staging/lustre: conflicting PW & PR extent locks on a client

PW lock isn't replayed once a lock is marked
LDLM_FL_CANCELING and glimpse lock doesn't wait for
conflicting locks on the client. So the server will
grant a PR lock in response to the glimpse lock request,
which conflicts with the PW lock in LDLM_FL_CANCELING
state on the client.

Lock in LDLM_FL_CANCELING state may still have pending IO,
so it should be replayed until LDLM_FL_BL_DONE is set to
avoid granted conflicting lock by a server.

Seagate-bug-id: MRP-3311
Signed-off-by: Andriy Skulysh <andriy.skulysh@seagate.com>
Reviewed-on: http://review.whamcloud.com/20345
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-8175Reviewed-by: Jinshan Xiong <jinshan.xiong@intel.com>
Signed-off-by: Oleg Drokin <green@linuxhacker.ru>
Reviewed-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/staging/lustre/lustre/include/obd_support.h

@@ -321,6 +321,8 @@ extern char obd_jobid_var[];
 #define OBD_FAIL_LDLM_CP_CB_WAIT4	 0x322
 #define OBD_FAIL_LDLM_CP_CB_WAIT5	 0x323
 
+#define OBD_FAIL_LDLM_GRANT_CHECK        0x32a
+
 /* LOCKLESS IO */
 #define OBD_FAIL_LDLM_SET_CONTENTION     0x385
 
@@ -343,6 +345,7 @@ extern char obd_jobid_var[];
 #define OBD_FAIL_OSC_CP_ENQ_RACE	 0x410
 #define OBD_FAIL_OSC_NO_GRANT	    0x411
 #define OBD_FAIL_OSC_DELAY_SETTIME	 0x412
+#define OBD_FAIL_OSC_DELAY_IO		 0x414
 
 #define OBD_FAIL_PTLRPC		  0x500
 #define OBD_FAIL_PTLRPC_ACK	      0x501

drivers/staging/lustre/lustre/ldlm/ldlm_extent.c

@@ -193,6 +193,26 @@ void ldlm_extent_add_lock(struct ldlm_resource *res,
 	 * add the locks into grant list, for debug purpose, ..
 	 */
 	ldlm_resource_add_lock(res, &res->lr_granted, lock);
+
+	if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_GRANT_CHECK)) {
+		struct ldlm_lock *lck;
+
+		list_for_each_entry_reverse(lck, &res->lr_granted,
+					    l_res_link) {
+			if (lck == lock)
+				continue;
+			if (lockmode_compat(lck->l_granted_mode,
+					    lock->l_granted_mode))
+				continue;
+			if (ldlm_extent_overlap(&lck->l_req_extent,
+						&lock->l_req_extent)) {
+				CDEBUG(D_ERROR, "granting conflicting lock %p %p\n",
+				       lck, lock);
+				ldlm_resource_dump(D_ERROR, res);
+				LBUG();
+			}
+		}
+	}
 }
 
 /** Remove cancelled lock from resource interval tree. */

drivers/staging/lustre/lustre/ldlm/ldlm_request.c

@@ -1846,7 +1846,7 @@ static int ldlm_chain_lock_for_replay(struct ldlm_lock *lock, void *closure)
 	 * bug 17614: locks being actively cancelled. Get a reference
 	 * on a lock so that it does not disappear under us (e.g. due to cancel)
 	 */
-	if (!(lock->l_flags & (LDLM_FL_FAILED | LDLM_FL_CANCELING))) {
+	if (!(lock->l_flags & (LDLM_FL_FAILED | LDLM_FL_BL_DONE))) {
 		list_add(&lock->l_pending_chain, list);
 		LDLM_LOCK_GET(lock);
 	}
@@ -1915,7 +1915,7 @@ static int replay_one_lock(struct obd_import *imp, struct ldlm_lock *lock)
 	int flags;
 
 	/* Bug 11974: Do not replay a lock which is actively being canceled */
-	if (ldlm_is_canceling(lock)) {
+	if (ldlm_is_bl_done(lock)) {
 		LDLM_DEBUG(lock, "Not replaying canceled lock:");
 		return 0;
 	}

drivers/staging/lustre/lustre/osc/osc_request.c

@@ -1823,6 +1823,7 @@ int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
 	DEBUG_REQ(D_INODE, req, "%d pages, aa %p. now %ur/%dw in flight",
 		  page_count, aa, cli->cl_r_in_flight,
 		  cli->cl_w_in_flight);
+	OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
 
 	ptlrpcd_add_req(req);
 	rc = 0;