Whamcloud - gitweb
LU-56 lnet: cleanup for rtrpool and LNet counter
[fs/lustre-release.git] / lnet / lnet / router.c
index ef1ff34..6faf95b 100644 (file)
 
 #if defined(__KERNEL__) && defined(LNET_ROUTER)
 
+#define LNET_NRB_TINY          1024
+#define LNET_NRB_SMALL         8192
+#define LNET_NRB_LARGE         512
+
 static char *forwarding = "";
 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
                 "Explicitly enable/disable forwarding between networks");
 
-static int tiny_router_buffers = 1024;
+static int tiny_router_buffers;
 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
-                "# of 0 payload messages to buffer in the router");
-static int small_router_buffers = 8192;
+               "# of 0 payload messages to buffer in the router");
+static int small_router_buffers;
 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
-                "# of small (1 page) messages to buffer in the router");
-static int large_router_buffers = 512;
+               "# of small (1 page) messages to buffer in the router");
+static int large_router_buffers;
 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
-                "# of large messages to buffer in the router");
+               "# of large messages to buffer in the router");
 static int peer_buffer_credits = 0;
 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
                 "# router buffer credits per peer");
@@ -122,6 +126,8 @@ lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
         lp->lp_alive = !(!alive);               /* 1 bit! */
         lp->lp_notify = 1;
         lp->lp_notifylnd |= notifylnd;
+       if (lp->lp_alive)
+               lp->lp_ping_version = LNET_PROTO_PING_UNKNOWN; /* reset */
 
         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
 }
@@ -198,6 +204,8 @@ lnet_rtr_decref_locked(lnet_peer_t *lp)
 
         lp->lp_rtr_refcount--;
         if (lp->lp_rtr_refcount == 0) {
+               LASSERT(cfs_list_empty(&lp->lp_routes));
+
                 if (lp->lp_rcd != NULL) {
                         cfs_list_add(&lp->lp_rcd->rcd_list,
                                     &the_lnet.ln_rcd_deathrow);
@@ -279,6 +287,7 @@ lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route)
                 offset--;
         }
         cfs_list_add(&route->lr_list, e);
+       cfs_list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 
         the_lnet.ln_remote_nets_version++;
         lnet_rtr_addref_locked(route->lr_gateway);
@@ -325,6 +334,7 @@ lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway)
         CFS_INIT_LIST_HEAD(&rnet->lrn_routes);
         rnet->lrn_net = net;
         route->lr_hops = hops;
+       route->lr_net = net;
 
         LNET_LOCK();
 
@@ -335,11 +345,13 @@ lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway)
                 LIBCFS_FREE(route, sizeof(*route));
                 LIBCFS_FREE(rnet, sizeof(*rnet));
 
-                if (rc == -EHOSTUNREACH)        /* gateway is not on a local net */
+               if (rc == -EHOSTUNREACH) { /* gateway is not on a local net */
                         return 0;               /* ignore the route entry */
-
-                CERROR("Error %d creating route %s %d %s\n", rc,
-                       libcfs_net2str(net), hops, libcfs_nid2str(gateway));
+               } else {
+                       CERROR("Error %d creating route %s %d %s\n", rc,
+                              libcfs_net2str(net), hops,
+                              libcfs_nid2str(gateway));
+               }
                 return rc;
         }
 
@@ -367,22 +379,25 @@ lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway)
         }
 
         if (add_route) {
-                ni = route->lr_gateway->lp_ni;
-                lnet_ni_addref_locked(ni);
+               lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
+               lnet_add_route_to_rnet(rnet2, route);
 
-                lnet_add_route_to_rnet(rnet2, route);
-                LNET_UNLOCK();
+               ni = route->lr_gateway->lp_ni;
+               LNET_UNLOCK();
 
-                /* XXX Assume alive */
-                if (ni->ni_lnd->lnd_notify != NULL)
-                        (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
+               /* XXX Assume alive */
+               if (ni->ni_lnd->lnd_notify != NULL)
+                       (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
 
