Whamcloud - gitweb
LU-9549 lnet: prevent assert on ln_state
[fs/lustre-release.git] / lnet / lnet / lib-move.c
index c98ec94..b05df55 100644 (file)
@@ -47,7 +47,7 @@ MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
 int
 lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
 {
-       lnet_test_peer_t *tp;
+       struct lnet_test_peer *tp;
        struct list_head *el;
        struct list_head *next;
        struct list_head  cull;
@@ -74,7 +74,7 @@ lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
        lnet_net_lock(0);
 
        list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
-               tp = list_entry(el, lnet_test_peer_t, tp_list);
+               tp = list_entry(el, struct lnet_test_peer, tp_list);
 
                if (tp->tp_threshold == 0 ||    /* needs culling anyway */
                    nid == LNET_NID_ANY ||      /* removing all entries */
@@ -87,7 +87,7 @@ lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
        lnet_net_unlock(0);
 
        while (!list_empty(&cull)) {
-               tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
+               tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 
                list_del(&tp->tp_list);
                LIBCFS_FREE(tp, sizeof(*tp));
@@ -98,7 +98,7 @@ lnet_fail_nid(lnet_nid_t nid, unsigned int threshold)
 static int
 fail_peer (lnet_nid_t nid, int outgoing)
 {
-       lnet_test_peer_t *tp;
+       struct lnet_test_peer *tp;
        struct list_head *el;
        struct list_head *next;
        struct list_head  cull;
@@ -110,7 +110,7 @@ fail_peer (lnet_nid_t nid, int outgoing)
        lnet_net_lock(0);
 
        list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
-               tp = list_entry(el, lnet_test_peer_t, tp_list);
+               tp = list_entry(el, struct lnet_test_peer, tp_list);
 
                if (tp->tp_threshold == 0) {
                        /* zombie entry */
@@ -144,7 +144,7 @@ fail_peer (lnet_nid_t nid, int outgoing)
        lnet_net_unlock(0);
 
        while (!list_empty(&cull)) {
-               tp = list_entry(cull.next, lnet_test_peer_t, tp_list);
+               tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
                list_del(&tp->tp_list);
 
                LIBCFS_FREE(tp, sizeof(*tp));
@@ -558,8 +558,9 @@ lnet_extract_kiov(int dst_niov, lnet_kiov_t *dst,
 EXPORT_SYMBOL(lnet_extract_kiov);
 
 void
-lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
-            unsigned int offset, unsigned int mlen, unsigned int rlen)
+lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
+            int delayed, unsigned int offset, unsigned int mlen,
+            unsigned int rlen)
 {
        unsigned int  niov = 0;
        struct kvec *iov = NULL;
@@ -593,13 +594,13 @@ lnet_ni_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
                                             niov, iov, kiov, offset, mlen,
                                             rlen);
        if (rc < 0)
-               lnet_finalize(ni, msg, rc);
+               lnet_finalize(msg, rc);
 }
 
 static void
-lnet_setpayloadbuffer(lnet_msg_t *msg)
+lnet_setpayloadbuffer(struct lnet_msg *msg)
 {
-       lnet_libmd_t *md = msg->msg_md;
+       struct lnet_libmd *md = msg->msg_md;
 
        LASSERT(msg->msg_len > 0);
        LASSERT(!msg->msg_routing);
@@ -616,7 +617,7 @@ lnet_setpayloadbuffer(lnet_msg_t *msg)
 }
 
 void
-lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
+lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
               unsigned int offset, unsigned int len)
 {
        msg->msg_type = type;
@@ -636,7 +637,7 @@ lnet_prep_send(lnet_msg_t *msg, int type, lnet_process_id_t target,
 }
 
 static void
-lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 {
        void   *priv = msg->msg_private;
        int     rc;
@@ -647,11 +648,11 @@ lnet_ni_send(lnet_ni_t *ni, lnet_msg_t *msg)
 
        rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
        if (rc < 0)
-               lnet_finalize(ni, msg, rc);
+               lnet_finalize(msg, rc);
 }
 
 static int
-lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 {
        int     rc;
 
@@ -683,7 +684,7 @@ lnet_ni_eager_recv(lnet_ni_t *ni, lnet_msg_t *msg)
  * lock before calling lnd_query()
  */
 static void
-lnet_ni_query_locked(lnet_ni_t *ni, struct lnet_peer_ni *lp)
+lnet_ni_query_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
 {
        cfs_time_t last_alive = 0;
        int cpt = lnet_cpt_of_nid_locked(lp->lpni_nid, ni);
@@ -802,7 +803,7 @@ lnet_peer_alive_locked (struct lnet_ni *ni, struct lnet_peer_ni *lp)
  * \retval -ECANCELED If the MD of the message has been unlinked.
  */
 static int
-lnet_post_send_locked(lnet_msg_t *msg, int do_send)
+lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 {
        struct lnet_peer_ni     *lp = msg->msg_txpeer;
        struct lnet_ni          *ni = msg->msg_txni;
@@ -828,7 +829,7 @@ lnet_post_send_locked(lnet_msg_t *msg, int do_send)
                CNETERR("Dropping message for %s: peer not alive\n",
                        libcfs_id2str(msg->msg_target));
                if (do_send)
-                       lnet_finalize(ni, msg, -EHOSTUNREACH);
+                       lnet_finalize(msg, -EHOSTUNREACH);
 
                lnet_net_lock(cpt);
                return -EHOSTUNREACH;
@@ -842,7 +843,7 @@ lnet_post_send_locked(lnet_msg_t *msg, int do_send)
                        "called on the MD/ME.\n",
                        libcfs_id2str(msg->msg_target));
                if (do_send)
-                       lnet_finalize(ni, msg, -ECANCELED);
+                       lnet_finalize(msg, -ECANCELED);
 
                lnet_net_lock(cpt);
                return -ECANCELED;
@@ -854,7 +855,7 @@ lnet_post_send_locked(lnet_msg_t *msg, int do_send)
                        !list_empty(&lp->lpni_txq));
 
                msg->msg_peertxcredit = 1;
-               lp->lpni_txqnob += msg->msg_len + sizeof(lnet_hdr_t);
+               lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
                lp->lpni_txcredits--;
 
                if (lp->lpni_txcredits < lp->lpni_mintxcredits)
@@ -896,10 +897,10 @@ lnet_post_send_locked(lnet_msg_t *msg, int do_send)
 }
 
