lkml.org 
[lkml]   [2008]   [Nov]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: regression introduced by - timers: fix itimer/many thread hang
From
Date
On Wed, 2008-11-05 at 17:58 -0800, Frank Mayhar wrote:
> On Tue, 2008-10-28 at 14:38 -0400, Doug Chapman wrote:
> > On Mon, 2008-10-27 at 11:39 -0700, Frank Mayhar wrote:
> > > On Wed, 2008-10-22 at 13:03 -0400, Doug Chapman wrote:
> > > > Unable to handle kernel paging request at virtual address
> > > > 94949494949494a4
> > >
> > > I take it this can be read as an uninitialized (or cleared) pointer?
> > >
> > > It certainly looks like this is a race in thread (process?) teardown. I
> > > don't have hardware on which to reproduce this but _looks_ like another
> > > thread has gotten in and torn down the process while we've been busy.
> >
> > I finally managed to get kdump working and caught this in the act. I
> > still need to dig into this more but I think these 2 threads will show
> > us the race condition. Note that this is a slightly hacked kernel in
> > that I removed "static" from a few functions to better see what was
> > going on but no real functional changes when compared to a recent (day
> > old or so) git pull from Linus's tree.
>
> After digging through this a bit, I've concluded that it's probably a
> race between process reap and the dequeue_entity() call to update_curr()
> combined with a side effect of the slab debug stuff. The
> account_group_exec_runtime() routine (like the rest of these routines)
> checks tsk->signal and tsk->signal->cputime.totals for NULL to make sure
> they're still valid. It looks like at this point tsk->signal is valid
> (since the tsk->signal->cputime dereference succeeded) but
> tsk->signal->cputime.totals is invalid. That can't happen unless the
> process is being reaped, and in fact can only happen in a narrow window
> in __cleanup_signal() between the call to thread_group_cputime_free()
> and the kmem_cache_free() of the signal struct itself. Without the slab
> debug stuff it would either skip the update (by noticing that pointers
> were NULL) or blithely update freed structures.
>
> I can't see anything that would prevent this from happening in the
> general case. I don't see what would stop the parent from coming in on
> another CPU and reaping the process after schedule() has picked it off
> the rq but before the deactivate_task->dequeue_task->dequeue_entity->
> update_curr chain completes. I see that schedule() disables preemption
> on that CPU but will that really do the job? I also suspect that with
> hyperthreading both of these processes are on the same silicon, meaning
> that one can be unexpectedly suspended while the other runs, thereby
> making the window wider.
>
> Unfortunately I don't know the code well enough to know what the right
> fix is. Maybe account_group_exec_runtime() should check for PF_EXITED?

That is just plain ugly. The right fix to me seems to destroy the
signal/thread group stuff _after_ the last task goes away.

> Maybe update_curr() should do that? I think that it makes more sense
> for dequeue_entity() to do the check and then not call update_curr() if
> the task is exiting but I defer to others with more knowledge of this
> area.

Hell no. Its a race in this signal/thread group stuff, fix it there.

But now you made me look at this patch:

commit f06febc96ba8e0af80bcc3eaec0a109e88275fac
Author: Frank Mayhar <fmayhar@google.com>
Date: Fri Sep 12 09:54:39 2008 -0700

timers: fix itimer/many thread hang


and I'm thinking you just rendered big iron unbootable.

You replaced a loop-over-threads by a loop-over-cpus, on every tick. Did
you stop to think what would happen on 4096 cpu machines?

Also, you just introduced per-cpu allocations for each thread-group,
while Christoph is reworking the per-cpu allocator, with one unfortunate
side-effect - its going to have a limited size pool. Therefore this will
limit the number of thread-groups we can have.

Hohumm, not at all liking this..


\
 
 \ /
  Last update: 2008-11-06 12:05    [W:1.359 / U:1.248 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site