Messages in this thread | | | Date | Tue, 17 Nov 2009 13:48:52 +0100 | From | Stanislaw Gruszka <> | Subject | Re: utime/stime decreasing on thread exit |
| |
On Tue, Nov 10, 2009 at 08:23:27PM +0100, Oleg Nesterov wrote: > > > Can't we ensure that fastpath_timer_check() never do while_each_thread() ? > > > > Removing possibility to call while_each_tread() from fastpath_timer_check() > > was exactly my intension here, perhaps I was not clear. > > Yes, yes, I understand. > > I meant, perhaps we can ensure this shouldn't happen "by design", instead > of checking ->running in fastpath_timer_check().
Rule "sig->cputimer_expire != zero implies sig->cputimer.running == true" is _almost_ assured (after fix in next mail). IMHO there is only one problem with that functions: posix_cpu_timer_set() and posix_cpu_timer_schedule().
These functions first call thread_group_cputimer() without tsk->sighand->siglock (only tasklist_lock is taken) and then arm_timer(), which setups list and cputime_expires cache.
When there is some timer expiring already we can have situation like below:
cpu_timer_sample_group() check_process_timers() stop_process_timers() arm_timer()
At the end we end with cputimer_expire != zero and ->running == false. Very unlikely situation indeed, but possible. To address this we can do in arm_timer() something like that:
if (unlikely(!sig->cputimer.running)) { cpu_timer_sample_group() bump_cpu_timer(); }
Since we have this we can do optimization, you proposed here:
http://lkml.org/lkml/2009/3/23/381
Use cputimer->running in fastpath_timer_check(). I'm going to work on it as well on some other optimizations in posix-cpu-timer.c
> Still. check_process_timers() updates sig->cputime_expires at the end, > but it never clears it. For example, > > if (sched_expires != 0 && > (sig->cputime_expires.sched_exp == 0 || > sig->cputime_expires.sched_exp > sched_expires)) > sig->cputime_expires.sched_exp = sched_expires; > > Why? > > Now suppose that (say) sig->cputime_expires.sched_exp != 0, there are > no cpu timers, ->running == F. > > In this case fastpath_timer_check() always returns T and triggers the > slow path which does nothing, not good.
This is real bug. I will fix it in the next patch.
Stanislaw
| |