NFSv4 fix CLOSE not waiting for direct IO compeletion
authorOlga Kornievskaia <olga.kornievskaia@gmail.com>
Wed, 24 Jun 2020 17:54:08 +0000 (13:54 -0400)
committerSasha Levin <sashal@kernel.org>
Tue, 30 Jun 2020 19:38:45 +0000 (15:38 -0400)
commit5c85c81e86340c50101fb03e9095150d64a0bf99
tree9f6dfb5d62601ff8ee6941321f29d315baabf45b
parent86dfbc17a336905d53835a532c85aa424e25ca0d
NFSv4 fix CLOSE not waiting for direct IO compeletion

commit d03727b248d0dae6199569a8d7b629a681154633 upstream.

Figuring out the root case for the REMOVE/CLOSE race and
suggesting the solution was done by Neil Brown.

Currently what happens is that direct IO calls hold a reference
on the open context which is decremented as an asynchronous task
in the nfs_direct_complete(). Before reference is decremented,
control is returned to the application which is free to close the
file. When close is being processed, it decrements its reference
on the open_context but since directIO still holds one, it doesn't
sent a close on the wire. It returns control to the application
which is free to do other operations. For instance, it can delete a
file. Direct IO is finally releasing its reference and triggering
an asynchronous close. Which races with the REMOVE. On the server,
REMOVE can be processed before the CLOSE, failing the REMOVE with
EACCES as the file is still opened.

Signed-off-by: Olga Kornievskaia <kolga@netapp.com>
Suggested-by: Neil Brown <neilb@suse.com>
CC: stable@vger.kernel.org
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
fs/nfs/direct.c
fs/nfs/file.c