-                lnet_ni_decref(ni);
-        } else {
-                lnet_peer_decref_locked(route->lr_gateway);
-                LNET_UNLOCK();
-                LIBCFS_FREE(route, sizeof(*route));
-        }
+               LNET_LOCK();
+       }
+
+       /* -1 for notify or !add_route */
+       lnet_peer_decref_locked(route->lr_gateway);
+       LNET_UNLOCK();
+
+       if (!add_route)
+               LIBCFS_FREE(route, sizeof(*route));
 
         if (rnet != rnet2)
                 LIBCFS_FREE(rnet, sizeof(*rnet));
@@ -406,20 +421,31 @@ lnet_check_routes (void)
 
                 route2 = NULL;
                 cfs_list_for_each (e2, &rnet->lrn_routes) {
+                       lnet_nid_t      nid1;
+                       lnet_nid_t      nid2;
+                       int             net;
+
                         route = cfs_list_entry(e2, lnet_route_t, lr_list);
 
-                        if (route2 == NULL)
-                                route2 = route;
-                        else if (route->lr_gateway->lp_ni !=
-                                 route2->lr_gateway->lp_ni) {
-                                LNET_UNLOCK();
+                       if (route2 == NULL) {
+                               route2 = route;
+                               continue;
+                       }
 
-                                CERROR("Routes to %s via %s and %s not supported\n",
-                                       libcfs_net2str(rnet->lrn_net),
-                                       libcfs_nid2str(route->lr_gateway->lp_nid),
-                                       libcfs_nid2str(route2->lr_gateway->lp_nid));
-                                return -EINVAL;
-                        }
+                       if (route->lr_gateway->lp_ni ==
+                           route2->lr_gateway->lp_ni)
+                               continue;
+
+                       nid1 = route->lr_gateway->lp_nid;
+                       nid2 = route2->lr_gateway->lp_nid;
+                       net = rnet->lrn_net;
+
+                       LNET_UNLOCK();
+
+                       CERROR("Routes to %s via %s and %s not supported\n",
+                              libcfs_net2str(net), libcfs_nid2str(nid1),
+                              libcfs_nid2str(nid2));
+                       return -EINVAL;
                 }
         }
 
@@ -430,6 +456,7 @@ lnet_check_routes (void)
 int
 lnet_del_route (__u32 net, lnet_nid_t gw_nid)
 {
+       struct lnet_peer        *gateway;
         lnet_remotenet_t    *rnet;
         lnet_route_t        *route;
         cfs_list_t          *e1;
@@ -455,11 +482,13 @@ lnet_del_route (__u32 net, lnet_nid_t gw_nid)
                 cfs_list_for_each (e2, &rnet->lrn_routes) {
                         route = cfs_list_entry(e2, lnet_route_t, lr_list);
 
-                        if (!(gw_nid == LNET_NID_ANY ||
-                              gw_nid == route->lr_gateway->lp_nid))
-                                continue;
+                       gateway = route->lr_gateway;
+                       if (!(gw_nid == LNET_NID_ANY ||
+                             gw_nid == gateway->lp_nid))
+                               continue;
 
-                        cfs_list_del(&route->lr_list);
+                       cfs_list_del(&route->lr_list);
+                       cfs_list_del(&route->lr_gwlist);
                         the_lnet.ln_remote_nets_version++;
 
                         if (cfs_list_empty(&rnet->lrn_routes))
@@ -467,8 +496,8 @@ lnet_del_route (__u32 net, lnet_nid_t gw_nid)
                         else
                                 rnet = NULL;
 
-                        lnet_rtr_decref_locked(route->lr_gateway);
-                        lnet_peer_decref_locked(route->lr_gateway);
+                       lnet_rtr_decref_locked(gateway);
+                       lnet_peer_decref_locked(gateway);
                         LNET_UNLOCK();
 
