Skip to contentSkip to conversation

hugetlb: force allocating surplus hugepages on mempolicy allowed nodes

PATCH In review Aristeu Rozanski 4 participants 8 messages Jun 21, 2024 – Jul 1, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 hugetlb: force allocating surplus hugepages on mempolicy allowed nodes

allowed_mems_nr() only accounts for the number of free hugepages in the nodes the current process belongs to. In case one or more of the requested surplus hugepages are allocated in a different node, the whole allocation will fail due allowed_mems_nr() returning a lower value.

So allocate surplus hugepages in one of the nodes the current process belongs to.

Easy way to reproduce this issue is to use a 2+ NUMA nodes system:

		# echo 0 >/proc/sys/vm/nr_hugepages
		# echo 1 >/proc/sys/vm/nr_overcommit_hugepages
		# numactl -m0 ./tools/testing/selftests/mm/map_hugetlb 2

will eventually fail when the hugepage ends up allocated in a different node.

Cc Muchun Song <[email protected]>
Cc Andrew Morton <[email protected]>
---
 mm/hugetlb.c |   46 +++++++++++++++++++++++++++-------------------
 1 file changed, 27 insertions(+), 19 deletions(-)
mm/hugetlb.c +27 -19
@@ -2618,6 +2618,23 @@ struct folio *alloc_hugetlb_folio_nodema
 	return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
 }
 
+static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
+{
+#ifdef CONFIG_NUMA
+	struct mempolicy *mpol = get_task_policy(current);
+
+	/*
+	 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
+	 * (from policy_nodemask) specifically for hugetlb case
+	 */
+	if (mpol->mode == MPOL_BIND &&
+		(apply_policy_zone(mpol, gfp_zone(gfp)) &&
+		 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
+		return &mpol->nodes;
+#endif
+	return NULL;
+}
+
 /*
  * Increase the hugetlb pool such that it can accommodate a reservation
  * of size 'delta'.
@@ -2631,6 +2648,8 @@ static int gather_surplus_pages(struct h
 	long i;
 	long needed, allocated;
 	bool alloc_ok = true;
+	int node;
+	nodemask_t *mbind_nodemask = policy_mbind_nodemask(htlb_alloc_mask(h));
 
 	lockdep_assert_held(&hugetlb_lock);
 	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
@@ -2645,8 +2664,14 @@ allocated = 0;
 retry:
 	spin_unlock_irq(&hugetlb_lock);
 	for (i = 0; i < needed; i++) {
-		folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
-				NUMA_NO_NODE, NULL);
+		for_each_node_mask(node, cpuset_current_mems_allowed) {
+			if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) {
+				folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
+						node, NULL);
+				if (folio)
+					break;
+			}
+		}
 		if (!folio) {
 			alloc_ok = false;
 			break;
@@ -4876,23 +4901,6 @@ default_hstate_max_huge_pages = 0;
 }
 __setup("default_hugepagesz=", default_hugepagesz_setup);
 
-static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
-{
-#ifdef CONFIG_NUMA
-	struct mempolicy *mpol = get_task_policy(current);
-
-	/*
-	 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
-	 * (from policy_nodemask) specifically for hugetlb case
-	 */
-	if (mpol->mode == MPOL_BIND &&
-		(apply_policy_zone(mpol, gfp_zone(gfp)) &&
-		 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
-		return &mpol->nodes;
-#endif
-	return NULL;
-}
-
 static unsigned int allowed_mems_nr(struct hstate *h)
 {
 	int node;
maintainer
#1 lore raw
Quoted 7 lines · On Fri, 21 Jun 2024 15:00:50 -0400 Aristeu Rozanski <[email protected]> wrote:

allowed_mems_nr() only accounts for the number of free hugepages in the nodes the current process belongs to. In case one or more of the requested surplus hugepages are allocated in a different node, the whole allocation will fail due allowed_mems_nr() returning a lower value.

So allocate surplus hugepages in one of the nodes the current process belongs to.

Quoted 8 lines

Easy way to reproduce this issue is to use a 2+ NUMA nodes system:

		# echo 0 >/proc/sys/vm/nr_hugepages
		# echo 1 >/proc/sys/vm/nr_overcommit_hugepages
		# numactl -m0 ./tools/testing/selftests/mm/map_hugetlb 2

will eventually fail when the hugepage ends up allocated in a different node.

Should we backport this into -stable kernels?

maintainer
#2 lore raw
Quoted 11 lines · On Fri, 21 Jun 2024 17:56:09 -0700 Andrew Morton <[email protected]>…
Quoted 9 lines · On Fri, 21 Jun 2024 15:00:50 -0400 Aristeu Rozanski <[email protected]> wrote:

allowed_mems_nr() only accounts for the number of free hugepages in the nodes the current process belongs to. In case one or more of the requested surplus hugepages are allocated in a different node, the whole allocation will fail due allowed_mems_nr() returning a lower value.

So allocate surplus hugepages in one of the nodes the current process belongs to.

Easy way to reproduce this issue is to use a 2+ NUMA nodes system:

Quoted 8 lines
Quoted 6 lines
		# echo 0 >/proc/sys/vm/nr_hugepages
		# echo 1 >/proc/sys/vm/nr_overcommit_hugepages
		# numactl -m0 ./tools/testing/selftests/mm/map_hugetlb 2

will eventually fail when the hugepage ends up allocated in a different node.

Should we backport this into -stable kernels?

Please?

Aristeu, a description of the userspace-visible effects of the bug will really help things along here.

