * - the queue is weight-raised (waiting for the request is necessary to |
2796 |
-+ * provide the queue with fairness and latency guarantees, see [1] for |
2797 |
-+ * details). |
2798 |
-+ */ |
2799 |
-+static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) |
2800 |
-+{ |
2801 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
2802 |
-+ |
2803 |
-+ return RB_EMPTY_ROOT(&bfqq->sort_list) && bfqd->bfq_slice_idle != 0 && |
2804 |
-+ bfq_bfqq_must_not_expire(bfqq) && |
2805 |
-+ !bfq_queue_nonrot_noidle(bfqd, bfqq); |
2806 |
-+} |
2807 |
-+ |
2808 |
-+/* |
2809 |
-+ * Select a queue for service. If we have a current queue in service, |
2810 |
-+ * check whether to continue servicing it, or retrieve and set a new one. |
2811 |
-+ */ |
2812 |
-+static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
2813 |
-+{ |
2814 |
-+ struct bfq_queue *bfqq, *new_bfqq = NULL; |
2815 |
-+ struct request *next_rq; |
2816 |
-+ enum bfqq_expiration reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
2817 |
-+ |
2818 |
-+ bfqq = bfqd->in_service_queue; |
2819 |
-+ if (bfqq == NULL) |
2820 |
-+ goto new_queue; |
2821 |
-+ |
2822 |
-+ bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); |
2823 |
-+ |
2824 |
-+ /* |
2825 |
-+ * If another queue has a request waiting within our mean seek |
2826 |
-+ * distance, let it run. The expire code will check for close |
2827 |
-+ * cooperators and put the close queue at the front of the |
2828 |
-+ * service tree. If possible, merge the expiring queue with the |
2829 |
-+ * new bfqq. |
2830 |
-+ */ |
2831 |
-+ new_bfqq = bfq_close_cooperator(bfqd, bfqq); |
2832 |
-+ if (new_bfqq != NULL && bfqq->new_bfqq == NULL) |
2833 |
-+ bfq_setup_merge(bfqq, new_bfqq); |
2834 |
-+ |
2835 |
-+ if (bfq_may_expire_for_budg_timeout(bfqq) && |
2836 |
-+ !timer_pending(&bfqd->idle_slice_timer) && |
2837 |
-+ !bfq_bfqq_must_idle(bfqq)) |
2838 |
-+ goto expire; |
2839 |
-+ |
2840 |
-+ next_rq = bfqq->next_rq; |
2841 |
-+ /* |
2842 |
-+ * If bfqq has requests queued and it has enough budget left to |
2843 |
-+ * serve them, keep the queue, otherwise expire it. |
2844 |
-+ */ |
2845 |
-+ if (next_rq != NULL) { |
2846 |
-+ if (bfq_serv_to_charge(next_rq, bfqq) > |
2847 |
-+ bfq_bfqq_budget_left(bfqq)) { |
2848 |
-+ reason = BFQ_BFQQ_BUDGET_EXHAUSTED; |
2849 |
-+ goto expire; |
2850 |
-+ } else { |
2851 |
-+ /* |
2852 |
-+ * The idle timer may be pending because we may not |
2853 |
-+ * disable disk idling even when a new request arrives |
2854 |
-+ */ |
2855 |
-+ if (timer_pending(&bfqd->idle_slice_timer)) { |
2856 |
-+ /* |
2857 |
-+ * If we get here: 1) at least a new request |
2858 |
-+ * has arrived but we have not disabled the |
2859 |
-+ * timer because the request was too small, |
2860 |
-+ * 2) then the block layer has unplugged the |
2861 |
-+ * device, causing the dispatch to be invoked. |
2862 |
-+ * |
2863 |
-+ * Since the device is unplugged, now the |
2864 |
-+ * requests are probably large enough to |
2865 |
-+ * provide a reasonable throughput. |
2866 |
-+ * So we disable idling. |
2867 |
-+ */ |
2868 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
2869 |
-+ del_timer(&bfqd->idle_slice_timer); |
2870 |
-+ } |
2871 |
-+ if (new_bfqq == NULL) |
2872 |
-+ goto keep_queue; |
2873 |
-+ else |
2874 |
-+ goto expire; |
2875 |
-+ } |
2876 |
-+ } |
2877 |
-+ |
2878 |
-+ /* |
2879 |
-+ * No requests pending. If the in-service queue has no cooperator and |
2880 |
-+ * still has requests in flight (possibly waiting for a completion) |
2881 |
-+ * or is idling for a new request, then keep it. |
2882 |
-+ */ |
2883 |
-+ if (new_bfqq == NULL && (timer_pending(&bfqd->idle_slice_timer) || |
2884 |
-+ (bfqq->dispatched != 0 && bfq_bfqq_must_not_expire(bfqq)))) { |
2885 |
-+ bfqq = NULL; |
2886 |
-+ goto keep_queue; |
2887 |
-+ } else if (new_bfqq != NULL && timer_pending(&bfqd->idle_slice_timer)) { |
2888 |
-+ /* |
2889 |
-+ * Expiring the queue because there is a close cooperator, |
2890 |
-+ * cancel timer. |
2891 |
-+ */ |
2892 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
2893 |
-+ del_timer(&bfqd->idle_slice_timer); |
2894 |
-+ } |
2895 |
-+ |
2896 |
-+ reason = BFQ_BFQQ_NO_MORE_REQUESTS; |
2897 |
-+expire: |
2898 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, reason); |
2899 |
-+new_queue: |
2900 |
-+ bfqq = bfq_set_in_service_queue(bfqd, new_bfqq); |
2901 |
-+ bfq_log(bfqd, "select_queue: new queue %d returned", |
2902 |
-+ bfqq != NULL ? bfqq->pid : 0); |
2903 |
-+keep_queue: |
2904 |
-+ return bfqq; |
2905 |
-+} |
2906 |
-+ |
2907 |
-+static void bfq_update_raising_data(struct bfq_data *bfqd, |
2908 |
-+ struct bfq_queue *bfqq) |
2909 |
-+{ |
2910 |
-+ if (bfqq->raising_coeff > 1) { /* queue is being boosted */ |
2911 |
-+ struct bfq_entity *entity = &bfqq->entity; |
2912 |
-+ |
2913 |
-+ bfq_log_bfqq(bfqd, bfqq, |
2914 |
-+ "raising period dur %u/%u msec, " |
2915 |
-+ "old raising coeff %u, w %d(%d)", |
2916 |
-+ jiffies_to_msecs(jiffies - |
2917 |
-+ bfqq->last_rais_start_finish), |
2918 |
-+ jiffies_to_msecs(bfqq->raising_cur_max_time), |
2919 |
-+ bfqq->raising_coeff, |
2920 |
-+ bfqq->entity.weight, bfqq->entity.orig_weight); |
2921 |
-+ |
2922 |
-+ BUG_ON(bfqq != bfqd->in_service_queue && entity->weight != |
2923 |
-+ entity->orig_weight * bfqq->raising_coeff); |
2924 |
-+ if (entity->ioprio_changed) |
2925 |
-+ bfq_log_bfqq(bfqd, bfqq, |
2926 |
-+ "WARN: pending prio change"); |
2927 |
-+ /* |
2928 |
-+ * If too much time has elapsed from the beginning |
2929 |
-+ * of this weight-raising, stop it. |
2930 |
-+ */ |
2931 |
-+ if (time_is_before_jiffies(bfqq->last_rais_start_finish + |
2932 |
-+ bfqq->raising_cur_max_time)) { |
2933 |
-+ bfqq->last_rais_start_finish = jiffies; |
2934 |
-+ bfq_log_bfqq(bfqd, bfqq, |
2935 |
-+ "wrais ending at %lu, " |
2936 |
-+ "rais_max_time %u", |
2937 |
-+ bfqq->last_rais_start_finish, |
2938 |
-+ jiffies_to_msecs(bfqq-> |
2939 |
-+ raising_cur_max_time)); |
2940 |
-+ bfq_bfqq_end_raising(bfqq); |
2941 |
-+ __bfq_entity_update_weight_prio( |
2942 |
-+ bfq_entity_service_tree(entity), |
2943 |
-+ entity); |
2944 |
-+ } |
2945 |
-+ } |
2946 |
-+} |
2947 |
-+ |
2948 |
-+/* |
2949 |
-+ * Dispatch one request from bfqq, moving it to the request queue |
2950 |
-+ * dispatch list. |
2951 |
-+ */ |
2952 |
-+static int bfq_dispatch_request(struct bfq_data *bfqd, |
2953 |
-+ struct bfq_queue *bfqq) |
2954 |
-+{ |
2955 |
-+ int dispatched = 0; |
2956 |
-+ struct request *rq; |
2957 |
-+ unsigned long service_to_charge; |
2958 |
-+ |
2959 |
-+ BUG_ON(RB_EMPTY_ROOT(&bfqq->sort_list)); |
2960 |
-+ |
2961 |
-+ /* Follow expired path, else get first next available. */ |
2962 |
-+ rq = bfq_check_fifo(bfqq); |
2963 |
-+ if (rq == NULL) |
2964 |
-+ rq = bfqq->next_rq; |
2965 |
-+ service_to_charge = bfq_serv_to_charge(rq, bfqq); |
2966 |
-+ |
2967 |
-+ if (service_to_charge > bfq_bfqq_budget_left(bfqq)) { |
2968 |
-+ /* |
2969 |
-+ * This may happen if the next rq is chosen |
2970 |
-+ * in fifo order instead of sector order. |
2971 |
-+ * The budget is properly dimensioned |
2972 |
-+ * to be always sufficient to serve the next request |
2973 |
-+ * only if it is chosen in sector order. The reason is |
2974 |
-+ * that it would be quite inefficient and little useful |
2975 |
-+ * to always make sure that the budget is large enough |
2976 |
-+ * to serve even the possible next rq in fifo order. |
2977 |
-+ * In fact, requests are seldom served in fifo order. |
2978 |
-+ * |
2979 |
-+ * Expire the queue for budget exhaustion, and |
2980 |
-+ * make sure that the next act_budget is enough |
2981 |
-+ * to serve the next request, even if it comes |
2982 |
-+ * from the fifo expired path. |
2983 |
-+ */ |
2984 |
-+ bfqq->next_rq = rq; |
2985 |
-+ /* |
2986 |
-+ * Since this dispatch is failed, make sure that |
2987 |
-+ * a new one will be performed |
2988 |
-+ */ |
2989 |
-+ if (!bfqd->rq_in_driver) |
2990 |
-+ bfq_schedule_dispatch(bfqd); |
2991 |
-+ goto expire; |
2992 |
-+ } |
2993 |
-+ |
2994 |
-+ /* Finally, insert request into driver dispatch list. */ |
2995 |
-+ bfq_bfqq_served(bfqq, service_to_charge); |
2996 |
-+ bfq_dispatch_insert(bfqd->queue, rq); |
2997 |
-+ |
2998 |
-+ bfq_update_raising_data(bfqd, bfqq); |
2999 |
-+ |
3000 |
-+ bfq_log_bfqq(bfqd, bfqq, |
3001 |
-+ "dispatched %u sec req (%llu), budg left %lu", |
3002 |
-+ blk_rq_sectors(rq), |
3003 |
-+ (long long unsigned)blk_rq_pos(rq), |
3004 |
-+ bfq_bfqq_budget_left(bfqq)); |
3005 |
-+ |
3006 |
-+ dispatched++; |
3007 |
-+ |
3008 |
-+ if (bfqd->in_service_bic == NULL) { |
3009 |
-+ atomic_long_inc(&RQ_BIC(rq)->icq.ioc->refcount); |
3010 |
-+ bfqd->in_service_bic = RQ_BIC(rq); |
3011 |
-+ } |
3012 |
-+ |
3013 |
-+ if (bfqd->busy_queues > 1 && ((!bfq_bfqq_sync(bfqq) && |
3014 |
-+ dispatched >= bfqd->bfq_max_budget_async_rq) || |
3015 |
-+ bfq_class_idle(bfqq))) |
3016 |
-+ goto expire; |
3017 |
-+ |
3018 |
-+ return dispatched; |
3019 |
-+ |
3020 |
-+expire: |
3021 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_EXHAUSTED); |
3022 |
-+ return dispatched; |
3023 |
-+} |
3024 |
-+ |
3025 |
-+static int __bfq_forced_dispatch_bfqq(struct bfq_queue *bfqq) |
3026 |
-+{ |
3027 |
-+ int dispatched = 0; |
3028 |
-+ |
3029 |
-+ while (bfqq->next_rq != NULL) { |
3030 |
-+ bfq_dispatch_insert(bfqq->bfqd->queue, bfqq->next_rq); |
3031 |
-+ dispatched++; |
3032 |
-+ } |
3033 |
-+ |
3034 |
-+ BUG_ON(!list_empty(&bfqq->fifo)); |
3035 |
-+ return dispatched; |
3036 |
-+} |
3037 |
-+ |
3038 |
-+/* |
3039 |
-+ * Drain our current requests. Used for barriers and when switching |
3040 |
-+ * io schedulers on-the-fly. |
3041 |
-+ */ |
3042 |
-+static int bfq_forced_dispatch(struct bfq_data *bfqd) |
3043 |
-+{ |
3044 |
-+ struct bfq_queue *bfqq, *n; |
3045 |
-+ struct bfq_service_tree *st; |
3046 |
-+ int dispatched = 0; |
3047 |
-+ |
3048 |
-+ bfqq = bfqd->in_service_queue; |
3049 |
-+ if (bfqq != NULL) |
3050 |
-+ __bfq_bfqq_expire(bfqd, bfqq); |
3051 |
-+ |
3052 |
-+ /* |
3053 |
-+ * Loop through classes, and be careful to leave the scheduler |
3054 |
-+ * in a consistent state, as feedback mechanisms and vtime |
3055 |
-+ * updates cannot be disabled during the process. |
3056 |
-+ */ |
3057 |
-+ list_for_each_entry_safe(bfqq, n, &bfqd->active_list, bfqq_list) { |
3058 |
-+ st = bfq_entity_service_tree(&bfqq->entity); |
3059 |
-+ |
3060 |
-+ dispatched += __bfq_forced_dispatch_bfqq(bfqq); |
3061 |
-+ bfqq->max_budget = bfq_max_budget(bfqd); |
3062 |
-+ |
3063 |
-+ bfq_forget_idle(st); |
3064 |
-+ } |
3065 |
-+ |
3066 |
-+ BUG_ON(bfqd->busy_queues != 0); |
3067 |
-+ |
3068 |
-+ return dispatched; |
3069 |
-+} |
3070 |
-+ |
3071 |
-+static int bfq_dispatch_requests(struct request_queue *q, int force) |
3072 |
-+{ |
3073 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
3074 |
-+ struct bfq_queue *bfqq; |
3075 |
-+ int max_dispatch; |
3076 |
-+ |
3077 |
-+ bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues); |
3078 |
-+ if (bfqd->busy_queues == 0) |
3079 |
-+ return 0; |
3080 |
-+ |
3081 |
-+ if (unlikely(force)) |
3082 |
-+ return bfq_forced_dispatch(bfqd); |
3083 |
-+ |
3084 |
-+ bfqq = bfq_select_queue(bfqd); |
3085 |
-+ if (bfqq == NULL) |
3086 |
-+ return 0; |
3087 |
-+ |
3088 |
-+ max_dispatch = bfqd->bfq_quantum; |
3089 |
-+ if (bfq_class_idle(bfqq)) |
3090 |
-+ max_dispatch = 1; |
3091 |
-+ |
3092 |
-+ if (!bfq_bfqq_sync(bfqq)) |
3093 |
-+ max_dispatch = bfqd->bfq_max_budget_async_rq; |
3094 |
-+ |
3095 |
-+ if (bfqq->dispatched >= max_dispatch) { |
3096 |
-+ if (bfqd->busy_queues > 1) |
3097 |
-+ return 0; |
3098 |
-+ if (bfqq->dispatched >= 4 * max_dispatch) |
3099 |
-+ return 0; |
3100 |
-+ } |
3101 |
-+ |
3102 |
-+ if (bfqd->sync_flight != 0 && !bfq_bfqq_sync(bfqq)) |
3103 |
-+ return 0; |
3104 |
-+ |
3105 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
3106 |
-+ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
3107 |
-+ |
3108 |
-+ if (!bfq_dispatch_request(bfqd, bfqq)) |
3109 |
-+ return 0; |
3110 |
-+ |
3111 |
-+ bfq_log_bfqq(bfqd, bfqq, "dispatched one request of %d (max_disp %d)", |
3112 |
-+ bfqq->pid, max_dispatch); |
3113 |
-+ |
3114 |
-+ return 1; |
3115 |
-+} |
3116 |
-+ |
3117 |
-+/* |
3118 |
-+ * Task holds one reference to the queue, dropped when task exits. Each rq |
3119 |
-+ * in-flight on this queue also holds a reference, dropped when rq is freed. |
3120 |
-+ * |
3121 |
-+ * Queue lock must be held here. |
3122 |
-+ */ |
3123 |
-+static void bfq_put_queue(struct bfq_queue *bfqq) |
3124 |
-+{ |
3125 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
3126 |
-+ |
3127 |
-+ BUG_ON(atomic_read(&bfqq->ref) <= 0); |
3128 |
-+ |
3129 |
-+ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p %d", bfqq, |
3130 |
-+ atomic_read(&bfqq->ref)); |
3131 |
-+ if (!atomic_dec_and_test(&bfqq->ref)) |
3132 |
-+ return; |
3133 |
-+ |
3134 |
-+ BUG_ON(rb_first(&bfqq->sort_list) != NULL); |
3135 |
-+ BUG_ON(bfqq->allocated[READ] + bfqq->allocated[WRITE] != 0); |
3136 |
-+ BUG_ON(bfqq->entity.tree != NULL); |
3137 |
-+ BUG_ON(bfq_bfqq_busy(bfqq)); |
3138 |
-+ BUG_ON(bfqd->in_service_queue == bfqq); |
3139 |
-+ |
3140 |
-+ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p freed", bfqq); |
3141 |
-+ |
3142 |
-+ kmem_cache_free(bfq_pool, bfqq); |
3143 |
-+} |
3144 |
-+ |
3145 |
-+static void bfq_put_cooperator(struct bfq_queue *bfqq) |
3146 |
-+{ |
3147 |
-+ struct bfq_queue *__bfqq, *next; |
3148 |
-+ |
3149 |
-+ /* |
3150 |
-+ * If this queue was scheduled to merge with another queue, be |
3151 |
-+ * sure to drop the reference taken on that queue (and others in |
3152 |
-+ * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs. |
3153 |
-+ */ |
3154 |
-+ __bfqq = bfqq->new_bfqq; |
3155 |
-+ while (__bfqq) { |
3156 |
-+ if (__bfqq == bfqq) { |
3157 |
-+ WARN(1, "bfqq->new_bfqq loop detected.\n"); |
3158 |
-+ break; |
3159 |
-+ } |
3160 |
-+ next = __bfqq->new_bfqq; |
3161 |
-+ bfq_put_queue(__bfqq); |
3162 |
-+ __bfqq = next; |
3163 |
-+ } |
3164 |
-+} |
3165 |
-+ |
3166 |
-+static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
3167 |
-+{ |
3168 |
-+ if (bfqq == bfqd->in_service_queue) { |
3169 |
-+ __bfq_bfqq_expire(bfqd, bfqq); |
3170 |
-+ bfq_schedule_dispatch(bfqd); |
3171 |
-+ } |
3172 |
-+ |
3173 |
-+ bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, |
3174 |
-+ atomic_read(&bfqq->ref)); |
3175 |
-+ |
3176 |
-+ bfq_put_cooperator(bfqq); |
3177 |
-+ |
3178 |
-+ bfq_put_queue(bfqq); |
3179 |
-+} |
3180 |
-+ |
3181 |
-+static void bfq_init_icq(struct io_cq *icq) |
3182 |
-+{ |
3183 |
-+ struct bfq_io_cq *bic = icq_to_bic(icq); |
3184 |
-+ |
3185 |
-+ bic->ttime.last_end_request = jiffies; |
3186 |
-+} |
3187 |
-+ |
3188 |
-+static void bfq_exit_icq(struct io_cq *icq) |
3189 |
-+{ |
3190 |
-+ struct bfq_io_cq *bic = icq_to_bic(icq); |
3191 |
-+ struct bfq_data *bfqd = bic_to_bfqd(bic); |
3192 |
-+ |
3193 |
-+ if (bic->bfqq[BLK_RW_ASYNC]) { |
3194 |
-+ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_ASYNC]); |
3195 |
-+ bic->bfqq[BLK_RW_ASYNC] = NULL; |
3196 |
-+ } |
3197 |
-+ |
3198 |
-+ if (bic->bfqq[BLK_RW_SYNC]) { |
3199 |
-+ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_SYNC]); |
3200 |
-+ bic->bfqq[BLK_RW_SYNC] = NULL; |
3201 |
-+ } |
3202 |
-+} |
3203 |
-+ |
3204 |
-+/* |
3205 |
-+ * Update the entity prio values; note that the new values will not |
3206 |
-+ * be used until the next (re)activation. |
3207 |
-+ */ |
3208 |
-+static void bfq_init_prio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) |
3209 |
-+{ |
3210 |
-+ struct task_struct *tsk = current; |
3211 |
-+ int ioprio_class; |
3212 |
-+ |
3213 |
-+ if (!bfq_bfqq_prio_changed(bfqq)) |
3214 |
-+ return; |
3215 |
-+ |
3216 |
-+ ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
3217 |
-+ switch (ioprio_class) { |
3218 |
-+ default: |
3219 |
-+ dev_err(bfqq->bfqd->queue->backing_dev_info.dev, |
3220 |
-+ "bfq: bad prio %x\n", ioprio_class); |
3221 |
-+ case IOPRIO_CLASS_NONE: |
3222 |
-+ /* |
3223 |
-+ * No prio set, inherit CPU scheduling settings. |
3224 |
-+ */ |
3225 |
-+ bfqq->entity.new_ioprio = task_nice_ioprio(tsk); |
3226 |
-+ bfqq->entity.new_ioprio_class = task_nice_ioclass(tsk); |
3227 |
-+ break; |
3228 |
-+ case IOPRIO_CLASS_RT: |
3229 |
-+ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
3230 |
-+ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_RT; |
3231 |
-+ break; |
3232 |
-+ case IOPRIO_CLASS_BE: |
3233 |
-+ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
3234 |
-+ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_BE; |
3235 |
-+ break; |
3236 |
-+ case IOPRIO_CLASS_IDLE: |
3237 |
-+ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_IDLE; |
3238 |
-+ bfqq->entity.new_ioprio = 7; |
3239 |
-+ bfq_clear_bfqq_idle_window(bfqq); |
3240 |
-+ break; |
3241 |
-+ } |
3242 |
-+ |
3243 |
-+ bfqq->entity.ioprio_changed = 1; |
3244 |
-+ |
3245 |
-+ /* |
3246 |
-+ * Keep track of original prio settings in case we have to temporarily |
3247 |
-+ * elevate the priority of this queue. |
3248 |
-+ */ |
3249 |
-+ bfqq->org_ioprio = bfqq->entity.new_ioprio; |
3250 |
-+ bfq_clear_bfqq_prio_changed(bfqq); |
3251 |
-+} |
3252 |
-+ |
3253 |
-+static void bfq_changed_ioprio(struct bfq_io_cq *bic) |
3254 |
-+{ |
3255 |
-+ struct bfq_data *bfqd; |
3256 |
-+ struct bfq_queue *bfqq, *new_bfqq; |
3257 |
-+ struct bfq_group *bfqg; |
3258 |
-+ unsigned long uninitialized_var(flags); |
3259 |
-+ int ioprio = bic->icq.ioc->ioprio; |
3260 |
-+ |
3261 |
-+ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), |
3262 |
-+ &flags); |
3263 |
-+ /* |
3264 |
-+ * This condition may trigger on a newly created bic, be sure to drop |
3265 |
-+ * the lock before returning. |
3266 |
-+ */ |
3267 |
-+ if (unlikely(bfqd == NULL) || likely(bic->ioprio == ioprio)) |
3268 |
-+ goto out; |
3269 |
-+ |
3270 |
-+ bfqq = bic->bfqq[BLK_RW_ASYNC]; |
3271 |
-+ if (bfqq != NULL) { |
3272 |
-+ bfqg = container_of(bfqq->entity.sched_data, struct bfq_group, |
3273 |
-+ sched_data); |
3274 |
-+ new_bfqq = bfq_get_queue(bfqd, bfqg, BLK_RW_ASYNC, bic, |
3275 |
-+ GFP_ATOMIC); |
3276 |
-+ if (new_bfqq != NULL) { |
3277 |
-+ bic->bfqq[BLK_RW_ASYNC] = new_bfqq; |
3278 |
-+ bfq_log_bfqq(bfqd, bfqq, |
3279 |
-+ "changed_ioprio: bfqq %p %d", |
3280 |
-+ bfqq, atomic_read(&bfqq->ref)); |
3281 |
-+ bfq_put_queue(bfqq); |
3282 |
-+ } |
3283 |
-+ } |
3284 |
-+ |
3285 |
-+ bfqq = bic->bfqq[BLK_RW_SYNC]; |
3286 |
-+ if (bfqq != NULL) |
3287 |
-+ bfq_mark_bfqq_prio_changed(bfqq); |
3288 |
-+ |
3289 |
-+ bic->ioprio = ioprio; |
3290 |
-+ |
3291 |
-+out: |
3292 |
-+ bfq_put_bfqd_unlock(bfqd, &flags); |
3293 |
-+} |
3294 |
-+ |
3295 |
-+static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
3296 |
-+ pid_t pid, int is_sync) |
3297 |
-+{ |
3298 |
-+ RB_CLEAR_NODE(&bfqq->entity.rb_node); |
3299 |
-+ INIT_LIST_HEAD(&bfqq->fifo); |
3300 |
-+ |
3301 |
-+ atomic_set(&bfqq->ref, 0); |
3302 |
-+ bfqq->bfqd = bfqd; |
3303 |
-+ |
3304 |
-+ bfq_mark_bfqq_prio_changed(bfqq); |
3305 |
-+ |
3306 |
-+ if (is_sync) { |
3307 |
-+ if (!bfq_class_idle(bfqq)) |
3308 |
-+ bfq_mark_bfqq_idle_window(bfqq); |
3309 |
-+ bfq_mark_bfqq_sync(bfqq); |
3310 |
-+ } |
3311 |
-+ |
3312 |
-+ /* Tentative initial value to trade off between thr and lat */ |
3313 |
-+ bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
3314 |
-+ bfqq->pid = pid; |
3315 |
-+ |
3316 |
-+ bfqq->raising_coeff = 1; |
3317 |
-+ bfqq->last_rais_start_finish = 0; |
3318 |
-+ /* |
3319 |
-+ * Set to the value for which bfqq will not be deemed as |
3320 |
-+ * soft rt when it becomes backlogged. |
3321 |
-+ */ |
3322 |
-+ bfqq->soft_rt_next_start = bfq_infinity_from_now(jiffies); |
3323 |
-+} |
3324 |
-+ |
3325 |
-+static struct bfq_queue *bfq_find_alloc_queue(struct bfq_data *bfqd, |
3326 |
-+ struct bfq_group *bfqg, |
3327 |
-+ int is_sync, |
3328 |
-+ struct bfq_io_cq *bic, |
3329 |
-+ gfp_t gfp_mask) |
3330 |
-+{ |
3331 |
-+ struct bfq_queue *bfqq, *new_bfqq = NULL; |
3332 |
-+ |
3333 |
-+retry: |
3334 |
-+ /* bic always exists here */ |
3335 |
-+ bfqq = bic_to_bfqq(bic, is_sync); |
3336 |
-+ |
3337 |
-+ /* |
3338 |
-+ * Always try a new alloc if we fall back to the OOM bfqq |
3339 |
-+ * originally, since it should just be a temporary situation. |
3340 |
-+ */ |
3341 |
-+ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
3342 |
-+ bfqq = NULL; |
3343 |
-+ if (new_bfqq != NULL) { |
3344 |
-+ bfqq = new_bfqq; |
3345 |
-+ new_bfqq = NULL; |
3346 |
-+ } else if (gfp_mask & __GFP_WAIT) { |
3347 |
-+ spin_unlock_irq(bfqd->queue->queue_lock); |
3348 |
-+ new_bfqq = kmem_cache_alloc_node(bfq_pool, |
3349 |
-+ gfp_mask | __GFP_ZERO, |
3350 |
-+ bfqd->queue->node); |
3351 |
-+ spin_lock_irq(bfqd->queue->queue_lock); |
3352 |
-+ if (new_bfqq != NULL) |
3353 |
-+ goto retry; |
3354 |
-+ } else { |
3355 |
-+ bfqq = kmem_cache_alloc_node(bfq_pool, |
3356 |
-+ gfp_mask | __GFP_ZERO, |
3357 |
-+ bfqd->queue->node); |
3358 |
-+ } |
3359 |
-+ |
3360 |
-+ if (bfqq != NULL) { |
3361 |
-+ bfq_init_bfqq(bfqd, bfqq, current->pid, is_sync); |
3362 |
-+ bfq_log_bfqq(bfqd, bfqq, "allocated"); |
3363 |
-+ } else { |
3364 |
-+ bfqq = &bfqd->oom_bfqq; |
3365 |
-+ bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); |
3366 |
-+ } |
3367 |
-+ |
3368 |
-+ bfq_init_prio_data(bfqq, bic); |
3369 |
-+ bfq_init_entity(&bfqq->entity, bfqg); |
3370 |
-+ } |
3371 |
-+ |
3372 |
-+ if (new_bfqq != NULL) |
3373 |
-+ kmem_cache_free(bfq_pool, new_bfqq); |
3374 |
-+ |
3375 |
-+ return bfqq; |
3376 |
-+} |
3377 |
-+ |
3378 |
-+static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, |
3379 |
-+ struct bfq_group *bfqg, |
3380 |
-+ int ioprio_class, int ioprio) |
3381 |
-+{ |
3382 |
-+ switch (ioprio_class) { |
3383 |
-+ case IOPRIO_CLASS_RT: |
3384 |
-+ return &bfqg->async_bfqq[0][ioprio]; |
3385 |
-+ case IOPRIO_CLASS_NONE: |
3386 |
-+ ioprio = IOPRIO_NORM; |
3387 |
-+ /* fall through */ |
3388 |
-+ case IOPRIO_CLASS_BE: |
3389 |
-+ return &bfqg->async_bfqq[1][ioprio]; |
3390 |
-+ case IOPRIO_CLASS_IDLE: |
3391 |
-+ return &bfqg->async_idle_bfqq; |
3392 |
-+ default: |
3393 |
-+ BUG(); |
3394 |
-+ } |
3395 |
-+} |
3396 |
-+ |
3397 |
-+static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
3398 |
-+ struct bfq_group *bfqg, int is_sync, |
3399 |
-+ struct bfq_io_cq *bic, gfp_t gfp_mask) |
3400 |
-+{ |
3401 |
-+ const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
3402 |
-+ const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
3403 |
-+ struct bfq_queue **async_bfqq = NULL; |
3404 |
-+ struct bfq_queue *bfqq = NULL; |
3405 |
-+ |
3406 |
-+ if (!is_sync) { |
3407 |
-+ async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
3408 |
-+ ioprio); |
3409 |
-+ bfqq = *async_bfqq; |
3410 |
-+ } |
3411 |
-+ |
3412 |
-+ if (bfqq == NULL) |
3413 |
-+ bfqq = bfq_find_alloc_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
3414 |
-+ |
3415 |
-+ /* |
3416 |
-+ * Pin the queue now that it's allocated, scheduler exit will prune it. |
3417 |
-+ */ |
3418 |
-+ if (!is_sync && *async_bfqq == NULL) { |
3419 |
-+ atomic_inc(&bfqq->ref); |
3420 |
-+ bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", |
3421 |
-+ bfqq, atomic_read(&bfqq->ref)); |
3422 |
-+ *async_bfqq = bfqq; |
3423 |
-+ } |
3424 |
-+ |
3425 |
-+ atomic_inc(&bfqq->ref); |
3426 |
-+ bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, |
3427 |
-+ atomic_read(&bfqq->ref)); |
3428 |
-+ return bfqq; |
3429 |
-+} |
3430 |
-+ |
3431 |
-+static void bfq_update_io_thinktime(struct bfq_data *bfqd, |
3432 |
-+ struct bfq_io_cq *bic) |
3433 |
-+{ |
3434 |
-+ unsigned long elapsed = jiffies - bic->ttime.last_end_request; |
3435 |
-+ unsigned long ttime = min(elapsed, 2UL * bfqd->bfq_slice_idle); |
3436 |
-+ |
3437 |
-+ bic->ttime.ttime_samples = (7*bic->ttime.ttime_samples + 256) / 8; |
3438 |
-+ bic->ttime.ttime_total = (7*bic->ttime.ttime_total + 256*ttime) / 8; |
3439 |
-+ bic->ttime.ttime_mean = (bic->ttime.ttime_total + 128) / |
3440 |
-+ bic->ttime.ttime_samples; |
3441 |
-+} |
3442 |
-+ |
3443 |
-+static void bfq_update_io_seektime(struct bfq_data *bfqd, |
3444 |
-+ struct bfq_queue *bfqq, |
3445 |
-+ struct request *rq) |
3446 |
-+{ |
3447 |
-+ sector_t sdist; |
3448 |
-+ u64 total; |
3449 |
-+ |
3450 |
-+ if (bfqq->last_request_pos < blk_rq_pos(rq)) |
3451 |
-+ sdist = blk_rq_pos(rq) - bfqq->last_request_pos; |
3452 |
-+ else |
3453 |
-+ sdist = bfqq->last_request_pos - blk_rq_pos(rq); |
3454 |
-+ |
3455 |
-+ /* |
3456 |
-+ * Don't allow the seek distance to get too large from the |
3457 |
-+ * odd fragment, pagein, etc. |
3458 |
-+ */ |
3459 |
-+ if (bfqq->seek_samples == 0) /* first request, not really a seek */ |
3460 |
-+ sdist = 0; |
3461 |
-+ else if (bfqq->seek_samples <= 60) /* second & third seek */ |
3462 |
-+ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*1024); |
3463 |
-+ else |
3464 |
-+ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*64); |
3465 |
-+ |
3466 |
-+ bfqq->seek_samples = (7*bfqq->seek_samples + 256) / 8; |
3467 |
-+ bfqq->seek_total = (7*bfqq->seek_total + (u64)256*sdist) / 8; |
3468 |
-+ total = bfqq->seek_total + (bfqq->seek_samples/2); |
3469 |
-+ do_div(total, bfqq->seek_samples); |
3470 |
-+ bfqq->seek_mean = (sector_t)total; |
3471 |
-+ |
3472 |
-+ bfq_log_bfqq(bfqd, bfqq, "dist=%llu mean=%llu", (u64)sdist, |
3473 |
-+ (u64)bfqq->seek_mean); |
3474 |
-+} |
3475 |
-+ |
3476 |
-+/* |
3477 |
-+ * Disable idle window if the process thinks too long or seeks so much that |
3478 |
-+ * it doesn't matter. |
3479 |
-+ */ |
3480 |
-+static void bfq_update_idle_window(struct bfq_data *bfqd, |
3481 |
-+ struct bfq_queue *bfqq, |
3482 |
-+ struct bfq_io_cq *bic) |
3483 |
-+{ |
3484 |
-+ int enable_idle; |
3485 |
-+ |
3486 |
-+ /* Don't idle for async or idle io prio class. */ |
3487 |
-+ if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq)) |
3488 |
-+ return; |
3489 |
-+ |
3490 |
-+ enable_idle = bfq_bfqq_idle_window(bfqq); |
3491 |
-+ |
3492 |
-+ if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
3493 |
-+ bfqd->bfq_slice_idle == 0 || |
3494 |
-+ (bfqd->hw_tag && BFQQ_SEEKY(bfqq) && |
3495 |
-+ bfqq->raising_coeff == 1)) |
3496 |
-+ enable_idle = 0; |
3497 |
-+ else if (bfq_sample_valid(bic->ttime.ttime_samples)) { |
3498 |
-+ if (bic->ttime.ttime_mean > bfqd->bfq_slice_idle && |
3499 |
-+ bfqq->raising_coeff == 1) |
3500 |
-+ enable_idle = 0; |
3501 |
-+ else |
3502 |
-+ enable_idle = 1; |
3503 |
-+ } |
3504 |
-+ bfq_log_bfqq(bfqd, bfqq, "update_idle_window: enable_idle %d", |
3505 |
-+ enable_idle); |
3506 |
-+ |
3507 |
-+ if (enable_idle) |
3508 |
-+ bfq_mark_bfqq_idle_window(bfqq); |
3509 |
-+ else |
3510 |
-+ bfq_clear_bfqq_idle_window(bfqq); |
3511 |
-+} |
3512 |
-+ |
3513 |
-+/* |
3514 |
-+ * Called when a new fs request (rq) is added to bfqq. Check if there's |
3515 |
-+ * something we should do about it. |
3516 |
-+ */ |
3517 |
-+static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
3518 |
-+ struct request *rq) |
3519 |
-+{ |
3520 |
-+ struct bfq_io_cq *bic = RQ_BIC(rq); |
3521 |
-+ |
3522 |
-+ if (rq->cmd_flags & REQ_META) |
3523 |
-+ bfqq->meta_pending++; |
3524 |
-+ |
3525 |
-+ bfq_update_io_thinktime(bfqd, bic); |
3526 |
-+ bfq_update_io_seektime(bfqd, bfqq, rq); |
3527 |
-+ if (bfqq->entity.service > bfq_max_budget(bfqd) / 8 || |
3528 |
-+ !BFQQ_SEEKY(bfqq)) |
3529 |
-+ bfq_update_idle_window(bfqd, bfqq, bic); |
3530 |
-+ |
3531 |
-+ bfq_log_bfqq(bfqd, bfqq, |
3532 |
-+ "rq_enqueued: idle_window=%d (seeky %d, mean %llu)", |
3533 |
-+ bfq_bfqq_idle_window(bfqq), BFQQ_SEEKY(bfqq), |
3534 |
-+ (long long unsigned)bfqq->seek_mean); |
3535 |
-+ |
3536 |
-+ bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
3537 |
-+ |
3538 |
-+ if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) { |
3539 |
-+ int small_req = bfqq->queued[rq_is_sync(rq)] == 1 && |
3540 |
-+ blk_rq_sectors(rq) < 32; |
3541 |
-+ int budget_timeout = bfq_bfqq_budget_timeout(bfqq); |
3542 |
-+ |
3543 |
-+ /* |
3544 |
-+ * There is just this request queued: if the request |
3545 |
-+ * is small and the queue is not to be expired, then |
3546 |
-+ * just exit. |
3547 |
-+ * |
3548 |
-+ * In this way, if the disk is being idled to wait for |
3549 |
-+ * a new request from the in-service queue, we avoid |
3550 |
-+ * unplugging the device and committing the disk to serve |
3551 |
-+ * just a small request. On the contrary, we wait for |
3552 |
-+ * the block layer to decide when to unplug the device: |
3553 |
-+ * hopefully, new requests will be merged to this one |
3554 |
-+ * quickly, then the device will be unplugged and |
3555 |
-+ * larger requests will be dispatched. |
3556 |
-+ */ |
3557 |
-+ if (small_req && !budget_timeout) |
3558 |
-+ return; |
3559 |
-+ |
3560 |
-+ /* |
3561 |
-+ * A large enough request arrived, or the queue is to |
3562 |
-+ * be expired: in both cases disk idling is to be |
3563 |
-+ * stopped, so clear wait_request flag and reset |
3564 |
-+ * timer. |
3565 |
-+ */ |
3566 |
-+ bfq_clear_bfqq_wait_request(bfqq); |
3567 |
-+ del_timer(&bfqd->idle_slice_timer); |
3568 |
-+ |
3569 |
-+ /* |
3570 |
-+ * The queue is not empty, because a new request just |
3571 |
-+ * arrived. Hence we can safely expire the queue, in |
3572 |
-+ * case of budget timeout, without risking that the |
3573 |
-+ * timestamps of the queue are not updated correctly. |
3574 |
-+ * See [1] for more details. |
3575 |
-+ */ |
3576 |
-+ if (budget_timeout) |
3577 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_TIMEOUT); |
3578 |
-+ |
3579 |
-+ /* |
3580 |
-+ * Let the request rip immediately, or let a new queue be |
3581 |
-+ * selected if bfqq has just been expired. |
3582 |
-+ */ |
3583 |
-+ __blk_run_queue(bfqd->queue); |
3584 |
-+ } |
3585 |
-+} |
3586 |
-+ |
3587 |
-+static void bfq_insert_request(struct request_queue *q, struct request *rq) |
3588 |
-+{ |
3589 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
3590 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
3591 |
-+ |
3592 |
-+ assert_spin_locked(bfqd->queue->queue_lock); |
3593 |
-+ bfq_init_prio_data(bfqq, RQ_BIC(rq)); |
3594 |
-+ |
3595 |
-+ bfq_add_rq_rb(rq); |
3596 |
-+ |
3597 |
-+ rq_set_fifo_time(rq, jiffies + bfqd->bfq_fifo_expire[rq_is_sync(rq)]); |
3598 |
-+ list_add_tail(&rq->queuelist, &bfqq->fifo); |
3599 |
-+ |
3600 |
-+ bfq_rq_enqueued(bfqd, bfqq, rq); |
3601 |
-+} |
3602 |
-+ |
3603 |
-+static void bfq_update_hw_tag(struct bfq_data *bfqd) |
3604 |
-+{ |
3605 |
-+ bfqd->max_rq_in_driver = max(bfqd->max_rq_in_driver, |
3606 |
-+ bfqd->rq_in_driver); |
3607 |
-+ |
3608 |
-+ if (bfqd->hw_tag == 1) |
3609 |
-+ return; |
3610 |
-+ |
3611 |
-+ /* |
3612 |
-+ * This sample is valid if the number of outstanding requests |
3613 |
-+ * is large enough to allow a queueing behavior. Note that the |
3614 |
-+ * sum is not exact, as it's not taking into account deactivated |
3615 |
-+ * requests. |
3616 |
-+ */ |
3617 |
-+ if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD) |
3618 |
-+ return; |
3619 |
-+ |
3620 |
-+ if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) |
3621 |
-+ return; |
3622 |
-+ |
3623 |
-+ bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; |
3624 |
-+ bfqd->max_rq_in_driver = 0; |
3625 |
-+ bfqd->hw_tag_samples = 0; |
3626 |
-+} |
3627 |
-+ |
3628 |
-+static void bfq_completed_request(struct request_queue *q, struct request *rq) |
3629 |
-+{ |
3630 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
3631 |
-+ struct bfq_data *bfqd = bfqq->bfqd; |
3632 |
-+ const int sync = rq_is_sync(rq); |
3633 |
-+ |
3634 |
-+ bfq_log_bfqq(bfqd, bfqq, "completed %u sects req (%d)", |
3635 |
-+ blk_rq_sectors(rq), sync); |
3636 |
-+ |
3637 |
-+ bfq_update_hw_tag(bfqd); |
3638 |
-+ |
3639 |
-+ WARN_ON(!bfqd->rq_in_driver); |
3640 |
-+ WARN_ON(!bfqq->dispatched); |
3641 |
-+ bfqd->rq_in_driver--; |
3642 |
-+ bfqq->dispatched--; |
3643 |
-+ |
3644 |
-+ if (bfq_bfqq_sync(bfqq)) |
3645 |
-+ bfqd->sync_flight--; |
3646 |
-+ |
3647 |
-+ if (sync) |
3648 |
-+ RQ_BIC(rq)->ttime.last_end_request = jiffies; |
3649 |
-+ |
3650 |
-+ /* |
3651 |
-+ * If we are waiting to discover whether the request pattern of the |
3652 |
-+ * task associated with the queue is actually isochronous, and |
3653 |
-+ * both requisites for this condition to hold are satisfied, then |
3654 |
-+ * compute soft_rt_next_start (see the comments to the function |
3655 |
-+ * bfq_bfqq_softrt_next_start()). |
3656 |
-+ */ |
3657 |
-+ if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 && |
3658 |
-+ RB_EMPTY_ROOT(&bfqq->sort_list)) |
3659 |
-+ bfqq->soft_rt_next_start = |
3660 |
-+ bfq_bfqq_softrt_next_start(bfqd, bfqq); |
3661 |
-+ |
3662 |
-+ /* |
3663 |
-+ * If this is the in-service queue, check if it needs to be expired, |
3664 |
-+ * or if we want to idle in case it has no pending requests. |
3665 |
-+ */ |
3666 |
-+ if (bfqd->in_service_queue == bfqq) { |
3667 |
-+ if (bfq_bfqq_budget_new(bfqq)) |
3668 |
-+ bfq_set_budget_timeout(bfqd); |
3669 |
-+ |
3670 |
-+ if (bfq_bfqq_must_idle(bfqq)) { |
3671 |
-+ bfq_arm_slice_timer(bfqd); |
3672 |
-+ goto out; |
3673 |
-+ } else if (bfq_may_expire_for_budg_timeout(bfqq)) |
3674 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_TIMEOUT); |
3675 |
-+ else if (RB_EMPTY_ROOT(&bfqq->sort_list) && |
3676 |
-+ (bfqq->dispatched == 0 || |
3677 |
-+ !bfq_bfqq_must_not_expire(bfqq))) |
3678 |
-+ bfq_bfqq_expire(bfqd, bfqq, 0, |
3679 |
-+ BFQ_BFQQ_NO_MORE_REQUESTS); |
3680 |
-+ } |
3681 |
-+ |
3682 |
-+ if (!bfqd->rq_in_driver) |
3683 |
-+ bfq_schedule_dispatch(bfqd); |
3684 |
-+ |
3685 |
-+out: |
3686 |
-+ return; |
3687 |
-+} |
3688 |
-+ |
3689 |
-+static inline int __bfq_may_queue(struct bfq_queue *bfqq) |
3690 |
-+{ |
3691 |
-+ if (bfq_bfqq_wait_request(bfqq) && bfq_bfqq_must_alloc(bfqq)) { |
3692 |
-+ bfq_clear_bfqq_must_alloc(bfqq); |
3693 |
-+ return ELV_MQUEUE_MUST; |
3694 |
-+ } |
3695 |
-+ |
3696 |
-+ return ELV_MQUEUE_MAY; |
3697 |
-+} |
3698 |
-+ |
3699 |
-+static int bfq_may_queue(struct request_queue *q, int rw) |
3700 |
-+{ |
3701 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
3702 |
-+ struct task_struct *tsk = current; |
3703 |
-+ struct bfq_io_cq *bic; |
3704 |
-+ struct bfq_queue *bfqq; |
3705 |
-+ |
3706 |
-+ /* |
3707 |
-+ * Don't force setup of a queue from here, as a call to may_queue |
3708 |
-+ * does not necessarily imply that a request actually will be queued. |
3709 |
-+ * So just lookup a possibly existing queue, or return 'may queue' |
3710 |
-+ * if that fails. |
3711 |
-+ */ |
3712 |
-+ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
3713 |
-+ if (bic == NULL) |
3714 |
-+ return ELV_MQUEUE_MAY; |
3715 |
-+ |
3716 |
-+ bfqq = bic_to_bfqq(bic, rw_is_sync(rw)); |
3717 |
-+ if (bfqq != NULL) { |
3718 |
-+ bfq_init_prio_data(bfqq, bic); |
3719 |
-+ |
3720 |
-+ return __bfq_may_queue(bfqq); |
3721 |
-+ } |
3722 |
-+ |
3723 |
-+ return ELV_MQUEUE_MAY; |
3724 |
-+} |
3725 |
-+ |
3726 |
-+/* |
3727 |
-+ * Queue lock held here. |
3728 |
-+ */ |
3729 |
-+static void bfq_put_request(struct request *rq) |
3730 |
-+{ |
3731 |
-+ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
3732 |
-+ |
3733 |
-+ if (bfqq != NULL) { |
3734 |
-+ const int rw = rq_data_dir(rq); |
3735 |
-+ |
3736 |
-+ BUG_ON(!bfqq->allocated[rw]); |
3737 |
-+ bfqq->allocated[rw]--; |
3738 |
-+ |
3739 |
-+ rq->elv.priv[0] = NULL; |
3740 |
-+ rq->elv.priv[1] = NULL; |
3741 |
-+ |
3742 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "put_request %p, %d", |
3743 |
-+ bfqq, atomic_read(&bfqq->ref)); |
3744 |
-+ bfq_put_queue(bfqq); |
3745 |
-+ } |
3746 |
-+} |
3747 |
-+ |
3748 |
-+static struct bfq_queue * |
3749 |
-+bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
3750 |
-+ struct bfq_queue *bfqq) |
3751 |
-+{ |
3752 |
-+ bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
3753 |
-+ (long unsigned)bfqq->new_bfqq->pid); |
3754 |
-+ bic_set_bfqq(bic, bfqq->new_bfqq, 1); |
3755 |
-+ bfq_mark_bfqq_coop(bfqq->new_bfqq); |
3756 |
-+ bfq_put_queue(bfqq); |
3757 |
-+ return bic_to_bfqq(bic, 1); |
3758 |
-+} |
3759 |
-+ |
3760 |
-+/* |
3761 |
-+ * Returns NULL if a new bfqq should be allocated, or the old bfqq if this |
3762 |
-+ * was the last process referring to said bfqq. |
3763 |
-+ */ |
3764 |
-+static struct bfq_queue * |
3765 |
-+bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) |
3766 |
-+{ |
3767 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); |
3768 |
-+ if (bfqq_process_refs(bfqq) == 1) { |
3769 |
-+ bfqq->pid = current->pid; |
3770 |
-+ bfq_clear_bfqq_coop(bfqq); |
3771 |
-+ bfq_clear_bfqq_split_coop(bfqq); |
3772 |
-+ return bfqq; |
3773 |
-+ } |
3774 |
-+ |
3775 |
-+ bic_set_bfqq(bic, NULL, 1); |
3776 |
-+ |
3777 |
-+ bfq_put_cooperator(bfqq); |
3778 |
-+ |
3779 |
-+ bfq_put_queue(bfqq); |
3780 |
-+ return NULL; |
3781 |
-+} |
3782 |
-+ |
3783 |
-+/* |
3784 |
-+ * Allocate bfq data structures associated with this request. |
3785 |
-+ */ |
3786 |
-+static int bfq_set_request(struct request_queue *q, struct request *rq, |
3787 |
-+ struct bio *bio, gfp_t gfp_mask) |
3788 |
-+{ |
3789 |
-+ struct bfq_data *bfqd = q->elevator->elevator_data; |
3790 |
-+ struct bfq_io_cq *bic = icq_to_bic(rq->elv.icq); |
3791 |
-+ const int rw = rq_data_dir(rq); |
3792 |
-+ const int is_sync = rq_is_sync(rq); |
3793 |
-+ struct bfq_queue *bfqq; |
3794 |
-+ struct bfq_group *bfqg; |
3795 |
-+ unsigned long flags; |
3796 |
-+ |
3797 |
-+ might_sleep_if(gfp_mask & __GFP_WAIT); |
3798 |
-+ |
3799 |
-+ bfq_changed_ioprio(bic); |
3800 |
-+ |
3801 |
-+ spin_lock_irqsave(q->queue_lock, flags); |
3802 |
-+ |
3803 |
-+ if (bic == NULL) |
3804 |
-+ goto queue_fail; |
3805 |
-+ |
3806 |
-+ bfqg = bfq_bic_update_cgroup(bic); |
3807 |
-+ |
3808 |
-+new_queue: |
3809 |
-+ bfqq = bic_to_bfqq(bic, is_sync); |
3810 |
-+ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
3811 |
-+ bfqq = bfq_get_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
3812 |
-+ bic_set_bfqq(bic, bfqq, is_sync); |
3813 |
-+ } else { |
3814 |
-+ /* |
3815 |
-+ * If the queue was seeky for too long, break it apart. |
3816 |
-+ */ |
3817 |
-+ if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { |
3818 |
-+ bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); |
3819 |
-+ bfqq = bfq_split_bfqq(bic, bfqq); |
3820 |
-+ if (!bfqq) |
3821 |
-+ goto new_queue; |
3822 |
-+ } |
3823 |
-+ |
3824 |
-+ /* |
3825 |
-+ * Check to see if this queue is scheduled to merge with |
3826 |
-+ * another closely cooperating queue. The merging of queues |
3827 |
-+ * happens here as it must be done in process context. |
3828 |
-+ * The reference on new_bfqq was taken in merge_bfqqs. |
3829 |
-+ */ |
3830 |
-+ if (bfqq->new_bfqq != NULL) |
3831 |
-+ bfqq = bfq_merge_bfqqs(bfqd, bic, bfqq); |
3832 |
-+ } |
3833 |
-+ |
3834 |
-+ bfqq->allocated[rw]++; |
3835 |
-+ atomic_inc(&bfqq->ref); |
3836 |
-+ bfq_log_bfqq(bfqd, bfqq, "set_request: bfqq %p, %d", bfqq, |
3837 |
-+ atomic_read(&bfqq->ref)); |
3838 |
-+ |
3839 |
-+ rq->elv.priv[0] = bic; |
3840 |
-+ rq->elv.priv[1] = bfqq; |
3841 |
-+ |
3842 |
-+ spin_unlock_irqrestore(q->queue_lock, flags); |
3843 |
-+ |
3844 |
-+ return 0; |
3845 |
-+ |
3846 |
-+queue_fail: |
3847 |
-+ bfq_schedule_dispatch(bfqd); |
3848 |
-+ spin_unlock_irqrestore(q->queue_lock, flags); |
3849 |
-+ |
3850 |
-+ return 1; |
3851 |
-+} |
3852 |
-+ |
3853 |
-+static void bfq_kick_queue(struct work_struct *work) |
3854 |
-+{ |
3855 |
-+ struct bfq_data *bfqd = |
3856 |
-+ container_of(work, struct bfq_data, unplug_work); |
3857 |
-+ struct request_queue *q = bfqd->queue; |
3858 |
-+ |
3859 |
-+ spin_lock_irq(q->queue_lock); |
3860 |
-+ __blk_run_queue(q); |
3861 |
-+ spin_unlock_irq(q->queue_lock); |
3862 |
-+} |
3863 |
-+ |
3864 |
-+/* |
3865 |
-+ * Handler of the expiration of the timer running if the in-service queue |
3866 |
-+ * is idling inside its time slice. |
3867 |
-+ */ |
3868 |
-+static void bfq_idle_slice_timer(unsigned long data) |
3869 |
-+{ |
3870 |
-+ struct bfq_data *bfqd = (struct bfq_data *)data; |
3871 |
-+ struct bfq_queue *bfqq; |
3872 |
-+ unsigned long flags; |
3873 |
-+ enum bfqq_expiration reason; |
3874 |
-+ |
3875 |
-+ spin_lock_irqsave(bfqd->queue->queue_lock, flags); |
3876 |
-+ |
3877 |
-+ bfqq = bfqd->in_service_queue; |
3878 |
-+ /* |
3879 |
-+ * Theoretical race here: the in-service queue can be NULL or different |
3880 |
-+ * from the queue that was idling if the timer handler spins on |
3881 |
-+ * the queue_lock and a new request arrives for the current |
3882 |
-+ * queue and there is a full dispatch cycle that changes the |
3883 |
-+ * in-service queue. This can hardly happen, but in the worst case |
3884 |
-+ * we just expire a queue too early. |
3885 |
-+ */ |
3886 |
-+ if (bfqq != NULL) { |
3887 |
-+ bfq_log_bfqq(bfqd, bfqq, "slice_timer expired"); |
3888 |
-+ if (bfq_bfqq_budget_timeout(bfqq)) |
3889 |
-+ /* |
3890 |
-+ * Also here the queue can be safely expired |
3891 |
-+ * for budget timeout without wasting |
3892 |
-+ * guarantees |
3893 |
-+ */ |
3894 |
-+ reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
3895 |
-+ else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) |
3896 |
-+ /* |
3897 |
-+ * The queue may not be empty upon timer expiration, |
3898 |
-+ * because we may not disable the timer when the first |
3899 |
-+ * request of the in-service queue arrives during |
3900 |
-+ * disk idling |
3901 |
-+ */ |
3902 |
-+ reason = BFQ_BFQQ_TOO_IDLE; |
3903 |
-+ else |
3904 |
-+ goto schedule_dispatch; |
3905 |
-+ |
3906 |
-+ bfq_bfqq_expire(bfqd, bfqq, 1, reason); |
3907 |
-+ } |
3908 |
-+ |
3909 |
-+schedule_dispatch: |
3910 |
-+ bfq_schedule_dispatch(bfqd); |
3911 |
-+ |
3912 |
-+ spin_unlock_irqrestore(bfqd->queue->queue_lock, flags); |
3913 |
-+} |
3914 |
-+ |
3915 |
-+static void bfq_shutdown_timer_wq(struct bfq_data *bfqd) |
3916 |
-+{ |
3917 |
-+ del_timer_sync(&bfqd->idle_slice_timer); |
3918 |
-+ cancel_work_sync(&bfqd->unplug_work); |
3919 |
-+} |
3920 |
-+ |
3921 |
-+static inline void __bfq_put_async_bfqq(struct bfq_data *bfqd, |
3922 |
-+ struct bfq_queue **bfqq_ptr) |
3923 |
-+{ |
3924 |
-+ struct bfq_group *root_group = bfqd->root_group; |
3925 |
-+ struct bfq_queue *bfqq = *bfqq_ptr; |
3926 |
-+ |
3927 |
-+ bfq_log(bfqd, "put_async_bfqq: %p", bfqq); |
3928 |
-+ if (bfqq != NULL) { |
3929 |
-+ bfq_bfqq_move(bfqd, bfqq, &bfqq->entity, root_group); |
3930 |
-+ bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
3931 |
-+ bfqq, atomic_read(&bfqq->ref)); |
3932 |
-+ bfq_put_queue(bfqq); |
3933 |
-+ *bfqq_ptr = NULL; |
3934 |
-+ } |
3935 |
-+} |
3936 |
-+ |
3937 |
-+/* |
3938 |
-+ * Release all the bfqg references to its async queues. If we are |
3939 |
-+ * deallocating the group these queues may still contain requests, so |
3940 |
-+ * we reparent them to the root cgroup (i.e., the only one that will |
3941 |
-+ * exist for sure until all the requests on a device are gone). |
3942 |
-+ */ |
3943 |
-+static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
3944 |
-+{ |
3945 |
-+ int i, j; |
3946 |
-+ |
3947 |
-+ for (i = 0; i < 2; i++) |
3948 |
-+ for (j = 0; j < IOPRIO_BE_NR; j++) |
3949 |
-+ __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]); |
3950 |
-+ |
3951 |
-+ __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq); |
3952 |
-+} |
3953 |
-+ |
3954 |
-+static void bfq_exit_queue(struct elevator_queue *e) |
3955 |
-+{ |
3956 |
-+ struct bfq_data *bfqd = e->elevator_data; |
3957 |
-+ struct request_queue *q = bfqd->queue; |
3958 |
-+ struct bfq_queue *bfqq, *n; |
3959 |
-+ |
3960 |
-+ bfq_shutdown_timer_wq(bfqd); |
3961 |
-+ |
3962 |
-+ spin_lock_irq(q->queue_lock); |
3963 |
-+ |
3964 |
-+ BUG_ON(bfqd->in_service_queue != NULL); |
3965 |
-+ list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) |
3966 |
-+ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
3967 |
-+ |
3968 |
-+ bfq_disconnect_groups(bfqd); |
3969 |
-+ spin_unlock_irq(q->queue_lock); |
3970 |
-+ |
3971 |
-+ bfq_shutdown_timer_wq(bfqd); |
3972 |
-+ |
3973 |
-+ synchronize_rcu(); |
3974 |
-+ |
3975 |
-+ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
3976 |
-+ |
3977 |
-+ bfq_free_root_group(bfqd); |
3978 |
-+ kfree(bfqd); |
3979 |
-+} |
3980 |
-+ |
3981 |
-+static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
3982 |
-+{ |
3983 |
-+ struct bfq_group *bfqg; |
3984 |
-+ struct bfq_data *bfqd; |
3985 |
-+ struct elevator_queue *eq; |
3986 |
-+ |
3987 |
-+ eq = elevator_alloc(q, e); |
3988 |
-+ if (eq == NULL) |
3989 |
-+ return -ENOMEM; |
3990 |
-+ |
3991 |
-+ bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node); |
3992 |
-+ if (bfqd == NULL) { |
3993 |
-+ kobject_put(&eq->kobj); |
3994 |
-+ return -ENOMEM; |
3995 |
-+ } |
3996 |
-+ eq->elevator_data = bfqd; |
3997 |
-+ |
3998 |
-+ /* |
3999 |
-+ * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. |
4000 |
-+ * Grab a permanent reference to it, so that the normal code flow |
4001 |
-+ * will not attempt to free it. |
4002 |
-+ */ |
4003 |
-+ bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, 1, 0); |
4004 |
-+ atomic_inc(&bfqd->oom_bfqq.ref); |
4005 |
-+ |
4006 |
-+ bfqd->queue = q; |
4007 |
-+ |
4008 |
-+ spin_lock_irq(q->queue_lock); |
4009 |
-+ q->elevator = eq; |
4010 |
-+ spin_unlock_irq(q->queue_lock); |
4011 |
-+ |
4012 |
-+ bfqg = bfq_alloc_root_group(bfqd, q->node); |
4013 |
-+ if (bfqg == NULL) { |
4014 |
-+ kfree(bfqd); |
4015 |
-+ kobject_put(&eq->kobj); |
4016 |
-+ return -ENOMEM; |
4017 |
-+ } |
4018 |
-+ |
4019 |
-+ bfqd->root_group = bfqg; |
4020 |
-+ |
4021 |
-+ init_timer(&bfqd->idle_slice_timer); |
4022 |
-+ bfqd->idle_slice_timer.function = bfq_idle_slice_timer; |
4023 |
-+ bfqd->idle_slice_timer.data = (unsigned long)bfqd; |
4024 |
-+ |
4025 |
-+ bfqd->rq_pos_tree = RB_ROOT; |
4026 |
-+ |
4027 |
-+ INIT_WORK(&bfqd->unplug_work, bfq_kick_queue); |
4028 |
-+ |
4029 |
-+ INIT_LIST_HEAD(&bfqd->active_list); |
4030 |
-+ INIT_LIST_HEAD(&bfqd->idle_list); |
4031 |
-+ |
4032 |
-+ bfqd->hw_tag = -1; |
4033 |
-+ |
4034 |
-+ bfqd->bfq_max_budget = bfq_default_max_budget; |
4035 |
-+ |
4036 |
-+ bfqd->bfq_quantum = bfq_quantum; |
4037 |
-+ bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; |
4038 |
-+ bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; |
4039 |
-+ bfqd->bfq_back_max = bfq_back_max; |
4040 |
-+ bfqd->bfq_back_penalty = bfq_back_penalty; |
4041 |
-+ bfqd->bfq_slice_idle = bfq_slice_idle; |
4042 |
-+ bfqd->bfq_class_idle_last_service = 0; |
4043 |
-+ bfqd->bfq_max_budget_async_rq = bfq_max_budget_async_rq; |
4044 |
-+ bfqd->bfq_timeout[BLK_RW_ASYNC] = bfq_timeout_async; |
4045 |
-+ bfqd->bfq_timeout[BLK_RW_SYNC] = bfq_timeout_sync; |
4046 |
-+ |
4047 |
-+ bfqd->low_latency = true; |
4048 |
-+ |
4049 |
-+ bfqd->bfq_raising_coeff = 20; |
4050 |
-+ bfqd->bfq_raising_rt_max_time = msecs_to_jiffies(300); |
4051 |
-+ bfqd->bfq_raising_max_time = 0; |
4052 |