 
-static lnet_rtrbufpool_t *
-lnet_msg2bufpool(lnet_msg_t *msg)
+static struct lnet_rtrbufpool *
+lnet_msg2bufpool(struct lnet_msg *msg)
 {
-       lnet_rtrbufpool_t       *rbp;
+       struct lnet_rtrbufpool  *rbp;
        int                     cpt;
 
        LASSERT(msg->msg_rx_committed);
@@ -917,15 +918,15 @@ lnet_msg2bufpool(lnet_msg_t *msg)
 }
 
 static int
-lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
+lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 {
        /* lnet_parse is going to lnet_net_unlock immediately after this, so it
         * sets do_recv FALSE and I don't do the unlock/send/lock bit.
         * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
         * received or OK to receive */
        struct lnet_peer_ni *lp = msg->msg_rxpeer;
-       lnet_rtrbufpool_t   *rbp;
-       lnet_rtrbuf_t       *rb;
+       struct lnet_rtrbufpool *rbp;
+       struct lnet_rtrbuf *rb;
 
        LASSERT (msg->msg_iov == NULL);
        LASSERT (msg->msg_kiov == NULL);
@@ -976,7 +977,7 @@ lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
        }
 
        LASSERT(!list_empty(&rbp->rbp_bufs));
-       rb = list_entry(rbp->rbp_bufs.next, lnet_rtrbuf_t, rb_list);
+       rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
        list_del(&rb->rb_list);
 
        msg->msg_niov = rbp->rbp_npages;
@@ -994,11 +995,11 @@ lnet_post_routed_recv_locked (lnet_msg_t *msg, int do_recv)
 }
 
 void
-lnet_return_tx_credits_locked(lnet_msg_t *msg)
+lnet_return_tx_credits_locked(struct lnet_msg *msg)
 {
        struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
        struct lnet_ni          *txni = msg->msg_txni;
-       lnet_msg_t              *msg2;
+       struct lnet_msg         *msg2;
 
        if (msg->msg_txcredit) {
                struct lnet_ni       *ni = msg->msg_txni;
@@ -1014,7 +1015,7 @@ lnet_return_tx_credits_locked(lnet_msg_t *msg)
                atomic_inc(&ni->ni_tx_credits);
                if (tq->tq_credits <= 0) {
                        msg2 = list_entry(tq->tq_delayed.next,
-                                         lnet_msg_t, msg_list);
+                                         struct lnet_msg, msg_list);
                        list_del(&msg2->msg_list);
 
                        LASSERT(msg2->msg_txni == ni);
@@ -1033,13 +1034,13 @@ lnet_return_tx_credits_locked(lnet_msg_t *msg)
                LASSERT((txpeer->lpni_txcredits < 0) ==
                        !list_empty(&txpeer->lpni_txq));
 
-               txpeer->lpni_txqnob -= msg->msg_len + sizeof(lnet_hdr_t);
+               txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
                LASSERT(txpeer->lpni_txqnob >= 0);
 
                txpeer->lpni_txcredits++;
                if (txpeer->lpni_txcredits <= 0) {
                        msg2 = list_entry(txpeer->lpni_txq.next,
-                                              lnet_msg_t, msg_list);
+                                             struct lnet_msg, msg_list);
                        list_del(&msg2->msg_list);
                        spin_unlock(&txpeer->lpni_lock);
 
@@ -1081,14 +1082,14 @@ lnet_return_tx_credits_locked(lnet_msg_t *msg)
 }
 
 void
-lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
+lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
 {
-       lnet_msg_t      *msg;
+       struct lnet_msg *msg;
 
        if (list_empty(&rbp->rbp_msgs))
                return;
        msg = list_entry(rbp->rbp_msgs.next,
-                        lnet_msg_t, msg_list);
+                        struct lnet_msg, msg_list);
        list_del(&msg->msg_list);
 
        (void)lnet_post_routed_recv_locked(msg, 1);
@@ -1097,8 +1098,8 @@ lnet_schedule_blocked_locked(lnet_rtrbufpool_t *rbp)
 void
 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
 {
-       lnet_msg_t *msg;
-       lnet_msg_t *tmp;
+       struct lnet_msg *msg;
+       struct lnet_msg *tmp;
 
        lnet_net_unlock(cpt);
 
@@ -1106,30 +1107,30 @@ lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
                lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
                             0, 0, 0, msg->msg_hdr.payload_length);
                list_del_init(&msg->msg_list);
-               lnet_finalize(NULL, msg, -ECANCELED);
+               lnet_finalize(msg, -ECANCELED);
        }
 
        lnet_net_lock(cpt);
 }
 
 void
-lnet_return_rx_credits_locked(lnet_msg_t *msg)
+lnet_return_rx_credits_locked(struct lnet_msg *msg)
 {
-       struct lnet_peer_ni     *rxpeer = msg->msg_rxpeer;
-       struct lnet_ni          *rxni = msg->msg_rxni;
-       lnet_msg_t              *msg2;
+       struct lnet_peer_ni *rxpeer = msg->msg_rxpeer;
+       struct lnet_ni *rxni = msg->msg_rxni;
+       struct lnet_msg *msg2;
 
        if (msg->msg_rtrcredit) {
                /* give back global router credits */
-               lnet_rtrbuf_t     *rb;
-               lnet_rtrbufpool_t *rbp;
+               struct lnet_rtrbuf *rb;
+               struct lnet_rtrbufpool *rbp;
 
