Whamcloud - gitweb
LU-9121 lnet: Select NI/peer NI with highest prio
[fs/lustre-release.git] / lnet / lnet / lib-move.c
index 8099a27..a82dbbe 100644 (file)
@@ -69,6 +69,30 @@ lnet_msg_is_response(struct lnet_msg *msg)
        return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
 }
 
+static inline bool
+lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
+{
+       if (md_options & LNET_MD_NO_TRACK_RESPONSE)
+               /* Explicitly disabled in MD options */
+               return false;
+
+       if (md_options & LNET_MD_TRACK_RESPONSE)
+               /* Explicity enabled in MD options */
+               return true;
+
+       if (lnet_response_tracking == 3)
+               /* Enabled for all message types */
+               return true;
+
+       if (msg_type == LNET_MSG_PUT)
+               return lnet_response_tracking == 2;
+
+       if (msg_type == LNET_MSG_GET)
+               return lnet_response_tracking == 1;
+
+       return false;
+}
+
 static inline struct lnet_comm_count *
 get_stats_counts(struct lnet_element_stats *stats,
                 enum lnet_stats_type stats_type)
@@ -338,54 +362,6 @@ lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 }
 EXPORT_SYMBOL(lnet_copy_iov2iov);
 
-int
-lnet_extract_iov(int dst_niov, struct kvec *dst,
-                int src_niov, struct kvec *src,
-                unsigned int offset, unsigned int len)
-{
-       /* Initialise 'dst' to the subset of 'src' starting at 'offset',
-        * for exactly 'len' bytes, and return the number of entries.
-        * NB not destructive to 'src' */
-       unsigned int    frag_len;
-       unsigned int    niov;
-
-       if (len == 0)                           /* no data => */
-               return (0);                     /* no frags */
-
-       LASSERT(src_niov > 0);
-       while (offset >= src->iov_len) {      /* skip initial frags */
-               offset -= src->iov_len;
-               src_niov--;
-               src++;
-               LASSERT(src_niov > 0);
-       }
-
-       niov = 1;
-       for (;;) {
-               LASSERT(src_niov > 0);
-               LASSERT((int)niov <= dst_niov);
-
-               frag_len = src->iov_len - offset;
-               dst->iov_base = ((char *)src->iov_base) + offset;
-
-               if (len <= frag_len) {
-                       dst->iov_len = len;
-                       return (niov);
-               }
-
-               dst->iov_len = frag_len;
-
-               len -= frag_len;
-               dst++;
-               src++;
-               niov++;
-               src_niov--;
-               offset = 0;
-       }
-}
-EXPORT_SYMBOL(lnet_extract_iov);
-
-
 unsigned int
 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
 {
@@ -1307,27 +1283,10 @@ routing_off:
        }
 }
 
-static int
-lnet_compare_gw_lpnis(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
-{
-       if (p1->lpni_txqnob < p2->lpni_txqnob)
-               return 1;
-
-       if (p1->lpni_txqnob > p2->lpni_txqnob)
-               return -1;
-
-       if (p1->lpni_txcredits > p2->lpni_txcredits)
-               return 1;
-
-       if (p1->lpni_txcredits < p2->lpni_txcredits)
-               return -1;
-
-       return 0;
-}
-
 static struct lnet_peer_ni *
 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
                    struct lnet_peer *peer,
+                   struct lnet_peer_ni *best_lpni,
                    struct lnet_peer_net *peer_net)
 {
        /*
@@ -1339,12 +1298,15 @@ lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
         * credits are equal, we round-robin over the peer_ni.
         */
        struct lnet_peer_ni *lpni = NULL;
-       struct lnet_peer_ni *best_lpni = NULL;
-       int best_lpni_credits = INT_MIN;
-       bool preferred = false;
-       bool ni_is_pref;
-       int best_lpni_healthv = 0;
+       int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
+               INT_MIN;
+       int best_lpni_healthv = (best_lpni) ?
+               atomic_read(&best_lpni->lpni_healthv) : 0;
+       bool best_lpni_is_preferred = false;
+       bool lpni_is_preferred;
        int lpni_healthv;
+       __u32 lpni_sel_prio;
+       __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
 
