Re: [PATCH mm-unstable 1/1] mm: fix deferred split queue races during migration

[Usama Arif <usama.arif@linux.dev>]

On Wed,  1 Apr 2026 21:10:32 +0800 Lance Yang <lance.yang@linux.dev> wrote:

> From: Lance Yang <lance.yang@linux.dev>
> 
> migrate_folio_move() records the deferred split queue state from src and
> replays it on dst. Replaying it after remove_migration_ptes(src, dst, 0)
> makes dst visible before it is requeued, so a concurrent rmap-removal path
> can mark dst partially mapped and trip the WARN in deferred_split_folio().
> 
> Move the requeue before remove_migration_ptes() so dst is back on the
> deferred split queue before it becomes visible again.
> 
> Because migration still holds dst locked at that point, teach
> deferred_split_scan() to requeue a folio when folio_trylock() fails.
> Otherwise a fully mapped underused folio can be dequeued by the shrinker
> and silently lost from split_queue.
> 
> Link: https://syzkaller.appspot.com/bug?extid=a7067a757858ac8eb085
> Fixes: 8a8ca142a488 ("mm: migrate: requeue destination folio on deferred split queue")
> Reported-by: syzbot+a7067a757858ac8eb085@syzkaller.appspotmail.com
> Closes: https://lore.kernel.org/linux-mm/69ccb65b.050a0220.183828.003a.GAE@google.com/
> Cc: <stable@vger.kernel.org>
> Suggested-by: David Hildenbrand (Arm) <david@kernel.org>
> Signed-off-by: Lance Yang <lance.yang@linux.dev>
> ---
> 
> [ Backport note ]
> This patch is a follow-up fix for 8a8ca142a488 ("mm: migrate: requeue
> destination folio on deferred split queue"), which is currently only in
> mm-stable, and should be backported together with it.
> 
> Credit for this fix goes to David, thanks!
> 
>  mm/huge_memory.c | 12 +++++++-----
>  mm/migrate.c     | 18 +++++++++---------
>  2 files changed, 16 insertions(+), 14 deletions(-)
> 


Thanks for the fix! And sorry for introducing the bug in
migrate_folio_move() :)

So I am happy with the migrate_folio_move() change, it makes sense.

The goto next if folio is locked in deferred_split_scan() was actually
on purpose. The reasoning was that if the folio is locked, we consider
it as in use by someone and therefore we shouldnt split it. Eventhough
thp_underused() does a zero-filled check, the whole point of the shrinker
was to split THPs that are "not in use", and in my mind, locked folio
is a folio in use. So not sure about that change..


> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index ff9a42abd1b6..ac6d823e351f 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -4558,7 +4558,7 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,
>  				goto next;
>  		}
>  		if (!folio_trylock(folio))
> -			goto next;
> +			goto requeue;
>  		if (!split_folio(folio)) {
>  			did_split = true;
>  			if (underused)
> @@ -4569,11 +4569,13 @@ static unsigned long deferred_split_scan(struct shrinker *shrink,
>  next:
>  		if (did_split || !folio_test_partially_mapped(folio))
>  			continue;
> +requeue:
>  		/*
> -		 * Only add back to the queue if folio is partially mapped.
> -		 * If thp_underused returns false, or if split_folio fails
> -		 * in the case it was underused, then consider it used and
> -		 * don't add it back to split_queue.
> +		 * Add back partially mapped folios, or underused folios
> +		 * that we could not lock this round.  If thp_underused()
> +		 * returns false, or if split_folio() succeeds, or if
> +		 * split_folio() fails in the case it was underused, then
> +		 * consider it used and don't add it back to split_queue.
>  		 */
>  		fqueue = folio_split_queue_lock_irqsave(folio, &flags);
>  		if (list_empty(&folio->_deferred_list)) {
> diff --git a/mm/migrate.c b/mm/migrate.c
> index 05cb408846f2..8a64291ab5b4 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -1385,6 +1385,15 @@ static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
>  	if (rc)
>  		goto out;
>  
> +	/*
> +	 * Requeue the destination folio on the deferred split queue if
> +	 * the source was on the queue.  The source is unqueued in
> +	 * __folio_migrate_mapping(), so we recorded the state from
> +	 * before move_to_new_folio().
> +	 */
> +	if (src_deferred_split)
> +		deferred_split_folio(dst, src_partially_mapped);
> +
>  	/*
>  	 * When successful, push dst to LRU immediately: so that if it
>  	 * turns out to be an mlocked page, remove_migration_ptes() will
> @@ -1401,15 +1410,6 @@ static int migrate_folio_move(free_folio_t put_new_folio, unsigned long private,
>  	if (old_page_state & PAGE_WAS_MAPPED)
>  		remove_migration_ptes(src, dst, 0);
>  
> -	/*
> -	 * Requeue the destination folio on the deferred split queue if
> -	 * the source was on the queue.  The source is unqueued in
> -	 * __folio_migrate_mapping(), so we recorded the state from
> -	 * before move_to_new_folio().
> -	 */
> -	if (src_deferred_split)
> -		deferred_split_folio(dst, src_partially_mapped);
> -
>  out_unlock_both:
>  	folio_unlock(dst);
>  	folio_set_owner_migrate_reason(dst, reason);
> -- 
> 2.49.0
> 
>