                /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
                 * there until it gets one allocated, or aborts the wait
                 * itself */
                LASSERT(msg->msg_kiov != NULL);
 
-               rb = list_entry(msg->msg_kiov, lnet_rtrbuf_t, rb_kiov[0]);
+               rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
                rbp = rb->rb_pool;
 
                msg->msg_kiov = NULL;
@@ -1184,7 +1185,7 @@ routing_off:
                        lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
                } else if (rxpeer->lpni_rtrcredits <= 0) {
                        msg2 = list_entry(rxpeer->lpni_rtrq.next,
-                                         lnet_msg_t, msg_list);
+                                         struct lnet_msg, msg_list);
                        list_del(&msg2->msg_list);
                        spin_unlock(&rxpeer->lpni_lock);
                        (void) lnet_post_routed_recv_locked(msg2, 1);
@@ -1221,7 +1222,7 @@ lnet_compare_peers(struct lnet_peer_ni *p1, struct lnet_peer_ni *p2)
 }
 
 static int
-lnet_compare_routes(lnet_route_t *r1, lnet_route_t *r2)
+lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
 {
        struct lnet_peer_ni *p1 = r1->lr_gateway;
        struct lnet_peer_ni *p2 = r2->lr_gateway;
@@ -1255,10 +1256,10 @@ static struct lnet_peer_ni *
 lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
                       lnet_nid_t rtr_nid)
 {
-       lnet_remotenet_t        *rnet;
-       lnet_route_t            *route;
-       lnet_route_t            *best_route;
-       lnet_route_t            *last_route;
+       struct lnet_remotenet   *rnet;
+       struct lnet_route               *route;
+       struct lnet_route               *best_route;
+       struct lnet_route               *last_route;
        struct lnet_peer_ni     *lpni_best;
        struct lnet_peer_ni     *lp;
        int                     rc;
@@ -1310,6 +1311,69 @@ lnet_find_route_locked(struct lnet_net *net, lnet_nid_t target,
        return lpni_best;
 }
 
+static struct lnet_ni *
+lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *cur_ni,
+                int md_cpt)
+{
+       struct lnet_ni *ni = NULL, *best_ni = cur_ni;
+       unsigned int shortest_distance;
+       int best_credits;
+
+       if (best_ni == NULL) {
+               shortest_distance = UINT_MAX;
+               best_credits = INT_MIN;
+       } else {
+               shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
+                                                    best_ni->ni_dev_cpt);
+               best_credits = atomic_read(&best_ni->ni_tx_credits);
+       }
+
+       while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
+               unsigned int distance;
+               int ni_credits;
+
+               if (!lnet_is_ni_healthy_locked(ni))
+                       continue;
+
+               ni_credits = atomic_read(&ni->ni_tx_credits);
+
+               /*
+                * calculate the distance from the CPT on which
+                * the message memory is allocated to the CPT of
+                * the NI's physical device
+                */
+               distance = cfs_cpt_distance(lnet_cpt_table(),
+                                           md_cpt,
+                                           ni->ni_dev_cpt);
+
+               /*
+                * All distances smaller than the NUMA range
+                * are treated equally.
+                */
+               if (distance < lnet_numa_range)
+                       distance = lnet_numa_range;
+
+               /*
+                * Select on shorter distance, then available
+                * credits, then round-robin.
+                */
+               if (distance > shortest_distance) {
+                       continue;
+               } else if (distance < shortest_distance) {
+                       shortest_distance = distance;
+               } else if (ni_credits < best_credits) {
+                       continue;
+               } else if (ni_credits == best_credits) {
+                       if (best_ni && (best_ni)->ni_seq <= ni->ni_seq)
+                               continue;
+               }
+               best_ni = ni;
+               best_credits = ni_credits;
+       }
+
+       return best_ni;
+}
+
 static int
 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
                    struct lnet_msg *msg, lnet_nid_t rtr_nid)
@@ -1318,19 +1382,19 @@ lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
        struct lnet_peer_ni     *best_lpni;
        struct lnet_peer_ni     *best_gw;
        struct lnet_peer_ni     *lpni;
+       struct lnet_peer_ni     *final_dst;
        struct lnet_peer        *peer;
        struct lnet_peer_net    *peer_net;
        struct lnet_net         *local_net;
-       struct lnet_ni          *ni;
        __u32                   seq;
        int                     cpt, cpt2, rc;
        bool                    routing;
+       bool                    routing2;
        bool                    ni_is_pref;
        bool                    preferred;
-       int                     best_credits;
+       bool                    local_found;
        int                     best_lpni_credits;
        int                     md_cpt;
-       int                     shortest_distance;
 
        /*
         * get an initial CPT to use for locking. The idea here is not to
@@ -1342,26 +1406,28 @@ lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
         * then we proceed, if there is, then we restart the operation.
         */
        cpt = lnet_net_lock_current();
+
+       md_cpt = lnet_cpt_of_md(msg->msg_md);
+       if (md_cpt == CFS_CPT_ANY)
+               md_cpt = cpt;
+
 again:
        best_ni = NULL;
        best_lpni = NULL;
        best_gw = NULL;
+       final_dst = NULL;
        local_net = NULL;
        routing = false;
+       routing2 = false;
+       local_found = false;
 
        seq = lnet_get_dlc_seq_locked();
 
-       if (the_lnet.ln_shutdown) {
+       if (the_lnet.ln_state != LNET_STATE_RUNNING) {
                lnet_net_unlock(cpt);
                return -ESHUTDOWN;
        }
 
