lockd: fix regression in lockd's handling of blocked locks

If a client requests a blocking lock, is denied, then requests it again,
then here in nlmsvc_lock() we will call vfs_lock_file() without FL_SLEEP
set, because we've already queued a block and don't need the locks code
to do it again.

But that means vfs_lock_file() will return -EAGAIN instead of
FILE_LOCK_DENIED.  So we still need to translate that -EAGAIN return
into a nlm_lck_blocked error in this case, and put ourselves back on
lockd's block list.

The bug was introduced by bde74e4 "locks: add special return
value for asynchronous locks".

Thanks to Frank van Maarseveen for the report; his original test
case was essentially

	for i in `seq 30`; do flock /nfsmount/foo sleep 10 & done

Tested-by: Frank van Maarseveen <[email protected]>
Reported-by: Frank van Maarseveen <[email protected]>
Cc: Miklos Szeredi <[email protected]>
Signed-off-by: J. Bruce Fields <[email protected]>

fs/lockd/svclock.c

@@ -427,7 +427,7 @@ nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
 			goto out;
 		case -EAGAIN:
 			ret = nlm_lck_denied;
-			goto out;
+			break;
 		case FILE_LOCK_DEFERRED:
 			if (wait)
 				break;
@@ -443,6 +443,10 @@ nlmsvc_lock(struct svc_rqst *rqstp, struct nlm_file *file,
 			goto out;
 	}
 
+	ret = nlm_lck_denied;
+	if (!wait)
+		goto out;
+
 	ret = nlm_lck_blocked;
 
 	/* Append to list of blocked */