fuse: fix possibly missed wake-up after abort
authorMiklos Szeredi <mszeredi@redhat.com>
Fri, 9 Nov 2018 14:52:16 +0000 (15:52 +0100)
committerMiklos Szeredi <mszeredi@redhat.com>
Fri, 9 Nov 2018 14:52:16 +0000 (15:52 +0100)
commit2d84a2d19b6150c6dbac1e6ebad9c82e4c123772
tree30b80353869c028c20faed5b553dbb40aa7442b7
parent7fabaf303458fcabb694999d6fa772cc13d4e217
fuse: fix possibly missed wake-up after abort

In current fuse_drop_waiting() implementation it's possible that
fuse_wait_aborted() will not be woken up in the unlikely case that
fuse_abort_conn() + fuse_wait_aborted() runs in between checking
fc->connected and calling atomic_dec(&fc->num_waiting).

Do the atomic_dec_and_test() unconditionally, which also provides the
necessary barrier against reordering with the fc->connected check.

The explicit smp_mb() in fuse_wait_aborted() is not actually needed, since
the spin_unlock() in fuse_abort_conn() provides the necessary RELEASE
barrier after resetting fc->connected.  However, this is not a performance
sensitive path, and adding the explicit barrier makes it easier to
document.

Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
Fixes: b8f95e5d13f5 ("fuse: umount should wait for all requests")
Cc: <stable@vger.kernel.org> #v4.19
fs/fuse/dev.c