        while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
                /*
@@ -1352,56 +1314,76 @@ lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
                 * preferred, then let's use it
                 */
                if (best_ni) {
-                       ni_is_pref = lnet_peer_is_pref_nid_locked(lpni,
+                       lpni_is_preferred = lnet_peer_is_pref_nid_locked(lpni,
                                                                best_ni->ni_nid);
-                       CDEBUG(D_NET, "%s ni_is_pref = %d\n",
-                              libcfs_nid2str(best_ni->ni_nid), ni_is_pref);
+                       CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
+                              libcfs_nid2str(best_ni->ni_nid),
+                              lpni_is_preferred);
                } else {
-                       ni_is_pref = false;
+                       lpni_is_preferred = false;
                }
 
                lpni_healthv = atomic_read(&lpni->lpni_healthv);
+               lpni_sel_prio = lpni->lpni_sel_priority;
 
                if (best_lpni)
-                       CDEBUG(D_NET, "%s c:[%d, %d], s:[%d, %d]\n",
+                       CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
                                libcfs_nid2str(lpni->lpni_nid),
+                               libcfs_nid2str(best_lpni->lpni_nid),
+                               lpni_healthv, best_lpni_healthv,
+                               lpni_sel_prio, best_sel_prio,
                                lpni->lpni_txcredits, best_lpni_credits,
                                lpni->lpni_seq, best_lpni->lpni_seq);
+               else
+                       goto select_lpni;
 
                /* pick the healthiest peer ni */
-               if (lpni_healthv < best_lpni_healthv) {
+               if (lpni_healthv < best_lpni_healthv)
                        continue;
-               } else if (lpni_healthv > best_lpni_healthv) {
-                       best_lpni_healthv = lpni_healthv;
+               else if (lpni_healthv > best_lpni_healthv) {
+                       if (best_lpni_is_preferred)
+                               best_lpni_is_preferred = false;
+                       goto select_lpni;
+               }
+
+               if (lpni_sel_prio > best_sel_prio)
+                       continue;
+               else if (lpni_sel_prio < best_sel_prio) {
+                       if (best_lpni_is_preferred)
+                               best_lpni_is_preferred = false;
+                       goto select_lpni;
+               }
+
                /* if this is a preferred peer use it */
-               } else if (!preferred && ni_is_pref) {
-                       preferred = true;
-               } else if (preferred && !ni_is_pref) {
-                       /*
-                        * this is not the preferred peer so let's ignore
+               if (!best_lpni_is_preferred && lpni_is_preferred) {
+                       best_lpni_is_preferred = true;
+                       goto select_lpni;
+               } else if (best_lpni_is_preferred && !lpni_is_preferred) {
+                       /* this is not the preferred peer so let's ignore
                         * it.
                         */
                        continue;
-               } else if (lpni->lpni_txcredits < best_lpni_credits) {
-                       /*
-                        * We already have a peer that has more credits
+               }
+
+               if (lpni->lpni_txcredits < best_lpni_credits)
+                       /* We already have a peer that has more credits
                         * available than this one. No need to consider
                         * this peer further.
                         */
                        continue;
-               } else if (lpni->lpni_txcredits == best_lpni_credits) {
-                       /*
-                        * The best peer found so far and the current peer
-                        * have the same number of available credits let's
-                        * make sure to select between them using Round
-                        * Robin
-                        */
-                       if (best_lpni) {
-                               if (best_lpni->lpni_seq <= lpni->lpni_seq)
-                                       continue;
-                       }
-               }
+               else if (lpni->lpni_txcredits > best_lpni_credits)
+                       goto select_lpni;
 
+               /* The best peer found so far and the current peer
+                * have the same number of available credits let's
+                * make sure to select between them using Round Robin
+                */
+               if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
+                       continue;
+select_lpni:
+               best_lpni_is_preferred = lpni_is_preferred;
+               best_lpni_healthv = lpni_healthv;
+               best_sel_prio = lpni_sel_prio;
                best_lpni = lpni;
                best_lpni_credits = lpni->lpni_txcredits;
        }
@@ -1423,20 +1405,58 @@ lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
 
 /*
  * Prerequisite: the best_ni should already be set in the sd
+ * Find the best lpni.
+ * If the net id is provided then restrict lpni selection on
+ * that particular net.
+ * Otherwise find any reachable lpni. When dealing with an MR
+ * gateway and it has multiple lpnis which we can use
+ * we want to select the best one from the list of reachable
+ * ones.
  */
 static inline struct lnet_peer_ni *
-lnet_find_best_lpni_on_net(struct lnet_ni *lni, lnet_nid_t dst_nid,
-                          struct lnet_peer *peer, __u32 net_id)
+lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
+                   struct lnet_peer *peer, __u32 net_id)
 {
        struct lnet_peer_net *peer_net;
 