                         LIBCFS_FREE(route, sizeof (*route));
@@ -529,6 +558,7 @@ lnet_swap_pinginfo(lnet_ping_info_t *info)
         int               i;
         lnet_ni_status_t *stat;
 
+       __swab32s(&info->pi_magic);
         __swab32s(&info->pi_version);
         __swab32s(&info->pi_pid);
         __swab32s(&info->pi_nnis);
@@ -540,84 +570,144 @@ lnet_swap_pinginfo(lnet_ping_info_t *info)
         return;
 }
 
-/* Returns # of down NIs, or negative error codes; ignore downed NIs
- * if a NI in 'net' is up */
-int
-lnet_router_down_ni(lnet_peer_t *rtr, __u32 net)
+/**
+ * parse router-checker pinginfo, record number of down NIs for remote
+ * networks on that router.
+ */
+static void
+lnet_parse_rc_info(lnet_rc_data_t *rcd)
 {
-        int               i;
-        int               down = 0;
-        int               ptl_up = 0;
-        int               ptl_down = 0;
-        lnet_ping_info_t *info;
+       lnet_ping_info_t        *info = rcd->rcd_pinginfo;
+       struct lnet_peer        *gw   = rcd->rcd_gateway;
+       lnet_route_t            *rtr;
 
-        if (!avoid_asym_router_failure)
-                return -ENOENT;
+       if (!gw->lp_alive)
+               return;
 
-        if (rtr->lp_rcd == NULL)
-                return -EINVAL;
+       if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
+               lnet_swap_pinginfo(info);
 
-        if (!rtr->lp_alive)
-                return -EINVAL;  /* stale lp_rcd */
+       /* NB always racing with network! */
+       if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
+               CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
+                      libcfs_nid2str(gw->lp_nid), info->pi_magic);
+               gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN;
+               return;
+       }
 
-        info = rtr->lp_rcd->rcd_pinginfo;
-        LASSERT (info != NULL);
+       gw->lp_ping_version = info->pi_version;
+       if (gw->lp_ping_version == LNET_PROTO_PING_VERSION_1)
+               return; /* v1 doesn't carry NI status info */
 
-        /* NB always racing with network! */
-        if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
-                lnet_swap_pinginfo(info);
-        } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
-                CNETERR("%s: Unexpected magic %08x\n",
-                        libcfs_nid2str(rtr->lp_nid), info->pi_magic);
-                return -EPROTO;
-        }
+       if (gw->lp_ping_version != LNET_PROTO_PING_VERSION) {
+               CDEBUG(D_NET, "%s: Unexpected version 0x%x\n",
+                      libcfs_nid2str(gw->lp_nid), gw->lp_ping_version);
+               gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN;
+               return;
+       }
 
-        if (info->pi_version == LNET_PROTO_PING_VERSION1)
-                return -ENOENT;  /* v1 doesn't carry NI status info */
+       cfs_list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) {
+               int     ptl_status = LNET_NI_STATUS_INVALID;
+               int     down = 0;
+               int     up = 0;
+               int     i;
 
-        if (info->pi_version != LNET_PROTO_PING_VERSION) {
-                CNETERR("%s: Unexpected version 0x%x\n",
-                        libcfs_nid2str(rtr->lp_nid), info->pi_version);
-                return -EPROTO;
-        }
+               for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
+                       lnet_ni_status_t *stat = &info->pi_ni[i];
+                       lnet_nid_t       nid = stat->ns_nid;
 
-        for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
-                lnet_ni_status_t *stat = &info->pi_ni[i];
-                lnet_nid_t        nid = stat->ns_nid;
+                       if (nid == LNET_NID_ANY) {
+                               CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
+                                      libcfs_nid2str(gw->lp_nid));
+                               gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN;
+                               return;
+                       }
 
-                if (nid == LNET_NID_ANY) {
-                        CNETERR("%s: unexpected LNET_NID_ANY\n",
-                                libcfs_nid2str(rtr->lp_nid));
-                        return -EPROTO;
-                }
+                       if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
+                               continue;
 
