Re: [patch V3 13/20] sched/mmcid: Provide precomputed maximal value

[Mathieu Desnoyers]

On 2025-10-31 11:06, Thomas Gleixner wrote:
> On Thu, Oct 30 2025 at 10:23, Mathieu Desnoyers wrote:
> > On 2025-10-29 09:09, Thomas Gleixner wrote:
> > >    struct mm_mm_cid {
> > >    	struct mm_cid_pcpu	__percpu *pcpu;
> > > +	unsigned int		max_cids;
> > >    	unsigned int		nr_cpus_allowed;
> > > +	unsigned int		users;
> >
> > I suspect this reintroduces false-sharing between the "users"
> > and "lock" fields (updated every time a thread is forked/exits)
> > and load of the pcpu pointer which is pretty much immutable.
> > This will slow down accesses to the percpu data in the scheduler
> > fast path.
>
> At this point yes, but when all bits are in place then the lock fields
> end up in a different cache line.
>
> The false sharing issue vs. *pcpu and max_cids is minor, but I can move
> the low frequency modified members past the work, so it does not matter
> at all. The work stuff is rarely used, so there is no point to worry
> about the occasional cache line contention on that.

If we have alignment requirements on fields that matter for performance,
I recommend using __cacheline_group_{begin,end}_aligned() to make this
explicit. See include/linux/cache.h.

Thanks,

Mathieu



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