lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2017, Intel Corporation.
   5  *
   6  *   This file is part of Lustre, https://wiki.whamcloud.com/
   7  *
   8  *   Portals is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Portals is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Portals; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_LNET
  24
  25 #include <linux/random.h>
  26 #include <lnet/lib-lnet.h>
  27
  28 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
  29 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
  30 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
  31 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
  32 #define LNET_NRB_SMALL_PAGES    1
  33 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  34 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  35 #define LNET_NRB_LARGE_PAGES    ((LNET_MTU + PAGE_SIZE - 1) >> \
  36                                   PAGE_SHIFT)
  37
  38 extern unsigned int lnet_current_net_count;
  39
  40 static char *forwarding = "";
  41 module_param(forwarding, charp, 0444);
  42 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
  43
  44 static int tiny_router_buffers;
  45 module_param(tiny_router_buffers, int, 0444);
  46 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
  47 static int small_router_buffers;
  48 module_param(small_router_buffers, int, 0444);
  49 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
  50 static int large_router_buffers;
  51 module_param(large_router_buffers, int, 0444);
  52 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
  53 static int peer_buffer_credits;
  54 module_param(peer_buffer_credits, int, 0444);
  55 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
  56
  57 static int auto_down = 1;
  58 module_param(auto_down, int, 0444);
  59 MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
  60
  61 int
  62 lnet_peer_buffer_credits(struct lnet_net *net)
  63 {
  64         /* NI option overrides LNet default */
  65         if (net->net_tunables.lct_peer_rtr_credits > 0)
  66                 return net->net_tunables.lct_peer_rtr_credits;
  67         if (peer_buffer_credits > 0)
  68                 return peer_buffer_credits;
  69
  70         /* As an approximation, allow this peer the same number of router
  71          * buffers as it is allowed outstanding sends */
  72         return net->net_tunables.lct_peer_tx_credits;
  73 }
  74
  75 static int check_routers_before_use;
  76 module_param(check_routers_before_use, int, 0444);
  77 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
  78
  79 int avoid_asym_router_failure = 1;
  80 module_param(avoid_asym_router_failure, int, 0644);
  81 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
  82
  83 int dead_router_check_interval = INT_MIN;
  84 module_param(dead_router_check_interval, int, 0444);
  85 MODULE_PARM_DESC(dead_router_check_interval, "(DEPRECATED - Use alive_router_check_interval)");
  86
  87 int live_router_check_interval = INT_MIN;
  88 module_param(live_router_check_interval, int, 0444);
  89 MODULE_PARM_DESC(live_router_check_interval, "(DEPRECATED - Use alive_router_check_interval)");
  90
  91 int alive_router_check_interval = 60;
  92 module_param(alive_router_check_interval, int, 0644);
  93 MODULE_PARM_DESC(alive_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
  94
  95 static int router_ping_timeout = 50;
  96 module_param(router_ping_timeout, int, 0644);
  97 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
  98
  99 /*
 100  * A value between 0 and 100. 0 meaning that even if router's interfaces
 101  * have the worse health still consider the gateway usable.
 102  * 100 means that at least one interface on the route's remote net is 100%
 103  * healthy to consider the route alive.
 104  * The default is set to 100 to ensure we maintain the original behavior.
 105  */
 106 unsigned int router_sensitivity_percentage = 100;
 107 static int rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
 108 static struct kernel_param_ops param_ops_rtr_sensitivity = {
 109         .set = rtr_sensitivity_set,
 110         .get = param_get_int,
 111 };
 112 #define param_check_rtr_sensitivity(name, p) \
 113                 __param_check(name, p, int)
 114 #ifdef HAVE_KERNEL_PARAM_OPS
 115 module_param(router_sensitivity_percentage, rtr_sensitivity, S_IRUGO|S_IWUSR);
 116 #else
 117 module_param_call(router_sensitivity_percentage, rtr_sensitivity_set, param_get_int,
 118                   &router_sensitivity_percentage, S_IRUGO|S_IWUSR);
 119 #endif
 120 MODULE_PARM_DESC(router_sensitivity_percentage,
 121                 "How healthy a gateway should be to be used in percent");
 122
 123 static int
 124 rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 125 {
 126         int rc;
 127         unsigned *sen = (unsigned *)kp->arg;
 128         unsigned long value;
 129
 130         rc = kstrtoul(val, 0, &value);
 131         if (rc) {
 132                 CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
 133                 return rc;
 134         }
 135
 136         if (value < 0 || value > 100) {
 137                 CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
 138                 return -EINVAL;
 139         }
 140
 141         /*
 142          * The purpose of locking the api_mutex here is to ensure that
 143          * the correct value ends up stored properly.