-       /*
-        * The gateway is Multi-Rail capable so now we must select the
-        * proper peer_ni
-        */
+       /* find the best_lpni on any local network */
+       if (net_id == LNET_NET_ANY) {
+               struct lnet_peer_ni *best_lpni = NULL;
+               struct lnet_peer_net *lpn;
+               list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
+                       /* no net specified find any reachable peer ni */
+                       if (!lnet_islocalnet_locked(lpn->lpn_net_id))
+                               continue;
+                       best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
+                                                       best_lpni, lpn);
+               }
+
+               return best_lpni;
+       }
+       /* restrict on the specified net */
        peer_net = lnet_peer_get_net_locked(peer, net_id);
+       if (peer_net)
+               return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
 
-       return lnet_select_peer_ni(lni, dst_nid, peer, peer_net);
+       return NULL;
+}
+
+static int
+lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
+{
+       if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
+               return 1;
+
+       if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
+               return -1;
+
+       if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
+               return 1;
+
+       if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
+               return -1;
+
+       return 0;
 }
 
 /* Compare route priorities and hop counts */
@@ -1463,6 +1483,7 @@ lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
 
 static struct lnet_route *
 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
+                      struct lnet_peer_ni *remote_lpni,
                       struct lnet_route **prev_route,
                       struct lnet_peer_ni **gwni)
 {
@@ -1471,46 +1492,85 @@ lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
        struct lnet_route *last_route;
        struct lnet_route *route;
        int rc;
+       bool best_rte_is_preferred = false;
+       lnet_nid_t gw_pnid;
+
+       CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
+              libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
 
        best_route = last_route = NULL;
        list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
                if (!lnet_is_route_alive(route))
                        continue;
+               gw_pnid = route->lr_gateway->lp_primary_nid;
 
-               /*
-                * Restrict the selection of the router NI on the src_net
-                * provided. If the src_net is LNET_NID_ANY, then select
-                * the best interface available.
+               /* no protection on below fields, but it's harmless */
+               if (last_route && (last_route->lr_seq - route->lr_seq < 0))
+                       last_route = route;
+
+               /* if the best route found is in the preferred list then
+                * tag it as preferred and use it later on. But if we
+                * didn't find any routes which are on the preferred list
+                * then just use the best route possible.
                 */
-               if (!best_route) {
-                       lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
-                                                         route->lr_gateway,
-                                                         src_net);
-                       if (lpni) {
-                               best_route = last_route = route;
-                               best_gw_ni = lpni;
-                       } else
-                               CERROR("Gateway %s does not have a peer NI on net %s\n",
-                                      libcfs_nid2str(route->lr_gateway->lp_primary_nid),
+               rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
+
+               if (!best_route || (rc && !best_rte_is_preferred)) {
+                       /* Restrict the selection of the router NI on the
+                        * src_net provided. If the src_net is LNET_NID_ANY,
+                        * then select the best interface available.
+                        */
+                       lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
+                                                  route->lr_gateway,
+                                                  src_net);
+                       if (!lpni) {
+                               CDEBUG(D_NET,
+                                      "Gateway %s does not have a peer NI on net %s\n",
+                                      libcfs_nid2str(gw_pnid),
                                       libcfs_net2str(src_net));
+                               continue;
+                       }
+               }
 
+               if (rc && !best_rte_is_preferred) {
+                       /* This is the first preferred route we found,
+                        * so it beats any route found previously
+                        */
+                       best_route = route;
+                       if (!last_route)
+                               last_route = route;
+                       best_gw_ni = lpni;
+                       best_rte_is_preferred = true;
+                       CDEBUG(D_NET, "preferred gw = %s\n",
+                              libcfs_nid2str(gw_pnid));
+                       continue;
+               } else if ((!rc) && best_rte_is_preferred)
+                       /* The best route we found so far is in the preferred
+                        * list, so it beats any non-preferred route
+                        */
                        continue;
-               }
 
-               /* no protection on below fields, but it's harmless */
-               if (last_route->lr_seq - route->lr_seq < 0)
-                       last_route = route;
+               if (!best_route) {
+                       best_route = last_route = route;
+                       best_gw_ni = lpni;
+                       continue;
+               }
 
                rc = lnet_compare_routes(route, best_route);
                if (rc == -1)
                        continue;
 
