1 ext4: fix bh->b_state corruption
3 From: Jan Kara <jack@suse.com>
5 ext4 can update bh->b_state non-atomically in _ext4_get_block() and
6 ext4_da_get_block_prep(). Usually this is fine since bh is just a
7 temporary storage for mapping information on stack but in some cases it
8 can be fully living bh attached to a page. In such case non-atomic
9 update of bh->b_state can race with an atomic update which then gets
10 lost. Usually when we are mapping bh and thus updating bh->b_state
11 non-atomically, nobody else touches the bh and so things work out fine
12 but there is one case to especially worry about: ext4_finish_bio() uses
13 BH_Uptodate_Lock on the first bh in the page to synchronize handling of
14 PageWriteback state. So when blocksize < pagesize, we can be atomically
15 modifying bh->b_state of a buffer that actually isn't under IO and thus
16 can race e.g. with delalloc trying to map that buffer. The result is
17 that we can mistakenly set / clear BH_Uptodate_Lock bit resulting in the
18 corruption of PageWriteback state or missed unlock of BH_Uptodate_Lock.
20 Fix the problem by always updating bh->b_state bits atomically.
22 CC: stable@vger.kernel.org
23 Reported-by: Nikolay Borisov <kernel@kyup.com>
24 Signed-off-by: Jan Kara <jack@suse.cz>
25 Signed-off-by: Theodore Ts'o <tytso@mit.edu>
27 fs/ext4/inode.c | 32 ++++++++++++++++++++++++++++++--
28 1 file changed, 30 insertions(+), 2 deletions(-)
30 diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
31 index f72dc04..e646abb 100644
34 @@ -690,6 +690,34 @@ out_sem:
39 + * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
40 + * we have to be careful as someone else may be manipulating b_state as well.
42 +static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
44 + unsigned long old_state;
45 + unsigned long new_state;
47 + flags &= EXT4_MAP_FLAGS;
49 + /* Dummy buffer_head? Set non-atomically. */
51 + bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
55 + * Someone else may be modifying b_state. Be careful! This is ugly but
56 + * once we get rid of using bh as a container for mapping information
57 + * to pass to / from get_block functions, this can go away.
60 + old_state = READ_ONCE(bh->b_state);
61 + new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
63 + cmpxchg(&bh->b_state, old_state, new_state) != old_state));
66 /* Maximum number of blocks we map for direct IO at once. */
67 #define DIO_MAX_BLOCKS 4096
69 @@ -726,7 +754,7 @@ static int _ext4_get_block(struct inode *inode, sector_t iblock,
70 ext4_io_end_t *io_end = ext4_inode_aio(inode);
72 map_bh(bh, inode->i_sb, map.m_pblk);
73 - bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
74 + ext4_update_bh_state(bh, map.m_flags);
75 if (io_end && io_end->flag & EXT4_IO_END_UNWRITTEN)
76 set_buffer_defer_completion(bh);
77 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
78 @@ -1692,7 +1720,7 @@ int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
81 map_bh(bh, inode->i_sb, map.m_pblk);
82 - bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
83 + ext4_update_bh_state(bh, map.m_flags);
85 if (buffer_unwritten(bh)) {
86 /* A delayed write to unwritten bh should be marked