Re: [patch V2 0/4] sched/mmcid: Cure mode transition woes

[Mathieu Desnoyers]

On 2026-02-02 07:54, Peter Zijlstra wrote:
> On Mon, Feb 02, 2026 at 06:46:34AM -0500, Mathieu Desnoyers wrote:
> > On 2026-02-02 05:14, Peter Zijlstra wrote:
> > > On Mon, Feb 02, 2026 at 10:39:35AM +0100, Thomas Gleixner wrote:
> > >
> > > > --- a/kernel/sched/core.c
> > > > +++ b/kernel/sched/core.c
> > > > @@ -10445,6 +10445,12 @@ static bool mm_update_max_cids(struct mm
> > > >    	/* Flip the mode and set the transition flag to bridge the transfer */
> > > >    	WRITE_ONCE(mc->mode, mc->mode ^ (MM_CID_TRANSIT | MM_CID_ONCPU));
> > > > +	/*
> > > > +	 * Order the store against the subsequent fixups so that
> > > > +	 * acquire(rq::lock) cannot be reordered by the CPU before the
> > > > +	 * store.
> > > > +	 */
> > > > +	smp_mb();
> > > >    	return true;
> > > >    }
> > > > @@ -10487,6 +10493,16 @@ static inline void mm_update_cpus_allowe
> > > >    	irq_work_queue(&mc->irq_work);
> > > >    }
> > > > +static inline void mm_cid_complete_transit(struct mm_struct *mm, unsigned int mode)
> > > > +{
> > > > +	/*
> > > > +	 * Ensure that the store removing the TRANSIT bit cannot be
> > > > +	 * reordered by the CPU before the fixups have been completed.
> > > > +	 */
> > > > +	smp_mb();
> > > > +	WRITE_ONCE(mm->mm_cid.mode, mode);
> > > > +}
> > >
> > > I think this could've been smp_store_release(), but this is the slow
> > > path so nobody cares and this is nicely symmetric.
> >
> > I'm not sure the store-release would work here. What load-acquire
> > would it pair with ?
>
> The purpose here -- per the comment is to ensure the fixup stuff is
> visible before the TRANSIT bit goes 0, store-release ensures that.
>
> That pairs with whatever cares about this barrier now.

Now that I think about it some more, I think my advice about adding
smp_mb() before rq lock/after rq unlock was wrong.

The store setting transit will be ordered by rq _unlock_ which acts
as a release barrier, and store clearing transit will be ordered by
rq _lock_ acting as an acquire barrier. Those pair with the respective
rq unlock/lock in the scheduler.

So AFAIU we can remove those two useless smp_mb().

Thoughts ?

Thanks,

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com