From: Joshua Hahn <joshua.hahnjy@gmail.com>
To: Vlastimil Babka <vbabka@suse.cz>
Cc: Andrew Morton <akpm@linux-foundation.org>,
Chris Mason <clm@fb.com>, Kiryl Shutsemau <kirill@shutemov.name>,
Brendan Jackman <jackmanb@google.com>,
Johannes Weiner <hannes@cmpxchg.org>,
Michal Hocko <mhocko@suse.com>,
Suren Baghdasaryan <surenb@google.com>, Zi Yan <ziy@nvidia.com>,
linux-kernel@vger.kernel.org, linux-mm@kvack.org,
kernel-team@meta.com
Subject: Re: [PATCH v3 3/3] mm/page_alloc: Batch page freeing in free_frozen_page_commit
Date: Fri, 10 Oct 2025 08:13:59 -0700 [thread overview]
Message-ID: <20251010151400.1925642-1-joshua.hahnjy@gmail.com> (raw)
In-Reply-To: <c48ae6d5-3a04-4c32-8084-33d2e10d8d2e@suse.cz>
On Fri, 10 Oct 2025 15:37:23 +0200 Vlastimil Babka <vbabka@suse.cz> wrote:
[...snip...]
> > high = nr_pcp_high(pcp, zone, batch, free_high);
> > if (pcp->count < high)
> > - return;
> > + return true;
> > +
> > + to_free = nr_pcp_free(pcp, batch, high, free_high);
> > + if (to_free == 0)
> > + return true;
> >
> > - free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
> > - pcp, pindex);
> > +free_batch:
> > + to_free_batched = min(to_free, batch);
> > + free_pcppages_bulk(zone, to_free_batched, pcp, pindex);
> > if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
Hello Vlastimil, thank you for your review!
> We could do this handling once after all batches. But maybe it's better to
> act as soon as it becomes true, and checking once ber batch isn't measurably
> slower, dunno.
My thinking was that if we are to release the pcp and zone lock and another
task remotely inspects this pcp in the middle of the loop, I would want the
ZONE_BELOW_HIGH bit to reflect the correct state of the pcp, as opposed to
doing it once at the end.
> > zone_watermark_ok(zone, 0, high_wmark_pages(zone),
> > ZONE_MOVABLE, 0)) {
> > @@ -2887,6 +2901,35 @@ static void free_frozen_page_commit(struct zone *zone,
> > next_memory_node(pgdat->node_id) < MAX_NUMNODES)
> > atomic_set(&pgdat->kswapd_failures, 0);
> > }
> > + high = nr_pcp_high(pcp, zone, batch, free_high);
>
> It's not clear why we recalculate this. Ah I see, the calculation involves a
> ZONE_BELOW_HIGH check which we might have changed above. So as a result ths
> patch isn't a straightforward "we free the same amount of pages but in
> smaller batches" but something different (and I'm not immediately sure what
> exactly).
Yes! So nr_pcp_high will check if we are now below high, and see what the
new high should be, and see if we need to free some more pages. My goal was
to prevent having to call free_frozen_page_commit again in the future once
we realize that there are still pages yet to be freed, and handle that case
within this loop as if it was what we originally decided to free.
> I think it's an argument for doing the ZONE_BELOW_HIGH test above just once
> in the end, and not recalculating "high" here. I'd just stick to the
> "to_free" calculated uprofront and decreasing it by "to_free_batched" each
> round.
Yes, this makes sense to me. The last sentence in my paragraph above seems
a little sneaky in doing more work than it originally was called to do, so
I do agree that we should do the ZONE_BELOW_HIGH check once, and also
calculate how many pages to free at once in the beginning.
> We should maybe also check that pcp->count went to 0 and bail out so we
> don't make useless iterations in rare cases when someone else drains the pcp
> between our batches (free_pcppages_bulk() protects itself from passing too
> high "count" so it would be just some wasted cpu otherwise, not a disaster).
Hopefully the check below is enough to handle this race case. When it
determines whether to iterate through in the loop again, we check that
to_free > 0 && pcp->count >= high, so if someone else were to drain the
drain the pcp->count, it would go under high (I don't think it's possible
for high to become negative here) it would break this check and prevent
another iteration.