-+ bfqd->bfq_raising_min_idle_time = msecs_to_jiffies(2000); |
4053 |
-+ bfqd->bfq_raising_min_inter_arr_async = msecs_to_jiffies(500); |
4054 |
-+ bfqd->bfq_raising_max_softrt_rate = 7000; /* |
4055 |
-+ * Approximate rate required |
4056 |
-+ * to playback or record a |
4057 |
-+ * high-definition compressed |
4058 |
-+ * video. |
4059 |
-+ */ |
4060 |
-+ bfqd->raised_busy_queues = 0; |
4061 |
-+ |
4062 |
-+ /* Initially estimate the device's peak rate as the reference rate */ |
4063 |
-+ if (blk_queue_nonrot(bfqd->queue)) { |
4064 |
-+ bfqd->RT_prod = R_nonrot * T_nonrot; |
4065 |
-+ bfqd->peak_rate = R_nonrot; |
4066 |
-+ } else { |
4067 |
-+ bfqd->RT_prod = R_rot * T_rot; |
4068 |
-+ bfqd->peak_rate = R_rot; |
4069 |
-+ } |
4070 |
-+ |
4071 |
-+ return 0; |
4072 |
-+} |
4073 |
-+ |
4074 |
-+static void bfq_slab_kill(void) |
4075 |
-+{ |
4076 |
-+ if (bfq_pool != NULL) |
4077 |
-+ kmem_cache_destroy(bfq_pool); |
4078 |
-+} |
4079 |
-+ |
4080 |
-+static int __init bfq_slab_setup(void) |
4081 |
-+{ |
4082 |
-+ bfq_pool = KMEM_CACHE(bfq_queue, 0); |
4083 |
-+ if (bfq_pool == NULL) |
4084 |
-+ return -ENOMEM; |
4085 |
-+ return 0; |
4086 |
-+} |
4087 |
-+ |
4088 |
-+static ssize_t bfq_var_show(unsigned int var, char *page) |
4089 |
-+{ |
4090 |
-+ return sprintf(page, "%d\n", var); |
4091 |
-+} |
4092 |
-+ |
4093 |
-+static ssize_t bfq_var_store(unsigned long *var, const char *page, size_t count) |
4094 |
-+{ |
4095 |
-+ unsigned long new_val; |
4096 |
-+ int ret = kstrtoul(page, 10, &new_val); |
4097 |
-+ |
4098 |
-+ if (ret == 0) |
4099 |
-+ *var = new_val; |
4100 |
-+ |
4101 |
-+ return count; |
4102 |
-+} |
4103 |
-+ |
4104 |
-+static ssize_t bfq_raising_max_time_show(struct elevator_queue *e, char *page) |
4105 |
-+{ |
4106 |
-+ struct bfq_data *bfqd = e->elevator_data; |
4107 |
-+ return sprintf(page, "%d\n", bfqd->bfq_raising_max_time > 0 ? |
4108 |
-+ jiffies_to_msecs(bfqd->bfq_raising_max_time) : |
4109 |
-+ jiffies_to_msecs(bfq_wrais_duration(bfqd))); |
4110 |
-+} |
4111 |
-+ |
4112 |
-+static ssize_t bfq_weights_show(struct elevator_queue *e, char *page) |
4113 |
-+{ |
4114 |
-+ struct bfq_queue *bfqq; |
4115 |
-+ struct bfq_data *bfqd = e->elevator_data; |
4116 |
-+ ssize_t num_char = 0; |
4117 |
-+ |
4118 |
-+ num_char += sprintf(page + num_char, "Tot reqs queued %d\n\n", |
4119 |
-+ bfqd->queued); |
4120 |
-+ |
4121 |
-+ spin_lock_irq(bfqd->queue->queue_lock); |
4122 |
-+ |
4123 |
-+ num_char += sprintf(page + num_char, "Active:\n"); |
4124 |
-+ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) { |
4125 |
-+ num_char += sprintf(page + num_char, |
4126 |
-+ "pid%d: weight %hu, nr_queued %d %d," |
4127 |
-+ " dur %d/%u\n", |
4128 |
-+ bfqq->pid, |
4129 |
-+ bfqq->entity.weight, |
4130 |
-+ bfqq->queued[0], |
4131 |
-+ bfqq->queued[1], |
4132 |
-+ jiffies_to_msecs(jiffies - |
4133 |
-+ bfqq->last_rais_start_finish), |
4134 |
-+ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
4135 |
-+ } |
4136 |
-+ |
4137 |
-+ num_char += sprintf(page + num_char, "Idle:\n"); |
4138 |
-+ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) { |
4139 |
-+ num_char += sprintf(page + num_char, |
4140 |
-+ "pid%d: weight %hu, dur %d/%u\n", |
4141 |
-+ bfqq->pid, |
4142 |
-+ bfqq->entity.weight, |
4143 |
-+ jiffies_to_msecs(jiffies - |
4144 |
-+ bfqq->last_rais_start_finish), |
4145 |
-+ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
4146 |
-+ } |
4147 |
-+ |
4148 |
-+ spin_unlock_irq(bfqd->queue->queue_lock); |
4149 |
-+ |
4150 |
-+ return num_char; |
4151 |
-+} |
4152 |
-+ |
4153 |
-+#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
4154 |
-+static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
4155 |
-+{ \ |
4156 |
-+ struct bfq_data *bfqd = e->elevator_data; \ |
4157 |
-+ unsigned int __data = __VAR; \ |
4158 |
-+ if (__CONV) \ |
4159 |
-+ __data = jiffies_to_msecs(__data); \ |
4160 |
-+ return bfq_var_show(__data, (page)); \ |
4161 |
-+} |
4162 |
-+SHOW_FUNCTION(bfq_quantum_show, bfqd->bfq_quantum, 0); |
4163 |
-+SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 1); |
4164 |
-+SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 1); |
4165 |
-+SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); |
4166 |
-+SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); |
4167 |
-+SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 1); |
4168 |
-+SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); |
4169 |
-+SHOW_FUNCTION(bfq_max_budget_async_rq_show, bfqd->bfq_max_budget_async_rq, 0); |
4170 |
-+SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout[BLK_RW_SYNC], 1); |
4171 |
-+SHOW_FUNCTION(bfq_timeout_async_show, bfqd->bfq_timeout[BLK_RW_ASYNC], 1); |
4172 |
-+SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
4173 |
-+SHOW_FUNCTION(bfq_raising_coeff_show, bfqd->bfq_raising_coeff, 0); |
4174 |
-+SHOW_FUNCTION(bfq_raising_rt_max_time_show, bfqd->bfq_raising_rt_max_time, 1); |
4175 |
-+SHOW_FUNCTION(bfq_raising_min_idle_time_show, bfqd->bfq_raising_min_idle_time, |
4176 |
-+ 1); |
4177 |
-+SHOW_FUNCTION(bfq_raising_min_inter_arr_async_show, |
4178 |
-+ bfqd->bfq_raising_min_inter_arr_async, |
4179 |
-+ 1); |
4180 |
-+SHOW_FUNCTION(bfq_raising_max_softrt_rate_show, |
4181 |
-+ bfqd->bfq_raising_max_softrt_rate, 0); |
4182 |
-+#undef SHOW_FUNCTION |
4183 |
-+ |
4184 |
-+#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ |
4185 |
-+static ssize_t \ |
4186 |
-+__FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
4187 |
-+{ \ |
4188 |
-+ struct bfq_data *bfqd = e->elevator_data; \ |
4189 |
-+ unsigned long uninitialized_var(__data); \ |
4190 |
-+ int ret = bfq_var_store(&__data, (page), count); \ |
4191 |
-+ if (__data < (MIN)) \ |
4192 |
-+ __data = (MIN); \ |
4193 |
-+ else if (__data > (MAX)) \ |
4194 |
-+ __data = (MAX); \ |
4195 |
-+ if (__CONV) \ |
4196 |
-+ *(__PTR) = msecs_to_jiffies(__data); \ |
4197 |
-+ else \ |
4198 |
-+ *(__PTR) = __data; \ |
4199 |
-+ return ret; \ |
4200 |
-+} |
4201 |
-+STORE_FUNCTION(bfq_quantum_store, &bfqd->bfq_quantum, 1, INT_MAX, 0); |
4202 |
-+STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, |
4203 |
-+ INT_MAX, 1); |
4204 |
-+STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, |
4205 |
-+ INT_MAX, 1); |
4206 |
-+STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); |
4207 |
-+STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, |
4208 |
-+ INT_MAX, 0); |
4209 |
-+STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 1); |
4210 |
-+STORE_FUNCTION(bfq_max_budget_async_rq_store, &bfqd->bfq_max_budget_async_rq, |
4211 |
-+ 1, INT_MAX, 0); |
4212 |
-+STORE_FUNCTION(bfq_timeout_async_store, &bfqd->bfq_timeout[BLK_RW_ASYNC], 0, |
4213 |
-+ INT_MAX, 1); |
4214 |
-+STORE_FUNCTION(bfq_raising_coeff_store, &bfqd->bfq_raising_coeff, 1, |
4215 |
-+ INT_MAX, 0); |
4216 |
-+STORE_FUNCTION(bfq_raising_max_time_store, &bfqd->bfq_raising_max_time, 0, |
4217 |
-+ INT_MAX, 1); |
4218 |
-+STORE_FUNCTION(bfq_raising_rt_max_time_store, &bfqd->bfq_raising_rt_max_time, 0, |
4219 |
-+ INT_MAX, 1); |
4220 |
-+STORE_FUNCTION(bfq_raising_min_idle_time_store, |
4221 |
-+ &bfqd->bfq_raising_min_idle_time, 0, INT_MAX, 1); |
4222 |
-+STORE_FUNCTION(bfq_raising_min_inter_arr_async_store, |
4223 |
-+ &bfqd->bfq_raising_min_inter_arr_async, 0, INT_MAX, 1); |
4224 |
-+STORE_FUNCTION(bfq_raising_max_softrt_rate_store, |
4225 |
-+ &bfqd->bfq_raising_max_softrt_rate, 0, INT_MAX, 0); |
4226 |
-+#undef STORE_FUNCTION |
4227 |
-+ |
4228 |
-+/* do nothing for the moment */ |
4229 |
-+static ssize_t bfq_weights_store(struct elevator_queue *e, |
4230 |
-+ const char *page, size_t count) |
4231 |
-+{ |
4232 |
-+ return count; |
4233 |
-+} |
4234 |
-+ |
4235 |
-+static inline unsigned long bfq_estimated_max_budget(struct bfq_data *bfqd) |
4236 |
-+{ |
4237 |
-+ u64 timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
4238 |
-+ |
4239 |
-+ if (bfqd->peak_rate_samples >= BFQ_PEAK_RATE_SAMPLES) |
4240 |
-+ return bfq_calc_max_budget(bfqd->peak_rate, timeout); |
4241 |
-+ else |
4242 |
-+ return bfq_default_max_budget; |
4243 |
-+} |
4244 |
-+ |
4245 |
-+static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
4246 |
-+ const char *page, size_t count) |
4247 |
-+{ |
4248 |
-+ struct bfq_data *bfqd = e->elevator_data; |
4249 |
-+ unsigned long uninitialized_var(__data); |
4250 |
-+ int ret = bfq_var_store(&__data, (page), count); |
4251 |
-+ |
4252 |
-+ if (__data == 0) |
4253 |
-+ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
4254 |
-+ else { |
4255 |
-+ if (__data > INT_MAX) |
4256 |
-+ __data = INT_MAX; |
4257 |
-+ bfqd->bfq_max_budget = __data; |
4258 |
-+ } |
4259 |
-+ |
4260 |
-+ bfqd->bfq_user_max_budget = __data; |
4261 |
-+ |
4262 |
-+ return ret; |
4263 |
-+} |
4264 |
-+ |
4265 |
-+static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, |
4266 |
-+ const char *page, size_t count) |
4267 |
-+{ |
4268 |
-+ struct bfq_data *bfqd = e->elevator_data; |
4269 |
-+ unsigned long uninitialized_var(__data); |
4270 |
-+ int ret = bfq_var_store(&__data, (page), count); |
4271 |
-+ |
4272 |
-+ if (__data < 1) |
4273 |
-+ __data = 1; |
4274 |
-+ else if (__data > INT_MAX) |
4275 |
-+ __data = INT_MAX; |
4276 |
-+ |
4277 |
-+ bfqd->bfq_timeout[BLK_RW_SYNC] = msecs_to_jiffies(__data); |
4278 |
-+ if (bfqd->bfq_user_max_budget == 0) |
4279 |
-+ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
4280 |
-+ |
4281 |
-+ return ret; |
4282 |
-+} |
4283 |
-+ |
4284 |
-+static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
4285 |
-+ const char *page, size_t count) |
4286 |
-+{ |
4287 |
-+ struct bfq_data *bfqd = e->elevator_data; |
4288 |
-+ unsigned long uninitialized_var(__data); |
4289 |
-+ int ret = bfq_var_store(&__data, (page), count); |
4290 |
-+ |
4291 |
-+ if (__data > 1) |
4292 |
-+ __data = 1; |
4293 |
-+ if (__data == 0 && bfqd->low_latency != 0) |
4294 |
-+ bfq_end_raising(bfqd); |
4295 |
-+ bfqd->low_latency = __data; |
4296 |
-+ |
4297 |
-+ return ret; |
4298 |
-+} |
4299 |
-+ |
4300 |
-+#define BFQ_ATTR(name) \ |
4301 |
-+ __ATTR(name, S_IRUGO|S_IWUSR, bfq_##name##_show, bfq_##name##_store) |
4302 |
-+ |
4303 |
-+static struct elv_fs_entry bfq_attrs[] = { |
4304 |
-+ BFQ_ATTR(quantum), |
4305 |
-+ BFQ_ATTR(fifo_expire_sync), |
4306 |
-+ BFQ_ATTR(fifo_expire_async), |
4307 |
-+ BFQ_ATTR(back_seek_max), |
4308 |
-+ BFQ_ATTR(back_seek_penalty), |
4309 |
-+ BFQ_ATTR(slice_idle), |
4310 |
-+ BFQ_ATTR(max_budget), |
4311 |
-+ BFQ_ATTR(max_budget_async_rq), |
4312 |
-+ BFQ_ATTR(timeout_sync), |
4313 |
-+ BFQ_ATTR(timeout_async), |
4314 |
-+ BFQ_ATTR(low_latency), |
4315 |
-+ BFQ_ATTR(raising_coeff), |
4316 |
-+ BFQ_ATTR(raising_max_time), |
4317 |
-+ BFQ_ATTR(raising_rt_max_time), |
4318 |
-+ BFQ_ATTR(raising_min_idle_time), |
4319 |
-+ BFQ_ATTR(raising_min_inter_arr_async), |
4320 |
-+ BFQ_ATTR(raising_max_softrt_rate), |
4321 |
-+ BFQ_ATTR(weights), |
4322 |
-+ __ATTR_NULL |
4323 |
-+}; |
4324 |
-+ |
4325 |
-+static struct elevator_type iosched_bfq = { |
4326 |
-+ .ops = { |
4327 |
-+ .elevator_merge_fn = bfq_merge, |
4328 |
-+ .elevator_merged_fn = bfq_merged_request, |
4329 |
-+ .elevator_merge_req_fn = bfq_merged_requests, |
4330 |
-+ .elevator_allow_merge_fn = bfq_allow_merge, |
4331 |
-+ .elevator_dispatch_fn = bfq_dispatch_requests, |
4332 |
-+ .elevator_add_req_fn = bfq_insert_request, |
4333 |
-+ .elevator_activate_req_fn = bfq_activate_request, |
4334 |
-+ .elevator_deactivate_req_fn = bfq_deactivate_request, |
4335 |
-+ .elevator_completed_req_fn = bfq_completed_request, |
4336 |
-+ .elevator_former_req_fn = elv_rb_former_request, |
4337 |
-+ .elevator_latter_req_fn = elv_rb_latter_request, |
4338 |
-+ .elevator_init_icq_fn = bfq_init_icq, |
4339 |
-+ .elevator_exit_icq_fn = bfq_exit_icq, |
4340 |
-+ .elevator_set_req_fn = bfq_set_request, |
4341 |
-+ .elevator_put_req_fn = bfq_put_request, |
4342 |
-+ .elevator_may_queue_fn = bfq_may_queue, |
4343 |
-+ .elevator_init_fn = bfq_init_queue, |
4344 |
-+ .elevator_exit_fn = bfq_exit_queue, |
4345 |
-+ }, |
4346 |
-+ .icq_size = sizeof(struct bfq_io_cq), |
4347 |
-+ .icq_align = __alignof__(struct bfq_io_cq), |
4348 |
-+ .elevator_attrs = bfq_attrs, |
4349 |
-+ .elevator_name = "bfq", |
4350 |
-+ .elevator_owner = THIS_MODULE, |
4351 |
-+}; |
4352 |
-+ |
4353 |
-+static int __init bfq_init(void) |
4354 |
-+{ |
4355 |
-+ /* |
4356 |
-+ * Can be 0 on HZ < 1000 setups. |
4357 |
-+ */ |
4358 |
-+ if (bfq_slice_idle == 0) |
4359 |
-+ bfq_slice_idle = 1; |
4360 |
-+ |
4361 |
-+ if (bfq_timeout_async == 0) |
4362 |
-+ bfq_timeout_async = 1; |
4363 |
-+ |
4364 |
-+ if (bfq_slab_setup()) |
4365 |
-+ return -ENOMEM; |
4366 |
-+ |
4367 |
-+ elv_register(&iosched_bfq); |
4368 |
-+ pr_info("BFQ I/O-scheduler version: v7r2"); |
4369 |
-+ |
4370 |
-+ return 0; |
4371 |
-+} |
4372 |
-+ |
4373 |
-+static void __exit bfq_exit(void) |
4374 |
-+{ |
4375 |
-+ elv_unregister(&iosched_bfq); |
4376 |
-+ bfq_slab_kill(); |
4377 |
-+} |
4378 |
-+ |
4379 |
-+module_init(bfq_init); |
4380 |
-+module_exit(bfq_exit); |
4381 |
-+ |
4382 |
-+MODULE_AUTHOR("Fabio Checconi, Paolo Valente"); |
4383 |
-diff --git a/block/bfq-sched.c b/block/bfq-sched.c |
4384 |
-new file mode 100644 |
4385 |
-index 0000000..999b475 |
4386 |
---- /dev/null |
4387 |
-+++ b/block/bfq-sched.c |
4388 |
-@@ -0,0 +1,1078 @@ |
4389 |
-+/* |
4390 |
-+ * BFQ: Hierarchical B-WF2Q+ scheduler. |
4391 |
-+ * |
4392 |
-+ * Based on ideas and code from CFQ: |
4393 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
4394 |
-+ * |
4395 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
4396 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
4397 |
-+ * |
4398 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
4399 |
-+ */ |
4400 |
-+ |
4401 |
-+#ifdef CONFIG_CGROUP_BFQIO |
4402 |
-+#define for_each_entity(entity) \ |
4403 |
-+ for (; entity != NULL; entity = entity->parent) |
4404 |
-+ |
4405 |
-+#define for_each_entity_safe(entity, parent) \ |
4406 |
-+ for (; entity && ({ parent = entity->parent; 1; }); entity = parent) |
4407 |
-+ |
4408 |
-+static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
4409 |
-+ int extract, |
4410 |
-+ struct bfq_data *bfqd); |
4411 |
-+ |
4412 |
-+static inline void bfq_update_budget(struct bfq_entity *next_in_service) |
4413 |
-+{ |
4414 |
-+ struct bfq_entity *bfqg_entity; |
4415 |
-+ struct bfq_group *bfqg; |
4416 |
-+ struct bfq_sched_data *group_sd; |
4417 |
-+ |
4418 |
-+ BUG_ON(next_in_service == NULL); |
4419 |
-+ |
4420 |
-+ group_sd = next_in_service->sched_data; |
4421 |
-+ |
4422 |
-+ bfqg = container_of(group_sd, struct bfq_group, sched_data); |
4423 |
-+ /* |
4424 |
-+ * bfq_group's my_entity field is not NULL only if the group |
4425 |
-+ * is not the root group. We must not touch the root entity |
4426 |
-+ * as it must never become an in-service entity. |
4427 |
-+ */ |
4428 |
-+ bfqg_entity = bfqg->my_entity; |
4429 |
-+ if (bfqg_entity != NULL) |
4430 |
-+ bfqg_entity->budget = next_in_service->budget; |
4431 |
-+} |
4432 |
-+ |
4433 |
-+static int bfq_update_next_in_service(struct bfq_sched_data *sd) |
4434 |
-+{ |
4435 |
-+ struct bfq_entity *next_in_service; |
4436 |
-+ |
4437 |
-+ if (sd->in_service_entity != NULL) |
4438 |
-+ /* will update/requeue at the end of service */ |
4439 |
-+ return 0; |
4440 |
-+ |
4441 |
-+ /* |
4442 |
-+ * NOTE: this can be improved in many ways, such as returning |
4443 |
-+ * 1 (and thus propagating upwards the update) only when the |
4444 |
-+ * budget changes, or caching the bfqq that will be scheduled |
4445 |
-+ * next from this subtree. By now we worry more about |
4446 |
-+ * correctness than about performance... |
4447 |
-+ */ |
4448 |
-+ next_in_service = bfq_lookup_next_entity(sd, 0, NULL); |
4449 |
-+ sd->next_in_service = next_in_service; |
4450 |
-+ |
4451 |
-+ if (next_in_service != NULL) |
4452 |
-+ bfq_update_budget(next_in_service); |
4453 |
-+ |
4454 |
-+ return 1; |
4455 |
-+} |
4456 |
-+ |
4457 |
-+static inline void bfq_check_next_in_service(struct bfq_sched_data *sd, |
4458 |
-+ struct bfq_entity *entity) |
4459 |
-+{ |
4460 |
-+ BUG_ON(sd->next_in_service != entity); |
4461 |
-+} |
4462 |
-+#else |
4463 |
-+#define for_each_entity(entity) \ |
4464 |
-+ for (; entity != NULL; entity = NULL) |
4465 |
-+ |
4466 |
-+#define for_each_entity_safe(entity, parent) \ |
4467 |
-+ for (parent = NULL; entity != NULL; entity = parent) |
4468 |
-+ |
4469 |
-+static inline int bfq_update_next_in_service(struct bfq_sched_data *sd) |
4470 |
-+{ |
4471 |
-+ return 0; |
4472 |
-+} |
4473 |
-+ |
4474 |
-+static inline void bfq_check_next_in_service(struct bfq_sched_data *sd, |
4475 |
-+ struct bfq_entity *entity) |
4476 |
-+{ |
4477 |
-+} |
4478 |
-+ |
4479 |
-+static inline void bfq_update_budget(struct bfq_entity *next_in_service) |
4480 |
-+{ |
4481 |
-+} |
4482 |
-+#endif |
4483 |
-+ |
4484 |
-+/* |
4485 |
-+ * Shift for timestamp calculations. This actually limits the maximum |
4486 |
-+ * service allowed in one timestamp delta (small shift values increase it), |
4487 |
-+ * the maximum total weight that can be used for the queues in the system |
4488 |
-+ * (big shift values increase it), and the period of virtual time wraparounds. |
4489 |
-+ */ |
4490 |
-+#define WFQ_SERVICE_SHIFT 22 |
4491 |
-+ |
4492 |
-+/** |
4493 |
-+ * bfq_gt - compare two timestamps. |
4494 |
-+ * @a: first ts. |
4495 |
-+ * @b: second ts. |
4496 |
-+ * |
4497 |
-+ * Return @a > @b, dealing with wrapping correctly. |
4498 |
-+ */ |
4499 |
-+static inline int bfq_gt(u64 a, u64 b) |
4500 |
-+{ |
4501 |
-+ return (s64)(a - b) > 0; |
4502 |
-+} |
4503 |
-+ |
4504 |
-+static inline struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity) |
4505 |
-+{ |
4506 |
-+ struct bfq_queue *bfqq = NULL; |
4507 |
-+ |
4508 |
-+ BUG_ON(entity == NULL); |
4509 |
-+ |
4510 |
-+ if (entity->my_sched_data == NULL) |
4511 |
-+ bfqq = container_of(entity, struct bfq_queue, entity); |
4512 |
-+ |
4513 |
-+ return bfqq; |
4514 |
-+} |
4515 |
-+ |
4516 |
-+ |
4517 |
-+/** |
4518 |
-+ * bfq_delta - map service into the virtual time domain. |
4519 |
-+ * @service: amount of service. |
4520 |
-+ * @weight: scale factor (weight of an entity or weight sum). |
4521 |
-+ */ |
4522 |
-+static inline u64 bfq_delta(unsigned long service, |
4523 |
-+ unsigned long weight) |
4524 |
-+{ |
4525 |
-+ u64 d = (u64)service << WFQ_SERVICE_SHIFT; |
4526 |
-+ |
4527 |
-+ do_div(d, weight); |
4528 |
-+ return d; |
4529 |
-+} |
4530 |
-+ |
4531 |
-+/** |
4532 |
-+ * bfq_calc_finish - assign the finish time to an entity. |
4533 |
-+ * @entity: the entity to act upon. |
4534 |
-+ * @service: the service to be charged to the entity. |
4535 |
-+ */ |
4536 |
-+static inline void bfq_calc_finish(struct bfq_entity *entity, |
4537 |
-+ unsigned long service) |
4538 |
-+{ |
4539 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4540 |
-+ |
4541 |
-+ BUG_ON(entity->weight == 0); |
4542 |
-+ |
4543 |
-+ entity->finish = entity->start + |
4544 |
-+ bfq_delta(service, entity->weight); |
4545 |
-+ |
4546 |
-+ if (bfqq != NULL) { |
4547 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, |
4548 |
-+ "calc_finish: serv %lu, w %d", |
4549 |
-+ service, entity->weight); |
4550 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, |
4551 |
-+ "calc_finish: start %llu, finish %llu, delta %llu", |
4552 |
-+ entity->start, entity->finish, |
4553 |
-+ bfq_delta(service, entity->weight)); |
4554 |
-+ } |
4555 |
-+} |
4556 |
-+ |
4557 |
-+/** |
4558 |
-+ * bfq_entity_of - get an entity from a node. |
4559 |
-+ * @node: the node field of the entity. |
4560 |
-+ * |
4561 |
-+ * Convert a node pointer to the relative entity. This is used only |
4562 |
-+ * to simplify the logic of some functions and not as the generic |
4563 |
-+ * conversion mechanism because, e.g., in the tree walking functions, |
4564 |
-+ * the check for a %NULL value would be redundant. |
4565 |
-+ */ |
4566 |
-+static inline struct bfq_entity *bfq_entity_of(struct rb_node *node) |
4567 |
-+{ |
4568 |
-+ struct bfq_entity *entity = NULL; |
4569 |
-+ |
4570 |
-+ if (node != NULL) |
4571 |
-+ entity = rb_entry(node, struct bfq_entity, rb_node); |
4572 |
-+ |
4573 |
-+ return entity; |
4574 |
-+} |
4575 |
-+ |
4576 |
-+/** |
4577 |
-+ * bfq_extract - remove an entity from a tree. |
4578 |
-+ * @root: the tree root. |
4579 |
-+ * @entity: the entity to remove. |
4580 |
-+ */ |
4581 |
-+static inline void bfq_extract(struct rb_root *root, |
4582 |
-+ struct bfq_entity *entity) |
4583 |
-+{ |
4584 |
-+ BUG_ON(entity->tree != root); |
4585 |
-+ |
4586 |
-+ entity->tree = NULL; |
4587 |
-+ rb_erase(&entity->rb_node, root); |
4588 |
-+} |
4589 |
-+ |
4590 |
-+/** |
4591 |
-+ * bfq_idle_extract - extract an entity from the idle tree. |
4592 |
-+ * @st: the service tree of the owning @entity. |
4593 |
-+ * @entity: the entity being removed. |
4594 |
-+ */ |
4595 |
-+static void bfq_idle_extract(struct bfq_service_tree *st, |
4596 |
-+ struct bfq_entity *entity) |
4597 |
-+{ |
4598 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4599 |
-+ struct rb_node *next; |
4600 |
-+ |
4601 |
-+ BUG_ON(entity->tree != &st->idle); |
4602 |
-+ |
4603 |
-+ if (entity == st->first_idle) { |
4604 |
-+ next = rb_next(&entity->rb_node); |
4605 |
-+ st->first_idle = bfq_entity_of(next); |
4606 |
-+ } |
4607 |
-+ |
4608 |
-+ if (entity == st->last_idle) { |
4609 |
-+ next = rb_prev(&entity->rb_node); |
4610 |
-+ st->last_idle = bfq_entity_of(next); |
4611 |
-+ } |
4612 |
-+ |
4613 |
-+ bfq_extract(&st->idle, entity); |
4614 |
-+ |
4615 |
-+ if (bfqq != NULL) |
4616 |
-+ list_del(&bfqq->bfqq_list); |
4617 |
-+} |
4618 |
-+ |
4619 |
-+/** |
4620 |
-+ * bfq_insert - generic tree insertion. |
4621 |
-+ * @root: tree root. |
4622 |
-+ * @entity: entity to insert. |
4623 |
-+ * |
4624 |
-+ * This is used for the idle and the active tree, since they are both |
4625 |
-+ * ordered by finish time. |
4626 |
-+ */ |
4627 |
-+static void bfq_insert(struct rb_root *root, struct bfq_entity *entity) |
4628 |
-+{ |
4629 |
-+ struct bfq_entity *entry; |
4630 |
-+ struct rb_node **node = &root->rb_node; |
4631 |
-+ struct rb_node *parent = NULL; |
4632 |
-+ |
4633 |
-+ BUG_ON(entity->tree != NULL); |
4634 |
-+ |
4635 |
-+ while (*node != NULL) { |
4636 |
-+ parent = *node; |
4637 |
-+ entry = rb_entry(parent, struct bfq_entity, rb_node); |
4638 |
-+ |
4639 |
-+ if (bfq_gt(entry->finish, entity->finish)) |
4640 |
-+ node = &parent->rb_left; |
4641 |
-+ else |
4642 |
-+ node = &parent->rb_right; |
4643 |
-+ } |
4644 |
-+ |
4645 |
-+ rb_link_node(&entity->rb_node, parent, node); |
4646 |
-+ rb_insert_color(&entity->rb_node, root); |
4647 |
-+ |
4648 |
-+ entity->tree = root; |
4649 |
-+} |
4650 |
-+ |
4651 |
-+/** |
4652 |
-+ * bfq_update_min - update the min_start field of a entity. |
4653 |
-+ * @entity: the entity to update. |
4654 |
-+ * @node: one of its children. |
4655 |
-+ * |
4656 |
-+ * This function is called when @entity may store an invalid value for |
4657 |
-+ * min_start due to updates to the active tree. The function assumes |
4658 |
-+ * that the subtree rooted at @node (which may be its left or its right |
4659 |
-+ * child) has a valid min_start value. |
4660 |
-+ */ |
4661 |
-+static inline void bfq_update_min(struct bfq_entity *entity, |
4662 |
-+ struct rb_node *node) |
4663 |
-+{ |
4664 |
-+ struct bfq_entity *child; |
4665 |
-+ |
4666 |
-+ if (node != NULL) { |
4667 |
-+ child = rb_entry(node, struct bfq_entity, rb_node); |
4668 |
-+ if (bfq_gt(entity->min_start, child->min_start)) |
4669 |
-+ entity->min_start = child->min_start; |
4670 |
-+ } |
4671 |
-+} |
4672 |
-+ |
4673 |
-+/** |
4674 |
-+ * bfq_update_active_node - recalculate min_start. |
4675 |
-+ * @node: the node to update. |
4676 |
-+ * |
4677 |
-+ * @node may have changed position or one of its children may have moved, |
4678 |
-+ * this function updates its min_start value. The left and right subtrees |
4679 |
-+ * are assumed to hold a correct min_start value. |
4680 |
-+ */ |
4681 |
-+static inline void bfq_update_active_node(struct rb_node *node) |
4682 |
-+{ |
4683 |
-+ struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node); |
4684 |
-+ |
4685 |
-+ entity->min_start = entity->start; |
4686 |
-+ bfq_update_min(entity, node->rb_right); |
4687 |
-+ bfq_update_min(entity, node->rb_left); |
4688 |
-+} |
4689 |
-+ |
4690 |
-+/** |
4691 |
-+ * bfq_update_active_tree - update min_start for the whole active tree. |
4692 |
-+ * @node: the starting node. |
4693 |
-+ * |
4694 |
-+ * @node must be the deepest modified node after an update. This function |
4695 |
-+ * updates its min_start using the values held by its children, assuming |
4696 |
-+ * that they did not change, and then updates all the nodes that may have |
4697 |
-+ * changed in the path to the root. The only nodes that may have changed |
4698 |
-+ * are the ones in the path or their siblings. |
4699 |
-+ */ |
4700 |
-+static void bfq_update_active_tree(struct rb_node *node) |
4701 |
-+{ |
4702 |
-+ struct rb_node *parent; |
4703 |
-+ |
4704 |
-+up: |
4705 |
-+ bfq_update_active_node(node); |
4706 |
-+ |
4707 |
-+ parent = rb_parent(node); |
4708 |
-+ if (parent == NULL) |
4709 |
-+ return; |
4710 |
-+ |
4711 |
-+ if (node == parent->rb_left && parent->rb_right != NULL) |
4712 |
-+ bfq_update_active_node(parent->rb_right); |
4713 |
-+ else if (parent->rb_left != NULL) |
4714 |
-+ bfq_update_active_node(parent->rb_left); |
4715 |
-+ |
4716 |
-+ node = parent; |
4717 |
-+ goto up; |
4718 |
-+} |
4719 |
-+ |
4720 |
-+/** |
4721 |
-+ * bfq_active_insert - insert an entity in the active tree of its group/device. |
4722 |
-+ * @st: the service tree of the entity. |
4723 |
-+ * @entity: the entity being inserted. |
4724 |
-+ * |
4725 |
-+ * The active tree is ordered by finish time, but an extra key is kept |
4726 |
-+ * per each node, containing the minimum value for the start times of |
4727 |
-+ * its children (and the node itself), so it's possible to search for |
4728 |
-+ * the eligible node with the lowest finish time in logarithmic time. |
4729 |
-+ */ |
4730 |
-+static void bfq_active_insert(struct bfq_service_tree *st, |
4731 |
-+ struct bfq_entity *entity) |
4732 |
-+{ |
4733 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4734 |
-+ struct rb_node *node = &entity->rb_node; |
4735 |
-+ |
4736 |
-+ bfq_insert(&st->active, entity); |
4737 |
-+ |
4738 |
-+ if (node->rb_left != NULL) |
4739 |
-+ node = node->rb_left; |
4740 |
-+ else if (node->rb_right != NULL) |
4741 |
-+ node = node->rb_right; |
4742 |
-+ |
4743 |
-+ bfq_update_active_tree(node); |
4744 |
-+ |
4745 |
-+ if (bfqq != NULL) |
4746 |
-+ list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list); |
4747 |
-+} |
4748 |
-+ |
4749 |
-+/** |
4750 |
-+ * bfq_ioprio_to_weight - calc a weight from an ioprio. |
4751 |
-+ * @ioprio: the ioprio value to convert. |
4752 |
-+ */ |
4753 |
-+static unsigned short bfq_ioprio_to_weight(int ioprio) |
4754 |
-+{ |
4755 |
-+ WARN_ON(ioprio < 0 || ioprio >= IOPRIO_BE_NR); |
4756 |
-+ return IOPRIO_BE_NR - ioprio; |
4757 |
-+} |
4758 |
-+ |
4759 |
-+/** |
4760 |
-+ * bfq_weight_to_ioprio - calc an ioprio from a weight. |
4761 |
-+ * @weight: the weight value to convert. |
4762 |
-+ * |
4763 |
-+ * To preserve as mush as possible the old only-ioprio user interface, |
4764 |
-+ * 0 is used as an escape ioprio value for weights (numerically) equal or |
4765 |
-+ * larger than IOPRIO_BE_NR |
4766 |
-+ */ |
4767 |
-+static unsigned short bfq_weight_to_ioprio(int weight) |
4768 |
-+{ |
4769 |
-+ WARN_ON(weight < BFQ_MIN_WEIGHT || weight > BFQ_MAX_WEIGHT); |
4770 |
-+ return IOPRIO_BE_NR - weight < 0 ? 0 : IOPRIO_BE_NR - weight; |
4771 |
-+} |
4772 |
-+ |
4773 |
-+static inline void bfq_get_entity(struct bfq_entity *entity) |
4774 |
-+{ |
4775 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4776 |
-+ struct bfq_sched_data *sd; |
4777 |
-+ |
4778 |
-+ if (bfqq != NULL) { |
4779 |
-+ sd = entity->sched_data; |
4780 |
-+ atomic_inc(&bfqq->ref); |
4781 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d", |
4782 |
-+ bfqq, atomic_read(&bfqq->ref)); |
4783 |
-+ } |
4784 |
-+} |
4785 |
-+ |
4786 |
-+/** |
4787 |
-+ * bfq_find_deepest - find the deepest node that an extraction can modify. |
4788 |
-+ * @node: the node being removed. |
4789 |
-+ * |
4790 |
-+ * Do the first step of an extraction in an rb tree, looking for the |
4791 |
-+ * node that will replace @node, and returning the deepest node that |
4792 |
-+ * the following modifications to the tree can touch. If @node is the |
4793 |
-+ * last node in the tree return %NULL. |
4794 |
-+ */ |
4795 |
-+static struct rb_node *bfq_find_deepest(struct rb_node *node) |
4796 |
-+{ |
4797 |
-+ struct rb_node *deepest; |
4798 |
-+ |
4799 |
-+ if (node->rb_right == NULL && node->rb_left == NULL) |
4800 |
-+ deepest = rb_parent(node); |
4801 |
-+ else if (node->rb_right == NULL) |
4802 |
-+ deepest = node->rb_left; |
4803 |
-+ else if (node->rb_left == NULL) |
4804 |
-+ deepest = node->rb_right; |
4805 |
-+ else { |
4806 |
-+ deepest = rb_next(node); |
4807 |
-+ if (deepest->rb_right != NULL) |
4808 |
-+ deepest = deepest->rb_right; |
4809 |
-+ else if (rb_parent(deepest) != node) |
4810 |
-+ deepest = rb_parent(deepest); |
4811 |
-+ } |
4812 |
-+ |
4813 |
-+ return deepest; |
4814 |
-+} |
4815 |
-+ |
4816 |
-+/** |
4817 |
-+ * bfq_active_extract - remove an entity from the active tree. |
4818 |
-+ * @st: the service_tree containing the tree. |
4819 |
-+ * @entity: the entity being removed. |
4820 |
-+ */ |
4821 |
-+static void bfq_active_extract(struct bfq_service_tree *st, |
4822 |
-+ struct bfq_entity *entity) |
4823 |
-+{ |
4824 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4825 |
-+ struct rb_node *node; |
4826 |
-+ |
4827 |
-+ node = bfq_find_deepest(&entity->rb_node); |
4828 |
-+ bfq_extract(&st->active, entity); |
4829 |
-+ |
4830 |
-+ if (node != NULL) |
4831 |
-+ bfq_update_active_tree(node); |
4832 |
-+ |
4833 |
-+ if (bfqq != NULL) |
4834 |
-+ list_del(&bfqq->bfqq_list); |
4835 |
-+} |
4836 |
-+ |
4837 |
-+/** |
4838 |
-+ * bfq_idle_insert - insert an entity into the idle tree. |
4839 |
-+ * @st: the service tree containing the tree. |
4840 |
-+ * @entity: the entity to insert. |
4841 |
-+ */ |
4842 |
-+static void bfq_idle_insert(struct bfq_service_tree *st, |
4843 |
-+ struct bfq_entity *entity) |
4844 |
-+{ |
4845 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4846 |
-+ struct bfq_entity *first_idle = st->first_idle; |
4847 |
-+ struct bfq_entity *last_idle = st->last_idle; |
4848 |
-+ |
4849 |
-+ if (first_idle == NULL || bfq_gt(first_idle->finish, entity->finish)) |
4850 |
-+ st->first_idle = entity; |
4851 |
-+ if (last_idle == NULL || bfq_gt(entity->finish, last_idle->finish)) |
4852 |
-+ st->last_idle = entity; |
4853 |
-+ |
4854 |
-+ bfq_insert(&st->idle, entity); |
4855 |
-+ |
4856 |
-+ if (bfqq != NULL) |
4857 |
-+ list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list); |
4858 |
-+} |
4859 |
-+ |
4860 |
-+/** |
4861 |
-+ * bfq_forget_entity - remove an entity from the wfq trees. |
4862 |
-+ * @st: the service tree. |
4863 |
-+ * @entity: the entity being removed. |
4864 |
-+ * |
4865 |
-+ * Update the device status and forget everything about @entity, putting |
4866 |
-+ * the device reference to it, if it is a queue. Entities belonging to |
4867 |
-+ * groups are not refcounted. |
4868 |
-+ */ |
4869 |
-+static void bfq_forget_entity(struct bfq_service_tree *st, |
4870 |
-+ struct bfq_entity *entity) |
4871 |
-+{ |
4872 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4873 |
-+ struct bfq_sched_data *sd; |
4874 |
-+ |
4875 |
-+ BUG_ON(!entity->on_st); |
4876 |
-+ |
4877 |
-+ entity->on_st = 0; |
4878 |
-+ st->wsum -= entity->weight; |
4879 |
-+ if (bfqq != NULL) { |
4880 |
-+ sd = entity->sched_data; |
4881 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "forget_entity: %p %d", |
4882 |
-+ bfqq, atomic_read(&bfqq->ref)); |
4883 |
-+ bfq_put_queue(bfqq); |
4884 |
-+ } |
4885 |
-+} |
4886 |
-+ |
4887 |
-+/** |
4888 |
-+ * bfq_put_idle_entity - release the idle tree ref of an entity. |
4889 |
-+ * @st: service tree for the entity. |
4890 |
-+ * @entity: the entity being released. |
4891 |
-+ */ |
4892 |
-+static void bfq_put_idle_entity(struct bfq_service_tree *st, |
4893 |
-+ struct bfq_entity *entity) |
4894 |
-+{ |
4895 |
-+ bfq_idle_extract(st, entity); |
4896 |
-+ bfq_forget_entity(st, entity); |
4897 |
-+} |
4898 |
-+ |
4899 |
-+/** |
4900 |
-+ * bfq_forget_idle - update the idle tree if necessary. |
4901 |
-+ * @st: the service tree to act upon. |
4902 |
-+ * |
4903 |
-+ * To preserve the global O(log N) complexity we only remove one entry here; |
4904 |
-+ * as the idle tree will not grow indefinitely this can be done safely. |
4905 |
-+ */ |
4906 |
-+static void bfq_forget_idle(struct bfq_service_tree *st) |
4907 |
-+{ |
4908 |
-+ struct bfq_entity *first_idle = st->first_idle; |
4909 |
-+ struct bfq_entity *last_idle = st->last_idle; |
4910 |
-+ |
4911 |
-+ if (RB_EMPTY_ROOT(&st->active) && last_idle != NULL && |
4912 |
-+ !bfq_gt(last_idle->finish, st->vtime)) { |
4913 |
-+ /* |
4914 |
-+ * Forget the whole idle tree, increasing the vtime past |
4915 |
-+ * the last finish time of idle entities. |
4916 |
-+ */ |
4917 |
-+ st->vtime = last_idle->finish; |
4918 |
-+ } |
4919 |
-+ |
4920 |
-+ if (first_idle != NULL && !bfq_gt(first_idle->finish, st->vtime)) |
4921 |
-+ bfq_put_idle_entity(st, first_idle); |
4922 |
-+} |
4923 |
-+ |
4924 |
-+static struct bfq_service_tree * |
4925 |
-+__bfq_entity_update_weight_prio(struct bfq_service_tree *old_st, |
4926 |
-+ struct bfq_entity *entity) |
4927 |
-+{ |
4928 |
-+ struct bfq_service_tree *new_st = old_st; |
4929 |
-+ |
4930 |
-+ if (entity->ioprio_changed) { |
4931 |
-+ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
4932 |
-+ |
4933 |
-+ BUG_ON(old_st->wsum < entity->weight); |
4934 |
-+ old_st->wsum -= entity->weight; |
4935 |
-+ |
4936 |
-+ if (entity->new_weight != entity->orig_weight) { |
4937 |
-+ entity->orig_weight = entity->new_weight; |
4938 |
-+ entity->ioprio = |
4939 |
-+ bfq_weight_to_ioprio(entity->orig_weight); |
4940 |
-+ } else if (entity->new_ioprio != entity->ioprio) { |
4941 |
-+ entity->ioprio = entity->new_ioprio; |
4942 |
-+ entity->orig_weight = |
4943 |
-+ bfq_ioprio_to_weight(entity->ioprio); |
4944 |
-+ } else |
4945 |
-+ entity->new_weight = entity->orig_weight = |
4946 |
-+ bfq_ioprio_to_weight(entity->ioprio); |
4947 |
-+ |
4948 |
-+ entity->ioprio_class = entity->new_ioprio_class; |
4949 |
-+ entity->ioprio_changed = 0; |
4950 |
-+ |
4951 |
-+ /* |
4952 |
-+ * NOTE: here we may be changing the weight too early, |
4953 |
-+ * this will cause unfairness. The correct approach |
4954 |
-+ * would have required additional complexity to defer |
4955 |
-+ * weight changes to the proper time instants (i.e., |
4956 |
-+ * when entity->finish <= old_st->vtime). |
4957 |
-+ */ |
4958 |
-+ new_st = bfq_entity_service_tree(entity); |
4959 |
-+ entity->weight = entity->orig_weight * |
4960 |
-+ (bfqq != NULL ? bfqq->raising_coeff : 1); |
4961 |
-+ new_st->wsum += entity->weight; |
4962 |
-+ |
4963 |
-+ if (new_st != old_st) |
4964 |
-+ entity->start = new_st->vtime; |
4965 |
-+ } |
4966 |
-+ |
4967 |
-+ return new_st; |
4968 |
-+} |
4969 |
-+ |
4970 |
-+/** |
4971 |
-+ * bfq_bfqq_served - update the scheduler status after selection for service. |
4972 |
-+ * @bfqq: the queue being served. |
4973 |
-+ * @served: bytes to transfer. |
4974 |
-+ * |
4975 |
-+ * NOTE: this can be optimized, as the timestamps of upper level entities |
4976 |
-+ * are synchronized every time a new bfqq is selected for service. By now, |
4977 |
-+ * we keep it to better check consistency. |
4978 |
-+ */ |
4979 |
-+static void bfq_bfqq_served(struct bfq_queue *bfqq, unsigned long served) |
4980 |
-+{ |
4981 |
-+ struct bfq_entity *entity = &bfqq->entity; |
4982 |
-+ struct bfq_service_tree *st; |
4983 |
-+ |
4984 |
-+ for_each_entity(entity) { |
4985 |
-+ st = bfq_entity_service_tree(entity); |
4986 |
-+ |
4987 |
-+ entity->service += served; |
4988 |
-+ BUG_ON(entity->service > entity->budget); |
4989 |
-+ BUG_ON(st->wsum == 0); |
4990 |
-+ |
4991 |
-+ st->vtime += bfq_delta(served, st->wsum); |
4992 |
-+ bfq_forget_idle(st); |
4993 |
-+ } |
4994 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %lu secs", served); |
4995 |
-+} |
4996 |
-+ |
4997 |
-+/** |
4998 |
-+ * bfq_bfqq_charge_full_budget - set the service to the entity budget. |
4999 |
-+ * @bfqq: the queue that needs a service update. |
5000 |
-+ * |
5001 |
-+ * When it's not possible to be fair in the service domain, because |
5002 |
-+ * a queue is not consuming its budget fast enough (the meaning of |
5003 |
-+ * fast depends on the timeout parameter), we charge it a full |
5004 |
-+ * budget. In this way we should obtain a sort of time-domain |
5005 |
-+ * fairness among all the seeky/slow queues. |
5006 |
-+ */ |
5007 |
-+static inline void bfq_bfqq_charge_full_budget(struct bfq_queue *bfqq) |
5008 |
-+{ |
5009 |
-+ struct bfq_entity *entity = &bfqq->entity; |
5010 |
-+ |
5011 |
-+ bfq_log_bfqq(bfqq->bfqd, bfqq, "charge_full_budget"); |
5012 |
-+ |
5013 |
-+ bfq_bfqq_served(bfqq, entity->budget - entity->service); |
5014 |
-+} |
5015 |
-+ |
5016 |
-+/** |
5017 |
-+ * __bfq_activate_entity - activate an entity. |
5018 |
-+ * @entity: the entity being activated. |
5019 |
-+ * |
5020 |
-+ * Called whenever an entity is activated, i.e., it is not active and one |
5021 |
-+ * of its children receives a new request, or has to be reactivated due to |
5022 |
-+ * budget exhaustion. It uses the current budget of the entity (and the |
5023 |
-+ * service received if @entity is active) of the queue to calculate its |
5024 |
-+ * timestamps. |
5025 |
-+ */ |
5026 |
-+static void __bfq_activate_entity(struct bfq_entity *entity) |
5027 |
-+{ |
5028 |
-+ struct bfq_sched_data *sd = entity->sched_data; |
5029 |
-+ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
5030 |
-+ |
5031 |
-+ if (entity == sd->in_service_entity) { |
5032 |
-+ BUG_ON(entity->tree != NULL); |
5033 |
-+ /* |
5034 |
-+ * If we are requeueing the current entity we have |
5035 |
-+ * to take care of not charging to it service it has |
5036 |
-+ * not received. |
5037 |
-+ */ |
5038 |
-+ bfq_calc_finish(entity, entity->service); |
5039 |
-+ entity->start = entity->finish; |
5040 |
-+ sd->in_service_entity = NULL; |
5041 |
-+ } else if (entity->tree == &st->active) { |
5042 |
-+ /* |
5043 |
-+ * Requeueing an entity due to a change of some |
5044 |
-+ * next_in_service entity below it. We reuse the |
5045 |
-+ * old start time. |
5046 |
-+ */ |
5047 |
-+ bfq_active_extract(st, entity); |
5048 |
-+ } else if (entity->tree == &st->idle) { |
5049 |
-+ /* |
5050 |
-+ * Must be on the idle tree, bfq_idle_extract() will |
5051 |
-+ * check for that. |
5052 |
-+ */ |
5053 |
-+ bfq_idle_extract(st, entity); |
5054 |
-+ entity->start = bfq_gt(st->vtime, entity->finish) ? |
5055 |
-+ st->vtime : entity->finish; |
5056 |
-+ } else { |
5057 |
-+ /* |
5058 |
-+ * The finish time of the entity may be invalid, and |
5059 |
-+ * it is in the past for sure, otherwise the queue |
5060 |
-+ * would have been on the idle tree. |
5061 |
-+ */ |
5062 |
-+ entity->start = st->vtime; |
5063 |
-+ st->wsum += entity->weight; |
5064 |
-+ bfq_get_entity(entity); |
5065 |
-+ |
5066 |
-+ BUG_ON(entity->on_st); |
5067 |
-+ entity->on_st = 1; |
5068 |
-+ } |
5069 |
-+ |
5070 |
-+ st = __bfq_entity_update_weight_prio(st, entity); |
5071 |
-+ bfq_calc_finish(entity, entity->budget); |
5072 |
-+ bfq_active_insert(st, entity); |
5073 |
-+} |
5074 |
-+ |
5075 |
-+/** |
5076 |
-+ * bfq_activate_entity - activate an entity and its ancestors if necessary. |
5077 |
-+ * @entity: the entity to activate. |
5078 |
-+ * |
5079 |
-+ * Activate @entity and all the entities on the path from it to the root. |
5080 |
-+ */ |
5081 |
-+static void bfq_activate_entity(struct bfq_entity *entity) |
5082 |
-+{ |
5083 |
-+ struct bfq_sched_data *sd; |
5084 |
-+ |
5085 |
-+ for_each_entity(entity) { |
5086 |
-+ __bfq_activate_entity(entity); |
5087 |
-+ |
5088 |
-+ sd = entity->sched_data; |
5089 |
-+ if (!bfq_update_next_in_service(sd)) |
5090 |
-+ /* |
5091 |
-+ * No need to propagate the activation to the |
5092 |
-+ * upper entities, as they will be updated when |
5093 |
-+ * the in-service entity is rescheduled. |
5094 |
-+ */ |
5095 |
-+ break; |
5096 |
-+ } |
5097 |
-+} |
5098 |
-+ |
5099 |
-+/** |
5100 |
-+ * __bfq_deactivate_entity - deactivate an entity from its service tree. |
5101 |
-+ * @entity: the entity to deactivate. |
5102 |
-+ * @requeue: if false, the entity will not be put into the idle tree. |
5103 |
-+ * |
5104 |
-+ * Deactivate an entity, independently from its previous state. If the |
5105 |
-+ * entity was not on a service tree just return, otherwise if it is on |
5106 |
-+ * any scheduler tree, extract it from that tree, and if necessary |
5107 |
-+ * and if the caller did not specify @requeue, put it on the idle tree. |
5108 |
-+ * |
5109 |
-+ * Return %1 if the caller should update the entity hierarchy, i.e., |
5110 |
-+ * if the entity was under service or if it was the next_in_service for |
5111 |
-+ * its sched_data; return %0 otherwise. |
5112 |
-+ */ |
5113 |
-+static int __bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
5114 |
-+{ |
5115 |
-+ struct bfq_sched_data *sd = entity->sched_data; |
5116 |
-+ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
5117 |
-+ int was_in_service = entity == sd->in_service_entity; |
5118 |
-+ int ret = 0; |
5119 |
-+ |
5120 |
-+ if (!entity->on_st) |
5121 |
-+ return 0; |
5122 |
-+ |
5123 |
-+ BUG_ON(was_in_service && entity->tree != NULL); |
5124 |
-+ |
5125 |
-+ if (was_in_service) { |
5126 |
-+ bfq_calc_finish(entity, entity->service); |
5127 |
-+ sd->in_service_entity = NULL; |
5128 |
-+ } else if (entity->tree == &st->active) |
5129 |
-+ bfq_active_extract(st, entity); |
5130 |
-+ else if (entity->tree == &st->idle) |
5131 |
-+ bfq_idle_extract(st, entity); |
5132 |
-+ else if (entity->tree != NULL) |
5133 |
-+ BUG(); |
5134 |
-+ |
5135 |
-+ if (was_in_service || sd->next_in_service == entity) |
5136 |
-+ ret = bfq_update_next_in_service(sd); |
5137 |
-+ |
5138 |
-+ if (!requeue || !bfq_gt(entity->finish, st->vtime)) |
5139 |
-+ bfq_forget_entity(st, entity); |
5140 |
-+ else |
5141 |
-+ bfq_idle_insert(st, entity); |
5142 |
-+ |
5143 |
-+ BUG_ON(sd->in_service_entity == entity); |
5144 |
-+ BUG_ON(sd->next_in_service == entity); |
5145 |
-+ |
5146 |
-+ return ret; |
5147 |
-+} |
5148 |
-+ |
5149 |
-+/** |
5150 |
-+ * bfq_deactivate_entity - deactivate an entity. |
5151 |
-+ * @entity: the entity to deactivate. |
5152 |
-+ * @requeue: true if the entity can be put on the idle tree |
5153 |
-+ */ |
5154 |
-+static void bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
5155 |
-+{ |
5156 |
-+ struct bfq_sched_data *sd; |
5157 |
-+ struct bfq_entity *parent; |
5158 |
-+ |
5159 |
-+ for_each_entity_safe(entity, parent) { |
5160 |
-+ sd = entity->sched_data; |
5161 |
-+ |
5162 |
-+ if (!__bfq_deactivate_entity(entity, requeue)) |
5163 |
-+ /* |
5164 |
-+ * The parent entity is still backlogged, and |
5165 |
-+ * we don't need to update it as it is still |
5166 |
-+ * under service. |
5167 |
-+ */ |
5168 |
-+ break; |
5169 |
-+ |
5170 |
-+ if (sd->next_in_service != NULL) |
5171 |
-+ /* |
5172 |
-+ * The parent entity is still backlogged and |
5173 |
-+ * the budgets on the path towards the root |
5174 |
-+ * need to be updated. |
5175 |
-+ */ |
5176 |
-+ goto update; |
5177 |
-+ |
5178 |
-+ /* |
5179 |
-+ * If we reach there the parent is no more backlogged and |
5180 |
-+ * we want to propagate the dequeue upwards. |
5181 |
-+ */ |
5182 |
-+ requeue = 1; |
5183 |
-+ } |
5184 |
-+ |
5185 |
-+ return; |
5186 |
-+ |
5187 |
-+update: |
5188 |
-+ entity = parent; |
5189 |
-+ for_each_entity(entity) { |
5190 |
-+ __bfq_activate_entity(entity); |
5191 |
-+ |
5192 |
-+ sd = entity->sched_data; |
5193 |
-+ if (!bfq_update_next_in_service(sd)) |
5194 |
-+ break; |
5195 |
-+ } |
5196 |
-+} |
5197 |
-+ |
5198 |
-+/** |
5199 |
-+ * bfq_update_vtime - update vtime if necessary. |
5200 |
-+ * @st: the service tree to act upon. |
5201 |
-+ * |
5202 |
-+ * If necessary update the service tree vtime to have at least one |
5203 |
-+ * eligible entity, skipping to its start time. Assumes that the |
5204 |
-+ * active tree of the device is not empty. |
5205 |
-+ * |
5206 |
-+ * NOTE: this hierarchical implementation updates vtimes quite often, |
5207 |
-+ * we may end up with reactivated tasks getting timestamps after a |
5208 |
-+ * vtime skip done because we needed a ->first_active entity on some |
5209 |
-+ * intermediate node. |
5210 |
-+ */ |
5211 |
-+static void bfq_update_vtime(struct bfq_service_tree *st) |
5212 |
-+{ |
5213 |
-+ struct bfq_entity *entry; |
5214 |
-+ struct rb_node *node = st->active.rb_node; |
5215 |
-+ |
5216 |
-+ entry = rb_entry(node, struct bfq_entity, rb_node); |
5217 |
-+ if (bfq_gt(entry->min_start, st->vtime)) { |
5218 |
-+ st->vtime = entry->min_start; |
5219 |
-+ bfq_forget_idle(st); |
5220 |
-+ } |
5221 |
-+} |
5222 |
-+ |
5223 |
-+/** |
5224 |
-+ * bfq_first_active_entity - find the eligible entity with |
5225 |
-+ * the smallest finish time |
5226 |
-+ * @st: the service tree to select from. |
5227 |
-+ * |
5228 |
-+ * This function searches the first schedulable entity, starting from the |
5229 |
-+ * root of the tree and going on the left every time on this side there is |
5230 |
-+ * a subtree with at least one eligible (start >= vtime) entity. The path |
5231 |
-+ * on the right is followed only if a) the left subtree contains no eligible |
5232 |
-+ * entities and b) no eligible entity has been found yet. |
5233 |
-+ */ |
5234 |
-+static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st) |
5235 |
-+{ |
5236 |
-+ struct bfq_entity *entry, *first = NULL; |
5237 |
-+ struct rb_node *node = st->active.rb_node; |
5238 |
-+ |
5239 |
-+ while (node != NULL) { |
5240 |
-+ entry = rb_entry(node, struct bfq_entity, rb_node); |
5241 |
-+left: |
5242 |
-+ if (!bfq_gt(entry->start, st->vtime)) |
5243 |
-+ first = entry; |
5244 |
-+ |
5245 |
-+ BUG_ON(bfq_gt(entry->min_start, st->vtime)); |
5246 |
-+ |
5247 |
-+ if (node->rb_left != NULL) { |
5248 |
-+ entry = rb_entry(node->rb_left, |
5249 |
-+ struct bfq_entity, rb_node); |
5250 |
-+ if (!bfq_gt(entry->min_start, st->vtime)) { |
5251 |
-+ node = node->rb_left; |
5252 |
-+ goto left; |
5253 |
-+ } |
5254 |
-+ } |
5255 |
-+ if (first != NULL) |
5256 |
-+ break; |
5257 |
-+ node = node->rb_right; |
5258 |
-+ } |
5259 |
-+ |
5260 |
-+ BUG_ON(first == NULL && !RB_EMPTY_ROOT(&st->active)); |
5261 |
-+ return first; |
