Re: [PATCH 6/6] mm: Drain LRUs upon resume to userspace on nohz_full CPUs

[Michal Hocko]

On Thu 03-07-25 15:28:23, Matthew Wilcox wrote:
> On Thu, Jul 03, 2025 at 04:07:17PM +0200, Frederic Weisbecker wrote:
> > +unsigned int folio_batch_add(struct folio_batch *fbatch,
> > +			     struct folio *folio)
> > +{
> > +	unsigned int ret;
> > +
> > +	fbatch->folios[fbatch->nr++] = folio;
> > +	ret = folio_batch_space(fbatch);
> > +	isolated_task_work_queue();
> 
> Umm.  LRUs use folio_batches, but they are definitely not the only user
> of folio_batches.  Maybe you want to add a new lru_batch_add()
> abstraction, because this call is definitely being done at the wrong
> level.

You have answered one of my question in other response. My initial
thought was that __lru_add_drain_all seems to be a better fit. But then
we have a problem that draining will become an unbounded operation which
will become a problem for lru_cache_disable which will never converge
until isolated workload does the draining. So it indeed seems like we
need to queue draining when a page is added. Are there other places
where we put folios into teh folio_batch than folio_batch_add? I cannot
seem to see any...

-- 
Michal Hocko
SUSE Labs