From: Kairui Song Only anon split needs the anon_vma, and it only needs it to unmap and remap. Move the folio_get_anon_vma()/anon_vma_lock_write() pair out of __folio_split() into the anon helper next to the folio_mapped() check that already gates unmap_folio(). This makes the anon_vma conditional on folio_mapped(), which is a behaviour change but should be fine. folio_get_anon_vma() returns NULL whenever !folio_mapped(), so an anon folio with folio_mapcount() == 0 used to get -EBUSY from split_huge_page() and is now split instead. A realistic case is a THP that has been fully swapped out and is still in the swap cache: swap PTEs do not contribute mapcount, so it is !folio_mapped() but still alive. That should be safe and right to have because: - folio_ref_freeze() below still rejects a folio that picked up any reference, a mapping or a GUP pin, in the meantime. - A parallel split is excluded by the folio lock. The anon_vma write lock was added to serialize split in commit 062f1af2170a ("mm: thp: acquire the anon_vma rwsem for write during split"), when split_huge_page() did not hold the folio lock throughout. commit e9b61f19858a ("thp: reintroduce split_huge_page()") later made the folio lock a caller requirement and added the folio_ref_freeze() scheme, so that has been covered ever since. - Unmapped path is already exercised by folio_split_unmapped(), and the swap cache split already runs well for a partially swapped-out mapped THP. - A !folio_mapped() folio cannot become mapped meanwhile: mapping it requires the folio lock. For mapped folios the anon_vma write lock is now released before __folio_split() unlocks the after-split sub-folios, where previously it was held across that loop; that window is harmless as the sub-folios stay folio-locked and referenced until it. Reviewed-by: Zi Yan Signed-off-by: Kairui Song --- mm/huge_memory.c | 58 ++++++++++++++++++++++++++++++-------------------------- 1 file changed, 31 insertions(+), 27 deletions(-) diff --git a/mm/huge_memory.c b/mm/huge_memory.c index 8fc1dd5b8354..a4c2e38f4a6d 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -4029,16 +4029,38 @@ static int __folio_freeze_split_anon(struct folio *folio, struct swap_cluster_info *ci = NULL; int old_order = folio_order(folio); struct folio *new_folio, *next; + struct anon_vma *anon_vma = NULL; enum ttu_flags ttu_flags = 0; struct lruvec *lruvec; - bool need_remap = false; int ret = 0; + /* + * Unmap/remap needs the anon_vma, so we first take a reference on + * it to prevent it from disappearing, and lock it for write here, + * letting unmap_folio() walk the rmap with TTU_RMAP_LOCKED. + * + * folio_mapped() is not stable here, but it can only change in + * one direction while the folio is locked. The mapcount can drop + * to zero at any time, zap_pte_range() takes no folio lock. It + * cannot go up: swapin, migration and uffd move all lock the folio + * before mapping it, and fork only copies PTEs that already exist. + * + * So if we see the folio mapped, the worst case is an empty rmap + * walk. If we see it unmapped, it stays unmapped and needs neither + * the reference nor the lock. Anything else needs a reference + * first and folio_ref_freeze() below catches it. + * + * Note a swapped-out THP counts as unmapped here as swap PTEs do + * not contribute mapcount, and they are splittable. + */ if (folio_mapped(folio)) { - need_remap = true; + anon_vma = folio_get_anon_vma(folio); + if (!anon_vma) + return -EBUSY; + anon_vma_lock_write(anon_vma); ret = unmap_folio(folio); if (ret) - return ret; + goto out_unlock; } local_irq_disable(); @@ -4096,11 +4118,16 @@ static int __folio_freeze_split_anon(struct folio *folio, swap_cluster_unlock(ci); out_no_split: local_irq_enable(); - if (need_remap) { + if (anon_vma) { if (!ret && !folio_is_device_private(folio)) ttu_flags = TTU_USE_SHARED_ZEROPAGE; remap_anon_folio(folio, 1 << old_order, ttu_flags); } +out_unlock: + if (anon_vma) { + anon_vma_unlock_write(anon_vma); + put_anon_vma(anon_vma); + } return ret; } @@ -4295,7 +4322,6 @@ static int __folio_split(struct folio *folio, unsigned int new_order, struct folio *end_folio = folio_next(folio); bool is_anon = folio_test_anon(folio); struct mem_cgroup *memcg, *old_memcg; - struct anon_vma *anon_vma = NULL; int old_order = folio_order(folio); struct folio *new_folio, *next; int ret; @@ -4326,23 +4352,6 @@ static int __folio_split(struct folio *folio, unsigned int new_order, memcg = get_mem_cgroup_from_folio(folio); old_memcg = set_active_memcg(memcg); - if (is_anon) { - /* - * The caller does not necessarily hold an mmap_lock that would - * prevent the anon_vma disappearing so we first we take a - * reference to it and then lock the anon_vma for write. This - * is similar to folio_lock_anon_vma_read except the write lock - * is taken to serialise against parallel split or collapse - * operations. - */ - anon_vma = folio_get_anon_vma(folio); - if (!anon_vma) { - ret = -EBUSY; - goto out; - } - anon_vma_lock_write(anon_vma); - } - if (is_anon) ret = __folio_freeze_split_anon(folio, new_order, split_at, true, list, split_type); @@ -4369,11 +4378,6 @@ static int __folio_split(struct folio *folio, unsigned int new_order, free_folio_and_swap_cache(new_folio); } - if (anon_vma) { - anon_vma_unlock_write(anon_vma); - put_anon_vma(anon_vma); - } -out: /* restore to caller's old_memcg */ set_active_memcg(old_memcg); mem_cgroup_put(memcg); -- 2.55.0