Skip to contentSkip to conversation

hugetlb, userfaultfd: Fix reservation restore on userfaultfd error

PATCH Applied Mina Almasry 2 participants 2 messages Nov 17, 2021

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

PATCH v2 1/1 hugetlb, userfaultfd: Fix reservation restore on userfaultfd error

Currently in the is_continue case in hugetlb_mcopy_atomic_pte(), if we bail out using "goto out_release_unlock;" in the cases where idx >= size, or !huge_pte_none(), the code will detect that new_pagecache_page == false, and so call restore_reserve_on_error().
In this case I see restore_reserve_on_error() delete the reservation, and the following call to remove_inode_hugepages() will increment h->resv_hugepages causing a 100% reproducible leak.

We should treat the is_continue case similar to adding a page into the pagecache and set new_pagecache_page to true, to indicate that there is no reservation to restore on the error path, and we need not call restore_reserve_on_error(). Rename new_pagecache_page to page_in_pagecache to make that clear.

Cc Wei Xu <[email protected]>
Fixes c7b1850dfb41 ("hugetlb: don't pass page cache pages to restore_reserve_on_error")

---

Changes in v2:
- Renamed new_pagecache_page to page_in_pagecache
- Removed unnecessary comment after the name update.
- Cc: stable

---
 mm/hugetlb.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
mm/hugetlb.c +4 -3
index e09159c957e3..e7ebc4b355cf 100644
@@ -5734,13 +5734,14 @@ int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
 	int ret = -ENOMEM;
 	struct page *page;
 	int writable;
-	bool new_pagecache_page = false;
+	bool page_in_pagecache = false;
 
 	if (is_continue) {
 		ret = -EFAULT;
 		page = find_lock_page(mapping, idx);
 		if (!page)
 			goto out;
+		page_in_pagecache = true;
 	} else if (!*pagep) {
 		/* If a page already exists, then it's UFFDIO_COPY for
 		 * a non-missing case. Return -EEXIST.
@@ -5828,7 +5829,7 @@ int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
 		ret = huge_add_to_page_cache(page, mapping, idx);
 		if (ret)
 			goto out_release_nounlock;
-		new_pagecache_page = true;
+		page_in_pagecache = true;
 	}
 
 	ptl = huge_pte_lockptr(h, dst_mm, dst_pte);
@@ -5892,7 +5893,7 @@ int hugetlb_mcopy_atomic_pte(struct mm_struct *dst_mm,
 	if (vm_shared || is_continue)
 		unlock_page(page);
 out_release_nounlock:
-	if (!new_pagecache_page)
+	if (!page_in_pagecache)
 		restore_reserve_on_error(h, dst_vma, dst_addr, page);
 	put_page(page);
 	goto out;

--
2.34.0.rc2.393.gf8c9666880-goog

Reviewed-by on v2 1/1
#1 lore raw
Quoted 23 lines · Mina Almasry

Currently in the is_continue case in hugetlb_mcopy_atomic_pte(), if we bail out using "goto out_release_unlock;" in the cases where idx >= size, or !huge_pte_none(), the code will detect that new_pagecache_page == false, and so call restore_reserve_on_error().
In this case I see restore_reserve_on_error() delete the reservation, and the following call to remove_inode_hugepages() will increment h->resv_hugepages causing a 100% reproducible leak.

We should treat the is_continue case similar to adding a page into the pagecache and set new_pagecache_page to true, to indicate that there is no reservation to restore on the error path, and we need not call restore_reserve_on_error(). Rename new_pagecache_page to page_in_pagecache to make that clear.

Cc Wei Xu <[email protected]>
Fixes c7b1850dfb41 ("hugetlb: don't pass page cache pages to restore_reserve_on_error")
Quoted 6 lines

---

Changes in v2:
- Renamed new_pagecache_page to page_in_pagecache
- Removed unnecessary comment after the name update.
- Cc: stable

Thanks for making the changes!

Signature · 1 line
Mike Kravetz