ext4: fix race in xattr block allocation path
authorEric Sandeen <sandeen@redhat.com>
Sat, 29 Oct 2011 14:15:35 +0000 (10:15 -0400)
committerTheodore Ts'o <tytso@mit.edu>
Sat, 29 Oct 2011 14:15:35 +0000 (10:15 -0400)
commit6d6a435190bdf2e04c9465cde5bdc3ac68cf11a4
treed20f0dc543e58f424f782ba70648c88b7fb0f3e5
parente7b319e39776bd0e9c0c7855b023dafed2c93d27
ext4: fix race in xattr block allocation path

Ceph users reported that when using Ceph on ext4, the filesystem
would often become corrupted, containing inodes with incorrect
i_blocks counters.

I managed to reproduce this with a very hacked-up "streamtest"
binary from the Ceph tree.

Ceph is doing a lot of xattr writes, to out-of-inode blocks.
There is also another thread which does sync_file_range and close,
of the same files.  The problem appears to happen due to this race:

sync/flush thread               xattr-set thread
-----------------               ----------------

do_writepages                   ext4_xattr_set
ext4_da_writepages              ext4_xattr_set_handle
mpage_da_map_blocks             ext4_xattr_block_set
        set DELALLOC_RESERVE
                                ext4_new_meta_blocks
                                        ext4_mb_new_blocks
                                                if (!i_delalloc_reserved_flag)
                                                        vfs_dq_alloc_block
ext4_get_blocks
down_write(i_data_sem)
        set i_delalloc_reserved_flag
...
up_write(i_data_sem)
                                        if (i_delalloc_reserved_flag)
                                                vfs_dq_alloc_block_nofail

In other words, the sync/flush thread pops in and sets
i_delalloc_reserved_flag on the inode, which makes the xattr thread
think that it's in a delalloc path in ext4_new_meta_blocks(),
and add the block for a second time, after already having added
it once in the !i_delalloc_reserved_flag case in ext4_mb_new_blocks

The real problem is that we shouldn't be using the DELALLOC_RESERVED
state flag, and instead we should be passing
EXT4_GET_BLOCKS_DELALLOC_RESERVE down to ext4_map_blocks() instead of
using an inode state flag.  We'll fix this for now with using
i_data_sem to prevent this race, but this is really not the right way
to fix things.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>
Cc: stable@kernel.org
fs/ext4/xattr.c