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

[Oleg Nesterov <oleg@redhat.com>]

On 08/14, Frederic Weisbecker wrote:
>
> 2014-08-14 15:22 GMT+02:00 Oleg Nesterov <oleg@redhat.com>:
> > On 08/13, Rik van Riel wrote:
> >>
> >> @@ -646,6 +646,7 @@ struct signal_struct {
> >>        * Live threads maintain their own counters and add to these
> >>        * in __exit_signal, except for the group leader.
> >>        */
> >> +     seqlock_t stats_lock;
> >
> > Ah. Somehow I thought that you were going to use seqcount_t and fallback
> > to taking ->siglock if seqcount_retry, but this patch adds the "full blown"
> > seqlock_t.
> >
> > OK, I won't argue, this can make the seqbegin_or_lock simpler...
>
> Is this really needed? seqlock are useful when we have concurrent
> updaters. But updaters of thread stats should be under the thread lock
> already, right? If we have only one updater at a time, seqcount should
> be enough.

Yes, this is what I meant. Although I can see 2 reasons to use seqlock_t:

	1. It can simplify the seqbegin-or-lock logic. If nothing else,
	   you simply can't use read_seqbegin_or_lock() to take ->siglock.
	   But this is just syntactic sugar.

	2. If we use ->siglock in fallback path, we need to verify that
	   thread_group_cputime() is never called with ->siglock held first.

	   Or, we need a fat comment to explain that need_seqrtry == T is not
	   possible if it is called under ->siglock, and thus "fallback to
	   lock_task_sighand" must be always safe. But in this case we need
	   to ensure that the caller didn't do write_seqcount_begin().

So perhaps seqlock_t makes more sense at least initially...

Oleg.