-                if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
-                        continue;
+                       if (stat->ns_status == LNET_NI_STATUS_DOWN) {
+                               if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND)
+                                       down++;
+                               else if (ptl_status != LNET_NI_STATUS_UP)
+                                       ptl_status = LNET_NI_STATUS_DOWN;
+                               continue;
+                       }
+
+                       if (stat->ns_status == LNET_NI_STATUS_UP) {
+                               if (LNET_NIDNET(nid) == rtr->lr_net) {
+                                       up = 1;
+                                       break;
+                               }
+                               /* ptl NIs are considered down only when
+                                * they're all down */
+                               if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
+                                       ptl_status = LNET_NI_STATUS_UP;
+                               continue;
+                       }
 
-                if (stat->ns_status == LNET_NI_STATUS_DOWN) {
-                        if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
-                                ptl_down = 1;
-                        else
-                                down++;
-                        continue;
-                }
+                       CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
+                              libcfs_nid2str(gw->lp_nid), stat->ns_status);
+                       gw->lp_ping_version = LNET_PROTO_PING_UNKNOWN;
+                       return;
+               }
 
-                if (stat->ns_status != LNET_NI_STATUS_UP) {
-                        CNETERR("%s: Unexpected status 0x%x\n",
-                                libcfs_nid2str(rtr->lp_nid), stat->ns_status);
-                        return -EPROTO;
-                }
+               if (up) { /* ignore downed NIs if NI for dest network is up */
+                       rtr->lr_downis = 0;
+                       continue;
+               }
+               rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN);
+       }
+}
 
-                /* ignore downed NIs if there's a NI up for dest network */
-                if (LNET_NIDNET(nid) == net)
-                        return 0;
+static void
+lnet_router_checker_event(lnet_event_t *event)
+{
+       lnet_rc_data_t          *rcd = event->md.user_ptr;
+       struct lnet_peer        *lp;
 
-                if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
-                        ptl_up = 1;
+       LASSERT(rcd != NULL);
+
+       if (event->unlinked) {
+               LNetInvalidateHandle(&rcd->rcd_mdh);
+               return;
         }
 
-        /* ptl NIs are considered down only when they're all down */
-        return down + (ptl_up ? 0 : ptl_down);
+       LASSERT(event->type == LNET_EVENT_SEND ||
+               event->type == LNET_EVENT_REPLY);
+
+       lp = rcd->rcd_gateway;
+       LASSERT(lp != NULL);
+
+       LNET_LOCK();
+       if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
+               /* ignore if no longer a router or rcd is replaced */
+               goto out;
+       }
+
+       if (event->type == LNET_EVENT_SEND) {
+               lp->lp_ping_notsent = 0;
+               if (event->status == 0)
+                       goto out;
+       }
+
+       /* LNET_EVENT_REPLY */
+       /* A successful REPLY means the router is up.  If _any_ comms
+        * to the router fail I assume it's down (this will happen if
+        * we ping alive routers to try to detect router death before
+        * apps get burned). */
+
+       lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
+       /* The router checker will wake up very shortly and do the
+        * actual notification.
+        * XXX If 'lp' stops being a router before then, it will still
+        * have the notification pending!!! */
+
+       if (avoid_asym_router_failure && event->status == 0)
+               lnet_parse_rc_info(rcd);
+
+ out:
+       LNET_UNLOCK();
 }
 