> > + to_free -= to_free_batched;
> > + if (pcp->count >= high) {
> > + pcp_spin_unlock(pcp);
> > + pcp_trylock_finish(*UP_flags);
> > +
> > + pcp_trylock_prepare(*UP_flags);
> > + pcp = pcp_spin_trylock(zone->per_cpu_pageset);
> > + if (!pcp) {
> > + pcp_trylock_finish(*UP_flags);
> > + return false;
> > + }
> > +
> > + /*
> > + * Check if this thread has been migrated to a different
> > + * CPU. If that is the case, give up and indicate that
> > + * the pcp is returned in an unlocked state.
> > + */
> > + if (smp_processor_id() != cpu) {
>
> We could have remembered the old pcp pointer and compare that instead of
> doing smp_processor_id(), although that should also work.
I think that would have also worked. To be completely honest, I wasn't
sure if there was some pcp stuff inside that would change the pointer,
so just went with using the CPU int instead.
> > + pcp_spin_unlock(pcp);
> > + pcp_trylock_finish(*UP_flags);
> > + return false;
> > + }
> > + }
> > +
> > + if (to_free > 0 && pcp->count >= high)
Check is here: ^^^^^^^^^^^^^^^^^^
> > + goto free_batch;
> > +
> > + return true;
> > }
> >
> > /*
> > @@ -2934,7 +2977,9 @@ static void __free_frozen_pages(struct page *page, unsigned int order,
> > pcp_trylock_prepare(UP_flags);
> > pcp = pcp_spin_trylock(zone->per_cpu_pageset);
> > if (pcp) {
> > - free_frozen_page_commit(zone, pcp, page, migratetype, order, fpi_flags);
> > + if (!free_frozen_page_commit(zone, pcp, page, migratetype,
> > + order, fpi_flags, &UP_flags))
> > + return;
> > pcp_spin_unlock(pcp);
> > } else {
> > free_one_page(zone, page, pfn, order, fpi_flags);
> > @@ -3034,8 +3079,11 @@ void free_unref_folios(struct folio_batch *folios)
> > migratetype = MIGRATE_MOVABLE;
> >
> > trace_mm_page_free_batched(&folio->page);
> > - free_frozen_page_commit(zone, pcp, &folio->page, migratetype,
> > - order, FPI_NONE);
> > + if (!free_frozen_page_commit(zone, pcp, &folio->page,
> > + migratetype, order, FPI_NONE, &UP_flags)) {
> > + pcp = NULL;
> > + locked_zone = NULL;
> > + }
> > }
> >
> > if (pcp) {
Thank you again for your review, Vlastimil. I agree with all of your
feedback, so I will make those changes (as well as the commit message
change for patch 2/3) and submit a new version. I hope you have a great day!
Joshua
prev parent reply other threads:[~2025-10-10 15:14 UTC|newest]
Thread overview: 9+ messages / expand[flat|nested] mbox.gz Atom feed top
2025-10-02 20:46 [PATCH v3 0/3] mm/page_alloc: Batch callers of free_pcppages_bulk Joshua Hahn
2025-10-02 20:46 ` [PATCH v3 1/3] mm/page_alloc/vmstat: Simplify refresh_cpu_vm_stats change detection Joshua Hahn
2025-10-10 12:44 ` Vlastimil Babka
2025-10-10 18:30 ` SeongJae Park
2025-10-02 20:46 ` [PATCH v3 2/3] mm/page_alloc: Batch page freeing in decay_pcp_high Joshua Hahn
2025-10-10 12:48 ` Vlastimil Babka
2025-10-02 20:46 ` [PATCH v3 3/3] mm/page_alloc: Batch page freeing in free_frozen_page_commit Joshua Hahn
2025-10-10 13:37 ` Vlastimil Babka
2025-10-10 15:13 ` Joshua Hahn [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20251010151400.1925642-1-joshua.hahnjy@gmail.com \
--to=joshua.hahnjy@gmail.com \
--cc=akpm@linux-foundation.org \
--cc=clm@fb.com \
--cc=hannes@cmpxchg.org \
--cc=jackmanb@google.com \
--cc=kernel-team@meta.com \
--cc=kirill@shutemov.name \
--cc=linux-kernel@vger.kernel.org \
--cc=linux-mm@kvack.org \
--cc=mhocko@suse.com \
--cc=surenb@google.com \
--cc=vbabka@suse.cz \
--cc=ziy@nvidia.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®