Message-ID: <jhjwo3e6zd1.mognet@arm.com>
Date:   Wed, 08 Jul 2020 12:05:30 +0100
From:   Valentin Schneider <valentin.schneider@....com>
To:     Qais Yousef <qais.yousef@....com>
Cc:     Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Doug Anderson <dianders@...omium.org>,
        Jonathan Corbet <corbet@....net>,
        Juri Lelli <juri.lelli@...hat.com>,
        Vincent Guittot <vincent.guittot@...aro.org>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Ben Segall <bsegall@...gle.com>, Mel Gorman <mgorman@...e.de>,
        Luis Chamberlain <mcgrof@...nel.org>,
        Kees Cook <keescook@...omium.org>,
        Iurii Zaikin <yzaikin@...gle.com>,
        Quentin Perret <qperret@...gle.com>,
        Patrick Bellasi <patrick.bellasi@...bug.net>,
        Pavan Kondeti <pkondeti@...eaurora.org>,
        linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-fsdevel@...r.kernel.org
Subject: Re: [PATCH v6 1/2] sched/uclamp: Add a new sysctl to control RT default boost value


On 07/07/20 13:36, Qais Yousef wrote:
> On 07/07/20 12:30, Valentin Schneider wrote:
>>
>> On 07/07/20 10:34, Qais Yousef wrote:
>> > On 07/06/20 16:49, Valentin Schneider wrote:
>> >>
>> >> On 06/07/20 15:28, Qais Yousef wrote:
>> >> > CC: linux-fsdevel@...r.kernel.org
>> >> > ---
>> >> >
>> >> > Peter
>> >> >
>> >> > I didn't do the
>> >> >
>> >> >       read_lock(&taslist_lock);
>> >> >       smp_mb__after_spinlock();
>> >> >       read_unlock(&tasklist_lock);
>> >> >
>> >> > dance you suggested on IRC as it didn't seem necessary. But maybe I missed
>> >> > something.
>> >> >
>> >>
>> >> So the annoying bit with just uclamp_fork() is that it happens *before* the
>> >> task is appended to the tasklist. This means without too much care we
>> >> would have (if we'd do a sync at uclamp_fork()):
>> >>
>> >>   CPU0 (sysctl write)                                CPU1 (concurrent forker)
>> >>
>> >>                                                        copy_process()
>> >>                                                          uclamp_fork()
>> >>                                                            p.uclamp_min = state
>> >>     state = foo
>> >>
>> >>     for_each_process_thread(p, t)
>> >>       update_state(t);
>> >>                                                          list_add(p)
>> >>
>> >> i.e. that newly forked process would entirely sidestep the update. Now,
>> >> with Peter's suggested approach we can be in a much better situation. If we
>> >> have this in the sysctl update:
>> >>
>> >>   state = foo;
>> >>
>> >>   read_lock(&taslist_lock);
>> >>   smp_mb__after_spinlock();
>> >>   read_unlock(&tasklist_lock);
>> >>
>> >>   for_each_process_thread(p, t)
>> >>     update_state(t);
>> >>
>> >> While having this in the fork:
>> >>
>> >>   write_lock(&tasklist_lock);
>> >>   list_add(p);
>> >>   write_unlock(&tasklist_lock);
>> >>
>> >>   sched_post_fork(p); // state re-read here; probably wants an mb first
>> >>
>> >> Then we can no longer miss an update. If the forked p doesn't see the new
>> >> value, it *must* have been added to the tasklist before the updater loops
>> >> over it, so the loop will catch it. If it sees the new value, we're done.
>> >
>> > uclamp_fork() has nothing to do with the race. If copy_process() duplicates the
>> > task_struct of an RT task, it'll copy the old value.
>> >
>>
>> Quite so; my point was if we were to use uclamp_fork() as to re-read the value.
>>
>> > I'd expect the newly introduced sched_post_fork() (also in copy_process() after
>> > the list update) to prevent this race altogether.
>> >
>> > Now we could end up with a problem if for_each_process_thread() doesn't see the
>> > newly forked task _after_ sched_post_fork(). Hence my question to Peter.
>> >
>>
>>
>> >>
>> >> AIUI, the above strategy doesn't require any use of RCU. The update_state()
>> >> and sched_post_fork() can race, but as per the above they should both be
>> >> writing the same value.
>> >
>> > for_each_process_thread() must be protected by either tasklist_lock or
>> > rcu_read_lock().
>> >
>>
>> Right
>>
>> > The other RCU logic I added is not to protect against the race above. I
>> > describe the other race condition in a comment.
>>
>> I take it that's the one in uclamp_sync_util_min_rt_default()?
>
> Correct.
>
>>
>> __setscheduler_uclamp() can't be preempted as we hold task_rq_lock(). It
>> can indeed race with the sync though, but again with the above suggested
>> setup it would either:
>> - see the old value, but be guaranteed to be iterated over later by the
>>   updater
>> - see the new value
>
> AFAIU rcu_read_lock() is light weight. So having the protection applied is more
> robust against future changes.

So I think the one thing you win by having this dance with mb's and the
suggested handling of the task list is that you do not need any
rcu_synchronize() anymore. Both approaches have merit, it's just that the
way I understood the suggestion to add sched_post_fork() was to simplify
the ordering of the update with the aforementioned scheme.

>
>>
>> sched_post_fork() being preempted out is a bit more annoying, but what
>> prevents us from making that bit preempt-disabled?
>
> preempt_disable() is not friendly to RT and heavy handed approach IMO.
>

True, but this is both an infrequent and slow sysctl path, so I don't think
RT would care much.

>>
>> I have to point out I'm assuming here updaters are serialized, which does
>> seem to be see the case (cf. uclamp_mutex).
>
> Correct.
>
> Thanks

Powered by blists - more mailing lists