5262 |
-+} |
5263 |
-+ |
5264 |
-+/** |
5265 |
-+ * __bfq_lookup_next_entity - return the first eligible entity in @st. |
5266 |
-+ * @st: the service tree. |
5267 |
-+ * |
5268 |
-+ * Update the virtual time in @st and return the first eligible entity |
5269 |
-+ * it contains. |
5270 |
-+ */ |
5271 |
-+static struct bfq_entity *__bfq_lookup_next_entity(struct bfq_service_tree *st, |
5272 |
-+ bool force) |
5273 |
-+{ |
5274 |
-+ struct bfq_entity *entity, *new_next_in_service = NULL; |
5275 |
-+ |
5276 |
-+ if (RB_EMPTY_ROOT(&st->active)) |
5277 |
-+ return NULL; |
5278 |
-+ |
5279 |
-+ bfq_update_vtime(st); |
5280 |
-+ entity = bfq_first_active_entity(st); |
5281 |
-+ BUG_ON(bfq_gt(entity->start, st->vtime)); |
5282 |
-+ |
5283 |
-+ /* |
5284 |
-+ * If the chosen entity does not match with the sched_data's |
5285 |
-+ * next_in_service and we are forcedly serving the IDLE priority |
5286 |
-+ * class tree, bubble up budget update. |
5287 |
-+ */ |
5288 |
-+ if (unlikely(force && entity != entity->sched_data->next_in_service)) { |
5289 |
-+ new_next_in_service = entity; |
5290 |
-+ for_each_entity(new_next_in_service) |
5291 |
-+ bfq_update_budget(new_next_in_service); |
5292 |
-+ } |
5293 |
-+ |
5294 |
-+ return entity; |
5295 |
-+} |
5296 |
-+ |
5297 |
-+/** |
5298 |
-+ * bfq_lookup_next_entity - return the first eligible entity in @sd. |
5299 |
-+ * @sd: the sched_data. |
5300 |
-+ * @extract: if true the returned entity will be also extracted from @sd. |
5301 |
-+ * |
5302 |
-+ * NOTE: since we cache the next_in_service entity at each level of the |
5303 |
-+ * hierarchy, the complexity of the lookup can be decreased with |
5304 |
-+ * absolutely no effort just returning the cached next_in_service value; |
5305 |
-+ * we prefer to do full lookups to test the consistency of * the data |
5306 |
-+ * structures. |
5307 |
-+ */ |
5308 |
-+static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
5309 |
-+ int extract, |
5310 |
-+ struct bfq_data *bfqd) |
5311 |
-+{ |
5312 |
-+ struct bfq_service_tree *st = sd->service_tree; |
5313 |
-+ struct bfq_entity *entity; |
5314 |
-+ int i = 0; |
5315 |
-+ |
5316 |
-+ BUG_ON(sd->in_service_entity != NULL); |
5317 |
-+ |
5318 |
-+ if (bfqd != NULL && |
5319 |
-+ jiffies - bfqd->bfq_class_idle_last_service > BFQ_CL_IDLE_TIMEOUT) { |
5320 |
-+ entity = __bfq_lookup_next_entity(st + BFQ_IOPRIO_CLASSES - 1, |
5321 |
-+ true); |
5322 |
-+ if (entity != NULL) { |
5323 |
-+ i = BFQ_IOPRIO_CLASSES - 1; |
5324 |
-+ bfqd->bfq_class_idle_last_service = jiffies; |
5325 |
-+ sd->next_in_service = entity; |
5326 |
-+ } |
5327 |
-+ } |
5328 |
-+ for (; i < BFQ_IOPRIO_CLASSES; i++) { |
5329 |
-+ entity = __bfq_lookup_next_entity(st + i, false); |
5330 |
-+ if (entity != NULL) { |
5331 |
-+ if (extract) { |
5332 |
-+ bfq_check_next_in_service(sd, entity); |
5333 |
-+ bfq_active_extract(st + i, entity); |
5334 |
-+ sd->in_service_entity = entity; |
5335 |
-+ sd->next_in_service = NULL; |
5336 |
-+ } |
5337 |
-+ break; |
5338 |
-+ } |
5339 |
-+ } |
5340 |
-+ |
5341 |
-+ return entity; |
5342 |
-+} |
5343 |
-+ |
5344 |
-+/* |
5345 |
-+ * Get next queue for service. |
5346 |
-+ */ |
5347 |
-+static struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd) |
5348 |
-+{ |
5349 |
-+ struct bfq_entity *entity = NULL; |
5350 |
-+ struct bfq_sched_data *sd; |
5351 |
-+ struct bfq_queue *bfqq; |
5352 |
-+ |
5353 |
-+ BUG_ON(bfqd->in_service_queue != NULL); |
5354 |
-+ |
5355 |
-+ if (bfqd->busy_queues == 0) |
5356 |
-+ return NULL; |
5357 |
-+ |
5358 |
-+ sd = &bfqd->root_group->sched_data; |
5359 |
-+ for (; sd != NULL; sd = entity->my_sched_data) { |
5360 |
-+ entity = bfq_lookup_next_entity(sd, 1, bfqd); |
5361 |
-+ BUG_ON(entity == NULL); |
5362 |
-+ entity->service = 0; |
5363 |
-+ } |
5364 |
-+ |
5365 |
-+ bfqq = bfq_entity_to_bfqq(entity); |
5366 |
-+ BUG_ON(bfqq == NULL); |
5367 |
-+ |
5368 |
-+ return bfqq; |
5369 |
-+} |
5370 |
-+ |
5371 |
-+/* |
5372 |
-+ * Forced extraction of the given queue. |
5373 |
-+ */ |
5374 |
-+static void bfq_get_next_queue_forced(struct bfq_data *bfqd, |
5375 |
-+ struct bfq_queue *bfqq) |
5376 |
-+{ |
5377 |
-+ struct bfq_entity *entity; |
5378 |
-+ struct bfq_sched_data *sd; |
5379 |
-+ |
5380 |
-+ BUG_ON(bfqd->in_service_queue != NULL); |
5381 |
-+ |
5382 |
-+ entity = &bfqq->entity; |
5383 |
-+ /* |
5384 |
-+ * Bubble up extraction/update from the leaf to the root. |
5385 |
-+ */ |
5386 |
-+ for_each_entity(entity) { |
5387 |
-+ sd = entity->sched_data; |
5388 |
-+ bfq_update_budget(entity); |
5389 |
-+ bfq_update_vtime(bfq_entity_service_tree(entity)); |
5390 |
-+ bfq_active_extract(bfq_entity_service_tree(entity), entity); |
5391 |
-+ sd->active_entity = entity; |
5392 |
-+ sd->next_active = NULL; |
5393 |
-+ entity->service = 0; |
5394 |
-+ } |
5395 |
-+ |
5396 |
-+ return; |
5397 |
-+} |
5398 |
-+ |
5399 |
-+static void __bfq_bfqd_reset_in_service(struct bfq_data *bfqd) |
5400 |
-+{ |
5401 |
-+ if (bfqd->in_service_bic != NULL) { |
5402 |
-+ put_io_context(bfqd->in_service_bic->icq.ioc); |
5403 |
-+ bfqd->in_service_bic = NULL; |
5404 |
-+ } |
5405 |
-+ |
5406 |
-+ bfqd->in_service_queue = NULL; |
5407 |
-+ del_timer(&bfqd->idle_slice_timer); |
5408 |
-+} |
5409 |
-+ |
5410 |
-+static void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
5411 |
-+ int requeue) |
5412 |
-+{ |
5413 |
-+ struct bfq_entity *entity = &bfqq->entity; |
5414 |
-+ |
5415 |
-+ if (bfqq == bfqd->in_service_queue) |
5416 |
-+ __bfq_bfqd_reset_in_service(bfqd); |
5417 |
-+ |
5418 |
-+ bfq_deactivate_entity(entity, requeue); |
5419 |
-+} |
5420 |
-+ |
5421 |
-+static void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
5422 |
-+{ |
5423 |
-+ struct bfq_entity *entity = &bfqq->entity; |
5424 |
-+ |
5425 |
-+ bfq_activate_entity(entity); |
5426 |
-+} |
5427 |
-+ |
5428 |
-+/* |
5429 |
-+ * Called when the bfqq no longer has requests pending, remove it from |
5430 |
-+ * the service tree. |
5431 |
-+ */ |
5432 |
-+static void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
5433 |
-+ int requeue) |
5434 |
-+{ |
5435 |
-+ BUG_ON(!bfq_bfqq_busy(bfqq)); |
5436 |
-+ BUG_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
5437 |
-+ |
5438 |
-+ bfq_log_bfqq(bfqd, bfqq, "del from busy"); |
5439 |
-+ |
5440 |
-+ bfq_clear_bfqq_busy(bfqq); |
5441 |
-+ |
5442 |
-+ BUG_ON(bfqd->busy_queues == 0); |
5443 |
-+ bfqd->busy_queues--; |
5444 |
-+ if (bfqq->raising_coeff > 1) |
5445 |
-+ bfqd->raised_busy_queues--; |
5446 |
-+ |
5447 |
-+ bfq_deactivate_bfqq(bfqd, bfqq, requeue); |
5448 |
-+} |
5449 |
-+ |
5450 |
-+/* |
5451 |
-+ * Called when an inactive queue receives a new request. |
5452 |
-+ */ |
5453 |
-+static void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
5454 |
-+{ |
5455 |
-+ BUG_ON(bfq_bfqq_busy(bfqq)); |
5456 |
-+ BUG_ON(bfqq == bfqd->in_service_queue); |
5457 |
-+ |
5458 |
-+ bfq_log_bfqq(bfqd, bfqq, "add to busy"); |
5459 |
-+ |
5460 |
-+ bfq_activate_bfqq(bfqd, bfqq); |
5461 |
-+ |
5462 |
-+ bfq_mark_bfqq_busy(bfqq); |
5463 |
-+ bfqd->busy_queues++; |
5464 |
-+ if (bfqq->raising_coeff > 1) |
5465 |
-+ bfqd->raised_busy_queues++; |
5466 |
-+} |
5467 |
-diff --git a/block/bfq.h b/block/bfq.h |
5468 |
-new file mode 100644 |
5469 |
-index 0000000..3ca8482 |
5470 |
---- /dev/null |
5471 |
-+++ b/block/bfq.h |
5472 |
-@@ -0,0 +1,622 @@ |
5473 |
-+/* |
5474 |
-+ * BFQ-v7r2 for 3.14.0: data structures and common functions prototypes. |
5475 |
-+ * |
5476 |
-+ * Based on ideas and code from CFQ: |
5477 |
-+ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
5478 |
-+ * |
5479 |
-+ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
5480 |
-+ * Paolo Valente <paolo.valente@×××××××.it> |
5481 |
-+ * |
5482 |
-+ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
5483 |
-+ */ |
5484 |
-+ |
5485 |
-+#ifndef _BFQ_H |
5486 |
-+#define _BFQ_H |
5487 |
-+ |
5488 |
-+#include <linux/blktrace_api.h> |
5489 |
-+#include <linux/hrtimer.h> |
5490 |
-+#include <linux/ioprio.h> |
5491 |
-+#include <linux/rbtree.h> |
5492 |
-+ |
5493 |
-+#define BFQ_IOPRIO_CLASSES 3 |
5494 |
-+#define BFQ_CL_IDLE_TIMEOUT (HZ/5) |
5495 |
-+ |
5496 |
-+#define BFQ_MIN_WEIGHT 1 |
5497 |
-+#define BFQ_MAX_WEIGHT 1000 |
5498 |
-+ |
5499 |
-+#define BFQ_DEFAULT_GRP_WEIGHT 10 |
5500 |
-+#define BFQ_DEFAULT_GRP_IOPRIO 0 |
5501 |
-+#define BFQ_DEFAULT_GRP_CLASS IOPRIO_CLASS_BE |
5502 |
-+ |
5503 |
-+struct bfq_entity; |
5504 |
-+ |
5505 |
-+/** |
5506 |
-+ * struct bfq_service_tree - per ioprio_class service tree. |
5507 |
-+ * @active: tree for active entities (i.e., those backlogged). |
5508 |
-+ * @idle: tree for idle entities (i.e., those not backlogged, with V <= F_i). |
5509 |
-+ * @first_idle: idle entity with minimum F_i. |
5510 |
-+ * @last_idle: idle entity with maximum F_i. |
5511 |
-+ * @vtime: scheduler virtual time. |
5512 |
-+ * @wsum: scheduler weight sum; active and idle entities contribute to it. |
5513 |
-+ * |
5514 |
-+ * Each service tree represents a B-WF2Q+ scheduler on its own. Each |
5515 |
-+ * ioprio_class has its own independent scheduler, and so its own |
5516 |
-+ * bfq_service_tree. All the fields are protected by the queue lock |
5517 |
-+ * of the containing bfqd. |
5518 |
-+ */ |
5519 |
-+struct bfq_service_tree { |
5520 |
-+ struct rb_root active; |
5521 |
-+ struct rb_root idle; |
5522 |
-+ |
5523 |
-+ struct bfq_entity *first_idle; |
5524 |
-+ struct bfq_entity *last_idle; |
5525 |
-+ |
5526 |
-+ u64 vtime; |
5527 |
-+ unsigned long wsum; |
5528 |
-+}; |
5529 |
-+ |
5530 |
-+/** |
5531 |
-+ * struct bfq_sched_data - multi-class scheduler. |
5532 |
-+ * @in_service_entity: entity under service. |
5533 |
-+ * @next_in_service: head-of-the-line entity in the scheduler. |
5534 |
-+ * @service_tree: array of service trees, one per ioprio_class. |
5535 |
-+ * |
5536 |
-+ * bfq_sched_data is the basic scheduler queue. It supports three |
5537 |
-+ * ioprio_classes, and can be used either as a toplevel queue or as |
5538 |
-+ * an intermediate queue on a hierarchical setup. |
5539 |
-+ * @next_in_service points to the active entity of the sched_data |
5540 |
-+ * service trees that will be scheduled next. |
5541 |
-+ * |
5542 |
-+ * The supported ioprio_classes are the same as in CFQ, in descending |
5543 |
-+ * priority order, IOPRIO_CLASS_RT, IOPRIO_CLASS_BE, IOPRIO_CLASS_IDLE. |
5544 |
-+ * Requests from higher priority queues are served before all the |
5545 |
-+ * requests from lower priority queues; among requests of the same |
5546 |
-+ * queue requests are served according to B-WF2Q+. |
5547 |
-+ * All the fields are protected by the queue lock of the containing bfqd. |
5548 |
-+ */ |
5549 |
-+struct bfq_sched_data { |
5550 |
-+ struct bfq_entity *in_service_entity; |
5551 |
-+ struct bfq_entity *next_in_service; |
5552 |
-+ struct bfq_service_tree service_tree[BFQ_IOPRIO_CLASSES]; |
5553 |
-+}; |
5554 |
-+ |
5555 |
-+/** |
5556 |
-+ * struct bfq_entity - schedulable entity. |
5557 |
-+ * @rb_node: service_tree member. |
5558 |
-+ * @on_st: flag, true if the entity is on a tree (either the active or |
5559 |
-+ * the idle one of its service_tree). |
5560 |
-+ * @finish: B-WF2Q+ finish timestamp (aka F_i). |
5561 |
-+ * @start: B-WF2Q+ start timestamp (aka S_i). |
5562 |
-+ * @tree: tree the entity is enqueued into; %NULL if not on a tree. |
5563 |
-+ * @min_start: minimum start time of the (active) subtree rooted at |
5564 |
-+ * this entity; used for O(log N) lookups into active trees. |
5565 |
-+ * @service: service received during the last round of service. |
5566 |
-+ * @budget: budget used to calculate F_i; F_i = S_i + @budget / @weight. |
5567 |
-+ * @weight: weight of the queue |
5568 |
-+ * @parent: parent entity, for hierarchical scheduling. |
5569 |
-+ * @my_sched_data: for non-leaf nodes in the cgroup hierarchy, the |
5570 |
-+ * associated scheduler queue, %NULL on leaf nodes. |
5571 |
-+ * @sched_data: the scheduler queue this entity belongs to. |
5572 |
-+ * @ioprio: the ioprio in use. |
5573 |
-+ * @new_weight: when a weight change is requested, the new weight value. |
5574 |
-+ * @orig_weight: original weight, used to implement weight boosting |
5575 |
-+ * @new_ioprio: when an ioprio change is requested, the new ioprio value. |
5576 |
-+ * @ioprio_class: the ioprio_class in use. |
5577 |
-+ * @new_ioprio_class: when an ioprio_class change is requested, the new |
5578 |
-+ * ioprio_class value. |
5579 |
-+ * @ioprio_changed: flag, true when the user requested a weight, ioprio or |
5580 |
-+ * ioprio_class change. |
5581 |
-+ * |
5582 |
-+ * A bfq_entity is used to represent either a bfq_queue (leaf node in the |
5583 |
-+ * cgroup hierarchy) or a bfq_group into the upper level scheduler. Each |
5584 |
-+ * entity belongs to the sched_data of the parent group in the cgroup |
5585 |
-+ * hierarchy. Non-leaf entities have also their own sched_data, stored |
5586 |
-+ * in @my_sched_data. |
5587 |
-+ * |
5588 |
-+ * Each entity stores independently its priority values; this would |
5589 |
-+ * allow different weights on different devices, but this |
5590 |
-+ * functionality is not exported to userspace by now. Priorities and |
5591 |
-+ * weights are updated lazily, first storing the new values into the |
5592 |
-+ * new_* fields, then setting the @ioprio_changed flag. As soon as |
5593 |
-+ * there is a transition in the entity state that allows the priority |
5594 |
-+ * update to take place the effective and the requested priority |
5595 |
-+ * values are synchronized. |
5596 |
-+ * |
5597 |
-+ * Unless cgroups are used, the weight value is calculated from the |
5598 |
-+ * ioprio to export the same interface as CFQ. When dealing with |
5599 |
-+ * ``well-behaved'' queues (i.e., queues that do not spend too much |
5600 |
-+ * time to consume their budget and have true sequential behavior, and |
5601 |
-+ * when there are no external factors breaking anticipation) the |
5602 |
-+ * relative weights at each level of the cgroups hierarchy should be |
5603 |
-+ * guaranteed. All the fields are protected by the queue lock of the |
5604 |
-+ * containing bfqd. |
5605 |
-+ */ |
5606 |
-+struct bfq_entity { |
5607 |
-+ struct rb_node rb_node; |
5608 |
-+ |
5609 |
-+ int on_st; |
5610 |
-+ |
5611 |
-+ u64 finish; |
5612 |
-+ u64 start; |
5613 |
-+ |
5614 |
-+ struct rb_root *tree; |
5615 |
-+ |
5616 |
-+ u64 min_start; |
5617 |
-+ |
5618 |
-+ unsigned long service, budget; |
5619 |
-+ unsigned short weight, new_weight; |
5620 |
-+ unsigned short orig_weight; |
5621 |
-+ |
5622 |
-+ struct bfq_entity *parent; |
5623 |
-+ |
5624 |
-+ struct bfq_sched_data *my_sched_data; |
5625 |
-+ struct bfq_sched_data *sched_data; |
5626 |
-+ |
5627 |
-+ unsigned short ioprio, new_ioprio; |
5628 |
-+ unsigned short ioprio_class, new_ioprio_class; |
5629 |
-+ |
5630 |
-+ int ioprio_changed; |
5631 |
-+}; |
5632 |
-+ |
5633 |
-+struct bfq_group; |
5634 |
-+ |
5635 |
-+/** |
5636 |
-+ * struct bfq_queue - leaf schedulable entity. |
5637 |
-+ * @ref: reference counter. |
5638 |
-+ * @bfqd: parent bfq_data. |
5639 |
-+ * @new_bfqq: shared bfq_queue if queue is cooperating with |
5640 |
-+ * one or more other queues. |
5641 |
-+ * @pos_node: request-position tree member (see bfq_data's @rq_pos_tree). |
5642 |
-+ * @pos_root: request-position tree root (see bfq_data's @rq_pos_tree). |
5643 |
-+ * @sort_list: sorted list of pending requests. |
5644 |
-+ * @next_rq: if fifo isn't expired, next request to serve. |
5645 |
-+ * @queued: nr of requests queued in @sort_list. |
5646 |
-+ * @allocated: currently allocated requests. |
5647 |
-+ * @meta_pending: pending metadata requests. |
5648 |
-+ * @fifo: fifo list of requests in sort_list. |
5649 |
-+ * @entity: entity representing this queue in the scheduler. |
5650 |
-+ * @max_budget: maximum budget allowed from the feedback mechanism. |
5651 |
-+ * @budget_timeout: budget expiration (in jiffies). |
5652 |
-+ * @dispatched: number of requests on the dispatch list or inside driver. |
5653 |
-+ * @org_ioprio: saved ioprio during boosted periods. |
5654 |
-+ * @flags: status flags. |
5655 |
-+ * @bfqq_list: node for active/idle bfqq list inside our bfqd. |
5656 |
-+ * @seek_samples: number of seeks sampled |
5657 |
-+ * @seek_total: sum of the distances of the seeks sampled |
5658 |
-+ * @seek_mean: mean seek distance |
5659 |
-+ * @last_request_pos: position of the last request enqueued |
5660 |
-+ * @pid: pid of the process owning the queue, used for logging purposes. |
5661 |
-+ * @last_rais_start_finish: start time of the current weight-raising period if |
5662 |
-+ * the @bfq-queue is being weight-raised, otherwise |
5663 |
-+ * finish time of the last weight-raising period |
5664 |
-+ * @raising_cur_max_time: current max raising time for this queue |
5665 |
-+ * @soft_rt_next_start: minimum time instant such that, only if a new request |
5666 |
-+ * is enqueued after this time instant in an idle |
5667 |
-+ * @bfq_queue with no outstanding requests, then the |
5668 |
-+ * task associated with the queue it is deemed as soft |
5669 |
-+ * real-time (see the comments to the function |
5670 |
-+ * bfq_bfqq_softrt_next_start()) |
5671 |
-+ * @last_idle_bklogged: time of the last transition of the @bfq_queue from |
5672 |
-+ * idle to backlogged |
5673 |
-+ * @service_from_backlogged: cumulative service received from the @bfq_queue |
5674 |
-+ * since the last transition from idle to backlogged |
5675 |
-+ * |
5676 |
-+ * A bfq_queue is a leaf request queue; it can be associated with an io_context |
5677 |
-+ * or more, if it is async or shared between cooperating processes. @cgroup |
5678 |
-+ * holds a reference to the cgroup, to be sure that it does not disappear while |
5679 |
-+ * a bfqq still references it (mostly to avoid races between request issuing and |
5680 |
-+ * task migration followed by cgroup destruction). |
5681 |
-+ * All the fields are protected by the queue lock of the containing bfqd. |
5682 |
-+ */ |
5683 |
-+struct bfq_queue { |
5684 |
-+ atomic_t ref; |
5685 |
-+ struct bfq_data *bfqd; |
5686 |
-+ |
5687 |
-+ /* fields for cooperating queues handling */ |
5688 |
-+ struct bfq_queue *new_bfqq; |
5689 |
-+ struct rb_node pos_node; |
5690 |
-+ struct rb_root *pos_root; |
5691 |
-+ |
5692 |
-+ struct rb_root sort_list; |
5693 |
-+ struct request *next_rq; |
5694 |
-+ int queued[2]; |
5695 |
-+ int allocated[2]; |
5696 |
-+ int meta_pending; |
5697 |
-+ struct list_head fifo; |
5698 |
-+ |
5699 |
-+ struct bfq_entity entity; |
5700 |
-+ |
5701 |
-+ unsigned long max_budget; |
5702 |
-+ unsigned long budget_timeout; |
5703 |
-+ |
5704 |
-+ int dispatched; |
5705 |
-+ |
5706 |
-+ unsigned short org_ioprio; |
5707 |
-+ |
5708 |
-+ unsigned int flags; |
5709 |
-+ |
5710 |
-+ struct list_head bfqq_list; |
5711 |
-+ |
5712 |
-+ unsigned int seek_samples; |
5713 |
-+ u64 seek_total; |
5714 |
-+ sector_t seek_mean; |
5715 |
-+ sector_t last_request_pos; |
5716 |
-+ |
5717 |
-+ pid_t pid; |
5718 |
-+ |
5719 |
-+ /* weight-raising fields */ |
5720 |
-+ unsigned long raising_cur_max_time; |
5721 |
-+ unsigned long soft_rt_next_start; |
5722 |
-+ unsigned long last_rais_start_finish; |
5723 |
-+ unsigned int raising_coeff; |
5724 |
-+ unsigned long last_idle_bklogged; |
5725 |
-+ unsigned long service_from_backlogged; |
5726 |
-+}; |
5727 |
-+ |
5728 |
-+/** |
5729 |
-+ * struct bfq_ttime - per process thinktime stats. |
5730 |
-+ * @ttime_total: total process thinktime |
5731 |
-+ * @ttime_samples: number of thinktime samples |
5732 |
-+ * @ttime_mean: average process thinktime |
5733 |
-+ */ |
5734 |
-+struct bfq_ttime { |
5735 |
-+ unsigned long last_end_request; |
5736 |
-+ |
5737 |
-+ unsigned long ttime_total; |
5738 |
-+ unsigned long ttime_samples; |
5739 |
-+ unsigned long ttime_mean; |
5740 |
-+}; |
5741 |
-+ |
5742 |
-+/** |
5743 |
-+ * struct bfq_io_cq - per (request_queue, io_context) structure. |
5744 |
-+ * @icq: associated io_cq structure |
5745 |
-+ * @bfqq: array of two process queues, the sync and the async |
5746 |
-+ * @ttime: associated @bfq_ttime struct |
5747 |
-+ */ |
5748 |
-+struct bfq_io_cq { |
5749 |
-+ struct io_cq icq; /* must be the first member */ |
5750 |
-+ struct bfq_queue *bfqq[2]; |
5751 |
-+ struct bfq_ttime ttime; |
5752 |
-+ int ioprio; |
5753 |
-+}; |
5754 |
-+ |
5755 |
-+/** |
5756 |
-+ * struct bfq_data - per device data structure. |
5757 |
-+ * @queue: request queue for the managed device. |
5758 |
-+ * @root_group: root bfq_group for the device. |
5759 |
-+ * @rq_pos_tree: rbtree sorted by next_request position, |
5760 |
-+ * used when determining if two or more queues |
5761 |
-+ * have interleaving requests (see bfq_close_cooperator). |
5762 |
-+ * @busy_queues: number of bfq_queues containing requests (including the |
5763 |
-+ * queue under service, even if it is idling). |
5764 |
-+ * @raised_busy_queues: number of weight-raised busy bfq_queues. |
5765 |
-+ * @queued: number of queued requests. |
5766 |
-+ * @rq_in_driver: number of requests dispatched and waiting for completion. |
5767 |
-+ * @sync_flight: number of sync requests in the driver. |
5768 |
-+ * @max_rq_in_driver: max number of reqs in driver in the last @hw_tag_samples |
5769 |
-+ * completed requests . |
5770 |
-+ * @hw_tag_samples: nr of samples used to calculate hw_tag. |
5771 |
-+ * @hw_tag: flag set to one if the driver is showing a queueing behavior. |
5772 |
-+ * @budgets_assigned: number of budgets assigned. |
5773 |
-+ * @idle_slice_timer: timer set when idling for the next sequential request |
5774 |
-+ * from the queue under service. |
5775 |
-+ * @unplug_work: delayed work to restart dispatching on the request queue. |
5776 |
-+ * @in_service_queue: bfq_queue under service. |
5777 |
-+ * @in_service_bic: bfq_io_cq (bic) associated with the @in_service_queue. |
5778 |
-+ * @last_position: on-disk position of the last served request. |
5779 |
-+ * @last_budget_start: beginning of the last budget. |
5780 |
-+ * @last_idling_start: beginning of the last idle slice. |
5781 |
-+ * @peak_rate: peak transfer rate observed for a budget. |
5782 |
-+ * @peak_rate_samples: number of samples used to calculate @peak_rate. |
5783 |
-+ * @bfq_max_budget: maximum budget allotted to a bfq_queue before rescheduling. |
5784 |
-+ * @group_list: list of all the bfq_groups active on the device. |
5785 |
-+ * @active_list: list of all the bfq_queues active on the device. |
5786 |
-+ * @idle_list: list of all the bfq_queues idle on the device. |
5787 |
-+ * @bfq_quantum: max number of requests dispatched per dispatch round. |
5788 |
-+ * @bfq_fifo_expire: timeout for async/sync requests; when it expires |
5789 |
-+ * requests are served in fifo order. |
5790 |
-+ * @bfq_back_penalty: weight of backward seeks wrt forward ones. |
5791 |
-+ * @bfq_back_max: maximum allowed backward seek. |
5792 |
-+ * @bfq_slice_idle: maximum idling time. |
5793 |
-+ * @bfq_user_max_budget: user-configured max budget value (0 for auto-tuning). |
5794 |
-+ * @bfq_max_budget_async_rq: maximum budget (in nr of requests) allotted to |
5795 |
-+ * async queues. |
5796 |
-+ * @bfq_timeout: timeout for bfq_queues to consume their budget; used to |
5797 |
-+ * to prevent seeky queues to impose long latencies to well |
5798 |
-+ * behaved ones (this also implies that seeky queues cannot |
5799 |
-+ * receive guarantees in the service domain; after a timeout |
5800 |
-+ * they are charged for the whole allocated budget, to try |
5801 |
-+ * to preserve a behavior reasonably fair among them, but |
5802 |
-+ * without service-domain guarantees). |
5803 |
-+ * @bfq_raising_coeff: Maximum factor by which the weight of a boosted |
5804 |
-+ * queue is multiplied |
5805 |
-+ * @bfq_raising_max_time: maximum duration of a weight-raising period (jiffies) |
5806 |
-+ * @bfq_raising_rt_max_time: maximum duration for soft real-time processes |
5807 |
-+ * @bfq_raising_min_idle_time: minimum idle period after which weight-raising |
5808 |
-+ * may be reactivated for a queue (in jiffies) |
5809 |
-+ * @bfq_raising_min_inter_arr_async: minimum period between request arrivals |
5810 |
-+ * after which weight-raising may be |
5811 |
-+ * reactivated for an already busy queue |
5812 |
-+ * (in jiffies) |
5813 |
-+ * @bfq_raising_max_softrt_rate: max service-rate for a soft real-time queue, |
5814 |
-+ * sectors per seconds |
5815 |
-+ * @RT_prod: cached value of the product R*T used for computing the maximum |
5816 |
-+ * duration of the weight raising automatically |
5817 |
-+ * @oom_bfqq: fallback dummy bfqq for extreme OOM conditions |
5818 |
-+ * |
5819 |
-+ * All the fields are protected by the @queue lock. |
5820 |
-+ */ |
5821 |
-+struct bfq_data { |
5822 |
-+ struct request_queue *queue; |
5823 |
-+ |
5824 |
-+ struct bfq_group *root_group; |
5825 |
-+ |
5826 |
-+ struct rb_root rq_pos_tree; |
5827 |
-+ |
5828 |
-+ int busy_queues; |
5829 |
-+ int raised_busy_queues; |
5830 |
-+ int queued; |
5831 |
-+ int rq_in_driver; |
5832 |
-+ int sync_flight; |
5833 |
-+ |
5834 |
-+ int max_rq_in_driver; |
5835 |
-+ int hw_tag_samples; |
5836 |
-+ int hw_tag; |
5837 |
-+ |
5838 |
-+ int budgets_assigned; |
5839 |
-+ |
5840 |
-+ struct timer_list idle_slice_timer; |
5841 |
-+ struct work_struct unplug_work; |
5842 |
-+ |
5843 |
-+ struct bfq_queue *in_service_queue; |
5844 |
-+ struct bfq_io_cq *in_service_bic; |
5845 |
-+ |
5846 |
-+ sector_t last_position; |
5847 |
-+ |
5848 |
-+ ktime_t last_budget_start; |
5849 |
-+ ktime_t last_idling_start; |
5850 |
-+ int peak_rate_samples; |
5851 |
-+ u64 peak_rate; |
5852 |
-+ unsigned long bfq_max_budget; |
5853 |
-+ |
5854 |
-+ struct hlist_head group_list; |
5855 |
-+ struct list_head active_list; |
5856 |
-+ struct list_head idle_list; |
5857 |
-+ |
5858 |
-+ unsigned int bfq_quantum; |
5859 |
-+ unsigned int bfq_fifo_expire[2]; |
5860 |
-+ unsigned int bfq_back_penalty; |
5861 |
-+ unsigned int bfq_back_max; |
5862 |
-+ unsigned int bfq_slice_idle; |
5863 |
-+ u64 bfq_class_idle_last_service; |
5864 |
-+ |
5865 |
-+ unsigned int bfq_user_max_budget; |
5866 |
-+ unsigned int bfq_max_budget_async_rq; |
5867 |
-+ unsigned int bfq_timeout[2]; |
5868 |
-+ |
5869 |
-+ bool low_latency; |
5870 |
-+ |
5871 |
-+ /* parameters of the low_latency heuristics */ |
5872 |
-+ unsigned int bfq_raising_coeff; |
5873 |
-+ unsigned int bfq_raising_max_time; |
5874 |
-+ unsigned int bfq_raising_rt_max_time; |
5875 |
-+ unsigned int bfq_raising_min_idle_time; |
5876 |
-+ unsigned long bfq_raising_min_inter_arr_async; |
5877 |
-+ unsigned int bfq_raising_max_softrt_rate; |
5878 |
-+ u64 RT_prod; |
5879 |
-+ |
5880 |
-+ struct bfq_queue oom_bfqq; |
5881 |
-+}; |
5882 |
-+ |
5883 |
-+enum bfqq_state_flags { |
5884 |
-+ BFQ_BFQQ_FLAG_busy = 0, /* has requests or is under service */ |
5885 |
-+ BFQ_BFQQ_FLAG_wait_request, /* waiting for a request */ |
5886 |
-+ BFQ_BFQQ_FLAG_must_alloc, /* must be allowed rq alloc */ |
5887 |
-+ BFQ_BFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
5888 |
-+ BFQ_BFQQ_FLAG_idle_window, /* slice idling enabled */ |
5889 |
-+ BFQ_BFQQ_FLAG_prio_changed, /* task priority has changed */ |
5890 |
-+ BFQ_BFQQ_FLAG_sync, /* synchronous queue */ |
5891 |
-+ BFQ_BFQQ_FLAG_budget_new, /* no completion with this budget */ |
5892 |
-+ BFQ_BFQQ_FLAG_coop, /* bfqq is shared */ |
5893 |
-+ BFQ_BFQQ_FLAG_split_coop, /* shared bfqq will be splitted */ |
5894 |
-+ BFQ_BFQQ_FLAG_softrt_update, /* needs softrt-next-start update */ |
5895 |
-+}; |
5896 |
-+ |
5897 |
-+#define BFQ_BFQQ_FNS(name) \ |
5898 |
-+static inline void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ |
5899 |
-+{ \ |
5900 |
-+ (bfqq)->flags |= (1 << BFQ_BFQQ_FLAG_##name); \ |
5901 |
-+} \ |
5902 |
-+static inline void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ |
5903 |
-+{ \ |
5904 |
-+ (bfqq)->flags &= ~(1 << BFQ_BFQQ_FLAG_##name); \ |
5905 |
-+} \ |
5906 |
-+static inline int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ |
5907 |
-+{ \ |
5908 |
-+ return ((bfqq)->flags & (1 << BFQ_BFQQ_FLAG_##name)) != 0; \ |
5909 |
-+} |
5910 |
-+ |
5911 |
-+BFQ_BFQQ_FNS(busy); |
5912 |
-+BFQ_BFQQ_FNS(wait_request); |
5913 |
-+BFQ_BFQQ_FNS(must_alloc); |
5914 |
-+BFQ_BFQQ_FNS(fifo_expire); |
5915 |
-+BFQ_BFQQ_FNS(idle_window); |
5916 |
-+BFQ_BFQQ_FNS(prio_changed); |
5917 |
-+BFQ_BFQQ_FNS(sync); |
5918 |
-+BFQ_BFQQ_FNS(budget_new); |
5919 |
-+BFQ_BFQQ_FNS(coop); |
5920 |
-+BFQ_BFQQ_FNS(split_coop); |
5921 |
-+BFQ_BFQQ_FNS(softrt_update); |
5922 |
-+#undef BFQ_BFQQ_FNS |
5923 |
-+ |
5924 |
-+/* Logging facilities. */ |
5925 |
-+#define bfq_log_bfqq(bfqd, bfqq, fmt, args...) \ |
5926 |
-+ blk_add_trace_msg((bfqd)->queue, "bfq%d " fmt, (bfqq)->pid, ##args) |
5927 |
-+ |
5928 |
-+#define bfq_log(bfqd, fmt, args...) \ |
5929 |
-+ blk_add_trace_msg((bfqd)->queue, "bfq " fmt, ##args) |
5930 |
-+ |
5931 |
-+/* Expiration reasons. */ |
5932 |
-+enum bfqq_expiration { |
5933 |
-+ BFQ_BFQQ_TOO_IDLE = 0, /* queue has been idling for too long */ |
5934 |
-+ BFQ_BFQQ_BUDGET_TIMEOUT, /* budget took too long to be used */ |
5935 |
-+ BFQ_BFQQ_BUDGET_EXHAUSTED, /* budget consumed */ |
5936 |
-+ BFQ_BFQQ_NO_MORE_REQUESTS, /* the queue has no more requests */ |
5937 |
-+}; |
5938 |
-+ |
5939 |
-+#ifdef CONFIG_CGROUP_BFQIO |
5940 |
-+/** |
5941 |
-+ * struct bfq_group - per (device, cgroup) data structure. |
5942 |
-+ * @entity: schedulable entity to insert into the parent group sched_data. |
5943 |
-+ * @sched_data: own sched_data, to contain child entities (they may be |
5944 |
-+ * both bfq_queues and bfq_groups). |
5945 |
-+ * @group_node: node to be inserted into the bfqio_cgroup->group_data |
5946 |
-+ * list of the containing cgroup's bfqio_cgroup. |
5947 |
-+ * @bfqd_node: node to be inserted into the @bfqd->group_list list |
5948 |
-+ * of the groups active on the same device; used for cleanup. |
5949 |
-+ * @bfqd: the bfq_data for the device this group acts upon. |
5950 |
-+ * @async_bfqq: array of async queues for all the tasks belonging to |
5951 |
-+ * the group, one queue per ioprio value per ioprio_class, |
5952 |
-+ * except for the idle class that has only one queue. |
5953 |
-+ * @async_idle_bfqq: async queue for the idle class (ioprio is ignored). |
5954 |
-+ * @my_entity: pointer to @entity, %NULL for the toplevel group; used |
5955 |
-+ * to avoid too many special cases during group creation/migration. |
5956 |
-+ * |
5957 |
-+ * Each (device, cgroup) pair has its own bfq_group, i.e., for each cgroup |
5958 |
-+ * there is a set of bfq_groups, each one collecting the lower-level |
5959 |
-+ * entities belonging to the group that are acting on the same device. |
5960 |
-+ * |
5961 |
-+ * Locking works as follows: |
5962 |
-+ * o @group_node is protected by the bfqio_cgroup lock, and is accessed |
5963 |
-+ * via RCU from its readers. |
5964 |
-+ * o @bfqd is protected by the queue lock, RCU is used to access it |
5965 |
-+ * from the readers. |
5966 |
-+ * o All the other fields are protected by the @bfqd queue lock. |
5967 |
-+ */ |
5968 |
-+struct bfq_group { |
5969 |
-+ struct bfq_entity entity; |
5970 |
-+ struct bfq_sched_data sched_data; |
5971 |
-+ |
5972 |
-+ struct hlist_node group_node; |
5973 |
-+ struct hlist_node bfqd_node; |
5974 |
-+ |
5975 |
-+ void *bfqd; |
5976 |
-+ |
5977 |
-+ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
5978 |
-+ struct bfq_queue *async_idle_bfqq; |
5979 |
-+ |
5980 |
-+ struct bfq_entity *my_entity; |
5981 |
-+}; |
5982 |
-+ |
5983 |
-+/** |
5984 |
-+ * struct bfqio_cgroup - bfq cgroup data structure. |
5985 |
-+ * @css: subsystem state for bfq in the containing cgroup. |
5986 |
-+ * @online: flag marked when the subsystem is inserted. |
5987 |
-+ * @weight: cgroup weight. |
5988 |
-+ * @ioprio: cgroup ioprio. |
5989 |
-+ * @ioprio_class: cgroup ioprio_class. |
5990 |
-+ * @lock: spinlock that protects @ioprio, @ioprio_class and @group_data. |
5991 |
-+ * @group_data: list containing the bfq_group belonging to this cgroup. |
5992 |
-+ * |
5993 |
-+ * @group_data is accessed using RCU, with @lock protecting the updates, |
5994 |
-+ * @ioprio and @ioprio_class are protected by @lock. |
5995 |
-+ */ |
5996 |
-+struct bfqio_cgroup { |
5997 |
-+ struct cgroup_subsys_state css; |
5998 |
-+ bool online; |
5999 |
-+ |
6000 |
-+ unsigned short weight, ioprio, ioprio_class; |
6001 |
-+ |
6002 |
-+ spinlock_t lock; |
6003 |
-+ struct hlist_head group_data; |
6004 |
-+}; |
6005 |
-+#else |
6006 |
-+struct bfq_group { |
6007 |
-+ struct bfq_sched_data sched_data; |
6008 |
-+ |
6009 |
-+ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
6010 |
-+ struct bfq_queue *async_idle_bfqq; |
6011 |
-+}; |
6012 |
-+#endif |
6013 |
-+ |
6014 |
-+static inline struct bfq_service_tree * |
6015 |
-+bfq_entity_service_tree(struct bfq_entity *entity) |
6016 |
-+{ |
6017 |
-+ struct bfq_sched_data *sched_data = entity->sched_data; |
6018 |
-+ unsigned int idx = entity->ioprio_class - 1; |
6019 |
-+ |
6020 |
-+ BUG_ON(idx >= BFQ_IOPRIO_CLASSES); |
6021 |
-+ BUG_ON(sched_data == NULL); |
6022 |
-+ |
6023 |
-+ return sched_data->service_tree + idx; |
6024 |
-+} |
6025 |
-+ |
6026 |
-+static inline struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, |
6027 |
-+ int is_sync) |
6028 |
-+{ |
6029 |
-+ return bic->bfqq[!!is_sync]; |
6030 |
-+} |
6031 |
-+ |
6032 |
-+static inline void bic_set_bfqq(struct bfq_io_cq *bic, |
6033 |
-+ struct bfq_queue *bfqq, int is_sync) |
6034 |
-+{ |
6035 |
-+ bic->bfqq[!!is_sync] = bfqq; |
6036 |
-+} |
6037 |
-+ |
6038 |
-+static inline struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) |
6039 |
-+{ |
6040 |
-+ return bic->icq.q->elevator->elevator_data; |
6041 |
-+} |
6042 |
-+ |
6043 |
-+/** |
6044 |
-+ * bfq_get_bfqd_locked - get a lock to a bfqd using a RCU protected pointer. |
6045 |
-+ * @ptr: a pointer to a bfqd. |
6046 |
-+ * @flags: storage for the flags to be saved. |
6047 |
-+ * |
6048 |
-+ * This function allows bfqg->bfqd to be protected by the |
6049 |
-+ * queue lock of the bfqd they reference; the pointer is dereferenced |
6050 |
-+ * under RCU, so the storage for bfqd is assured to be safe as long |
6051 |
-+ * as the RCU read side critical section does not end. After the |
6052 |
-+ * bfqd->queue->queue_lock is taken the pointer is rechecked, to be |
6053 |
-+ * sure that no other writer accessed it. If we raced with a writer, |
6054 |
-+ * the function returns NULL, with the queue unlocked, otherwise it |
6055 |
-+ * returns the dereferenced pointer, with the queue locked. |
6056 |
-+ */ |
6057 |
-+static inline struct bfq_data *bfq_get_bfqd_locked(void **ptr, |
6058 |
-+ unsigned long *flags) |
6059 |
-+{ |
6060 |
-+ struct bfq_data *bfqd; |
6061 |
-+ |
6062 |
-+ rcu_read_lock(); |
6063 |
-+ bfqd = rcu_dereference(*(struct bfq_data **)ptr); |
6064 |
-+ |
6065 |
-+ if (bfqd != NULL) { |
6066 |
-+ spin_lock_irqsave(bfqd->queue->queue_lock, *flags); |
6067 |
-+ if (*ptr == bfqd) |
6068 |
-+ goto out; |
6069 |
-+ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
6070 |
-+ } |
6071 |
-+ |
6072 |
-+ bfqd = NULL; |
6073 |
-+out: |
6074 |
-+ rcu_read_unlock(); |
6075 |
-+ return bfqd; |
6076 |
-+} |
6077 |
-+ |
6078 |
-+static inline void bfq_put_bfqd_unlock(struct bfq_data *bfqd, |
6079 |
-+ unsigned long *flags) |
6080 |
-+{ |
6081 |
-+ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
6082 |
-+} |
6083 |
-+ |
6084 |
-+static void bfq_changed_ioprio(struct bfq_io_cq *bic); |
6085 |
-+static void bfq_put_queue(struct bfq_queue *bfqq); |
6086 |
-+static void bfq_dispatch_insert(struct request_queue *q, struct request *rq); |
6087 |
-+static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
6088 |
-+ struct bfq_group *bfqg, int is_sync, |
6089 |
-+ struct bfq_io_cq *bic, gfp_t gfp_mask); |
6090 |
-+static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
6091 |
-+ struct bfq_group *bfqg); |
6092 |
-+static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg); |
6093 |
-+static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq); |
6094 |
-+#endif |
6095 |
--- |
6096 |
-1.9.0 |
6097 |
- |
6098 |
|
6099 |
Added: genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch1 |
6100 |
=================================================================== |
6101 |
--- genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch1 (rev 0) |
6102 |
+++ genpatches-2.6/trunk/3.14/5002_BFQ-2-block-introduce-the-BFQ-v7r2-I-O-sched-for-3.14.patch1 2014-04-01 18:44:20 UTC (rev 2727) |
6103 |
@@ -0,0 +1,6065 @@ |
6104 |
+From 5055277df59d9280da6b60cf90bed8e5e57dc44d Mon Sep 17 00:00:00 2001 |
6105 |
+From: Paolo Valente <paolo.valente@×××××××.it> |
6106 |
+Date: Thu, 9 May 2013 19:10:02 +0200 |
6107 |
+Subject: [PATCH 2/3] block: introduce the BFQ-v7r2 I/O sched for 3.14 |
6108 |
+ |
6109 |
+Add the BFQ-v7r2 I/O scheduler to 3.14. |
6110 |
+The general structure is borrowed from CFQ, as much of the code for |
6111 |
+handling I/O contexts. Over time, several useful features have been |
6112 |
+ported from CFQ as well (details in the changelog in README.BFQ). A |
6113 |
+(bfq_)queue is associated to each task doing I/O on a device, and each |
6114 |
+time a scheduling decision has to be made a queue is selected and served |
6115 |
+until it expires. |
6116 |
+ |
6117 |
+ - Slices are given in the service domain: tasks are assigned |
6118 |
+ budgets, measured in number of sectors. Once got the disk, a task |
6119 |
+ must however consume its assigned budget within a configurable |
6120 |
+ maximum time (by default, the maximum possible value of the |
6121 |
+ budgets is automatically computed to comply with this timeout). |
6122 |
+ This allows the desired latency vs "throughput boosting" tradeoff |
6123 |
+ to be set. |
6124 |
+ |
6125 |
+ - Budgets are scheduled according to a variant of WF2Q+, implemented |
6126 |
+ using an augmented rb-tree to take eligibility into account while |
6127 |
+ preserving an O(log N) overall complexity. |
6128 |
+ |
6129 |
+ - A low-latency tunable is provided; if enabled, both interactive |
6130 |
+ and soft real-time applications are guaranteed a very low latency. |
6131 |
+ |
6132 |
+ - Latency guarantees are preserved also in the presence of NCQ. |
6133 |
+ |
6134 |
+ - Also with flash-based devices, a high throughput is achieved |
6135 |
+ while still preserving latency guarantees. |
6136 |
+ |
6137 |
+ - BFQ features Early Queue Merge (EQM), a sort of fusion of the |
6138 |
+ cooperating-queue-merging and the preemption mechanisms present |
6139 |
+ in CFQ. EQM is in fact a unified mechanism that tries to get a |
6140 |
+ sequential read pattern, and hence a high throughput, with any |
6141 |
+ set of processes performing interleaved I/O over a contiguous |
6142 |
+ sequence of sectors. |
6143 |
+ |
6144 |
+ - BFQ supports full hierarchical scheduling, exporting a cgroups |
6145 |
+ interface. Since each node has a full scheduler, each group can |
6146 |
+ be assigned its own weight. |
6147 |
+ |
6148 |
+ - If the cgroups interface is not used, only I/O priorities can be |
6149 |
+ assigned to processes, with ioprio values mapped to weights |
6150 |
+ with the relation weight = IOPRIO_BE_NR - ioprio. |
6151 |
+ |
6152 |
+ - ioprio classes are served in strict priority order, i.e., lower |
6153 |
+ priority queues are not served as long as there are higher |
6154 |
+ priority queues. Among queues in the same class the bandwidth is |
6155 |
+ distributed in proportion to the weight of each queue. A very |
6156 |
+ thin extra bandwidth is however guaranteed to the Idle class, to |
6157 |
+ prevent it from starving. |
6158 |
+ |
6159 |
+Signed-off-by: Paolo Valente <paolo.valente@×××××××.it> |
6160 |
+Signed-off-by: Arianna Avanzini <avanzini.arianna@×××××.com> |
6161 |
+--- |
6162 |
+ block/bfq-cgroup.c | 926 +++++++++++++++ |
6163 |
+ block/bfq-ioc.c | 36 + |
6164 |
+ block/bfq-iosched.c | 3300 +++++++++++++++++++++++++++++++++++++++++++++++++++ |
6165 |
+ block/bfq-sched.c | 1078 +++++++++++++++++ |
6166 |
+ block/bfq.h | 622 ++++++++++ |
6167 |
+ 5 files changed, 5962 insertions(+) |
6168 |
+ create mode 100644 block/bfq-cgroup.c |
6169 |
+ create mode 100644 block/bfq-ioc.c |
6170 |
+ create mode 100644 block/bfq-iosched.c |
6171 |
+ create mode 100644 block/bfq-sched.c |
6172 |
+ create mode 100644 block/bfq.h |
6173 |
+ |
6174 |
+diff --git a/block/bfq-cgroup.c b/block/bfq-cgroup.c |
6175 |
+new file mode 100644 |
6176 |
+index 0000000..bcecdb4 |
6177 |
+--- /dev/null |
6178 |
++++ b/block/bfq-cgroup.c |
6179 |
+@@ -0,0 +1,926 @@ |
6180 |
++/* |
6181 |
++ * BFQ: CGROUPS support. |
6182 |
++ * |
6183 |
++ * Based on ideas and code from CFQ: |
6184 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
6185 |
++ * |
6186 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
6187 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
6188 |
++ * |
6189 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
6190 |
++ * |
6191 |
++ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
6192 |
++ */ |
6193 |
++ |
6194 |
++#ifdef CONFIG_CGROUP_BFQIO |
6195 |
++ |
6196 |
++static DEFINE_MUTEX(bfqio_mutex); |
6197 |
++ |
6198 |
++static bool bfqio_is_removed(struct bfqio_cgroup *bgrp) |
6199 |
++{ |
6200 |
++ return bgrp ? !bgrp->online : false; |
6201 |
++} |
6202 |
++ |
6203 |
++static struct bfqio_cgroup bfqio_root_cgroup = { |
6204 |
++ .weight = BFQ_DEFAULT_GRP_WEIGHT, |
6205 |
++ .ioprio = BFQ_DEFAULT_GRP_IOPRIO, |
6206 |
++ .ioprio_class = BFQ_DEFAULT_GRP_CLASS, |
6207 |
++}; |
6208 |
++ |
6209 |
++static inline void bfq_init_entity(struct bfq_entity *entity, |
6210 |
++ struct bfq_group *bfqg) |
6211 |
++{ |
6212 |
++ entity->weight = entity->new_weight; |
6213 |
++ entity->orig_weight = entity->new_weight; |
6214 |
++ entity->ioprio = entity->new_ioprio; |
6215 |
++ entity->ioprio_class = entity->new_ioprio_class; |
6216 |
++ entity->parent = bfqg->my_entity; |
6217 |
++ entity->sched_data = &bfqg->sched_data; |
6218 |
++} |
6219 |
++ |
6220 |
++static struct bfqio_cgroup *css_to_bfqio(struct cgroup_subsys_state *css) |
6221 |
++{ |
6222 |
++ return css ? container_of(css, struct bfqio_cgroup, css) : NULL; |
6223 |
++} |
6224 |
++ |
6225 |
++/* |
6226 |
++ * Search the bfq_group for bfqd into the hash table (by now only a list) |
6227 |
++ * of bgrp. Must be called under rcu_read_lock(). |
6228 |
++ */ |
6229 |
++static struct bfq_group *bfqio_lookup_group(struct bfqio_cgroup *bgrp, |
6230 |
++ struct bfq_data *bfqd) |
6231 |
++{ |
6232 |
++ struct bfq_group *bfqg; |
6233 |
++ void *key; |
6234 |
++ |
6235 |
++ hlist_for_each_entry_rcu(bfqg, &bgrp->group_data, group_node) { |
6236 |
++ key = rcu_dereference(bfqg->bfqd); |
6237 |
++ if (key == bfqd) |
6238 |
++ return bfqg; |
6239 |
++ } |
6240 |
++ |
6241 |
++ return NULL; |
6242 |
++} |
6243 |
++ |
6244 |
++static inline void bfq_group_init_entity(struct bfqio_cgroup *bgrp, |
6245 |
++ struct bfq_group *bfqg) |
6246 |
++{ |
6247 |
++ struct bfq_entity *entity = &bfqg->entity; |
6248 |
++ |
6249 |
++ /* |
6250 |
++ * If the weight of the entity has never been set via the sysfs |
6251 |
++ * interface, then bgrp->weight == 0. In this case we initialize |
6252 |
++ * the weight from the current ioprio value. Otherwise, the group |
6253 |
++ * weight, if set, has priority over the ioprio value. |
6254 |
++ */ |
6255 |
++ if (bgrp->weight == 0) { |
6256 |
++ entity->new_weight = bfq_ioprio_to_weight(bgrp->ioprio); |
6257 |
++ entity->new_ioprio = bgrp->ioprio; |
6258 |
++ } else { |
6259 |
++ entity->new_weight = bgrp->weight; |
6260 |
++ entity->new_ioprio = bfq_weight_to_ioprio(bgrp->weight); |
6261 |
++ } |
6262 |
++ entity->orig_weight = entity->weight = entity->new_weight; |
6263 |
++ entity->ioprio = entity->new_ioprio; |
6264 |
++ entity->ioprio_class = entity->new_ioprio_class = bgrp->ioprio_class; |
6265 |
++ entity->my_sched_data = &bfqg->sched_data; |
6266 |
++} |
6267 |
++ |
6268 |
++static inline void bfq_group_set_parent(struct bfq_group *bfqg, |
6269 |
++ struct bfq_group *parent) |
6270 |
++{ |
6271 |
++ struct bfq_entity *entity; |
6272 |
++ |
6273 |
++ BUG_ON(parent == NULL); |
6274 |
++ BUG_ON(bfqg == NULL); |
6275 |
++ |
6276 |
++ entity = &bfqg->entity; |
6277 |
++ entity->parent = parent->my_entity; |
6278 |
++ entity->sched_data = &parent->sched_data; |
6279 |
++} |
6280 |
++ |
6281 |
++/** |
6282 |
++ * bfq_group_chain_alloc - allocate a chain of groups. |
6283 |
++ * @bfqd: queue descriptor. |
6284 |
++ * @css: the leaf cgroup_subsys_state this chain starts from. |
6285 |
++ * |
6286 |
++ * Allocate a chain of groups starting from the one belonging to |
6287 |
++ * @cgroup up to the root cgroup. Stop if a cgroup on the chain |
6288 |
++ * to the root has already an allocated group on @bfqd. |
6289 |
++ */ |
6290 |
++static struct bfq_group *bfq_group_chain_alloc(struct bfq_data *bfqd, |
6291 |
++ struct cgroup_subsys_state *css) |
6292 |
++{ |
6293 |
++ struct bfqio_cgroup *bgrp; |
6294 |
++ struct bfq_group *bfqg, *prev = NULL, *leaf = NULL; |
6295 |
++ |
6296 |
++ for (; css != NULL; css = css->parent) { |
6297 |
++ bgrp = css_to_bfqio(css); |
6298 |
++ |
6299 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6300 |
++ if (bfqg != NULL) { |
6301 |
++ /* |
6302 |
++ * All the cgroups in the path from there to the |
6303 |
++ * root must have a bfq_group for bfqd, so we don't |
6304 |
++ * need any more allocations. |
6305 |
++ */ |
6306 |
++ break; |
6307 |
++ } |
6308 |
++ |
6309 |
++ bfqg = kzalloc(sizeof(*bfqg), GFP_ATOMIC); |
6310 |
++ if (bfqg == NULL) |
6311 |
++ goto cleanup; |
6312 |
++ |
6313 |
++ bfq_group_init_entity(bgrp, bfqg); |
6314 |
++ bfqg->my_entity = &bfqg->entity; |
6315 |
++ |
6316 |
++ if (leaf == NULL) { |
6317 |
++ leaf = bfqg; |
6318 |
++ prev = leaf; |
6319 |
++ } else { |
6320 |
++ bfq_group_set_parent(prev, bfqg); |
6321 |
++ /* |
6322 |
++ * Build a list of allocated nodes using the bfqd |
6323 |
++ * filed, that is still unused and will be initialized |
6324 |
++ * only after the node will be connected. |
6325 |
++ */ |
6326 |
++ prev->bfqd = bfqg; |
6327 |
++ prev = bfqg; |
6328 |
++ } |
6329 |
++ } |
6330 |
++ |
6331 |
++ return leaf; |
6332 |
++ |
6333 |
++cleanup: |
6334 |
++ while (leaf != NULL) { |
6335 |
++ prev = leaf; |
6336 |
++ leaf = leaf->bfqd; |
6337 |
++ kfree(prev); |
6338 |
++ } |
6339 |
++ |
6340 |
++ return NULL; |
6341 |
++} |
6342 |
++ |
6343 |
++/** |
6344 |
++ * bfq_group_chain_link - link an allocated group chain to a cgroup hierarchy. |
6345 |
++ * @bfqd: the queue descriptor. |
6346 |
++ * @css: the leaf cgroup_subsys_state to start from. |
6347 |
++ * @leaf: the leaf group (to be associated to @cgroup). |
6348 |
++ * |
6349 |
++ * Try to link a chain of groups to a cgroup hierarchy, connecting the |
6350 |