 void
@@ -654,100 +744,36 @@ lnet_wait_known_routerstate(void)
         }
 }
 
-static void
-lnet_router_checker_event (lnet_event_t *event)
-{
-        /* CAVEAT EMPTOR: I'm called with LNET_LOCKed and I'm not allowed to
-         * drop it (that's how come I see _every_ event, even ones that would
-         * overflow my EQ) */
-        lnet_rc_data_t *rcd = event->md.user_ptr;
-        lnet_peer_t    *lp;
-        lnet_nid_t      nid;
-
-       LASSERT(rcd != NULL);
-
-       if (event->unlinked) {
-               LNetInvalidateHandle(&rcd->rcd_mdh);
-               return;
-       }
-
-        LASSERT (event->type == LNET_EVENT_SEND ||
-                 event->type == LNET_EVENT_REPLY);
-
-        nid = (event->type == LNET_EVENT_SEND) ?
-              event->target.nid : event->initiator.nid;
-
-        lp = lnet_find_peer_locked(nid);
-        if (lp == NULL) {
-                /* router may have been removed */
-                CDEBUG(D_NET, "Router %s not found\n", libcfs_nid2str(nid));
-                return;
-        }
-
-        if (event->type == LNET_EVENT_SEND)     /* re-enable another ping */
-                lp->lp_ping_notsent = 0;
-
-        if (lnet_isrouter(lp) &&                /* ignore if no longer a router */
-            (event->status != 0 ||
-             event->type == LNET_EVENT_REPLY)) {
-
-                /* A successful REPLY means the router is up.  If _any_ comms
-                 * to the router fail I assume it's down (this will happen if
-                 * we ping alive routers to try to detect router death before
-                 * apps get burned). */
-
-                lnet_notify_locked(lp, 1, (event->status == 0),
-                                   cfs_time_current());
-
-                /* The router checker will wake up very shortly and do the
-                 * actual notification.  
-                 * XXX If 'lp' stops being a router before then, it will still
-                 * have the notification pending!!! */
-        }
-
-        /* This decref will NOT drop LNET_LOCK (it had to have 1 ref when it
-         * was in the peer table and I've not dropped the lock, so no-one else
-         * can have reduced the refcount) */
-        LASSERT(lp->lp_refcount > 1);
-
-        lnet_peer_decref_locked(lp);
-}
-
 void
-lnet_update_ni_status(void)
+lnet_update_ni_status_locked(void)
 {
-        cfs_time_t now = cfs_time_current();
-        lnet_ni_t *ni;
-        int        status;
-        int        timeout;
-
-        LASSERT (the_lnet.ln_routing);
-
-        timeout = router_ping_timeout +
-                  MAX(live_router_check_interval, dead_router_check_interval);
+       lnet_ni_t       *ni;
+       long            now;
+       int             timeout;
 
-        LNET_LOCK();
+       LASSERT(the_lnet.ln_routing);
 
-        cfs_list_for_each_entry (ni, &the_lnet.ln_nis, ni_list) {
-                lnet_ni_status_t *ns = ni->ni_status;
+       timeout = router_ping_timeout +
+                 MAX(live_router_check_interval, dead_router_check_interval);
 
-                LASSERT (ns != NULL);
+       now = cfs_time_current_sec();
+       cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
+               if (ni->ni_lnd->lnd_type == LOLND)
+                       continue;
 
-                status = LNET_NI_STATUS_UP;
-                if (ni->ni_lnd->lnd_type != LOLND &&  /* @lo forever alive */
-                    cfs_time_after(now, cfs_time_add(ni->ni_last_alive,
-                                                     cfs_time_seconds(timeout))))
-                        status = LNET_NI_STATUS_DOWN;
+               if (now < ni->ni_last_alive + timeout)
+                       continue;
 
-                if (ns->ns_status != status) {
-                        ns->ns_status = status;
-                        CDEBUG(D_NET, "NI(%s:%d) status changed to %s\n",
-                               libcfs_nid2str(ni->ni_nid), timeout,
-                               status == LNET_NI_STATUS_UP ? "up" : "down");
-                }
-        }
+               LASSERT(ni->ni_status != NULL);
 