 144          */
 145         mutex_lock(&the_lnet.ln_api_mutex);
 146
 147         *sen = value;
 148
 149         mutex_unlock(&the_lnet.ln_api_mutex);
 150
 151         return 0;
 152 }
 153
 154 void
 155 lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target)
 156 {
 157         struct lnet_route *route;
 158
 159         lnet_net_lock(LNET_LOCK_EX);
 160         target->lp_rtr_refcount += src->lp_rtr_refcount;
 161         /* move the list of queued messages to the new peer */
 162         list_splice_init(&src->lp_rtrq, &target->lp_rtrq);
 163         /* move all the routes that reference the peer */
 164         list_splice_init(&src->lp_routes, &target->lp_routes);
 165         /* update all the routes to point to the new peer */
 166         list_for_each_entry(route, &target->lp_routes, lr_gwlist)
 167                 route->lr_gateway = target;
 168         /* remove the old peer from the ln_routers list */
 169         list_del_init(&src->lp_rtr_list);
 170         /* add the new peer to the ln_routers list */
 171         if (list_empty(&target->lp_rtr_list)) {
 172                 lnet_peer_addref_locked(target);
 173                 list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers);
 174         }
 175         /* reset the ref count on the old peer and decrement its ref count */
 176         src->lp_rtr_refcount = 0;
 177         lnet_peer_decref_locked(src);
 178         /* update the router version */
 179         the_lnet.ln_routers_version++;
 180         lnet_net_unlock(LNET_LOCK_EX);
 181 }
 182
 183 int
 184 lnet_peers_start_down(void)
 185 {
 186         return check_routers_before_use;
 187 }
 188
 189 /*
 190  * A net is alive if at least one gateway NI on the network is alive.
 191  */
 192 static bool
 193 lnet_is_gateway_net_alive(struct lnet_peer_net *lpn)
 194 {
 195         struct lnet_peer_ni *lpni;
 196
 197         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 198                 if (lnet_is_peer_ni_alive(lpni))
 199                         return true;
 200         }
 201
 202         return false;
 203 }
 204
 205 /*
 206  * a gateway is alive only if all its nets are alive
 207  * called with cpt lock held
 208  */
 209 bool lnet_is_gateway_alive(struct lnet_peer *gw)
 210 {
 211         struct lnet_peer_net *lpn;
 212
 213         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 214                 if (!lnet_is_gateway_net_alive(lpn))
 215                         return false;
 216         }
 217
 218         return true;
 219 }
 220
 221 /*
 222  * lnet_is_route_alive() needs to be called with cpt lock held
 223  * A route is alive if the gateway can route between the local network and
 224  * the remote network of the route.
 225  * This means at least one NI is alive on each of the local and remote
 226  * networks of the gateway.
 227  */
 228 bool lnet_is_route_alive(struct lnet_route *route)
 229 {
 230         struct lnet_peer *gw = route->lr_gateway;
 231         struct lnet_peer_net *llpn;
 232         struct lnet_peer_net *rlpn;
 233         bool route_alive;
 234
 235         /*
 236          * if discovery is disabled then rely on the cached aliveness
 237          * information. This is handicapped information which we log when
 238          * we receive the discovery ping response. The most uptodate
 239          * aliveness information can only be obtained when discovery is
 240          * enabled.
 241          */
 242         if (lnet_is_discovery_disabled(gw))
 243                 return route->lr_alive;
 244
 245         /*
 246          * check the gateway's interfaces on the route rnet to make sure
 247          * that the gateway is viable.