++ * nodes bottom-up, so we can be sure that when we find a cgroup in the |
6351 |
++ * hierarchy that already as a group associated to @bfqd all the nodes |
6352 |
++ * in the path to the root cgroup have one too. |
6353 |
++ * |
6354 |
++ * On locking: the queue lock protects the hierarchy (there is a hierarchy |
6355 |
++ * per device) while the bfqio_cgroup lock protects the list of groups |
6356 |
++ * belonging to the same cgroup. |
6357 |
++ */ |
6358 |
++static void bfq_group_chain_link(struct bfq_data *bfqd, |
6359 |
++ struct cgroup_subsys_state *css, |
6360 |
++ struct bfq_group *leaf) |
6361 |
++{ |
6362 |
++ struct bfqio_cgroup *bgrp; |
6363 |
++ struct bfq_group *bfqg, *next, *prev = NULL; |
6364 |
++ unsigned long flags; |
6365 |
++ |
6366 |
++ assert_spin_locked(bfqd->queue->queue_lock); |
6367 |
++ |
6368 |
++ for (; css != NULL && leaf != NULL; css = css->parent) { |
6369 |
++ bgrp = css_to_bfqio(css); |
6370 |
++ next = leaf->bfqd; |
6371 |
++ |
6372 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6373 |
++ BUG_ON(bfqg != NULL); |
6374 |
++ |
6375 |
++ spin_lock_irqsave(&bgrp->lock, flags); |
6376 |
++ |
6377 |
++ rcu_assign_pointer(leaf->bfqd, bfqd); |
6378 |
++ hlist_add_head_rcu(&leaf->group_node, &bgrp->group_data); |
6379 |
++ hlist_add_head(&leaf->bfqd_node, &bfqd->group_list); |
6380 |
++ |
6381 |
++ spin_unlock_irqrestore(&bgrp->lock, flags); |
6382 |
++ |
6383 |
++ prev = leaf; |
6384 |
++ leaf = next; |
6385 |
++ } |
6386 |
++ |
6387 |
++ BUG_ON(css == NULL && leaf != NULL); |
6388 |
++ if (css != NULL && prev != NULL) { |
6389 |
++ bgrp = css_to_bfqio(css); |
6390 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6391 |
++ bfq_group_set_parent(prev, bfqg); |
6392 |
++ } |
6393 |
++} |
6394 |
++ |
6395 |
++/** |
6396 |
++ * bfq_find_alloc_group - return the group associated to @bfqd in @cgroup. |
6397 |
++ * @bfqd: queue descriptor. |
6398 |
++ * @cgroup: cgroup being searched for. |
6399 |
++ * |
6400 |
++ * Return a group associated to @bfqd in @cgroup, allocating one if |
6401 |
++ * necessary. When a group is returned all the cgroups in the path |
6402 |
++ * to the root have a group associated to @bfqd. |
6403 |
++ * |
6404 |
++ * If the allocation fails, return the root group: this breaks guarantees |
6405 |
++ * but is a safe fallback. If this loss becomes a problem it can be |
6406 |
++ * mitigated using the equivalent weight (given by the product of the |
6407 |
++ * weights of the groups in the path from @group to the root) in the |
6408 |
++ * root scheduler. |
6409 |
++ * |
6410 |
++ * We allocate all the missing nodes in the path from the leaf cgroup |
6411 |
++ * to the root and we connect the nodes only after all the allocations |
6412 |
++ * have been successful. |
6413 |
++ */ |
6414 |
++static struct bfq_group *bfq_find_alloc_group(struct bfq_data *bfqd, |
6415 |
++ struct cgroup_subsys_state *css) |
6416 |
++{ |
6417 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
6418 |
++ struct bfq_group *bfqg; |
6419 |
++ |
6420 |
++ bfqg = bfqio_lookup_group(bgrp, bfqd); |
6421 |
++ if (bfqg != NULL) |
6422 |
++ return bfqg; |
6423 |
++ |
6424 |
++ bfqg = bfq_group_chain_alloc(bfqd, css); |
6425 |
++ if (bfqg != NULL) |
6426 |
++ bfq_group_chain_link(bfqd, css, bfqg); |
6427 |
++ else |
6428 |
++ bfqg = bfqd->root_group; |
6429 |
++ |
6430 |
++ return bfqg; |
6431 |
++} |
6432 |
++ |
6433 |
++/** |
6434 |
++ * bfq_bfqq_move - migrate @bfqq to @bfqg. |
6435 |
++ * @bfqd: queue descriptor. |
6436 |
++ * @bfqq: the queue to move. |
6437 |
++ * @entity: @bfqq's entity. |
6438 |
++ * @bfqg: the group to move to. |
6439 |
++ * |
6440 |
++ * Move @bfqq to @bfqg, deactivating it from its old group and reactivating |
6441 |
++ * it on the new one. Avoid putting the entity on the old group idle tree. |
6442 |
++ * |
6443 |
++ * Must be called under the queue lock; the cgroup owning @bfqg must |
6444 |
++ * not disappear (by now this just means that we are called under |
6445 |
++ * rcu_read_lock()). |
6446 |
++ */ |
6447 |
++static void bfq_bfqq_move(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
6448 |
++ struct bfq_entity *entity, struct bfq_group *bfqg) |
6449 |
++{ |
6450 |
++ int busy, resume; |
6451 |
++ |
6452 |
++ busy = bfq_bfqq_busy(bfqq); |
6453 |
++ resume = !RB_EMPTY_ROOT(&bfqq->sort_list); |
6454 |
++ |
6455 |
++ BUG_ON(resume && !entity->on_st); |
6456 |
++ BUG_ON(busy && !resume && entity->on_st && |
6457 |
++ bfqq != bfqd->in_service_queue); |
6458 |
++ |
6459 |
++ if (busy) { |
6460 |
++ BUG_ON(atomic_read(&bfqq->ref) < 2); |
6461 |
++ |
6462 |
++ if (!resume) |
6463 |
++ bfq_del_bfqq_busy(bfqd, bfqq, 0); |
6464 |
++ else |
6465 |
++ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
6466 |
++ } else if (entity->on_st) |
6467 |
++ bfq_put_idle_entity(bfq_entity_service_tree(entity), entity); |
6468 |
++ |
6469 |
++ /* |
6470 |
++ * Here we use a reference to bfqg. We don't need a refcounter |
6471 |
++ * as the cgroup reference will not be dropped, so that its |
6472 |
++ * destroy() callback will not be invoked. |
6473 |
++ */ |
6474 |
++ entity->parent = bfqg->my_entity; |
6475 |
++ entity->sched_data = &bfqg->sched_data; |
6476 |
++ |
6477 |
++ if (busy && resume) |
6478 |
++ bfq_activate_bfqq(bfqd, bfqq); |
6479 |
++ |
6480 |
++ if (bfqd->in_service_queue == NULL && !bfqd->rq_in_driver) |
6481 |
++ bfq_schedule_dispatch(bfqd); |
6482 |
++} |
6483 |
++ |
6484 |
++/** |
6485 |
++ * __bfq_bic_change_cgroup - move @bic to @cgroup. |
6486 |
++ * @bfqd: the queue descriptor. |
6487 |
++ * @bic: the bic to move. |
6488 |
++ * @cgroup: the cgroup to move to. |
6489 |
++ * |
6490 |
++ * Move bic to cgroup, assuming that bfqd->queue is locked; the caller |
6491 |
++ * has to make sure that the reference to cgroup is valid across the call. |
6492 |
++ * |
6493 |
++ * NOTE: an alternative approach might have been to store the current |
6494 |
++ * cgroup in bfqq and getting a reference to it, reducing the lookup |
6495 |
++ * time here, at the price of slightly more complex code. |
6496 |
++ */ |
6497 |
++static struct bfq_group *__bfq_bic_change_cgroup(struct bfq_data *bfqd, |
6498 |
++ struct bfq_io_cq *bic, |
6499 |
++ struct cgroup_subsys_state *css) |
6500 |
++{ |
6501 |
++ struct bfq_queue *async_bfqq = bic_to_bfqq(bic, 0); |
6502 |
++ struct bfq_queue *sync_bfqq = bic_to_bfqq(bic, 1); |
6503 |
++ struct bfq_entity *entity; |
6504 |
++ struct bfq_group *bfqg; |
6505 |
++ struct bfqio_cgroup *bgrp; |
6506 |
++ |
6507 |
++ bgrp = css_to_bfqio(css); |
6508 |
++ |
6509 |
++ bfqg = bfq_find_alloc_group(bfqd, css); |
6510 |
++ if (async_bfqq != NULL) { |
6511 |
++ entity = &async_bfqq->entity; |
6512 |
++ |
6513 |
++ if (entity->sched_data != &bfqg->sched_data) { |
6514 |
++ bic_set_bfqq(bic, NULL, 0); |
6515 |
++ bfq_log_bfqq(bfqd, async_bfqq, |
6516 |
++ "bic_change_group: %p %d", |
6517 |
++ async_bfqq, atomic_read(&async_bfqq->ref)); |
6518 |
++ bfq_put_queue(async_bfqq); |
6519 |
++ } |
6520 |
++ } |
6521 |
++ |
6522 |
++ if (sync_bfqq != NULL) { |
6523 |
++ entity = &sync_bfqq->entity; |
6524 |
++ if (entity->sched_data != &bfqg->sched_data) |
6525 |
++ bfq_bfqq_move(bfqd, sync_bfqq, entity, bfqg); |
6526 |
++ } |
6527 |
++ |
6528 |
++ return bfqg; |
6529 |
++} |
6530 |
++ |
6531 |
++/** |
6532 |
++ * bfq_bic_change_cgroup - move @bic to @cgroup. |
6533 |
++ * @bic: the bic being migrated. |
6534 |
++ * @cgroup: the destination cgroup. |
6535 |
++ * |
6536 |
++ * When the task owning @bic is moved to @cgroup, @bic is immediately |
6537 |
++ * moved into its new parent group. |
6538 |
++ */ |
6539 |
++static void bfq_bic_change_cgroup(struct bfq_io_cq *bic, |
6540 |
++ struct cgroup_subsys_state *css) |
6541 |
++{ |
6542 |
++ struct bfq_data *bfqd; |
6543 |
++ unsigned long uninitialized_var(flags); |
6544 |
++ |
6545 |
++ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), |
6546 |
++ &flags); |
6547 |
++ if (bfqd != NULL) { |
6548 |
++ __bfq_bic_change_cgroup(bfqd, bic, css); |
6549 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
6550 |
++ } |
6551 |
++} |
6552 |
++ |
6553 |
++/** |
6554 |
++ * bfq_bic_update_cgroup - update the cgroup of @bic. |
6555 |
++ * @bic: the @bic to update. |
6556 |
++ * |
6557 |
++ * Make sure that @bic is enqueued in the cgroup of the current task. |
6558 |
++ * We need this in addition to moving bics during the cgroup attach |
6559 |
++ * phase because the task owning @bic could be at its first disk |
6560 |
++ * access or we may end up in the root cgroup as the result of a |
6561 |
++ * memory allocation failure and here we try to move to the right |
6562 |
++ * group. |
6563 |
++ * |
6564 |
++ * Must be called under the queue lock. It is safe to use the returned |
6565 |
++ * value even after the rcu_read_unlock() as the migration/destruction |
6566 |
++ * paths act under the queue lock too. IOW it is impossible to race with |
6567 |
++ * group migration/destruction and end up with an invalid group as: |
6568 |
++ * a) here cgroup has not yet been destroyed, nor its destroy callback |
6569 |
++ * has started execution, as current holds a reference to it, |
6570 |
++ * b) if it is destroyed after rcu_read_unlock() [after current is |
6571 |
++ * migrated to a different cgroup] its attach() callback will have |
6572 |
++ * taken care of remove all the references to the old cgroup data. |
6573 |
++ */ |
6574 |
++static struct bfq_group *bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
6575 |
++{ |
6576 |
++ struct bfq_data *bfqd = bic_to_bfqd(bic); |
6577 |
++ struct bfq_group *bfqg; |
6578 |
++ struct cgroup_subsys_state *css; |
6579 |
++ |
6580 |
++ BUG_ON(bfqd == NULL); |
6581 |
++ |
6582 |
++ rcu_read_lock(); |
6583 |
++ css = task_css(current, bfqio_subsys_id); |
6584 |
++ bfqg = __bfq_bic_change_cgroup(bfqd, bic, css); |
6585 |
++ rcu_read_unlock(); |
6586 |
++ |
6587 |
++ return bfqg; |
6588 |
++} |
6589 |
++ |
6590 |
++/** |
6591 |
++ * bfq_flush_idle_tree - deactivate any entity on the idle tree of @st. |
6592 |
++ * @st: the service tree being flushed. |
6593 |
++ */ |
6594 |
++static inline void bfq_flush_idle_tree(struct bfq_service_tree *st) |
6595 |
++{ |
6596 |
++ struct bfq_entity *entity = st->first_idle; |
6597 |
++ |
6598 |
++ for (; entity != NULL; entity = st->first_idle) |
6599 |
++ __bfq_deactivate_entity(entity, 0); |
6600 |
++} |
6601 |
++ |
6602 |
++/** |
6603 |
++ * bfq_reparent_leaf_entity - move leaf entity to the root_group. |
6604 |
++ * @bfqd: the device data structure with the root group. |
6605 |
++ * @entity: the entity to move. |
6606 |
++ */ |
6607 |
++static inline void bfq_reparent_leaf_entity(struct bfq_data *bfqd, |
6608 |
++ struct bfq_entity *entity) |
6609 |
++{ |
6610 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
6611 |
++ |
6612 |
++ BUG_ON(bfqq == NULL); |
6613 |
++ bfq_bfqq_move(bfqd, bfqq, entity, bfqd->root_group); |
6614 |
++ return; |
6615 |
++} |
6616 |
++ |
6617 |
++/** |
6618 |
++ * bfq_reparent_active_entities - move to the root group all active entities. |
6619 |
++ * @bfqd: the device data structure with the root group. |
6620 |
++ * @bfqg: the group to move from. |
6621 |
++ * @st: the service tree with the entities. |
6622 |
++ * |
6623 |
++ * Needs queue_lock to be taken and reference to be valid over the call. |
6624 |
++ */ |
6625 |
++static inline void bfq_reparent_active_entities(struct bfq_data *bfqd, |
6626 |
++ struct bfq_group *bfqg, |
6627 |
++ struct bfq_service_tree *st) |
6628 |
++{ |
6629 |
++ struct rb_root *active = &st->active; |
6630 |
++ struct bfq_entity *entity = NULL; |
6631 |
++ |
6632 |
++ if (!RB_EMPTY_ROOT(&st->active)) |
6633 |
++ entity = bfq_entity_of(rb_first(active)); |
6634 |
++ |
6635 |
++ for (; entity != NULL; entity = bfq_entity_of(rb_first(active))) |
6636 |
++ bfq_reparent_leaf_entity(bfqd, entity); |
6637 |
++ |
6638 |
++ if (bfqg->sched_data.in_service_entity != NULL) |
6639 |
++ bfq_reparent_leaf_entity(bfqd, |
6640 |
++ bfqg->sched_data.in_service_entity); |
6641 |
++ |
6642 |
++ return; |
6643 |
++} |
6644 |
++ |
6645 |
++/** |
6646 |
++ * bfq_destroy_group - destroy @bfqg. |
6647 |
++ * @bgrp: the bfqio_cgroup containing @bfqg. |
6648 |
++ * @bfqg: the group being destroyed. |
6649 |
++ * |
6650 |
++ * Destroy @bfqg, making sure that it is not referenced from its parent. |
6651 |
++ */ |
6652 |
++static void bfq_destroy_group(struct bfqio_cgroup *bgrp, struct bfq_group *bfqg) |
6653 |
++{ |
6654 |
++ struct bfq_data *bfqd; |
6655 |
++ struct bfq_service_tree *st; |
6656 |
++ struct bfq_entity *entity = bfqg->my_entity; |
6657 |
++ unsigned long uninitialized_var(flags); |
6658 |
++ int i; |
6659 |
++ |
6660 |
++ hlist_del(&bfqg->group_node); |
6661 |
++ |
6662 |
++ /* |
6663 |
++ * Empty all service_trees belonging to this group before deactivating |
6664 |
++ * the group itself. |
6665 |
++ */ |
6666 |
++ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) { |
6667 |
++ st = bfqg->sched_data.service_tree + i; |
6668 |
++ |
6669 |
++ /* |
6670 |
++ * The idle tree may still contain bfq_queues belonging |
6671 |
++ * to exited task because they never migrated to a different |
6672 |
++ * cgroup from the one being destroyed now. No one else |
6673 |
++ * can access them so it's safe to act without any lock. |
6674 |
++ */ |
6675 |
++ bfq_flush_idle_tree(st); |
6676 |
++ |
6677 |
++ /* |
6678 |
++ * It may happen that some queues are still active |
6679 |
++ * (busy) upon group destruction (if the corresponding |
6680 |
++ * processes have been forced to terminate). We move |
6681 |
++ * all the leaf entities corresponding to these queues |
6682 |
++ * to the root_group. |
6683 |
++ * Also, it may happen that the group has an entity |
6684 |
++ * under service, which is disconnected from the active |
6685 |
++ * tree: it must be moved, too. |
6686 |
++ * There is no need to put the sync queues, as the |
6687 |
++ * scheduler has taken no reference. |
6688 |
++ */ |
6689 |
++ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
6690 |
++ if (bfqd != NULL) { |
6691 |
++ bfq_reparent_active_entities(bfqd, bfqg, st); |
6692 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
6693 |
++ } |
6694 |
++ BUG_ON(!RB_EMPTY_ROOT(&st->active)); |
6695 |
++ BUG_ON(!RB_EMPTY_ROOT(&st->idle)); |
6696 |
++ } |
6697 |
++ BUG_ON(bfqg->sched_data.next_in_service != NULL); |
6698 |
++ BUG_ON(bfqg->sched_data.in_service_entity != NULL); |
6699 |
++ |
6700 |
++ /* |
6701 |
++ * We may race with device destruction, take extra care when |
6702 |
++ * dereferencing bfqg->bfqd. |
6703 |
++ */ |
6704 |
++ bfqd = bfq_get_bfqd_locked(&bfqg->bfqd, &flags); |
6705 |
++ if (bfqd != NULL) { |
6706 |
++ hlist_del(&bfqg->bfqd_node); |
6707 |
++ __bfq_deactivate_entity(entity, 0); |
6708 |
++ bfq_put_async_queues(bfqd, bfqg); |
6709 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
6710 |
++ } |
6711 |
++ BUG_ON(entity->tree != NULL); |
6712 |
++ |
6713 |
++ /* |
6714 |
++ * No need to defer the kfree() to the end of the RCU grace |
6715 |
++ * period: we are called from the destroy() callback of our |
6716 |
++ * cgroup, so we can be sure that no one is a) still using |
6717 |
++ * this cgroup or b) doing lookups in it. |
6718 |
++ */ |
6719 |
++ kfree(bfqg); |
6720 |
++} |
6721 |
++ |
6722 |
++static void bfq_end_raising_async(struct bfq_data *bfqd) |
6723 |
++{ |
6724 |
++ struct hlist_node *tmp; |
6725 |
++ struct bfq_group *bfqg; |
6726 |
++ |
6727 |
++ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) |
6728 |
++ bfq_end_raising_async_queues(bfqd, bfqg); |
6729 |
++ bfq_end_raising_async_queues(bfqd, bfqd->root_group); |
6730 |
++} |
6731 |
++ |
6732 |
++/** |
6733 |
++ * bfq_disconnect_groups - disconnect @bfqd from all its groups. |
6734 |
++ * @bfqd: the device descriptor being exited. |
6735 |
++ * |
6736 |
++ * When the device exits we just make sure that no lookup can return |
6737 |
++ * the now unused group structures. They will be deallocated on cgroup |
6738 |
++ * destruction. |
6739 |
++ */ |
6740 |
++static void bfq_disconnect_groups(struct bfq_data *bfqd) |
6741 |
++{ |
6742 |
++ struct hlist_node *tmp; |
6743 |
++ struct bfq_group *bfqg; |
6744 |
++ |
6745 |
++ bfq_log(bfqd, "disconnect_groups beginning"); |
6746 |
++ hlist_for_each_entry_safe(bfqg, tmp, &bfqd->group_list, bfqd_node) { |
6747 |
++ hlist_del(&bfqg->bfqd_node); |
6748 |
++ |
6749 |
++ __bfq_deactivate_entity(bfqg->my_entity, 0); |
6750 |
++ |
6751 |
++ /* |
6752 |
++ * Don't remove from the group hash, just set an |
6753 |
++ * invalid key. No lookups can race with the |
6754 |
++ * assignment as bfqd is being destroyed; this |
6755 |
++ * implies also that new elements cannot be added |
6756 |
++ * to the list. |
6757 |
++ */ |
6758 |
++ rcu_assign_pointer(bfqg->bfqd, NULL); |
6759 |
++ |
6760 |
++ bfq_log(bfqd, "disconnect_groups: put async for group %p", |
6761 |
++ bfqg); |
6762 |
++ bfq_put_async_queues(bfqd, bfqg); |
6763 |
++ } |
6764 |
++} |
6765 |
++ |
6766 |
++static inline void bfq_free_root_group(struct bfq_data *bfqd) |
6767 |
++{ |
6768 |
++ struct bfqio_cgroup *bgrp = &bfqio_root_cgroup; |
6769 |
++ struct bfq_group *bfqg = bfqd->root_group; |
6770 |
++ |
6771 |
++ bfq_put_async_queues(bfqd, bfqg); |
6772 |
++ |
6773 |
++ spin_lock_irq(&bgrp->lock); |
6774 |
++ hlist_del_rcu(&bfqg->group_node); |
6775 |
++ spin_unlock_irq(&bgrp->lock); |
6776 |
++ |
6777 |
++ /* |
6778 |
++ * No need to synchronize_rcu() here: since the device is gone |
6779 |
++ * there cannot be any read-side access to its root_group. |
6780 |
++ */ |
6781 |
++ kfree(bfqg); |
6782 |
++} |
6783 |
++ |
6784 |
++static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
6785 |
++{ |
6786 |
++ struct bfq_group *bfqg; |
6787 |
++ struct bfqio_cgroup *bgrp; |
6788 |
++ int i; |
6789 |
++ |
6790 |
++ bfqg = kzalloc_node(sizeof(*bfqg), GFP_KERNEL, node); |
6791 |
++ if (bfqg == NULL) |
6792 |
++ return NULL; |
6793 |
++ |
6794 |
++ bfqg->entity.parent = NULL; |
6795 |
++ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
6796 |
++ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
6797 |
++ |
6798 |
++ bgrp = &bfqio_root_cgroup; |
6799 |
++ spin_lock_irq(&bgrp->lock); |
6800 |
++ rcu_assign_pointer(bfqg->bfqd, bfqd); |
6801 |
++ hlist_add_head_rcu(&bfqg->group_node, &bgrp->group_data); |
6802 |
++ spin_unlock_irq(&bgrp->lock); |
6803 |
++ |
6804 |
++ return bfqg; |
6805 |
++} |
6806 |
++ |
6807 |
++#define SHOW_FUNCTION(__VAR) \ |
6808 |
++static u64 bfqio_cgroup_##__VAR##_read(struct cgroup_subsys_state *css, \ |
6809 |
++ struct cftype *cftype) \ |
6810 |
++{ \ |
6811 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
6812 |
++ u64 ret = -ENODEV; \ |
6813 |
++ \ |
6814 |
++ mutex_lock(&bfqio_mutex); \ |
6815 |
++ if (bfqio_is_removed(bgrp)) \ |
6816 |
++ goto out_unlock; \ |
6817 |
++ \ |
6818 |
++ spin_lock_irq(&bgrp->lock); \ |
6819 |
++ ret = bgrp->__VAR; \ |
6820 |
++ spin_unlock_irq(&bgrp->lock); \ |
6821 |
++ \ |
6822 |
++out_unlock: \ |
6823 |
++ mutex_unlock(&bfqio_mutex); \ |
6824 |
++ return ret; \ |
6825 |
++} |
6826 |
++ |
6827 |
++SHOW_FUNCTION(weight); |
6828 |
++SHOW_FUNCTION(ioprio); |
6829 |
++SHOW_FUNCTION(ioprio_class); |
6830 |
++#undef SHOW_FUNCTION |
6831 |
++ |
6832 |
++#define STORE_FUNCTION(__VAR, __MIN, __MAX) \ |
6833 |
++static int bfqio_cgroup_##__VAR##_write(struct cgroup_subsys_state *css,\ |
6834 |
++ struct cftype *cftype, \ |
6835 |
++ u64 val) \ |
6836 |
++{ \ |
6837 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); \ |
6838 |
++ struct bfq_group *bfqg; \ |
6839 |
++ int ret = -EINVAL; \ |
6840 |
++ \ |
6841 |
++ if (val < (__MIN) || val > (__MAX)) \ |
6842 |
++ return ret; \ |
6843 |
++ \ |
6844 |
++ ret = -ENODEV; \ |
6845 |
++ mutex_lock(&bfqio_mutex); \ |
6846 |
++ if (bfqio_is_removed(bgrp)) \ |
6847 |
++ goto out_unlock; \ |
6848 |
++ ret = 0; \ |
6849 |
++ \ |
6850 |
++ spin_lock_irq(&bgrp->lock); \ |
6851 |
++ bgrp->__VAR = (unsigned short)val; \ |
6852 |
++ hlist_for_each_entry(bfqg, &bgrp->group_data, group_node) { \ |
6853 |
++ /* \ |
6854 |
++ * Setting the ioprio_changed flag of the entity \ |
6855 |
++ * to 1 with new_##__VAR == ##__VAR would re-set \ |
6856 |
++ * the value of the weight to its ioprio mapping. \ |
6857 |
++ * Set the flag only if necessary. \ |
6858 |
++ */ \ |
6859 |
++ if ((unsigned short)val != bfqg->entity.new_##__VAR) { \ |
6860 |
++ bfqg->entity.new_##__VAR = (unsigned short)val; \ |
6861 |
++ /* \ |
6862 |
++ * Make sure that the above new value has been \ |
6863 |
++ * stored in bfqg->entity.new_##__VAR before \ |
6864 |
++ * setting the ioprio_changed flag. In fact, \ |
6865 |
++ * this flag may be read asynchronously (in \ |
6866 |
++ * critical sections protected by a different \ |
6867 |
++ * lock than that held here), and finding this \ |
6868 |
++ * flag set may cause the execution of the code \ |
6869 |
++ * for updating parameters whose value may \ |
6870 |
++ * depend also on bfqg->entity.new_##__VAR (in \ |
6871 |
++ * __bfq_entity_update_weight_prio). \ |
6872 |
++ * This barrier makes sure that the new value \ |
6873 |
++ * of bfqg->entity.new_##__VAR is correctly \ |
6874 |
++ * seen in that code. \ |
6875 |
++ */ \ |
6876 |
++ smp_wmb(); \ |
6877 |
++ bfqg->entity.ioprio_changed = 1; \ |
6878 |
++ } \ |
6879 |
++ } \ |
6880 |
++ spin_unlock_irq(&bgrp->lock); \ |
6881 |
++ \ |
6882 |
++out_unlock: \ |
6883 |
++ mutex_unlock(&bfqio_mutex); \ |
6884 |
++ return ret; \ |
6885 |
++} |
6886 |
++ |
6887 |
++STORE_FUNCTION(weight, BFQ_MIN_WEIGHT, BFQ_MAX_WEIGHT); |
6888 |
++STORE_FUNCTION(ioprio, 0, IOPRIO_BE_NR - 1); |
6889 |
++STORE_FUNCTION(ioprio_class, IOPRIO_CLASS_RT, IOPRIO_CLASS_IDLE); |
6890 |
++#undef STORE_FUNCTION |
6891 |
++ |
6892 |
++static struct cftype bfqio_files[] = { |
6893 |
++ { |
6894 |
++ .name = "weight", |
6895 |
++ .read_u64 = bfqio_cgroup_weight_read, |
6896 |
++ .write_u64 = bfqio_cgroup_weight_write, |
6897 |
++ }, |
6898 |
++ { |
6899 |
++ .name = "ioprio", |
6900 |
++ .read_u64 = bfqio_cgroup_ioprio_read, |
6901 |
++ .write_u64 = bfqio_cgroup_ioprio_write, |
6902 |
++ }, |
6903 |
++ { |
6904 |
++ .name = "ioprio_class", |
6905 |
++ .read_u64 = bfqio_cgroup_ioprio_class_read, |
6906 |
++ .write_u64 = bfqio_cgroup_ioprio_class_write, |
6907 |
++ }, |
6908 |
++ { }, /* terminate */ |
6909 |
++}; |
6910 |
++ |
6911 |
++static struct cgroup_subsys_state *bfqio_create(struct cgroup_subsys_state |
6912 |
++ *parent_css) |
6913 |
++{ |
6914 |
++ struct bfqio_cgroup *bgrp; |
6915 |
++ |
6916 |
++ if (parent_css != NULL) { |
6917 |
++ bgrp = kzalloc(sizeof(*bgrp), GFP_KERNEL); |
6918 |
++ if (bgrp == NULL) |
6919 |
++ return ERR_PTR(-ENOMEM); |
6920 |
++ } else |
6921 |
++ bgrp = &bfqio_root_cgroup; |
6922 |
++ |
6923 |
++ spin_lock_init(&bgrp->lock); |
6924 |
++ INIT_HLIST_HEAD(&bgrp->group_data); |
6925 |
++ bgrp->ioprio = BFQ_DEFAULT_GRP_IOPRIO; |
6926 |
++ bgrp->ioprio_class = BFQ_DEFAULT_GRP_CLASS; |
6927 |
++ |
6928 |
++ return &bgrp->css; |
6929 |
++} |
6930 |
++ |
6931 |
++/* |
6932 |
++ * We cannot support shared io contexts, as we have no means to support |
6933 |
++ * two tasks with the same ioc in two different groups without major rework |
6934 |
++ * of the main bic/bfqq data structures. By now we allow a task to change |
6935 |
++ * its cgroup only if it's the only owner of its ioc; the drawback of this |
6936 |
++ * behavior is that a group containing a task that forked using CLONE_IO |
6937 |
++ * will not be destroyed until the tasks sharing the ioc die. |
6938 |
++ */ |
6939 |
++static int bfqio_can_attach(struct cgroup_subsys_state *css, |
6940 |
++ struct cgroup_taskset *tset) |
6941 |
++{ |
6942 |
++ struct task_struct *task; |
6943 |
++ struct io_context *ioc; |
6944 |
++ int ret = 0; |
6945 |
++ |
6946 |
++ cgroup_taskset_for_each(task, css, tset) { |
6947 |
++ /* |
6948 |
++ * task_lock() is needed to avoid races with |
6949 |
++ * exit_io_context() |
6950 |
++ */ |
6951 |
++ task_lock(task); |
6952 |
++ ioc = task->io_context; |
6953 |
++ if (ioc != NULL && atomic_read(&ioc->nr_tasks) > 1) |
6954 |
++ /* |
6955 |
++ * ioc == NULL means that the task is either too young |
6956 |
++ * or exiting: if it has still no ioc the ioc can't be |
6957 |
++ * shared, if the task is exiting the attach will fail |
6958 |
++ * anyway, no matter what we return here. |
6959 |
++ */ |
6960 |
++ ret = -EINVAL; |
6961 |
++ task_unlock(task); |
6962 |
++ if (ret) |
6963 |
++ break; |
6964 |
++ } |
6965 |
++ |
6966 |
++ return ret; |
6967 |
++} |
6968 |
++ |
6969 |
++static void bfqio_attach(struct cgroup_subsys_state *css, |
6970 |
++ struct cgroup_taskset *tset) |
6971 |
++{ |
6972 |
++ struct task_struct *task; |
6973 |
++ struct io_context *ioc; |
6974 |
++ struct io_cq *icq; |
6975 |
++ |
6976 |
++ /* |
6977 |
++ * IMPORTANT NOTE: The move of more than one process at a time to a |
6978 |
++ * new group has not yet been tested. |
6979 |
++ */ |
6980 |
++ cgroup_taskset_for_each(task, css, tset) { |
6981 |
++ ioc = get_task_io_context(task, GFP_ATOMIC, NUMA_NO_NODE); |
6982 |
++ if (ioc) { |
6983 |
++ /* |
6984 |
++ * Handle cgroup change here. |
6985 |
++ */ |
6986 |
++ rcu_read_lock(); |
6987 |
++ hlist_for_each_entry_rcu(icq, &ioc->icq_list, ioc_node) |
6988 |
++ if (!strncmp( |
6989 |
++ icq->q->elevator->type->elevator_name, |
6990 |
++ "bfq", ELV_NAME_MAX)) |
6991 |
++ bfq_bic_change_cgroup(icq_to_bic(icq), |
6992 |
++ css); |
6993 |
++ rcu_read_unlock(); |
6994 |
++ put_io_context(ioc); |
6995 |
++ } |
6996 |
++ } |
6997 |
++} |
6998 |
++ |
6999 |
++static void bfqio_destroy(struct cgroup_subsys_state *css) |
7000 |
++{ |
7001 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
7002 |
++ struct hlist_node *tmp; |
7003 |
++ struct bfq_group *bfqg; |
7004 |
++ |
7005 |
++ /* |
7006 |
++ * Since we are destroying the cgroup, there are no more tasks |
7007 |
++ * referencing it, and all the RCU grace periods that may have |
7008 |
++ * referenced it are ended (as the destruction of the parent |
7009 |
++ * cgroup is RCU-safe); bgrp->group_data will not be accessed by |
7010 |
++ * anything else and we don't need any synchronization. |
7011 |
++ */ |
7012 |
++ hlist_for_each_entry_safe(bfqg, tmp, &bgrp->group_data, group_node) |
7013 |
++ bfq_destroy_group(bgrp, bfqg); |
7014 |
++ |
7015 |
++ BUG_ON(!hlist_empty(&bgrp->group_data)); |
7016 |
++ |
7017 |
++ kfree(bgrp); |
7018 |
++} |
7019 |
++ |
7020 |
++static int bfqio_css_online(struct cgroup_subsys_state *css) |
7021 |
++{ |
7022 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
7023 |
++ |
7024 |
++ mutex_lock(&bfqio_mutex); |
7025 |
++ bgrp->online = true; |
7026 |
++ mutex_unlock(&bfqio_mutex); |
7027 |
++ |
7028 |
++ return 0; |
7029 |
++} |
7030 |
++ |
7031 |
++static void bfqio_css_offline(struct cgroup_subsys_state *css) |
7032 |
++{ |
7033 |
++ struct bfqio_cgroup *bgrp = css_to_bfqio(css); |
7034 |
++ |
7035 |
++ mutex_lock(&bfqio_mutex); |
7036 |
++ bgrp->online = false; |
7037 |
++ mutex_unlock(&bfqio_mutex); |
7038 |
++} |
7039 |
++ |
7040 |
++struct cgroup_subsys bfqio_subsys = { |
7041 |
++ .name = "bfqio", |
7042 |
++ .css_alloc = bfqio_create, |
7043 |
++ .css_online = bfqio_css_online, |
7044 |
++ .css_offline = bfqio_css_offline, |
7045 |
++ .can_attach = bfqio_can_attach, |
7046 |
++ .attach = bfqio_attach, |
7047 |
++ .css_free = bfqio_destroy, |
7048 |
++ .subsys_id = bfqio_subsys_id, |
7049 |
++ .base_cftypes = bfqio_files, |
7050 |
++}; |
7051 |
++#else |
7052 |
++static inline void bfq_init_entity(struct bfq_entity *entity, |
7053 |
++ struct bfq_group *bfqg) |
7054 |
++{ |
7055 |
++ entity->weight = entity->new_weight; |
7056 |
++ entity->orig_weight = entity->new_weight; |
7057 |
++ entity->ioprio = entity->new_ioprio; |
7058 |
++ entity->ioprio_class = entity->new_ioprio_class; |
7059 |
++ entity->sched_data = &bfqg->sched_data; |
7060 |
++} |
7061 |
++ |
7062 |
++static inline struct bfq_group * |
7063 |
++bfq_bic_update_cgroup(struct bfq_io_cq *bic) |
7064 |
++{ |
7065 |
++ struct bfq_data *bfqd = bic_to_bfqd(bic); |
7066 |
++ return bfqd->root_group; |
7067 |
++} |
7068 |
++ |
7069 |
++static inline void bfq_bfqq_move(struct bfq_data *bfqd, |
7070 |
++ struct bfq_queue *bfqq, |
7071 |
++ struct bfq_entity *entity, |
7072 |
++ struct bfq_group *bfqg) |
7073 |
++{ |
7074 |
++} |
7075 |
++ |
7076 |
++static void bfq_end_raising_async(struct bfq_data *bfqd) |
7077 |
++{ |
7078 |
++ bfq_end_raising_async_queues(bfqd, bfqd->root_group); |
7079 |
++} |
7080 |
++ |
7081 |
++static inline void bfq_disconnect_groups(struct bfq_data *bfqd) |
7082 |
++{ |
7083 |
++ bfq_put_async_queues(bfqd, bfqd->root_group); |
7084 |
++} |
7085 |
++ |
7086 |
++static inline void bfq_free_root_group(struct bfq_data *bfqd) |
7087 |
++{ |
7088 |
++ kfree(bfqd->root_group); |
7089 |
++} |
7090 |
++ |
7091 |
++static struct bfq_group *bfq_alloc_root_group(struct bfq_data *bfqd, int node) |
7092 |
++{ |
7093 |
++ struct bfq_group *bfqg; |
7094 |
++ int i; |
7095 |
++ |
7096 |
++ bfqg = kmalloc_node(sizeof(*bfqg), GFP_KERNEL | __GFP_ZERO, node); |
7097 |
++ if (bfqg == NULL) |
7098 |
++ return NULL; |
7099 |
++ |
7100 |
++ for (i = 0; i < BFQ_IOPRIO_CLASSES; i++) |
7101 |
++ bfqg->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT; |
7102 |
++ |
7103 |
++ return bfqg; |
7104 |
++} |
7105 |
++#endif |
7106 |
+diff --git a/block/bfq-ioc.c b/block/bfq-ioc.c |
7107 |
+new file mode 100644 |
7108 |
+index 0000000..7f6b000 |
7109 |
+--- /dev/null |
7110 |
++++ b/block/bfq-ioc.c |
7111 |
+@@ -0,0 +1,36 @@ |
7112 |
++/* |
7113 |
++ * BFQ: I/O context handling. |
7114 |
++ * |
7115 |
++ * Based on ideas and code from CFQ: |
7116 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
7117 |
++ * |
7118 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
7119 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
7120 |
++ * |
7121 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
7122 |
++ */ |
7123 |
++ |
7124 |
++/** |
7125 |
++ * icq_to_bic - convert iocontext queue structure to bfq_io_cq. |
7126 |
++ * @icq: the iocontext queue. |
7127 |
++ */ |
7128 |
++static inline struct bfq_io_cq *icq_to_bic(struct io_cq *icq) |
7129 |
++{ |
7130 |
++ /* bic->icq is the first member, %NULL will convert to %NULL */ |
7131 |
++ return container_of(icq, struct bfq_io_cq, icq); |
7132 |
++} |
7133 |
++ |
7134 |
++/** |
7135 |
++ * bfq_bic_lookup - search into @ioc a bic associated to @bfqd. |
7136 |
++ * @bfqd: the lookup key. |
7137 |
++ * @ioc: the io_context of the process doing I/O. |
7138 |
++ * |
7139 |
++ * Queue lock must be held. |
7140 |
++ */ |
7141 |
++static inline struct bfq_io_cq *bfq_bic_lookup(struct bfq_data *bfqd, |
7142 |
++ struct io_context *ioc) |
7143 |
++{ |
7144 |
++ if (ioc) |
7145 |
++ return icq_to_bic(ioc_lookup_icq(ioc, bfqd->queue)); |
7146 |
++ return NULL; |
7147 |
++} |
7148 |
+diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c |
7149 |
+new file mode 100644 |
7150 |
+index 0000000..f5f71e4 |
7151 |
+--- /dev/null |
7152 |
++++ b/block/bfq-iosched.c |
7153 |
+@@ -0,0 +1,3300 @@ |
7154 |
++/* |
7155 |
++ * Budget Fair Queueing (BFQ) disk scheduler. |
7156 |
++ * |
7157 |
++ * Based on ideas and code from CFQ: |
7158 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
7159 |
++ * |
7160 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
7161 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
7162 |
++ * |
7163 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
7164 |
++ * |
7165 |
++ * Licensed under the GPL-2 as detailed in the accompanying COPYING.BFQ file. |
7166 |
++ * |
7167 |
++ * BFQ is a proportional share disk scheduling algorithm based on the |
7168 |
++ * slice-by-slice service scheme of CFQ. But BFQ assigns budgets, measured in |
7169 |
++ * number of sectors, to tasks instead of time slices. The disk is not granted |
7170 |
++ * to the in-service task for a given time slice, but until it has exhausted |
7171 |
++ * its assigned budget. This change from the time to the service domain allows |
7172 |
++ * BFQ to distribute the disk bandwidth among tasks as desired, without any |
7173 |
++ * distortion due to ZBR, workload fluctuations or other factors. BFQ uses an |
7174 |
++ * ad hoc internal scheduler, called B-WF2Q+, to schedule tasks according to |
7175 |
++ * their budgets (more precisely BFQ schedules queues associated to tasks). |
7176 |
++ * Thanks to this accurate scheduler, BFQ can afford to assign high budgets to |
7177 |
++ * disk-bound non-seeky tasks (to boost the throughput), and yet guarantee low |
7178 |
++ * latencies to interactive and soft real-time applications. |
7179 |
++ * |
7180 |
++ * BFQ is described in [1], where also a reference to the initial, more |
7181 |
++ * theoretical paper on BFQ can be found. The interested reader can find in |
7182 |
++ * the latter paper full details on the main algorithm as well as formulas of |
7183 |
++ * the guarantees, plus formal proofs of all the properties. With respect to |
7184 |
++ * the version of BFQ presented in these papers, this implementation adds a |
7185 |
++ * few more heuristics, such as the one that guarantees a low latency to soft |
7186 |
++ * real-time applications, and a hierarchical extension based on H-WF2Q+. |
7187 |
++ * |
7188 |
++ * B-WF2Q+ is based on WF2Q+, that is described in [2], together with |
7189 |
++ * H-WF2Q+, while the augmented tree used to implement B-WF2Q+ with O(log N) |
7190 |
++ * complexity derives from the one introduced with EEVDF in [3]. |
7191 |
++ * |
7192 |
++ * [1] P. Valente and M. Andreolini, ``Improving Application Responsiveness |
7193 |
++ * with the BFQ Disk I/O Scheduler'', |
7194 |
++ * Proceedings of the 5th Annual International Systems and Storage |
7195 |
++ * Conference (SYSTOR '12), June 2012. |
7196 |
++ * |
7197 |
++ * http://algogroup.unimo.it/people/paolo/disk_sched/bf1-v1-suite-results.pdf |
7198 |
++ * |
7199 |
++ * [2] Jon C.R. Bennett and H. Zhang, ``Hierarchical Packet Fair Queueing |
7200 |
++ * Algorithms,'' IEEE/ACM Transactions on Networking, 5(5):675-689, |
7201 |
++ * Oct 1997. |
7202 |
++ * |
7203 |
++ * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz |
7204 |
++ * |
7205 |
++ * [3] I. Stoica and H. Abdel-Wahab, ``Earliest Eligible Virtual Deadline |
7206 |
++ * First: A Flexible and Accurate Mechanism for Proportional Share |
7207 |
++ * Resource Allocation,'' technical report. |
7208 |
++ * |
7209 |
++ * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf |
7210 |
++ */ |
7211 |
++#include <linux/module.h> |
7212 |
++#include <linux/slab.h> |
7213 |
++#include <linux/blkdev.h> |
7214 |
++#include <linux/cgroup.h> |
7215 |
++#include <linux/elevator.h> |
7216 |
++#include <linux/jiffies.h> |
7217 |
++#include <linux/rbtree.h> |
7218 |
++#include <linux/ioprio.h> |
7219 |
++#include "bfq.h" |
7220 |
++#include "blk.h" |
7221 |
++ |
7222 |
++/* Max number of dispatches in one round of service. */ |
7223 |
++static const int bfq_quantum = 4; |
7224 |
++ |
7225 |
++/* Expiration time of sync (0) and async (1) requests, in jiffies. */ |
7226 |
++static const int bfq_fifo_expire[2] = { HZ / 4, HZ / 8 }; |
7227 |
++ |
7228 |
++/* Maximum backwards seek, in KiB. */ |
7229 |
++static const int bfq_back_max = 16 * 1024; |
7230 |
++ |
7231 |
++/* Penalty of a backwards seek, in number of sectors. */ |
7232 |
++static const int bfq_back_penalty = 2; |
7233 |
++ |
7234 |
++/* Idling period duration, in jiffies. */ |
7235 |
++static int bfq_slice_idle = HZ / 125; |
7236 |
++ |
7237 |
++/* Default maximum budget values, in sectors and number of requests. */ |
7238 |
++static const int bfq_default_max_budget = 16 * 1024; |
7239 |
++static const int bfq_max_budget_async_rq = 4; |
7240 |
++ |
7241 |
++/* |
7242 |
++ * Async to sync throughput distribution is controlled as follows: |
7243 |
++ * when an async request is served, the entity is charged the number |
7244 |
++ * of sectors of the request, multiplied by the factor below |
7245 |
++ */ |
7246 |
++static const int bfq_async_charge_factor = 10; |
7247 |
++ |
7248 |
++/* Default timeout values, in jiffies, approximating CFQ defaults. */ |
7249 |
++static const int bfq_timeout_sync = HZ / 8; |
7250 |
++static int bfq_timeout_async = HZ / 25; |
7251 |
++ |
7252 |
++struct kmem_cache *bfq_pool; |
7253 |
++ |
7254 |
++/* Below this threshold (in ms), we consider thinktime immediate. */ |
7255 |
++#define BFQ_MIN_TT 2 |
7256 |
++ |
7257 |
++/* hw_tag detection: parallel requests threshold and min samples needed. */ |
7258 |
++#define BFQ_HW_QUEUE_THRESHOLD 4 |
7259 |
++#define BFQ_HW_QUEUE_SAMPLES 32 |
7260 |
++ |
7261 |
++#define BFQQ_SEEK_THR (sector_t)(8 * 1024) |
7262 |
++#define BFQQ_SEEKY(bfqq) ((bfqq)->seek_mean > BFQQ_SEEK_THR) |
7263 |
++ |
7264 |
++/* Min samples used for peak rate estimation (for autotuning). */ |
7265 |
++#define BFQ_PEAK_RATE_SAMPLES 32 |
7266 |
++ |
7267 |
++/* Shift used for peak rate fixed precision calculations. */ |
7268 |
++#define BFQ_RATE_SHIFT 16 |
7269 |
++ |
7270 |
++/* |
7271 |
++ * The duration of the weight raising for interactive applications is |
7272 |
++ * computed automatically (as default behaviour), using the following |
7273 |
++ * formula: duration = (R / r) * T, where r is the peak rate of the |
7274 |
++ * disk, and R and T are two reference parameters. In particular, R is |
7275 |
++ * the peak rate of a reference disk, and T is about the maximum time |
7276 |
++ * for starting popular large applications on that disk, under BFQ and |
7277 |
++ * while reading two files in parallel. Finally, BFQ uses two |
7278 |
++ * different pairs (R, T) depending on whether the disk is rotational |
7279 |
++ * or non-rotational. |
7280 |
++ */ |
7281 |
++#define T_rot (msecs_to_jiffies(5500)) |
7282 |
++#define T_nonrot (msecs_to_jiffies(2000)) |
7283 |
++/* Next two quantities are in sectors/usec, left-shifted by BFQ_RATE_SHIFT */ |
7284 |
++#define R_rot 17415 |
7285 |
++#define R_nonrot 34791 |
7286 |
++ |
7287 |
++#define BFQ_SERVICE_TREE_INIT ((struct bfq_service_tree) \ |
7288 |
++ { RB_ROOT, RB_ROOT, NULL, NULL, 0, 0 }) |
7289 |
++ |
7290 |
++#define RQ_BIC(rq) ((struct bfq_io_cq *) (rq)->elv.priv[0]) |
7291 |
++#define RQ_BFQQ(rq) ((rq)->elv.priv[1]) |
7292 |
++ |
7293 |
++static inline void bfq_schedule_dispatch(struct bfq_data *bfqd); |
7294 |
++ |
7295 |
++#include "bfq-ioc.c" |
7296 |
++#include "bfq-sched.c" |
7297 |
++#include "bfq-cgroup.c" |
7298 |
++ |
7299 |
++#define bfq_class_idle(bfqq) ((bfqq)->entity.ioprio_class ==\ |
7300 |
++ IOPRIO_CLASS_IDLE) |
7301 |
++#define bfq_class_rt(bfqq) ((bfqq)->entity.ioprio_class ==\ |
7302 |
++ IOPRIO_CLASS_RT) |
7303 |
++ |
7304 |
++#define bfq_sample_valid(samples) ((samples) > 80) |
7305 |
++ |
7306 |
++/* |
7307 |
++ * We regard a request as SYNC, if either it's a read or has the SYNC bit |
7308 |
++ * set (in which case it could also be a direct WRITE). |
7309 |
++ */ |
7310 |
++static inline int bfq_bio_sync(struct bio *bio) |
7311 |
++{ |
7312 |
++ if (bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC)) |
7313 |
++ return 1; |
7314 |
++ |
7315 |
++ return 0; |
7316 |
++} |
7317 |
++ |
7318 |
++/* |
7319 |
++ * Scheduler run of queue, if there are requests pending and no one in the |
7320 |
++ * driver that will restart queueing. |
7321 |
++ */ |
7322 |
++static inline void bfq_schedule_dispatch(struct bfq_data *bfqd) |
7323 |
++{ |
7324 |
++ if (bfqd->queued != 0) { |
7325 |
++ bfq_log(bfqd, "schedule dispatch"); |
7326 |
++ kblockd_schedule_work(bfqd->queue, &bfqd->unplug_work); |
7327 |
++ } |
7328 |
++} |
7329 |
++ |
7330 |
++/* |
7331 |
++ * Lifted from AS - choose which of rq1 and rq2 that is best served now. |
7332 |
++ * We choose the request that is closesr to the head right now. Distance |
7333 |
++ * behind the head is penalized and only allowed to a certain extent. |
7334 |
++ */ |
7335 |
++static struct request *bfq_choose_req(struct bfq_data *bfqd, |
7336 |
++ struct request *rq1, |
7337 |
++ struct request *rq2, |
7338 |
++ sector_t last) |
7339 |
++{ |
7340 |
++ sector_t s1, s2, d1 = 0, d2 = 0; |
7341 |
++ unsigned long back_max; |
7342 |
++#define BFQ_RQ1_WRAP 0x01 /* request 1 wraps */ |
7343 |
++#define BFQ_RQ2_WRAP 0x02 /* request 2 wraps */ |
7344 |
++ unsigned wrap = 0; /* bit mask: requests behind the disk head? */ |
7345 |
++ |
7346 |
++ if (rq1 == NULL || rq1 == rq2) |
7347 |
++ return rq2; |
7348 |
++ if (rq2 == NULL) |
7349 |
++ return rq1; |
7350 |
++ |
7351 |
++ if (rq_is_sync(rq1) && !rq_is_sync(rq2)) |
7352 |
++ return rq1; |
7353 |
++ else if (rq_is_sync(rq2) && !rq_is_sync(rq1)) |
7354 |
++ return rq2; |
7355 |
++ if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META)) |
7356 |
++ return rq1; |
7357 |
++ else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META)) |
7358 |
++ return rq2; |
7359 |
++ |
7360 |
++ s1 = blk_rq_pos(rq1); |
7361 |
++ s2 = blk_rq_pos(rq2); |
7362 |
++ |
7363 |
++ /* |
7364 |
++ * By definition, 1KiB is 2 sectors. |
7365 |
++ */ |
7366 |
++ back_max = bfqd->bfq_back_max * 2; |
7367 |
++ |
7368 |
++ /* |
7369 |
++ * Strict one way elevator _except_ in the case where we allow |
7370 |
++ * short backward seeks which are biased as twice the cost of a |
7371 |
++ * similar forward seek. |
7372 |
++ */ |
7373 |
++ if (s1 >= last) |
7374 |
++ d1 = s1 - last; |
7375 |
++ else if (s1 + back_max >= last) |
7376 |
++ d1 = (last - s1) * bfqd->bfq_back_penalty; |
7377 |
++ else |
7378 |
++ wrap |= BFQ_RQ1_WRAP; |
7379 |
++ |
7380 |
++ if (s2 >= last) |
7381 |
++ d2 = s2 - last; |
7382 |
++ else if (s2 + back_max >= last) |
7383 |
++ d2 = (last - s2) * bfqd->bfq_back_penalty; |
7384 |
++ else |
7385 |
++ wrap |= BFQ_RQ2_WRAP; |
7386 |
++ |
7387 |
++ /* Found required data */ |
7388 |
++ |
7389 |
++ /* |
7390 |
++ * By doing switch() on the bit mask "wrap" we avoid having to |
7391 |
++ * check two variables for all permutations: --> faster! |
7392 |
++ */ |
7393 |
++ switch (wrap) { |
7394 |
++ case 0: /* common case for CFQ: rq1 and rq2 not wrapped */ |
7395 |
++ if (d1 < d2) |
7396 |
++ return rq1; |
7397 |
++ else if (d2 < d1) |
7398 |
++ return rq2; |
7399 |
++ else { |
7400 |
++ if (s1 >= s2) |
7401 |
++ return rq1; |
7402 |
++ else |
7403 |
++ return rq2; |
7404 |
++ } |
7405 |
++ |
7406 |
++ case BFQ_RQ2_WRAP: |
7407 |
++ return rq1; |
7408 |
++ case BFQ_RQ1_WRAP: |
7409 |
++ return rq2; |
7410 |
++ case (BFQ_RQ1_WRAP|BFQ_RQ2_WRAP): /* both rqs wrapped */ |
7411 |
++ default: |
7412 |
++ /* |
7413 |
++ * Since both rqs are wrapped, |
7414 |
++ * start with the one that's further behind head |
7415 |
++ * (--> only *one* back seek required), |
7416 |
++ * since back seek takes more time than forward. |
7417 |
++ */ |
7418 |
++ if (s1 <= s2) |
7419 |
++ return rq1; |
7420 |
++ else |
7421 |
++ return rq2; |
7422 |
++ } |
7423 |
++} |
7424 |
++ |
7425 |
++static struct bfq_queue * |
7426 |
++bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root, |
7427 |
++ sector_t sector, struct rb_node **ret_parent, |
7428 |
++ struct rb_node ***rb_link) |
7429 |
++{ |
7430 |
++ struct rb_node **p, *parent; |
7431 |
++ struct bfq_queue *bfqq = NULL; |
7432 |
++ |
7433 |
++ parent = NULL; |
7434 |
++ p = &root->rb_node; |
7435 |
++ while (*p) { |
7436 |
++ struct rb_node **n; |
7437 |
++ |
7438 |
++ parent = *p; |
7439 |
++ bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
7440 |
++ |
7441 |
++ /* |
7442 |
++ * Sort strictly based on sector. Smallest to the left, |
7443 |
++ * largest to the right. |
7444 |
++ */ |
7445 |
++ if (sector > blk_rq_pos(bfqq->next_rq)) |
7446 |
++ n = &(*p)->rb_right; |
7447 |
++ else if (sector < blk_rq_pos(bfqq->next_rq)) |
7448 |
++ n = &(*p)->rb_left; |
7449 |
++ else |
7450 |
++ break; |
7451 |
++ p = n; |
7452 |
++ bfqq = NULL; |
7453 |
++ } |
7454 |
++ |
7455 |
++ *ret_parent = parent; |
7456 |
++ if (rb_link) |
7457 |
++ *rb_link = p; |
7458 |
++ |
7459 |
++ bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d", |
7460 |
++ (long long unsigned)sector, |
7461 |
++ bfqq != NULL ? bfqq->pid : 0); |
7462 |
++ |
7463 |
++ return bfqq; |
7464 |
++} |
7465 |
++ |
7466 |
++static void bfq_rq_pos_tree_add(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
7467 |
++{ |
7468 |
++ struct rb_node **p, *parent; |
7469 |
++ struct bfq_queue *__bfqq; |
7470 |
++ |
7471 |
++ if (bfqq->pos_root != NULL) { |
7472 |
++ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
7473 |
++ bfqq->pos_root = NULL; |
7474 |
++ } |
7475 |
++ |
7476 |
++ if (bfq_class_idle(bfqq)) |
7477 |
++ return; |
7478 |
++ if (!bfqq->next_rq) |
7479 |
++ return; |
7480 |
++ |
7481 |
++ bfqq->pos_root = &bfqd->rq_pos_tree; |
7482 |
++ __bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root, |
7483 |
++ blk_rq_pos(bfqq->next_rq), &parent, &p); |
7484 |
++ if (__bfqq == NULL) { |
7485 |
++ rb_link_node(&bfqq->pos_node, parent, p); |
7486 |
++ rb_insert_color(&bfqq->pos_node, bfqq->pos_root); |
7487 |
++ } else |
7488 |
++ bfqq->pos_root = NULL; |
7489 |
++} |
7490 |
++ |
7491 |
++static struct request *bfq_find_next_rq(struct bfq_data *bfqd, |
7492 |
++ struct bfq_queue *bfqq, |
7493 |
++ struct request *last) |
7494 |
++{ |
7495 |
++ struct rb_node *rbnext = rb_next(&last->rb_node); |
7496 |
++ struct rb_node *rbprev = rb_prev(&last->rb_node); |
7497 |
++ struct request *next = NULL, *prev = NULL; |
7498 |
++ |
7499 |
++ BUG_ON(RB_EMPTY_NODE(&last->rb_node)); |
7500 |
++ |
7501 |
++ if (rbprev != NULL) |
7502 |
++ prev = rb_entry_rq(rbprev); |
7503 |
++ |
7504 |
++ if (rbnext != NULL) |
7505 |
++ next = rb_entry_rq(rbnext); |
7506 |
++ else { |
7507 |
++ rbnext = rb_first(&bfqq->sort_list); |
7508 |
++ if (rbnext && rbnext != &last->rb_node) |
7509 |
++ next = rb_entry_rq(rbnext); |
7510 |
++ } |
7511 |
++ |
7512 |
++ return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last)); |
7513 |
++} |
7514 |
++ |
7515 |
++static void bfq_del_rq_rb(struct request *rq) |
7516 |
++{ |
7517 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7518 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
7519 |
++ const int sync = rq_is_sync(rq); |
7520 |
++ |
7521 |
++ BUG_ON(bfqq->queued[sync] == 0); |
7522 |
++ bfqq->queued[sync]--; |
7523 |
++ bfqd->queued--; |
7524 |
++ |
7525 |
++ elv_rb_del(&bfqq->sort_list, rq); |
7526 |
++ |
7527 |
++ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
7528 |
++ if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) |
7529 |
++ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
7530 |
++ /* |
7531 |
++ * Remove queue from request-position tree as it is empty. |
7532 |
++ */ |
7533 |
++ if (bfqq->pos_root != NULL) { |
7534 |
++ rb_erase(&bfqq->pos_node, bfqq->pos_root); |
7535 |
++ bfqq->pos_root = NULL; |
7536 |
++ } |
7537 |
++ } |
7538 |
++} |
7539 |
++ |
7540 |
