2.6: races between modprobe and depmod in rc.sysinit

2.6: races between modprobe and depmod in rc.sysinit

[Andrey Borzenkov]

This started when I noticed that coldplugging of my USB flash did not work. I 
added call to /etc/hotplug/usb.rc in rc.sysinit but it still did not work 
because scsi.agent was not ready for 2.6. I fixed scsi.agent to work for 2.6 
- and it still did not work with nice message "sd_mod not found". I added 
strace to scsi.agent - and wow! it opens /lib/modules/`uname -r`/modules.dep 
and reads 0 bytes ... it adds /lib/modules/`uname -r`/modules.alias -and read 
0 bytes ...

the only reason I can imagine is that it hits depmod -A that runs in 
rc.sysinit and does

truncate /lib/modules/`uname -r`/modules.*
scan modules
write files

So, quesitons

- would anybody (Rusty?) object if I change depmod to do 
  build temp name
  move temp name into original

to avoid such races?

- Chmouel, Fred - is depmod on every boot really neccessary? When people 
install modules they are expected to run it actually ...

-andrey

Re: 2.6: races between modprobe and depmod in rc.sysinit

[Rusty Russell]

In message <200307312145.01711.arvidjaar@mail.ru> you write:
> the only reason I can imagine is that it hits depmod -A that runs in 
> rc.sysinit and does
> 
> truncate /lib/modules/`uname -r`/modules.*
> scan modules
> write files
> 
> So, quesitons
> 
> - would anybody (Rusty?) object if I change depmod to do 
>   build temp name
>   move temp name into original

No, that's definitely my preferred method anyway: please send patch.
Although I double this causes your problem here.

> - Chmouel, Fred - is depmod on every boot really neccessary? When people 
> install modules they are expected to run it actually ...

People get it wrong, and you still want them to boot.  Increasingly we
can live without depmod though: I might fix this in a near-future
version to live without it.

Cheers,
Rusty.
--
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.

[PATCH] 2.6: races between modprobe and depmod in rc.sysinit

[Andrey Borzenkov]

[-- Attachment #1: Type: text/plain, Size: 1346 bytes --]

On Friday 01 August 2003 10:34, Rusty Russell wrote:
> In message <200307312145.01711.arvidjaar@mail.ru> you write:
> > the only reason I can imagine is that it hits depmod -A that runs in
> > rc.sysinit and does
> >
> > truncate /lib/modules/`uname -r`/modules.*
> > scan modules
> > write files
> >
> > So, quesitons
> >
> > - would anybody (Rusty?) object if I change depmod to do
> >   build temp name
> >   move temp name into original
>
> No, that's definitely my preferred method anyway: please send patch.
> Although I double this causes your problem here.
>

this does :) I have seen even more incredible races during boot recently. It 
affects other modules as well, like input loaded by hotplug triggered by USB 
load. depmod on current 2.6 takes some not negligible time.

anyway here is patch against 0.9.13-pre that makes depmod output in temp file 
and rename it after. It fixed the problem for sure for me (like commenting 
out depmod in rc.sysinit did but most people seem to object to it).

> > - Chmouel, Fred - is depmod on every boot really neccessary? When people
> > install modules they are expected to run it actually ...
>
> People get it wrong, and you still want them to boot.  Increasingly we
> can live without depmod though: I might fix this in a near-future
> version to live without it.
>

would be nice.

-andrey

[-- Attachment #2: module-init-tools-0.9.13-tempfile.patch.bz2 --]
[-- Type: application/x-bzip2, Size: 1681 bytes --]

Re: [PATCH] 2.6: races between modprobe and depmod in rc.sysinit

[Rusty Russell]

In message <200308020122.55169.arvidjaar@mail.ru> you write:
> anyway here is patch against 0.9.13-pre that makes depmod output in temp file
> and rename it after. It fixed the problem for sure for me (like commenting 
> out depmod in rc.sysinit did but most people seem to object to it).

Thanks, I've applied this with a fix (you left the declaration of
skipchars in main(), so it shadowed the new global one).

I also repaired the testsuite after this patch (and found another bug
along the way), which also would have caught the mistake (Hint hint!).

I've put the lot up in the usual place, masquerading as 0.9.13-pre2.

	http://www.kernel.org/pub/linux/kernel/people/rusty/modules/module-init-tools-0.9.13-pre2.tar.gz

Feedback welcome!
Rusty.
--
  Anyone who quotes me in their sig is an idiot. -- Rusty Russell.

Re: [PATCH] 2.6: races between modprobe and depmod in rc.sysinit

[Andrey Borzenkov]

On Saturday 02 August 2003 04:29, Rusty Russell wrote:
> In message <200308020122.55169.arvidjaar@mail.ru> you write:
> > anyway here is patch against 0.9.13-pre that makes depmod output in temp
> > file and rename it after. It fixed the problem for sure for me (like
> > commenting out depmod in rc.sysinit did but most people seem to object to
> > it).
>
> Thanks, I've applied this with a fix (you left the declaration of
> skipchars in main(), so it shadowed the new global one).
>
> I also repaired the testsuite after this patch (and found another bug
> along the way), which also would have caught the mistake (Hint hint!).
>

thank you :)

> I've put the lot up in the usual place, masquerading as 0.9.13-pre2.
>
> 	http://www.kernel.org/pub/linux/kernel/people/rusty/modules/module-init-to
>ols-0.9.13-pre2.tar.gz
>
> Feedback welcome!

runs just fine. May I humbly ask for consistent naming scheme in future? 
Jumping from "pre" to "pre2" does not make packaging task easier :)

-andrey