lkml.org 
[lkml]   [2010]   [Mar]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC,PATCH 1/2] cputimers/proc: do_task_stat()->task_times() can race with getrusage()
On 03/26, Balbir Singh wrote:
>
> * Oleg Nesterov <oleg@redhat.com> [2010-03-24 21:45:50]:
>
> > do_task_stat()->task_times() can race with getrusage(), they both can
> > try to update task->prev_Xtime at the same time.
> >
> > Remove this bit of d180c5bc "sched: Introduce task_times() to replace
> > task_{u,s}time()".
>
> One of the reasons for adding this accuracy was to avoid sampling
> based noise and errors that occur with utime and stime.
>
> As long as there is no preemption during the assignment, I think we
> should be OK.

I don't think preemp_disable() can help. Probably we can use task_lock().


As for do_task_stat()->thread_group_times(), I think we can make it
rc-safe without breaking /bin/top.

1. add spin_lock_irqsave(&sig->cputimer.lock) around
sig->prev_Xtime = max(...)

2. Add a couple of barriers into thread_group_cputime()
and __exit_signal() so that without ->siglock we can
never overestimate utime/stime if we race with exit.

If we underestimate these values, this should be fine:

- the error can't be "systematic", the next read from
/prod/pid/stat will see the updated values

- the prev_Xtime logic in thread_group_times() ensures
the reported time can never go back.

IOW: at worse, cat /proc/pid/stat can miss the time
which the exited thread spent on CPU after the previous
read of /proc/pid/stat. This looks absolutely harmless,
the next read will see this time.

Probably we can even detect this case if we look at
sig->nr_threads and retry.

I'll try to make patches unless someone has a better idea.

I just can't accept the fact that we are doing while_each_thread()
under ->siglock here ;)

Oleg.



\
 
 \ /
  Last update: 2010-03-26 22:53    [from the cache]
©2003-2014 Jasper Spaans. Advertise on this site