Skip to contentSkip to conversation

hugetlbfs: move lock assertions after early returns in huge_pmd_unshare()

PATCH Applied Deepanshu Kartikey 4 participants 4 messages Oct 14, 2025 – Oct 24, 2025

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

PATCH v8 1/1 hugetlbfs: move lock assertions after early returns in huge_pmd_unshare()

When hugetlb_vmdelete_list() processes VMAs during truncate operations, it may encounter VMAs where huge_pmd_unshare() is called without the required shareable lock. This triggers an assertion failure in hugetlb_vma_assert_locked().

The previous fix in commit dd83609b8898 ("hugetlbfs: skip VMAs without shareable locks in hugetlb_vmdelete_list") skipped entire VMAs without shareable locks to avoid the assertion. However, this prevented pages from being unmapped and freed, causing a regression in fallocate(PUNCH_HOLE) operations where pages were not freed immediately, as reported by Mark Brown.

Instead of checking locks in the caller or skipping VMAs, move the lock assertions in huge_pmd_unshare() to after the early return checks. The assertions are only needed when actual PMD unsharing work will be performed. If the function returns early because sz != PMD_SIZE or the PMD is not shared, no locks are required and assertions should not fire.

This approach reverts the VMA skipping logic from commit dd83609b8898 ("hugetlbfs: skip VMAs without shareable locks in hugetlb_vmdelete_list") while moving the assertions to avoid the assertion failure, keeping all the logic within huge_pmd_unshare() itself and allowing page unmapping and freeing to proceed for all VMAs.

---
Changes in v8:
- Add missing revert of skip code in fs/hugetlbfs/inode.c that was introduced in commit dd83609b8898
- Move lock assertions after early returns in huge_pmd_unshare()
- Based on mm/mm-stable

Changes in v7:
- Rebased on mm/mm-stable per David's request
- Forgot to revert skip code from fs/hugetlbfs/inode.c

Changes in v6:
- Remove __vma_shareable_lock() check from __unmap_hugepage_range()
- Move lock assertions after early returns in huge_pmd_unshare()

Changes in v5:
- Incomplete: only moved assertions, forgot to remove v4 check

Changes in v4:
- Check __vma_shareable_lock() in __unmap_hugepage_range() before calling huge_pmd_unshare() per Oscar's suggestion

Changes in v3:
- Add ZAP_FLAG_NO_UNSHARE to skip only PMD unsharing

Changes in v2:
- Skip entire VMAs without shareable locks (caused PUNCH_HOLE regression)

Changes in v1:
- Initial fix attempt

---
 fs/hugetlbfs/inode.c | 9 ---------
 mm/hugetlb.c         | 5 ++---
 2 files changed, 2 insertions(+), 12 deletions(-)
fs/hugetlbfs/inode.c +0 -9
index 9c94ed8c3ab0..f42548ee9083 100644
@@ -478,14 +478,6 @@ hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 		if (!hugetlb_vma_trylock_write(vma))
 			continue;
 
-		/*
-		 * Skip VMAs without shareable locks. Per the design in commit
-		 * 40549ba8f8e0, these will be handled by remove_inode_hugepages()
-		 * called after this function with proper locking.
-		 */
-		if (!__vma_shareable_lock(vma))
-			goto skip;
-
 		v_start = vma_offset_start(vma, start);
 		v_end = vma_offset_end(vma, end);
 
@@ -496,7 +488,6 @@ hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 		 * vmas.  Therefore, lock is not held when calling
 		 * unmap_hugepage_range for private vmas.
 		 */
-skip:
 		hugetlb_vma_unlock_write(vma);
 	}
 }
mm/hugetlb.c +2 -3
index 795ee393eac0..0455119716ec 100644
@@ -7614,13 +7614,12 @@ int huge_pmd_unshare(struct mm_struct *mm, struct vm_area_struct *vma,
 	p4d_t *p4d = p4d_offset(pgd, addr);
 	pud_t *pud = pud_offset(p4d, addr);
 
-	i_mmap_assert_write_locked(vma->vm_file->f_mapping);
-	hugetlb_vma_assert_locked(vma);
 	if (sz != PMD_SIZE)
 		return 0;
 	if (!ptdesc_pmd_is_shared(virt_to_ptdesc(ptep)))
 		return 0;
-
+	i_mmap_assert_write_locked(vma->vm_file->f_mapping);
+	hugetlb_vma_assert_locked(vma);
 	pud_clear(pud);
 	/*
 	 * Once our caller drops the rmap lock, some other process might be
Signature · 1 line
2.34.1
Acked-by on v8 1/1
#1 lore raw
Quoted 38 lines · Deepanshu Kartikey

When hugetlb_vmdelete_list() processes VMAs during truncate operations, it may encounter VMAs where huge_pmd_unshare() is called without the required shareable lock. This triggers an assertion failure in hugetlb_vma_assert_locked().