++/* see the definition of bfq_async_charge_factor for details */ |
7541 |
++static inline unsigned long bfq_serv_to_charge(struct request *rq, |
7542 |
++ struct bfq_queue *bfqq) |
7543 |
++{ |
7544 |
++ return blk_rq_sectors(rq) * |
7545 |
++ (1 + ((!bfq_bfqq_sync(bfqq)) * (bfqq->raising_coeff == 1) * |
7546 |
++ bfq_async_charge_factor)); |
7547 |
++} |
7548 |
++ |
7549 |
++/** |
7550 |
++ * bfq_updated_next_req - update the queue after a new next_rq selection. |
7551 |
++ * @bfqd: the device data the queue belongs to. |
7552 |
++ * @bfqq: the queue to update. |
7553 |
++ * |
7554 |
++ * If the first request of a queue changes we make sure that the queue |
7555 |
++ * has enough budget to serve at least its first request (if the |
7556 |
++ * request has grown). We do this because if the queue has not enough |
7557 |
++ * budget for its first request, it has to go through two dispatch |
7558 |
++ * rounds to actually get it dispatched. |
7559 |
++ */ |
7560 |
++static void bfq_updated_next_req(struct bfq_data *bfqd, |
7561 |
++ struct bfq_queue *bfqq) |
7562 |
++{ |
7563 |
++ struct bfq_entity *entity = &bfqq->entity; |
7564 |
++ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
7565 |
++ struct request *next_rq = bfqq->next_rq; |
7566 |
++ unsigned long new_budget; |
7567 |
++ |
7568 |
++ if (next_rq == NULL) |
7569 |
++ return; |
7570 |
++ |
7571 |
++ if (bfqq == bfqd->in_service_queue) |
7572 |
++ /* |
7573 |
++ * In order not to break guarantees, budgets cannot be |
7574 |
++ * changed after an entity has been selected. |
7575 |
++ */ |
7576 |
++ return; |
7577 |
++ |
7578 |
++ BUG_ON(entity->tree != &st->active); |
7579 |
++ BUG_ON(entity == entity->sched_data->in_service_entity); |
7580 |
++ |
7581 |
++ new_budget = max_t(unsigned long, bfqq->max_budget, |
7582 |
++ bfq_serv_to_charge(next_rq, bfqq)); |
7583 |
++ entity->budget = new_budget; |
7584 |
++ bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu", new_budget); |
7585 |
++ bfq_activate_bfqq(bfqd, bfqq); |
7586 |
++} |
7587 |
++ |
7588 |
++static inline unsigned int bfq_wrais_duration(struct bfq_data *bfqd) |
7589 |
++{ |
7590 |
++ u64 dur; |
7591 |
++ |
7592 |
++ if (bfqd->bfq_raising_max_time > 0) |
7593 |
++ return bfqd->bfq_raising_max_time; |
7594 |
++ |
7595 |
++ dur = bfqd->RT_prod; |
7596 |
++ do_div(dur, bfqd->peak_rate); |
7597 |
++ |
7598 |
++ return dur; |
7599 |
++} |
7600 |
++ |
7601 |
++static void bfq_add_rq_rb(struct request *rq) |
7602 |
++{ |
7603 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7604 |
++ struct bfq_entity *entity = &bfqq->entity; |
7605 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
7606 |
++ struct request *next_rq, *prev; |
7607 |
++ unsigned long old_raising_coeff = bfqq->raising_coeff; |
7608 |
++ int idle_for_long_time = 0; |
7609 |
++ |
7610 |
++ bfq_log_bfqq(bfqd, bfqq, "add_rq_rb %d", rq_is_sync(rq)); |
7611 |
++ bfqq->queued[rq_is_sync(rq)]++; |
7612 |
++ bfqd->queued++; |
7613 |
++ |
7614 |
++ elv_rb_add(&bfqq->sort_list, rq); |
7615 |
++ |
7616 |
++ /* |
7617 |
++ * Check if this request is a better next-serve candidate. |
7618 |
++ */ |
7619 |
++ prev = bfqq->next_rq; |
7620 |
++ next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position); |
7621 |
++ BUG_ON(next_rq == NULL); |
7622 |
++ bfqq->next_rq = next_rq; |
7623 |
++ |
7624 |
++ /* |
7625 |
++ * Adjust priority tree position, if next_rq changes. |
7626 |
++ */ |
7627 |
++ if (prev != bfqq->next_rq) |
7628 |
++ bfq_rq_pos_tree_add(bfqd, bfqq); |
7629 |
++ |
7630 |
++ if (!bfq_bfqq_busy(bfqq)) { |
7631 |
++ int soft_rt = bfqd->bfq_raising_max_softrt_rate > 0 && |
7632 |
++ time_is_before_jiffies(bfqq->soft_rt_next_start); |
7633 |
++ idle_for_long_time = time_is_before_jiffies( |
7634 |
++ bfqq->budget_timeout + |
7635 |
++ bfqd->bfq_raising_min_idle_time); |
7636 |
++ entity->budget = max_t(unsigned long, bfqq->max_budget, |
7637 |
++ bfq_serv_to_charge(next_rq, bfqq)); |
7638 |
++ |
7639 |
++ if (!bfqd->low_latency) |
7640 |
++ goto add_bfqq_busy; |
7641 |
++ |
7642 |
++ /* |
7643 |
++ * If the queue is not being boosted and has been idle |
7644 |
++ * for enough time, start a weight-raising period |
7645 |
++ */ |
7646 |
++ if (old_raising_coeff == 1 && |
7647 |
++ (idle_for_long_time || soft_rt)) { |
7648 |
++ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
7649 |
++ if (idle_for_long_time) |
7650 |
++ bfqq->raising_cur_max_time = |
7651 |
++ bfq_wrais_duration(bfqd); |
7652 |
++ else |
7653 |
++ bfqq->raising_cur_max_time = |
7654 |
++ bfqd->bfq_raising_rt_max_time; |
7655 |
++ bfq_log_bfqq(bfqd, bfqq, |
7656 |
++ "wrais starting at %lu, " |
7657 |
++ "rais_max_time %u", |
7658 |
++ jiffies, |
7659 |
++ jiffies_to_msecs(bfqq-> |
7660 |
++ raising_cur_max_time)); |
7661 |
++ } else if (old_raising_coeff > 1) { |
7662 |
++ if (idle_for_long_time) |
7663 |
++ bfqq->raising_cur_max_time = |
7664 |
++ bfq_wrais_duration(bfqd); |
7665 |
++ else if (bfqq->raising_cur_max_time == |
7666 |
++ bfqd->bfq_raising_rt_max_time && |
7667 |
++ !soft_rt) { |
7668 |
++ bfqq->raising_coeff = 1; |
7669 |
++ bfq_log_bfqq(bfqd, bfqq, |
7670 |
++ "wrais ending at %lu, " |
7671 |
++ "rais_max_time %u", |
7672 |
++ jiffies, |
7673 |
++ jiffies_to_msecs(bfqq-> |
7674 |
++ raising_cur_max_time)); |
7675 |
++ } else if (time_before( |
7676 |
++ bfqq->last_rais_start_finish + |
7677 |
++ bfqq->raising_cur_max_time, |
7678 |
++ jiffies + |
7679 |
++ bfqd->bfq_raising_rt_max_time) && |
7680 |
++ soft_rt) { |
7681 |
++ /* |
7682 |
++ * |
7683 |
++ * The remaining weight-raising time is lower |
7684 |
++ * than bfqd->bfq_raising_rt_max_time, which |
7685 |
++ * means that the application is enjoying |
7686 |
++ * weight raising either because deemed soft- |
7687 |
++ * rt in the near past, or because deemed |
7688 |
++ * interactive a long ago. In both cases, |
7689 |
++ * resetting now the current remaining weight- |
7690 |
++ * raising time for the application to the |
7691 |
++ * weight-raising duration for soft rt |
7692 |
++ * applications would not cause any latency |
7693 |
++ * increase for the application (as the new |
7694 |
++ * duration would be higher than the remaining |
7695 |
++ * time). |
7696 |
++ * |
7697 |
++ * In addition, the application is now meeting |
7698 |
++ * the requirements for being deemed soft rt. |
7699 |
++ * In the end we can correctly and safely |
7700 |
++ * (re)charge the weight-raising duration for |
7701 |
++ * the application with the weight-raising |
7702 |
++ * duration for soft rt applications. |
7703 |
++ * |
7704 |
++ * In particular, doing this recharge now, i.e., |
7705 |
++ * before the weight-raising period for the |
7706 |
++ * application finishes, reduces the probability |
7707 |
++ * of the following negative scenario: |
7708 |
++ * 1) the weight of a soft rt application is |
7709 |
++ * raised at startup (as for any newly |
7710 |
++ * created application), |
7711 |
++ * 2) since the application is not interactive, |
7712 |
++ * at a certain time weight-raising is |
7713 |
++ * stopped for the application, |
7714 |
++ * 3) at that time the application happens to |
7715 |
++ * still have pending requests, and hence |
7716 |
++ * is destined to not have a chance to be |
7717 |
++ * deemed soft rt before these requests are |
7718 |
++ * completed (see the comments to the |
7719 |
++ * function bfq_bfqq_softrt_next_start() |
7720 |
++ * for details on soft rt detection), |
7721 |
++ * 4) these pending requests experience a high |
7722 |
++ * latency because the application is not |
7723 |
++ * weight-raised while they are pending. |
7724 |
++ */ |
7725 |
++ bfqq->last_rais_start_finish = jiffies; |
7726 |
++ bfqq->raising_cur_max_time = |
7727 |
++ bfqd->bfq_raising_rt_max_time; |
7728 |
++ } |
7729 |
++ } |
7730 |
++ if (old_raising_coeff != bfqq->raising_coeff) |
7731 |
++ entity->ioprio_changed = 1; |
7732 |
++add_bfqq_busy: |
7733 |
++ bfqq->last_idle_bklogged = jiffies; |
7734 |
++ bfqq->service_from_backlogged = 0; |
7735 |
++ bfq_clear_bfqq_softrt_update(bfqq); |
7736 |
++ bfq_add_bfqq_busy(bfqd, bfqq); |
7737 |
++ } else { |
7738 |
++ if (bfqd->low_latency && old_raising_coeff == 1 && |
7739 |
++ !rq_is_sync(rq) && |
7740 |
++ time_is_before_jiffies( |
7741 |
++ bfqq->last_rais_start_finish + |
7742 |
++ bfqd->bfq_raising_min_inter_arr_async)) { |
7743 |
++ bfqq->raising_coeff = bfqd->bfq_raising_coeff; |
7744 |
++ bfqq->raising_cur_max_time = bfq_wrais_duration(bfqd); |
7745 |
++ |
7746 |
++ bfqd->raised_busy_queues++; |
7747 |
++ entity->ioprio_changed = 1; |
7748 |
++ bfq_log_bfqq(bfqd, bfqq, |
7749 |
++ "non-idle wrais starting at %lu, " |
7750 |
++ "rais_max_time %u", |
7751 |
++ jiffies, |
7752 |
++ jiffies_to_msecs(bfqq-> |
7753 |
++ raising_cur_max_time)); |
7754 |
++ } |
7755 |
++ bfq_updated_next_req(bfqd, bfqq); |
7756 |
++ } |
7757 |
++ |
7758 |
++ if (bfqd->low_latency && |
7759 |
++ (old_raising_coeff == 1 || bfqq->raising_coeff == 1 || |
7760 |
++ idle_for_long_time)) |
7761 |
++ bfqq->last_rais_start_finish = jiffies; |
7762 |
++} |
7763 |
++ |
7764 |
++static void bfq_reposition_rq_rb(struct bfq_queue *bfqq, struct request *rq) |
7765 |
++{ |
7766 |
++ elv_rb_del(&bfqq->sort_list, rq); |
7767 |
++ bfqq->queued[rq_is_sync(rq)]--; |
7768 |
++ bfqq->bfqd->queued--; |
7769 |
++ bfq_add_rq_rb(rq); |
7770 |
++} |
7771 |
++ |
7772 |
++static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd, |
7773 |
++ struct bio *bio) |
7774 |
++{ |
7775 |
++ struct task_struct *tsk = current; |
7776 |
++ struct bfq_io_cq *bic; |
7777 |
++ struct bfq_queue *bfqq; |
7778 |
++ |
7779 |
++ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
7780 |
++ if (bic == NULL) |
7781 |
++ return NULL; |
7782 |
++ |
7783 |
++ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
7784 |
++ if (bfqq != NULL) |
7785 |
++ return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio)); |
7786 |
++ |
7787 |
++ return NULL; |
7788 |
++} |
7789 |
++ |
7790 |
++static void bfq_activate_request(struct request_queue *q, struct request *rq) |
7791 |
++{ |
7792 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
7793 |
++ |
7794 |
++ bfqd->rq_in_driver++; |
7795 |
++ bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq); |
7796 |
++ bfq_log(bfqd, "activate_request: new bfqd->last_position %llu", |
7797 |
++ (long long unsigned)bfqd->last_position); |
7798 |
++} |
7799 |
++ |
7800 |
++static void bfq_deactivate_request(struct request_queue *q, struct request *rq) |
7801 |
++{ |
7802 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
7803 |
++ |
7804 |
++ WARN_ON(bfqd->rq_in_driver == 0); |
7805 |
++ bfqd->rq_in_driver--; |
7806 |
++} |
7807 |
++ |
7808 |
++static void bfq_remove_request(struct request *rq) |
7809 |
++{ |
7810 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7811 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
7812 |
++ |
7813 |
++ if (bfqq->next_rq == rq) { |
7814 |
++ bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq); |
7815 |
++ bfq_updated_next_req(bfqd, bfqq); |
7816 |
++ } |
7817 |
++ |
7818 |
++ list_del_init(&rq->queuelist); |
7819 |
++ bfq_del_rq_rb(rq); |
7820 |
++ |
7821 |
++ if (rq->cmd_flags & REQ_META) { |
7822 |
++ WARN_ON(bfqq->meta_pending == 0); |
7823 |
++ bfqq->meta_pending--; |
7824 |
++ } |
7825 |
++} |
7826 |
++ |
7827 |
++static int bfq_merge(struct request_queue *q, struct request **req, |
7828 |
++ struct bio *bio) |
7829 |
++{ |
7830 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
7831 |
++ struct request *__rq; |
7832 |
++ |
7833 |
++ __rq = bfq_find_rq_fmerge(bfqd, bio); |
7834 |
++ if (__rq != NULL && elv_rq_merge_ok(__rq, bio)) { |
7835 |
++ *req = __rq; |
7836 |
++ return ELEVATOR_FRONT_MERGE; |
7837 |
++ } |
7838 |
++ |
7839 |
++ return ELEVATOR_NO_MERGE; |
7840 |
++} |
7841 |
++ |
7842 |
++static void bfq_merged_request(struct request_queue *q, struct request *req, |
7843 |
++ int type) |
7844 |
++{ |
7845 |
++ if (type == ELEVATOR_FRONT_MERGE) { |
7846 |
++ struct bfq_queue *bfqq = RQ_BFQQ(req); |
7847 |
++ |
7848 |
++ bfq_reposition_rq_rb(bfqq, req); |
7849 |
++ } |
7850 |
++} |
7851 |
++ |
7852 |
++static void bfq_merged_requests(struct request_queue *q, struct request *rq, |
7853 |
++ struct request *next) |
7854 |
++{ |
7855 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
7856 |
++ |
7857 |
++ /* |
7858 |
++ * Reposition in fifo if next is older than rq. |
7859 |
++ */ |
7860 |
++ if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) && |
7861 |
++ time_before(rq_fifo_time(next), rq_fifo_time(rq))) { |
7862 |
++ list_move(&rq->queuelist, &next->queuelist); |
7863 |
++ rq_set_fifo_time(rq, rq_fifo_time(next)); |
7864 |
++ } |
7865 |
++ |
7866 |
++ if (bfqq->next_rq == next) |
7867 |
++ bfqq->next_rq = rq; |
7868 |
++ |
7869 |
++ bfq_remove_request(next); |
7870 |
++} |
7871 |
++ |
7872 |
++/* Must be called with bfqq != NULL */ |
7873 |
++static inline void bfq_bfqq_end_raising(struct bfq_queue *bfqq) |
7874 |
++{ |
7875 |
++ BUG_ON(bfqq == NULL); |
7876 |
++ if (bfq_bfqq_busy(bfqq)) |
7877 |
++ bfqq->bfqd->raised_busy_queues--; |
7878 |
++ bfqq->raising_coeff = 1; |
7879 |
++ bfqq->raising_cur_max_time = 0; |
7880 |
++ /* Trigger a weight change on the next activation of the queue */ |
7881 |
++ bfqq->entity.ioprio_changed = 1; |
7882 |
++} |
7883 |
++ |
7884 |
++static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
7885 |
++ struct bfq_group *bfqg) |
7886 |
++{ |
7887 |
++ int i, j; |
7888 |
++ |
7889 |
++ for (i = 0; i < 2; i++) |
7890 |
++ for (j = 0; j < IOPRIO_BE_NR; j++) |
7891 |
++ if (bfqg->async_bfqq[i][j] != NULL) |
7892 |
++ bfq_bfqq_end_raising(bfqg->async_bfqq[i][j]); |
7893 |
++ if (bfqg->async_idle_bfqq != NULL) |
7894 |
++ bfq_bfqq_end_raising(bfqg->async_idle_bfqq); |
7895 |
++} |
7896 |
++ |
7897 |
++static void bfq_end_raising(struct bfq_data *bfqd) |
7898 |
++{ |
7899 |
++ struct bfq_queue *bfqq; |
7900 |
++ |
7901 |
++ spin_lock_irq(bfqd->queue->queue_lock); |
7902 |
++ |
7903 |
++ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) |
7904 |
++ bfq_bfqq_end_raising(bfqq); |
7905 |
++ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) |
7906 |
++ bfq_bfqq_end_raising(bfqq); |
7907 |
++ bfq_end_raising_async(bfqd); |
7908 |
++ |
7909 |
++ spin_unlock_irq(bfqd->queue->queue_lock); |
7910 |
++} |
7911 |
++ |
7912 |
++static int bfq_allow_merge(struct request_queue *q, struct request *rq, |
7913 |
++ struct bio *bio) |
7914 |
++{ |
7915 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
7916 |
++ struct bfq_io_cq *bic; |
7917 |
++ struct bfq_queue *bfqq; |
7918 |
++ |
7919 |
++ /* |
7920 |
++ * Disallow merge of a sync bio into an async request. |
7921 |
++ */ |
7922 |
++ if (bfq_bio_sync(bio) && !rq_is_sync(rq)) |
7923 |
++ return 0; |
7924 |
++ |
7925 |
++ /* |
7926 |
++ * Lookup the bfqq that this bio will be queued with. Allow |
7927 |
++ * merge only if rq is queued there. |
7928 |
++ * Queue lock is held here. |
7929 |
++ */ |
7930 |
++ bic = bfq_bic_lookup(bfqd, current->io_context); |
7931 |
++ if (bic == NULL) |
7932 |
++ return 0; |
7933 |
++ |
7934 |
++ bfqq = bic_to_bfqq(bic, bfq_bio_sync(bio)); |
7935 |
++ return bfqq == RQ_BFQQ(rq); |
7936 |
++} |
7937 |
++ |
7938 |
++static void __bfq_set_in_service_queue(struct bfq_data *bfqd, |
7939 |
++ struct bfq_queue *bfqq) |
7940 |
++{ |
7941 |
++ if (bfqq != NULL) { |
7942 |
++ bfq_mark_bfqq_must_alloc(bfqq); |
7943 |
++ bfq_mark_bfqq_budget_new(bfqq); |
7944 |
++ bfq_clear_bfqq_fifo_expire(bfqq); |
7945 |
++ |
7946 |
++ bfqd->budgets_assigned = (bfqd->budgets_assigned*7 + 256) / 8; |
7947 |
++ |
7948 |
++ bfq_log_bfqq(bfqd, bfqq, |
7949 |
++ "set_in_service_queue, cur-budget = %lu", |
7950 |
++ bfqq->entity.budget); |
7951 |
++ } |
7952 |
++ |
7953 |
++ bfqd->in_service_queue = bfqq; |
7954 |
++} |
7955 |
++ |
7956 |
++/* |
7957 |
++ * Get and set a new queue for service. |
7958 |
++ */ |
7959 |
++static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd, |
7960 |
++ struct bfq_queue *bfqq) |
7961 |
++{ |
7962 |
++ if (!bfqq) |
7963 |
++ bfqq = bfq_get_next_queue(bfqd); |
7964 |
++ else |
7965 |
++ bfq_get_next_queue_forced(bfqd, bfqq); |
7966 |
++ |
7967 |
++ __bfq_set_in_service_queue(bfqd, bfqq); |
7968 |
++ return bfqq; |
7969 |
++} |
7970 |
++ |
7971 |
++static inline sector_t bfq_dist_from_last(struct bfq_data *bfqd, |
7972 |
++ struct request *rq) |
7973 |
++{ |
7974 |
++ if (blk_rq_pos(rq) >= bfqd->last_position) |
7975 |
++ return blk_rq_pos(rq) - bfqd->last_position; |
7976 |
++ else |
7977 |
++ return bfqd->last_position - blk_rq_pos(rq); |
7978 |
++} |
7979 |
++ |
7980 |
++/* |
7981 |
++ * Return true if bfqq has no request pending and rq is close enough to |
7982 |
++ * bfqd->last_position, or if rq is closer to bfqd->last_position than |
7983 |
++ * bfqq->next_rq |
7984 |
++ */ |
7985 |
++static inline int bfq_rq_close(struct bfq_data *bfqd, struct request *rq) |
7986 |
++{ |
7987 |
++ return bfq_dist_from_last(bfqd, rq) <= BFQQ_SEEK_THR; |
7988 |
++} |
7989 |
++ |
7990 |
++static struct bfq_queue *bfqq_close(struct bfq_data *bfqd) |
7991 |
++{ |
7992 |
++ struct rb_root *root = &bfqd->rq_pos_tree; |
7993 |
++ struct rb_node *parent, *node; |
7994 |
++ struct bfq_queue *__bfqq; |
7995 |
++ sector_t sector = bfqd->last_position; |
7996 |
++ |
7997 |
++ if (RB_EMPTY_ROOT(root)) |
7998 |
++ return NULL; |
7999 |
++ |
8000 |
++ /* |
8001 |
++ * First, if we find a request starting at the end of the last |
8002 |
++ * request, choose it. |
8003 |
++ */ |
8004 |
++ __bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL); |
8005 |
++ if (__bfqq != NULL) |
8006 |
++ return __bfqq; |
8007 |
++ |
8008 |
++ /* |
8009 |
++ * If the exact sector wasn't found, the parent of the NULL leaf |
8010 |
++ * will contain the closest sector (rq_pos_tree sorted by next_request |
8011 |
++ * position). |
8012 |
++ */ |
8013 |
++ __bfqq = rb_entry(parent, struct bfq_queue, pos_node); |
8014 |
++ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
8015 |
++ return __bfqq; |
8016 |
++ |
8017 |
++ if (blk_rq_pos(__bfqq->next_rq) < sector) |
8018 |
++ node = rb_next(&__bfqq->pos_node); |
8019 |
++ else |
8020 |
++ node = rb_prev(&__bfqq->pos_node); |
8021 |
++ if (node == NULL) |
8022 |
++ return NULL; |
8023 |
++ |
8024 |
++ __bfqq = rb_entry(node, struct bfq_queue, pos_node); |
8025 |
++ if (bfq_rq_close(bfqd, __bfqq->next_rq)) |
8026 |
++ return __bfqq; |
8027 |
++ |
8028 |
++ return NULL; |
8029 |
++} |
8030 |
++ |
8031 |
++/* |
8032 |
++ * bfqd - obvious |
8033 |
++ * cur_bfqq - passed in so that we don't decide that the current queue |
8034 |
++ * is closely cooperating with itself. |
8035 |
++ * |
8036 |
++ * We are assuming that cur_bfqq has dispatched at least one request, |
8037 |
++ * and that bfqd->last_position reflects a position on the disk associated |
8038 |
++ * with the I/O issued by cur_bfqq. |
8039 |
++ */ |
8040 |
++static struct bfq_queue *bfq_close_cooperator(struct bfq_data *bfqd, |
8041 |
++ struct bfq_queue *cur_bfqq) |
8042 |
++{ |
8043 |
++ struct bfq_queue *bfqq; |
8044 |
++ |
8045 |
++ if (bfq_class_idle(cur_bfqq)) |
8046 |
++ return NULL; |
8047 |
++ if (!bfq_bfqq_sync(cur_bfqq)) |
8048 |
++ return NULL; |
8049 |
++ if (BFQQ_SEEKY(cur_bfqq)) |
8050 |
++ return NULL; |
8051 |
++ |
8052 |
++ /* If device has only one backlogged bfq_queue, don't search. */ |
8053 |
++ if (bfqd->busy_queues == 1) |
8054 |
++ return NULL; |
8055 |
++ |
8056 |
++ /* |
8057 |
++ * We should notice if some of the queues are cooperating, e.g. |
8058 |
++ * working closely on the same area of the disk. In that case, |
8059 |
++ * we can group them together and don't waste time idling. |
8060 |
++ */ |
8061 |
++ bfqq = bfqq_close(bfqd); |
8062 |
++ if (bfqq == NULL || bfqq == cur_bfqq) |
8063 |
++ return NULL; |
8064 |
++ |
8065 |
++ /* |
8066 |
++ * Do not merge queues from different bfq_groups. |
8067 |
++ */ |
8068 |
++ if (bfqq->entity.parent != cur_bfqq->entity.parent) |
8069 |
++ return NULL; |
8070 |
++ |
8071 |
++ /* |
8072 |
++ * It only makes sense to merge sync queues. |
8073 |
++ */ |
8074 |
++ if (!bfq_bfqq_sync(bfqq)) |
8075 |
++ return NULL; |
8076 |
++ if (BFQQ_SEEKY(bfqq)) |
8077 |
++ return NULL; |
8078 |
++ |
8079 |
++ /* |
8080 |
++ * Do not merge queues of different priority classes. |
8081 |
++ */ |
8082 |
++ if (bfq_class_rt(bfqq) != bfq_class_rt(cur_bfqq)) |
8083 |
++ return NULL; |
8084 |
++ |
8085 |
++ return bfqq; |
8086 |
++} |
8087 |
++ |
8088 |
++/* |
8089 |
++ * If enough samples have been computed, return the current max budget |
8090 |
++ * stored in bfqd, which is dynamically updated according to the |
8091 |
++ * estimated disk peak rate; otherwise return the default max budget |
8092 |
++ */ |
8093 |
++static inline unsigned long bfq_max_budget(struct bfq_data *bfqd) |
8094 |
++{ |
8095 |
++ if (bfqd->budgets_assigned < 194) |
8096 |
++ return bfq_default_max_budget; |
8097 |
++ else |
8098 |
++ return bfqd->bfq_max_budget; |
8099 |
++} |
8100 |
++ |
8101 |
++/* |
8102 |
++ * Return min budget, which is a fraction of the current or default |
8103 |
++ * max budget (trying with 1/32) |
8104 |
++ */ |
8105 |
++static inline unsigned long bfq_min_budget(struct bfq_data *bfqd) |
8106 |
++{ |
8107 |
++ if (bfqd->budgets_assigned < 194) |
8108 |
++ return bfq_default_max_budget / 32; |
8109 |
++ else |
8110 |
++ return bfqd->bfq_max_budget / 32; |
8111 |
++} |
8112 |
++ |
8113 |
++/* |
8114 |
++ * Decides whether idling should be done for given device and |
8115 |
++ * given in-service queue. |
8116 |
++ */ |
8117 |
++static inline bool bfq_queue_nonrot_noidle(struct bfq_data *bfqd, |
8118 |
++ struct bfq_queue *in_service_bfqq) |
8119 |
++{ |
8120 |
++ if (in_service_bfqq == NULL) |
8121 |
++ return false; |
8122 |
++ /* |
8123 |
++ * If the device is non-rotational, and hence has no seek penalty, |
8124 |
++ * disable idling; but do so only if: |
8125 |
++ * - device does not support queuing, otherwise we still have |
8126 |
++ * a problem with sync vs async workloads; |
8127 |
++ * - the queue is not weight-raised, to preserve guarantees. |
8128 |
++ */ |
8129 |
++ return blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag && |
8130 |
++ (in_service_bfqq->raising_coeff == 1); |
8131 |
++} |
8132 |
++ |
8133 |
++static void bfq_arm_slice_timer(struct bfq_data *bfqd) |
8134 |
++{ |
8135 |
++ struct bfq_queue *bfqq = bfqd->in_service_queue; |
8136 |
++ struct bfq_io_cq *bic; |
8137 |
++ unsigned long sl; |
8138 |
++ |
8139 |
++ WARN_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
8140 |
++ |
8141 |
++ /* Tasks have exited, don't wait. */ |
8142 |
++ bic = bfqd->in_service_bic; |
8143 |
++ if (bic == NULL || atomic_read(&bic->icq.ioc->active_ref) == 0) |
8144 |
++ return; |
8145 |
++ |
8146 |
++ bfq_mark_bfqq_wait_request(bfqq); |
8147 |
++ |
8148 |
++ /* |
8149 |
++ * We don't want to idle for seeks, but we do want to allow |
8150 |
++ * fair distribution of slice time for a process doing back-to-back |
8151 |
++ * seeks. So allow a little bit of time for him to submit a new rq. |
8152 |
++ * |
8153 |
++ * To prevent processes with (partly) seeky workloads from |
8154 |
++ * being too ill-treated, grant them a small fraction of the |
8155 |
++ * assigned budget before reducing the waiting time to |
8156 |
++ * BFQ_MIN_TT. This happened to help reduce latency. |
8157 |
++ */ |
8158 |
++ sl = bfqd->bfq_slice_idle; |
8159 |
++ if (bfq_sample_valid(bfqq->seek_samples) && BFQQ_SEEKY(bfqq) && |
8160 |
++ bfqq->entity.service > bfq_max_budget(bfqd) / 8 && |
8161 |
++ bfqq->raising_coeff == 1) |
8162 |
++ sl = min(sl, msecs_to_jiffies(BFQ_MIN_TT)); |
8163 |
++ else if (bfqq->raising_coeff > 1) |
8164 |
++ sl = sl * 3; |
8165 |
++ bfqd->last_idling_start = ktime_get(); |
8166 |
++ mod_timer(&bfqd->idle_slice_timer, jiffies + sl); |
8167 |
++ bfq_log(bfqd, "arm idle: %u/%u ms", |
8168 |
++ jiffies_to_msecs(sl), jiffies_to_msecs(bfqd->bfq_slice_idle)); |
8169 |
++} |
8170 |
++ |
8171 |
++/* |
8172 |
++ * Set the maximum time for the in-service queue to consume its |
8173 |
++ * budget. This prevents seeky processes from lowering the disk |
8174 |
++ * throughput (always guaranteed with a time slice scheme as in CFQ). |
8175 |
++ */ |
8176 |
++static void bfq_set_budget_timeout(struct bfq_data *bfqd) |
8177 |
++{ |
8178 |
++ struct bfq_queue *bfqq = bfqd->in_service_queue; |
8179 |
++ unsigned int timeout_coeff; |
8180 |
++ if (bfqq->raising_cur_max_time == bfqd->bfq_raising_rt_max_time) |
8181 |
++ timeout_coeff = 1; |
8182 |
++ else |
8183 |
++ timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight; |
8184 |
++ |
8185 |
++ bfqd->last_budget_start = ktime_get(); |
8186 |
++ |
8187 |
++ bfq_clear_bfqq_budget_new(bfqq); |
8188 |
++ bfqq->budget_timeout = jiffies + |
8189 |
++ bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * timeout_coeff; |
8190 |
++ |
8191 |
++ bfq_log_bfqq(bfqd, bfqq, "set budget_timeout %u", |
8192 |
++ jiffies_to_msecs(bfqd->bfq_timeout[bfq_bfqq_sync(bfqq)] * |
8193 |
++ timeout_coeff)); |
8194 |
++} |
8195 |
++ |
8196 |
++/* |
8197 |
++ * Move request from internal lists to the request queue dispatch list. |
8198 |
++ */ |
8199 |
++static void bfq_dispatch_insert(struct request_queue *q, struct request *rq) |
8200 |
++{ |
8201 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
8202 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
8203 |
++ |
8204 |
++ bfq_remove_request(rq); |
8205 |
++ bfqq->dispatched++; |
8206 |
++ elv_dispatch_sort(q, rq); |
8207 |
++ |
8208 |
++ if (bfq_bfqq_sync(bfqq)) |
8209 |
++ bfqd->sync_flight++; |
8210 |
++} |
8211 |
++ |
8212 |
++/* |
8213 |
++ * Return expired entry, or NULL to just start from scratch in rbtree. |
8214 |
++ */ |
8215 |
++static struct request *bfq_check_fifo(struct bfq_queue *bfqq) |
8216 |
++{ |
8217 |
++ struct request *rq = NULL; |
8218 |
++ |
8219 |
++ if (bfq_bfqq_fifo_expire(bfqq)) |
8220 |
++ return NULL; |
8221 |
++ |
8222 |
++ bfq_mark_bfqq_fifo_expire(bfqq); |
8223 |
++ |
8224 |
++ if (list_empty(&bfqq->fifo)) |
8225 |
++ return NULL; |
8226 |
++ |
8227 |
++ rq = rq_entry_fifo(bfqq->fifo.next); |
8228 |
++ |
8229 |
++ if (time_before(jiffies, rq_fifo_time(rq))) |
8230 |
++ return NULL; |
8231 |
++ |
8232 |
++ return rq; |
8233 |
++} |
8234 |
++ |
8235 |
++/* |
8236 |
++ * Must be called with the queue_lock held. |
8237 |
++ */ |
8238 |
++static int bfqq_process_refs(struct bfq_queue *bfqq) |
8239 |
++{ |
8240 |
++ int process_refs, io_refs; |
8241 |
++ |
8242 |
++ io_refs = bfqq->allocated[READ] + bfqq->allocated[WRITE]; |
8243 |
++ process_refs = atomic_read(&bfqq->ref) - io_refs - bfqq->entity.on_st; |
8244 |
++ BUG_ON(process_refs < 0); |
8245 |
++ return process_refs; |
8246 |
++} |
8247 |
++ |
8248 |
++static void bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq) |
8249 |
++{ |
8250 |
++ int process_refs, new_process_refs; |
8251 |
++ struct bfq_queue *__bfqq; |
8252 |
++ |
8253 |
++ /* |
8254 |
++ * If there are no process references on the new_bfqq, then it is |
8255 |
++ * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain |
8256 |
++ * may have dropped their last reference (not just their last process |
8257 |
++ * reference). |
8258 |
++ */ |
8259 |
++ if (!bfqq_process_refs(new_bfqq)) |
8260 |
++ return; |
8261 |
++ |
8262 |
++ /* Avoid a circular list and skip interim queue merges. */ |
8263 |
++ while ((__bfqq = new_bfqq->new_bfqq)) { |
8264 |
++ if (__bfqq == bfqq) |
8265 |
++ return; |
8266 |
++ new_bfqq = __bfqq; |
8267 |
++ } |
8268 |
++ |
8269 |
++ process_refs = bfqq_process_refs(bfqq); |
8270 |
++ new_process_refs = bfqq_process_refs(new_bfqq); |
8271 |
++ /* |
8272 |
++ * If the process for the bfqq has gone away, there is no |
8273 |
++ * sense in merging the queues. |
8274 |
++ */ |
8275 |
++ if (process_refs == 0 || new_process_refs == 0) |
8276 |
++ return; |
8277 |
++ |
8278 |
++ /* |
8279 |
++ * Merge in the direction of the lesser amount of work. |
8280 |
++ */ |
8281 |
++ if (new_process_refs >= process_refs) { |
8282 |
++ bfqq->new_bfqq = new_bfqq; |
8283 |
++ atomic_add(process_refs, &new_bfqq->ref); |
8284 |
++ } else { |
8285 |
++ new_bfqq->new_bfqq = bfqq; |
8286 |
++ atomic_add(new_process_refs, &bfqq->ref); |
8287 |
++ } |
8288 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d", |
8289 |
++ new_bfqq->pid); |
8290 |
++} |
8291 |
++ |
8292 |
++static inline unsigned long bfq_bfqq_budget_left(struct bfq_queue *bfqq) |
8293 |
++{ |
8294 |
++ struct bfq_entity *entity = &bfqq->entity; |
8295 |
++ return entity->budget - entity->service; |
8296 |
++} |
8297 |
++ |
8298 |
++static void __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
8299 |
++{ |
8300 |
++ BUG_ON(bfqq != bfqd->in_service_queue); |
8301 |
++ |
8302 |
++ __bfq_bfqd_reset_in_service(bfqd); |
8303 |
++ |
8304 |
++ /* |
8305 |
++ * If this bfqq is shared between multiple processes, check |
8306 |
++ * to make sure that those processes are still issuing I/Os |
8307 |
++ * within the mean seek distance. If not, it may be time to |
8308 |
++ * break the queues apart again. |
8309 |
++ */ |
8310 |
++ if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq)) |
8311 |
++ bfq_mark_bfqq_split_coop(bfqq); |
8312 |
++ |
8313 |
++ if (RB_EMPTY_ROOT(&bfqq->sort_list)) { |
8314 |
++ /* |
8315 |
++ * overloading budget_timeout field to store when |
8316 |
++ * the queue remains with no backlog, used by |
8317 |
++ * the weight-raising mechanism |
8318 |
++ */ |
8319 |
++ bfqq->budget_timeout = jiffies; |
8320 |
++ bfq_del_bfqq_busy(bfqd, bfqq, 1); |
8321 |
++ } else { |
8322 |
++ bfq_activate_bfqq(bfqd, bfqq); |
8323 |
++ /* |
8324 |
++ * Resort priority tree of potential close cooperators. |
8325 |
++ */ |
8326 |
++ bfq_rq_pos_tree_add(bfqd, bfqq); |
8327 |
++ } |
8328 |
++} |
8329 |
++ |
8330 |
++/** |
8331 |
++ * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior. |
8332 |
++ * @bfqd: device data. |
8333 |
++ * @bfqq: queue to update. |
8334 |
++ * @reason: reason for expiration. |
8335 |
++ * |
8336 |
++ * Handle the feedback on @bfqq budget. See the body for detailed |
8337 |
++ * comments. |
8338 |
++ */ |
8339 |
++static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd, |
8340 |
++ struct bfq_queue *bfqq, |
8341 |
++ enum bfqq_expiration reason) |
8342 |
++{ |
8343 |
++ struct request *next_rq; |
8344 |
++ unsigned long budget, min_budget; |
8345 |
++ |
8346 |
++ budget = bfqq->max_budget; |
8347 |
++ min_budget = bfq_min_budget(bfqd); |
8348 |
++ |
8349 |
++ BUG_ON(bfqq != bfqd->in_service_queue); |
8350 |
++ |
8351 |
++ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %lu, budg left %lu", |
8352 |
++ bfqq->entity.budget, bfq_bfqq_budget_left(bfqq)); |
8353 |
++ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %lu, min budg %lu", |
8354 |
++ budget, bfq_min_budget(bfqd)); |
8355 |
++ bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d", |
8356 |
++ bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue)); |
8357 |
++ |
8358 |
++ if (bfq_bfqq_sync(bfqq)) { |
8359 |
++ switch (reason) { |
8360 |
++ /* |
8361 |
++ * Caveat: in all the following cases we trade latency |
8362 |
++ * for throughput. |
8363 |
++ */ |
8364 |
++ case BFQ_BFQQ_TOO_IDLE: |
8365 |
++ /* |
8366 |
++ * This is the only case where we may reduce |
8367 |
++ * the budget: if there is no request of the |
8368 |
++ * process still waiting for completion, then |
8369 |
++ * we assume (tentatively) that the timer has |
8370 |
++ * expired because the batch of requests of |
8371 |
++ * the process could have been served with a |
8372 |
++ * smaller budget. Hence, betting that |
8373 |
++ * process will behave in the same way when it |
8374 |
++ * becomes backlogged again, we reduce its |
8375 |
++ * next budget. As long as we guess right, |
8376 |
++ * this budget cut reduces the latency |
8377 |
++ * experienced by the process. |
8378 |
++ * |
8379 |
++ * However, if there are still outstanding |
8380 |
++ * requests, then the process may have not yet |
8381 |
++ * issued its next request just because it is |
8382 |
++ * still waiting for the completion of some of |
8383 |
++ * the still outstanding ones. So in this |
8384 |
++ * subcase we do not reduce its budget, on the |
8385 |
++ * contrary we increase it to possibly boost |
8386 |
++ * the throughput, as discussed in the |
8387 |
++ * comments to the BUDGET_TIMEOUT case. |
8388 |
++ */ |
8389 |
++ if (bfqq->dispatched > 0) /* still outstanding reqs */ |
8390 |
++ budget = min(budget * 2, bfqd->bfq_max_budget); |
8391 |
++ else { |
8392 |
++ if (budget > 5 * min_budget) |
8393 |
++ budget -= 4 * min_budget; |
8394 |
++ else |
8395 |
++ budget = min_budget; |
8396 |
++ } |
8397 |
++ break; |
8398 |
++ case BFQ_BFQQ_BUDGET_TIMEOUT: |
8399 |
++ /* |
8400 |
++ * We double the budget here because: 1) it |
8401 |
++ * gives the chance to boost the throughput if |
8402 |
++ * this is not a seeky process (which may have |
8403 |
++ * bumped into this timeout because of, e.g., |
8404 |
++ * ZBR), 2) together with charge_full_budget |
8405 |
++ * it helps give seeky processes higher |
8406 |
++ * timestamps, and hence be served less |
8407 |
++ * frequently. |
8408 |
++ */ |
8409 |
++ budget = min(budget * 2, bfqd->bfq_max_budget); |
8410 |
++ break; |
8411 |
++ case BFQ_BFQQ_BUDGET_EXHAUSTED: |
8412 |
++ /* |
8413 |
++ * The process still has backlog, and did not |
8414 |
++ * let either the budget timeout or the disk |
8415 |
++ * idling timeout expire. Hence it is not |
8416 |
++ * seeky, has a short thinktime and may be |
8417 |
++ * happy with a higher budget too. So |
8418 |
++ * definitely increase the budget of this good |
8419 |
++ * candidate to boost the disk throughput. |
8420 |
++ */ |
8421 |
++ budget = min(budget * 4, bfqd->bfq_max_budget); |
8422 |
++ break; |
8423 |
++ case BFQ_BFQQ_NO_MORE_REQUESTS: |
8424 |
++ /* |
8425 |
++ * Leave the budget unchanged. |
8426 |
++ */ |
8427 |
++ default: |
8428 |
++ return; |
8429 |
++ } |
8430 |
++ } else /* async queue */ |
8431 |
++ /* async queues get always the maximum possible budget |
8432 |
++ * (their ability to dispatch is limited by |
8433 |
++ * @bfqd->bfq_max_budget_async_rq). |
8434 |
++ */ |
8435 |
++ budget = bfqd->bfq_max_budget; |
8436 |
++ |
8437 |
++ bfqq->max_budget = budget; |
8438 |
++ |
8439 |
++ if (bfqd->budgets_assigned >= 194 && bfqd->bfq_user_max_budget == 0 && |
8440 |
++ bfqq->max_budget > bfqd->bfq_max_budget) |
8441 |
++ bfqq->max_budget = bfqd->bfq_max_budget; |
8442 |
++ |
8443 |
++ /* |
8444 |
++ * Make sure that we have enough budget for the next request. |
8445 |
++ * Since the finish time of the bfqq must be kept in sync with |
8446 |
++ * the budget, be sure to call __bfq_bfqq_expire() after the |
8447 |
++ * update. |
8448 |
++ */ |
8449 |
++ next_rq = bfqq->next_rq; |
8450 |
++ if (next_rq != NULL) |
8451 |
++ bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget, |
8452 |
++ bfq_serv_to_charge(next_rq, bfqq)); |
8453 |
++ else |
8454 |
++ bfqq->entity.budget = bfqq->max_budget; |
8455 |
++ |
8456 |
++ bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %lu", |
8457 |
++ next_rq != NULL ? blk_rq_sectors(next_rq) : 0, |
8458 |
++ bfqq->entity.budget); |
8459 |
++} |
8460 |
++ |
8461 |
++static unsigned long bfq_calc_max_budget(u64 peak_rate, u64 timeout) |
8462 |
++{ |
8463 |
++ unsigned long max_budget; |
8464 |
++ |
8465 |
++ /* |
8466 |
++ * The max_budget calculated when autotuning is equal to the |
8467 |
++ * amount of sectors transfered in timeout_sync at the |
8468 |
++ * estimated peak rate. |
8469 |
++ */ |
8470 |
++ max_budget = (unsigned long)(peak_rate * 1000 * |
8471 |
++ timeout >> BFQ_RATE_SHIFT); |
8472 |
++ |
8473 |
++ return max_budget; |
8474 |
++} |
8475 |
++ |
8476 |
++/* |
8477 |
++ * In addition to updating the peak rate, checks whether the process |
8478 |
++ * is "slow", and returns 1 if so. This slow flag is used, in addition |
8479 |
++ * to the budget timeout, to reduce the amount of service provided to |
8480 |
++ * seeky processes, and hence reduce their chances to lower the |
8481 |
++ * throughput. See the code for more details. |
8482 |
++ */ |
8483 |
++static int bfq_update_peak_rate(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
8484 |
++ int compensate, enum bfqq_expiration reason) |
8485 |
++{ |
8486 |
++ u64 bw, usecs, expected, timeout; |
8487 |
++ ktime_t delta; |
8488 |
++ int update = 0; |
8489 |
++ |
8490 |
++ if (!bfq_bfqq_sync(bfqq) || bfq_bfqq_budget_new(bfqq)) |
8491 |
++ return 0; |
8492 |
++ |
8493 |
++ if (compensate) |
8494 |
++ delta = bfqd->last_idling_start; |
8495 |
++ else |
8496 |
++ delta = ktime_get(); |
8497 |
++ delta = ktime_sub(delta, bfqd->last_budget_start); |
8498 |
++ usecs = ktime_to_us(delta); |
8499 |
++ |
8500 |
++ /* Don't trust short/unrealistic values. */ |
8501 |
++ if (usecs < 100 || usecs >= LONG_MAX) |
8502 |
++ return 0; |
8503 |
++ |
8504 |
++ /* |
8505 |
++ * Calculate the bandwidth for the last slice. We use a 64 bit |
8506 |
++ * value to store the peak rate, in sectors per usec in fixed |
8507 |
++ * point math. We do so to have enough precision in the estimate |
8508 |
++ * and to avoid overflows. |
8509 |
++ */ |
8510 |
++ bw = (u64)bfqq->entity.service << BFQ_RATE_SHIFT; |
8511 |
++ do_div(bw, (unsigned long)usecs); |
8512 |
++ |
8513 |
++ timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
8514 |
++ |
8515 |
++ /* |
8516 |
++ * Use only long (> 20ms) intervals to filter out spikes for |
8517 |
++ * the peak rate estimation. |
8518 |
++ */ |
8519 |
++ if (usecs > 20000) { |
8520 |
++ if (bw > bfqd->peak_rate || |
8521 |
++ (!BFQQ_SEEKY(bfqq) && |
8522 |
++ reason == BFQ_BFQQ_BUDGET_TIMEOUT)) { |
8523 |
++ bfq_log(bfqd, "measured bw =%llu", bw); |
8524 |
++ /* |
8525 |
++ * To smooth oscillations use a low-pass filter with |
8526 |
++ * alpha=7/8, i.e., |
8527 |
++ * new_rate = (7/8) * old_rate + (1/8) * bw |
8528 |
++ */ |
8529 |
++ do_div(bw, 8); |
8530 |
++ if (bw == 0) |
8531 |
++ return 0; |
8532 |
++ bfqd->peak_rate *= 7; |
8533 |
++ do_div(bfqd->peak_rate, 8); |
8534 |
++ bfqd->peak_rate += bw; |
8535 |
++ update = 1; |
8536 |
++ bfq_log(bfqd, "new peak_rate=%llu", bfqd->peak_rate); |
8537 |
++ } |
8538 |
++ |
8539 |
++ update |= bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES - 1; |
8540 |
++ |
8541 |
++ if (bfqd->peak_rate_samples < BFQ_PEAK_RATE_SAMPLES) |
8542 |
++ bfqd->peak_rate_samples++; |
8543 |
++ |
8544 |
++ if (bfqd->peak_rate_samples == BFQ_PEAK_RATE_SAMPLES && |
8545 |
++ update && bfqd->bfq_user_max_budget == 0) { |
8546 |
++ bfqd->bfq_max_budget = |
8547 |
++ bfq_calc_max_budget(bfqd->peak_rate, timeout); |
8548 |
++ bfq_log(bfqd, "new max_budget=%lu", |
8549 |
++ bfqd->bfq_max_budget); |
8550 |
++ } |
8551 |
++ } |
8552 |
++ |
8553 |
++ /* |
8554 |
++ * If the process has been served for a too short time |
8555 |
++ * interval to let its possible sequential accesses prevail on |
8556 |
++ * the initial seek time needed to move the disk head on the |
8557 |
++ * first sector it requested, then give the process a chance |
8558 |
++ * and for the moment return false. |
8559 |
++ */ |
8560 |
++ if (bfqq->entity.budget <= bfq_max_budget(bfqd) / 8) |
8561 |
++ return 0; |
8562 |
++ |
8563 |
++ /* |
8564 |
++ * A process is considered ``slow'' (i.e., seeky, so that we |
8565 |
++ * cannot treat it fairly in the service domain, as it would |
8566 |
++ * slow down too much the other processes) if, when a slice |
8567 |
++ * ends for whatever reason, it has received service at a |
8568 |
++ * rate that would not be high enough to complete the budget |
8569 |
++ * before the budget timeout expiration. |
8570 |
++ */ |
8571 |
++ expected = bw * 1000 * timeout >> BFQ_RATE_SHIFT; |
8572 |
++ |
8573 |
++ /* |
8574 |
++ * Caveat: processes doing IO in the slower disk zones will |
8575 |
++ * tend to be slow(er) even if not seeky. And the estimated |
8576 |
++ * peak rate will actually be an average over the disk |
8577 |
++ * surface. Hence, to not be too harsh with unlucky processes, |
8578 |
++ * we keep a budget/3 margin of safety before declaring a |
8579 |
++ * process slow. |
8580 |
++ */ |
8581 |
++ return expected > (4 * bfqq->entity.budget) / 3; |
8582 |
++} |
8583 |
++ |
8584 |
++/* |
8585 |
++ * To be deemed as soft real-time, an application must meet two requirements. |
8586 |
++ * First, the application must not require an average bandwidth higher than |
8587 |
++ * the approximate bandwidth required to playback or record a compressed high- |
8588 |
++ * definition video. |
8589 |
++ * The next function is invoked on the completion of the last request of a |
8590 |
++ * batch, to compute the next-start time instant, soft_rt_next_start, such |
8591 |
++ * that, if the next request of the application does not arrive before |
8592 |
++ * soft_rt_next_start, then the above requirement on the bandwidth is met. |
8593 |
++ * |
8594 |
++ * The second requirement is that the request pattern of the application is |
8595 |
++ * isochronous, i.e., that, after issuing a request or a batch of requests, |
8596 |
++ * the application stops issuing new requests until all its pending requests |
8597 |
++ * have been completed. After that, the application may issue a new batch, |
8598 |
++ * and so on. |
8599 |
++ * For this reason the next function is invoked to compute soft_rt_next_start |
8600 |
++ * only for applications that meet this requirement, whereas soft_rt_next_start |
8601 |
++ * is set to infinity for applications that do not. |
8602 |
++ * |
8603 |
++ * Unfortunately, even a greedy application may happen to behave in an |
8604 |
++ * isochronous way if the CPU load is high. In fact, the application may stop |
8605 |
++ * issuing requests while the CPUs are busy serving other processes, then |
8606 |
++ * restart, then stop again for a while, and so on. In addition, if the disk |
8607 |
++ * achieves a low enough throughput with the request pattern issued by the |
8608 |
++ * application (e.g., because the request pattern is random and/or the device |
8609 |
++ * is slow), then the application may meet the above bandwidth requirement too. |
8610 |
++ * To prevent such a greedy application to be deemed as soft real-time, a |
8611 |
++ * further rule is used in the computation of soft_rt_next_start: |
8612 |
++ * soft_rt_next_start must be higher than the current time plus the maximum |
8613 |
++ * time for which the arrival of a request is waited for when a sync queue |
8614 |
++ * becomes idle, namely bfqd->bfq_slice_idle. |
8615 |
++ * This filters out greedy applications, as the latter issue instead their next |
8616 |
++ * request as soon as possible after the last one has been completed (in |
8617 |
++ * contrast, when a batch of requests is completed, a soft real-time application |
8618 |
++ * spends some time processing data). |
8619 |
++ * |
8620 |
++ * Unfortunately, the last filter may easily generate false positives if only |
8621 |
++ * bfqd->bfq_slice_idle is used as a reference time interval and one or both |
8622 |
++ * the following cases occur: |
8623 |
++ * 1) HZ is so low that the duration of a jiffy is comparable to or higher |
8624 |
++ * than bfqd->bfq_slice_idle. This happens, e.g., on slow devices with |
8625 |
++ * HZ=100. |
8626 |
++ * 2) jiffies, instead of increasing at a constant rate, may stop increasing |
8627 |
++ * for a while, then suddenly 'jump' by several units to recover the lost |
8628 |
++ * increments. This seems to happen, e.g., inside virtual machines. |
8629 |
++ * To address this issue, we do not use as a reference time interval just |
8630 |
++ * bfqd->bfq_slice_idle, but bfqd->bfq_slice_idle plus a few jiffies. In |
8631 |
++ * particular we add the minimum number of jiffies for which the filter seems |
8632 |
++ * to be quite precise also in embedded systems and KVM/QEMU virtual machines. |
8633 |
++ */ |
8634 |
++static inline unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd, |
8635 |
++ struct bfq_queue *bfqq) |
8636 |
++{ |
8637 |
++ return max(bfqq->last_idle_bklogged + |
8638 |
++ HZ * bfqq->service_from_backlogged / |
8639 |
++ bfqd->bfq_raising_max_softrt_rate, |
8640 |
++ jiffies + bfqq->bfqd->bfq_slice_idle + 4); |
8641 |
++} |
8642 |
++ |
8643 |
++/* |
8644 |
++ * Return the largest-possible time instant such that, for as long as possible, |
8645 |
++ * the current time will be lower than this time instant according to the macro |
8646 |
++ * time_is_before_jiffies(). |
8647 |
++ */ |
8648 |
++static inline unsigned long bfq_infinity_from_now(unsigned long now) |
8649 |
++{ |
8650 |
++ return now + ULONG_MAX / 2; |
8651 |
++} |
8652 |
++ |
8653 |
++/** |
8654 |
++ * bfq_bfqq_expire - expire a queue. |
8655 |
++ * @bfqd: device owning the queue. |
8656 |
++ * @bfqq: the queue to expire. |
8657 |
++ * @compensate: if true, compensate for the time spent idling. |
8658 |
++ * @reason: the reason causing the expiration. |
8659 |
++ * |
8660 |
++ * |
8661 |
++ * If the process associated to the queue is slow (i.e., seeky), or in |
8662 |
++ * case of budget timeout, or, finally, if it is async, we |
8663 |
++ * artificially charge it an entire budget (independently of the |
8664 |
++ * actual service it received). As a consequence, the queue will get |
8665 |
++ * higher timestamps than the correct ones upon reactivation, and |
8666 |
++ * hence it will be rescheduled as if it had received more service |
8667 |
++ * than what it actually received. In the end, this class of processes |
8668 |
++ * will receive less service in proportion to how slowly they consume |
8669 |
++ * their budgets (and hence how seriously they tend to lower the |
8670 |
++ * throughput). |
8671 |
++ * |
8672 |
++ * In contrast, when a queue expires because it has been idling for |
8673 |
++ * too much or because it exhausted its budget, we do not touch the |
8674 |
++ * amount of service it has received. Hence when the queue will be |
8675 |
++ * reactivated and its timestamps updated, the latter will be in sync |
8676 |
++ * with the actual service received by the queue until expiration. |
8677 |
++ * |
8678 |
++ * Charging a full budget to the first type of queues and the exact |
8679 |
++ * service to the others has the effect of using the WF2Q+ policy to |
8680 |
++ * schedule the former on a timeslice basis, without violating the |
8681 |
++ * service domain guarantees of the latter. |
8682 |
++ */ |
8683 |
++static void bfq_bfqq_expire(struct bfq_data *bfqd, |
8684 |
++ struct bfq_queue *bfqq, |
8685 |
++ int compensate, |
8686 |
++ enum bfqq_expiration reason) |
8687 |
++{ |
8688 |
++ int slow; |
8689 |
++ BUG_ON(bfqq != bfqd->in_service_queue); |
8690 |
++ |
8691 |
++ /* Update disk peak rate for autotuning and check whether the |
