1 |
On 08/04/2011 05:30 AM, Michał Górny wrote: |
2 |
> On Sat, 30 Jul 2011 10:27:27 +0300 |
3 |
> Samuli Suominen <ssuominen@g.o> wrote: |
4 |
> |
5 |
>> Since running separate /usr without mounting it from initramfs on top |
6 |
>> of / before init is and has been broken with udev for a long time |
7 |
>> now[1][2][3] |
8 |
>> |
9 |
>> [1] http://bugs.gentoo.org/show_bug.cgi?id=364235 |
10 |
>> [2] http://fedoraproject.org/wiki/Features/UsrMove#Move_all_to_.2Fusr |
11 |
>> [3] |
12 |
>> http://www.freedesktop.org/wiki/Software/systemd/separate-usr-is-broken |
13 |
>> |
14 |
>> Can we warn users about not doing the separate /usr mistake in the |
15 |
>> handbook? |
16 |
> |
17 |
> So, let's sum up a little. |
18 |
> |
19 |
> The most common argument against separate /usr requiring a proper |
20 |
> initramfs is 'it works now, thus it's great'. That is practically |
21 |
> understandable that people don't like to switch things upside down like |
22 |
> that, especially when machines are not locally reachable. |
23 |
> |
24 |
> What's the exact differences between an initramfs and an early bootup |
25 |
> setup in rootfs? As I see it: |
26 |
> - initramfs is a small fs which is used for a short while on boot, to |
27 |
> setup the system necessarily for the early bootup sequence, |
28 |
> - while initial rootfs is a rather large piece of fs which is supposed |
29 |
> to contain random stuff necessary for the early bootup to be able to |
30 |
> proceed and mount the necessary remaining stuff before the actual |
31 |
> bootup begins. And we're mostly stuck with it for the whole runtime. |
32 |
> |
33 |
> As I see it, I see no reason to keep forcing things like complete glibc, |
34 |
> ncurses and the whole other lot of libraries for the early bootup if |
35 |
> all needed is some kind of minimal 'mount' program (for instance). |
36 |
> |
37 |
> In the ol' days I tried building a NFS-shared system and the main |
38 |
> problem was that some of early run tools relied heavily on the local |
39 |
> system libs and files before they were replaced by NFS mounts. And I |
40 |
> had to keep them in sync manually which is not the most comfortable |
41 |
> thing. |
42 |
> |
43 |
> I don't see how trying to fit the best set of libs and files into |
44 |
> rootfs can solve it. You either want for the system to be clean or |
45 |
> weirdly split to support various possible configurations. And decide |
46 |
> which are not 'weird enough' not to support. |
47 |
> |
48 |
> And really, most of the things about separate /usr are hacks which were |
49 |
> introduced because the system was incapable of a proper rootfs. |
50 |
> Read-only /usr should be read-only rootfs with writable mounts on top |
51 |
> of it. NFS-mounted /usr should be the whole system part network-mounted |
52 |
> (which would be easier if everything went into /usr rather than being |
53 |
> split). |
54 |
> |
55 |
|
56 |
It seems what we need is an migration plan. Sending out a Portage News |
57 |
item, and correcting documentation as first step. |
58 |
|
59 |
Then giving people enough time to migrate. This would give us plenty of |
60 |
time to work on the details for moving the files over from / to /usr. |
61 |
|
62 |
It seems non-problematic for new installs, as stages could ship the |
63 |
symlinks and files get installed to /usr through them, even before the |
64 |
packages are changed. |