-               lpni = lnet_find_best_lpni_on_net(NULL, LNET_NID_ANY,
-                                                 route->lr_gateway,
-                                                 src_net);
+               /* Restrict the selection of the router NI on the
+                * src_net provided. If the src_net is LNET_NID_ANY,
+                * then select the best interface available.
+                */
+               lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
+                                          route->lr_gateway,
+                                          src_net);
                if (!lpni) {
-                       CERROR("Gateway %s does not have a peer NI on net %s\n",
-                              libcfs_nid2str(route->lr_gateway->lp_primary_nid),
+                       CDEBUG(D_NET,
+                              "Gateway %s does not have a peer NI on net %s\n",
+                              libcfs_nid2str(gw_pnid),
                               libcfs_net2str(src_net));
                        continue;
                }
@@ -1547,6 +1607,7 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
        unsigned int shortest_distance;
        int best_credits;
        int best_healthv;
+       __u32 best_sel_prio;
 
        /*
         * If there is no peer_ni that we can send to on this network,
@@ -1556,6 +1617,7 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
                return best_ni;
 
        if (best_ni == NULL) {
+               best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
                shortest_distance = UINT_MAX;
                best_credits = INT_MIN;
                best_healthv = 0;
@@ -1564,6 +1626,7 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
                                                     best_ni->ni_dev_cpt);
                best_credits = atomic_read(&best_ni->ni_tx_credits);
                best_healthv = atomic_read(&best_ni->ni_healthv);
+               best_sel_prio = best_ni->ni_sel_priority;
        }
 
        while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
@@ -1571,10 +1634,12 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
                int ni_credits;
                int ni_healthv;
                int ni_fatal;
+               __u32 ni_sel_prio;
 
                ni_credits = atomic_read(&ni->ni_tx_credits);
                ni_healthv = atomic_read(&ni->ni_healthv);
                ni_fatal = atomic_read(&ni->ni_fatal_error_on);
+               ni_sel_prio = ni->ni_sel_priority;
 
                /*
                 * calculate the distance from the CPT on which
@@ -1585,12 +1650,6 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
                                            md_cpt,
                                            ni->ni_dev_cpt);
 
-               CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d] with best_ni %s [c:%d, d:%d, s:%d]\n",
-                      libcfs_nid2str(ni->ni_nid), ni_credits, distance,
-                      ni->ni_seq, (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
-                       : "not seleced", best_credits, shortest_distance,
-                       (best_ni) ? best_ni->ni_seq : 0);
-
                /*
                 * All distances smaller than the NUMA range
                 * are treated equally.
@@ -1602,31 +1661,47 @@ lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
                 * Select on health, shorter distance, available
                 * credits, then round-robin.
                 */
-               if (ni_fatal) {
+               if (ni_fatal)
                        continue;
-               } else if (ni_healthv < best_healthv) {
+
+               if (best_ni)
+                       CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d, p:%u] with best_ni %s [c:%d, d:%d, s:%d, p:%u]\n",
+                              libcfs_nid2str(ni->ni_nid), ni_credits, distance,
+                              ni->ni_seq, ni_sel_prio,
+                              (best_ni) ? libcfs_nid2str(best_ni->ni_nid)
+                              : "not selected", best_credits, shortest_distance,
+                              (best_ni) ? best_ni->ni_seq : 0,
+                              best_sel_prio);
+               else
+                       goto select_ni;
+
+               if (ni_healthv < best_healthv)
                        continue;
-               } else if (ni_healthv > best_healthv) {
-                       best_healthv = ni_healthv;
-                       /*
-                        * If we're going to prefer this ni because it's
-                        * the healthiest, then we should set the
-                        * shortest_distance in the algorithm in case
-                        * there are multiple NIs with the same health but
-                        * different distances.
-                        */
-                       if (distance < shortest_distance)
-                               shortest_distance = distance;
-               } else if (distance > shortest_distance) {
+               else if (ni_healthv > best_healthv)
+                       goto select_ni;
+
+               if (ni_sel_prio > best_sel_prio)
                        continue;
-               } else if (distance < shortest_distance) {
-                       shortest_distance = distance;
-               } else if (ni_credits < best_credits) {
+               else if (ni_sel_prio < best_sel_prio)
+                       goto select_ni;
+
+               if (distance > shortest_distance)
                        continue;
-               } else if (ni_credits == best_credits) {
-                       if (best_ni && best_ni->ni_seq <= ni->ni_seq)
-                               continue;
-               }
+               else if (distance < shortest_distance)
+                       goto select_ni;
+
+               if (ni_credits < best_credits)
+                       continue;
+               else if (ni_credits > best_credits)
+                       goto select_ni;
+
+               if (best_ni && best_ni->ni_seq <= ni->ni_seq)
+                       continue;
+
+select_ni:
+               best_sel_prio = ni_sel_prio;
+               shortest_distance = distance;
+               best_healthv = ni_healthv;
                best_ni = ni;
                best_credits = ni_credits;
        }
@@ -1836,7 +1911,8 @@ static inline void
 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
                         struct lnet_msg *msg)
 {
-       if (!lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
+       if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
+           !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
                CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
                       libcfs_nid2str(lni->ni_nid),
                       libcfs_nid2str(lpni->lpni_nid));
@@ -2001,7 +2077,15 @@ lnet_handle_find_routed_path(struct lnet_send_data *sd,
        struct lnet_route *last_route = NULL;
        struct lnet_peer_ni *lpni = NULL;
        struct lnet_peer_ni *gwni = NULL;
-       lnet_nid_t src_nid = sd->sd_src_nid;
+       bool route_found = false;
+       lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
+               (sd->sd_best_ni != NULL) ? sd->sd_best_ni->ni_nid :
+               LNET_NID_ANY;
+       int best_lpn_healthv = 0;
+       __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
+
+       CDEBUG(D_NET, "using src nid %s for route restriction\n",
+              libcfs_nid2str(src_nid));
 
