ext3: fix lock inversion in direct IO
commitbd1939de9061dbc5cac44ffb4425aaf4c9b894f1
authorJan Kara <jack@suse.cz>
Wed, 6 Feb 2008 09:40:21 +0000 (6 01:40 -0800)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Wed, 6 Feb 2008 18:41:21 +0000 (6 10:41 -0800)
tree99cc44f2b92db9dba1391c81638f7755603c0199
parentd8fd66aaea7fe3e4f1ea044a563f129e3b9f05ff
ext3: fix lock inversion in direct IO

We cannot start transaction in ext3_direct_IO() and just let it last during
the whole write because dio_get_page() acquires mmap_sem which ranks above
transaction start (e.g.  because we have dependency chain
mmap_sem->PageLock->journal_start, or because we update atime while holding
mmap_sem) and thus deadlocks could happen.  We solve the problem by
starting a transaction separately for each ext3_get_block() call.

We *could* have a problem that we allocate a block and before its data are
written out the machine crashes and thus we expose stale data.  But that
does not happen because for hole-filling generic code falls back to
buffered writes and for file extension, we add inode to orphan list and
thus in case of crash, journal replay will truncate inode back to the
original size.

[akpm@linux-foundation.org: build fix]
Signed-off-by: Jan Kara <jack@suse.cz>
Cc: <linux-ext4@vger.kernel.org>
Cc: Zach Brown <zach.brown@oracle.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/ext3/inode.c