Re: [PATCH RFC] time,signal: protect resource use statistics with seqlock

[Rik van Riel <riel@redhat.com>]

On 08/14/2014 12:12 PM, Oleg Nesterov wrote:
> On 08/14, Rik van Riel wrote:
>>
>> -----BEGIN PGP SIGNED MESSAGE-----
>> Hash: SHA1
>>
>> On 08/14/2014 10:24 AM, Oleg Nesterov wrote:
>>> On 08/13, Rik van Riel wrote:
>>>>
>>>> @@ -862,11 +862,9 @@ void do_sys_times(struct tms *tms) {
>>>> cputime_t tgutime, tgstime, cutime, cstime;
>>>>
>>>> -	spin_lock_irq(&current->sighand->siglock);
>>>> thread_group_cputime_adjusted(current, &tgutime, &tgstime);
>>>> cutime = current->signal->cutime; cstime =
>>>> current->signal->cstime; -
>>>> spin_unlock_irq(&current->sighand->siglock);
>>>
>>> Ah, wait, there is another problem afaics...
>>
>> Last night I worked on another problem with this code.
>>
>> After propagating the stats from a dying task to the signal struct,
>> we need to make sure that that task's stats are not counted twice.
> 
> Heh indeed ;) Can't understand how I missed that.
> 
>> This requires zeroing the stats under the write_seqlock, which was
>> easy enough to add.
> 
> Or you can expand the scope of write_seqlock/write_sequnlock, so that
> __unhash_process in called from inside the critical section. This looks
> simpler at first glance.

The problem with that is that wait_task_zombie() calls
thread_group_cputime_adjusted() in that if() branch, and
that code ends up taking the seqlock for read...

However, in __exit_signal that approach should work.

> Hmm, wait, it seems there is yet another problem ;) Afaics, you also
> need to modify __exit_signal() so that ->sum_sched_runtime/etc are
> accounted unconditionally, even if the group leader exits.
> 
> Probably this is not a big problem, and sys_times() or clock_gettime()
> do not care at all because they use current.
> 
> But without this change thread_group_cputime(reaped_zombie) won't look
> at this task_struct at all, this can lead to non-monotonic result if
> it was previously called when this task was alive (non-reaped).

You mean this whole block needs to run regardless of whether
the group is dead?

                task_cputime(tsk, &utime, &stime);
                write_seqlock(&sig->stats_lock);
                sig->utime += utime;
                sig->stime += stime;
                sig->gtime += task_gtime(tsk);
                sig->min_flt += tsk->min_flt;
                sig->maj_flt += tsk->maj_flt;
                sig->nvcsw += tsk->nvcsw;
                sig->nivcsw += tsk->nivcsw;
                sig->inblock += task_io_get_inblock(tsk);
                sig->oublock += task_io_get_oublock(tsk);
                task_io_accounting_add(&sig->ioac, &tsk->ioac);
                sig->sum_sched_runtime += tsk->se.sum_exec_runtime;

How does that square with wait_task_zombie reaping the
statistics of the whole group with thread_group_cputime_adjusted()
when the group leader is exiting?

Could that lead to things being double-counted?

Or do you mean ONLY ->sum_sched_runtime is unconditionally
accounted in __exit_signal(), because wait_task_zombie() seems
to be missing that one?