8692 |
++ * process is slow (see bfq_update_peak_rate). |
8693 |
++ */ |
8694 |
++ slow = bfq_update_peak_rate(bfqd, bfqq, compensate, reason); |
8695 |
++ |
8696 |
++ /* |
8697 |
++ * As above explained, 'punish' slow (i.e., seeky), timed-out |
8698 |
++ * and async queues, to favor sequential sync workloads. |
8699 |
++ * |
8700 |
++ * Processes doing IO in the slower disk zones will tend to be |
8701 |
++ * slow(er) even if not seeky. Hence, since the estimated peak |
8702 |
++ * rate is actually an average over the disk surface, these |
8703 |
++ * processes may timeout just for bad luck. To avoid punishing |
8704 |
++ * them we do not charge a full budget to a process that |
8705 |
++ * succeeded in consuming at least 2/3 of its budget. |
8706 |
++ */ |
8707 |
++ if (slow || (reason == BFQ_BFQQ_BUDGET_TIMEOUT && |
8708 |
++ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3)) |
8709 |
++ bfq_bfqq_charge_full_budget(bfqq); |
8710 |
++ |
8711 |
++ bfqq->service_from_backlogged += bfqq->entity.service; |
8712 |
++ |
8713 |
++ if (bfqd->low_latency && bfqq->raising_coeff == 1) |
8714 |
++ bfqq->last_rais_start_finish = jiffies; |
8715 |
++ |
8716 |
++ if (bfqd->low_latency && bfqd->bfq_raising_max_softrt_rate > 0 && |
8717 |
++ RB_EMPTY_ROOT(&bfqq->sort_list)) { |
8718 |
++ /* |
8719 |
++ * If we get here, and there are no outstanding requests, |
8720 |
++ * then the request pattern is isochronous (see the comments |
8721 |
++ * to the function bfq_bfqq_softrt_next_start()). Hence we can |
8722 |
++ * compute soft_rt_next_start. If, instead, the queue still |
8723 |
++ * has outstanding requests, then we have to wait for the |
8724 |
++ * completion of all the outstanding requests to discover |
8725 |
++ * whether the request pattern is actually isochronous. |
8726 |
++ */ |
8727 |
++ if (bfqq->dispatched == 0) |
8728 |
++ bfqq->soft_rt_next_start = |
8729 |
++ bfq_bfqq_softrt_next_start(bfqd, bfqq); |
8730 |
++ else { |
8731 |
++ /* |
8732 |
++ * The application is still waiting for the |
8733 |
++ * completion of one or more requests: |
8734 |
++ * prevent it from possibly being incorrectly |
8735 |
++ * deemed as soft real-time by setting its |
8736 |
++ * soft_rt_next_start to infinity. In fact, |
8737 |
++ * without this assignment, the application |
8738 |
++ * would be incorrectly deemed as soft |
8739 |
++ * real-time if: |
8740 |
++ * 1) it issued a new request before the |
8741 |
++ * completion of all its in-flight |
8742 |
++ * requests, and |
8743 |
++ * 2) at that time, its soft_rt_next_start |
8744 |
++ * happened to be in the past. |
8745 |
++ */ |
8746 |
++ bfqq->soft_rt_next_start = |
8747 |
++ bfq_infinity_from_now(jiffies); |
8748 |
++ /* |
8749 |
++ * Schedule an update of soft_rt_next_start to when |
8750 |
++ * the task may be discovered to be isochronous. |
8751 |
++ */ |
8752 |
++ bfq_mark_bfqq_softrt_update(bfqq); |
8753 |
++ } |
8754 |
++ } |
8755 |
++ |
8756 |
++ bfq_log_bfqq(bfqd, bfqq, |
8757 |
++ "expire (%d, slow %d, num_disp %d, idle_win %d)", reason, slow, |
8758 |
++ bfqq->dispatched, bfq_bfqq_idle_window(bfqq)); |
8759 |
++ |
8760 |
++ /* Increase, decrease or leave budget unchanged according to reason */ |
8761 |
++ __bfq_bfqq_recalc_budget(bfqd, bfqq, reason); |
8762 |
++ __bfq_bfqq_expire(bfqd, bfqq); |
8763 |
++} |
8764 |
++ |
8765 |
++/* |
8766 |
++ * Budget timeout is not implemented through a dedicated timer, but |
8767 |
++ * just checked on request arrivals and completions, as well as on |
8768 |
++ * idle timer expirations. |
8769 |
++ */ |
8770 |
++static int bfq_bfqq_budget_timeout(struct bfq_queue *bfqq) |
8771 |
++{ |
8772 |
++ if (bfq_bfqq_budget_new(bfqq)) |
8773 |
++ return 0; |
8774 |
++ |
8775 |
++ if (time_before(jiffies, bfqq->budget_timeout)) |
8776 |
++ return 0; |
8777 |
++ |
8778 |
++ return 1; |
8779 |
++} |
8780 |
++ |
8781 |
++/* |
8782 |
++ * If we expire a queue that is waiting for the arrival of a new |
8783 |
++ * request, we may prevent the fictitious timestamp back-shifting that |
8784 |
++ * allows the guarantees of the queue to be preserved (see [1] for |
8785 |
++ * this tricky aspect). Hence we return true only if this condition |
8786 |
++ * does not hold, or if the queue is slow enough to deserve only to be |
8787 |
++ * kicked off for preserving a high throughput. |
8788 |
++*/ |
8789 |
++static inline int bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq) |
8790 |
++{ |
8791 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, |
8792 |
++ "may_budget_timeout: wr %d left %d timeout %d", |
8793 |
++ bfq_bfqq_wait_request(bfqq), |
8794 |
++ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3, |
8795 |
++ bfq_bfqq_budget_timeout(bfqq)); |
8796 |
++ |
8797 |
++ return (!bfq_bfqq_wait_request(bfqq) || |
8798 |
++ bfq_bfqq_budget_left(bfqq) >= bfqq->entity.budget / 3) |
8799 |
++ && |
8800 |
++ bfq_bfqq_budget_timeout(bfqq); |
8801 |
++} |
8802 |
++ |
8803 |
++/* |
8804 |
++ * For weight-raised queues issuing sync requests, idling is always performed, |
8805 |
++ * as this is instrumental in guaranteeing a high fraction of the throughput |
8806 |
++ * to these queues, and hence in guaranteeing a lower latency for their |
8807 |
++ * requests. See [1] for details. |
8808 |
++ * |
8809 |
++ * For non-weight-raised queues, idling is instead disabled if the device is |
8810 |
++ * NCQ-enabled and non-rotational, as this boosts the throughput on such |
8811 |
++ * devices. |
8812 |
++ */ |
8813 |
++static inline bool bfq_bfqq_must_not_expire(struct bfq_queue *bfqq) |
8814 |
++{ |
8815 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
8816 |
++ |
8817 |
++ return bfq_bfqq_sync(bfqq) && ( |
8818 |
++ bfqq->raising_coeff > 1 || |
8819 |
++ (bfq_bfqq_idle_window(bfqq) && |
8820 |
++ !(bfqd->hw_tag && |
8821 |
++ (blk_queue_nonrot(bfqd->queue) || |
8822 |
++ /* |
8823 |
++ * If there are weight-raised busy queues, then do not idle |
8824 |
++ * the disk for a sync non-weight-raised queue, and hence |
8825 |
++ * expire the queue immediately if empty. Combined with the |
8826 |
++ * timestamping rules of BFQ (see [1] for details), this |
8827 |
++ * causes sync non-weight-raised queues to get a lower |
8828 |
++ * fraction of the disk throughput, and hence reduces the rate |
8829 |
++ * at which the processes associated to these queues ask for |
8830 |
++ * requests from the request pool. |
8831 |
++ * |
8832 |
++ * This is beneficial for weight-raised processes, when the |
8833 |
++ * system operates in request-pool saturation conditions |
8834 |
++ * (e.g., in the presence of write hogs). In fact, if |
8835 |
++ * non-weight-raised processes ask for requests at a lower |
8836 |
++ * rate, then weight-raised processes have a higher |
8837 |
++ * probability to get a request from the pool immediately |
8838 |
++ * (or at least soon) when they need one. Hence they have a |
8839 |
++ * higher probability to actually get a fraction of the disk |
8840 |
++ * throughput proportional to their high weight. This is |
8841 |
++ * especially true with NCQ-enabled drives, which enqueue |
8842 |
++ * several requests in advance and further reorder |
8843 |
++ * internally-queued requests. |
8844 |
++ * |
8845 |
++ * Mistreating non-weight-raised queues in the above-described |
8846 |
++ * way, when there are busy weight-raised queues, seems to |
8847 |
++ * mitigate starvation problems in the presence of heavy write |
8848 |
++ * workloads and NCQ, and hence to guarantee a higher |
8849 |
++ * application and system responsiveness in these hostile |
8850 |
++ * scenarios. |
8851 |
++ */ |
8852 |
++ bfqd->raised_busy_queues > 0) |
8853 |
++ ) |
8854 |
++ ) |
8855 |
++ ); |
8856 |
++} |
8857 |
++ |
8858 |
++/* |
8859 |
++ * If the in-service queue is empty, but it is sync and either of the following |
8860 |
++ * conditions holds, then: 1) the queue must remain in service and cannot be |
8861 |
++ * expired, and 2) the disk must be idled to wait for the possible arrival |
8862 |
++ * of a new request for the queue. The conditions are: |
8863 |
++ * - the device is rotational and not performing NCQ, and the queue has its |
8864 |
++ * idle window set (in this case, waiting for a new request for the queue |
8865 |
++ * is likely to boost the disk throughput); |
8866 |
++ * - the queue is weight-raised (waiting for the request is necessary to |
8867 |
++ * provide the queue with fairness and latency guarantees, see [1] for |
8868 |
++ * details). |
8869 |
++ */ |
8870 |
++static inline bool bfq_bfqq_must_idle(struct bfq_queue *bfqq) |
8871 |
++{ |
8872 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
8873 |
++ |
8874 |
++ return RB_EMPTY_ROOT(&bfqq->sort_list) && bfqd->bfq_slice_idle != 0 && |
8875 |
++ bfq_bfqq_must_not_expire(bfqq) && |
8876 |
++ !bfq_queue_nonrot_noidle(bfqd, bfqq); |
8877 |
++} |
8878 |
++ |
8879 |
++/* |
8880 |
++ * Select a queue for service. If we have a current queue in service, |
8881 |
++ * check whether to continue servicing it, or retrieve and set a new one. |
8882 |
++ */ |
8883 |
++static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd) |
8884 |
++{ |
8885 |
++ struct bfq_queue *bfqq, *new_bfqq = NULL; |
8886 |
++ struct request *next_rq; |
8887 |
++ enum bfqq_expiration reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
8888 |
++ |
8889 |
++ bfqq = bfqd->in_service_queue; |
8890 |
++ if (bfqq == NULL) |
8891 |
++ goto new_queue; |
8892 |
++ |
8893 |
++ bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue"); |
8894 |
++ |
8895 |
++ /* |
8896 |
++ * If another queue has a request waiting within our mean seek |
8897 |
++ * distance, let it run. The expire code will check for close |
8898 |
++ * cooperators and put the close queue at the front of the |
8899 |
++ * service tree. If possible, merge the expiring queue with the |
8900 |
++ * new bfqq. |
8901 |
++ */ |
8902 |
++ new_bfqq = bfq_close_cooperator(bfqd, bfqq); |
8903 |
++ if (new_bfqq != NULL && bfqq->new_bfqq == NULL) |
8904 |
++ bfq_setup_merge(bfqq, new_bfqq); |
8905 |
++ |
8906 |
++ if (bfq_may_expire_for_budg_timeout(bfqq) && |
8907 |
++ !timer_pending(&bfqd->idle_slice_timer) && |
8908 |
++ !bfq_bfqq_must_idle(bfqq)) |
8909 |
++ goto expire; |
8910 |
++ |
8911 |
++ next_rq = bfqq->next_rq; |
8912 |
++ /* |
8913 |
++ * If bfqq has requests queued and it has enough budget left to |
8914 |
++ * serve them, keep the queue, otherwise expire it. |
8915 |
++ */ |
8916 |
++ if (next_rq != NULL) { |
8917 |
++ if (bfq_serv_to_charge(next_rq, bfqq) > |
8918 |
++ bfq_bfqq_budget_left(bfqq)) { |
8919 |
++ reason = BFQ_BFQQ_BUDGET_EXHAUSTED; |
8920 |
++ goto expire; |
8921 |
++ } else { |
8922 |
++ /* |
8923 |
++ * The idle timer may be pending because we may not |
8924 |
++ * disable disk idling even when a new request arrives |
8925 |
++ */ |
8926 |
++ if (timer_pending(&bfqd->idle_slice_timer)) { |
8927 |
++ /* |
8928 |
++ * If we get here: 1) at least a new request |
8929 |
++ * has arrived but we have not disabled the |
8930 |
++ * timer because the request was too small, |
8931 |
++ * 2) then the block layer has unplugged the |
8932 |
++ * device, causing the dispatch to be invoked. |
8933 |
++ * |
8934 |
++ * Since the device is unplugged, now the |
8935 |
++ * requests are probably large enough to |
8936 |
++ * provide a reasonable throughput. |
8937 |
++ * So we disable idling. |
8938 |
++ */ |
8939 |
++ bfq_clear_bfqq_wait_request(bfqq); |
8940 |
++ del_timer(&bfqd->idle_slice_timer); |
8941 |
++ } |
8942 |
++ if (new_bfqq == NULL) |
8943 |
++ goto keep_queue; |
8944 |
++ else |
8945 |
++ goto expire; |
8946 |
++ } |
8947 |
++ } |
8948 |
++ |
8949 |
++ /* |
8950 |
++ * No requests pending. If the in-service queue has no cooperator and |
8951 |
++ * still has requests in flight (possibly waiting for a completion) |
8952 |
++ * or is idling for a new request, then keep it. |
8953 |
++ */ |
8954 |
++ if (new_bfqq == NULL && (timer_pending(&bfqd->idle_slice_timer) || |
8955 |
++ (bfqq->dispatched != 0 && bfq_bfqq_must_not_expire(bfqq)))) { |
8956 |
++ bfqq = NULL; |
8957 |
++ goto keep_queue; |
8958 |
++ } else if (new_bfqq != NULL && timer_pending(&bfqd->idle_slice_timer)) { |
8959 |
++ /* |
8960 |
++ * Expiring the queue because there is a close cooperator, |
8961 |
++ * cancel timer. |
8962 |
++ */ |
8963 |
++ bfq_clear_bfqq_wait_request(bfqq); |
8964 |
++ del_timer(&bfqd->idle_slice_timer); |
8965 |
++ } |
8966 |
++ |
8967 |
++ reason = BFQ_BFQQ_NO_MORE_REQUESTS; |
8968 |
++expire: |
8969 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, reason); |
8970 |
++new_queue: |
8971 |
++ bfqq = bfq_set_in_service_queue(bfqd, new_bfqq); |
8972 |
++ bfq_log(bfqd, "select_queue: new queue %d returned", |
8973 |
++ bfqq != NULL ? bfqq->pid : 0); |
8974 |
++keep_queue: |
8975 |
++ return bfqq; |
8976 |
++} |
8977 |
++ |
8978 |
++static void bfq_update_raising_data(struct bfq_data *bfqd, |
8979 |
++ struct bfq_queue *bfqq) |
8980 |
++{ |
8981 |
++ if (bfqq->raising_coeff > 1) { /* queue is being boosted */ |
8982 |
++ struct bfq_entity *entity = &bfqq->entity; |
8983 |
++ |
8984 |
++ bfq_log_bfqq(bfqd, bfqq, |
8985 |
++ "raising period dur %u/%u msec, " |
8986 |
++ "old raising coeff %u, w %d(%d)", |
8987 |
++ jiffies_to_msecs(jiffies - |
8988 |
++ bfqq->last_rais_start_finish), |
8989 |
++ jiffies_to_msecs(bfqq->raising_cur_max_time), |
8990 |
++ bfqq->raising_coeff, |
8991 |
++ bfqq->entity.weight, bfqq->entity.orig_weight); |
8992 |
++ |
8993 |
++ BUG_ON(bfqq != bfqd->in_service_queue && entity->weight != |
8994 |
++ entity->orig_weight * bfqq->raising_coeff); |
8995 |
++ if (entity->ioprio_changed) |
8996 |
++ bfq_log_bfqq(bfqd, bfqq, |
8997 |
++ "WARN: pending prio change"); |
8998 |
++ /* |
8999 |
++ * If too much time has elapsed from the beginning |
9000 |
++ * of this weight-raising, stop it. |
9001 |
++ */ |
9002 |
++ if (time_is_before_jiffies(bfqq->last_rais_start_finish + |
9003 |
++ bfqq->raising_cur_max_time)) { |
9004 |
++ bfqq->last_rais_start_finish = jiffies; |
9005 |
++ bfq_log_bfqq(bfqd, bfqq, |
9006 |
++ "wrais ending at %lu, " |
9007 |
++ "rais_max_time %u", |
9008 |
++ bfqq->last_rais_start_finish, |
9009 |
++ jiffies_to_msecs(bfqq-> |
9010 |
++ raising_cur_max_time)); |
9011 |
++ bfq_bfqq_end_raising(bfqq); |
9012 |
++ __bfq_entity_update_weight_prio( |
9013 |
++ bfq_entity_service_tree(entity), |
9014 |
++ entity); |
9015 |
++ } |
9016 |
++ } |
9017 |
++} |
9018 |
++ |
9019 |
++/* |
9020 |
++ * Dispatch one request from bfqq, moving it to the request queue |
9021 |
++ * dispatch list. |
9022 |
++ */ |
9023 |
++static int bfq_dispatch_request(struct bfq_data *bfqd, |
9024 |
++ struct bfq_queue *bfqq) |
9025 |
++{ |
9026 |
++ int dispatched = 0; |
9027 |
++ struct request *rq; |
9028 |
++ unsigned long service_to_charge; |
9029 |
++ |
9030 |
++ BUG_ON(RB_EMPTY_ROOT(&bfqq->sort_list)); |
9031 |
++ |
9032 |
++ /* Follow expired path, else get first next available. */ |
9033 |
++ rq = bfq_check_fifo(bfqq); |
9034 |
++ if (rq == NULL) |
9035 |
++ rq = bfqq->next_rq; |
9036 |
++ service_to_charge = bfq_serv_to_charge(rq, bfqq); |
9037 |
++ |
9038 |
++ if (service_to_charge > bfq_bfqq_budget_left(bfqq)) { |
9039 |
++ /* |
9040 |
++ * This may happen if the next rq is chosen |
9041 |
++ * in fifo order instead of sector order. |
9042 |
++ * The budget is properly dimensioned |
9043 |
++ * to be always sufficient to serve the next request |
9044 |
++ * only if it is chosen in sector order. The reason is |
9045 |
++ * that it would be quite inefficient and little useful |
9046 |
++ * to always make sure that the budget is large enough |
9047 |
++ * to serve even the possible next rq in fifo order. |
9048 |
++ * In fact, requests are seldom served in fifo order. |
9049 |
++ * |
9050 |
++ * Expire the queue for budget exhaustion, and |
9051 |
++ * make sure that the next act_budget is enough |
9052 |
++ * to serve the next request, even if it comes |
9053 |
++ * from the fifo expired path. |
9054 |
++ */ |
9055 |
++ bfqq->next_rq = rq; |
9056 |
++ /* |
9057 |
++ * Since this dispatch is failed, make sure that |
9058 |
++ * a new one will be performed |
9059 |
++ */ |
9060 |
++ if (!bfqd->rq_in_driver) |
9061 |
++ bfq_schedule_dispatch(bfqd); |
9062 |
++ goto expire; |
9063 |
++ } |
9064 |
++ |
9065 |
++ /* Finally, insert request into driver dispatch list. */ |
9066 |
++ bfq_bfqq_served(bfqq, service_to_charge); |
9067 |
++ bfq_dispatch_insert(bfqd->queue, rq); |
9068 |
++ |
9069 |
++ bfq_update_raising_data(bfqd, bfqq); |
9070 |
++ |
9071 |
++ bfq_log_bfqq(bfqd, bfqq, |
9072 |
++ "dispatched %u sec req (%llu), budg left %lu", |
9073 |
++ blk_rq_sectors(rq), |
9074 |
++ (long long unsigned)blk_rq_pos(rq), |
9075 |
++ bfq_bfqq_budget_left(bfqq)); |
9076 |
++ |
9077 |
++ dispatched++; |
9078 |
++ |
9079 |
++ if (bfqd->in_service_bic == NULL) { |
9080 |
++ atomic_long_inc(&RQ_BIC(rq)->icq.ioc->refcount); |
9081 |
++ bfqd->in_service_bic = RQ_BIC(rq); |
9082 |
++ } |
9083 |
++ |
9084 |
++ if (bfqd->busy_queues > 1 && ((!bfq_bfqq_sync(bfqq) && |
9085 |
++ dispatched >= bfqd->bfq_max_budget_async_rq) || |
9086 |
++ bfq_class_idle(bfqq))) |
9087 |
++ goto expire; |
9088 |
++ |
9089 |
++ return dispatched; |
9090 |
++ |
9091 |
++expire: |
9092 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_EXHAUSTED); |
9093 |
++ return dispatched; |
9094 |
++} |
9095 |
++ |
9096 |
++static int __bfq_forced_dispatch_bfqq(struct bfq_queue *bfqq) |
9097 |
++{ |
9098 |
++ int dispatched = 0; |
9099 |
++ |
9100 |
++ while (bfqq->next_rq != NULL) { |
9101 |
++ bfq_dispatch_insert(bfqq->bfqd->queue, bfqq->next_rq); |
9102 |
++ dispatched++; |
9103 |
++ } |
9104 |
++ |
9105 |
++ BUG_ON(!list_empty(&bfqq->fifo)); |
9106 |
++ return dispatched; |
9107 |
++} |
9108 |
++ |
9109 |
++/* |
9110 |
++ * Drain our current requests. Used for barriers and when switching |
9111 |
++ * io schedulers on-the-fly. |
9112 |
++ */ |
9113 |
++static int bfq_forced_dispatch(struct bfq_data *bfqd) |
9114 |
++{ |
9115 |
++ struct bfq_queue *bfqq, *n; |
9116 |
++ struct bfq_service_tree *st; |
9117 |
++ int dispatched = 0; |
9118 |
++ |
9119 |
++ bfqq = bfqd->in_service_queue; |
9120 |
++ if (bfqq != NULL) |
9121 |
++ __bfq_bfqq_expire(bfqd, bfqq); |
9122 |
++ |
9123 |
++ /* |
9124 |
++ * Loop through classes, and be careful to leave the scheduler |
9125 |
++ * in a consistent state, as feedback mechanisms and vtime |
9126 |
++ * updates cannot be disabled during the process. |
9127 |
++ */ |
9128 |
++ list_for_each_entry_safe(bfqq, n, &bfqd->active_list, bfqq_list) { |
9129 |
++ st = bfq_entity_service_tree(&bfqq->entity); |
9130 |
++ |
9131 |
++ dispatched += __bfq_forced_dispatch_bfqq(bfqq); |
9132 |
++ bfqq->max_budget = bfq_max_budget(bfqd); |
9133 |
++ |
9134 |
++ bfq_forget_idle(st); |
9135 |
++ } |
9136 |
++ |
9137 |
++ BUG_ON(bfqd->busy_queues != 0); |
9138 |
++ |
9139 |
++ return dispatched; |
9140 |
++} |
9141 |
++ |
9142 |
++static int bfq_dispatch_requests(struct request_queue *q, int force) |
9143 |
++{ |
9144 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
9145 |
++ struct bfq_queue *bfqq; |
9146 |
++ int max_dispatch; |
9147 |
++ |
9148 |
++ bfq_log(bfqd, "dispatch requests: %d busy queues", bfqd->busy_queues); |
9149 |
++ if (bfqd->busy_queues == 0) |
9150 |
++ return 0; |
9151 |
++ |
9152 |
++ if (unlikely(force)) |
9153 |
++ return bfq_forced_dispatch(bfqd); |
9154 |
++ |
9155 |
++ bfqq = bfq_select_queue(bfqd); |
9156 |
++ if (bfqq == NULL) |
9157 |
++ return 0; |
9158 |
++ |
9159 |
++ max_dispatch = bfqd->bfq_quantum; |
9160 |
++ if (bfq_class_idle(bfqq)) |
9161 |
++ max_dispatch = 1; |
9162 |
++ |
9163 |
++ if (!bfq_bfqq_sync(bfqq)) |
9164 |
++ max_dispatch = bfqd->bfq_max_budget_async_rq; |
9165 |
++ |
9166 |
++ if (bfqq->dispatched >= max_dispatch) { |
9167 |
++ if (bfqd->busy_queues > 1) |
9168 |
++ return 0; |
9169 |
++ if (bfqq->dispatched >= 4 * max_dispatch) |
9170 |
++ return 0; |
9171 |
++ } |
9172 |
++ |
9173 |
++ if (bfqd->sync_flight != 0 && !bfq_bfqq_sync(bfqq)) |
9174 |
++ return 0; |
9175 |
++ |
9176 |
++ bfq_clear_bfqq_wait_request(bfqq); |
9177 |
++ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
9178 |
++ |
9179 |
++ if (!bfq_dispatch_request(bfqd, bfqq)) |
9180 |
++ return 0; |
9181 |
++ |
9182 |
++ bfq_log_bfqq(bfqd, bfqq, "dispatched one request of %d (max_disp %d)", |
9183 |
++ bfqq->pid, max_dispatch); |
9184 |
++ |
9185 |
++ return 1; |
9186 |
++} |
9187 |
++ |
9188 |
++/* |
9189 |
++ * Task holds one reference to the queue, dropped when task exits. Each rq |
9190 |
++ * in-flight on this queue also holds a reference, dropped when rq is freed. |
9191 |
++ * |
9192 |
++ * Queue lock must be held here. |
9193 |
++ */ |
9194 |
++static void bfq_put_queue(struct bfq_queue *bfqq) |
9195 |
++{ |
9196 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
9197 |
++ |
9198 |
++ BUG_ON(atomic_read(&bfqq->ref) <= 0); |
9199 |
++ |
9200 |
++ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p %d", bfqq, |
9201 |
++ atomic_read(&bfqq->ref)); |
9202 |
++ if (!atomic_dec_and_test(&bfqq->ref)) |
9203 |
++ return; |
9204 |
++ |
9205 |
++ BUG_ON(rb_first(&bfqq->sort_list) != NULL); |
9206 |
++ BUG_ON(bfqq->allocated[READ] + bfqq->allocated[WRITE] != 0); |
9207 |
++ BUG_ON(bfqq->entity.tree != NULL); |
9208 |
++ BUG_ON(bfq_bfqq_busy(bfqq)); |
9209 |
++ BUG_ON(bfqd->in_service_queue == bfqq); |
9210 |
++ |
9211 |
++ bfq_log_bfqq(bfqd, bfqq, "put_queue: %p freed", bfqq); |
9212 |
++ |
9213 |
++ kmem_cache_free(bfq_pool, bfqq); |
9214 |
++} |
9215 |
++ |
9216 |
++static void bfq_put_cooperator(struct bfq_queue *bfqq) |
9217 |
++{ |
9218 |
++ struct bfq_queue *__bfqq, *next; |
9219 |
++ |
9220 |
++ /* |
9221 |
++ * If this queue was scheduled to merge with another queue, be |
9222 |
++ * sure to drop the reference taken on that queue (and others in |
9223 |
++ * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs. |
9224 |
++ */ |
9225 |
++ __bfqq = bfqq->new_bfqq; |
9226 |
++ while (__bfqq) { |
9227 |
++ if (__bfqq == bfqq) { |
9228 |
++ WARN(1, "bfqq->new_bfqq loop detected.\n"); |
9229 |
++ break; |
9230 |
++ } |
9231 |
++ next = __bfqq->new_bfqq; |
9232 |
++ bfq_put_queue(__bfqq); |
9233 |
++ __bfqq = next; |
9234 |
++ } |
9235 |
++} |
9236 |
++ |
9237 |
++static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
9238 |
++{ |
9239 |
++ if (bfqq == bfqd->in_service_queue) { |
9240 |
++ __bfq_bfqq_expire(bfqd, bfqq); |
9241 |
++ bfq_schedule_dispatch(bfqd); |
9242 |
++ } |
9243 |
++ |
9244 |
++ bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, |
9245 |
++ atomic_read(&bfqq->ref)); |
9246 |
++ |
9247 |
++ bfq_put_cooperator(bfqq); |
9248 |
++ |
9249 |
++ bfq_put_queue(bfqq); |
9250 |
++} |
9251 |
++ |
9252 |
++static void bfq_init_icq(struct io_cq *icq) |
9253 |
++{ |
9254 |
++ struct bfq_io_cq *bic = icq_to_bic(icq); |
9255 |
++ |
9256 |
++ bic->ttime.last_end_request = jiffies; |
9257 |
++} |
9258 |
++ |
9259 |
++static void bfq_exit_icq(struct io_cq *icq) |
9260 |
++{ |
9261 |
++ struct bfq_io_cq *bic = icq_to_bic(icq); |
9262 |
++ struct bfq_data *bfqd = bic_to_bfqd(bic); |
9263 |
++ |
9264 |
++ if (bic->bfqq[BLK_RW_ASYNC]) { |
9265 |
++ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_ASYNC]); |
9266 |
++ bic->bfqq[BLK_RW_ASYNC] = NULL; |
9267 |
++ } |
9268 |
++ |
9269 |
++ if (bic->bfqq[BLK_RW_SYNC]) { |
9270 |
++ bfq_exit_bfqq(bfqd, bic->bfqq[BLK_RW_SYNC]); |
9271 |
++ bic->bfqq[BLK_RW_SYNC] = NULL; |
9272 |
++ } |
9273 |
++} |
9274 |
++ |
9275 |
++/* |
9276 |
++ * Update the entity prio values; note that the new values will not |
9277 |
++ * be used until the next (re)activation. |
9278 |
++ */ |
9279 |
++static void bfq_init_prio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic) |
9280 |
++{ |
9281 |
++ struct task_struct *tsk = current; |
9282 |
++ int ioprio_class; |
9283 |
++ |
9284 |
++ if (!bfq_bfqq_prio_changed(bfqq)) |
9285 |
++ return; |
9286 |
++ |
9287 |
++ ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
9288 |
++ switch (ioprio_class) { |
9289 |
++ default: |
9290 |
++ dev_err(bfqq->bfqd->queue->backing_dev_info.dev, |
9291 |
++ "bfq: bad prio %x\n", ioprio_class); |
9292 |
++ case IOPRIO_CLASS_NONE: |
9293 |
++ /* |
9294 |
++ * No prio set, inherit CPU scheduling settings. |
9295 |
++ */ |
9296 |
++ bfqq->entity.new_ioprio = task_nice_ioprio(tsk); |
9297 |
++ bfqq->entity.new_ioprio_class = task_nice_ioclass(tsk); |
9298 |
++ break; |
9299 |
++ case IOPRIO_CLASS_RT: |
9300 |
++ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
9301 |
++ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_RT; |
9302 |
++ break; |
9303 |
++ case IOPRIO_CLASS_BE: |
9304 |
++ bfqq->entity.new_ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
9305 |
++ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_BE; |
9306 |
++ break; |
9307 |
++ case IOPRIO_CLASS_IDLE: |
9308 |
++ bfqq->entity.new_ioprio_class = IOPRIO_CLASS_IDLE; |
9309 |
++ bfqq->entity.new_ioprio = 7; |
9310 |
++ bfq_clear_bfqq_idle_window(bfqq); |
9311 |
++ break; |
9312 |
++ } |
9313 |
++ |
9314 |
++ bfqq->entity.ioprio_changed = 1; |
9315 |
++ |
9316 |
++ /* |
9317 |
++ * Keep track of original prio settings in case we have to temporarily |
9318 |
++ * elevate the priority of this queue. |
9319 |
++ */ |
9320 |
++ bfqq->org_ioprio = bfqq->entity.new_ioprio; |
9321 |
++ bfq_clear_bfqq_prio_changed(bfqq); |
9322 |
++} |
9323 |
++ |
9324 |
++static void bfq_changed_ioprio(struct bfq_io_cq *bic) |
9325 |
++{ |
9326 |
++ struct bfq_data *bfqd; |
9327 |
++ struct bfq_queue *bfqq, *new_bfqq; |
9328 |
++ struct bfq_group *bfqg; |
9329 |
++ unsigned long uninitialized_var(flags); |
9330 |
++ int ioprio = bic->icq.ioc->ioprio; |
9331 |
++ |
9332 |
++ bfqd = bfq_get_bfqd_locked(&(bic->icq.q->elevator->elevator_data), |
9333 |
++ &flags); |
9334 |
++ /* |
9335 |
++ * This condition may trigger on a newly created bic, be sure to drop |
9336 |
++ * the lock before returning. |
9337 |
++ */ |
9338 |
++ if (unlikely(bfqd == NULL) || likely(bic->ioprio == ioprio)) |
9339 |
++ goto out; |
9340 |
++ |
9341 |
++ bfqq = bic->bfqq[BLK_RW_ASYNC]; |
9342 |
++ if (bfqq != NULL) { |
9343 |
++ bfqg = container_of(bfqq->entity.sched_data, struct bfq_group, |
9344 |
++ sched_data); |
9345 |
++ new_bfqq = bfq_get_queue(bfqd, bfqg, BLK_RW_ASYNC, bic, |
9346 |
++ GFP_ATOMIC); |
9347 |
++ if (new_bfqq != NULL) { |
9348 |
++ bic->bfqq[BLK_RW_ASYNC] = new_bfqq; |
9349 |
++ bfq_log_bfqq(bfqd, bfqq, |
9350 |
++ "changed_ioprio: bfqq %p %d", |
9351 |
++ bfqq, atomic_read(&bfqq->ref)); |
9352 |
++ bfq_put_queue(bfqq); |
9353 |
++ } |
9354 |
++ } |
9355 |
++ |
9356 |
++ bfqq = bic->bfqq[BLK_RW_SYNC]; |
9357 |
++ if (bfqq != NULL) |
9358 |
++ bfq_mark_bfqq_prio_changed(bfqq); |
9359 |
++ |
9360 |
++ bic->ioprio = ioprio; |
9361 |
++ |
9362 |
++out: |
9363 |
++ bfq_put_bfqd_unlock(bfqd, &flags); |
9364 |
++} |
9365 |
++ |
9366 |
++static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
9367 |
++ pid_t pid, int is_sync) |
9368 |
++{ |
9369 |
++ RB_CLEAR_NODE(&bfqq->entity.rb_node); |
9370 |
++ INIT_LIST_HEAD(&bfqq->fifo); |
9371 |
++ |
9372 |
++ atomic_set(&bfqq->ref, 0); |
9373 |
++ bfqq->bfqd = bfqd; |
9374 |
++ |
9375 |
++ bfq_mark_bfqq_prio_changed(bfqq); |
9376 |
++ |
9377 |
++ if (is_sync) { |
9378 |
++ if (!bfq_class_idle(bfqq)) |
9379 |
++ bfq_mark_bfqq_idle_window(bfqq); |
9380 |
++ bfq_mark_bfqq_sync(bfqq); |
9381 |
++ } |
9382 |
++ |
9383 |
++ /* Tentative initial value to trade off between thr and lat */ |
9384 |
++ bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3; |
9385 |
++ bfqq->pid = pid; |
9386 |
++ |
9387 |
++ bfqq->raising_coeff = 1; |
9388 |
++ bfqq->last_rais_start_finish = 0; |
9389 |
++ /* |
9390 |
++ * Set to the value for which bfqq will not be deemed as |
9391 |
++ * soft rt when it becomes backlogged. |
9392 |
++ */ |
9393 |
++ bfqq->soft_rt_next_start = bfq_infinity_from_now(jiffies); |
9394 |
++} |
9395 |
++ |
9396 |
++static struct bfq_queue *bfq_find_alloc_queue(struct bfq_data *bfqd, |
9397 |
++ struct bfq_group *bfqg, |
9398 |
++ int is_sync, |
9399 |
++ struct bfq_io_cq *bic, |
9400 |
++ gfp_t gfp_mask) |
9401 |
++{ |
9402 |
++ struct bfq_queue *bfqq, *new_bfqq = NULL; |
9403 |
++ |
9404 |
++retry: |
9405 |
++ /* bic always exists here */ |
9406 |
++ bfqq = bic_to_bfqq(bic, is_sync); |
9407 |
++ |
9408 |
++ /* |
9409 |
++ * Always try a new alloc if we fall back to the OOM bfqq |
9410 |
++ * originally, since it should just be a temporary situation. |
9411 |
++ */ |
9412 |
++ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
9413 |
++ bfqq = NULL; |
9414 |
++ if (new_bfqq != NULL) { |
9415 |
++ bfqq = new_bfqq; |
9416 |
++ new_bfqq = NULL; |
9417 |
++ } else if (gfp_mask & __GFP_WAIT) { |
9418 |
++ spin_unlock_irq(bfqd->queue->queue_lock); |
9419 |
++ new_bfqq = kmem_cache_alloc_node(bfq_pool, |
9420 |
++ gfp_mask | __GFP_ZERO, |
9421 |
++ bfqd->queue->node); |
9422 |
++ spin_lock_irq(bfqd->queue->queue_lock); |
9423 |
++ if (new_bfqq != NULL) |
9424 |
++ goto retry; |
9425 |
++ } else { |
9426 |
++ bfqq = kmem_cache_alloc_node(bfq_pool, |
9427 |
++ gfp_mask | __GFP_ZERO, |
9428 |
++ bfqd->queue->node); |
9429 |
++ } |
9430 |
++ |
9431 |
++ if (bfqq != NULL) { |
9432 |
++ bfq_init_bfqq(bfqd, bfqq, current->pid, is_sync); |
9433 |
++ bfq_log_bfqq(bfqd, bfqq, "allocated"); |
9434 |
++ } else { |
9435 |
++ bfqq = &bfqd->oom_bfqq; |
9436 |
++ bfq_log_bfqq(bfqd, bfqq, "using oom bfqq"); |
9437 |
++ } |
9438 |
++ |
9439 |
++ bfq_init_prio_data(bfqq, bic); |
9440 |
++ bfq_init_entity(&bfqq->entity, bfqg); |
9441 |
++ } |
9442 |
++ |
9443 |
++ if (new_bfqq != NULL) |
9444 |
++ kmem_cache_free(bfq_pool, new_bfqq); |
9445 |
++ |
9446 |
++ return bfqq; |
9447 |
++} |
9448 |
++ |
9449 |
++static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd, |
9450 |
++ struct bfq_group *bfqg, |
9451 |
++ int ioprio_class, int ioprio) |
9452 |
++{ |
9453 |
++ switch (ioprio_class) { |
9454 |
++ case IOPRIO_CLASS_RT: |
9455 |
++ return &bfqg->async_bfqq[0][ioprio]; |
9456 |
++ case IOPRIO_CLASS_NONE: |
9457 |
++ ioprio = IOPRIO_NORM; |
9458 |
++ /* fall through */ |
9459 |
++ case IOPRIO_CLASS_BE: |
9460 |
++ return &bfqg->async_bfqq[1][ioprio]; |
9461 |
++ case IOPRIO_CLASS_IDLE: |
9462 |
++ return &bfqg->async_idle_bfqq; |
9463 |
++ default: |
9464 |
++ BUG(); |
9465 |
++ } |
9466 |
++} |
9467 |
++ |
9468 |
++static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
9469 |
++ struct bfq_group *bfqg, int is_sync, |
9470 |
++ struct bfq_io_cq *bic, gfp_t gfp_mask) |
9471 |
++{ |
9472 |
++ const int ioprio = IOPRIO_PRIO_DATA(bic->ioprio); |
9473 |
++ const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio); |
9474 |
++ struct bfq_queue **async_bfqq = NULL; |
9475 |
++ struct bfq_queue *bfqq = NULL; |
9476 |
++ |
9477 |
++ if (!is_sync) { |
9478 |
++ async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class, |
9479 |
++ ioprio); |
9480 |
++ bfqq = *async_bfqq; |
9481 |
++ } |
9482 |
++ |
9483 |
++ if (bfqq == NULL) |
9484 |
++ bfqq = bfq_find_alloc_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
9485 |
++ |
9486 |
++ /* |
9487 |
++ * Pin the queue now that it's allocated, scheduler exit will prune it. |
9488 |
++ */ |
9489 |
++ if (!is_sync && *async_bfqq == NULL) { |
9490 |
++ atomic_inc(&bfqq->ref); |
9491 |
++ bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d", |
9492 |
++ bfqq, atomic_read(&bfqq->ref)); |
9493 |
++ *async_bfqq = bfqq; |
9494 |
++ } |
9495 |
++ |
9496 |
++ atomic_inc(&bfqq->ref); |
9497 |
++ bfq_log_bfqq(bfqd, bfqq, "get_queue, at end: %p, %d", bfqq, |
9498 |
++ atomic_read(&bfqq->ref)); |
9499 |
++ return bfqq; |
9500 |
++} |
9501 |
++ |
9502 |
++static void bfq_update_io_thinktime(struct bfq_data *bfqd, |
9503 |
++ struct bfq_io_cq *bic) |
9504 |
++{ |
9505 |
++ unsigned long elapsed = jiffies - bic->ttime.last_end_request; |
9506 |
++ unsigned long ttime = min(elapsed, 2UL * bfqd->bfq_slice_idle); |
9507 |
++ |
9508 |
++ bic->ttime.ttime_samples = (7*bic->ttime.ttime_samples + 256) / 8; |
9509 |
++ bic->ttime.ttime_total = (7*bic->ttime.ttime_total + 256*ttime) / 8; |
9510 |
++ bic->ttime.ttime_mean = (bic->ttime.ttime_total + 128) / |
9511 |
++ bic->ttime.ttime_samples; |
9512 |
++} |
9513 |
++ |
9514 |
++static void bfq_update_io_seektime(struct bfq_data *bfqd, |
9515 |
++ struct bfq_queue *bfqq, |
9516 |
++ struct request *rq) |
9517 |
++{ |
9518 |
++ sector_t sdist; |
9519 |
++ u64 total; |
9520 |
++ |
9521 |
++ if (bfqq->last_request_pos < blk_rq_pos(rq)) |
9522 |
++ sdist = blk_rq_pos(rq) - bfqq->last_request_pos; |
9523 |
++ else |
9524 |
++ sdist = bfqq->last_request_pos - blk_rq_pos(rq); |
9525 |
++ |
9526 |
++ /* |
9527 |
++ * Don't allow the seek distance to get too large from the |
9528 |
++ * odd fragment, pagein, etc. |
9529 |
++ */ |
9530 |
++ if (bfqq->seek_samples == 0) /* first request, not really a seek */ |
9531 |
++ sdist = 0; |
9532 |
++ else if (bfqq->seek_samples <= 60) /* second & third seek */ |
9533 |
++ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*1024); |
9534 |
++ else |
9535 |
++ sdist = min(sdist, (bfqq->seek_mean * 4) + 2*1024*64); |
9536 |
++ |
9537 |
++ bfqq->seek_samples = (7*bfqq->seek_samples + 256) / 8; |
9538 |
++ bfqq->seek_total = (7*bfqq->seek_total + (u64)256*sdist) / 8; |
9539 |
++ total = bfqq->seek_total + (bfqq->seek_samples/2); |
9540 |
++ do_div(total, bfqq->seek_samples); |
9541 |
++ bfqq->seek_mean = (sector_t)total; |
9542 |
++ |
9543 |
++ bfq_log_bfqq(bfqd, bfqq, "dist=%llu mean=%llu", (u64)sdist, |
9544 |
++ (u64)bfqq->seek_mean); |
9545 |
++} |
9546 |
++ |
9547 |
++/* |
9548 |
++ * Disable idle window if the process thinks too long or seeks so much that |
9549 |
++ * it doesn't matter. |
9550 |
++ */ |
9551 |
++static void bfq_update_idle_window(struct bfq_data *bfqd, |
9552 |
++ struct bfq_queue *bfqq, |
9553 |
++ struct bfq_io_cq *bic) |
9554 |
++{ |
9555 |
++ int enable_idle; |
9556 |
++ |
9557 |
++ /* Don't idle for async or idle io prio class. */ |
9558 |
++ if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq)) |
9559 |
++ return; |
9560 |
++ |
9561 |
++ enable_idle = bfq_bfqq_idle_window(bfqq); |
9562 |
++ |
9563 |
++ if (atomic_read(&bic->icq.ioc->active_ref) == 0 || |
9564 |
++ bfqd->bfq_slice_idle == 0 || |
9565 |
++ (bfqd->hw_tag && BFQQ_SEEKY(bfqq) && |
9566 |
++ bfqq->raising_coeff == 1)) |
9567 |
++ enable_idle = 0; |
9568 |
++ else if (bfq_sample_valid(bic->ttime.ttime_samples)) { |
9569 |
++ if (bic->ttime.ttime_mean > bfqd->bfq_slice_idle && |
9570 |
++ bfqq->raising_coeff == 1) |
9571 |
++ enable_idle = 0; |
9572 |
++ else |
9573 |
++ enable_idle = 1; |
9574 |
++ } |
9575 |
++ bfq_log_bfqq(bfqd, bfqq, "update_idle_window: enable_idle %d", |
9576 |
++ enable_idle); |
9577 |
++ |
9578 |
++ if (enable_idle) |
9579 |
++ bfq_mark_bfqq_idle_window(bfqq); |
9580 |
++ else |
9581 |
++ bfq_clear_bfqq_idle_window(bfqq); |
9582 |
++} |
9583 |
++ |
9584 |
++/* |
9585 |
++ * Called when a new fs request (rq) is added to bfqq. Check if there's |
9586 |
++ * something we should do about it. |
9587 |
++ */ |
9588 |
++static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
9589 |
++ struct request *rq) |
9590 |
++{ |
9591 |
++ struct bfq_io_cq *bic = RQ_BIC(rq); |
9592 |
++ |
9593 |
++ if (rq->cmd_flags & REQ_META) |
9594 |
++ bfqq->meta_pending++; |
9595 |
++ |
9596 |
++ bfq_update_io_thinktime(bfqd, bic); |
9597 |
++ bfq_update_io_seektime(bfqd, bfqq, rq); |
9598 |
++ if (bfqq->entity.service > bfq_max_budget(bfqd) / 8 || |
9599 |
++ !BFQQ_SEEKY(bfqq)) |
9600 |
++ bfq_update_idle_window(bfqd, bfqq, bic); |
9601 |
++ |
9602 |
++ bfq_log_bfqq(bfqd, bfqq, |
9603 |
++ "rq_enqueued: idle_window=%d (seeky %d, mean %llu)", |
9604 |
++ bfq_bfqq_idle_window(bfqq), BFQQ_SEEKY(bfqq), |
9605 |
++ (long long unsigned)bfqq->seek_mean); |
9606 |
++ |
9607 |
++ bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq); |
9608 |
++ |
9609 |
++ if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) { |
9610 |
++ int small_req = bfqq->queued[rq_is_sync(rq)] == 1 && |
9611 |
++ blk_rq_sectors(rq) < 32; |
9612 |
++ int budget_timeout = bfq_bfqq_budget_timeout(bfqq); |
9613 |
++ |
9614 |
++ /* |
9615 |
++ * There is just this request queued: if the request |
9616 |
++ * is small and the queue is not to be expired, then |
9617 |
++ * just exit. |
9618 |
++ * |
9619 |
++ * In this way, if the disk is being idled to wait for |
9620 |
++ * a new request from the in-service queue, we avoid |
9621 |
++ * unplugging the device and committing the disk to serve |
9622 |
++ * just a small request. On the contrary, we wait for |
9623 |
++ * the block layer to decide when to unplug the device: |
9624 |
++ * hopefully, new requests will be merged to this one |
9625 |
++ * quickly, then the device will be unplugged and |
9626 |
++ * larger requests will be dispatched. |
9627 |
++ */ |
9628 |
++ if (small_req && !budget_timeout) |
9629 |
++ return; |
9630 |
++ |
9631 |
++ /* |
9632 |
++ * A large enough request arrived, or the queue is to |
9633 |
++ * be expired: in both cases disk idling is to be |
9634 |
++ * stopped, so clear wait_request flag and reset |
9635 |
++ * timer. |
9636 |
++ */ |
9637 |
++ bfq_clear_bfqq_wait_request(bfqq); |
9638 |
++ del_timer(&bfqd->idle_slice_timer); |
9639 |
++ |
9640 |
++ /* |
9641 |
++ * The queue is not empty, because a new request just |
9642 |
++ * arrived. Hence we can safely expire the queue, in |
9643 |
++ * case of budget timeout, without risking that the |
9644 |
++ * timestamps of the queue are not updated correctly. |
9645 |
++ * See [1] for more details. |
9646 |
++ */ |
9647 |
++ if (budget_timeout) |
9648 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_TIMEOUT); |
9649 |
++ |
9650 |
++ /* |
9651 |
++ * Let the request rip immediately, or let a new queue be |
9652 |
++ * selected if bfqq has just been expired. |
9653 |
++ */ |
9654 |
++ __blk_run_queue(bfqd->queue); |
9655 |
++ } |
9656 |
++} |
9657 |
++ |
9658 |
++static void bfq_insert_request(struct request_queue *q, struct request *rq) |
9659 |
++{ |
9660 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
9661 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
9662 |
++ |
9663 |
++ assert_spin_locked(bfqd->queue->queue_lock); |
9664 |
++ bfq_init_prio_data(bfqq, RQ_BIC(rq)); |
9665 |
++ |
9666 |
++ bfq_add_rq_rb(rq); |
9667 |
++ |
9668 |
++ rq_set_fifo_time(rq, jiffies + bfqd->bfq_fifo_expire[rq_is_sync(rq)]); |
9669 |
++ list_add_tail(&rq->queuelist, &bfqq->fifo); |
9670 |
++ |
9671 |
++ bfq_rq_enqueued(bfqd, bfqq, rq); |
9672 |
++} |
9673 |
++ |
9674 |
++static void bfq_update_hw_tag(struct bfq_data *bfqd) |
9675 |
++{ |
9676 |
++ bfqd->max_rq_in_driver = max(bfqd->max_rq_in_driver, |
9677 |
++ bfqd->rq_in_driver); |
9678 |
++ |
9679 |
++ if (bfqd->hw_tag == 1) |
9680 |
++ return; |
9681 |
++ |
9682 |
++ /* |
9683 |
++ * This sample is valid if the number of outstanding requests |
9684 |
++ * is large enough to allow a queueing behavior. Note that the |
9685 |
++ * sum is not exact, as it's not taking into account deactivated |
9686 |
++ * requests. |
9687 |
++ */ |
9688 |
++ if (bfqd->rq_in_driver + bfqd->queued < BFQ_HW_QUEUE_THRESHOLD) |
9689 |
++ return; |
9690 |
++ |
9691 |
++ if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES) |
9692 |
++ return; |
9693 |
++ |
9694 |
++ bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD; |
9695 |
++ bfqd->max_rq_in_driver = 0; |
9696 |
++ bfqd->hw_tag_samples = 0; |
9697 |
++} |
9698 |
++ |
9699 |
++static void bfq_completed_request(struct request_queue *q, struct request *rq) |
9700 |
++{ |
9701 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
9702 |
++ struct bfq_data *bfqd = bfqq->bfqd; |
9703 |
++ const int sync = rq_is_sync(rq); |
9704 |
++ |
9705 |
++ bfq_log_bfqq(bfqd, bfqq, "completed %u sects req (%d)", |
9706 |
++ blk_rq_sectors(rq), sync); |
9707 |
++ |
9708 |
++ bfq_update_hw_tag(bfqd); |
9709 |
++ |
9710 |
++ WARN_ON(!bfqd->rq_in_driver); |
9711 |
++ WARN_ON(!bfqq->dispatched); |
9712 |
++ bfqd->rq_in_driver--; |
9713 |
++ bfqq->dispatched--; |
9714 |
++ |
9715 |
++ if (bfq_bfqq_sync(bfqq)) |
9716 |
++ bfqd->sync_flight--; |
9717 |
++ |
9718 |
++ if (sync) |
9719 |
++ RQ_BIC(rq)->ttime.last_end_request = jiffies; |
9720 |
++ |
9721 |
++ /* |
9722 |
++ * If we are waiting to discover whether the request pattern of the |
9723 |
++ * task associated with the queue is actually isochronous, and |
9724 |
++ * both requisites for this condition to hold are satisfied, then |
9725 |
++ * compute soft_rt_next_start (see the comments to the function |
9726 |
++ * bfq_bfqq_softrt_next_start()). |
9727 |
++ */ |
9728 |
++ if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 && |
9729 |
++ RB_EMPTY_ROOT(&bfqq->sort_list)) |
9730 |
++ bfqq->soft_rt_next_start = |
9731 |
++ bfq_bfqq_softrt_next_start(bfqd, bfqq); |
9732 |
++ |
9733 |
++ /* |
9734 |
++ * If this is the in-service queue, check if it needs to be expired, |
9735 |
++ * or if we want to idle in case it has no pending requests. |
9736 |
++ */ |
9737 |
++ if (bfqd->in_service_queue == bfqq) { |
9738 |
++ if (bfq_bfqq_budget_new(bfqq)) |
9739 |
++ bfq_set_budget_timeout(bfqd); |
9740 |
++ |
9741 |
++ if (bfq_bfqq_must_idle(bfqq)) { |
9742 |
++ bfq_arm_slice_timer(bfqd); |
9743 |
++ goto out; |
9744 |
++ } else if (bfq_may_expire_for_budg_timeout(bfqq)) |
9745 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, BFQ_BFQQ_BUDGET_TIMEOUT); |
9746 |
++ else if (RB_EMPTY_ROOT(&bfqq->sort_list) && |
9747 |
++ (bfqq->dispatched == 0 || |
9748 |
++ !bfq_bfqq_must_not_expire(bfqq))) |
9749 |
++ bfq_bfqq_expire(bfqd, bfqq, 0, |
9750 |
++ BFQ_BFQQ_NO_MORE_REQUESTS); |
9751 |
++ } |
9752 |
++ |
9753 |
++ if (!bfqd->rq_in_driver) |
9754 |
++ bfq_schedule_dispatch(bfqd); |
9755 |
++ |
9756 |
++out: |
9757 |
++ return; |
9758 |
++} |
9759 |
++ |
9760 |
++static inline int __bfq_may_queue(struct bfq_queue *bfqq) |
9761 |
++{ |
9762 |
++ if (bfq_bfqq_wait_request(bfqq) && bfq_bfqq_must_alloc(bfqq)) { |
9763 |
++ bfq_clear_bfqq_must_alloc(bfqq); |
9764 |
++ return ELV_MQUEUE_MUST; |
9765 |
++ } |
9766 |
++ |
9767 |
++ return ELV_MQUEUE_MAY; |
9768 |
++} |
9769 |
++ |
9770 |
++static int bfq_may_queue(struct request_queue *q, int rw) |
9771 |
++{ |
9772 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
9773 |
++ struct task_struct *tsk = current; |
9774 |
++ struct bfq_io_cq *bic; |
9775 |
++ struct bfq_queue *bfqq; |
9776 |
++ |
9777 |
++ /* |
9778 |
++ * Don't force setup of a queue from here, as a call to may_queue |
9779 |
++ * does not necessarily imply that a request actually will be queued. |
9780 |
++ * So just lookup a possibly existing queue, or return 'may queue' |
9781 |
++ * if that fails. |
9782 |
++ */ |
9783 |
++ bic = bfq_bic_lookup(bfqd, tsk->io_context); |
9784 |
++ if (bic == NULL) |
9785 |
++ return ELV_MQUEUE_MAY; |
9786 |
++ |
9787 |
++ bfqq = bic_to_bfqq(bic, rw_is_sync(rw)); |
9788 |
++ if (bfqq != NULL) { |
9789 |
++ bfq_init_prio_data(bfqq, bic); |
9790 |
++ |
9791 |
++ return __bfq_may_queue(bfqq); |
9792 |
++ } |
9793 |
++ |
9794 |
++ return ELV_MQUEUE_MAY; |
9795 |
++} |
9796 |
++ |
9797 |
++/* |
9798 |
++ * Queue lock held here. |
9799 |
++ */ |
9800 |
++static void bfq_put_request(struct request *rq) |
9801 |
++{ |
9802 |
++ struct bfq_queue *bfqq = RQ_BFQQ(rq); |
9803 |
++ |
9804 |
++ if (bfqq != NULL) { |
9805 |
++ const int rw = rq_data_dir(rq); |
9806 |
++ |
9807 |
++ BUG_ON(!bfqq->allocated[rw]); |
9808 |
++ bfqq->allocated[rw]--; |
9809 |
++ |
9810 |
++ rq->elv.priv[0] = NULL; |
9811 |
++ rq->elv.priv[1] = NULL; |
9812 |
++ |
9813 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "put_request %p, %d", |
9814 |
++ bfqq, atomic_read(&bfqq->ref)); |
9815 |
++ bfq_put_queue(bfqq); |
9816 |
++ } |
9817 |
++} |
9818 |
++ |
9819 |
++static struct bfq_queue * |
9820 |
++bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic, |
9821 |
++ struct bfq_queue *bfqq) |
9822 |
++{ |
9823 |
++ bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu", |
9824 |
++ (long unsigned)bfqq->new_bfqq->pid); |
9825 |
++ bic_set_bfqq(bic, bfqq->new_bfqq, 1); |
9826 |
++ bfq_mark_bfqq_coop(bfqq->new_bfqq); |
9827 |
++ bfq_put_queue(bfqq); |
9828 |
++ return bic_to_bfqq(bic, 1); |
9829 |
++} |
9830 |
++ |
9831 |
++/* |
9832 |
++ * Returns NULL if a new bfqq should be allocated, or the old bfqq if this |
9833 |
++ * was the last process referring to said bfqq. |
9834 |
++ */ |
9835 |
++static struct bfq_queue * |
9836 |
++bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq) |
9837 |
++{ |
9838 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue"); |
9839 |
++ if (bfqq_process_refs(bfqq) == 1) { |
9840 |
++ bfqq->pid = current->pid; |
9841 |
++ bfq_clear_bfqq_coop(bfqq); |
9842 |
++ bfq_clear_bfqq_split_coop(bfqq); |
9843 |
++ return bfqq; |
9844 |
++ } |
9845 |
++ |
9846 |
++ bic_set_bfqq(bic, NULL, 1); |
9847 |
++ |
9848 |
++ bfq_put_cooperator(bfqq); |
9849 |
++ |
9850 |
++ bfq_put_queue(bfqq); |
9851 |
++ return NULL; |
9852 |
++} |
9853 |
++ |
9854 |
++/* |
9855 |
++ * Allocate bfq data structures associated with this request. |
9856 |
++ */ |
9857 |
++static int bfq_set_request(struct request_queue *q, struct request *rq, |
9858 |
++ struct bio *bio, gfp_t gfp_mask) |
9859 |
++{ |
9860 |
++ struct bfq_data *bfqd = q->elevator->elevator_data; |
9861 |
++ struct bfq_io_cq *bic = icq_to_bic(rq->elv.icq); |
9862 |
++ const int rw = rq_data_dir(rq); |
9863 |
++ const int is_sync = rq_is_sync(rq); |
9864 |
++ struct bfq_queue *bfqq; |
9865 |
++ struct bfq_group *bfqg; |
9866 |
++ unsigned long flags; |
9867 |
++ |
9868 |
++ might_sleep_if(gfp_mask & __GFP_WAIT); |
9869 |
++ |
9870 |
++ bfq_changed_ioprio(bic); |
9871 |
++ |
9872 |
++ spin_lock_irqsave(q->queue_lock, flags); |
9873 |
++ |
9874 |
++ if (bic == NULL) |
9875 |
++ goto queue_fail; |
9876 |
++ |
9877 |
++ bfqg = bfq_bic_update_cgroup(bic); |
9878 |
++ |
9879 |
++new_queue: |
9880 |
++ bfqq = bic_to_bfqq(bic, is_sync); |
9881 |
++ if (bfqq == NULL || bfqq == &bfqd->oom_bfqq) { |
9882 |
++ bfqq = bfq_get_queue(bfqd, bfqg, is_sync, bic, gfp_mask); |
9883 |
++ bic_set_bfqq(bic, bfqq, is_sync); |
9884 |
++ } else { |
9885 |
++ /* |
9886 |
++ * If the queue was seeky for too long, break it apart. |
9887 |
++ */ |
9888 |
++ if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq)) { |
9889 |
++ bfq_log_bfqq(bfqd, bfqq, "breaking apart bfqq"); |
9890 |
++ bfqq = bfq_split_bfqq(bic, bfqq); |
9891 |
++ if (!bfqq) |
9892 |
++ goto new_queue; |
9893 |
++ } |
9894 |
++ |
9895 |
++ /* |
9896 |
++ * Check to see if this queue is scheduled to merge with |
9897 |
++ * another closely cooperating queue. The merging of queues |
9898 |
++ * happens here as it must be done in process context. |
