Re: [RFC PATCH] sched: Fix performance regression introduced by mm_cid

[Mathieu Desnoyers]

On 2023-04-04 05:10, Peter Zijlstra wrote:
> On Tue, Apr 04, 2023 at 10:57:49AM +0200, Peter Zijlstra wrote:
> > Sorry for being a little late to the party.
> >
> > On Mon, Apr 03, 2023 at 02:13:42PM -0400, Mathieu Desnoyers wrote:
> >
> > > diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> > > index 0722859c3647..21466fdc4dc6 100644
> > > --- a/include/linux/mm_types.h
> > > +++ b/include/linux/mm_types.h
> > > @@ -609,6 +609,14 @@ struct mm_struct {
> > >   		 * were being concurrently updated by the updaters.
> > >   		 */
> > >   		raw_spinlock_t cid_lock;
> > > +		/**
> > > +		 * @pcpu_cid: Per-cpu current cid.
> > > +		 *
> > > +		 * Keep track of the currently allocated mm_cid for each cpu.
> > > +		 * The per-cpu mm_cid values are serialized by their respective
> > > +		 * runqueue locks.
> > > +		 */
> > > +		int __percpu *pcpu_cid;
> >
> > This *might* be a problem... Consider running *many* single threaded
> > processes, this could exhaust the limited per-cpu storage (32bit) or use
> > excessive memory on large CPU systems.
> >
> > Consider having to allocate per-cpu storage on a 4K CPU system while
> > only running single threaded processes -- but *LOTS* of them..
>
> Ah, mm_struct::rss_stat[] beat us and set precedent, so not our problem I
> suppose :-)

Yes, exactly. When I've seen that mm_struct::rss_stat[] did not care 
about having per-mm/per-cpu data, I went the same way.

Thanks,

Mathieu


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