From 92635a3ffe8bb48304f50625dd591a27d7505698 Mon Sep 17 00:00:00 2001 From: Hongfu Li Date: Wed, 23 Sep 2026 10:05:15 +0800 Subject: [PATCH] mm/hugetlb: account migration target folio in per-node NR_HUGETLB vmstat Patch series "mm: fix hugetlb NR_HUGETLB accounting on folio migration", v2. The hugeTLB counters added by 05d4532b60e3 ("memcg/hugetlb: add hugeTLB counters to memcg") are maintained in two per-node places: - the per-node vmstat counter NR_HUGETLB, exposed as nr_hugetlb in /proc/vmstat; - the per-node memcg lruvec stat, exposed via memory.numa_stat. Both are accounted against the folio's node, and both drift when a hugetlb folio is migrated, though in different ways. A migration target folio is allocated by alloc_hugetlb_folio_nodemask() and inherits the old folio's state without ever being accounted, while the old folio is freed right after and its free is accounted. That alone loses vmstat accounting: the target node has no matching increment for the decrement on the old node, so /proc/vmstat's nr_hugetlb shrinks by nr_pages per migration. Patch 1 accounts the folio where it is obtained, so the increment pairs with the free in free_huge_folio() on the successful as well as the failed migration path. The memfd page cache preallocation helper has the same asymmetry and is fixed in the same patch. The per-node lruvec stat breaks differently. mem_cgroup_migrate() moves the charge to the new folio and drops the old folio's memcg data, so the old folio's free right after migration skips the memcg per-node decrement; the count stays attributed to the old node for the rest of the charge's life, and the target folio never gets an increment on its new node. Patch 2 moves that per-node accounting along with the charge, in move_hugetlb_state(). This patch (of 2): The NR_HUGETLB vmstat counter is maintained per folio's node: incremented when a huge page is handed to a user via hugetlb_alloc_folio() and decremented when it is returned to the pool via free_huge_folio(). A folio obtained by alloc_hugetlb_folio_nodemask() never goes through hugetlb_alloc_folio(), so it is never accounted, while its free always is. For a migration target this means the target node gets no matching increment for the decrement on the old node, so the global nr_hugetlb in /proc/vmstat drops by nr_pages for each migration. The same asymmetry affects the failed migration path, which frees the target again right away, and the temporary folio hugetlb_mfill_atomic_pte() takes from the same helper. alloc_hugetlb_folio_reserve(), used to preallocate the memfd page cache folios, has the same asymmetry: the folio is handed to a user without being accounted, while its free is accounted through free_huge_folio(). Account the folio where it is obtained, in alloc_hugetlb_folio_nodemask() and alloc_hugetlb_folio_reserve(), so that the increment pairs with the decrement in free_huge_folio(): a successful migration hands the folio to a user, a failed one frees it again. Link: https://lore.kernel.org/20260923-for-hugetlb_state3-v2-0-e8a36245bfab@kylinos.cn Link: https://lore.kernel.org/20260923-for-hugetlb_state3-v2-1-e8a36245bfab@kylinos.cn Fixes: 05d4532b60e3 ("memcg/hugetlb: add hugeTLB counters to memcg") Signed-off-by: Hongfu Li Signed-off-by: Andrew Morton Tested-by: Joshua Hahn Reviewed-by: Joshua Hahn Acked-by: Muchun Song Acked-by: Oscar Salvador Cc: David Hildenbrand Cc: Shakeel Butt Cc: Michal Hocko Cc: Johannes Weiner Cc: Nhat Pham Cc: Michal Hocko Cc: Roman Gushchin Cc: Chris Down Cc: --- mm/hugetlb.c | 35 +++++++++++++++++++++++------------ 1 file changed, 23 insertions(+), 12 deletions(-) diff --git a/mm/hugetlb.c b/mm/hugetlb.c index a949b446cbea..4a07eada7779 100644 --- a/mm/hugetlb.c +++ b/mm/hugetlb.c @@ -2206,6 +2206,11 @@ struct folio *alloc_hugetlb_folio_reserve(struct hstate *h, int preferred_nid, } spin_unlock_irq(&hugetlb_lock); + + if (folio) + lruvec_stat_mod_folio(folio, NR_HUGETLB, + folio_nr_pages(folio)); + return folio; } @@ -2213,24 +2218,30 @@ struct folio *alloc_hugetlb_folio_reserve(struct hstate *h, int preferred_nid, struct folio *alloc_hugetlb_folio_nodemask(struct hstate *h, int preferred_nid, nodemask_t *nmask, gfp_t gfp_mask, bool allow_alloc_fallback) { - spin_lock_irq(&hugetlb_lock); - if (available_huge_pages(h)) { - struct folio *folio; + struct folio *folio = NULL; + spin_lock_irq(&hugetlb_lock); + if (available_huge_pages(h)) folio = dequeue_hugetlb_folio_nodemask(h, gfp_mask, preferred_nid, nmask); - if (folio) { - spin_unlock_irq(&hugetlb_lock); - return folio; - } - } spin_unlock_irq(&hugetlb_lock); - /* We cannot fallback to other nodes, as we could break the per-node pool. */ - if (!allow_alloc_fallback) - gfp_mask |= __GFP_THISNODE; + if (!folio) { + /* + * We cannot fallback to other nodes, as we could break the + * per-node pool. + */ + if (!allow_alloc_fallback) + gfp_mask |= __GFP_THISNODE; - return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask); + folio = alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, + nmask); + } + + if (folio) + lruvec_stat_mod_folio(folio, NR_HUGETLB, folio_nr_pages(folio)); + + return folio; } static nodemask_t *policy_mbind_nodemask(gfp_t gfp)