9899 |
++ * The reference on new_bfqq was taken in merge_bfqqs. |
9900 |
++ */ |
9901 |
++ if (bfqq->new_bfqq != NULL) |
9902 |
++ bfqq = bfq_merge_bfqqs(bfqd, bic, bfqq); |
9903 |
++ } |
9904 |
++ |
9905 |
++ bfqq->allocated[rw]++; |
9906 |
++ atomic_inc(&bfqq->ref); |
9907 |
++ bfq_log_bfqq(bfqd, bfqq, "set_request: bfqq %p, %d", bfqq, |
9908 |
++ atomic_read(&bfqq->ref)); |
9909 |
++ |
9910 |
++ rq->elv.priv[0] = bic; |
9911 |
++ rq->elv.priv[1] = bfqq; |
9912 |
++ |
9913 |
++ spin_unlock_irqrestore(q->queue_lock, flags); |
9914 |
++ |
9915 |
++ return 0; |
9916 |
++ |
9917 |
++queue_fail: |
9918 |
++ bfq_schedule_dispatch(bfqd); |
9919 |
++ spin_unlock_irqrestore(q->queue_lock, flags); |
9920 |
++ |
9921 |
++ return 1; |
9922 |
++} |
9923 |
++ |
9924 |
++static void bfq_kick_queue(struct work_struct *work) |
9925 |
++{ |
9926 |
++ struct bfq_data *bfqd = |
9927 |
++ container_of(work, struct bfq_data, unplug_work); |
9928 |
++ struct request_queue *q = bfqd->queue; |
9929 |
++ |
9930 |
++ spin_lock_irq(q->queue_lock); |
9931 |
++ __blk_run_queue(q); |
9932 |
++ spin_unlock_irq(q->queue_lock); |
9933 |
++} |
9934 |
++ |
9935 |
++/* |
9936 |
++ * Handler of the expiration of the timer running if the in-service queue |
9937 |
++ * is idling inside its time slice. |
9938 |
++ */ |
9939 |
++static void bfq_idle_slice_timer(unsigned long data) |
9940 |
++{ |
9941 |
++ struct bfq_data *bfqd = (struct bfq_data *)data; |
9942 |
++ struct bfq_queue *bfqq; |
9943 |
++ unsigned long flags; |
9944 |
++ enum bfqq_expiration reason; |
9945 |
++ |
9946 |
++ spin_lock_irqsave(bfqd->queue->queue_lock, flags); |
9947 |
++ |
9948 |
++ bfqq = bfqd->in_service_queue; |
9949 |
++ /* |
9950 |
++ * Theoretical race here: the in-service queue can be NULL or different |
9951 |
++ * from the queue that was idling if the timer handler spins on |
9952 |
++ * the queue_lock and a new request arrives for the current |
9953 |
++ * queue and there is a full dispatch cycle that changes the |
9954 |
++ * in-service queue. This can hardly happen, but in the worst case |
9955 |
++ * we just expire a queue too early. |
9956 |
++ */ |
9957 |
++ if (bfqq != NULL) { |
9958 |
++ bfq_log_bfqq(bfqd, bfqq, "slice_timer expired"); |
9959 |
++ if (bfq_bfqq_budget_timeout(bfqq)) |
9960 |
++ /* |
9961 |
++ * Also here the queue can be safely expired |
9962 |
++ * for budget timeout without wasting |
9963 |
++ * guarantees |
9964 |
++ */ |
9965 |
++ reason = BFQ_BFQQ_BUDGET_TIMEOUT; |
9966 |
++ else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0) |
9967 |
++ /* |
9968 |
++ * The queue may not be empty upon timer expiration, |
9969 |
++ * because we may not disable the timer when the first |
9970 |
++ * request of the in-service queue arrives during |
9971 |
++ * disk idling |
9972 |
++ */ |
9973 |
++ reason = BFQ_BFQQ_TOO_IDLE; |
9974 |
++ else |
9975 |
++ goto schedule_dispatch; |
9976 |
++ |
9977 |
++ bfq_bfqq_expire(bfqd, bfqq, 1, reason); |
9978 |
++ } |
9979 |
++ |
9980 |
++schedule_dispatch: |
9981 |
++ bfq_schedule_dispatch(bfqd); |
9982 |
++ |
9983 |
++ spin_unlock_irqrestore(bfqd->queue->queue_lock, flags); |
9984 |
++} |
9985 |
++ |
9986 |
++static void bfq_shutdown_timer_wq(struct bfq_data *bfqd) |
9987 |
++{ |
9988 |
++ del_timer_sync(&bfqd->idle_slice_timer); |
9989 |
++ cancel_work_sync(&bfqd->unplug_work); |
9990 |
++} |
9991 |
++ |
9992 |
++static inline void __bfq_put_async_bfqq(struct bfq_data *bfqd, |
9993 |
++ struct bfq_queue **bfqq_ptr) |
9994 |
++{ |
9995 |
++ struct bfq_group *root_group = bfqd->root_group; |
9996 |
++ struct bfq_queue *bfqq = *bfqq_ptr; |
9997 |
++ |
9998 |
++ bfq_log(bfqd, "put_async_bfqq: %p", bfqq); |
9999 |
++ if (bfqq != NULL) { |
10000 |
++ bfq_bfqq_move(bfqd, bfqq, &bfqq->entity, root_group); |
10001 |
++ bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d", |
10002 |
++ bfqq, atomic_read(&bfqq->ref)); |
10003 |
++ bfq_put_queue(bfqq); |
10004 |
++ *bfqq_ptr = NULL; |
10005 |
++ } |
10006 |
++} |
10007 |
++ |
10008 |
++/* |
10009 |
++ * Release all the bfqg references to its async queues. If we are |
10010 |
++ * deallocating the group these queues may still contain requests, so |
10011 |
++ * we reparent them to the root cgroup (i.e., the only one that will |
10012 |
++ * exist for sure until all the requests on a device are gone). |
10013 |
++ */ |
10014 |
++static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg) |
10015 |
++{ |
10016 |
++ int i, j; |
10017 |
++ |
10018 |
++ for (i = 0; i < 2; i++) |
10019 |
++ for (j = 0; j < IOPRIO_BE_NR; j++) |
10020 |
++ __bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j]); |
10021 |
++ |
10022 |
++ __bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq); |
10023 |
++} |
10024 |
++ |
10025 |
++static void bfq_exit_queue(struct elevator_queue *e) |
10026 |
++{ |
10027 |
++ struct bfq_data *bfqd = e->elevator_data; |
10028 |
++ struct request_queue *q = bfqd->queue; |
10029 |
++ struct bfq_queue *bfqq, *n; |
10030 |
++ |
10031 |
++ bfq_shutdown_timer_wq(bfqd); |
10032 |
++ |
10033 |
++ spin_lock_irq(q->queue_lock); |
10034 |
++ |
10035 |
++ BUG_ON(bfqd->in_service_queue != NULL); |
10036 |
++ list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list) |
10037 |
++ bfq_deactivate_bfqq(bfqd, bfqq, 0); |
10038 |
++ |
10039 |
++ bfq_disconnect_groups(bfqd); |
10040 |
++ spin_unlock_irq(q->queue_lock); |
10041 |
++ |
10042 |
++ bfq_shutdown_timer_wq(bfqd); |
10043 |
++ |
10044 |
++ synchronize_rcu(); |
10045 |
++ |
10046 |
++ BUG_ON(timer_pending(&bfqd->idle_slice_timer)); |
10047 |
++ |
10048 |
++ bfq_free_root_group(bfqd); |
10049 |
++ kfree(bfqd); |
10050 |
++} |
10051 |
++ |
10052 |
++static int bfq_init_queue(struct request_queue *q, struct elevator_type *e) |
10053 |
++{ |
10054 |
++ struct bfq_group *bfqg; |
10055 |
++ struct bfq_data *bfqd; |
10056 |
++ struct elevator_queue *eq; |
10057 |
++ |
10058 |
++ eq = elevator_alloc(q, e); |
10059 |
++ if (eq == NULL) |
10060 |
++ return -ENOMEM; |
10061 |
++ |
10062 |
++ bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node); |
10063 |
++ if (bfqd == NULL) { |
10064 |
++ kobject_put(&eq->kobj); |
10065 |
++ return -ENOMEM; |
10066 |
++ } |
10067 |
++ eq->elevator_data = bfqd; |
10068 |
++ |
10069 |
++ /* |
10070 |
++ * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues. |
10071 |
++ * Grab a permanent reference to it, so that the normal code flow |
10072 |
++ * will not attempt to free it. |
10073 |
++ */ |
10074 |
++ bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, 1, 0); |
10075 |
++ atomic_inc(&bfqd->oom_bfqq.ref); |
10076 |
++ |
10077 |
++ bfqd->queue = q; |
10078 |
++ |
10079 |
++ spin_lock_irq(q->queue_lock); |
10080 |
++ q->elevator = eq; |
10081 |
++ spin_unlock_irq(q->queue_lock); |
10082 |
++ |
10083 |
++ bfqg = bfq_alloc_root_group(bfqd, q->node); |
10084 |
++ if (bfqg == NULL) { |
10085 |
++ kfree(bfqd); |
10086 |
++ kobject_put(&eq->kobj); |
10087 |
++ return -ENOMEM; |
10088 |
++ } |
10089 |
++ |
10090 |
++ bfqd->root_group = bfqg; |
10091 |
++ |
10092 |
++ init_timer(&bfqd->idle_slice_timer); |
10093 |
++ bfqd->idle_slice_timer.function = bfq_idle_slice_timer; |
10094 |
++ bfqd->idle_slice_timer.data = (unsigned long)bfqd; |
10095 |
++ |
10096 |
++ bfqd->rq_pos_tree = RB_ROOT; |
10097 |
++ |
10098 |
++ INIT_WORK(&bfqd->unplug_work, bfq_kick_queue); |
10099 |
++ |
10100 |
++ INIT_LIST_HEAD(&bfqd->active_list); |
10101 |
++ INIT_LIST_HEAD(&bfqd->idle_list); |
10102 |
++ |
10103 |
++ bfqd->hw_tag = -1; |
10104 |
++ |
10105 |
++ bfqd->bfq_max_budget = bfq_default_max_budget; |
10106 |
++ |
10107 |
++ bfqd->bfq_quantum = bfq_quantum; |
10108 |
++ bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0]; |
10109 |
++ bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1]; |
10110 |
++ bfqd->bfq_back_max = bfq_back_max; |
10111 |
++ bfqd->bfq_back_penalty = bfq_back_penalty; |
10112 |
++ bfqd->bfq_slice_idle = bfq_slice_idle; |
10113 |
++ bfqd->bfq_class_idle_last_service = 0; |
10114 |
++ bfqd->bfq_max_budget_async_rq = bfq_max_budget_async_rq; |
10115 |
++ bfqd->bfq_timeout[BLK_RW_ASYNC] = bfq_timeout_async; |
10116 |
++ bfqd->bfq_timeout[BLK_RW_SYNC] = bfq_timeout_sync; |
10117 |
++ |
10118 |
++ bfqd->low_latency = true; |
10119 |
++ |
10120 |
++ bfqd->bfq_raising_coeff = 20; |
10121 |
++ bfqd->bfq_raising_rt_max_time = msecs_to_jiffies(300); |
10122 |
++ bfqd->bfq_raising_max_time = 0; |
10123 |
++ bfqd->bfq_raising_min_idle_time = msecs_to_jiffies(2000); |
10124 |
++ bfqd->bfq_raising_min_inter_arr_async = msecs_to_jiffies(500); |
10125 |
++ bfqd->bfq_raising_max_softrt_rate = 7000; /* |
10126 |
++ * Approximate rate required |
10127 |
++ * to playback or record a |
10128 |
++ * high-definition compressed |
10129 |
++ * video. |
10130 |
++ */ |
10131 |
++ bfqd->raised_busy_queues = 0; |
10132 |
++ |
10133 |
++ /* Initially estimate the device's peak rate as the reference rate */ |
10134 |
++ if (blk_queue_nonrot(bfqd->queue)) { |
10135 |
++ bfqd->RT_prod = R_nonrot * T_nonrot; |
10136 |
++ bfqd->peak_rate = R_nonrot; |
10137 |
++ } else { |
10138 |
++ bfqd->RT_prod = R_rot * T_rot; |
10139 |
++ bfqd->peak_rate = R_rot; |
10140 |
++ } |
10141 |
++ |
10142 |
++ return 0; |
10143 |
++} |
10144 |
++ |
10145 |
++static void bfq_slab_kill(void) |
10146 |
++{ |
10147 |
++ if (bfq_pool != NULL) |
10148 |
++ kmem_cache_destroy(bfq_pool); |
10149 |
++} |
10150 |
++ |
10151 |
++static int __init bfq_slab_setup(void) |
10152 |
++{ |
10153 |
++ bfq_pool = KMEM_CACHE(bfq_queue, 0); |
10154 |
++ if (bfq_pool == NULL) |
10155 |
++ return -ENOMEM; |
10156 |
++ return 0; |
10157 |
++} |
10158 |
++ |
10159 |
++static ssize_t bfq_var_show(unsigned int var, char *page) |
10160 |
++{ |
10161 |
++ return sprintf(page, "%d\n", var); |
10162 |
++} |
10163 |
++ |
10164 |
++static ssize_t bfq_var_store(unsigned long *var, const char *page, size_t count) |
10165 |
++{ |
10166 |
++ unsigned long new_val; |
10167 |
++ int ret = kstrtoul(page, 10, &new_val); |
10168 |
++ |
10169 |
++ if (ret == 0) |
10170 |
++ *var = new_val; |
10171 |
++ |
10172 |
++ return count; |
10173 |
++} |
10174 |
++ |
10175 |
++static ssize_t bfq_raising_max_time_show(struct elevator_queue *e, char *page) |
10176 |
++{ |
10177 |
++ struct bfq_data *bfqd = e->elevator_data; |
10178 |
++ return sprintf(page, "%d\n", bfqd->bfq_raising_max_time > 0 ? |
10179 |
++ jiffies_to_msecs(bfqd->bfq_raising_max_time) : |
10180 |
++ jiffies_to_msecs(bfq_wrais_duration(bfqd))); |
10181 |
++} |
10182 |
++ |
10183 |
++static ssize_t bfq_weights_show(struct elevator_queue *e, char *page) |
10184 |
++{ |
10185 |
++ struct bfq_queue *bfqq; |
10186 |
++ struct bfq_data *bfqd = e->elevator_data; |
10187 |
++ ssize_t num_char = 0; |
10188 |
++ |
10189 |
++ num_char += sprintf(page + num_char, "Tot reqs queued %d\n\n", |
10190 |
++ bfqd->queued); |
10191 |
++ |
10192 |
++ spin_lock_irq(bfqd->queue->queue_lock); |
10193 |
++ |
10194 |
++ num_char += sprintf(page + num_char, "Active:\n"); |
10195 |
++ list_for_each_entry(bfqq, &bfqd->active_list, bfqq_list) { |
10196 |
++ num_char += sprintf(page + num_char, |
10197 |
++ "pid%d: weight %hu, nr_queued %d %d," |
10198 |
++ " dur %d/%u\n", |
10199 |
++ bfqq->pid, |
10200 |
++ bfqq->entity.weight, |
10201 |
++ bfqq->queued[0], |
10202 |
++ bfqq->queued[1], |
10203 |
++ jiffies_to_msecs(jiffies - |
10204 |
++ bfqq->last_rais_start_finish), |
10205 |
++ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
10206 |
++ } |
10207 |
++ |
10208 |
++ num_char += sprintf(page + num_char, "Idle:\n"); |
10209 |
++ list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list) { |
10210 |
++ num_char += sprintf(page + num_char, |
10211 |
++ "pid%d: weight %hu, dur %d/%u\n", |
10212 |
++ bfqq->pid, |
10213 |
++ bfqq->entity.weight, |
10214 |
++ jiffies_to_msecs(jiffies - |
10215 |
++ bfqq->last_rais_start_finish), |
10216 |
++ jiffies_to_msecs(bfqq->raising_cur_max_time)); |
10217 |
++ } |
10218 |
++ |
10219 |
++ spin_unlock_irq(bfqd->queue->queue_lock); |
10220 |
++ |
10221 |
++ return num_char; |
10222 |
++} |
10223 |
++ |
10224 |
++#define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \ |
10225 |
++static ssize_t __FUNC(struct elevator_queue *e, char *page) \ |
10226 |
++{ \ |
10227 |
++ struct bfq_data *bfqd = e->elevator_data; \ |
10228 |
++ unsigned int __data = __VAR; \ |
10229 |
++ if (__CONV) \ |
10230 |
++ __data = jiffies_to_msecs(__data); \ |
10231 |
++ return bfq_var_show(__data, (page)); \ |
10232 |
++} |
10233 |
++SHOW_FUNCTION(bfq_quantum_show, bfqd->bfq_quantum, 0); |
10234 |
++SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 1); |
10235 |
++SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 1); |
10236 |
++SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0); |
10237 |
++SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0); |
10238 |
++SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 1); |
10239 |
++SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0); |
10240 |
++SHOW_FUNCTION(bfq_max_budget_async_rq_show, bfqd->bfq_max_budget_async_rq, 0); |
10241 |
++SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout[BLK_RW_SYNC], 1); |
10242 |
++SHOW_FUNCTION(bfq_timeout_async_show, bfqd->bfq_timeout[BLK_RW_ASYNC], 1); |
10243 |
++SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0); |
10244 |
++SHOW_FUNCTION(bfq_raising_coeff_show, bfqd->bfq_raising_coeff, 0); |
10245 |
++SHOW_FUNCTION(bfq_raising_rt_max_time_show, bfqd->bfq_raising_rt_max_time, 1); |
10246 |
++SHOW_FUNCTION(bfq_raising_min_idle_time_show, bfqd->bfq_raising_min_idle_time, |
10247 |
++ 1); |
10248 |
++SHOW_FUNCTION(bfq_raising_min_inter_arr_async_show, |
10249 |
++ bfqd->bfq_raising_min_inter_arr_async, |
10250 |
++ 1); |
10251 |
++SHOW_FUNCTION(bfq_raising_max_softrt_rate_show, |
10252 |
++ bfqd->bfq_raising_max_softrt_rate, 0); |
10253 |
++#undef SHOW_FUNCTION |
10254 |
++ |
10255 |
++#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \ |
10256 |
++static ssize_t \ |
10257 |
++__FUNC(struct elevator_queue *e, const char *page, size_t count) \ |
10258 |
++{ \ |
10259 |
++ struct bfq_data *bfqd = e->elevator_data; \ |
10260 |
++ unsigned long uninitialized_var(__data); \ |
10261 |
++ int ret = bfq_var_store(&__data, (page), count); \ |
10262 |
++ if (__data < (MIN)) \ |
10263 |
++ __data = (MIN); \ |
10264 |
++ else if (__data > (MAX)) \ |
10265 |
++ __data = (MAX); \ |
10266 |
++ if (__CONV) \ |
10267 |
++ *(__PTR) = msecs_to_jiffies(__data); \ |
10268 |
++ else \ |
10269 |
++ *(__PTR) = __data; \ |
10270 |
++ return ret; \ |
10271 |
++} |
10272 |
++STORE_FUNCTION(bfq_quantum_store, &bfqd->bfq_quantum, 1, INT_MAX, 0); |
10273 |
++STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1, |
10274 |
++ INT_MAX, 1); |
10275 |
++STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1, |
10276 |
++ INT_MAX, 1); |
10277 |
++STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0); |
10278 |
++STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1, |
10279 |
++ INT_MAX, 0); |
10280 |
++STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 1); |
10281 |
++STORE_FUNCTION(bfq_max_budget_async_rq_store, &bfqd->bfq_max_budget_async_rq, |
10282 |
++ 1, INT_MAX, 0); |
10283 |
++STORE_FUNCTION(bfq_timeout_async_store, &bfqd->bfq_timeout[BLK_RW_ASYNC], 0, |
10284 |
++ INT_MAX, 1); |
10285 |
++STORE_FUNCTION(bfq_raising_coeff_store, &bfqd->bfq_raising_coeff, 1, |
10286 |
++ INT_MAX, 0); |
10287 |
++STORE_FUNCTION(bfq_raising_max_time_store, &bfqd->bfq_raising_max_time, 0, |
10288 |
++ INT_MAX, 1); |
10289 |
++STORE_FUNCTION(bfq_raising_rt_max_time_store, &bfqd->bfq_raising_rt_max_time, 0, |
10290 |
++ INT_MAX, 1); |
10291 |
++STORE_FUNCTION(bfq_raising_min_idle_time_store, |
10292 |
++ &bfqd->bfq_raising_min_idle_time, 0, INT_MAX, 1); |
10293 |
++STORE_FUNCTION(bfq_raising_min_inter_arr_async_store, |
10294 |
++ &bfqd->bfq_raising_min_inter_arr_async, 0, INT_MAX, 1); |
10295 |
++STORE_FUNCTION(bfq_raising_max_softrt_rate_store, |
10296 |
++ &bfqd->bfq_raising_max_softrt_rate, 0, INT_MAX, 0); |
10297 |
++#undef STORE_FUNCTION |
10298 |
++ |
10299 |
++/* do nothing for the moment */ |
10300 |
++static ssize_t bfq_weights_store(struct elevator_queue *e, |
10301 |
++ const char *page, size_t count) |
10302 |
++{ |
10303 |
++ return count; |
10304 |
++} |
10305 |
++ |
10306 |
++static inline unsigned long bfq_estimated_max_budget(struct bfq_data *bfqd) |
10307 |
++{ |
10308 |
++ u64 timeout = jiffies_to_msecs(bfqd->bfq_timeout[BLK_RW_SYNC]); |
10309 |
++ |
10310 |
++ if (bfqd->peak_rate_samples >= BFQ_PEAK_RATE_SAMPLES) |
10311 |
++ return bfq_calc_max_budget(bfqd->peak_rate, timeout); |
10312 |
++ else |
10313 |
++ return bfq_default_max_budget; |
10314 |
++} |
10315 |
++ |
10316 |
++static ssize_t bfq_max_budget_store(struct elevator_queue *e, |
10317 |
++ const char *page, size_t count) |
10318 |
++{ |
10319 |
++ struct bfq_data *bfqd = e->elevator_data; |
10320 |
++ unsigned long uninitialized_var(__data); |
10321 |
++ int ret = bfq_var_store(&__data, (page), count); |
10322 |
++ |
10323 |
++ if (__data == 0) |
10324 |
++ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
10325 |
++ else { |
10326 |
++ if (__data > INT_MAX) |
10327 |
++ __data = INT_MAX; |
10328 |
++ bfqd->bfq_max_budget = __data; |
10329 |
++ } |
10330 |
++ |
10331 |
++ bfqd->bfq_user_max_budget = __data; |
10332 |
++ |
10333 |
++ return ret; |
10334 |
++} |
10335 |
++ |
10336 |
++static ssize_t bfq_timeout_sync_store(struct elevator_queue *e, |
10337 |
++ const char *page, size_t count) |
10338 |
++{ |
10339 |
++ struct bfq_data *bfqd = e->elevator_data; |
10340 |
++ unsigned long uninitialized_var(__data); |
10341 |
++ int ret = bfq_var_store(&__data, (page), count); |
10342 |
++ |
10343 |
++ if (__data < 1) |
10344 |
++ __data = 1; |
10345 |
++ else if (__data > INT_MAX) |
10346 |
++ __data = INT_MAX; |
10347 |
++ |
10348 |
++ bfqd->bfq_timeout[BLK_RW_SYNC] = msecs_to_jiffies(__data); |
10349 |
++ if (bfqd->bfq_user_max_budget == 0) |
10350 |
++ bfqd->bfq_max_budget = bfq_estimated_max_budget(bfqd); |
10351 |
++ |
10352 |
++ return ret; |
10353 |
++} |
10354 |
++ |
10355 |
++static ssize_t bfq_low_latency_store(struct elevator_queue *e, |
10356 |
++ const char *page, size_t count) |
10357 |
++{ |
10358 |
++ struct bfq_data *bfqd = e->elevator_data; |
10359 |
++ unsigned long uninitialized_var(__data); |
10360 |
++ int ret = bfq_var_store(&__data, (page), count); |
10361 |
++ |
10362 |
++ if (__data > 1) |
10363 |
++ __data = 1; |
10364 |
++ if (__data == 0 && bfqd->low_latency != 0) |
10365 |
++ bfq_end_raising(bfqd); |
10366 |
++ bfqd->low_latency = __data; |
10367 |
++ |
10368 |
++ return ret; |
10369 |
++} |
10370 |
++ |
10371 |
++#define BFQ_ATTR(name) \ |
10372 |
++ __ATTR(name, S_IRUGO|S_IWUSR, bfq_##name##_show, bfq_##name##_store) |
10373 |
++ |
10374 |
++static struct elv_fs_entry bfq_attrs[] = { |
10375 |
++ BFQ_ATTR(quantum), |
10376 |
++ BFQ_ATTR(fifo_expire_sync), |
10377 |
++ BFQ_ATTR(fifo_expire_async), |
10378 |
++ BFQ_ATTR(back_seek_max), |
10379 |
++ BFQ_ATTR(back_seek_penalty), |
10380 |
++ BFQ_ATTR(slice_idle), |
10381 |
++ BFQ_ATTR(max_budget), |
10382 |
++ BFQ_ATTR(max_budget_async_rq), |
10383 |
++ BFQ_ATTR(timeout_sync), |
10384 |
++ BFQ_ATTR(timeout_async), |
10385 |
++ BFQ_ATTR(low_latency), |
10386 |
++ BFQ_ATTR(raising_coeff), |
10387 |
++ BFQ_ATTR(raising_max_time), |
10388 |
++ BFQ_ATTR(raising_rt_max_time), |
10389 |
++ BFQ_ATTR(raising_min_idle_time), |
10390 |
++ BFQ_ATTR(raising_min_inter_arr_async), |
10391 |
++ BFQ_ATTR(raising_max_softrt_rate), |
10392 |
++ BFQ_ATTR(weights), |
10393 |
++ __ATTR_NULL |
10394 |
++}; |
10395 |
++ |
10396 |
++static struct elevator_type iosched_bfq = { |
10397 |
++ .ops = { |
10398 |
++ .elevator_merge_fn = bfq_merge, |
10399 |
++ .elevator_merged_fn = bfq_merged_request, |
10400 |
++ .elevator_merge_req_fn = bfq_merged_requests, |
10401 |
++ .elevator_allow_merge_fn = bfq_allow_merge, |
10402 |
++ .elevator_dispatch_fn = bfq_dispatch_requests, |
10403 |
++ .elevator_add_req_fn = bfq_insert_request, |
10404 |
++ .elevator_activate_req_fn = bfq_activate_request, |
10405 |
++ .elevator_deactivate_req_fn = bfq_deactivate_request, |
10406 |
++ .elevator_completed_req_fn = bfq_completed_request, |
10407 |
++ .elevator_former_req_fn = elv_rb_former_request, |
10408 |
++ .elevator_latter_req_fn = elv_rb_latter_request, |
10409 |
++ .elevator_init_icq_fn = bfq_init_icq, |
10410 |
++ .elevator_exit_icq_fn = bfq_exit_icq, |
10411 |
++ .elevator_set_req_fn = bfq_set_request, |
10412 |
++ .elevator_put_req_fn = bfq_put_request, |
10413 |
++ .elevator_may_queue_fn = bfq_may_queue, |
10414 |
++ .elevator_init_fn = bfq_init_queue, |
10415 |
++ .elevator_exit_fn = bfq_exit_queue, |
10416 |
++ }, |
10417 |
++ .icq_size = sizeof(struct bfq_io_cq), |
10418 |
++ .icq_align = __alignof__(struct bfq_io_cq), |
10419 |
++ .elevator_attrs = bfq_attrs, |
10420 |
++ .elevator_name = "bfq", |
10421 |
++ .elevator_owner = THIS_MODULE, |
10422 |
++}; |
10423 |
++ |
10424 |
++static int __init bfq_init(void) |
10425 |
++{ |
10426 |
++ /* |
10427 |
++ * Can be 0 on HZ < 1000 setups. |
10428 |
++ */ |
10429 |
++ if (bfq_slice_idle == 0) |
10430 |
++ bfq_slice_idle = 1; |
10431 |
++ |
10432 |
++ if (bfq_timeout_async == 0) |
10433 |
++ bfq_timeout_async = 1; |
10434 |
++ |
10435 |
++ if (bfq_slab_setup()) |
10436 |
++ return -ENOMEM; |
10437 |
++ |
10438 |
++ elv_register(&iosched_bfq); |
10439 |
++ pr_info("BFQ I/O-scheduler version: v7r2"); |
10440 |
++ |
10441 |
++ return 0; |
10442 |
++} |
10443 |
++ |
10444 |
++static void __exit bfq_exit(void) |
10445 |
++{ |
10446 |
++ elv_unregister(&iosched_bfq); |
10447 |
++ bfq_slab_kill(); |
10448 |
++} |
10449 |
++ |
10450 |
++module_init(bfq_init); |
10451 |
++module_exit(bfq_exit); |
10452 |
++ |
10453 |
++MODULE_AUTHOR("Fabio Checconi, Paolo Valente"); |
10454 |
+diff --git a/block/bfq-sched.c b/block/bfq-sched.c |
10455 |
+new file mode 100644 |
10456 |
+index 0000000..999b475 |
10457 |
+--- /dev/null |
10458 |
++++ b/block/bfq-sched.c |
10459 |
+@@ -0,0 +1,1078 @@ |
10460 |
++/* |
10461 |
++ * BFQ: Hierarchical B-WF2Q+ scheduler. |
10462 |
++ * |
10463 |
++ * Based on ideas and code from CFQ: |
10464 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
10465 |
++ * |
10466 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
10467 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
10468 |
++ * |
10469 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
10470 |
++ */ |
10471 |
++ |
10472 |
++#ifdef CONFIG_CGROUP_BFQIO |
10473 |
++#define for_each_entity(entity) \ |
10474 |
++ for (; entity != NULL; entity = entity->parent) |
10475 |
++ |
10476 |
++#define for_each_entity_safe(entity, parent) \ |
10477 |
++ for (; entity && ({ parent = entity->parent; 1; }); entity = parent) |
10478 |
++ |
10479 |
++static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
10480 |
++ int extract, |
10481 |
++ struct bfq_data *bfqd); |
10482 |
++ |
10483 |
++static inline void bfq_update_budget(struct bfq_entity *next_in_service) |
10484 |
++{ |
10485 |
++ struct bfq_entity *bfqg_entity; |
10486 |
++ struct bfq_group *bfqg; |
10487 |
++ struct bfq_sched_data *group_sd; |
10488 |
++ |
10489 |
++ BUG_ON(next_in_service == NULL); |
10490 |
++ |
10491 |
++ group_sd = next_in_service->sched_data; |
10492 |
++ |
10493 |
++ bfqg = container_of(group_sd, struct bfq_group, sched_data); |
10494 |
++ /* |
10495 |
++ * bfq_group's my_entity field is not NULL only if the group |
10496 |
++ * is not the root group. We must not touch the root entity |
10497 |
++ * as it must never become an in-service entity. |
10498 |
++ */ |
10499 |
++ bfqg_entity = bfqg->my_entity; |
10500 |
++ if (bfqg_entity != NULL) |
10501 |
++ bfqg_entity->budget = next_in_service->budget; |
10502 |
++} |
10503 |
++ |
10504 |
++static int bfq_update_next_in_service(struct bfq_sched_data *sd) |
10505 |
++{ |
10506 |
++ struct bfq_entity *next_in_service; |
10507 |
++ |
10508 |
++ if (sd->in_service_entity != NULL) |
10509 |
++ /* will update/requeue at the end of service */ |
10510 |
++ return 0; |
10511 |
++ |
10512 |
++ /* |
10513 |
++ * NOTE: this can be improved in many ways, such as returning |
10514 |
++ * 1 (and thus propagating upwards the update) only when the |
10515 |
++ * budget changes, or caching the bfqq that will be scheduled |
10516 |
++ * next from this subtree. By now we worry more about |
10517 |
++ * correctness than about performance... |
10518 |
++ */ |
10519 |
++ next_in_service = bfq_lookup_next_entity(sd, 0, NULL); |
10520 |
++ sd->next_in_service = next_in_service; |
10521 |
++ |
10522 |
++ if (next_in_service != NULL) |
10523 |
++ bfq_update_budget(next_in_service); |
10524 |
++ |
10525 |
++ return 1; |
10526 |
++} |
10527 |
++ |
10528 |
++static inline void bfq_check_next_in_service(struct bfq_sched_data *sd, |
10529 |
++ struct bfq_entity *entity) |
10530 |
++{ |
10531 |
++ BUG_ON(sd->next_in_service != entity); |
10532 |
++} |
10533 |
++#else |
10534 |
++#define for_each_entity(entity) \ |
10535 |
++ for (; entity != NULL; entity = NULL) |
10536 |
++ |
10537 |
++#define for_each_entity_safe(entity, parent) \ |
10538 |
++ for (parent = NULL; entity != NULL; entity = parent) |
10539 |
++ |
10540 |
++static inline int bfq_update_next_in_service(struct bfq_sched_data *sd) |
10541 |
++{ |
10542 |
++ return 0; |
10543 |
++} |
10544 |
++ |
10545 |
++static inline void bfq_check_next_in_service(struct bfq_sched_data *sd, |
10546 |
++ struct bfq_entity *entity) |
10547 |
++{ |
10548 |
++} |
10549 |
++ |
10550 |
++static inline void bfq_update_budget(struct bfq_entity *next_in_service) |
10551 |
++{ |
10552 |
++} |
10553 |
++#endif |
10554 |
++ |
10555 |
++/* |
10556 |
++ * Shift for timestamp calculations. This actually limits the maximum |
10557 |
++ * service allowed in one timestamp delta (small shift values increase it), |
10558 |
++ * the maximum total weight that can be used for the queues in the system |
10559 |
++ * (big shift values increase it), and the period of virtual time wraparounds. |
10560 |
++ */ |
10561 |
++#define WFQ_SERVICE_SHIFT 22 |
10562 |
++ |
10563 |
++/** |
10564 |
++ * bfq_gt - compare two timestamps. |
10565 |
++ * @a: first ts. |
10566 |
++ * @b: second ts. |
10567 |
++ * |
10568 |
++ * Return @a > @b, dealing with wrapping correctly. |
10569 |
++ */ |
10570 |
++static inline int bfq_gt(u64 a, u64 b) |
10571 |
++{ |
10572 |
++ return (s64)(a - b) > 0; |
10573 |
++} |
10574 |
++ |
10575 |
++static inline struct bfq_queue *bfq_entity_to_bfqq(struct bfq_entity *entity) |
10576 |
++{ |
10577 |
++ struct bfq_queue *bfqq = NULL; |
10578 |
++ |
10579 |
++ BUG_ON(entity == NULL); |
10580 |
++ |
10581 |
++ if (entity->my_sched_data == NULL) |
10582 |
++ bfqq = container_of(entity, struct bfq_queue, entity); |
10583 |
++ |
10584 |
++ return bfqq; |
10585 |
++} |
10586 |
++ |
10587 |
++ |
10588 |
++/** |
10589 |
++ * bfq_delta - map service into the virtual time domain. |
10590 |
++ * @service: amount of service. |
10591 |
++ * @weight: scale factor (weight of an entity or weight sum). |
10592 |
++ */ |
10593 |
++static inline u64 bfq_delta(unsigned long service, |
10594 |
++ unsigned long weight) |
10595 |
++{ |
10596 |
++ u64 d = (u64)service << WFQ_SERVICE_SHIFT; |
10597 |
++ |
10598 |
++ do_div(d, weight); |
10599 |
++ return d; |
10600 |
++} |
10601 |
++ |
10602 |
++/** |
10603 |
++ * bfq_calc_finish - assign the finish time to an entity. |
10604 |
++ * @entity: the entity to act upon. |
10605 |
++ * @service: the service to be charged to the entity. |
10606 |
++ */ |
10607 |
++static inline void bfq_calc_finish(struct bfq_entity *entity, |
10608 |
++ unsigned long service) |
10609 |
++{ |
10610 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10611 |
++ |
10612 |
++ BUG_ON(entity->weight == 0); |
10613 |
++ |
10614 |
++ entity->finish = entity->start + |
10615 |
++ bfq_delta(service, entity->weight); |
10616 |
++ |
10617 |
++ if (bfqq != NULL) { |
10618 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, |
10619 |
++ "calc_finish: serv %lu, w %d", |
10620 |
++ service, entity->weight); |
10621 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, |
10622 |
++ "calc_finish: start %llu, finish %llu, delta %llu", |
10623 |
++ entity->start, entity->finish, |
10624 |
++ bfq_delta(service, entity->weight)); |
10625 |
++ } |
10626 |
++} |
10627 |
++ |
10628 |
++/** |
10629 |
++ * bfq_entity_of - get an entity from a node. |
10630 |
++ * @node: the node field of the entity. |
10631 |
++ * |
10632 |
++ * Convert a node pointer to the relative entity. This is used only |
10633 |
++ * to simplify the logic of some functions and not as the generic |
10634 |
++ * conversion mechanism because, e.g., in the tree walking functions, |
10635 |
++ * the check for a %NULL value would be redundant. |
10636 |
++ */ |
10637 |
++static inline struct bfq_entity *bfq_entity_of(struct rb_node *node) |
10638 |
++{ |
10639 |
++ struct bfq_entity *entity = NULL; |
10640 |
++ |
10641 |
++ if (node != NULL) |
10642 |
++ entity = rb_entry(node, struct bfq_entity, rb_node); |
10643 |
++ |
10644 |
++ return entity; |
10645 |
++} |
10646 |
++ |
10647 |
++/** |
10648 |
++ * bfq_extract - remove an entity from a tree. |
10649 |
++ * @root: the tree root. |
10650 |
++ * @entity: the entity to remove. |
10651 |
++ */ |
10652 |
++static inline void bfq_extract(struct rb_root *root, |
10653 |
++ struct bfq_entity *entity) |
10654 |
++{ |
10655 |
++ BUG_ON(entity->tree != root); |
10656 |
++ |
10657 |
++ entity->tree = NULL; |
10658 |
++ rb_erase(&entity->rb_node, root); |
10659 |
++} |
10660 |
++ |
10661 |
++/** |
10662 |
++ * bfq_idle_extract - extract an entity from the idle tree. |
10663 |
++ * @st: the service tree of the owning @entity. |
10664 |
++ * @entity: the entity being removed. |
10665 |
++ */ |
10666 |
++static void bfq_idle_extract(struct bfq_service_tree *st, |
10667 |
++ struct bfq_entity *entity) |
10668 |
++{ |
10669 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10670 |
++ struct rb_node *next; |
10671 |
++ |
10672 |
++ BUG_ON(entity->tree != &st->idle); |
10673 |
++ |
10674 |
++ if (entity == st->first_idle) { |
10675 |
++ next = rb_next(&entity->rb_node); |
10676 |
++ st->first_idle = bfq_entity_of(next); |
10677 |
++ } |
10678 |
++ |
10679 |
++ if (entity == st->last_idle) { |
10680 |
++ next = rb_prev(&entity->rb_node); |
10681 |
++ st->last_idle = bfq_entity_of(next); |
10682 |
++ } |
10683 |
++ |
10684 |
++ bfq_extract(&st->idle, entity); |
10685 |
++ |
10686 |
++ if (bfqq != NULL) |
10687 |
++ list_del(&bfqq->bfqq_list); |
10688 |
++} |
10689 |
++ |
10690 |
++/** |
10691 |
++ * bfq_insert - generic tree insertion. |
10692 |
++ * @root: tree root. |
10693 |
++ * @entity: entity to insert. |
10694 |
++ * |
10695 |
++ * This is used for the idle and the active tree, since they are both |
10696 |
++ * ordered by finish time. |
10697 |
++ */ |
10698 |
++static void bfq_insert(struct rb_root *root, struct bfq_entity *entity) |
10699 |
++{ |
10700 |
++ struct bfq_entity *entry; |
10701 |
++ struct rb_node **node = &root->rb_node; |
10702 |
++ struct rb_node *parent = NULL; |
10703 |
++ |
10704 |
++ BUG_ON(entity->tree != NULL); |
10705 |
++ |
10706 |
++ while (*node != NULL) { |
10707 |
++ parent = *node; |
10708 |
++ entry = rb_entry(parent, struct bfq_entity, rb_node); |
10709 |
++ |
10710 |
++ if (bfq_gt(entry->finish, entity->finish)) |
10711 |
++ node = &parent->rb_left; |
10712 |
++ else |
10713 |
++ node = &parent->rb_right; |
10714 |
++ } |
10715 |
++ |
10716 |
++ rb_link_node(&entity->rb_node, parent, node); |
10717 |
++ rb_insert_color(&entity->rb_node, root); |
10718 |
++ |
10719 |
++ entity->tree = root; |
10720 |
++} |
10721 |
++ |
10722 |
++/** |
10723 |
++ * bfq_update_min - update the min_start field of a entity. |
10724 |
++ * @entity: the entity to update. |
10725 |
++ * @node: one of its children. |
10726 |
++ * |
10727 |
++ * This function is called when @entity may store an invalid value for |
10728 |
++ * min_start due to updates to the active tree. The function assumes |
10729 |
++ * that the subtree rooted at @node (which may be its left or its right |
10730 |
++ * child) has a valid min_start value. |
10731 |
++ */ |
10732 |
++static inline void bfq_update_min(struct bfq_entity *entity, |
10733 |
++ struct rb_node *node) |
10734 |
++{ |
10735 |
++ struct bfq_entity *child; |
10736 |
++ |
10737 |
++ if (node != NULL) { |
10738 |
++ child = rb_entry(node, struct bfq_entity, rb_node); |
10739 |
++ if (bfq_gt(entity->min_start, child->min_start)) |
10740 |
++ entity->min_start = child->min_start; |
10741 |
++ } |
10742 |
++} |
10743 |
++ |
10744 |
++/** |
10745 |
++ * bfq_update_active_node - recalculate min_start. |
10746 |
++ * @node: the node to update. |
10747 |
++ * |
10748 |
++ * @node may have changed position or one of its children may have moved, |
10749 |
++ * this function updates its min_start value. The left and right subtrees |
10750 |
++ * are assumed to hold a correct min_start value. |
10751 |
++ */ |
10752 |
++static inline void bfq_update_active_node(struct rb_node *node) |
10753 |
++{ |
10754 |
++ struct bfq_entity *entity = rb_entry(node, struct bfq_entity, rb_node); |
10755 |
++ |
10756 |
++ entity->min_start = entity->start; |
10757 |
++ bfq_update_min(entity, node->rb_right); |
10758 |
++ bfq_update_min(entity, node->rb_left); |
10759 |
++} |
10760 |
++ |
10761 |
++/** |
10762 |
++ * bfq_update_active_tree - update min_start for the whole active tree. |
10763 |
++ * @node: the starting node. |
10764 |
++ * |
10765 |
++ * @node must be the deepest modified node after an update. This function |
10766 |
++ * updates its min_start using the values held by its children, assuming |
10767 |
++ * that they did not change, and then updates all the nodes that may have |
10768 |
++ * changed in the path to the root. The only nodes that may have changed |
10769 |
++ * are the ones in the path or their siblings. |
10770 |
++ */ |
10771 |
++static void bfq_update_active_tree(struct rb_node *node) |
10772 |
++{ |
10773 |
++ struct rb_node *parent; |
10774 |
++ |
10775 |
++up: |
10776 |
++ bfq_update_active_node(node); |
10777 |
++ |
10778 |
++ parent = rb_parent(node); |
10779 |
++ if (parent == NULL) |
10780 |
++ return; |
10781 |
++ |
10782 |
++ if (node == parent->rb_left && parent->rb_right != NULL) |
10783 |
++ bfq_update_active_node(parent->rb_right); |
10784 |
++ else if (parent->rb_left != NULL) |
10785 |
++ bfq_update_active_node(parent->rb_left); |
10786 |
++ |
10787 |
++ node = parent; |
10788 |
++ goto up; |
10789 |
++} |
10790 |
++ |
10791 |
++/** |
10792 |
++ * bfq_active_insert - insert an entity in the active tree of its group/device. |
10793 |
++ * @st: the service tree of the entity. |
10794 |
++ * @entity: the entity being inserted. |
10795 |
++ * |
10796 |
++ * The active tree is ordered by finish time, but an extra key is kept |
10797 |
++ * per each node, containing the minimum value for the start times of |
10798 |
++ * its children (and the node itself), so it's possible to search for |
10799 |
++ * the eligible node with the lowest finish time in logarithmic time. |
10800 |
++ */ |
10801 |
++static void bfq_active_insert(struct bfq_service_tree *st, |
10802 |
++ struct bfq_entity *entity) |
10803 |
++{ |
10804 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10805 |
++ struct rb_node *node = &entity->rb_node; |
10806 |
++ |
10807 |
++ bfq_insert(&st->active, entity); |
10808 |
++ |
10809 |
++ if (node->rb_left != NULL) |
10810 |
++ node = node->rb_left; |
10811 |
++ else if (node->rb_right != NULL) |
10812 |
++ node = node->rb_right; |
10813 |
++ |
10814 |
++ bfq_update_active_tree(node); |
10815 |
++ |
10816 |
++ if (bfqq != NULL) |
10817 |
++ list_add(&bfqq->bfqq_list, &bfqq->bfqd->active_list); |
10818 |
++} |
10819 |
++ |
10820 |
++/** |
10821 |
++ * bfq_ioprio_to_weight - calc a weight from an ioprio. |
10822 |
++ * @ioprio: the ioprio value to convert. |
10823 |
++ */ |
10824 |
++static unsigned short bfq_ioprio_to_weight(int ioprio) |
10825 |
++{ |
10826 |
++ WARN_ON(ioprio < 0 || ioprio >= IOPRIO_BE_NR); |
10827 |
++ return IOPRIO_BE_NR - ioprio; |
10828 |
++} |
10829 |
++ |
10830 |
++/** |
10831 |
++ * bfq_weight_to_ioprio - calc an ioprio from a weight. |
10832 |
++ * @weight: the weight value to convert. |
10833 |
++ * |
10834 |
++ * To preserve as mush as possible the old only-ioprio user interface, |
10835 |
++ * 0 is used as an escape ioprio value for weights (numerically) equal or |
10836 |
++ * larger than IOPRIO_BE_NR |
10837 |
++ */ |
10838 |
++static unsigned short bfq_weight_to_ioprio(int weight) |
10839 |
++{ |
10840 |
++ WARN_ON(weight < BFQ_MIN_WEIGHT || weight > BFQ_MAX_WEIGHT); |
10841 |
++ return IOPRIO_BE_NR - weight < 0 ? 0 : IOPRIO_BE_NR - weight; |
10842 |
++} |
10843 |
++ |
10844 |
++static inline void bfq_get_entity(struct bfq_entity *entity) |
10845 |
++{ |
10846 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10847 |
++ struct bfq_sched_data *sd; |
10848 |
++ |
10849 |
++ if (bfqq != NULL) { |
10850 |
++ sd = entity->sched_data; |
10851 |
++ atomic_inc(&bfqq->ref); |
10852 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "get_entity: %p %d", |
10853 |
++ bfqq, atomic_read(&bfqq->ref)); |
10854 |
++ } |
10855 |
++} |
10856 |
++ |
10857 |
++/** |
10858 |
++ * bfq_find_deepest - find the deepest node that an extraction can modify. |
10859 |
++ * @node: the node being removed. |
10860 |
++ * |
10861 |
++ * Do the first step of an extraction in an rb tree, looking for the |
10862 |
++ * node that will replace @node, and returning the deepest node that |
10863 |
++ * the following modifications to the tree can touch. If @node is the |
10864 |
++ * last node in the tree return %NULL. |
10865 |
++ */ |
10866 |
++static struct rb_node *bfq_find_deepest(struct rb_node *node) |
10867 |
++{ |
10868 |
++ struct rb_node *deepest; |
10869 |
++ |
10870 |
++ if (node->rb_right == NULL && node->rb_left == NULL) |
10871 |
++ deepest = rb_parent(node); |
10872 |
++ else if (node->rb_right == NULL) |
10873 |
++ deepest = node->rb_left; |
10874 |
++ else if (node->rb_left == NULL) |
10875 |
++ deepest = node->rb_right; |
10876 |
++ else { |
10877 |
++ deepest = rb_next(node); |
10878 |
++ if (deepest->rb_right != NULL) |
10879 |
++ deepest = deepest->rb_right; |
10880 |
++ else if (rb_parent(deepest) != node) |
10881 |
++ deepest = rb_parent(deepest); |
10882 |
++ } |
10883 |
++ |
10884 |
++ return deepest; |
10885 |
++} |
10886 |
++ |
10887 |
++/** |
10888 |
++ * bfq_active_extract - remove an entity from the active tree. |
10889 |
++ * @st: the service_tree containing the tree. |
10890 |
++ * @entity: the entity being removed. |
10891 |
++ */ |
10892 |
++static void bfq_active_extract(struct bfq_service_tree *st, |
10893 |
++ struct bfq_entity *entity) |
10894 |
++{ |
10895 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10896 |
++ struct rb_node *node; |
10897 |
++ |
10898 |
++ node = bfq_find_deepest(&entity->rb_node); |
10899 |
++ bfq_extract(&st->active, entity); |
10900 |
++ |
10901 |
++ if (node != NULL) |
10902 |
++ bfq_update_active_tree(node); |
10903 |
++ |
10904 |
++ if (bfqq != NULL) |
10905 |
++ list_del(&bfqq->bfqq_list); |
10906 |
++} |
10907 |
++ |
10908 |
++/** |
10909 |
++ * bfq_idle_insert - insert an entity into the idle tree. |
10910 |
++ * @st: the service tree containing the tree. |
10911 |
++ * @entity: the entity to insert. |
10912 |
++ */ |
10913 |
++static void bfq_idle_insert(struct bfq_service_tree *st, |
10914 |
++ struct bfq_entity *entity) |
10915 |
++{ |
10916 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10917 |
++ struct bfq_entity *first_idle = st->first_idle; |
10918 |
++ struct bfq_entity *last_idle = st->last_idle; |
10919 |
++ |
10920 |
++ if (first_idle == NULL || bfq_gt(first_idle->finish, entity->finish)) |
10921 |
++ st->first_idle = entity; |
10922 |
++ if (last_idle == NULL || bfq_gt(entity->finish, last_idle->finish)) |
10923 |
++ st->last_idle = entity; |
10924 |
++ |
10925 |
++ bfq_insert(&st->idle, entity); |
10926 |
++ |
10927 |
++ if (bfqq != NULL) |
10928 |
++ list_add(&bfqq->bfqq_list, &bfqq->bfqd->idle_list); |
10929 |
++} |
10930 |
++ |
10931 |
++/** |
10932 |
++ * bfq_forget_entity - remove an entity from the wfq trees. |
10933 |
++ * @st: the service tree. |
10934 |
++ * @entity: the entity being removed. |
10935 |
++ * |
10936 |
++ * Update the device status and forget everything about @entity, putting |
10937 |
++ * the device reference to it, if it is a queue. Entities belonging to |
10938 |
++ * groups are not refcounted. |
10939 |
++ */ |
10940 |
++static void bfq_forget_entity(struct bfq_service_tree *st, |
10941 |
++ struct bfq_entity *entity) |
10942 |
++{ |
10943 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
10944 |
++ struct bfq_sched_data *sd; |
10945 |
++ |
10946 |
++ BUG_ON(!entity->on_st); |
10947 |
++ |
10948 |
++ entity->on_st = 0; |
10949 |
++ st->wsum -= entity->weight; |
10950 |
++ if (bfqq != NULL) { |
10951 |
++ sd = entity->sched_data; |
10952 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "forget_entity: %p %d", |
10953 |
++ bfqq, atomic_read(&bfqq->ref)); |
10954 |
++ bfq_put_queue(bfqq); |
10955 |
++ } |
10956 |
++} |
10957 |
++ |
10958 |
++/** |
10959 |
++ * bfq_put_idle_entity - release the idle tree ref of an entity. |
10960 |
++ * @st: service tree for the entity. |
10961 |
++ * @entity: the entity being released. |
10962 |
++ */ |
10963 |
++static void bfq_put_idle_entity(struct bfq_service_tree *st, |
10964 |
++ struct bfq_entity *entity) |
10965 |
++{ |
10966 |
++ bfq_idle_extract(st, entity); |
10967 |
++ bfq_forget_entity(st, entity); |
10968 |
++} |
10969 |
++ |
10970 |
++/** |
10971 |
++ * bfq_forget_idle - update the idle tree if necessary. |
10972 |
++ * @st: the service tree to act upon. |
10973 |
++ * |
10974 |
++ * To preserve the global O(log N) complexity we only remove one entry here; |
10975 |
++ * as the idle tree will not grow indefinitely this can be done safely. |
10976 |
++ */ |
10977 |
++static void bfq_forget_idle(struct bfq_service_tree *st) |
10978 |
++{ |
10979 |
++ struct bfq_entity *first_idle = st->first_idle; |
10980 |
++ struct bfq_entity *last_idle = st->last_idle; |
10981 |
++ |
10982 |
++ if (RB_EMPTY_ROOT(&st->active) && last_idle != NULL && |
10983 |
++ !bfq_gt(last_idle->finish, st->vtime)) { |
10984 |
++ /* |
10985 |
++ * Forget the whole idle tree, increasing the vtime past |
10986 |
++ * the last finish time of idle entities. |
10987 |
++ */ |
10988 |
++ st->vtime = last_idle->finish; |
10989 |
++ } |
10990 |
++ |
10991 |
++ if (first_idle != NULL && !bfq_gt(first_idle->finish, st->vtime)) |
10992 |
++ bfq_put_idle_entity(st, first_idle); |
10993 |
++} |
10994 |
++ |
10995 |
++static struct bfq_service_tree * |
10996 |
++__bfq_entity_update_weight_prio(struct bfq_service_tree *old_st, |
10997 |
++ struct bfq_entity *entity) |
10998 |
++{ |
10999 |
++ struct bfq_service_tree *new_st = old_st; |
11000 |
++ |
11001 |
++ if (entity->ioprio_changed) { |
11002 |
++ struct bfq_queue *bfqq = bfq_entity_to_bfqq(entity); |
11003 |
++ |
11004 |
++ BUG_ON(old_st->wsum < entity->weight); |
11005 |
++ old_st->wsum -= entity->weight; |
11006 |
++ |
11007 |
++ if (entity->new_weight != entity->orig_weight) { |
11008 |
++ entity->orig_weight = entity->new_weight; |
11009 |
++ entity->ioprio = |
11010 |
++ bfq_weight_to_ioprio(entity->orig_weight); |
11011 |
++ } else if (entity->new_ioprio != entity->ioprio) { |
11012 |
++ entity->ioprio = entity->new_ioprio; |
11013 |
++ entity->orig_weight = |
11014 |
++ bfq_ioprio_to_weight(entity->ioprio); |
11015 |
++ } else |
11016 |
++ entity->new_weight = entity->orig_weight = |
11017 |
++ bfq_ioprio_to_weight(entity->ioprio); |
11018 |
++ |
11019 |
++ entity->ioprio_class = entity->new_ioprio_class; |
11020 |
++ entity->ioprio_changed = 0; |
11021 |
++ |
11022 |
++ /* |
11023 |
++ * NOTE: here we may be changing the weight too early, |
11024 |
++ * this will cause unfairness. The correct approach |
11025 |
++ * would have required additional complexity to defer |
11026 |
++ * weight changes to the proper time instants (i.e., |
11027 |
++ * when entity->finish <= old_st->vtime). |
11028 |
++ */ |
11029 |
++ new_st = bfq_entity_service_tree(entity); |
11030 |
++ entity->weight = entity->orig_weight * |
11031 |
++ (bfqq != NULL ? bfqq->raising_coeff : 1); |
11032 |
++ new_st->wsum += entity->weight; |
11033 |
++ |
11034 |
++ if (new_st != old_st) |
11035 |
++ entity->start = new_st->vtime; |
11036 |
++ } |
11037 |
++ |
11038 |
++ return new_st; |
11039 |
++} |
11040 |
++ |
11041 |
++/** |
11042 |
++ * bfq_bfqq_served - update the scheduler status after selection for service. |
11043 |
++ * @bfqq: the queue being served. |
11044 |
++ * @served: bytes to transfer. |
11045 |
++ * |
11046 |
++ * NOTE: this can be optimized, as the timestamps of upper level entities |
11047 |
++ * are synchronized every time a new bfqq is selected for service. By now, |
11048 |
++ * we keep it to better check consistency. |
11049 |
++ */ |
11050 |
++static void bfq_bfqq_served(struct bfq_queue *bfqq, unsigned long served) |
11051 |
++{ |
11052 |
++ struct bfq_entity *entity = &bfqq->entity; |
11053 |
++ struct bfq_service_tree *st; |
11054 |
++ |
11055 |
++ for_each_entity(entity) { |
11056 |
++ st = bfq_entity_service_tree(entity); |
11057 |
++ |
11058 |
++ entity->service += served; |
11059 |
++ BUG_ON(entity->service > entity->budget); |
11060 |
++ BUG_ON(st->wsum == 0); |
11061 |
++ |
11062 |
++ st->vtime += bfq_delta(served, st->wsum); |
11063 |
++ bfq_forget_idle(st); |
11064 |
++ } |
11065 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "bfqq_served %lu secs", served); |
11066 |
++} |
11067 |
++ |
11068 |
++/** |
11069 |
++ * bfq_bfqq_charge_full_budget - set the service to the entity budget. |
11070 |
++ * @bfqq: the queue that needs a service update. |
11071 |
++ * |
11072 |
++ * When it's not possible to be fair in the service domain, because |
11073 |
++ * a queue is not consuming its budget fast enough (the meaning of |
11074 |
++ * fast depends on the timeout parameter), we charge it a full |
11075 |
++ * budget. In this way we should obtain a sort of time-domain |
11076 |
++ * fairness among all the seeky/slow queues. |
11077 |
++ */ |
11078 |
++static inline void bfq_bfqq_charge_full_budget(struct bfq_queue *bfqq) |
11079 |
++{ |
11080 |
++ struct bfq_entity *entity = &bfqq->entity; |
11081 |
++ |
11082 |
++ bfq_log_bfqq(bfqq->bfqd, bfqq, "charge_full_budget"); |
11083 |
++ |
11084 |
++ bfq_bfqq_served(bfqq, entity->budget - entity->service); |
11085 |
++} |
11086 |
++ |
11087 |
++/** |