 248          */
 249         llpn = lnet_peer_get_net_locked(gw, route->lr_lnet);
 250         if (!llpn)
 251                 return false;
 252
 253         route_alive = lnet_is_gateway_net_alive(llpn);
 254
 255         if (avoid_asym_router_failure) {
 256                 rlpn = lnet_peer_get_net_locked(gw, route->lr_net);
 257                 if (!rlpn)
 258                         return false;
 259                 route_alive = route_alive &&
 260                               lnet_is_gateway_net_alive(rlpn);
 261         }
 262
 263         if (!route_alive)
 264                 return route_alive;
 265
 266         spin_lock(&gw->lp_lock);
 267         if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) {
 268                 if (gw->lp_rtr_refcount > 0)
 269                         CERROR("peer %s is being used as a gateway but routing feature is not turned on\n",
 270                                libcfs_nid2str(gw->lp_primary_nid));
 271                 route_alive = false;
 272         }
 273         spin_unlock(&gw->lp_lock);
 274
 275         return route_alive;
 276 }
 277
 278 void
 279 lnet_consolidate_routes_locked(struct lnet_peer *orig_lp,
 280                                struct lnet_peer *new_lp)
 281 {
 282         struct lnet_peer_ni *lpni;
 283         struct lnet_route *route;
 284
 285         /*
 286          * Although a route is correlated with a peer, but when it's added
 287          * a specific NID is used. That NID refers to a peer_ni within
 288          * a peer. There could be other peer_nis on the same net, which
 289          * can be used to send to that gateway. However when we are
 290          * consolidating gateways because of discovery, the nid used to
 291          * add the route might've moved between gateway peers. In this
 292          * case we want to move the route to the new gateway as well. The
 293          * intent here is not to confuse the user who added the route.
 294          */
 295         list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) {
 296                 lpni = lnet_peer_get_ni_locked(orig_lp, route->lr_nid);
 297                 if (!lpni) {
 298                         lnet_net_lock(LNET_LOCK_EX);
 299                         list_move(&route->lr_gwlist, &new_lp->lp_routes);
 300                         lnet_net_unlock(LNET_LOCK_EX);
 301                 }
 302         }
 303
 304 }
 305
 306 static inline void
 307 lnet_set_route_aliveness(struct lnet_route *route, bool alive)
 308 {
 309         /* Log when there's a state change */
 310         if (route->lr_alive != alive) {
 311                 CERROR("route to %s through %s has gone from %s to %s\n",
 312                        libcfs_net2str(route->lr_net),
 313                        libcfs_nid2str(route->lr_gateway->lp_primary_nid),
 314                        (route->lr_alive) ? "up" : "down",
 315                        alive ? "up" : "down");
 316                 route->lr_alive = alive;
 317         }
 318 }
 319
 320 void
 321 lnet_router_discovery_ping_reply(struct lnet_peer *lp)
 322 {
 323         struct lnet_ping_buffer *pbuf = lp->lp_data;
 324         struct lnet_remotenet *rnet;
 325         struct lnet_peer_net *llpn;
 326         struct lnet_route *route;
 327         bool net_up = false;
 328         unsigned lp_state;
 329         __u32 net, net2;
 330         int i, j;
 331
 332
 333         spin_lock(&lp->lp_lock);
 334         lp_state = lp->lp_state;
 335
 336         /* only handle replies if discovery is disabled. */
 337         if (!lnet_is_discovery_disabled_locked(lp)) {
 338                 spin_unlock(&lp->lp_lock);
 339                 return;
 340         }
 341
 342         spin_unlock(&lp->lp_lock);
 343
 344         if (lp_state & LNET_PEER_PING_FAILED) {
 345                 CDEBUG(D_NET,
 346                        "Ping failed with %d. Set routes down for gw %s\n",
 347                        lp->lp_ping_error, libcfs_nid2str(lp->lp_primary_nid));
 348                 /* If the ping failed then mark the routes served by this
 349                  * peer down
 350                  */
 351                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 352                         lnet_set_route_aliveness(route, false);
 353                 return;
 354         }
 355
 356         CDEBUG(D_NET, "Discovery is disabled. Processing reply for gw: %s\n",
 357                libcfs_nid2str(lp->lp_primary_nid));
 358
 359         /*
 360          * examine the ping response:
 361          * For each NID in the ping response, extract the net
 362          * if the net exists on our remote net list then
 363          * iterate over the routes on the rnet and if:
 364          *      The route's local net is healthy and
 365          *      The remote net status is UP, then mark the route up
 366          * otherwise mark the route down
 367          */
 368         for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
 369                 net = LNET_NIDNET(pbuf->pb_info.pi_ni[i].ns_nid);
 370                 rnet = lnet_find_rnet_locked(net);
 371                 if (!rnet)
 372                         continue;
 373                 list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
 374                         /* check if this is the route's gateway */
 375                         if (lp->lp_primary_nid !=
 376                             route->lr_gateway->lp_primary_nid)
 377                                 continue;
 378
 379                         /* gateway has the routing feature disabled */
 380                         if (pbuf->pb_info.pi_features &
 381                               LNET_PING_FEAT_RTE_DISABLED) {
 382                                 lnet_set_route_aliveness(route, false);
 383                                 continue;
 384                         }
 385
 386                         llpn = lnet_peer_get_net_locked(lp, route->lr_lnet);
 387                         if (!llpn) {
 388                                 lnet_set_route_aliveness(route, false);
 389                                 continue;
 390                         }
 391
 392                         if (!lnet_is_gateway_net_alive(llpn)) {
 393                                 lnet_set_route_aliveness(route, false);
 394                                 continue;
 395                         }
 396
 397                         if (avoid_asym_router_failure &&
 398                             pbuf->pb_info.pi_ni[i].ns_status !=
 399                                 LNET_NI_STATUS_UP) {
 400                                 net_up = false;
 401
 402                                 /*
 403                                  * revisit all previous NIDs and check if
 404                                  * any on the network we're examining is
 405                                  * up. If at least one is up then we consider
 406                                  * the route to be alive.
 407                                  */
 408                                 for (j = 1; j < i; j++) {
 409                                         net2 = LNET_NIDNET(pbuf->pb_info.
 410                                                            pi_ni[j].ns_nid);
 411                                         if (net2 == net &&
 412                                             pbuf->pb_info.pi_ni[j].ns_status ==
 413                                                 LNET_NI_STATUS_UP)
 414                                                 net_up = true;
 415                                 }
 416                                 if (!net_up) {
 417                                         lnet_set_route_aliveness(route, false);
 418                                         continue;
 419                                 }
 420                         }
 421
 422                         lnet_set_route_aliveness(route, true);
 423                 }
 424         }
 425 }
 426
 427 void
 428 lnet_router_discovery_complete(struct lnet_peer *lp)
 429 {
 430         struct lnet_peer_ni *lpni = NULL;
 431         struct lnet_route *route;
 432
 433         spin_lock(&lp->lp_lock);
 434         lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
 435         spin_unlock(&lp->lp_lock);
 436
 437         /*
 438          * Router discovery successful? All peer information would've been
 439          * updated already. No need to do any more processing
 440          */
 441         if (!lp->lp_dc_error)
 442                 return;
 443         /*
 444          * discovery failed? then we need to set the status of each lpni
 445          * to DOWN. It will be updated the next time we discover the
 446          * router. For router peer NIs not on local networks, we never send
 447          * messages directly to them, so their health will always remain
 448          * at maximum. We can only tell if they are up or down from the
 449          * status returned in the PING response. If we fail to get that
 450          * status in our scheduled router discovery, then we'll assume
 451          * it's down until we're told otherwise.
