(no subject)

[Shakeel Butt]

Bcc:
Subject: Re: [PATCH v2] mm: skip folio_activate() for mlocked folios
Message-ID: <mkfwoqg3k66632ltgg2t6rp3l2pmkrhu2jfgip2nsn7qp7aveo@l3ajbhzvzuns>
Reply-To:
In-Reply-To: <aOPDRmk2Zd20qxfk@shell.ilvokhin.com>

CC Huge, yangge, David

On Mon, Oct 06, 2025 at 01:25:26PM +0000, Dmitry Ilvokhin wrote:
> __mlock_folio() does not move folio to unevicable LRU, when
> folio_activate() removes folio from LRU.
> 
> To prevent this case also check for folio_test_mlocked() in
> folio_mark_accessed(). If folio is not yet marked as unevictable, but
> already marked as mlocked, then skip folio_activate() call to allow
> __mlock_folio() to make all necessary updates. It should be safe to skip
> folio_activate() here, because mlocked folio should end up in
> unevictable LRU eventually anyway.
> 
> To observe the problem mmap() and mlock() big file and check Unevictable
> and Mlocked values from /proc/meminfo. On freshly booted system without
> any other mlocked memory we expect them to match or be quite close.
> 
> See below for more detailed reproduction steps. Source code of stat.c is
> available at [1].
> 
>   $ head -c 8G < /dev/urandom > /tmp/random.bin
> 
>   $ cc -pedantic -Wall -std=c99 stat.c -O3 -o /tmp/stat
>   $ /tmp/stat
>   Unevictable:     8389668 kB
>   Mlocked:         8389700 kB
> 
>   Need to run binary twice. Problem does not reproduce on the first run,
>   but always reproduces on the second run.
> 
>   $ /tmp/stat
>   Unevictable:     5374676 kB
>   Mlocked:         8389332 kB
> 
> [1]: https://gist.github.com/ilvokhin/e50c3d2ff5d9f70dcbb378c6695386dd
> 
> Co-developed-by: Kiryl Shutsemau <kas@kernel.org>
> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
> Signed-off-by: Dmitry Ilvokhin <d@ilvokhin.com>
> Acked-by: Usama Arif <usamaarif642@gmail.com>
> ---
> Changes in v2:
>   - Rephrase commit message: frame it in terms of unevicable LRU, not stat
>     accounting.
> 
>  mm/swap.c | 10 ++++++++++
>  1 file changed, 10 insertions(+)
> 
> diff --git a/mm/swap.c b/mm/swap.c
> index 2260dcd2775e..f682f070160b 100644
> --- a/mm/swap.c
> +++ b/mm/swap.c
> @@ -469,6 +469,16 @@ void folio_mark_accessed(struct folio *folio)
>  		 * this list is never rotated or maintained, so marking an
>  		 * unevictable page accessed has no effect.
>  		 */
> +	} else if (folio_test_mlocked(folio)) {
> +		/*
> +		 * Pages that are mlocked, but not yet on unevictable LRU.
> +		 * They might be still in mlock_fbatch waiting to be processed
> +		 * and activating it here might interfere with
> +		 * mlock_folio_batch(). __mlock_folio() will fail
> +		 * folio_test_clear_lru() check and give up. It happens because
> +		 * __folio_batch_add_and_move() clears LRU flag, when adding
> +		 * folio to activate batch.
> +		 */

This makes sense as activating an mlocked folio should be a noop but I
am wondering why we are seeing this now. By this, I mean mlock()ed
memory being delayed to get to unevictable LRU. Also I remember Hugh
recently [1] removed the difference betwen mlock percpu cache and other
percpu caches of clearing LRU bit on entry. Does you repro work even
with Hugh's changes or without it?

[1] https://lore.kernel.org/all/05905d7b-ed14-68b1-79d8-bdec30367eba@google.com/