-       if (msg->msg_md != NULL)
-               /* get the cpt of the MD, used during NUMA based selection */
-               md_cpt = lnet_cpt_of_cookie(msg->msg_md->md_lh.lh_cookie);
-       else
-               md_cpt = CFS_CPT_ANY;
-
        peer = lnet_find_or_create_peer_locked(dst_nid, cpt);
        if (IS_ERR(peer)) {
                lnet_net_unlock(cpt);
@@ -1382,7 +1448,7 @@ again:
        }
 
        /*
-        * STEP 1: first jab at determineing best_ni
+        * STEP 1: first jab at determining best_ni
         * if src_nid is explicitly specified, then best_ni is already
         * pre-determiend for us. Otherwise we need to select the best
         * one to use later on
@@ -1396,77 +1462,177 @@ again:
                                      libcfs_nid2str(src_nid));
                        return -EINVAL;
                }
+       }
+
+       if (msg->msg_type == LNET_MSG_REPLY ||
+           msg->msg_type == LNET_MSG_ACK ||
+           !peer->lp_multi_rail ||
+           best_ni) {
+               /*
+                * for replies we want to respond on the same peer_ni we
+                * received the message on if possible. If not, then pick
+                * a peer_ni to send to
+                *
+                * if the peer is non-multi-rail then you want to send to
+                * the dst_nid provided as well.
+                *
+                * If the best_ni has already been determined, IE the
+                * src_nid has been specified, then use the
+                * destination_nid provided as well, since we're
+                * continuing a series of related messages for the same
+                * RPC.
+                *
+                * It is expected to find the lpni using dst_nid, since we
+                * created it earlier.
+                */
+               best_lpni = lnet_find_peer_ni_locked(dst_nid);
+               if (best_lpni)
+                       lnet_peer_ni_decref_locked(best_lpni);
+
+               if (best_lpni && !lnet_get_net_locked(LNET_NIDNET(dst_nid))) {
+                       /*
+                        * this lpni is not on a local network so we need
+                        * to route this reply.
+                        */
+                       best_gw = lnet_find_route_locked(NULL,
+                                                        best_lpni->lpni_nid,
+                                                        rtr_nid);
+                       if (best_gw) {
+                               /*
+                               * RULE: Each node considers only the next-hop
+                               *
+                               * We're going to route the message, so change the peer to
+                               * the router.
+                               */
+                               LASSERT(best_gw->lpni_peer_net);
+                               LASSERT(best_gw->lpni_peer_net->lpn_peer);
+                               peer = best_gw->lpni_peer_net->lpn_peer;
+
+                               /*
+                               * if the router is not multi-rail then use the best_gw
+                               * found to send the message to
+                               */
+                               if (!peer->lp_multi_rail)
+                                       best_lpni = best_gw;
+                               else
+                                       best_lpni = NULL;
 
-               if (best_ni->ni_net->net_id != LNET_NIDNET(dst_nid)) {
+                               routing = true;
+                       } else {
+                               best_lpni = NULL;
+                       }
+               } else if (!best_lpni) {
                        lnet_net_unlock(cpt);
-                       LCONSOLE_WARN("No route to %s via from %s\n",
-                                     libcfs_nid2str(dst_nid),
-                                     libcfs_nid2str(src_nid));
+                       CERROR("unable to send msg_type %d to "
+                             "originating %s. Destination NID not in DB\n",
+                             msg->msg_type, libcfs_nid2str(dst_nid));
                        return -EINVAL;
                }
-               goto pick_peer;
        }
 
        /*
-        * Decide whether we need to route to peer_ni.
-        * Get the local net that I need to be on to be able to directly
-        * send to that peer.
+        * if the peer is not MR capable, then we should always send to it
+        * using the first NI in the NET we determined.
+        */
+       if (!peer->lp_multi_rail) {
+               if (!best_lpni) {
+                       lnet_net_unlock(cpt);
+                       CERROR("no route to %s\n",
+                              libcfs_nid2str(dst_nid));
+                       return -EHOSTUNREACH;
+               }
+
+               /* best ni could be set because src_nid was provided */
+               if (!best_ni) {
+                       best_ni = lnet_net2ni_locked(best_lpni->lpni_net->net_id, cpt);
+                       if (!best_ni) {
+                               lnet_net_unlock(cpt);
+                               CERROR("no path to %s from net %s\n",
+                               libcfs_nid2str(best_lpni->lpni_nid),
+                               libcfs_net2str(best_lpni->lpni_net->net_id));
+                               return -EHOSTUNREACH;
+                       }
+               }
+       }
+
+       /*
+        * if we already found a best_ni because src_nid is specified and
+        * best_lpni because we are replying to a message then just send
+        * the message
+        */
+       if (best_ni && best_lpni)
+               goto send;
+
+       /*
+        * If we already found a best_ni because src_nid is specified then
+        * pick the peer then send the message
+        */
+       if (best_ni)
+               goto pick_peer;
+
+       /*
+        * pick the best_ni by going through all the possible networks of
+        * that peer and see which local NI is best suited to talk to that
+        * peer.
         *
-        * a. Find the peer which the dst_nid belongs to.
-        * b. Iterate through each of the peer_nets/nis to decide
-        * the best peer/local_ni pair to use
+        * Locally connected networks will always be preferred over
+        * a routed network. If there are only routed paths to the peer,
+        * then the best route is chosen. If all routes are equal then
+        * they are used in round robin.
         */
-       shortest_distance = INT_MAX;
-       best_credits = INT_MIN;
        list_for_each_entry(peer_net, &peer->lp_peer_nets, lpn_on_peer_list) {
                if (!lnet_is_peer_net_healthy_locked(peer_net))
                        continue;
 
                local_net = lnet_get_net_locked(peer_net->lpn_net_id);
-               if (!local_net) {
+               if (!local_net && !routing && !local_found) {
                        struct lnet_peer_ni *net_gw;
-                       /*
-                        * go through each peer_ni on that peer_net and
-                        * determine the best possible gw to go through
-                        */
-                       list_for_each_entry(lpni, &peer_net->lpn_peer_nis,
-                                           lpni_on_peer_net_list) {
-                               net_gw = lnet_find_route_locked(NULL,
-                                                               lpni->lpni_nid,
-                                                               rtr_nid);
 
