btrfs: fix deadlock when writing out space cache
authorJosef Bacik <jbacik@fb.com>
Wed, 15 Nov 2017 21:20:52 +0000 (16:20 -0500)
committerDavid Sterba <dsterba@suse.com>
Mon, 27 Nov 2017 14:50:07 +0000 (15:50 +0100)
If we fail to prepare our pages for whatever reason (out of memory in
our case) we need to make sure to drop the block_group->data_rwsem,
otherwise hilarity ensues.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
[ add label and use existing unlocking code ]
Signed-off-by: David Sterba <dsterba@suse.com>
fs/btrfs/free-space-cache.c

index cdc9f4015ec36c08688a81bb1ba2bf657a9845ef..4426d1c73e50f1d1b1105a182d9c982a28b8ff08 100644 (file)
@@ -1264,7 +1264,7 @@ static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
        /* Lock all pages first so we can lock the extent safely. */
        ret = io_ctl_prepare_pages(io_ctl, inode, 0);
        if (ret)
-               goto out;
+               goto out_unlock;
 
        lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
                         &cached_state);
@@ -1358,6 +1358,7 @@ out_nospc_locked:
 out_nospc:
        cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
 
+out_unlock:
        if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
                up_write(&block_group->data_rwsem);