btrfs: fix race between page release and a fast fsync

When releasing an extent map, done through the page release callback, we
can race with an ongoing fast fsync and cause the fsync to miss a new
extent and not log it. The steps for this to happen are the following:

1) A page is dirtied for some inode I;

2) Writeback for that page is triggered by a path other than fsync, for
   example by the system due to memory pressure;

3) When the ordered extent for the extent (a single 4K page) finishes,
   we unpin the corresponding extent map and set its generation to N,
   the current transaction's generation;

4) The btrfs_releasepage() callback is invoked by the system due to
   memory pressure for that no longer dirty page of inode I;

5) At the same time, some task calls fsync on inode I, joins transaction
   N, and at btrfs_log_inode() it sees that the inode does not have the
   full sync flag set, so we proceed with a fast fsync. But before we get
   into btrfs_log_changed_extents() and lock the inode's extent map tree:

6) Through btrfs_releasepage() we end up at try_release_extent_mapping()
   and we remove the extent map for the new 4Kb extent, because it is
   neither pinned anymore nor locked. By calling remove_extent_mapping(),
   we remove the extent map from the list of modified extents, since the
   extent map does not have the logging flag set. We unlock the inode's
   extent map tree;

7) The task doing the fast fsync now enters btrfs_log_changed_extents(),
   locks the inode's extent map tree and iterates its list of modified
   extents, which no longer has the 4Kb extent in it, so it does not log
   the extent;

8) The fsync finishes;

9) Before transaction N is committed, a power failure happens. After
   replaying the log, the 4K extent of inode I will be missing, since
   it was not logged due to the race with try_release_extent_mapping().

So fix this by teaching try_release_extent_mapping() to not remove an
extent map if it's still in the list of modified extents.

Fixes: ff44c6e3 ("Btrfs: do not hold the write_lock on the extent tree while logging")
CC: stable@vger.kernel.org # 5.4+
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>

fs/btrfs/extent_io.c

@@ -4494,15 +4494,25 @@ int try_release_extent_mapping(struct page *page, gfp_t mask)
 				free_extent_map(em);
 				break;
 			}
-			if (!test_range_bit(tree, em->start,
-					    extent_map_end(em) - 1,
-					    EXTENT_LOCKED, 0, NULL)) {
+			if (test_range_bit(tree, em->start,
+					   extent_map_end(em) - 1,
+					   EXTENT_LOCKED, 0, NULL))
+				goto next;
+			/*
+			 * If it's not in the list of modified extents, used
+			 * by a fast fsync, we can remove it. If it's being
+			 * logged we can safely remove it since fsync took an
+			 * extra reference on the em.
+			 */
+			if (list_empty(&em->list) ||
+			    test_bit(EXTENT_FLAG_LOGGING, &em->flags)) {
 				set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
 					&btrfs_inode->runtime_flags);
 				remove_extent_mapping(map, em);
 				/* once for the rb tree */
 				free_extent_map(em);
 			}
+next:
 			start = extent_map_end(em);
 			write_unlock(&map->lock);