 452          */
 453         CDEBUG(D_NET, "%s: Router discovery failed %d\n",
 454                libcfs_nid2str(lp->lp_primary_nid), lp->lp_dc_error);
 455         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 456                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 457
 458         list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 459                 lnet_set_route_aliveness(route, false);
 460 }
 461
 462 static void
 463 lnet_rtr_addref_locked(struct lnet_peer *lp)
 464 {
 465         LASSERT(lp->lp_rtr_refcount >= 0);
 466
 467         /* lnet_net_lock must be exclusively locked */
 468         lp->lp_rtr_refcount++;
 469         if (lp->lp_rtr_refcount == 1) {
 470                 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
 471                 /* addref for the_lnet.ln_routers */
 472                 lnet_peer_addref_locked(lp);
 473                 the_lnet.ln_routers_version++;
 474         }
 475 }
 476
 477 static void
 478 lnet_rtr_decref_locked(struct lnet_peer *lp)
 479 {
 480         LASSERT(atomic_read(&lp->lp_refcount) > 0);
 481         LASSERT(lp->lp_rtr_refcount > 0);
 482
 483         /* lnet_net_lock must be exclusively locked */
 484         lp->lp_rtr_refcount--;
 485         if (lp->lp_rtr_refcount == 0) {
 486                 LASSERT(list_empty(&lp->lp_routes));
 487
 488                 list_del(&lp->lp_rtr_list);
 489                 /* decref for the_lnet.ln_routers */
 490                 lnet_peer_decref_locked(lp);
 491                 the_lnet.ln_routers_version++;
 492         }
 493 }
 494
 495 struct lnet_remotenet *
 496 lnet_find_rnet_locked(__u32 net)
 497 {
 498         struct lnet_remotenet *rnet;
 499         struct list_head *tmp;
 500         struct list_head *rn_list;
 501
 502         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 503
 504         rn_list = lnet_net2rnethash(net);
 505         list_for_each(tmp, rn_list) {
 506                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 507
 508                 if (rnet->lrn_net == net)
 509                         return rnet;
 510         }
 511         return NULL;
 512 }
 513
 514 static void lnet_shuffle_seed(void)
 515 {
 516         static int seeded;
 517         struct lnet_ni *ni = NULL;
 518
 519         if (seeded)
 520                 return;
 521
 522         /* Nodes with small feet have little entropy
 523          * the NID for this node gives the most entropy in the low bits */
 524         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 525                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 526
 527         seeded = 1;
 528         return;
 529 }
 530
 531 /* NB expects LNET_LOCK held */
 532 static void
 533 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 534 {
 535         struct lnet_peer_net *lpn;
 536         unsigned int offset = 0;
 537         unsigned int len = 0;
 538         struct list_head *e;
 539
 540         lnet_shuffle_seed();
 541
 542         list_for_each(e, &rnet->lrn_routes)
 543                 len++;
 544
 545         /*
 546          * Randomly adding routes to the list is done to ensure that when
 547          * different nodes are using the same list of routers, they end up
 548          * preferring different routers.
 549          */
 550         offset = prandom_u32_max(len + 1);
 551         list_for_each(e, &rnet->lrn_routes) {
 552                 if (offset == 0)
 553                         break;
 554                 offset--;
 555         }
 556         list_add(&route->lr_list, e);
 557         /*
 558          * force a router check on the gateway to make sure the route is
 559          * alive
 560          */
 561         list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets,
 562                             lpn_peer_nets) {
 563                 lpn->lpn_rtrcheck_timestamp = 0;
 564         }
 565
 566         the_lnet.ln_remote_nets_version++;
 567
 568         /* add the route on the gateway list */
 569         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 570
 571         /* take a router reference count on the gateway */
 572         lnet_rtr_addref_locked(route->lr_gateway);
 573 }
 574
 575 int
 576 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 577                __u32 priority, __u32 sensitivity)
 578 {
 579         struct list_head *route_entry;
 580         struct lnet_remotenet *rnet;
 581         struct lnet_remotenet *rnet2;
 582         struct lnet_route *route;
 583         struct lnet_peer_ni *lpni;
 584         struct lnet_peer *gw;
 585         int add_route;
 586         int rc;