+                       lpni = list_entry(peer_net->lpn_peer_nis.next,
+                                         struct lnet_peer_ni,
+                                         lpni_on_peer_net_list);
+
+                       net_gw = lnet_find_route_locked(NULL,
+                                                       lpni->lpni_nid,
+                                                       rtr_nid);
+                       if (!net_gw)
+                               continue;
+
+                       if (best_gw) {
                                /*
-                                * if no route is found for that network then
-                                * move onto the next peer_ni in the peer
+                                * lnet_find_route_locked() call
+                                * will return the best_Gw on the
+                                * lpni->lpni_nid network.
+                                * However, best_gw and net_gw can
+                                * be on different networks.
+                                * Therefore need to compare them
+                                * to pick the better of either.
                                 */
-                               if (!net_gw)
+                               if (lnet_compare_peers(best_gw, net_gw) > 0)
+                                       continue;
+                               if (best_gw->lpni_gw_seq <= net_gw->lpni_gw_seq)
                                        continue;
-
-                               if (!best_gw) {
-                                       best_gw = net_gw;
-                                       best_lpni = lpni;
-                               } else  {
-                                       rc = lnet_compare_peers(net_gw,
-                                                               best_gw);
-                                       if (rc > 0) {
-                                               best_gw = net_gw;
-                                               best_lpni = lpni;
-                                       }
-                               }
                        }
+                       best_gw = net_gw;
+                       final_dst = lpni;
 
-                       if (!best_gw)
-                               continue;
-
-                       local_net = lnet_get_net_locked
-                                       (LNET_NIDNET(best_gw->lpni_nid));
-                       routing = true;
+                       routing2 = true;
                } else {
-                       routing = false;
                        best_gw = NULL;
+                       final_dst = NULL;
+                       routing2 = false;
+                       local_found = true;
                }
 
-               /* no routable net found go on to a different net */
+               /*
+                * a gw on this network is found, but there could be
+                * other better gateways on other networks. So don't pick
+                * the best_ni until we determine the best_gw.
+                */
+               if (best_gw)
+                       continue;
+
+               /* if no local_net found continue */
                if (!local_net)
                        continue;
 
@@ -1478,65 +1644,30 @@ again:
                 *      2. NI available credits
                 *      3. Round Robin
                 */
-               ni = NULL;
-               while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
-                       int ni_credits;
-                       int distance;
-
-                       if (!lnet_is_ni_healthy_locked(ni))
-                               continue;
-
-                       ni_credits = atomic_read(&ni->ni_tx_credits);
-
-                       /*
-                        * calculate the distance from the CPT on which
-                        * the message memory is allocated to the CPT of
-                        * the NI's physical device
-                        */
-                       distance = cfs_cpt_distance(lnet_cpt_table(),
-                                                   md_cpt,
-                                                   ni->dev_cpt);
-
-                       /*
-                        * All distances smaller than the NUMA range
-                        * are treated equally.
-                        */
-                       if (distance < lnet_get_numa_range())
-                               distance = lnet_get_numa_range();
-
-                       /*
-                        * Select on shorter distance, then available
-                        * credits, then round-robin.
-                        */
-                       if (distance > shortest_distance) {
-                               continue;
-                       } else if (distance < shortest_distance) {
-                               shortest_distance = distance;
-                       } else if (ni_credits < best_credits) {
-                               continue;
-                       } else if (ni_credits == best_credits) {
-                               if (best_ni && best_ni->ni_seq <= ni->ni_seq)
-                                       continue;
-                       }
-                       best_ni = ni;
-                       best_credits = ni_credits;
-               }
+               best_ni = lnet_get_best_ni(local_net, best_ni, md_cpt);
        }
 
-       /*
-        * if the peer is not MR capable, then we should always send to it
-        * using the first NI in the NET we determined.
-        */
-       if (!peer->lp_multi_rail && local_net != NULL)
-               best_ni = lnet_net2ni_locked(local_net->net_id, cpt);
-
-       if (!best_ni) {
+       if (!best_ni && !best_gw) {
                lnet_net_unlock(cpt);
                LCONSOLE_WARN("No local ni found to send from to %s\n",
                        libcfs_nid2str(dst_nid));
                return -EINVAL;
        }
 
+       if (!best_ni) {
+               best_ni = lnet_get_best_ni(best_gw->lpni_net, best_ni, md_cpt);
+               LASSERT(best_gw && best_ni);
+
+               /*
+                * We're going to route the message, so change the peer to
+                * the router.
+                */
+               LASSERT(best_gw->lpni_peer_net);
+               LASSERT(best_gw->lpni_peer_net->lpn_peer);
+               best_gw->lpni_gw_seq++;
+               peer = best_gw->lpni_peer_net->lpn_peer;
+       }
+
        /*
         * Now that we selected the NI to use increment its sequence
         * number so the Round Robin algorithm will detect that it has
@@ -1544,42 +1675,11 @@ again:
         */
        best_ni->ni_seq++;
 
