Skip to contentSkip to conversation

hugetlb: fix subpool release race

PATCH In review Yichong Chen 3 participants 3 messages Jul 20

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

PATCH 1/1 hugetlb: fix subpool release race

unlock_or_release_subpool() drops spool->lock before checking whether the subpool can be freed. However, subpool_is_free() reads fields that are updated under spool->lock, including count, used_hpages and rsv_hpages.

Another thread can update those fields before the first thread evaluates subpool_is_free(), allowing both threads to observe the final freeable state and release the subpool.

Make the free decision while still holding spool->lock. Keep the actual hugetlb_acct_memory() and kfree() calls after dropping the lock.

Fixes 90481622d757 ("hugepages: fix use after free bug in "quota" handling")
---
 mm/hugetlb.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
mm/hugetlb.c +4 -2
index e319c6a00555..46ab702c0fc0 100644
@@ -140,12 +140,14 @@ static inline bool subpool_is_free(struct hugepage_subpool *spool)
 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
 						unsigned long irq_flags)
 {
-	spin_unlock_irqrestore(&spool->lock, irq_flags);
+	bool free_subpool = subpool_is_free(spool);
 
 	/* If no pages are used, and no other handles to the subpool
 	 * remain, give up any reservations based on minimum size and
 	 * free the subpool */
-	if (subpool_is_free(spool)) {
+	spin_unlock_irqrestore(&spool->lock, irq_flags);
+
+	if (free_subpool) {
 		if (spool->min_hpages != -1)
 			hugetlb_acct_memory(spool->hstate,
 						-spool->min_hpages);
Signature · 1 line
2.51.0
Reviewed-by on v1 1/1
#1 lore raw

Hello Yichong, thank you for the patch!

Quoted 7 lines

unlock_or_release_subpool() drops spool->lock before checking whether the subpool can be freed. However, subpool_is_free() reads fields that are updated under spool->lock, including count, used_hpages and rsv_hpages.

Another thread can update those fields before the first thread evaluates subpool_is_free(), allowing both threads to observe the final freeable state and release the subpool.

I don't think that the race you are mentioning can be hit in practice (i.e. there is no user-visible change for this) since two threads cannot actually observe the freeable state at the same time.

subpool_is_free() does an early exit on spool->count so it will never reach the inside if block and have two threads try to decrement the spool->hstate and free the spool twice. I'm wondering if we need the Fixes tag in this case.

I do think that this is not obvious though, so having the subpool_is_free check inside the locked section makes sense to me too.

Have a great day!
Joshua

Quoted 32 lines

Make the free decision while still holding spool->lock. Keep the actual hugetlb_acct_memory() and kfree() calls after dropping the lock.

Fixes 90481622d757 ("hugepages: fix use after free bug in "quota" handling")
---
 mm/hugetlb.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
mm/hugetlb.c +4 -2
index e319c6a00555..46ab702c0fc0 100644
@@ -140,12 +140,14 @@ static inline bool subpool_is_free(struct hugepage_subpool *spool)
 static inline void unlock_or_release_subpool(struct hugepage_subpool *spool,
 						unsigned long irq_flags)
 {
-	spin_unlock_irqrestore(&spool->lock, irq_flags);
+	bool free_subpool = subpool_is_free(spool);
 
 	/* If no pages are used, and no other handles to the subpool
 	 * remain, give up any reservations based on minimum size and
 	 * free the subpool */
-	if (subpool_is_free(spool)) {
+	spin_unlock_irqrestore(&spool->lock, irq_flags);
+
+	if (free_subpool) {
 		if (spool->min_hpages != -1)
 			hugetlb_acct_memory(spool->hstate,
 						-spool->min_hpages);
Signature · 1 line
2.51.0
maintainer
#2 lore raw
Quoted 19 lines · On Mon, 20 Jul 2026 07:31:29 -0700 Joshua Hahn <[email protected]> wrote:

Hello Yichong, thank you for the patch!

Quoted 7 lines

unlock_or_release_subpool() drops spool->lock before checking whether the subpool can be freed. However, subpool_is_free() reads fields that are updated under spool->lock, including count, used_hpages and rsv_hpages.

Another thread can update those fields before the first thread evaluates subpool_is_free(), allowing both threads to observe the final freeable state and release the subpool.

I don't think that the race you are mentioning can be hit in practice (i.e. there is no user-visible change for this) since two threads cannot actually observe the freeable state at the same time.

subpool_is_free() does an early exit on spool->count so it will never reach the inside if block and have two threads try to decrement the spool->hstate and free the spool twice. I'm wondering if we need the Fixes tag in this case.

Quoted 4 lines

I do think that this is not obvious though, so having the subpool_is_free check inside the locked section makes sense to me too.

Thanks. Yichong, can you please consider Joshua's observations and perhaps resend with a more accurate changelog?