Whamcloud - gitweb
LU-6655 ptlrpc: skip delayed replay requests
[fs/lustre-release.git] / lustre / target / tgt_handler.c
index 4a52875..24e6936 100644 (file)
@@ -661,6 +661,19 @@ static int process_req_last_xid(struct ptlrpc_request *req)
                        RETURN(-EPROTO);
        }
 
                        RETURN(-EPROTO);
        }
 
+       /* The "last_xid" is the minimum xid among unreplied requests,
+        * if the request is from the previous connection, its xid can
+        * still be larger than "exp_last_xid", then the above check of
+        * xid is not enough to determine whether the request is delayed.
+        *
+        * For example, if some replay request was delayed and caused
+        * timeout at client and the replay is restarted, the delayed
+        * replay request will have the larger xid than "exp_last_xid"
+        */
+       if (req->rq_export->exp_conn_cnt >
+           lustre_msg_get_conn_cnt(req->rq_reqmsg))
+               RETURN(-ESTALE);
+
        /* try to release in-memory reply data */
        if (tgt_is_multimodrpcs_client(req->rq_export)) {
                tgt_handle_received_xid(req->rq_export,
        /* try to release in-memory reply data */
        if (tgt_is_multimodrpcs_client(req->rq_export)) {
                tgt_handle_received_xid(req->rq_export,
@@ -687,6 +700,19 @@ int tgt_request_handle(struct ptlrpc_request *req)
        bool                     is_connect = false;
        ENTRY;
 
        bool                     is_connect = false;
        ENTRY;
 
+       if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
+               if (cfs_fail_val == 0 &&
+                   lustre_msg_get_opc(msg) != OBD_PING &&
+                   lustre_msg_get_flags(msg) & MSG_REQ_REPLAY_DONE) {
+                       struct l_wait_info lwi =  { 0 };
+
+                       cfs_fail_val = 1;
+                       cfs_race_state = 0;
+                       l_wait_event(cfs_race_waitq, (cfs_race_state == 1),
+                                    &lwi);
+               }
+       }
+
        /* Refill the context, to make sure all thread keys are allocated */
        lu_env_refill(req->rq_svc_thread->t_env);
 
        /* Refill the context, to make sure all thread keys are allocated */
        lu_env_refill(req->rq_svc_thread->t_env);