-       if (routing)
-               goto send;
-
 pick_peer:
-       if (best_ni == the_lnet.ln_loni) {
-               /* No send credit hassles with LOLND */
-               lnet_ni_addref_locked(best_ni, cpt);
-               msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
-               if (!msg->msg_routing)
-                       msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
-               msg->msg_target.nid = best_ni->ni_nid;
-               lnet_msg_commit(msg, cpt);
-               msg->msg_txni = best_ni;
-               lnet_net_unlock(cpt);
-
-               return LNET_CREDIT_OK;
-       }
-
-       if (msg->msg_type == LNET_MSG_REPLY ||
-           msg->msg_type == LNET_MSG_ACK) {
-               /*
-                * for replies we want to respond on the same peer_ni we
-                * received the message on if possible. If not, then pick
-                * a peer_ni to send to
-                */
-               best_lpni = lnet_find_peer_ni_locked(dst_nid);
-               if (best_lpni) {
-                       lnet_peer_ni_decref_locked(best_lpni);
-                       goto send;
-               } else {
-                       CDEBUG(D_NET, "unable to send msg_type %d to "
-                             "originating %s\n", msg->msg_type,
-                             libcfs_nid2str(dst_nid));
-               }
-       }
-
+       /*
+        * At this point the best_ni is on a local network on which
+        * the peer has a peer_ni as well
+        */
        peer_net = lnet_peer_get_net_locked(peer,
                                            best_ni->ni_net->net_id);
        /*
@@ -1609,13 +1709,16 @@ pick_peer:
                        libcfs_nid2str(best_gw->lpni_nid),
                        lnet_msgtyp2str(msg->msg_type), msg->msg_len);
 
-               best_lpni = lnet_find_peer_ni_locked(dst_nid);
-               LASSERT(best_lpni != NULL);
-               lnet_peer_ni_decref_locked(best_lpni);
-
-               routing = true;
-
-               goto send;
+               routing2 = true;
+               /*
+                * RULE: Each node considers only the next-hop
+                *
+                * We're going to route the message, so change the peer to
+                * the router.
+                */
+               LASSERT(best_gw->lpni_peer_net);
+               LASSERT(best_gw->lpni_peer_net->lpn_peer);
+               peer = best_gw->lpni_peer_net->lpn_peer;
        } else if (!lnet_is_peer_net_healthy_locked(peer_net)) {
                /*
                 * this peer_net is unhealthy but we still have an opportunity
@@ -1638,6 +1741,7 @@ pick_peer:
        lpni = NULL;
        best_lpni_credits = INT_MIN;
        preferred = false;
+       best_lpni = NULL;
        while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
                /*
                 * if this peer ni is not healthy just skip it, no point in
@@ -1682,14 +1786,33 @@ pick_peer:
 
        /* if we still can't find a peer ni then we can't reach it */
        if (!best_lpni) {
-               __u32 net_id = peer_net->lpn_net_id;
+               __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
+                       LNET_NIDNET(dst_nid);
                lnet_net_unlock(cpt);
                LCONSOLE_WARN("no peer_ni found on peer net %s\n",
                                libcfs_net2str(net_id));
                return -EHOSTUNREACH;
        }
 
+
 send:
+       /* Shortcut for loopback. */
+       if (best_ni == the_lnet.ln_loni) {
+               /* No send credit hassles with LOLND */
+               lnet_ni_addref_locked(best_ni, cpt);
+               msg->msg_hdr.dest_nid = cpu_to_le64(best_ni->ni_nid);
+               if (!msg->msg_routing)
+                       msg->msg_hdr.src_nid = cpu_to_le64(best_ni->ni_nid);
+               msg->msg_target.nid = best_ni->ni_nid;
+               lnet_msg_commit(msg, cpt);
+               msg->msg_txni = best_ni;
+               lnet_net_unlock(cpt);
+
+               return LNET_CREDIT_OK;
+       }
+
+       routing = routing || routing2;
+
        /*
         * Increment sequence number of the peer selected so that we
         * pick the next one in Round Robin.
@@ -1697,13 +1820,6 @@ send:
        best_lpni->lpni_seq++;
 
        /*
-        * When routing the best gateway found acts as the best peer
-        * NI to send to.
-        */
-       if (routing)
-               best_lpni = best_gw;
-
-       /*
         * grab a reference on the peer_ni so it sticks around even if
         * we need to drop and relock the lnet_net_lock below.
         */
@@ -1775,7 +1891,8 @@ send:
                 * the router receives this message it knows how to route
                 * it.
                 */
-               msg->msg_hdr.dest_nid = cpu_to_le64(dst_nid);
+               msg->msg_hdr.dest_nid =
+                       cpu_to_le64(final_dst ? final_dst->lpni_nid : dst_nid);
        } else {
                /*
                 * if we're not routing set the dest_nid to the best peer
@@ -1792,7 +1909,7 @@ send:
 }
 
 int
-lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
+lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
 {
        lnet_nid_t              dst_nid = msg->msg_target.nid;
        int                     rc;
@@ -1824,7 +1941,7 @@ lnet_send(lnet_nid_t src_nid, lnet_msg_t *msg, lnet_nid_t rtr_nid)
 }
 
 void
-lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
+lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob)
 {
        lnet_net_lock(cpt);
        the_lnet.ln_counters[cpt]->drop_count++;
@@ -1835,9 +1952,9 @@ lnet_drop_message(lnet_ni_t *ni, int cpt, void *private, unsigned int nob)
 }
 
 static void
-lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
 {
-       lnet_hdr_t      *hdr = &msg->msg_hdr;
+       struct lnet_hdr *hdr = &msg->msg_hdr;
 
        if (msg->msg_wanted != 0)
                lnet_setpayloadbuffer(msg);
@@ -1854,9 +1971,9 @@ lnet_recv_put(lnet_ni_t *ni, lnet_msg_t *msg)
 }
 
 static int
-lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
 {
-       lnet_hdr_t              *hdr = &msg->msg_hdr;
+       struct lnet_hdr         *hdr = &msg->msg_hdr;
        struct lnet_match_info  info;
        int                     rc;
        bool                    ready_delay;
@@ -1913,13 +2030,13 @@ lnet_parse_put(lnet_ni_t *ni, lnet_msg_t *msg)
 }
 
 static int
-lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
+lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
 {
-       struct lnet_match_info  info;
-       lnet_hdr_t              *hdr = &msg->msg_hdr;
-       lnet_process_id_t       source_id;
+       struct lnet_match_info info;
+       struct lnet_hdr *hdr = &msg->msg_hdr;
+       struct lnet_process_id source_id;
        struct lnet_handle_wire reply_wmd;
-       int                     rc;
+       int rc;
 