        /* If a router nid was specified then we are replying to a GET or
         * sending an ACK. In this case we use the gateway associated with the
@@ -2009,72 +2093,110 @@ lnet_handle_find_routed_path(struct lnet_send_data *sd,
         */
        if (sd->sd_rtr_nid != LNET_NID_ANY) {
                gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
-               if (!gwni) {
-                       CERROR("No peer NI for gateway %s\n",
+               if (gwni) {
+                       gw = gwni->lpni_peer_net->lpn_peer;
+                       lnet_peer_ni_decref_locked(gwni);
+                       if (gw->lp_rtr_refcount) {
+                               local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
+                               route_found = true;
+                       }
+               } else {
+                       CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
                               libcfs_nid2str(sd->sd_rtr_nid));
-                       return -EHOSTUNREACH;
                }
-               gw = gwni->lpni_peer_net->lpn_peer;
-               lnet_peer_ni_decref_locked(gwni);
-               local_lnet = LNET_NIDNET(sd->sd_rtr_nid);
-       } else {
-               /* we've already looked up the initial lpni using dst_nid */
-               lpni = sd->sd_best_lpni;
-               /* the peer tree must be in existence */
-               LASSERT(lpni && lpni->lpni_peer_net &&
-                       lpni->lpni_peer_net->lpn_peer);
-               lp = lpni->lpni_peer_net->lpn_peer;
-
-               list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
-                       /* is this remote network reachable?  */
-                       rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
-                       if (!rnet)
-                               continue;
+       }
 
-                       if (!best_lpn) {
+       if (!route_found) {
+               if (sd->sd_msg->msg_routing) {
+                       /* If I'm routing this message then I need to find the
+                        * next hop based on the destination NID
+                        */
+                       best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
+                       if (!best_rnet) {
+                               CERROR("Unable to route message to %s - Route table may be misconfigured\n",
+                                      libcfs_nid2str(sd->sd_dst_nid));
+                               return -EHOSTUNREACH;
+                       }
+               } else {
+                       /* we've already looked up the initial lpni using
+                        * dst_nid
+                        */
+                       lpni = sd->sd_best_lpni;
+                       /* the peer tree must be in existence */
+                       LASSERT(lpni && lpni->lpni_peer_net &&
+                               lpni->lpni_peer_net->lpn_peer);
+                       lp = lpni->lpni_peer_net->lpn_peer;
+
+                       list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
+                               /* is this remote network reachable?  */
+                               rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
+                               if (!rnet)
+                                       continue;
+
+                               if (!best_lpn) {
+                                       best_lpn = lpn;
+                                       best_rnet = rnet;
+                               }
+
+                               /* select the preferred peer net */
+                               if (best_lpn_healthv > lpn->lpn_healthv)
+                                       continue;
+                               else if (best_lpn_healthv < lpn->lpn_healthv)
+                                       goto use_lpn;
+
+                               if (best_lpn_sel_prio < lpn->lpn_sel_priority)
+                                       continue;
+                               else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
+                                       goto use_lpn;
+
+                               if (best_lpn->lpn_seq <= lpn->lpn_seq)
+                                       continue;
+use_lpn:
+                               best_lpn_healthv = lpn->lpn_healthv;
+                               best_lpn_sel_prio = lpn->lpn_sel_priority;
                                best_lpn = lpn;
                                best_rnet = rnet;
                        }
 
