mm/memfd: don't unreserve hugetlb pages on -EEXIST in error path

[PATCH] mm/memfd: don't unreserve hugetlb pages on -EEXIST in error path

Posted by Hongfu Li 3 weeks, 5 days ago

From: Hongfu Li <lihongfu@kylinos.cn>

If hugetlb_add_to_page_cache() fails with -EEXIST, a concurrent fault has
already instantiated the folio in the page cache, and the reservation now
belongs to that folio. Calling hugetlb_unreserve_pages() in that case
incorrectly removes the region backing the cached folio, and a later
truncate or inode eviction passes a negative (chg - freed) into
hugepage_subpool_put_pages(), corrupting subpool and resv_huge_pages
accounting.

Skip the unreserve on -EEXIST; failures other than -EEXIST leave the
reservation unconsumed and still unreserve it.

Fixes: 717cf9357325 ("mm/memfd: reserve hugetlb folios before allocation")
Signed-off-by: Hongfu Li <lihongfu@kylinos.cn>
---
 mm/memfd.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/mm/memfd.c b/mm/memfd.c
index c708d92533f4..1b65bc22fb19 100644
--- a/mm/memfd.c
+++ b/mm/memfd.c
@@ -128,6 +128,12 @@ struct folio *memfd_alloc_folio(struct file *memfd, pgoff_t idx)
 
 			if (err) {
 				folio_put(folio);
+				/*
+				 * On -EEXIST, a concurrent fault has cached the folio,
+				 * which now owns the reservation; don't unreserve.
+				 */
+				if (err == -EEXIST)
+					return ERR_PTR(err);
 				goto err_unresv;
 			}
 
-- 
2.54.0

Re: [PATCH] mm/memfd: don't unreserve hugetlb pages on -EEXIST in error path

Posted by Andrew Morton 3 weeks, 4 days ago

On Mon, 31 Aug 2026 17:06:31 +0800 Hongfu Li <hongfu.li@linux.dev> wrote:

> From: Hongfu Li <lihongfu@kylinos.cn>
> 
> If hugetlb_add_to_page_cache() fails with -EEXIST, a concurrent fault has
> already instantiated the folio in the page cache, and the reservation now
> belongs to that folio. Calling hugetlb_unreserve_pages() in that case
> incorrectly removes the region backing the cached folio, and a later
> truncate or inode eviction passes a negative (chg - freed) into
> hugepage_subpool_put_pages(), corrupting subpool and resv_huge_pages
> accounting.
> 
> Skip the unreserve on -EEXIST; failures other than -EEXIST leave the
> reservation unconsumed and still unreserve it.

Thanks.

AI review might have found another bug and it suggests that your
proposal only partially fixes this bug:

	https://sashiko.dev/#/patchset/20260831090631.29227-1-hongfu.li@linux.dev

Re: [PATCH] mm/memfd: don't unreserve hugetlb pages on -EEXIST in error path

Posted by Hongfu Li 3 weeks, 4 days ago

On 9/1/26 8:35 AM, Andrew Morton wrote:
> On Mon, 31 Aug 2026 17:06:31 +0800 Hongfu Li <hongfu.li@linux.dev> wrote:
>
>> From: Hongfu Li <lihongfu@kylinos.cn>
>>
>> If hugetlb_add_to_page_cache() fails with -EEXIST, a concurrent fault has
>> already instantiated the folio in the page cache, and the reservation now
>> belongs to that folio. Calling hugetlb_unreserve_pages() in that case
>> incorrectly removes the region backing the cached folio, and a later
>> truncate or inode eviction passes a negative (chg - freed) into
>> hugepage_subpool_put_pages(), corrupting subpool and resv_huge_pages
>> accounting.
>>
>> Skip the unreserve on -EEXIST; failures other than -EEXIST leave the
>> reservation unconsumed and still unreserve it.
> Thanks.
>
> AI review might have found another bug and it suggests that your
> proposal only partially fixes this bug:
>
> 	https://sashiko.dev/#/patchset/20260831090631.29227-1-hongfu.li@linux.dev

Thank you for the reminder.

I have carefully gone through the Sashiko review report.
The issue it pointed out is valid. I will address this
remaining problem and send out a v2 patch shortly.

-- 
Best regards,
Hongfu