        /* Convert get fields to host byte order */
        hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
@@ -1976,19 +2093,19 @@ lnet_parse_get(lnet_ni_t *ni, lnet_msg_t *msg, int rdma_get)
                       libcfs_nid2str(ni->ni_nid),
                       libcfs_id2str(info.mi_id), rc);
 
-               lnet_finalize(ni, msg, rc);
+               lnet_finalize(msg, rc);
        }
 
        return 0;
 }
 
 static int
-lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
 {
        void             *private = msg->msg_private;
-       lnet_hdr_t       *hdr = &msg->msg_hdr;
-       lnet_process_id_t src = {0};
-       lnet_libmd_t     *md;
+       struct lnet_hdr  *hdr = &msg->msg_hdr;
+       struct lnet_process_id src = {0};
+       struct lnet_libmd        *md;
        int               rlength;
        int               mlength;
        int                     cpt;
@@ -2050,11 +2167,11 @@ lnet_parse_reply(lnet_ni_t *ni, lnet_msg_t *msg)
 }
 
 static int
-lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
 {
-       lnet_hdr_t       *hdr = &msg->msg_hdr;
-       lnet_process_id_t src = {0};
-       lnet_libmd_t     *md;
+       struct lnet_hdr  *hdr = &msg->msg_hdr;
+       struct lnet_process_id src = {0};
+       struct lnet_libmd        *md;
        int                     cpt;
 
        src.nid = hdr->src_nid;
@@ -2105,7 +2222,7 @@ lnet_parse_ack(lnet_ni_t *ni, lnet_msg_t *msg)
  * \retval -ve                 error code
  */
 int
-lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
 {
        int     rc = 0;
 
@@ -2129,7 +2246,7 @@ lnet_parse_forward_locked(lnet_ni_t *ni, lnet_msg_t *msg)
 }
 
 int
-lnet_parse_local(lnet_ni_t *ni, lnet_msg_t *msg)
+lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
 {
        int     rc;
 
@@ -2175,18 +2292,18 @@ lnet_msgtyp2str (int type)
 }
 
 void
-lnet_print_hdr(lnet_hdr_t * hdr)
+lnet_print_hdr(struct lnet_hdr *hdr)
 {
-       lnet_process_id_t src = {0};
-       lnet_process_id_t dst = {0};
+       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);
 
-       src.nid = hdr->src_nid;
-       src.pid = hdr->src_pid;
-
-       dst.nid = hdr->dest_nid;
-       dst.pid = hdr->dest_pid;
-
        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));
@@ -2237,7 +2354,7 @@ lnet_print_hdr(lnet_hdr_t * hdr)
 }
 
 int
-lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
+lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid,
           void *private, int rdma_req)
 {
        int             rc = 0;
@@ -2407,8 +2524,6 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
                msg->msg_hdr.dest_pid   = dest_pid;
                msg->msg_hdr.payload_length = payload_length;
        }
-       /* Multi-Rail: Primary NID of source. */
-       msg->msg_initiator = lnet_peer_primary_nid(src_nid);
 
        lnet_net_lock(cpt);
        lpni = lnet_nid2peerni_locked(from_nid, cpt);
@@ -2427,6 +2542,8 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
        msg->msg_rxpeer = lpni;
        msg->msg_rxni = ni;
        lnet_ni_addref_locked(ni, cpt);
+       /* Multi-Rail: Primary NID of source. */
+       msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
 
        if (lnet_isrouter(msg->msg_rxpeer)) {
                lnet_peer_set_alive(msg->msg_rxpeer);
@@ -2473,7 +2590,7 @@ lnet_parse(lnet_ni_t *ni, lnet_hdr_t *hdr, lnet_nid_t from_nid,
 
  free_drop:
        LASSERT(msg->msg_md == NULL);
-       lnet_finalize(ni, msg, rc);
+       lnet_finalize(msg, rc);
 
  drop:
        lnet_drop_message(ni, cpt, private, payload_length);
@@ -2485,10 +2602,10 @@ void
 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
 {
        while (!list_empty(head)) {
-               lnet_process_id_t       id = {0};
-               lnet_msg_t              *msg;
+               struct lnet_process_id id = {0};
+               struct lnet_msg *msg;
 
-               msg = list_entry(head->next, lnet_msg_t, msg_list);
+               msg = list_entry(head->next, struct lnet_msg, msg_list);
                list_del(&msg->msg_list);
 
                id.nid = msg->msg_hdr.src_nid;
@@ -2518,7 +2635,7 @@ lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
                 * but we still should give error code so lnet_msg_decommit()
                 * can skip counters operations and other checks.
                 */
-               lnet_finalize(msg->msg_rxni, msg, -ENOENT);
+               lnet_finalize(msg, -ENOENT);
        }
 }
 
@@ -2526,10 +2643,10 @@ void
 lnet_recv_delayed_msg_list(struct list_head *head)
 {
        while (!list_empty(head)) {
-               lnet_msg_t        *msg;
-               lnet_process_id_t  id;
+               struct lnet_msg *msg;
+               struct lnet_process_id id;
 
-               msg = list_entry(head->next, lnet_msg_t, msg_list);
+               msg = list_entry(head->next, struct lnet_msg, msg_list);
                list_del(&msg->msg_list);
 