The previous fix in commit dd83609b8898 ("hugetlbfs: skip VMAs without shareable locks in hugetlb_vmdelete_list") skipped entire VMAs without shareable locks to avoid the assertion. However, this prevented pages from being unmapped and freed, causing a regression in fallocate(PUNCH_HOLE) operations where pages were not freed immediately, as reported by Mark Brown.

Instead of checking locks in the caller or skipping VMAs, move the lock assertions in huge_pmd_unshare() to after the early return checks. The assertions are only needed when actual PMD unsharing work will be performed. If the function returns early because sz != PMD_SIZE or the PMD is not shared, no locks are required and assertions should not fire.

This approach reverts the VMA skipping logic from commit dd83609b8898 ("hugetlbfs: skip VMAs without shareable locks in hugetlb_vmdelete_list") while moving the assertions to avoid the assertion failure, keeping all the logic within huge_pmd_unshare() itself and allowing page unmapping and freeing to proceed for all VMAs.

Quoted 1 line

---

Signature · 3 lines
Cheers

David / dhildenb
#2 lore raw
In reply to Deepanshu Kartikey
Quoted 4 lines · Deepanshu Kartikey

When hugetlb_vmdelete_list() processes VMAs during truncate operations, it may encounter VMAs where huge_pmd_unshare() is called without the required shareable lock. This triggers an assertion failure in hugetlb_vma_assert_locked().

On current mainline this translates the test from a fail to a skip on the Raspberry Pi 4:

# # -------------------------
# # running ./hugetlb-madvise
# # -------------------------
# # [SKIP]

# ok 6 hugetlb-madvise # SKIP
# # hugepages not supported

due to:

[    0.000000] Kernel command line: console=ttyS1,115200n8 root=/dev/nfs rw nfsroot=172.16.0.2:/var/lib/lava/dispatcher/tmp/1991635/extract-nfsrootfs-9ik0m73w,tcp,hard secretmem.enable hugepagesz=32M hugepages=0:4 default_hugepagesz=2M hugepages=0:128 hugepagesz=64K hugepages=0:4 kpti=off ip=dhcp
[    0.000000] Unknown kernel command line parameters "hugepagesz=64K hugepages=0:4 default_hugepagesz=2M", will be passed to user space.

which used to DTRT but I doubt is due to this specific patch...

maintainer Acked-by on v8 1/1
#3 lore raw
In reply to Deepanshu Kartikey
Quoted 23 lines · Deepanshu Kartikey

When hugetlb_vmdelete_list() processes VMAs during truncate operations, it may encounter VMAs where huge_pmd_unshare() is called without the required shareable lock. This triggers an assertion failure in hugetlb_vma_assert_locked().

The previous fix in commit dd83609b8898 ("hugetlbfs: skip VMAs without shareable locks in hugetlb_vmdelete_list") skipped entire VMAs without shareable locks to avoid the assertion. However, this prevented pages from being unmapped and freed, causing a regression in fallocate(PUNCH_HOLE) operations where pages were not freed immediately, as reported by Mark Brown.

Instead of checking locks in the caller or skipping VMAs, move the lock assertions in huge_pmd_unshare() to after the early return checks. The assertions are only needed when actual PMD unsharing work will be performed. If the function returns early because sz != PMD_SIZE or the PMD is not shared, no locks are required and assertions should not fire.

This approach reverts the VMA skipping logic from commit dd83609b8898 ("hugetlbfs: skip VMAs without shareable locks in hugetlb_vmdelete_list") while moving the assertions to avoid the assertion failure, keeping all the logic within huge_pmd_unshare() itself and allowing page unmapping and freeing to proceed for all VMAs.

Quoted 15 lines

I got lost in the flow of multiple patches:

Signature · 2 lines
Oscar Salvador
SUSE Labs