-        LNET_UNLOCK();
+               if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
+                       CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
+                              libcfs_nid2str(ni->ni_nid), timeout);
+                       /* NB: so far, this is the only place to set
+                        * NI status to "down" */
+                       ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
+               }
+       }
 }
 
 void
@@ -757,6 +783,12 @@ lnet_destroy_rc_data (lnet_rc_data_t *rcd)
        /* detached from network */
        LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 
+       if (rcd->rcd_gateway != NULL) {
+               LNET_LOCK();
+               lnet_peer_decref_locked(rcd->rcd_gateway);
+               LNET_UNLOCK();
+       }
+
        if (rcd->rcd_pinginfo != NULL)
                LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 
@@ -813,7 +845,11 @@ lnet_create_rc_data_locked(lnet_peer_t *gateway)
                goto out;
        }
 
+       lnet_peer_addref_locked(gateway);
+       rcd->rcd_gateway = gateway;
        gateway->lp_rcd = rcd;
+       gateway->lp_ping_notsent = 0;
+
        return rcd;
 
  out:
@@ -1166,10 +1202,10 @@ rescan:
                         }
                 }
 
-                LNET_UNLOCK();
+               if (the_lnet.ln_routing)
+                       lnet_update_ni_status_locked();
 
-                if (the_lnet.ln_routing)
-                        lnet_update_ni_status();
+               LNET_UNLOCK();
 
                lnet_prune_rc_data(0); /* don't wait for UNLINK */
 
@@ -1237,9 +1273,12 @@ lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp)
 void
 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
 {
-        int            npages = rbp->rbp_npages;
-        int            nbuffers = 0;
-        lnet_rtrbuf_t *rb;
+       int             npages = rbp->rbp_npages;
+       int             nbuffers = 0;
+       lnet_rtrbuf_t   *rb;
+
+       if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
+               return;
 
         LASSERT (cfs_list_empty(&rbp->rbp_msgs));
         LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
@@ -1306,29 +1345,72 @@ lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
 }
 
 void
-lnet_free_rtrpools(void)
+lnet_rtrpools_free(void)
 {
-        lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[0]);
-        lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[1]);
-        lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[2]);
+       if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
+               return;
+
+       lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[0]);
+       lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[1]);
+       lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[2]);
+
+       LIBCFS_FREE(the_lnet.ln_rtrpools,
+                   sizeof(lnet_rtrbufpool_t) * LNET_NRBPOOLS);
+       the_lnet.ln_rtrpools = NULL;
 }
 
-void
-lnet_init_rtrpools(void)
+static int
+lnet_nrb_tiny_calculate(int npages)
 {
-        int small_pages = 1;
-        int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
+       if (tiny_router_buffers > 0)
+               return tiny_router_buffers;
 
-        lnet_rtrpool_init(&the_lnet.ln_rtrpools[0], 0);
-        lnet_rtrpool_init(&the_lnet.ln_rtrpools[1], small_pages);
-        lnet_rtrpool_init(&the_lnet.ln_rtrpools[2], large_pages);
+       if (tiny_router_buffers == 0)
+               return LNET_NRB_TINY;
+
+       LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when "
+                                 "routing enabled\n", tiny_router_buffers);
+       return -1;
 }
 
+static int
+lnet_nrb_small_calculate(int npages)
+{
+       if (small_router_buffers > 0)
+               return tiny_router_buffers;
+
+       if (small_router_buffers == 0)
+               return LNET_NRB_SMALL;
+
+       LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when "
+                                 "routing enabled\n", small_router_buffers);
+       return -1;
+}
+
+static int
+lnet_nrb_large_calculate(int npages)
+{
+       if (large_router_buffers > 0)
+               return large_router_buffers;
+
+       if (large_router_buffers == 0)
+               return LNET_NRB_LARGE;
+
+       LCONSOLE_ERROR_MSG(0x10e, "large_router_buffers=%d invalid when"
+                                 " routing enabled\n", large_router_buffers);
+       return -1;
+}
 
 int
-lnet_alloc_rtrpools(int im_a_router)
+lnet_rtrpools_alloc(int im_a_router)
 {
-        int       rc;
+       lnet_rtrbufpool_t *rtrp;
+       int     large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
+       int     small_pages = 1;
+       int     nrb_tiny;
+       int     nrb_small;
+       int     nrb_large;
+       int     rc;
 
         if (!strcmp(forwarding, "")) {
                 /* not set either way */
@@ -1345,51 +1427,54 @@ lnet_alloc_rtrpools(int im_a_router)
                 return -EINVAL;
         }
 