 587
 588         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 589                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 590
 591         if (gateway == LNET_NID_ANY ||
 592             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 593             net == LNET_NIDNET(LNET_NID_ANY) ||
 594             LNET_NETTYP(net) == LOLND ||
 595             LNET_NIDNET(gateway) == net ||
 596             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 597                 return -EINVAL;
 598
 599         /* it's a local network */
 600         if (lnet_islocalnet(net))
 601                 return -EEXIST;
 602
 603         if (!lnet_islocalnet(LNET_NIDNET(gateway))) {
 604                 CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n",
 605                        libcfs_nid2str(gateway),
 606                        libcfs_net2str(LNET_NIDNET(gateway)));
 607                 return -EHOSTUNREACH;
 608         }
 609
 610         /* Assume net, route, all new */
 611         LIBCFS_ALLOC(route, sizeof(*route));
 612         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 613         if (route == NULL || rnet == NULL) {
 614                 CERROR("Out of memory creating route %s %d %s\n",
 615                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 616                 if (route != NULL)
 617                         LIBCFS_FREE(route, sizeof(*route));
 618                 if (rnet != NULL)
 619                         LIBCFS_FREE(rnet, sizeof(*rnet));
 620                 return -ENOMEM;
 621         }
 622
 623         INIT_LIST_HEAD(&rnet->lrn_routes);
 624         rnet->lrn_net = net;
 625         /* store the local and remote net that the route represents */
 626         route->lr_lnet = LNET_NIDNET(gateway);
 627         route->lr_net = net;
 628         route->lr_nid = gateway;
 629         route->lr_priority = priority;
 630         route->lr_hops = hops;
 631
 632         lnet_net_lock(LNET_LOCK_EX);
 633
 634         /*
 635          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 636          * lose that once we're done
 637          */
 638         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 639         if (IS_ERR(lpni)) {
 640                 lnet_net_unlock(LNET_LOCK_EX);
 641
 642                 LIBCFS_FREE(route, sizeof(*route));
 643                 LIBCFS_FREE(rnet, sizeof(*rnet));
 644
 645                 rc = PTR_ERR(lpni);
 646                 CERROR("Error %d creating route %s %d %s\n", rc,
 647                         libcfs_net2str(net), hops,
 648                         libcfs_nid2str(gateway));
 649                 return rc;
 650         }
 651
 652         LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
 653         gw = lpni->lpni_peer_net->lpn_peer;
 654
 655         route->lr_gateway = gw;
 656
 657         rnet2 = lnet_find_rnet_locked(net);
 658         if (rnet2 == NULL) {
 659                 /* new network */
 660                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 661                 rnet2 = rnet;
 662         }
 663
 664         /* Search for a duplicate route (it's a NOOP if it is) */
 665         add_route = 1;
 666         list_for_each(route_entry, &rnet2->lrn_routes) {
 667                 struct lnet_route *route2;
 668
 669                 route2 = list_entry(route_entry, struct lnet_route, lr_list);
 670                 if (route2->lr_gateway == route->lr_gateway) {
 671                         add_route = 0;
 672                         break;
 673                 }
 674
 675                 /* our lookups must be true */
 676                 LASSERT(route2->lr_gateway->lp_primary_nid != gateway);
 677         }
 678
 679         /*
 680          * It is possible to add multiple routes through the same peer,
 681          * but it'll be using a different NID of that peer. When the
 682          * gateway is discovered, discovery will consolidate the different
 683          * peers into one peer. In this case the discovery code will have
 684          * to move the routes from the peer that's being deleted to the
 685          * consolidated peer lp_routes list
 686          */
 687         if (add_route) {
 688                 gw->lp_health_sensitivity = sensitivity;
 689                 lnet_add_route_to_rnet(rnet2, route);
 690                 if (lnet_peer_discovery_disabled)
 691                         CWARN("Consider turning discovery on to enable full "
 692                               "Multi-Rail routing functionality\n");
 693         }
 694
 695         /*
 696          * get rid of the reference on the lpni.