-                       if (best_lpn->lpn_seq <= lpn->lpn_seq)
-                               continue;
+                       if (!best_lpn) {
+                               CERROR("peer %s has no available nets\n",
+                                      libcfs_nid2str(sd->sd_dst_nid));
+                               return -EHOSTUNREACH;
+                       }
 
-                       best_lpn = lpn;
-                       best_rnet = rnet;
-               }
+                       sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
+                                                              sd->sd_dst_nid,
+                                                              lp,
+                                                              best_lpn->lpn_net_id);
+                       if (!sd->sd_best_lpni) {
+                               CERROR("peer %s is unreachable\n",
+                                      libcfs_nid2str(sd->sd_dst_nid));
+                               return -EHOSTUNREACH;
+                       }
 
-               if (!best_lpn) {
-                       CERROR("peer %s has no available nets\n",
-                              libcfs_nid2str(sd->sd_dst_nid));
-                       return -EHOSTUNREACH;
-               }
+                       /* We're attempting to round robin over the remote peer
+                        * NI's so update the final destination we selected
+                        */
+                       sd->sd_final_dst_lpni = sd->sd_best_lpni;
 
-               sd->sd_best_lpni = lnet_find_best_lpni_on_net(sd->sd_best_ni,
-                                                             sd->sd_dst_nid,
-                                                             lp,
-                                                             best_lpn->lpn_net_id);
-               if (!sd->sd_best_lpni) {
-                       CERROR("peer %s down\n",
-                              libcfs_nid2str(sd->sd_dst_nid));
-                       return -EHOSTUNREACH;
+                       /* Increment the sequence number of the remote lpni so
+                        * we can round robin over the different interfaces of
+                        * the remote lpni
+                        */
+                       sd->sd_best_lpni->lpni_seq++;
                }
 
                /*
-                * We're attempting to round robin over the remote peer
-                * NI's so update the final destination we selected
-                */
-               sd->sd_final_dst_lpni = sd->sd_best_lpni;
-
-               /*
                 * find the best route. Restrict the selection on the net of the
                 * local NI if we've already picked the local NI to send from.
                 * Otherwise, let's pick any route we can find and then find
-                * a local NI we can reach the route's gateway on. Any route we select
-                * will be reachable by virtue of the restriction we have when
-                * adding a route.
+                * a local NI we can reach the route's gateway on. Any route we
+                * select will be reachable by virtue of the restriction we have
+                * when adding a route.
                 */
                best_route = lnet_find_route_locked(best_rnet,
                                                    LNET_NIDNET(src_nid),
+                                                   sd->sd_best_lpni,
                                                    &last_route, &gwni);
 
                if (!best_route) {
@@ -2094,13 +2216,6 @@ lnet_handle_find_routed_path(struct lnet_send_data *sd,
                gw = best_route->lr_gateway;
                LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
                local_lnet = best_route->lr_lnet;
-
-               /*
-                * Increment the sequence number of the remote lpni so we
-                * can round robin over the different interfaces of the
-                * remote lpni
-                */
-               sd->sd_best_lpni->lpni_seq++;
        }
 
        /*
@@ -2137,7 +2252,8 @@ lnet_handle_find_routed_path(struct lnet_send_data *sd,
        if (sd->sd_rtr_nid == LNET_NID_ANY) {
                LASSERT(best_route && last_route);
                best_route->lr_seq = last_route->lr_seq + 1;
-               best_lpn->lpn_seq++;
+               if (best_lpn)
+                       best_lpn->lpn_seq++;
        }
 
        return 0;
@@ -2335,7 +2451,7 @@ lnet_select_preferred_best_ni(struct lnet_send_data *sd)
 static int
 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
 {
-       int rc;
+       int rc = 0;
 
