lkml.org 
[lkml]   [2011]   [Oct]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Linux 3.1-rc9
On Mon, 17 Oct 2011, Peter Zijlstra wrote:

> On Mon, 2011-10-17 at 11:31 -0700, Linus Torvalds wrote:
> > On Mon, Oct 17, 2011 at 10:54 AM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > >
> > > I could of course propose this... but I really won't since I'm half
> > > retching by now.. ;-)
> >
> > Wow. Is this "ugly and fragile code week" and I just didn't get the memo?
>
> Do I get a price?
>
> > I do wonder if we might not fix the problem by just taking the
> > *existing* lock in the right order?
> >
> > IOW, how nasty would be it be to make "scheduler_tick()" just get the
> > cputimer->lock outside or rq->lock?
> >
> > Sure, we'd hold that lock *much* longer than we need, but how much do
> > we care? Is that a lock that gets contention? It migth be the simple
> > solution for now - I *would* like to get 3.1 out..
>
> Ah, sadly the tick isn't the only one with the inverted callchain,
> pretty much every callchain in the scheduler ends up in update_curr()
> one way or another.
>
> The easier way around might be something like this... even when two
> threads in a process race to enable this clock the the wasted time is
> pretty much of the same order as we would otherwise have wasted spinning
> on the lock and the update_gt_cputime() think would end up moving the
> clock fwd to the latest outcome any which way.
>
> Humm,. Thomas anything?

No, that should work. It does not make that call path more racy
against exit, which is another trainwreck at least on 32bit machines
which I discovered while looking for the problems with your patch.

thread_group_cputime() reads task->signal->utime/stime/sum_sched_runtime

These fields are updated in __exit_signal() w/o holding
task->signal->cputimer.lock. So nothing prevents that these values
change while we read them.

All callers of thread_group_cputime() except the scheduler callpath
hold sighand lock, which is also taken in __exit_signal().

So your patch does not make that particular case worse.

That said, I really need some sleep before I can make a final
judgement on that horror. The call paths are such an intermingled mess
that it's not funny anymore. I do that tomorrow morning first thing.

Thanks,

tglx


\
 
 \ /
  Last update: 2011-10-17 23:03    [W:0.126 / U:0.348 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site