 697          */
 698         lnet_peer_ni_decref_locked(lpni);
 699         lnet_net_unlock(LNET_LOCK_EX);
 700
 701         rc = 0;
 702
 703         if (!add_route) {
 704                 rc = -EEXIST;
 705                 LIBCFS_FREE(route, sizeof(*route));
 706         }
 707
 708         if (rnet != rnet2)
 709                 LIBCFS_FREE(rnet, sizeof(*rnet));
 710
 711         /* kick start the monitor thread to handle the added route */
 712         complete(&the_lnet.ln_mt_wait_complete);
 713
 714         return rc;
 715 }
 716
 717 static void
 718 lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list,
 719                          struct list_head *zombies)
 720 {
 721         struct lnet_peer *gateway;
 722         struct lnet_route *route;
 723         struct lnet_route *tmp;
 724
 725         list_for_each_entry_safe(route, tmp, route_list, lr_list) {
 726                 gateway = route->lr_gateway;
 727                 if (gw_nid != LNET_NID_ANY &&
 728                     gw_nid != gateway->lp_primary_nid)
 729                         continue;
 730
 731                 /*
 732                  * move to zombie to delete outside the lock
 733                  * Note that this function is called with the
 734                  * ln_api_mutex held as well as the exclusive net
 735                  * lock. Adding to the remote net list happens
 736                  * under the same conditions. Same goes for the
 737                  * gateway router list
 738                  */
 739                 list_move(&route->lr_list, zombies);
 740                 the_lnet.ln_remote_nets_version++;
 741
 742                 list_del(&route->lr_gwlist);
 743                 lnet_rtr_decref_locked(gateway);
 744         }
 745 }
 746
 747 int
 748 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 749 {
 750         struct list_head rnet_zombies;
 751         struct lnet_remotenet *rnet;
 752         struct lnet_remotenet *tmp;
 753         struct list_head *rn_list;
 754         struct lnet_peer_ni *lpni;
 755         struct lnet_route *route;
 756         struct list_head zombies;
 757         struct lnet_peer *lp = NULL;
 758         int i = 0;
 759
 760         INIT_LIST_HEAD(&rnet_zombies);
 761         INIT_LIST_HEAD(&zombies);
 762
 763         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 764                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 765
 766         /* NB Caller may specify either all routes via the given gateway
 767          * or a specific route entry actual NIDs) */
 768
 769         lnet_net_lock(LNET_LOCK_EX);
 770
 771         lpni = lnet_find_peer_ni_locked(gw_nid);
 772         if (lpni) {
 773                 lp = lpni->lpni_peer_net->lpn_peer;
 774                 LASSERT(lp);
 775                 gw_nid = lp->lp_primary_nid;
 776                 lnet_peer_ni_decref_locked(lpni);
 777         }
 778
 779         if (net != LNET_NIDNET(LNET_NID_ANY)) {
 780                 rnet = lnet_find_rnet_locked(net);
 781                 if (!rnet) {
 782                         lnet_net_unlock(LNET_LOCK_EX);
 783                         return -ENOENT;
 784                 }
 785                 lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 786                                          &zombies);
 787                 if (list_empty(&rnet->lrn_routes))
 788                         list_move(&rnet->lrn_list, &rnet_zombies);
 789                 goto delete_zombies;
 790         }
 791
 792         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 793                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 794
 795                 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
 796                         lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 797                                                  &zombies);
 798                         if (list_empty(&rnet->lrn_routes))
 799                                 list_move(&rnet->lrn_list, &rnet_zombies);
 800                 }
 801         }
 802
 803 delete_zombies:
 804         /*
 805          * check if there are any routes remaining on the gateway
 806          * If there are no more routes make sure to set the peer's
 807          * lp_disc_net_id to 0 (invalid), in case we add more routes in
 808          * the future on that gateway, then we start our discovery process
 809          * from scratch
 810          */
 811         if (lpni) {
 812                 if (list_empty(&lp->lp_routes))
 813                         lp->lp_disc_net_id = 0;
 814         }
 815
 816         lnet_net_unlock(LNET_LOCK_EX);
 817
 818         while (!list_empty(&zombies)) {
 819                 route = list_first_entry(&zombies, struct lnet_route, lr_list);
 820                 list_del(&route->lr_list);
 821                 LIBCFS_FREE(route, sizeof(*route));
 822         }
 823
 824         while (!list_empty(&rnet_zombies)) {
 825                 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
 826                                         lrn_list);
 827                 list_del(&rnet->lrn_list);
 828                 LIBCFS_FREE(rnet, sizeof(*rnet));
 829         }
 830
 831         return 0;
 832 }
 833
 834 void
 835 lnet_destroy_routes (void)
 836 {
 837         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 838 }
 839
 840 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 841 {
 842         struct lnet_rtrbufpool *rbp;
 843         int i, rc = -ENOENT, j;
 844
 845         if (the_lnet.ln_rtrpools == NULL)
 846                 return rc;
 847
 848
 849         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 850                 if (i != cpt)
