Re: [PATCH 04/34] sched: add cpumask_find_and_set() and use it in __mm_cid_get()

[Yury Norov]

On Mon, Nov 20, 2023 at 11:17:32AM -0500, Mathieu Desnoyers wrote:

...

> > > diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
> > > index 2e5a95486a42..b2f095a9fc40 100644
> > > --- a/kernel/sched/sched.h
> > > +++ b/kernel/sched/sched.h
> > > @@ -3345,28 +3345,6 @@ static inline void mm_cid_put(struct mm_struct *mm)
> > >   	__mm_cid_put(mm, mm_cid_clear_lazy_put(cid));
> > >   }
> > > -static inline int __mm_cid_try_get(struct mm_struct *mm)
> > > -{
> > > -	struct cpumask *cpumask;
> > > -	int cid;
> > > -
> > > -	cpumask = mm_cidmask(mm);
> > > -	/*
> > > -	 * Retry finding first zero bit if the mask is temporarily
> > > -	 * filled. This only happens during concurrent remote-clear
> > > -	 * which owns a cid without holding a rq lock.
> > > -	 */
> > > -	for (;;) {
> > > -		cid = cpumask_first_zero(cpumask);
> > > -		if (cid < nr_cpu_ids)
> > > -			break;
> > > -		cpu_relax();
> > > -	}
> > > -	if (cpumask_test_and_set_cpu(cid, cpumask))
> > > -		return -1;
> 
> This was split in find / test_and_set on purpose because following
> patches I have (implementing numa-aware mm_cid) have a scan which
> needs to scan sets of two cpumasks in parallel (with "and" and
> and_not" operators).
> 
> Moreover, the "mask full" scenario only happens while a concurrent
> remote-clear temporarily owns a cid without rq lock. See
> sched_mm_cid_remote_clear():
> 
>         /*
>          * The cid is unused, so it can be unset.
>          * Disable interrupts to keep the window of cid ownership without rq
>          * lock small.
>          */
>         local_irq_save(flags);
>         if (try_cmpxchg(&pcpu_cid->cid, &lazy_cid, MM_CID_UNSET))
>                 __mm_cid_put(mm, cid);
>         local_irq_restore(flags);
> 
> The proposed patch here turns this scenario into something heavier
> (setting the use_cid_lock) rather than just retrying. I guess the
> question to ask here is whether it is theoretically possible to cause
> __mm_cid_try_get() to fail to have forward progress if we have a high
> rate of sched_mm_cid_remote_clear. If we decide that this is indeed
> a possible progress-failure scenario, then it makes sense to fallback
> to use_cid_lock as soon as a full mask is encountered.
> 
> However, removing the __mm_cid_try_get() helper will make it harder to
> integrate the following numa-awareness patches I have on top.
> 
> I am not against using cpumask_find_and_set, but can we keep the
> __mm_cid_try_get() helper to facilitate integration of future work ?
> We just have to make it use cpumask_find_and_set, which should be
> easy.

Sure, I can. Can you point me to the work you mention here?