Whamcloud - gitweb
Revert "LU-5570 lnet: check router aliveness timestamp"
[fs/lustre-release.git] / lnet / lnet / lib-move.c
index 5a5e14d..8973813 100644 (file)
@@ -746,6 +746,8 @@ lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
         int        alive;
         cfs_time_t deadline;
 
+        LASSERT (lnet_peer_aliveness_enabled(lp));
+
         /* Trust lnet_notify() if it has more recent aliveness news, but
          * ignore the initial assumed death (see lnet_peers_start_down()).
          */
@@ -774,28 +776,37 @@ lnet_peer_is_alive (lnet_peer_t *lp, cfs_time_t now)
 int
 lnet_peer_alive_locked (lnet_peer_t *lp)
 {
-       cfs_time_t now = cfs_time_current();
-       bool       query;
+        cfs_time_t now = cfs_time_current();
 
-       if (!lnet_peer_aliveness_enabled(lp))
-               return -ENODEV;
+        if (!lnet_peer_aliveness_enabled(lp))
+                return -ENODEV;
 
-       if (lp->lp_last_query == 0) {
-               query = true;
-       } else {
-               /* Peer appears dead, but we should avoid frequent NI queries
-                * (at most once per ni_query_interval seconds). */
-               static const int ni_query_interval = 1;
-               cfs_time_t       next_query;
-
-               next_query = cfs_time_add(lp->lp_last_query,
-                                         cfs_time_seconds(ni_query_interval));
-               query = cfs_time_aftereq(now, next_query);
+        if (lnet_peer_is_alive(lp, now))
+                return 1;
+
+        /* Peer appears dead, but we should avoid frequent NI queries (at
+         * most once per lnet_queryinterval seconds). */
+        if (lp->lp_last_query != 0) {
+                static const int lnet_queryinterval = 1;
+
+                cfs_time_t next_query =
+                           cfs_time_add(lp->lp_last_query,
+                                        cfs_time_seconds(lnet_queryinterval));
+
+                if (cfs_time_before(now, next_query)) {
+                        if (lp->lp_alive)
+                                CWARN("Unexpected aliveness of peer %s: "
+                                      "%d < %d (%d/%d)\n",
+                                      libcfs_nid2str(lp->lp_nid),
+                                      (int)now, (int)next_query,
+                                      lnet_queryinterval,
+                                      lp->lp_ni->ni_peertimeout);
+                        return 0;
+                }
         }
 
         /* query NI for latest aliveness news */
-       if (query)
-               lnet_ni_query_locked(lp->lp_ni, lp);
+       lnet_ni_query_locked(lp->lp_ni, lp);
 
         if (lnet_peer_is_alive(lp, now))
                 return 1;
@@ -830,8 +841,7 @@ lnet_post_send_locked(lnet_msg_t *msg, int do_send)
 
        /* NB 'lp' is always the next hop */
        if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
-           lnet_peer_alive_locked(lp) == 0 &&
-           !lnet_msg_is_rc_ping(msg)) { /* send RC ping even for dead router */
+           lnet_peer_alive_locked(lp) == 0) {
                the_lnet.ln_counters[cpt]->drop_count++;
                the_lnet.ln_counters[cpt]->drop_length += msg->msg_len;
                lnet_net_unlock(cpt);
@@ -1184,30 +1194,6 @@ lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
        lnet_peer_t *p1 = r1->lr_gateway;
        lnet_peer_t *p2 = r2->lr_gateway;
 
-       if (p1->lp_ni->ni_peertimeout > 0 &&
-           p2->lp_ni->ni_peertimeout > 0) {
-               /* if a router has queued bytes but no aliveness update for
-                * the last 10 seconds, it could be potentially dead or
-                * congested, so we prefer not to choose it even its status
-                * is still alive.
-                */
-               int        router_slow = cfs_time_seconds(10);
-               bool       r1_slow;
-               bool       r2_slow;
-               cfs_time_t now = cfs_time_current();
-
-               r1_slow = p1->lp_txqnob != 0 &&
-                       cfs_time_aftereq(now, p1->lp_last_alive + router_slow);
-               r2_slow = p2->lp_txqnob != 0 &&
-                       cfs_time_aftereq(now, p2->lp_last_alive + router_slow);
-
-               if (!r1_slow && r2_slow)
-                       return 1;
-
-               if (r1_slow && !r2_slow)
-                       return -1;
-       }
-
        if (r1->lr_priority < r2->lr_priority)
                return 1;
 
@@ -1890,18 +1876,17 @@ int
 lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
           void *private, int rdma_req)
 {
+       int             rc = 0;
+       int             cpt;
+       int             for_me;
        struct lnet_msg *msg;
-       lnet_peer_t     *rxpeer;
-       lnet_pid_t       dest_pid;
-       lnet_nid_t       dest_nid;
-       lnet_nid_t       src_nid;
-       __u32            payload_length;
-       __u32            type;
-       int              for_me;
-       int              cpt;
-       int              rc = 0;
-
-       LASSERT(!in_interrupt());
+       lnet_pid_t     dest_pid;
+       lnet_nid_t     dest_nid;
+       lnet_nid_t     src_nid;
+       __u32          payload_length;
+       __u32          type;
+
+       LASSERT (!in_interrupt ());
 
        type = le32_to_cpu(hdr->type);
        src_nid = le64_to_cpu(hdr->src_nid);
@@ -2084,18 +2069,6 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
        }
 
        lnet_msg_commit(msg, cpt);
-       /* LND just notified me for incoming message from rxpeer, so assume
-        * it is alive */
-       rxpeer = msg->msg_rxpeer;
-       rxpeer->lp_last_alive = rxpeer->lp_last_query = cfs_time_current();
-       if (!rxpeer->lp_alive)
-               lnet_notify_locked(rxpeer, 0, 1, rxpeer->lp_last_alive);
-
-       if (lnet_isrouter(msg->msg_rxpeer) &&
-           LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid)) {
-               lnet_router_ni_update_locked(msg->msg_rxpeer,
-                                            LNET_NIDNET(src_nid));
-       }
 
        /* message delay simulation */
        if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&