tmpfs: fix race between umount and writepage

Konstanin Khlebnikov reports that a dangerous race between umount and
shmem_writepage can be reproduced by this script:

  for i in {1..300} ; do
	mkdir $i
	while true ; do
		mount -t tmpfs none $i
		dd if=/dev/zero of=$i/test bs=1M count=$(($RANDOM % 100))
		umount $i
	done &
  done

on a 6xCPU node with 8Gb RAM: kernel very unstable after this accident. =)

Kernel log:

  VFS: Busy inodes after unmount of tmpfs.
                 Self-destruct in 5 seconds.  Have a nice day...

  WARNING: at lib/list_debug.c:53 __list_del_entry+0x8d/0x98()
  list_del corruption. prev->next should be ffff880222fdaac8, but was (null)
  Pid: 11222, comm: mount.tmpfs Not tainted 2.6.39-rc2+ #4
  Call Trace:
   warn_slowpath_common+0x80/0x98
   warn_slowpath_fmt+0x41/0x43
   __list_del_entry+0x8d/0x98
   evict+0x50/0x113
   iput+0x138/0x141
  ...
  BUG: unable to handle kernel paging request at ffffffffffffffff
  IP: shmem_free_blocks+0x18/0x4c
  Pid: 10422, comm: dd Tainted: G        W   2.6.39-rc2+ #4
  Call Trace:
   shmem_recalc_inode+0x61/0x66
   shmem_writepage+0xba/0x1dc
   pageout+0x13c/0x24c
   shrink_page_list+0x28e/0x4be
   shrink_inactive_list+0x21f/0x382
  ...

shmem_writepage() calls igrab() on the inode for the page which came from
page reclaim, to add it later into shmem_swaplist for swapoff operation.

This igrab() can race with super-block deactivating process:

  shrink_inactive_list()          deactivate_super()
  pageout()                       tmpfs_fs_type->kill_sb()
  shmem_writepage()               kill_litter_super()
                                  generic_shutdown_super()
                                   evict_inodes()
   igrab()
                                    atomic_read(&inode->i_count)
                                     skip-inode
   iput()
                                   if (!list_empty(&sb->s_inodes))
                                          printk("VFS: Busy inodes after...

This igrap-iput pair was added in commit 1b1b32f2c6f6 "tmpfs: fix
shmem_swaplist races" based on incorrect assumptions: igrab() protects the
inode from concurrent eviction by deletion, but it does nothing to protect
it from concurrent unmounting, which goes ahead despite the raised
i_count.

So this use of igrab() was wrong all along, but the race made much worse
in 2.6.37 when commit 63997e98a3be "split invalidate_inodes()" replaced
two attempts at invalidate_inodes() by a single evict_inodes().

Konstantin posted a plausible patch, raising sb->s_active too: I'm unsure
whether it was correct or not; but burnt once by igrab(), I am sure that
we don't want to rely more deeply upon externals here.

Fix it by adding the inode to shmem_swaplist earlier, while the page lock
on page in page cache still secures the inode against eviction, without
artifically raising i_count.  It was originally added later because
shmem_unuse_inode() is liable to remove an inode from the list while it's
unswapped; but we can guard against that by taking spinlock before
dropping mutex.

Reported-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Signed-off-by: Hugh Dickins <hughd@google.com>
Tested-by: Konstantin Khlebnikov <khlebnikov@openvz.org>
Cc: <stable@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 20 insertions, 11 deletions

diff --git a/mm/shmem.c b/mm/shmem.c
index 8fa27e4e582..262d7117344 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1039,6 +1039,7 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 	struct address_space *mapping;
 	unsigned long index;
 	struct inode *inode;
+	bool unlock_mutex = false;
 
 	BUG_ON(!PageLocked(page));
 	mapping = page->mapping;
@@ -1064,7 +1065,26 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 	else
 		swap.val = 0;
 
+	/*
+	 * Add inode to shmem_unuse()'s list of swapped-out inodes,
+	 * if it's not already there.  Do it now because we cannot take
+	 * mutex while holding spinlock, and must do so before the page
+	 * is moved to swap cache, when its pagelock no longer protects
+	 * the inode from eviction.  But don't unlock the mutex until
+	 * we've taken the spinlock, because shmem_unuse_inode() will
+	 * prune a !swapped inode from the swaplist under both locks.
+	 */
+	if (swap.val && list_empty(&info->swaplist)) {
+		mutex_lock(&shmem_swaplist_mutex);
+		/* move instead of add in case we're racing */
+		list_move_tail(&info->swaplist, &shmem_swaplist);
+		unlock_mutex = true;
+	}
+
 	spin_lock(&info->lock);
+	if (unlock_mutex)
+		mutex_unlock(&shmem_swaplist_mutex);
+
 	if (index >= info->next_index) {
 		BUG_ON(!(info->flags & SHMEM_TRUNCATE));
 		goto unlock;
@@ -1084,21 +1104,10 @@ static int shmem_writepage(struct page *page, struct writeback_control *wbc)
 		delete_from_page_cache(page);
 		shmem_swp_set(info, entry, swap.val);
 		shmem_swp_unmap(entry);
-		if (list_empty(&info->swaplist))
-			inode = igrab(inode);
-		else
-			inode = NULL;
 		spin_unlock(&info->lock);
 		swap_shmem_alloc(swap);
 		BUG_ON(page_mapped(page));
 		swap_writepage(page, wbc);
-		if (inode) {
-			mutex_lock(&shmem_swaplist_mutex);
-			/* move instead of add in case we're racing */
-			list_move_tail(&info->swaplist, &shmem_swaplist);
-			mutex_unlock(&shmem_swaplist_mutex);
-			iput(inode);
-		}
 		return 0;
 	}