                /* md won't disappear under me, since each msg
@@ -2597,11 +2714,11 @@ lnet_recv_delayed_msg_list(struct list_head *head)
  * \retval -ENOMEM Memory allocation failure.
  * \retval -ENOENT Invalid MD object.
  *
- * \see lnet_event_t::hdr_data and lnet_event_kind_t.
+ * \see struct lnet_event::hdr_data and lnet_event_kind_t.
  */
 int
-LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
-       lnet_process_id_t target, unsigned int portal,
+LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
+       struct lnet_process_id target, unsigned int portal,
        __u64 match_bits, unsigned int offset,
        __u64 hdr_data)
 {
@@ -2621,7 +2738,7 @@ LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
 
        msg = lnet_msg_alloc();
        if (msg == NULL) {
-               CERROR("Dropping PUT to %s: ENOMEM on lnet_msg_t\n",
+               CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
                       libcfs_id2str(target));
                return -ENOMEM;
        }
@@ -2676,7 +2793,7 @@ LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
        if (rc != 0) {
                CNETERR("Error sending PUT to %s: %d\n",
                        libcfs_id2str(target), rc);
-               lnet_finalize(NULL, msg, rc);
+               lnet_finalize(msg, rc);
        }
 
        /* completion will be signalled by an event */
@@ -2684,20 +2801,21 @@ LNetPut(lnet_nid_t self, lnet_handle_md_t mdh, lnet_ack_req_t ack,
 }
 EXPORT_SYMBOL(LNetPut);
 
-lnet_msg_t *
-lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
+/*
+ * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
+ * returns a msg for the LND to pass to lnet_finalize() when the sink
+ * data has been received.
+ *
+ * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
+ * lnet_finalize() is called on it, so the LND must call this first
+ */
+struct lnet_msg *
+lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
 {
-       /* The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
-        * returns a msg for the LND to pass to lnet_finalize() when the sink
-        * data has been received.
-        *
-        * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
-        * lnet_finalize() is called on it, so the LND must call this first */
-
-       struct lnet_msg         *msg = lnet_msg_alloc();
-       struct lnet_libmd       *getmd = getmsg->msg_md;
-       lnet_process_id_t       peer_id = getmsg->msg_target;
-       int                     cpt;
+       struct lnet_msg *msg = lnet_msg_alloc();
+       struct lnet_libmd *getmd = getmsg->msg_md;
+       struct lnet_process_id peer_id = getmsg->msg_target;
+       int cpt;
 
        LASSERT(!getmsg->msg_target_is_router);
        LASSERT(!getmsg->msg_routing);
@@ -2727,8 +2845,7 @@ lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
               libcfs_nid2str(ni->ni_nid), libcfs_id2str(peer_id), getmd);
 
        /* setup information for lnet_build_msg_event */
-       msg->msg_initiator = lnet_peer_primary_nid(peer_id.nid);
-       /* Cheaper: msg->msg_initiator = getmsg->msg_txpeer->lp_nid; */
+       msg->msg_initiator = getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid;
        msg->msg_from = peer_id.nid;
        msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
        msg->msg_hdr.src_nid = peer_id.nid;
@@ -2764,7 +2881,8 @@ lnet_create_reply_msg (lnet_ni_t *ni, lnet_msg_t *getmsg)
 EXPORT_SYMBOL(lnet_create_reply_msg);
 
 void
-lnet_set_reply_msg_len(lnet_ni_t *ni, lnet_msg_t *reply, unsigned int len)
+lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
+                      unsigned int len)
 {
        /* Set the REPLY length, now the RDMA that elides the REPLY message has
         * completed and I know it. */
@@ -2801,8 +2919,8 @@ EXPORT_SYMBOL(lnet_set_reply_msg_len);
  * \retval -ENOENT Invalid MD object.
  */
 int
-LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
-       lnet_process_id_t target, unsigned int portal,
+LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
+       struct lnet_process_id target, unsigned int portal,
        __u64 match_bits, unsigned int offset)
 {
        struct lnet_msg         *msg;
@@ -2822,7 +2940,7 @@ LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
 
        msg = lnet_msg_alloc();
        if (msg == NULL) {
-               CERROR("Dropping GET to %s: ENOMEM on lnet_msg_t\n",
+               CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
                       libcfs_id2str(target));
                return -ENOMEM;
        }
@@ -2870,7 +2988,7 @@ LNetGet(lnet_nid_t self, lnet_handle_md_t mdh,
        if (rc < 0) {
                CNETERR("Error sending GET to %s: %d\n",
                        libcfs_id2str(target), rc);
-               lnet_finalize(NULL, msg, rc);
+               lnet_finalize(msg, rc);
        }
 
        /* completion will be signalled by an event */
@@ -2896,8 +3014,8 @@ int
 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
 {
        struct list_head        *e;
-       struct lnet_ni          *ni = NULL;
-       lnet_remotenet_t        *rnet;
+       struct lnet_ni *ni = NULL;
+       struct lnet_remotenet *rnet;
        __u32                   dstnet = LNET_NIDNET(dstnid);
        int                     hops;
        int                     cpt;
@@ -2949,11 +3067,11 @@ LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
 
        rn_list = lnet_net2rnethash(dstnet);
        list_for_each(e, rn_list) {
-               rnet = list_entry(e, lnet_remotenet_t, lrn_list);
+               rnet = list_entry(e, struct lnet_remotenet, lrn_list);
 
                if (rnet->lrn_net == dstnet) {
-                       lnet_route_t *route;
-                       lnet_route_t *shortest = NULL;
+                       struct lnet_route *route;
+                       struct lnet_route *shortest = NULL;
                        __u32 shortest_hops = LNET_UNDEFINED_HOPS;
                        __u32 route_hops;