Re: [RFC PATCH 12/20] kthread: Implement preferred affinity

[Frederic Weisbecker <frederic@kernel.org>]

Le Mon, Aug 05, 2024 at 04:53:51PM +0200, Vlastimil Babka a écrit :
> If you mean the loop in kcompactd_init() then indeed, but we also have a
> hook in online_pages() to start new threads on newly onlined nodes, so
> that's not a problem.
> 
> The problem (I think) I see is cpumask_of_node(pgdat->node_id) is a snapshot
> of cpus running on the NUMA node the time, and is never updated later as new
> cpus might be brought up.

Oh I see now...

> 
> kcompactd_cpu_online() does try to update that when cpus are onlined (in a
> clumsy way), there was nothing like that for kswapd and after your series
> this update is also removed for kcompactd.

Ok...

> 
> > If all users of preferred affinity were to use NUMA nodes, it could be
> > a good idea to do a flavour of kernel/smpboot.c which would handle
> > per-node kthreads instead of per-cpu kthreads. I initially thought
> > about that. It would have handled all the lifecycle of those kthreads,
> > including creation, against hotplug. Unfortunately RCU doesn't rely on
> > per-NUMA nodes but rather use its own tree.
> > 
> > If there be more users of real per NUMA nodes kthreads than kswapd and
> > kcompactd, of course that would be much worth considering.
> 
> Yeah it's not that compelling, but a way to update the preferred affine mask
> in response to cpu hotplug events, that kswapd and kcompactd could use,
> would be sufficient. And maybe more widely useful.
> 
> I guess there could be a callback defined for kthread to provide a new
> preferred_affinity, that you'd call from kthreads_hotplug_update() ?
> And kcompactd and kswapd could both use the same callback that interprets
> kthread_data() as pgdat and fetches a new cpumask of it?

It's too bad we don't have a way to have a cpumask_possible_of_node(). I've
looked into the guts of numa but that doesn't look easy to do.

Or there could be kthread_set_preferred_node()... ?

Thanks.

> 
> > Thanks.
>