Re: [PATCH 02/20 v3] sched/eevdf: Reset lag at wakeup when prev CPU has been idle

[Vincent Guittot]

On Wed, 11 Nov 2026 at 10:03, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Mon, Nov 09, 2026 at 05:27:57PM +0100, Vincent Guittot wrote:
> > On Mon, 9 Nov 2026 at 16:45, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > > On Mon, Nov 09, 2026 at 04:27:14PM +0100, Vincent Guittot wrote:
> > > > On Fri, 6 Nov 2026 at 17:22, Kayra Cizmeci <kayracizmeci@xxxxxxxxx> wrote:
> > > > >
> > > > > On Fri, Nov 06, 2026 at 04:19:29PM +0100, Vincent Guittot wrote:
> > > > > > @@ -8376,6 +8383,9 @@ static bool __dequeue_task(struct rq *rq, struct task_struct *p, int flags)
> > > > > >
> > > > > >       dequeue_hierarchy(p, flags);
> > > > > >
> > > > > > +     if (!(flags & DEQUEUE_SAVE) && !cfs_rq->h_nr_queued)
> > > > > > +             cfs_rq->idle_seq++;
> > > > >
> > > > > This needs WRITE_ONCE, no?  3/20 uses READ_ONCE().
> > > >
> > > > I hesistated to make it WRITE_ONCE but i concluded that it was not
> > > > strictly necessary because the READ_ONCE is only used to compare the
> > > > value but not use it for any computation or sync so if the compiler
> > > > wants to do any kind of optimization what we really care is : is it
> > > > equal or not ? so any intermediate value that we would read would be
> > > > ok
> > >
> > > Right, so without WRITE_ONCE() the compiler is allowed to write the word
> > > in random order byte writes. When comparing only cares about equality,
> > > any intermediate messed up state is not equal. And since we were
> > > incrementing, we're not supposed to be equal to any previous snapshot.
> > >
> > > Therefore all is well.
> > >
> > > That said; we should either document the above reasoning, or just do
> > > WRITE_ONCE() and get all the sanitizers off our back, because I'm
> > > certain that KCSAN is going to complain about this.
> >
> > I agree that it needs at least a comment. I can use WRITE_ONCE but I
> > remember seeing somethign about its cost. But does it worth having
> > sanitizers complaining ... I will run some tests and use WRITE_ONCE if
> > I don't see difference
>
> Mostly it doesn't actually cost anything. The thing with WRITE_ONCE() is
> that the compiler cannot elide or merge the stores, it is forced to
> actually emit the store. But here we need a store anyway.
>
> So the worst WRITE_ONCE() does it get in the way of compiler
> optimizations, but I don't think any of them are relevant here.

Ack