        /* sd->sd_best_lpni is already set to the final destination */
 
@@ -2352,7 +2468,23 @@ lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
                return -EFAULT;
        }
 
-       rc = lnet_select_preferred_best_ni(sd);
+       if (sd->sd_msg->msg_routing) {
+               /* If I'm forwarding this message then I can choose any NI
+                * on the destination peer net
+                */
+               sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
+                                                              sd->sd_peer,
+                                                              sd->sd_best_lpni->lpni_peer_net,
+                                                              sd->sd_md_cpt,
+                                                              true);
+               if (!sd->sd_best_ni) {
+                       CERROR("Unable to forward message to %s. No local NI available\n",
+                              libcfs_nid2str(sd->sd_dst_nid));
+                       rc = -EHOSTUNREACH;
+               }
+       } else
+               rc = lnet_select_preferred_best_ni(sd);
+
        if (!rc)
                rc = lnet_handle_send(sd);
 
@@ -2404,9 +2536,9 @@ lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
                                        lnet_msg_discovery(sd->sd_msg));
        if (sd->sd_best_ni) {
                sd->sd_best_lpni =
-                 lnet_find_best_lpni_on_net(sd->sd_best_ni, sd->sd_dst_nid,
-                                            sd->sd_peer,
-                                            sd->sd_best_ni->ni_net->net_id);
+                 lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
+                                     sd->sd_peer,
+                                     sd->sd_best_ni->ni_net->net_id);
 
                /*
                 * if we're successful in selecting a peer_ni on the local
@@ -2629,6 +2761,8 @@ lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
        int cpt, rc;
        int md_cpt;
        __u32 send_case = 0;
+       bool final_hop;
+       bool mr_forwarding_allowed;
 
        memset(&send_data, 0, sizeof(send_data));
 
@@ -2712,17 +2846,51 @@ again:
        else
                send_case |= REMOTE_DST;
 
+       final_hop = false;
+       if (msg->msg_routing && (send_case & LOCAL_DST))
+               final_hop = true;
+
+       /* Determine whether to allow MR forwarding for this message.
+        * NB: MR forwarding is allowed if the message originator and the
+        * destination are both MR capable, and the destination lpni that was
+        * originally chosen by the originator is unhealthy or down.
+        * We check the MR capability of the destination further below
+        */
+       mr_forwarding_allowed = false;
+       if (final_hop) {
+               struct lnet_peer *src_lp;
+               struct lnet_peer_ni *src_lpni;
+
+               src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
+                                                 LNET_NID_ANY, cpt);
+               /* We don't fail the send if we hit any errors here. We'll just
+                * try to send it via non-multi-rail criteria
+                */
+               if (!IS_ERR(src_lpni)) {
+                       /* Drop ref taken by lnet_nid2peerni_locked() */
+                       lnet_peer_ni_decref_locked(src_lpni);
+                       src_lp = lpni->lpni_peer_net->lpn_peer;
+                       if (lnet_peer_is_multi_rail(src_lp) &&
+                           !lnet_is_peer_ni_alive(lpni))
+                               mr_forwarding_allowed = true;
+
+               }
+               CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
+                      mr_forwarding_allowed ? "allowed" : "not allowed");
+       }
+
        /*
         * Deal with the peer as NMR in the following cases:
         * 1. the peer is NMR
         * 2. We're trying to recover a specific peer NI
-        * 3. I'm a router sending to the final destination
+        * 3. I'm a router sending to the final destination and MR forwarding is
+        *    not allowed for this message (as determined above).
         *    In this case the source of the message would've
         *    already selected the final destination so my job
         *    is to honor the selection.
         */
        if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
-           (msg->msg_routing && (send_case & LOCAL_DST)))
+           (final_hop && !mr_forwarding_allowed))
                send_case |= NMR_DST;
        else
                send_case |= MR_DST;
@@ -3584,11 +3752,11 @@ lnet_send_ping(lnet_nid_t dest_nid,
        md.length    = LNET_PING_INFO_SIZE(nnis);
        md.threshold = 2; /* GET/REPLY */
        md.max_size  = 0;
-       md.options   = LNET_MD_TRUNCATE;
+       md.options   = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
        md.user_ptr  = user_data;
        md.handler   = handler;
 
