cifs: clear PF_MEMALLOC before exiting demultiplex thread
authorEric Biggers <ebiggers@google.com>
Mon, 9 Mar 2020 05:58:20 +0000 (22:58 -0700)
committerSteve French <stfrench@microsoft.com>
Mon, 23 Mar 2020 03:49:10 +0000 (22:49 -0500)
commitdc920277f1e59600c7cfd97438ee604ef6affdbd
treee1ba4dd2475e90b7649bd42a6da257c01d2974e1
parent266b9fecc58ca6b69a8852008896358830c0ba0f
cifs: clear PF_MEMALLOC before exiting demultiplex thread

Leaving PF_MEMALLOC set when exiting a kthread causes it to remain set
during do_exit().  That can confuse things.  For example, if BSD process
accounting is enabled and the accounting file has FS_SYNC_FL set and is
located on an ext4 filesystem without a journal, then do_exit() can end
up calling ext4_write_inode().  That triggers the
WARN_ON_ONCE(current->flags & PF_MEMALLOC) there, as it assumes
(appropriately) that inodes aren't written when allocating memory.

This was originally reported for another kernel thread, xfsaild() [1].
cifs_demultiplex_thread() also exits with PF_MEMALLOC set, so it's
potentially subject to this same class of issue -- though I haven't been
able to reproduce the WARN_ON_ONCE() via CIFS, since unlike xfsaild(),
cifs_demultiplex_thread() is sent SIGKILL before exiting, and that
interrupts the write to the BSD process accounting file.

Either way, leaving PF_MEMALLOC set is potentially problematic.  Let's
clean this up by properly saving and restoring PF_MEMALLOC.

[1] https://lore.kernel.org/r/0000000000000e7156059f751d7b@google.com

Signed-off-by: Eric Biggers <ebiggers@google.com>
Signed-off-by: Steve French <stfrench@microsoft.com>
fs/cifs/connect.c