Skip to contentSkip to conversation

hugetlb: Don't restore vmemmap of non-HVOed folios on bulk restore error

PATCH In review James Houghton 2 participants 2 messages Nov 16 – Nov 17

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 hugetlb: Don't restore vmemmap of non-HVOed folios on bulk restore error

When hugetlb_vmemmap_restore_folios() fails, it stops processing the list, so folio_list may still contain folios that were never vmemmap optimized, e.g. because HVO failed when they were allocated. Their hugetlb flag has already been cleared by remove_hugetlb_folio().

If no folios could be restored, bulk_vmemmap_restore_error() then calls hugetlb_vmemmap_restore_folio() on every folio left on folio_list, including those, which trips the !folio_test_hugetlb() check in __hugetlb_vmemmap_restore_folio():

  page dumped because: VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio))
  WARNING: mm/hugetlb_vmemmap.c:503 at __hugetlb_vmemmap_restore_folio+0x270/0x300
  ...
   __hugetlb_vmemmap_restore_folio+0x270/0x300 (P)
   hugetlb_vmemmap_restore_folio+0x1c/0x30
   update_and_free_pages_bulk+0x2c0/0x368
   __nr_hugepages_store_common+0x2e8/0x3f0

Otherwise this is benign: the restore is a no-op returning success for non-optimized folios, which are then freed as intended.

Skip the restore for folios that are not vmemmap optimized; they can be freed right away.

This is very unlikely to happen without fault injection, as it requires both HVO and a subsequent vmemmap restore to fail.

Fixes cfb8c75099db ("hugetlb: perform vmemmap restoration on a list of pages")

---
This patch was originally part of the series that enables the HugeTLB vmemmap optimization (HVO) on arm64, here[1]. It has been modified given the feedback received at [1]. Based on top of mm-unstable.

[1] https://lore.kernel.org/linux-mm/[email protected]/

---
 mm/hugetlb.c | 29 ++++++++++++++++++-----------
 1 file changed, 18 insertions(+), 11 deletions(-)
mm/hugetlb.c +18 -11
index cb2de24affdc..6e9f3fb04a3c 100644
@@ -1628,21 +1628,28 @@ static void bulk_vmemmap_restore_error(struct hstate *h,
 		 * If are able to restore vmemmap and free one hugetlb page, we
 		 * quit processing the list to retry the bulk operation.
 		 */
-		list_for_each_entry_safe(folio, t_folio, folio_list, lru)
-			if (hugetlb_vmemmap_restore_folio(h, folio)) {
-				list_del(&folio->lru);
-				spin_lock_irq(&hugetlb_lock);
-				add_hugetlb_folio(h, folio, true);
-				spin_unlock_irq(&hugetlb_lock);
-			} else {
-				list_del(&folio->lru);
+		list_for_each_entry_safe(folio, t_folio, folio_list, lru) {
+			list_del(&folio->lru);
+			/*
+			 * remove_hugetlb_folio() already cleared the hugetlb flag
+			 * for folios whose vmemmap was not optimized, so they can
+			 * be freed without restoring vmemmap.
+			 */
+			if (folio_test_hugetlb(folio)) {
+				if (hugetlb_vmemmap_restore_folio(h, folio)) {
+					spin_lock_irq(&hugetlb_lock);
+					add_hugetlb_folio(h, folio, true);
+					spin_unlock_irq(&hugetlb_lock);
+					continue;
+				}
 				spin_lock_irq(&hugetlb_lock);
 				__folio_clear_hugetlb(folio);
 				spin_unlock_irq(&hugetlb_lock);
-				update_and_free_hugetlb_folio(h, folio, false);
-				cond_resched();
-				break;
 			}
+			update_and_free_hugetlb_folio(h, folio, false);
+			cond_resched();
+			break;
+		}
 	}
 }
 

base-commit: e1f34fb16e0236c4e60b5f392f1cf5aee76f7d0f

Signature · 1 line
2.56.0.679.gc527c65306-goog
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 25 lines · James Houghton

When hugetlb_vmemmap_restore_folios() fails, it stops processing the list, so folio_list may still contain folios that were never vmemmap optimized, e.g. because HVO failed when they were allocated. Their hugetlb flag has already been cleared by remove_hugetlb_folio().

If no folios could be restored, bulk_vmemmap_restore_error() then calls hugetlb_vmemmap_restore_folio() on every folio left on folio_list, including those, which trips the !folio_test_hugetlb() check in __hugetlb_vmemmap_restore_folio():

 page dumped because: VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio))
 WARNING: mm/hugetlb_vmemmap.c:503 at __hugetlb_vmemmap_restore_folio+0x270/0x300
 ...
  __hugetlb_vmemmap_restore_folio+0x270/0x300 (P)
  hugetlb_vmemmap_restore_folio+0x1c/0x30
  update_and_free_pages_bulk+0x2c0/0x368
  __nr_hugepages_store_common+0x2e8/0x3f0

Otherwise this is benign: the restore is a no-op returning success for non-optimized folios, which are then freed as intended.

Quoted 8 lines

Skip the restore for folios that are not vmemmap optimized; they can be freed right away.

This is very unlikely to happen without fault injection, as it requires both HVO and a subsequent vmemmap restore to fail.

Fixes cfb8c75099db ("hugetlb: perform vmemmap restoration on a list of pages")

Thanks.