mm: thp: fix MADV_REMOVE deadlock on shmem THP

commit 1c2f67308af4c102b4e1e6cd6f69819ae59408e0 upstream.

Sergey reported deadlock between kswapd correctly doing its usual
lock_page(page) followed by down_read(page->mapping->i_mmap_rwsem), and
madvise(MADV_REMOVE) on an madvise(MADV_HUGEPAGE) area doing
down_write(page->mapping->i_mmap_rwsem) followed by lock_page(page).

This happened when shmem_fallocate(punch hole)'s unmap_mapping_range()
reaches zap_pmd_range()'s call to __split_huge_pmd().  The same deadlock
could occur when partially truncating a mapped huge tmpfs file, or using
fallocate(FALLOC_FL_PUNCH_HOLE) on it.

__split_huge_pmd()'s page lock was added in 5.8, to make sure that any
concurrent use of reuse_swap_page() (holding page lock) could not catch
the anon THP's mapcounts and swapcounts while they were being split.

Fortunately, reuse_swap_page() is never applied to a shmem or file THP
(not even by khugepaged, which checks PageSwapCache before calling), and
anonymous THPs are never created in shmem or file areas: so that
__split_huge_pmd()'s page lock can only be necessary for anonymous THPs,
on which there is no risk of deadlock with i_mmap_rwsem.

Link: https://lkml.kernel.org/r/alpine.LSU.2.11.2101161409470.2022@eggly.anvils
Fixes: c444eb564fb1 ("mm: thp: make the THP mapcount atomic against __split_huge_pmd_locked()")
Signed-off-by: Hugh Dickins <hughd@google.com>
Reported-by: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

mm/huge_memory.c

@@ -2198,7 +2198,7 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 	spinlock_t *ptl;
 	struct mm_struct *mm = vma->vm_mm;
 	unsigned long haddr = address & HPAGE_PMD_MASK;
-	bool was_locked = false;
+	bool do_unlock_page = false;
 	pmd_t _pmd;
 
 	mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
@@ -2211,7 +2211,6 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 	VM_BUG_ON(freeze && !page);
 	if (page) {
 		VM_WARN_ON_ONCE(!PageLocked(page));
-		was_locked = true;
 		if (page != pmd_page(*pmd))
 			goto out;
 	}
@@ -2220,19 +2219,29 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 	if (pmd_trans_huge(*pmd)) {
 		if (!page) {
 			page = pmd_page(*pmd);
-			if (unlikely(!trylock_page(page))) {
-				get_page(page);
-				_pmd = *pmd;
-				spin_unlock(ptl);
-				lock_page(page);
-				spin_lock(ptl);
-				if (unlikely(!pmd_same(*pmd, _pmd))) {
-					unlock_page(page);
+			/*
+			 * An anonymous page must be locked, to ensure that a
+			 * concurrent reuse_swap_page() sees stable mapcount;
+			 * but reuse_swap_page() is not used on shmem or file,
+			 * and page lock must not be taken when zap_pmd_range()
+			 * calls __split_huge_pmd() while i_mmap_lock is held.
+			 */
+			if (PageAnon(page)) {
+				if (unlikely(!trylock_page(page))) {
+					get_page(page);
+					_pmd = *pmd;
+					spin_unlock(ptl);
+					lock_page(page);
+					spin_lock(ptl);
+					if (unlikely(!pmd_same(*pmd, _pmd))) {
+						unlock_page(page);
+						put_page(page);
+						page = NULL;
+						goto repeat;
+					}
 					put_page(page);
-					page = NULL;
-					goto repeat;
 				}
-				put_page(page);
+				do_unlock_page = true;
 			}
 		}
 		if (PageMlocked(page))
@@ -2242,7 +2251,7 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 	__split_huge_pmd_locked(vma, pmd, haddr, freeze);
 out:
 	spin_unlock(ptl);
-	if (!was_locked && page)
+	if (do_unlock_page)
 		unlock_page(page);
 	mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
 }