[PATCH 3.16 020/157] NFS: fix mount/umount race in nlmclnt.

[Ben Hutchings]

3.16.72-rc1 review patch.  If anyone has any objections, please let me know.

------------------

From: NeilBrown <neilb@xxxxxxxx>

commit 4a9be28c45bf02fa0436808bb6c0baeba30e120e upstream.

If the last NFSv3 unmount from a given host races with a mount from the
same host, we can destroy an nlm_host that is still in use.

Specifically nlmclnt_lookup_host() can increment h_count on
an nlm_host that nlmclnt_release_host() has just successfully called
refcount_dec_and_test() on.
Once nlmclnt_lookup_host() drops the mutex, nlm_destroy_host_lock()
will be called to destroy the nlmclnt which is now in use again.

The cause of the problem is that the dec_and_test happens outside the
locked region.  This is easily fixed by using
refcount_dec_and_mutex_lock().

Fixes: 8ea6ecc8b075 ("lockd: Create client-side nlm_host cache")
Signed-off-by: NeilBrown <neilb@xxxxxxxx>
Signed-off-by: Trond Myklebust <trond.myklebust@xxxxxxxxxxxxxxx>
[bwh: Backported to 3.16: use atomic instead of refcount API]
Signed-off-by: Ben Hutchings <ben@xxxxxxxxxxxxxxx>
---
 fs/lockd/host.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

--- a/fs/lockd/host.c
+++ b/fs/lockd/host.c
@@ -288,12 +288,11 @@ void nlmclnt_release_host(struct nlm_hos
 
 	WARN_ON_ONCE(host->h_server);
 
-	if (atomic_dec_and_test(&host->h_count)) {
+	if (atomic_dec_and_mutex_lock(&host->h_count, &nlm_host_mutex)) {
 		WARN_ON_ONCE(!list_empty(&host->h_lockowners));
 		WARN_ON_ONCE(!list_empty(&host->h_granted));
 		WARN_ON_ONCE(!list_empty(&host->h_reclaim));
 
-		mutex_lock(&nlm_host_mutex);
 		nlm_destroy_host_locked(host);
 		mutex_unlock(&nlm_host_mutex);
 	}