-       rc = LNetMDBind(md, LNET_UNLINK, mdh);
+       rc = LNetMDBind(&md, LNET_UNLINK, mdh);
        if (rc) {
                lnet_ping_buffer_decref(pbuf);
                CERROR("Can't bind MD: %d\n", rc);
@@ -4172,68 +4340,6 @@ lnet_msgtyp2str (int type)
        }
 }
 
-void
-lnet_print_hdr(struct lnet_hdr *hdr)
-{
-       struct lnet_process_id src = {
-               .nid = hdr->src_nid,
-               .pid = hdr->src_pid,
-       };
-       struct lnet_process_id dst = {
-               .nid = hdr->dest_nid,
-               .pid = hdr->dest_pid,
-       };
-       char *type_str = lnet_msgtyp2str(hdr->type);
-
-       CWARN("P3 Header at %p of type %s\n", hdr, type_str);
-       CWARN("    From %s\n", libcfs_id2str(src));
-       CWARN("    To   %s\n", libcfs_id2str(dst));
-
-       switch (hdr->type) {
-       default:
-               break;
-
-       case LNET_MSG_PUT:
-               CWARN("    Ptl index %d, ack md %#llx.%#llx, "
-                     "match bits %llu\n",
-                     hdr->msg.put.ptl_index,
-                     hdr->msg.put.ack_wmd.wh_interface_cookie,
-                     hdr->msg.put.ack_wmd.wh_object_cookie,
-                     hdr->msg.put.match_bits);
-               CWARN("    Length %d, offset %d, hdr data %#llx\n",
-                     hdr->payload_length, hdr->msg.put.offset,
-                     hdr->msg.put.hdr_data);
-               break;
-
-       case LNET_MSG_GET:
-               CWARN("    Ptl index %d, return md %#llx.%#llx, "
-                     "match bits %llu\n", hdr->msg.get.ptl_index,
-                     hdr->msg.get.return_wmd.wh_interface_cookie,
-                     hdr->msg.get.return_wmd.wh_object_cookie,
-                     hdr->msg.get.match_bits);
-               CWARN("    Length %d, src offset %d\n",
-                     hdr->msg.get.sink_length,
-                     hdr->msg.get.src_offset);
-               break;
-
-       case LNET_MSG_ACK:
-               CWARN("    dst md %#llx.%#llx, "
-                     "manipulated length %d\n",
-                     hdr->msg.ack.dst_wmd.wh_interface_cookie,
-                     hdr->msg.ack.dst_wmd.wh_object_cookie,
-                     hdr->msg.ack.mlength);
-               break;
-
-       case LNET_MSG_REPLY:
-               CWARN("    dst md %#llx.%#llx, "
-                     "length %d\n",
-                     hdr->msg.reply.dst_wmd.wh_interface_cookie,
-                     hdr->msg.reply.dst_wmd.wh_object_cookie,
-                     hdr->payload_length);
-       }
-
-}
-
 int
 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
           void *private, int rdma_req)
@@ -4308,11 +4414,7 @@ lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
                spin_lock(&ni->ni_net->net_lock);
                ni->ni_net->net_last_alive = ktime_get_real_seconds();
                spin_unlock(&ni->ni_net->net_lock);
-               if (ni->ni_status != NULL &&
-                   ni->ni_status->ns_status == LNET_NI_STATUS_DOWN) {
-                       ni->ni_status->ns_status = LNET_NI_STATUS_UP;
-                       push = true;
-               }
+               push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
                lnet_ni_unlock(ni);
        }
 
@@ -4763,7 +4865,9 @@ LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
                               md->md_me->me_portal);
                lnet_res_unlock(cpt);
 
-               lnet_rspt_free(rspt, cpt);
+               if (rspt)
+                       lnet_rspt_free(rspt, cpt);
+
                lnet_msg_free(msg);
                return -ENOENT;
        }
@@ -4796,8 +4900,11 @@ LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
 
        lnet_build_msg_event(msg, LNET_EVENT_SEND);
 
-       if (ack == LNET_ACK_REQ)
+       if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
+                                                  md->md_options))
                lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
+       else if (rspt)
+               lnet_rspt_free(rspt, cpt);
 
        if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
                                 CFS_FAIL_ONCE))
@@ -5015,7 +5122,10 @@ LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
 
        lnet_build_msg_event(msg, LNET_EVENT_SEND);
 
-       lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
+       if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
+               lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
+       else
+               lnet_rspt_free(rspt, cpt);
 
        rc = lnet_send(self, msg, LNET_NID_ANY);
        if (rc < 0) {