-        if (tiny_router_buffers <= 0) {
-                LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when "
-                                   "routing enabled\n", tiny_router_buffers);
-                rc = -EINVAL;
-                goto failed;
-        }
+       nrb_tiny = lnet_nrb_tiny_calculate(0);
+       if (nrb_tiny < 0)
+               return -EINVAL;
 
-        rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[0],
-                                     tiny_router_buffers);
-        if (rc != 0)
-                goto failed;
+       nrb_small = lnet_nrb_small_calculate(small_pages);
+       if (nrb_small < 0)
+               return -EINVAL;
 
-        if (small_router_buffers <= 0) {
-                LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when"
-                                   " routing enabled\n", small_router_buffers);
-                rc = -EINVAL;
-                goto failed;
-        }
+       nrb_large = lnet_nrb_large_calculate(large_pages);
+       if (nrb_large < 0)
+               return -EINVAL;
 
-        rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[1],
-                                     small_router_buffers);
-        if (rc != 0)
-                goto failed;
+       LIBCFS_ALLOC(the_lnet.ln_rtrpools,
+                    sizeof(lnet_rtrbufpool_t) * LNET_NRBPOOLS);
+       if (the_lnet.ln_rtrpools == NULL) {
+               LCONSOLE_ERROR_MSG(0x10c,
+                                  "Failed to initialize router buffe pool\n");
+               return -ENOMEM;
+       }
 
-        if (large_router_buffers <= 0) {
-                LCONSOLE_ERROR_MSG(0x10e, "large_router_buffers=%d invalid when"
-                                   " routing enabled\n", large_router_buffers);
-                rc = -EINVAL;
-                goto failed;
-        }
+       do {    /* iterate over rtrpools on all CPTs in upcoming patches */
+               rtrp = the_lnet.ln_rtrpools;
 
-        rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[2],
-                                     large_router_buffers);
-        if (rc != 0)
-                goto failed;
+               lnet_rtrpool_init(&rtrp[0], 0);
+               rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny);
+               if (rc != 0)
+                       goto failed;
 
-        LNET_LOCK();
-        the_lnet.ln_routing = 1;
-        LNET_UNLOCK();
+               lnet_rtrpool_init(&rtrp[1], small_pages);
+               rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small);
+               if (rc != 0)
+                       goto failed;
 
-        return 0;
+               lnet_rtrpool_init(&rtrp[2], large_pages);
+               rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large);
+               if (rc != 0)
+                       goto failed;
+       } while (0);
+
+       LNET_LOCK();
+       the_lnet.ln_routing = 1;
+       LNET_UNLOCK();
+
+       return 0;
 
  failed:
-        lnet_free_rtrpools();
-        return rc;
+       lnet_rtrpools_free();
+       return rc;
 }
 
 int
@@ -1524,9 +1609,7 @@ lnet_router_checker (void)
 
                 LASSERT (rc == 1);
 
-                LNET_LOCK();
                 lnet_router_checker_event(&ev);
-                LNET_UNLOCK();
         }
 
        if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
@@ -1575,17 +1658,12 @@ lnet_get_tunables (void)
 }
 
 void
-lnet_free_rtrpools (void)
-{
-}
-
-void
-lnet_init_rtrpools (void)
+lnet_rtrpools_free(void)
 {
 }
 
 int
-lnet_alloc_rtrpools (int im_a_arouter)
+lnet_rtrpools_alloc(int im_a_arouter)
 {
         return 0;
 }