#3 lore raw
Quoted 2 lines · Andrew Morton

Aristeu, a description of the userspace-visible effects of the bug will really help things along here.

Will update with the reproducer and resubmit.

Signature · 1 line
Aristeu
#4 lore raw
Quoted 5 lines · Aristeu Rozanski
Quoted 2 lines · Andrew Morton

Aristeu, a description of the userspace-visible effects of the bug will really help things along here.

Will update with the reproducer and resubmit.

Could you look into this warning[1] and include the fix (if necessary) when you resubmit?

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

Quoted 2 lines
#6 lore raw

PATCH v2 1/1 hugetlb: force allocating surplus hugepages on mempolicy allowed nodes

In reply to Vishal Moola

When trying to allocate a hugepage with none reserved ones free, it may be allowed in case a number of overcommit hugepages was configured (using /proc/sys/vm/nr_overcommit_hugepages) and that number wasn't reached. This allows for a behavior of having extra hugepages allocated dynamically, if there're resources for it. Some sysadmins even prefer not reserving any hugepages and setting a big number of overcommit hugepages.

But while attempting to allocate overcommit hugepages in a multi node system (either NUMA or mempolicy/cpuset) said allocations might randomly fail even when there're resources available for the allocation.

This happens due allowed_mems_nr() only accounting for the number of free hugepages in the nodes the current process belongs to and the surplus hugepage allocation is done so it can be allocated in any node. In case one or more of the requested surplus hugepages are allocated in a different node, the whole allocation will fail due allowed_mems_nr() returning a lower value.

So allocate surplus hugepages in one of the nodes the current process belongs to.

Easy way to reproduce this issue is to use a 2+ NUMA nodes system:

	# echo 0 >/proc/sys/vm/nr_hugepages
	# echo 1 >/proc/sys/vm/nr_overcommit_hugepages
	# numactl -m0 ./tools/testing/selftests/mm/map_hugetlb 2

Repeating the execution of map_hugetlb test application will eventually fail when the hugepage ends up allocated in a different node.

v2: - attempt to make the description more clear
- prevent unitialized usage of folio in case current process isn't part of any nodes with memory

Cc Vishal Moola <[email protected]>
Cc David Hildenbrand <[email protected]>
Cc Aristeu Rozanski <[email protected]>
Cc Muchun Song <[email protected]>
Cc Andrew Morton <[email protected]>
---
 mm/hugetlb.c |   47 ++++++++++++++++++++++++++++-------------------
 1 file changed, 28 insertions(+), 19 deletions(-)
mm/hugetlb.c +28 -19
@@ -2618,6 +2618,23 @@ struct folio *alloc_hugetlb_folio_nodema
 	return alloc_migrate_hugetlb_folio(h, gfp_mask, preferred_nid, nmask);
 }
 
+static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
+{
+#ifdef CONFIG_NUMA
+	struct mempolicy *mpol = get_task_policy(current);
+
+	/*
+	 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
+	 * (from policy_nodemask) specifically for hugetlb case
+	 */
+	if (mpol->mode == MPOL_BIND &&
+		(apply_policy_zone(mpol, gfp_zone(gfp)) &&
+		 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
+		return &mpol->nodes;
+#endif
+	return NULL;
+}
+
 /*
  * Increase the hugetlb pool such that it can accommodate a reservation
  * of size 'delta'.
@@ -2631,6 +2648,8 @@ static int gather_surplus_pages(struct h
 	long i;
 	long needed, allocated;
 	bool alloc_ok = true;
+	int node;
+	nodemask_t *mbind_nodemask = policy_mbind_nodemask(htlb_alloc_mask(h));
 
 	lockdep_assert_held(&hugetlb_lock);
 	needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
@@ -2645,8 +2664,15 @@ allocated = 0;
 retry:
 	spin_unlock_irq(&hugetlb_lock);
 	for (i = 0; i < needed; i++) {
-		folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
-				NUMA_NO_NODE, NULL);
+		folio = NULL;
+		for_each_node_mask(node, cpuset_current_mems_allowed) {
+			if (!mbind_nodemask || node_isset(node, *mbind_nodemask)) {
+				folio = alloc_surplus_hugetlb_folio(h, htlb_alloc_mask(h),
+						node, NULL);
+				if (folio)
+					break;
+			}
+		}
 		if (!folio) {
 			alloc_ok = false;
 			break;
@@ -4876,23 +4902,6 @@ default_hstate_max_huge_pages = 0;
 }
 __setup("default_hugepagesz=", default_hugepagesz_setup);
 
-static nodemask_t *policy_mbind_nodemask(gfp_t gfp)
-{
-#ifdef CONFIG_NUMA
-	struct mempolicy *mpol = get_task_policy(current);
-
-	/*
-	 * Only enforce MPOL_BIND policy which overlaps with cpuset policy
-	 * (from policy_nodemask) specifically for hugetlb case
-	 */
-	if (mpol->mode == MPOL_BIND &&
-		(apply_policy_zone(mpol, gfp_zone(gfp)) &&
-		 cpuset_nodemask_valid_mems_allowed(&mpol->nodes)))
-		return &mpol->nodes;
-#endif
-	return NULL;
-}
-
 static unsigned int allowed_mems_nr(struct hstate *h)
 {
 	int node;
#7 lore raw
In reply to Andrew Morton

Hi Andrew,

Quoted 1 line · Andrew Morton

Should we backport this into -stable kernels?

sorry, stupid email server.

Yes, that would qualify for -stable kernels.

Signature · 1 line
Aristeu