tree 8201f2fc124d34098776799f8cec89a8f8b4f8bb
parent 8009f9fb3067fef6c2ca0c16f6bac786ae28639d
author Jan Kara <jack@suse.cz> 1202623718 -0500
committer Theodore Ts'o <tytso@mit.edu> 1202623718 -0500

ext4: Fix Direct I/O locking

We cannot start transaction in ext4_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 ext4_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.

Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Mingming Cao <cmm@us.ibm.com>
Signed-off-by: "Theodore Ts'o" <tytso@mit.edu>