[PATCH] mm/migrate: move NR_FILE_THPS on cross-node migration

[PATCH] mm/migrate: move NR_FILE_THPS on cross-node migration

[Ridong Chen]

From: Ridong Chen <chenridong@xiaomi.com>

When a page cache folio is migrated to another node,
__folio_migrate_mapping() moves the per-node and per-memcg counters
that are keyed off the folio's node: NR_FILE_PAGES, NR_SHMEM,
NR_SHMEM_THPS, NR_SWAPCACHE and NR_FILE_DIRTY.  NR_FILE_THPS is not
moved, although __filemap_add_folio() charges it to the source node and
filemap_unaccount_folio() takes it off whatever node the folio lives on
when it is removed.

Commit 0b52c420350e ("mm: fix shmem THP counters on migration") added
the NR_SHMEM_THPS case but not its file-backed counterpart.  Now that
ext4, xfs, btrfs and erofs can all have PMD-sized page cache folios,
migrating them via NUMA balancing, mbind(), migratepages, demotion or
memory offlining leaks FileHugePages on the source node and underflows
it on the destination, in both /sys/devices/system/node/nodeN/meminfo
and memory.numa_stat.  The global value is unaffected, which is why this
went unnoticed.

Reproducer, on a 2-node QEMU guest with THP set to "always":

  1. read a 256M ext4 file through a MADV_HUGEPAGE mapping while bound
     to node 0, so its page cache is 128 PMD-sized folios on node 0;
  2. move the mapping to node 1 with mbind(MPOL_MF_MOVE);
  3. unmap the file and drop its page cache.

nr_file_hugepages in /sys/devices/system/node/nodeN/vmstat after each
step (in units of PMD-sized folios):

                    without this patch     with this patch
                     node0     node1       node0     node1
  1. cached           128         0         128         0
  2. migrated         128         0           0       128
  3. dropped          128         0 (*)       0         0

  (*) the counter has underflowed to -128; negative values read as 0.

Without the patch the 128 folios stay accounted to node 0 forever, and
every further round adds another 128.  The same happens to the per-memcg
per-node counter: after step 3, memory.stat of the cgroup shows
"file_thp 0", while memory.numa_stat still shows
"file_thp N0=268435456 N1=0".  With the patch both are 0.

Move NR_FILE_THPS together with NR_FILE_PAGES, using the same condition
filemap_unaccount_folio() uses to take it off, so that THPs in the swap
cache are left alone.

Fixes: d68eccad3706 ("mm/filemap: Allow large folios to be added to the page cache")
Signed-off-by: Ridong Chen <chenridong@xiaomi.com>
---
Note: prepared with help from an LLM-based assistant (Kiro, claude-opus-5.5).

 mm/migrate.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/mm/migrate.c b/mm/migrate.c
index 7bdcdb57652f..d55f576a2579 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -700,6 +700,10 @@ static int __folio_migrate_mapping(struct address_space *mapping,
 				mod_lruvec_state(old_lruvec, NR_SHMEM_THPS, -nr);
 				mod_lruvec_state(new_lruvec, NR_SHMEM_THPS, nr);
 			}
+		} else if (!folio_test_swapbacked(folio) &&
+			   folio_test_pmd_mappable(folio)) {
+			mod_lruvec_state(old_lruvec, NR_FILE_THPS, -nr);
+			mod_lruvec_state(new_lruvec, NR_FILE_THPS, nr);
 		}
 #ifdef CONFIG_SWAP
 		if (folio_test_swapcache(folio)) {

base-commit: bd1ecbb61afefc6a8b876ccbc31e9228b44e9eab
-- 
2.34.1