lkml.org 
[lkml]   [2009]   [Nov]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: utime/stime decreasing on thread exit
On Mon, Nov 09, 2009 at 08:23:55PM +0100, Oleg Nesterov wrote:
> On 11/09, Peter Zijlstra wrote:
> >
> > On Mon, 2009-11-09 at 18:20 +0100, Oleg Nesterov wrote:
> >
> > > Stanislaw has already sent the patch, but I don't know what happened
> > > with this patch:
> > >
> > > [PATCH 1/2] posix-cpu-timers: avoid do_sys_times() races with __exit_signal()
> > > http://marc.info/?l=linux-kernel&m=124505545131145

I did not resubmit it enough times :) But I didn't thought it worth to,
since performance can be degraded.

> > That patch has the siglock in the function calling
> > thread_group_cputime(), the 22 code had it near the loop proper, which
> > to me seems a more sensible thing, since there could be more callers,
> > no?
>
> Well, we can't take ->siglock in thread_group_cputime(), sometimes it
> is called under ->siglock. do_task_stat(), get_cpu_itimer() at least.
>
> IIRC, Stanislaw verified other callers have no problems with this helper.

Actually I didn't. Try to do now.

Most calls are done with sighand or tasklist lock taken, except

Cscope tag: thread_group_cputime
# line filename / context / line
1 1353 fs/binfmt_elf.c <<fill_prstatus>>
thread_group_cputime(p, &cputime);
2 1403 fs/binfmt_elf_fdpic.c <<fill_prstatus>>
thread_group_cputime(p, &cputime);
10 129 mm/oom_kill.c <<badness>>
thread_group_cputime(p, &task_time);

I do not think in case 1 and 2 lock is needed since it seems to be core dump
with all threads dead. Not sure about 10 - oom killer.

One other exception is:
fastpath_timer_check() -> thread_group_cputimer() -> thread_group_cputime()

We can solve this like that:

--- a/kernel/posix-cpu-timers.c
+++ b/kernel/posix-cpu-timers.c
@@ -1375,13 +1375,8 @@ static inline int fastpath_timer_check(struct task_struct *tsk)
}

sig = tsk->signal;
- if (!task_cputime_zero(&sig->cputime_expires)) {
- struct task_cputime group_sample;
-
- thread_group_cputimer(tsk, &group_sample);
- if (task_cputime_expired(&group_sample, &sig->cputime_expires))
- return 1;
- }
+ if (!task_cputime_zero(&sig->cputime_expires))
+ return 1;

return sig->rlim[RLIMIT_CPU].rlim_cur != RLIM_INFINITY;
}
Or stay with task_cputime_expired() but only if cputimer is currently running.

Cheers
Stanislaw


\
 
 \ /
  Last update: 2009-11-10 11:49    [W:0.935 / U:0.056 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site