mm: thp: fix MADV_REMOVE deadlock on shmem THP
authorHugh Dickins <hughd@google.com>
Fri, 5 Feb 2021 02:32:31 +0000 (18:32 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 5 Feb 2021 19:03:47 +0000 (11:03 -0800)
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>
mm/huge_memory.c

index 9237976..91ca9b1 100644 (file)
@@ -2202,7 +2202,7 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 {
        spinlock_t *ptl;
        struct mmu_notifier_range range;
-       bool was_locked = false;
+       bool do_unlock_page = false;
        pmd_t _pmd;
 
        mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
@@ -2218,7 +2218,6 @@ 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;
        }
@@ -2227,19 +2226,29 @@ repeat:
        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))
@@ -2249,7 +2258,7 @@ repeat:
        __split_huge_pmd_locked(vma, pmd, range.start, freeze);
 out:
        spin_unlock(ptl);
-       if (!was_locked && page)
+       if (do_unlock_page)
                unlock_page(page);
        /*
         * No need to double call mmu_notifier->invalidate_range() callback.