Re: [PATCH 2/2] mm/migrate: allocate destination folios before locking a batch
From: Gregory Price (Meta)
Date: Sun Nov 22 2026 - 14:02:43 EST
On 2026-11-22 13:42 +0100, David Hildenbrand (Arm) wrote:
> Migration locks folios without holding a reference on their inode.
> migrate_pages_batch() keeps a whole batch locked while allocating the
> destinations of the next folios. If such an allocation ends up in direct
> reclaim with __GFP_FS, reclaim can evict an inode through the dcache
> shrinker, and wait in truncate_inode_pages_final() for a folio that we
> hold locked:
>
> migrate_pages_batch()
> migrate_folio_unmap(A)
> folio_trylock(A)
> migrate_folio_unmap(B)
> get_new_folio()
> try_to_free_pages()
> super_cache_scan()
> prune_dcache_sb()
> iput_final()
> evict()
> truncate_inode_pages_final()
> folio_lock(A)
>
> This requires that nothing else holds a reference on the inode, and can
> be triggered most easily with FUSE, which evicts inodes on the last
> reference.
>
> Consequently, allocations under the folio lock are dangerous if we
> cannot prevent that scenario through an inode reference; and we want to
> avoid grabbing inode references, because we might then drop the last
> one, and would have to evict the inode from migration context, ideally
> without holding any folio locks.
>
> Allocating in a NOFS scope would be one option, but it would have rather
> bad side effects, such as direct compaction not migrating file folios
> and the OOM killer not being invoked.
>
> So let's fix it by allocating all destination folios of a batch before
> locking any source folio. Now, we no longer allocate destination folios
> while holding folio locks, which also avoids holding these locks during
> direct reclaim or compaction for these allocations: an improvement on
> its own.
>
> We keep the destination folio of a folio that we have to retry, instead
> of freeing it and allocating a new one. If we would have to allocate
> while already holding folio locks, which can happen when retrying a
> failed split for longterm pinning, we leave the folio to the synchronous
> fallback in migrate_pages_sync() instead, which migrates folios one by
> one.
>
> Note that nfs_migrate_folio() still writes back the folio while holding
> its lock in MIGRATE_SYNC mode, which allocates memory. As NFS only
> evicts stale or unlinked inodes on the last reference, this is unlikely
> to deadlock; this must be sorted out in NFS code independently.
>
> This has medium urgency.
>
> Fixes: 5dfab109d519 ("migrate_pages: batch _unmap and _move")
> Cc: stable@xxxxxxxxxxxxxxx
> Reported-by: Chris Mason <mason@xxxxxxxxxx>
> Signed-off-by: David Hildenbrand (Arm) <david@xxxxxxxxxx>
I was able to reproduce the issue and validate that this
change fixes it. I used FUSE to do multiple allocations and
migrations while triggering the shrinkers to reclaim dentries /
cause FUSE to reclaim the inodes.
In the base we end up deadlocked, with patch 1 and 2, the
migrations succeed but the reclaimed folios will just end up
freed at the end of it all anyway. Some migration work gets
wasted either way, but that's fine.
I agree this is better than the NOFS option.
Tested-by: Gregory Price (Meta) <gourry@xxxxxxxxxx>
Reviewed-by: Gregory Price (Meta) <gourry@xxxxxxxxxx>
--
~Gregory