11088 |
++ * __bfq_activate_entity - activate an entity. |
11089 |
++ * @entity: the entity being activated. |
11090 |
++ * |
11091 |
++ * Called whenever an entity is activated, i.e., it is not active and one |
11092 |
++ * of its children receives a new request, or has to be reactivated due to |
11093 |
++ * budget exhaustion. It uses the current budget of the entity (and the |
11094 |
++ * service received if @entity is active) of the queue to calculate its |
11095 |
++ * timestamps. |
11096 |
++ */ |
11097 |
++static void __bfq_activate_entity(struct bfq_entity *entity) |
11098 |
++{ |
11099 |
++ struct bfq_sched_data *sd = entity->sched_data; |
11100 |
++ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
11101 |
++ |
11102 |
++ if (entity == sd->in_service_entity) { |
11103 |
++ BUG_ON(entity->tree != NULL); |
11104 |
++ /* |
11105 |
++ * If we are requeueing the current entity we have |
11106 |
++ * to take care of not charging to it service it has |
11107 |
++ * not received. |
11108 |
++ */ |
11109 |
++ bfq_calc_finish(entity, entity->service); |
11110 |
++ entity->start = entity->finish; |
11111 |
++ sd->in_service_entity = NULL; |
11112 |
++ } else if (entity->tree == &st->active) { |
11113 |
++ /* |
11114 |
++ * Requeueing an entity due to a change of some |
11115 |
++ * next_in_service entity below it. We reuse the |
11116 |
++ * old start time. |
11117 |
++ */ |
11118 |
++ bfq_active_extract(st, entity); |
11119 |
++ } else if (entity->tree == &st->idle) { |
11120 |
++ /* |
11121 |
++ * Must be on the idle tree, bfq_idle_extract() will |
11122 |
++ * check for that. |
11123 |
++ */ |
11124 |
++ bfq_idle_extract(st, entity); |
11125 |
++ entity->start = bfq_gt(st->vtime, entity->finish) ? |
11126 |
++ st->vtime : entity->finish; |
11127 |
++ } else { |
11128 |
++ /* |
11129 |
++ * The finish time of the entity may be invalid, and |
11130 |
++ * it is in the past for sure, otherwise the queue |
11131 |
++ * would have been on the idle tree. |
11132 |
++ */ |
11133 |
++ entity->start = st->vtime; |
11134 |
++ st->wsum += entity->weight; |
11135 |
++ bfq_get_entity(entity); |
11136 |
++ |
11137 |
++ BUG_ON(entity->on_st); |
11138 |
++ entity->on_st = 1; |
11139 |
++ } |
11140 |
++ |
11141 |
++ st = __bfq_entity_update_weight_prio(st, entity); |
11142 |
++ bfq_calc_finish(entity, entity->budget); |
11143 |
++ bfq_active_insert(st, entity); |
11144 |
++} |
11145 |
++ |
11146 |
++/** |
11147 |
++ * bfq_activate_entity - activate an entity and its ancestors if necessary. |
11148 |
++ * @entity: the entity to activate. |
11149 |
++ * |
11150 |
++ * Activate @entity and all the entities on the path from it to the root. |
11151 |
++ */ |
11152 |
++static void bfq_activate_entity(struct bfq_entity *entity) |
11153 |
++{ |
11154 |
++ struct bfq_sched_data *sd; |
11155 |
++ |
11156 |
++ for_each_entity(entity) { |
11157 |
++ __bfq_activate_entity(entity); |
11158 |
++ |
11159 |
++ sd = entity->sched_data; |
11160 |
++ if (!bfq_update_next_in_service(sd)) |
11161 |
++ /* |
11162 |
++ * No need to propagate the activation to the |
11163 |
++ * upper entities, as they will be updated when |
11164 |
++ * the in-service entity is rescheduled. |
11165 |
++ */ |
11166 |
++ break; |
11167 |
++ } |
11168 |
++} |
11169 |
++ |
11170 |
++/** |
11171 |
++ * __bfq_deactivate_entity - deactivate an entity from its service tree. |
11172 |
++ * @entity: the entity to deactivate. |
11173 |
++ * @requeue: if false, the entity will not be put into the idle tree. |
11174 |
++ * |
11175 |
++ * Deactivate an entity, independently from its previous state. If the |
11176 |
++ * entity was not on a service tree just return, otherwise if it is on |
11177 |
++ * any scheduler tree, extract it from that tree, and if necessary |
11178 |
++ * and if the caller did not specify @requeue, put it on the idle tree. |
11179 |
++ * |
11180 |
++ * Return %1 if the caller should update the entity hierarchy, i.e., |
11181 |
++ * if the entity was under service or if it was the next_in_service for |
11182 |
++ * its sched_data; return %0 otherwise. |
11183 |
++ */ |
11184 |
++static int __bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
11185 |
++{ |
11186 |
++ struct bfq_sched_data *sd = entity->sched_data; |
11187 |
++ struct bfq_service_tree *st = bfq_entity_service_tree(entity); |
11188 |
++ int was_in_service = entity == sd->in_service_entity; |
11189 |
++ int ret = 0; |
11190 |
++ |
11191 |
++ if (!entity->on_st) |
11192 |
++ return 0; |
11193 |
++ |
11194 |
++ BUG_ON(was_in_service && entity->tree != NULL); |
11195 |
++ |
11196 |
++ if (was_in_service) { |
11197 |
++ bfq_calc_finish(entity, entity->service); |
11198 |
++ sd->in_service_entity = NULL; |
11199 |
++ } else if (entity->tree == &st->active) |
11200 |
++ bfq_active_extract(st, entity); |
11201 |
++ else if (entity->tree == &st->idle) |
11202 |
++ bfq_idle_extract(st, entity); |
11203 |
++ else if (entity->tree != NULL) |
11204 |
++ BUG(); |
11205 |
++ |
11206 |
++ if (was_in_service || sd->next_in_service == entity) |
11207 |
++ ret = bfq_update_next_in_service(sd); |
11208 |
++ |
11209 |
++ if (!requeue || !bfq_gt(entity->finish, st->vtime)) |
11210 |
++ bfq_forget_entity(st, entity); |
11211 |
++ else |
11212 |
++ bfq_idle_insert(st, entity); |
11213 |
++ |
11214 |
++ BUG_ON(sd->in_service_entity == entity); |
11215 |
++ BUG_ON(sd->next_in_service == entity); |
11216 |
++ |
11217 |
++ return ret; |
11218 |
++} |
11219 |
++ |
11220 |
++/** |
11221 |
++ * bfq_deactivate_entity - deactivate an entity. |
11222 |
++ * @entity: the entity to deactivate. |
11223 |
++ * @requeue: true if the entity can be put on the idle tree |
11224 |
++ */ |
11225 |
++static void bfq_deactivate_entity(struct bfq_entity *entity, int requeue) |
11226 |
++{ |
11227 |
++ struct bfq_sched_data *sd; |
11228 |
++ struct bfq_entity *parent; |
11229 |
++ |
11230 |
++ for_each_entity_safe(entity, parent) { |
11231 |
++ sd = entity->sched_data; |
11232 |
++ |
11233 |
++ if (!__bfq_deactivate_entity(entity, requeue)) |
11234 |
++ /* |
11235 |
++ * The parent entity is still backlogged, and |
11236 |
++ * we don't need to update it as it is still |
11237 |
++ * under service. |
11238 |
++ */ |
11239 |
++ break; |
11240 |
++ |
11241 |
++ if (sd->next_in_service != NULL) |
11242 |
++ /* |
11243 |
++ * The parent entity is still backlogged and |
11244 |
++ * the budgets on the path towards the root |
11245 |
++ * need to be updated. |
11246 |
++ */ |
11247 |
++ goto update; |
11248 |
++ |
11249 |
++ /* |
11250 |
++ * If we reach there the parent is no more backlogged and |
11251 |
++ * we want to propagate the dequeue upwards. |
11252 |
++ */ |
11253 |
++ requeue = 1; |
11254 |
++ } |
11255 |
++ |
11256 |
++ return; |
11257 |
++ |
11258 |
++update: |
11259 |
++ entity = parent; |
11260 |
++ for_each_entity(entity) { |
11261 |
++ __bfq_activate_entity(entity); |
11262 |
++ |
11263 |
++ sd = entity->sched_data; |
11264 |
++ if (!bfq_update_next_in_service(sd)) |
11265 |
++ break; |
11266 |
++ } |
11267 |
++} |
11268 |
++ |
11269 |
++/** |
11270 |
++ * bfq_update_vtime - update vtime if necessary. |
11271 |
++ * @st: the service tree to act upon. |
11272 |
++ * |
11273 |
++ * If necessary update the service tree vtime to have at least one |
11274 |
++ * eligible entity, skipping to its start time. Assumes that the |
11275 |
++ * active tree of the device is not empty. |
11276 |
++ * |
11277 |
++ * NOTE: this hierarchical implementation updates vtimes quite often, |
11278 |
++ * we may end up with reactivated tasks getting timestamps after a |
11279 |
++ * vtime skip done because we needed a ->first_active entity on some |
11280 |
++ * intermediate node. |
11281 |
++ */ |
11282 |
++static void bfq_update_vtime(struct bfq_service_tree *st) |
11283 |
++{ |
11284 |
++ struct bfq_entity *entry; |
11285 |
++ struct rb_node *node = st->active.rb_node; |
11286 |
++ |
11287 |
++ entry = rb_entry(node, struct bfq_entity, rb_node); |
11288 |
++ if (bfq_gt(entry->min_start, st->vtime)) { |
11289 |
++ st->vtime = entry->min_start; |
11290 |
++ bfq_forget_idle(st); |
11291 |
++ } |
11292 |
++} |
11293 |
++ |
11294 |
++/** |
11295 |
++ * bfq_first_active_entity - find the eligible entity with |
11296 |
++ * the smallest finish time |
11297 |
++ * @st: the service tree to select from. |
11298 |
++ * |
11299 |
++ * This function searches the first schedulable entity, starting from the |
11300 |
++ * root of the tree and going on the left every time on this side there is |
11301 |
++ * a subtree with at least one eligible (start >= vtime) entity. The path |
11302 |
++ * on the right is followed only if a) the left subtree contains no eligible |
11303 |
++ * entities and b) no eligible entity has been found yet. |
11304 |
++ */ |
11305 |
++static struct bfq_entity *bfq_first_active_entity(struct bfq_service_tree *st) |
11306 |
++{ |
11307 |
++ struct bfq_entity *entry, *first = NULL; |
11308 |
++ struct rb_node *node = st->active.rb_node; |
11309 |
++ |
11310 |
++ while (node != NULL) { |
11311 |
++ entry = rb_entry(node, struct bfq_entity, rb_node); |
11312 |
++left: |
11313 |
++ if (!bfq_gt(entry->start, st->vtime)) |
11314 |
++ first = entry; |
11315 |
++ |
11316 |
++ BUG_ON(bfq_gt(entry->min_start, st->vtime)); |
11317 |
++ |
11318 |
++ if (node->rb_left != NULL) { |
11319 |
++ entry = rb_entry(node->rb_left, |
11320 |
++ struct bfq_entity, rb_node); |
11321 |
++ if (!bfq_gt(entry->min_start, st->vtime)) { |
11322 |
++ node = node->rb_left; |
11323 |
++ goto left; |
11324 |
++ } |
11325 |
++ } |
11326 |
++ if (first != NULL) |
11327 |
++ break; |
11328 |
++ node = node->rb_right; |
11329 |
++ } |
11330 |
++ |
11331 |
++ BUG_ON(first == NULL && !RB_EMPTY_ROOT(&st->active)); |
11332 |
++ return first; |
11333 |
++} |
11334 |
++ |
11335 |
++/** |
11336 |
++ * __bfq_lookup_next_entity - return the first eligible entity in @st. |
11337 |
++ * @st: the service tree. |
11338 |
++ * |
11339 |
++ * Update the virtual time in @st and return the first eligible entity |
11340 |
++ * it contains. |
11341 |
++ */ |
11342 |
++static struct bfq_entity *__bfq_lookup_next_entity(struct bfq_service_tree *st, |
11343 |
++ bool force) |
11344 |
++{ |
11345 |
++ struct bfq_entity *entity, *new_next_in_service = NULL; |
11346 |
++ |
11347 |
++ if (RB_EMPTY_ROOT(&st->active)) |
11348 |
++ return NULL; |
11349 |
++ |
11350 |
++ bfq_update_vtime(st); |
11351 |
++ entity = bfq_first_active_entity(st); |
11352 |
++ BUG_ON(bfq_gt(entity->start, st->vtime)); |
11353 |
++ |
11354 |
++ /* |
11355 |
++ * If the chosen entity does not match with the sched_data's |
11356 |
++ * next_in_service and we are forcedly serving the IDLE priority |
11357 |
++ * class tree, bubble up budget update. |
11358 |
++ */ |
11359 |
++ if (unlikely(force && entity != entity->sched_data->next_in_service)) { |
11360 |
++ new_next_in_service = entity; |
11361 |
++ for_each_entity(new_next_in_service) |
11362 |
++ bfq_update_budget(new_next_in_service); |
11363 |
++ } |
11364 |
++ |
11365 |
++ return entity; |
11366 |
++} |
11367 |
++ |
11368 |
++/** |
11369 |
++ * bfq_lookup_next_entity - return the first eligible entity in @sd. |
11370 |
++ * @sd: the sched_data. |
11371 |
++ * @extract: if true the returned entity will be also extracted from @sd. |
11372 |
++ * |
11373 |
++ * NOTE: since we cache the next_in_service entity at each level of the |
11374 |
++ * hierarchy, the complexity of the lookup can be decreased with |
11375 |
++ * absolutely no effort just returning the cached next_in_service value; |
11376 |
++ * we prefer to do full lookups to test the consistency of * the data |
11377 |
++ * structures. |
11378 |
++ */ |
11379 |
++static struct bfq_entity *bfq_lookup_next_entity(struct bfq_sched_data *sd, |
11380 |
++ int extract, |
11381 |
++ struct bfq_data *bfqd) |
11382 |
++{ |
11383 |
++ struct bfq_service_tree *st = sd->service_tree; |
11384 |
++ struct bfq_entity *entity; |
11385 |
++ int i = 0; |
11386 |
++ |
11387 |
++ BUG_ON(sd->in_service_entity != NULL); |
11388 |
++ |
11389 |
++ if (bfqd != NULL && |
11390 |
++ jiffies - bfqd->bfq_class_idle_last_service > BFQ_CL_IDLE_TIMEOUT) { |
11391 |
++ entity = __bfq_lookup_next_entity(st + BFQ_IOPRIO_CLASSES - 1, |
11392 |
++ true); |
11393 |
++ if (entity != NULL) { |
11394 |
++ i = BFQ_IOPRIO_CLASSES - 1; |
11395 |
++ bfqd->bfq_class_idle_last_service = jiffies; |
11396 |
++ sd->next_in_service = entity; |
11397 |
++ } |
11398 |
++ } |
11399 |
++ for (; i < BFQ_IOPRIO_CLASSES; i++) { |
11400 |
++ entity = __bfq_lookup_next_entity(st + i, false); |
11401 |
++ if (entity != NULL) { |
11402 |
++ if (extract) { |
11403 |
++ bfq_check_next_in_service(sd, entity); |
11404 |
++ bfq_active_extract(st + i, entity); |
11405 |
++ sd->in_service_entity = entity; |
11406 |
++ sd->next_in_service = NULL; |
11407 |
++ } |
11408 |
++ break; |
11409 |
++ } |
11410 |
++ } |
11411 |
++ |
11412 |
++ return entity; |
11413 |
++} |
11414 |
++ |
11415 |
++/* |
11416 |
++ * Get next queue for service. |
11417 |
++ */ |
11418 |
++static struct bfq_queue *bfq_get_next_queue(struct bfq_data *bfqd) |
11419 |
++{ |
11420 |
++ struct bfq_entity *entity = NULL; |
11421 |
++ struct bfq_sched_data *sd; |
11422 |
++ struct bfq_queue *bfqq; |
11423 |
++ |
11424 |
++ BUG_ON(bfqd->in_service_queue != NULL); |
11425 |
++ |
11426 |
++ if (bfqd->busy_queues == 0) |
11427 |
++ return NULL; |
11428 |
++ |
11429 |
++ sd = &bfqd->root_group->sched_data; |
11430 |
++ for (; sd != NULL; sd = entity->my_sched_data) { |
11431 |
++ entity = bfq_lookup_next_entity(sd, 1, bfqd); |
11432 |
++ BUG_ON(entity == NULL); |
11433 |
++ entity->service = 0; |
11434 |
++ } |
11435 |
++ |
11436 |
++ bfqq = bfq_entity_to_bfqq(entity); |
11437 |
++ BUG_ON(bfqq == NULL); |
11438 |
++ |
11439 |
++ return bfqq; |
11440 |
++} |
11441 |
++ |
11442 |
++/* |
11443 |
++ * Forced extraction of the given queue. |
11444 |
++ */ |
11445 |
++static void bfq_get_next_queue_forced(struct bfq_data *bfqd, |
11446 |
++ struct bfq_queue *bfqq) |
11447 |
++{ |
11448 |
++ struct bfq_entity *entity; |
11449 |
++ struct bfq_sched_data *sd; |
11450 |
++ |
11451 |
++ BUG_ON(bfqd->in_service_queue != NULL); |
11452 |
++ |
11453 |
++ entity = &bfqq->entity; |
11454 |
++ /* |
11455 |
++ * Bubble up extraction/update from the leaf to the root. |
11456 |
++ */ |
11457 |
++ for_each_entity(entity) { |
11458 |
++ sd = entity->sched_data; |
11459 |
++ bfq_update_budget(entity); |
11460 |
++ bfq_update_vtime(bfq_entity_service_tree(entity)); |
11461 |
++ bfq_active_extract(bfq_entity_service_tree(entity), entity); |
11462 |
++ sd->active_entity = entity; |
11463 |
++ sd->next_active = NULL; |
11464 |
++ entity->service = 0; |
11465 |
++ } |
11466 |
++ |
11467 |
++ return; |
11468 |
++} |
11469 |
++ |
11470 |
++static void __bfq_bfqd_reset_in_service(struct bfq_data *bfqd) |
11471 |
++{ |
11472 |
++ if (bfqd->in_service_bic != NULL) { |
11473 |
++ put_io_context(bfqd->in_service_bic->icq.ioc); |
11474 |
++ bfqd->in_service_bic = NULL; |
11475 |
++ } |
11476 |
++ |
11477 |
++ bfqd->in_service_queue = NULL; |
11478 |
++ del_timer(&bfqd->idle_slice_timer); |
11479 |
++} |
11480 |
++ |
11481 |
++static void bfq_deactivate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
11482 |
++ int requeue) |
11483 |
++{ |
11484 |
++ struct bfq_entity *entity = &bfqq->entity; |
11485 |
++ |
11486 |
++ if (bfqq == bfqd->in_service_queue) |
11487 |
++ __bfq_bfqd_reset_in_service(bfqd); |
11488 |
++ |
11489 |
++ bfq_deactivate_entity(entity, requeue); |
11490 |
++} |
11491 |
++ |
11492 |
++static void bfq_activate_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
11493 |
++{ |
11494 |
++ struct bfq_entity *entity = &bfqq->entity; |
11495 |
++ |
11496 |
++ bfq_activate_entity(entity); |
11497 |
++} |
11498 |
++ |
11499 |
++/* |
11500 |
++ * Called when the bfqq no longer has requests pending, remove it from |
11501 |
++ * the service tree. |
11502 |
++ */ |
11503 |
++static void bfq_del_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq, |
11504 |
++ int requeue) |
11505 |
++{ |
11506 |
++ BUG_ON(!bfq_bfqq_busy(bfqq)); |
11507 |
++ BUG_ON(!RB_EMPTY_ROOT(&bfqq->sort_list)); |
11508 |
++ |
11509 |
++ bfq_log_bfqq(bfqd, bfqq, "del from busy"); |
11510 |
++ |
11511 |
++ bfq_clear_bfqq_busy(bfqq); |
11512 |
++ |
11513 |
++ BUG_ON(bfqd->busy_queues == 0); |
11514 |
++ bfqd->busy_queues--; |
11515 |
++ if (bfqq->raising_coeff > 1) |
11516 |
++ bfqd->raised_busy_queues--; |
11517 |
++ |
11518 |
++ bfq_deactivate_bfqq(bfqd, bfqq, requeue); |
11519 |
++} |
11520 |
++ |
11521 |
++/* |
11522 |
++ * Called when an inactive queue receives a new request. |
11523 |
++ */ |
11524 |
++static void bfq_add_bfqq_busy(struct bfq_data *bfqd, struct bfq_queue *bfqq) |
11525 |
++{ |
11526 |
++ BUG_ON(bfq_bfqq_busy(bfqq)); |
11527 |
++ BUG_ON(bfqq == bfqd->in_service_queue); |
11528 |
++ |
11529 |
++ bfq_log_bfqq(bfqd, bfqq, "add to busy"); |
11530 |
++ |
11531 |
++ bfq_activate_bfqq(bfqd, bfqq); |
11532 |
++ |
11533 |
++ bfq_mark_bfqq_busy(bfqq); |
11534 |
++ bfqd->busy_queues++; |
11535 |
++ if (bfqq->raising_coeff > 1) |
11536 |
++ bfqd->raised_busy_queues++; |
11537 |
++} |
11538 |
+diff --git a/block/bfq.h b/block/bfq.h |
11539 |
+new file mode 100644 |
11540 |
+index 0000000..3ca8482 |
11541 |
+--- /dev/null |
11542 |
++++ b/block/bfq.h |
11543 |
+@@ -0,0 +1,622 @@ |
11544 |
++/* |
11545 |
++ * BFQ-v7r2 for 3.14.0: data structures and common functions prototypes. |
11546 |
++ * |
11547 |
++ * Based on ideas and code from CFQ: |
11548 |
++ * Copyright (C) 2003 Jens Axboe <axboe@××××××.dk> |
11549 |
++ * |
11550 |
++ * Copyright (C) 2008 Fabio Checconi <fabio@×××××××××××××.it> |
11551 |
++ * Paolo Valente <paolo.valente@×××××××.it> |
11552 |
++ * |
11553 |
++ * Copyright (C) 2010 Paolo Valente <paolo.valente@×××××××.it> |
11554 |
++ */ |
11555 |
++ |
11556 |
++#ifndef _BFQ_H |
11557 |
++#define _BFQ_H |
11558 |
++ |
11559 |
++#include <linux/blktrace_api.h> |
11560 |
++#include <linux/hrtimer.h> |
11561 |
++#include <linux/ioprio.h> |
11562 |
++#include <linux/rbtree.h> |
11563 |
++ |
11564 |
++#define BFQ_IOPRIO_CLASSES 3 |
11565 |
++#define BFQ_CL_IDLE_TIMEOUT (HZ/5) |
11566 |
++ |
11567 |
++#define BFQ_MIN_WEIGHT 1 |
11568 |
++#define BFQ_MAX_WEIGHT 1000 |
11569 |
++ |
11570 |
++#define BFQ_DEFAULT_GRP_WEIGHT 10 |
11571 |
++#define BFQ_DEFAULT_GRP_IOPRIO 0 |
11572 |
++#define BFQ_DEFAULT_GRP_CLASS IOPRIO_CLASS_BE |
11573 |
++ |
11574 |
++struct bfq_entity; |
11575 |
++ |
11576 |
++/** |
11577 |
++ * struct bfq_service_tree - per ioprio_class service tree. |
11578 |
++ * @active: tree for active entities (i.e., those backlogged). |
11579 |
++ * @idle: tree for idle entities (i.e., those not backlogged, with V <= F_i). |
11580 |
++ * @first_idle: idle entity with minimum F_i. |
11581 |
++ * @last_idle: idle entity with maximum F_i. |
11582 |
++ * @vtime: scheduler virtual time. |
11583 |
++ * @wsum: scheduler weight sum; active and idle entities contribute to it. |
11584 |
++ * |
11585 |
++ * Each service tree represents a B-WF2Q+ scheduler on its own. Each |
11586 |
++ * ioprio_class has its own independent scheduler, and so its own |
11587 |
++ * bfq_service_tree. All the fields are protected by the queue lock |
11588 |
++ * of the containing bfqd. |
11589 |
++ */ |
11590 |
++struct bfq_service_tree { |
11591 |
++ struct rb_root active; |
11592 |
++ struct rb_root idle; |
11593 |
++ |
11594 |
++ struct bfq_entity *first_idle; |
11595 |
++ struct bfq_entity *last_idle; |
11596 |
++ |
11597 |
++ u64 vtime; |
11598 |
++ unsigned long wsum; |
11599 |
++}; |
11600 |
++ |
11601 |
++/** |
11602 |
++ * struct bfq_sched_data - multi-class scheduler. |
11603 |
++ * @in_service_entity: entity under service. |
11604 |
++ * @next_in_service: head-of-the-line entity in the scheduler. |
11605 |
++ * @service_tree: array of service trees, one per ioprio_class. |
11606 |
++ * |
11607 |
++ * bfq_sched_data is the basic scheduler queue. It supports three |
11608 |
++ * ioprio_classes, and can be used either as a toplevel queue or as |
11609 |
++ * an intermediate queue on a hierarchical setup. |
11610 |
++ * @next_in_service points to the active entity of the sched_data |
11611 |
++ * service trees that will be scheduled next. |
11612 |
++ * |
11613 |
++ * The supported ioprio_classes are the same as in CFQ, in descending |
11614 |
++ * priority order, IOPRIO_CLASS_RT, IOPRIO_CLASS_BE, IOPRIO_CLASS_IDLE. |
11615 |
++ * Requests from higher priority queues are served before all the |
11616 |
++ * requests from lower priority queues; among requests of the same |
11617 |
++ * queue requests are served according to B-WF2Q+. |
11618 |
++ * All the fields are protected by the queue lock of the containing bfqd. |
11619 |
++ */ |
11620 |
++struct bfq_sched_data { |
11621 |
++ struct bfq_entity *in_service_entity; |
11622 |
++ struct bfq_entity *next_in_service; |
11623 |
++ struct bfq_service_tree service_tree[BFQ_IOPRIO_CLASSES]; |
11624 |
++}; |
11625 |
++ |
11626 |
++/** |
11627 |
++ * struct bfq_entity - schedulable entity. |
11628 |
++ * @rb_node: service_tree member. |
11629 |
++ * @on_st: flag, true if the entity is on a tree (either the active or |
11630 |
++ * the idle one of its service_tree). |
11631 |
++ * @finish: B-WF2Q+ finish timestamp (aka F_i). |
11632 |
++ * @start: B-WF2Q+ start timestamp (aka S_i). |
11633 |
++ * @tree: tree the entity is enqueued into; %NULL if not on a tree. |
11634 |
++ * @min_start: minimum start time of the (active) subtree rooted at |
11635 |
++ * this entity; used for O(log N) lookups into active trees. |
11636 |
++ * @service: service received during the last round of service. |
11637 |
++ * @budget: budget used to calculate F_i; F_i = S_i + @budget / @weight. |
11638 |
++ * @weight: weight of the queue |
11639 |
++ * @parent: parent entity, for hierarchical scheduling. |
11640 |
++ * @my_sched_data: for non-leaf nodes in the cgroup hierarchy, the |
11641 |
++ * associated scheduler queue, %NULL on leaf nodes. |
11642 |
++ * @sched_data: the scheduler queue this entity belongs to. |
11643 |
++ * @ioprio: the ioprio in use. |
11644 |
++ * @new_weight: when a weight change is requested, the new weight value. |
11645 |
++ * @orig_weight: original weight, used to implement weight boosting |
11646 |
++ * @new_ioprio: when an ioprio change is requested, the new ioprio value. |
11647 |
++ * @ioprio_class: the ioprio_class in use. |
11648 |
++ * @new_ioprio_class: when an ioprio_class change is requested, the new |
11649 |
++ * ioprio_class value. |
11650 |
++ * @ioprio_changed: flag, true when the user requested a weight, ioprio or |
11651 |
++ * ioprio_class change. |
11652 |
++ * |
11653 |
++ * A bfq_entity is used to represent either a bfq_queue (leaf node in the |
11654 |
++ * cgroup hierarchy) or a bfq_group into the upper level scheduler. Each |
11655 |
++ * entity belongs to the sched_data of the parent group in the cgroup |
11656 |
++ * hierarchy. Non-leaf entities have also their own sched_data, stored |
11657 |
++ * in @my_sched_data. |
11658 |
++ * |
11659 |
++ * Each entity stores independently its priority values; this would |
11660 |
++ * allow different weights on different devices, but this |
11661 |
++ * functionality is not exported to userspace by now. Priorities and |
11662 |
++ * weights are updated lazily, first storing the new values into the |
11663 |
++ * new_* fields, then setting the @ioprio_changed flag. As soon as |
11664 |
++ * there is a transition in the entity state that allows the priority |
11665 |
++ * update to take place the effective and the requested priority |
11666 |
++ * values are synchronized. |
11667 |
++ * |
11668 |
++ * Unless cgroups are used, the weight value is calculated from the |
11669 |
++ * ioprio to export the same interface as CFQ. When dealing with |
11670 |
++ * ``well-behaved'' queues (i.e., queues that do not spend too much |
11671 |
++ * time to consume their budget and have true sequential behavior, and |
11672 |
++ * when there are no external factors breaking anticipation) the |
11673 |
++ * relative weights at each level of the cgroups hierarchy should be |
11674 |
++ * guaranteed. All the fields are protected by the queue lock of the |
11675 |
++ * containing bfqd. |
11676 |
++ */ |
11677 |
++struct bfq_entity { |
11678 |
++ struct rb_node rb_node; |
11679 |
++ |
11680 |
++ int on_st; |
11681 |
++ |
11682 |
++ u64 finish; |
11683 |
++ u64 start; |
11684 |
++ |
11685 |
++ struct rb_root *tree; |
11686 |
++ |
11687 |
++ u64 min_start; |
11688 |
++ |
11689 |
++ unsigned long service, budget; |
11690 |
++ unsigned short weight, new_weight; |
11691 |
++ unsigned short orig_weight; |
11692 |
++ |
11693 |
++ struct bfq_entity *parent; |
11694 |
++ |
11695 |
++ struct bfq_sched_data *my_sched_data; |
11696 |
++ struct bfq_sched_data *sched_data; |
11697 |
++ |
11698 |
++ unsigned short ioprio, new_ioprio; |
11699 |
++ unsigned short ioprio_class, new_ioprio_class; |
11700 |
++ |
11701 |
++ int ioprio_changed; |
11702 |
++}; |
11703 |
++ |
11704 |
++struct bfq_group; |
11705 |
++ |
11706 |
++/** |
11707 |
++ * struct bfq_queue - leaf schedulable entity. |
11708 |
++ * @ref: reference counter. |
11709 |
++ * @bfqd: parent bfq_data. |
11710 |
++ * @new_bfqq: shared bfq_queue if queue is cooperating with |
11711 |
++ * one or more other queues. |
11712 |
++ * @pos_node: request-position tree member (see bfq_data's @rq_pos_tree). |
11713 |
++ * @pos_root: request-position tree root (see bfq_data's @rq_pos_tree). |
11714 |
++ * @sort_list: sorted list of pending requests. |
11715 |
++ * @next_rq: if fifo isn't expired, next request to serve. |
11716 |
++ * @queued: nr of requests queued in @sort_list. |
11717 |
++ * @allocated: currently allocated requests. |
11718 |
++ * @meta_pending: pending metadata requests. |
11719 |
++ * @fifo: fifo list of requests in sort_list. |
11720 |
++ * @entity: entity representing this queue in the scheduler. |
11721 |
++ * @max_budget: maximum budget allowed from the feedback mechanism. |
11722 |
++ * @budget_timeout: budget expiration (in jiffies). |
11723 |
++ * @dispatched: number of requests on the dispatch list or inside driver. |
11724 |
++ * @org_ioprio: saved ioprio during boosted periods. |
11725 |
++ * @flags: status flags. |
11726 |
++ * @bfqq_list: node for active/idle bfqq list inside our bfqd. |
11727 |
++ * @seek_samples: number of seeks sampled |
11728 |
++ * @seek_total: sum of the distances of the seeks sampled |
11729 |
++ * @seek_mean: mean seek distance |
11730 |
++ * @last_request_pos: position of the last request enqueued |
11731 |
++ * @pid: pid of the process owning the queue, used for logging purposes. |
11732 |
++ * @last_rais_start_finish: start time of the current weight-raising period if |
11733 |
++ * the @bfq-queue is being weight-raised, otherwise |
11734 |
++ * finish time of the last weight-raising period |
11735 |
++ * @raising_cur_max_time: current max raising time for this queue |
11736 |
++ * @soft_rt_next_start: minimum time instant such that, only if a new request |
11737 |
++ * is enqueued after this time instant in an idle |
11738 |
++ * @bfq_queue with no outstanding requests, then the |
11739 |
++ * task associated with the queue it is deemed as soft |
11740 |
++ * real-time (see the comments to the function |
11741 |
++ * bfq_bfqq_softrt_next_start()) |
11742 |
++ * @last_idle_bklogged: time of the last transition of the @bfq_queue from |
11743 |
++ * idle to backlogged |
11744 |
++ * @service_from_backlogged: cumulative service received from the @bfq_queue |
11745 |
++ * since the last transition from idle to backlogged |
11746 |
++ * |
11747 |
++ * A bfq_queue is a leaf request queue; it can be associated with an io_context |
11748 |
++ * or more, if it is async or shared between cooperating processes. @cgroup |
11749 |
++ * holds a reference to the cgroup, to be sure that it does not disappear while |
11750 |
++ * a bfqq still references it (mostly to avoid races between request issuing and |
11751 |
++ * task migration followed by cgroup destruction). |
11752 |
++ * All the fields are protected by the queue lock of the containing bfqd. |
11753 |
++ */ |
11754 |
++struct bfq_queue { |
11755 |
++ atomic_t ref; |
11756 |
++ struct bfq_data *bfqd; |
11757 |
++ |
11758 |
++ /* fields for cooperating queues handling */ |
11759 |
++ struct bfq_queue *new_bfqq; |
11760 |
++ struct rb_node pos_node; |
11761 |
++ struct rb_root *pos_root; |
11762 |
++ |
11763 |
++ struct rb_root sort_list; |
11764 |
++ struct request *next_rq; |
11765 |
++ int queued[2]; |
11766 |
++ int allocated[2]; |
11767 |
++ int meta_pending; |
11768 |
++ struct list_head fifo; |
11769 |
++ |
11770 |
++ struct bfq_entity entity; |
11771 |
++ |
11772 |
++ unsigned long max_budget; |
11773 |
++ unsigned long budget_timeout; |
11774 |
++ |
11775 |
++ int dispatched; |
11776 |
++ |
11777 |
++ unsigned short org_ioprio; |
11778 |
++ |
11779 |
++ unsigned int flags; |
11780 |
++ |
11781 |
++ struct list_head bfqq_list; |
11782 |
++ |
11783 |
++ unsigned int seek_samples; |
11784 |
++ u64 seek_total; |
11785 |
++ sector_t seek_mean; |
11786 |
++ sector_t last_request_pos; |
11787 |
++ |
11788 |
++ pid_t pid; |
11789 |
++ |
11790 |
++ /* weight-raising fields */ |
11791 |
++ unsigned long raising_cur_max_time; |
11792 |
++ unsigned long soft_rt_next_start; |
11793 |
++ unsigned long last_rais_start_finish; |
11794 |
++ unsigned int raising_coeff; |
11795 |
++ unsigned long last_idle_bklogged; |
11796 |
++ unsigned long service_from_backlogged; |
11797 |
++}; |
11798 |
++ |
11799 |
++/** |
11800 |
++ * struct bfq_ttime - per process thinktime stats. |
11801 |
++ * @ttime_total: total process thinktime |
11802 |
++ * @ttime_samples: number of thinktime samples |
11803 |
++ * @ttime_mean: average process thinktime |
11804 |
++ */ |
11805 |
++struct bfq_ttime { |
11806 |
++ unsigned long last_end_request; |
11807 |
++ |
11808 |
++ unsigned long ttime_total; |
11809 |
++ unsigned long ttime_samples; |
11810 |
++ unsigned long ttime_mean; |
11811 |
++}; |
11812 |
++ |
11813 |
++/** |
11814 |
++ * struct bfq_io_cq - per (request_queue, io_context) structure. |
11815 |
++ * @icq: associated io_cq structure |
11816 |
++ * @bfqq: array of two process queues, the sync and the async |
11817 |
++ * @ttime: associated @bfq_ttime struct |
11818 |
++ */ |
11819 |
++struct bfq_io_cq { |
11820 |
++ struct io_cq icq; /* must be the first member */ |
11821 |
++ struct bfq_queue *bfqq[2]; |
11822 |
++ struct bfq_ttime ttime; |
11823 |
++ int ioprio; |
11824 |
++}; |
11825 |
++ |
11826 |
++/** |
11827 |
++ * struct bfq_data - per device data structure. |
11828 |
++ * @queue: request queue for the managed device. |
11829 |
++ * @root_group: root bfq_group for the device. |
11830 |
++ * @rq_pos_tree: rbtree sorted by next_request position, |
11831 |
++ * used when determining if two or more queues |
11832 |
++ * have interleaving requests (see bfq_close_cooperator). |
11833 |
++ * @busy_queues: number of bfq_queues containing requests (including the |
11834 |
++ * queue under service, even if it is idling). |
11835 |
++ * @raised_busy_queues: number of weight-raised busy bfq_queues. |
11836 |
++ * @queued: number of queued requests. |
11837 |
++ * @rq_in_driver: number of requests dispatched and waiting for completion. |
11838 |
++ * @sync_flight: number of sync requests in the driver. |
11839 |
++ * @max_rq_in_driver: max number of reqs in driver in the last @hw_tag_samples |
11840 |
++ * completed requests . |
11841 |
++ * @hw_tag_samples: nr of samples used to calculate hw_tag. |
11842 |
++ * @hw_tag: flag set to one if the driver is showing a queueing behavior. |
11843 |
++ * @budgets_assigned: number of budgets assigned. |
11844 |
++ * @idle_slice_timer: timer set when idling for the next sequential request |
11845 |
++ * from the queue under service. |
11846 |
++ * @unplug_work: delayed work to restart dispatching on the request queue. |
11847 |
++ * @in_service_queue: bfq_queue under service. |
11848 |
++ * @in_service_bic: bfq_io_cq (bic) associated with the @in_service_queue. |
11849 |
++ * @last_position: on-disk position of the last served request. |
11850 |
++ * @last_budget_start: beginning of the last budget. |
11851 |
++ * @last_idling_start: beginning of the last idle slice. |
11852 |
++ * @peak_rate: peak transfer rate observed for a budget. |
11853 |
++ * @peak_rate_samples: number of samples used to calculate @peak_rate. |
11854 |
++ * @bfq_max_budget: maximum budget allotted to a bfq_queue before rescheduling. |
11855 |
++ * @group_list: list of all the bfq_groups active on the device. |
11856 |
++ * @active_list: list of all the bfq_queues active on the device. |
11857 |
++ * @idle_list: list of all the bfq_queues idle on the device. |
11858 |
++ * @bfq_quantum: max number of requests dispatched per dispatch round. |
11859 |
++ * @bfq_fifo_expire: timeout for async/sync requests; when it expires |
11860 |
++ * requests are served in fifo order. |
11861 |
++ * @bfq_back_penalty: weight of backward seeks wrt forward ones. |
11862 |
++ * @bfq_back_max: maximum allowed backward seek. |
11863 |
++ * @bfq_slice_idle: maximum idling time. |
11864 |
++ * @bfq_user_max_budget: user-configured max budget value (0 for auto-tuning). |
11865 |
++ * @bfq_max_budget_async_rq: maximum budget (in nr of requests) allotted to |
11866 |
++ * async queues. |
11867 |
++ * @bfq_timeout: timeout for bfq_queues to consume their budget; used to |
11868 |
++ * to prevent seeky queues to impose long latencies to well |
11869 |
++ * behaved ones (this also implies that seeky queues cannot |
11870 |
++ * receive guarantees in the service domain; after a timeout |
11871 |
++ * they are charged for the whole allocated budget, to try |
11872 |
++ * to preserve a behavior reasonably fair among them, but |
11873 |
++ * without service-domain guarantees). |
11874 |
++ * @bfq_raising_coeff: Maximum factor by which the weight of a boosted |
11875 |
++ * queue is multiplied |
11876 |
++ * @bfq_raising_max_time: maximum duration of a weight-raising period (jiffies) |
11877 |
++ * @bfq_raising_rt_max_time: maximum duration for soft real-time processes |
11878 |
++ * @bfq_raising_min_idle_time: minimum idle period after which weight-raising |
11879 |
++ * may be reactivated for a queue (in jiffies) |
11880 |
++ * @bfq_raising_min_inter_arr_async: minimum period between request arrivals |
11881 |
++ * after which weight-raising may be |
11882 |
++ * reactivated for an already busy queue |
11883 |
++ * (in jiffies) |
11884 |
++ * @bfq_raising_max_softrt_rate: max service-rate for a soft real-time queue, |
11885 |
++ * sectors per seconds |
11886 |
++ * @RT_prod: cached value of the product R*T used for computing the maximum |
11887 |
++ * duration of the weight raising automatically |
11888 |
++ * @oom_bfqq: fallback dummy bfqq for extreme OOM conditions |
11889 |
++ * |
11890 |
++ * All the fields are protected by the @queue lock. |
11891 |
++ */ |
11892 |
++struct bfq_data { |
11893 |
++ struct request_queue *queue; |
11894 |
++ |
11895 |
++ struct bfq_group *root_group; |
11896 |
++ |
11897 |
++ struct rb_root rq_pos_tree; |
11898 |
++ |
11899 |
++ int busy_queues; |
11900 |
++ int raised_busy_queues; |
11901 |
++ int queued; |
11902 |
++ int rq_in_driver; |
11903 |
++ int sync_flight; |
11904 |
++ |
11905 |
++ int max_rq_in_driver; |
11906 |
++ int hw_tag_samples; |
11907 |
++ int hw_tag; |
11908 |
++ |
11909 |
++ int budgets_assigned; |
11910 |
++ |
11911 |
++ struct timer_list idle_slice_timer; |
11912 |
++ struct work_struct unplug_work; |
11913 |
++ |
11914 |
++ struct bfq_queue *in_service_queue; |
11915 |
++ struct bfq_io_cq *in_service_bic; |
11916 |
++ |
11917 |
++ sector_t last_position; |
11918 |
++ |
11919 |
++ ktime_t last_budget_start; |
11920 |
++ ktime_t last_idling_start; |
11921 |
++ int peak_rate_samples; |
11922 |
++ u64 peak_rate; |
11923 |
++ unsigned long bfq_max_budget; |
11924 |
++ |
11925 |
++ struct hlist_head group_list; |
11926 |
++ struct list_head active_list; |
11927 |
++ struct list_head idle_list; |
11928 |
++ |
11929 |
++ unsigned int bfq_quantum; |
11930 |
++ unsigned int bfq_fifo_expire[2]; |
11931 |
++ unsigned int bfq_back_penalty; |
11932 |
++ unsigned int bfq_back_max; |
11933 |
++ unsigned int bfq_slice_idle; |
11934 |
++ u64 bfq_class_idle_last_service; |
11935 |
++ |
11936 |
++ unsigned int bfq_user_max_budget; |
11937 |
++ unsigned int bfq_max_budget_async_rq; |
11938 |
++ unsigned int bfq_timeout[2]; |
11939 |
++ |
11940 |
++ bool low_latency; |
11941 |
++ |
11942 |
++ /* parameters of the low_latency heuristics */ |
11943 |
++ unsigned int bfq_raising_coeff; |
11944 |
++ unsigned int bfq_raising_max_time; |
11945 |
++ unsigned int bfq_raising_rt_max_time; |
11946 |
++ unsigned int bfq_raising_min_idle_time; |
11947 |
++ unsigned long bfq_raising_min_inter_arr_async; |
11948 |
++ unsigned int bfq_raising_max_softrt_rate; |
11949 |
++ u64 RT_prod; |
11950 |
++ |
11951 |
++ struct bfq_queue oom_bfqq; |
11952 |
++}; |
11953 |
++ |
11954 |
++enum bfqq_state_flags { |
11955 |
++ BFQ_BFQQ_FLAG_busy = 0, /* has requests or is under service */ |
11956 |
++ BFQ_BFQQ_FLAG_wait_request, /* waiting for a request */ |
11957 |
++ BFQ_BFQQ_FLAG_must_alloc, /* must be allowed rq alloc */ |
11958 |
++ BFQ_BFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */ |
11959 |
++ BFQ_BFQQ_FLAG_idle_window, /* slice idling enabled */ |
11960 |
++ BFQ_BFQQ_FLAG_prio_changed, /* task priority has changed */ |
11961 |
++ BFQ_BFQQ_FLAG_sync, /* synchronous queue */ |
11962 |
++ BFQ_BFQQ_FLAG_budget_new, /* no completion with this budget */ |
11963 |
++ BFQ_BFQQ_FLAG_coop, /* bfqq is shared */ |
11964 |
++ BFQ_BFQQ_FLAG_split_coop, /* shared bfqq will be splitted */ |
11965 |
++ BFQ_BFQQ_FLAG_softrt_update, /* needs softrt-next-start update */ |
11966 |
++}; |
11967 |
++ |
11968 |
++#define BFQ_BFQQ_FNS(name) \ |
11969 |
++static inline void bfq_mark_bfqq_##name(struct bfq_queue *bfqq) \ |
11970 |
++{ \ |
11971 |
++ (bfqq)->flags |= (1 << BFQ_BFQQ_FLAG_##name); \ |
11972 |
++} \ |
11973 |
++static inline void bfq_clear_bfqq_##name(struct bfq_queue *bfqq) \ |
11974 |
++{ \ |
11975 |
++ (bfqq)->flags &= ~(1 << BFQ_BFQQ_FLAG_##name); \ |
11976 |
++} \ |
11977 |
++static inline int bfq_bfqq_##name(const struct bfq_queue *bfqq) \ |
11978 |
++{ \ |
11979 |
++ return ((bfqq)->flags & (1 << BFQ_BFQQ_FLAG_##name)) != 0; \ |
11980 |
++} |
11981 |
++ |
11982 |
++BFQ_BFQQ_FNS(busy); |
11983 |
++BFQ_BFQQ_FNS(wait_request); |
11984 |
++BFQ_BFQQ_FNS(must_alloc); |
11985 |
++BFQ_BFQQ_FNS(fifo_expire); |
11986 |
++BFQ_BFQQ_FNS(idle_window); |
11987 |
++BFQ_BFQQ_FNS(prio_changed); |
11988 |
++BFQ_BFQQ_FNS(sync); |
11989 |
++BFQ_BFQQ_FNS(budget_new); |
11990 |
++BFQ_BFQQ_FNS(coop); |
11991 |
++BFQ_BFQQ_FNS(split_coop); |
11992 |
++BFQ_BFQQ_FNS(softrt_update); |
11993 |
++#undef BFQ_BFQQ_FNS |
11994 |
++ |
11995 |
++/* Logging facilities. */ |
11996 |
++#define bfq_log_bfqq(bfqd, bfqq, fmt, args...) \ |
11997 |
++ blk_add_trace_msg((bfqd)->queue, "bfq%d " fmt, (bfqq)->pid, ##args) |
11998 |
++ |
11999 |
++#define bfq_log(bfqd, fmt, args...) \ |
12000 |
++ blk_add_trace_msg((bfqd)->queue, "bfq " fmt, ##args) |
12001 |
++ |
12002 |
++/* Expiration reasons. */ |
12003 |
++enum bfqq_expiration { |
12004 |
++ BFQ_BFQQ_TOO_IDLE = 0, /* queue has been idling for too long */ |
12005 |
++ BFQ_BFQQ_BUDGET_TIMEOUT, /* budget took too long to be used */ |
12006 |
++ BFQ_BFQQ_BUDGET_EXHAUSTED, /* budget consumed */ |
12007 |
++ BFQ_BFQQ_NO_MORE_REQUESTS, /* the queue has no more requests */ |
12008 |
++}; |
12009 |
++ |
12010 |
++#ifdef CONFIG_CGROUP_BFQIO |
12011 |
++/** |
12012 |
++ * struct bfq_group - per (device, cgroup) data structure. |
12013 |
++ * @entity: schedulable entity to insert into the parent group sched_data. |
12014 |
++ * @sched_data: own sched_data, to contain child entities (they may be |
12015 |
++ * both bfq_queues and bfq_groups). |
12016 |
++ * @group_node: node to be inserted into the bfqio_cgroup->group_data |
12017 |
++ * list of the containing cgroup's bfqio_cgroup. |
12018 |
++ * @bfqd_node: node to be inserted into the @bfqd->group_list list |
12019 |
++ * of the groups active on the same device; used for cleanup. |
12020 |
++ * @bfqd: the bfq_data for the device this group acts upon. |
12021 |
++ * @async_bfqq: array of async queues for all the tasks belonging to |
12022 |
++ * the group, one queue per ioprio value per ioprio_class, |
12023 |
++ * except for the idle class that has only one queue. |
12024 |
++ * @async_idle_bfqq: async queue for the idle class (ioprio is ignored). |
12025 |
++ * @my_entity: pointer to @entity, %NULL for the toplevel group; used |
12026 |
++ * to avoid too many special cases during group creation/migration. |
12027 |
++ * |
12028 |
++ * Each (device, cgroup) pair has its own bfq_group, i.e., for each cgroup |
12029 |
++ * there is a set of bfq_groups, each one collecting the lower-level |
12030 |
++ * entities belonging to the group that are acting on the same device. |
12031 |
++ * |
12032 |
++ * Locking works as follows: |
12033 |
++ * o @group_node is protected by the bfqio_cgroup lock, and is accessed |
12034 |
++ * via RCU from its readers. |
12035 |
++ * o @bfqd is protected by the queue lock, RCU is used to access it |
12036 |
++ * from the readers. |
12037 |
++ * o All the other fields are protected by the @bfqd queue lock. |
12038 |
++ */ |
12039 |
++struct bfq_group { |
12040 |
++ struct bfq_entity entity; |
12041 |
++ struct bfq_sched_data sched_data; |
12042 |
++ |
12043 |
++ struct hlist_node group_node; |
12044 |
++ struct hlist_node bfqd_node; |
12045 |
++ |
12046 |
++ void *bfqd; |
12047 |
++ |
12048 |
++ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
12049 |
++ struct bfq_queue *async_idle_bfqq; |
12050 |
++ |
12051 |
++ struct bfq_entity *my_entity; |
12052 |
++}; |
12053 |
++ |
12054 |
++/** |
12055 |
++ * struct bfqio_cgroup - bfq cgroup data structure. |
12056 |
++ * @css: subsystem state for bfq in the containing cgroup. |
12057 |
++ * @online: flag marked when the subsystem is inserted. |
12058 |
++ * @weight: cgroup weight. |
12059 |
++ * @ioprio: cgroup ioprio. |
12060 |
++ * @ioprio_class: cgroup ioprio_class. |
12061 |
++ * @lock: spinlock that protects @ioprio, @ioprio_class and @group_data. |
12062 |
++ * @group_data: list containing the bfq_group belonging to this cgroup. |
12063 |
++ * |
12064 |
++ * @group_data is accessed using RCU, with @lock protecting the updates, |
12065 |
++ * @ioprio and @ioprio_class are protected by @lock. |
12066 |
++ */ |
12067 |
++struct bfqio_cgroup { |
12068 |
++ struct cgroup_subsys_state css; |
12069 |
++ bool online; |
12070 |
++ |
12071 |
++ unsigned short weight, ioprio, ioprio_class; |
12072 |
++ |
12073 |
++ spinlock_t lock; |
12074 |
++ struct hlist_head group_data; |
12075 |
++}; |
12076 |
++#else |
12077 |
++struct bfq_group { |
12078 |
++ struct bfq_sched_data sched_data; |
12079 |
++ |
12080 |
++ struct bfq_queue *async_bfqq[2][IOPRIO_BE_NR]; |
12081 |
++ struct bfq_queue *async_idle_bfqq; |
12082 |
++}; |
12083 |
++#endif |
12084 |
++ |
12085 |
++static inline struct bfq_service_tree * |
12086 |
++bfq_entity_service_tree(struct bfq_entity *entity) |
12087 |
++{ |
12088 |
++ struct bfq_sched_data *sched_data = entity->sched_data; |
12089 |
++ unsigned int idx = entity->ioprio_class - 1; |
12090 |
++ |
12091 |
++ BUG_ON(idx >= BFQ_IOPRIO_CLASSES); |
12092 |
++ BUG_ON(sched_data == NULL); |
12093 |
++ |
12094 |
++ return sched_data->service_tree + idx; |
12095 |
++} |
12096 |
++ |
12097 |
++static inline struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, |
12098 |
++ int is_sync) |
12099 |
++{ |
12100 |
++ return bic->bfqq[!!is_sync]; |
12101 |
++} |
12102 |
++ |
12103 |
++static inline void bic_set_bfqq(struct bfq_io_cq *bic, |
12104 |
++ struct bfq_queue *bfqq, int is_sync) |
12105 |
++{ |
12106 |
++ bic->bfqq[!!is_sync] = bfqq; |
12107 |
++} |
12108 |
++ |
12109 |
++static inline struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic) |
12110 |
++{ |
12111 |
++ return bic->icq.q->elevator->elevator_data; |
12112 |
++} |
12113 |
++ |
12114 |
++/** |
12115 |
++ * bfq_get_bfqd_locked - get a lock to a bfqd using a RCU protected pointer. |
12116 |
++ * @ptr: a pointer to a bfqd. |
12117 |
++ * @flags: storage for the flags to be saved. |
12118 |
++ * |
12119 |
++ * This function allows bfqg->bfqd to be protected by the |
12120 |
++ * queue lock of the bfqd they reference; the pointer is dereferenced |
12121 |
++ * under RCU, so the storage for bfqd is assured to be safe as long |
12122 |
++ * as the RCU read side critical section does not end. After the |
12123 |
++ * bfqd->queue->queue_lock is taken the pointer is rechecked, to be |
12124 |
++ * sure that no other writer accessed it. If we raced with a writer, |
12125 |
++ * the function returns NULL, with the queue unlocked, otherwise it |
12126 |
++ * returns the dereferenced pointer, with the queue locked. |
12127 |
++ */ |
12128 |
++static inline struct bfq_data *bfq_get_bfqd_locked(void **ptr, |
12129 |
++ unsigned long *flags) |
12130 |
++{ |
12131 |
++ struct bfq_data *bfqd; |
12132 |
++ |
12133 |
++ rcu_read_lock(); |
12134 |
++ bfqd = rcu_dereference(*(struct bfq_data **)ptr); |
12135 |
++ |
12136 |
++ if (bfqd != NULL) { |
12137 |
++ spin_lock_irqsave(bfqd->queue->queue_lock, *flags); |
12138 |
++ if (*ptr == bfqd) |
12139 |
++ goto out; |
12140 |
++ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
12141 |
++ } |
12142 |
++ |
12143 |
++ bfqd = NULL; |
12144 |
++out: |
12145 |
++ rcu_read_unlock(); |
12146 |
++ return bfqd; |
12147 |
++} |
12148 |
++ |
12149 |
++static inline void bfq_put_bfqd_unlock(struct bfq_data *bfqd, |
12150 |
++ unsigned long *flags) |
12151 |
++{ |
12152 |
++ spin_unlock_irqrestore(bfqd->queue->queue_lock, *flags); |
12153 |
++} |
12154 |
++ |
12155 |
++static void bfq_changed_ioprio(struct bfq_io_cq *bic); |
12156 |
++static void bfq_put_queue(struct bfq_queue *bfqq); |
12157 |
++static void bfq_dispatch_insert(struct request_queue *q, struct request *rq); |
12158 |
++static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd, |
12159 |
++ struct bfq_group *bfqg, int is_sync, |
12160 |
++ struct bfq_io_cq *bic, gfp_t gfp_mask); |
12161 |
++static void bfq_end_raising_async_queues(struct bfq_data *bfqd, |
12162 |
++ struct bfq_group *bfqg); |
12163 |
++static void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg); |
12164 |
++static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq); |
12165 |
++#endif |
12166 |
+-- |
12167 |
+1.9.0 |
12168 |
+ |