 851                         continue;
 852
 853                 lnet_net_lock(i);
 854                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 855                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 856                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 857                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 858                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 859                 }
 860                 lnet_net_unlock(i);
 861                 rc = 0;
 862                 break;
 863         }
 864
 865         lnet_net_lock(LNET_LOCK_EX);
 866         pool_cfg->pl_routing = the_lnet.ln_routing;
 867         lnet_net_unlock(LNET_LOCK_EX);
 868
 869         return rc;
 870 }
 871
 872 int
 873 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 874                lnet_nid_t *gateway, __u32 *alive, __u32 *priority, __u32 *sensitivity)
 875 {
 876         struct lnet_remotenet *rnet;
 877         struct list_head *rn_list;
 878         struct lnet_route *route;
 879         struct list_head *e1;
 880         struct list_head *e2;
 881         int cpt;
 882         int i;
 883
 884         cpt = lnet_net_lock_current();
 885
 886         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 887                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 888                 list_for_each(e1, rn_list) {
 889                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 890
 891                         list_for_each(e2, &rnet->lrn_routes) {
 892                                 route = list_entry(e2, struct lnet_route,
 893                                                    lr_list);
 894
 895                                 if (idx-- == 0) {
 896                                         *net      = rnet->lrn_net;
 897                                         *gateway  = route->lr_nid;
 898                                         *hops     = route->lr_hops;
 899                                         *priority = route->lr_priority;
 900                                         *sensitivity = route->lr_gateway->
 901                                                 lp_health_sensitivity;
 902                                         *alive    = lnet_is_route_alive(route);
 903                                         lnet_net_unlock(cpt);
 904                                         return 0;
 905                                 }
 906                         }
 907                 }
 908         }
 909
 910         lnet_net_unlock(cpt);
 911         return -ENOENT;
 912 }
 913
 914 static void
 915 lnet_wait_known_routerstate(void)
 916 {
 917         struct lnet_peer *rtr;
 918         struct list_head *entry;
 919         int all_known;
 920
 921         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
 922
 923         for (;;) {
 924                 int cpt = lnet_net_lock_current();
 925
 926                 all_known = 1;
 927                 list_for_each(entry, &the_lnet.ln_routers) {
 928                         rtr = list_entry(entry, struct lnet_peer,
 929                                          lp_rtr_list);
 930
 931                         spin_lock(&rtr->lp_lock);
 932
 933                         if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
 934                                 all_known = 0;
 935                                 spin_unlock(&rtr->lp_lock);
 936                                 break;
 937                         }
 938                         spin_unlock(&rtr->lp_lock);
 939                 }
 940
 941                 lnet_net_unlock(cpt);
 942
 943                 if (all_known)
 944                         return;
 945
 946                 set_current_state(TASK_UNINTERRUPTIBLE);
 947                 schedule_timeout(cfs_time_seconds(1));
 948         }
 949 }
 950
 951 static inline bool
 952 lnet_net_set_status_locked(struct lnet_net *net, __u32 status)
 953 {
 954         struct lnet_ni *ni;
 955         bool update = false;
 956
 957         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
 958                 lnet_ni_lock(ni);
 959                 if (ni->ni_status &&
 960                     ni->ni_status->ns_status != status) {
 961                     ni->ni_status->ns_status = status;
 962                     update = true;
 963                 }
 964                 lnet_ni_unlock(ni);
 965         }
 966
 967         return update;
 968 }
 969
 970 static bool
 971 lnet_update_ni_status_locked(void)
 972 {
 973         struct lnet_net *net;
 974         bool push = false;
 975         time64_t now;
 976         time64_t timeout;
 977
 978         LASSERT(the_lnet.ln_routing);
 979
 980         timeout = router_ping_timeout + alive_router_check_interval;
 981
 982         now = ktime_get_real_seconds();
 983         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 984                 if (net->net_lnd->lnd_type == LOLND)
 985                         continue;
 986
 987                 if (now < net->net_last_alive + timeout)
 988                         continue;
 989
 990                 spin_lock(&net->net_lock);
 991                 /* re-check with lock */
 992                 if (now < net->net_last_alive + timeout) {
 993                         spin_unlock(&net->net_lock);
 994                         continue;
 995                 }
 996                 spin_unlock(&net->net_lock);
 997
 998                 /*
 999                  * if the net didn't receive any traffic for past the
1000                  * timeout on any of its constituent NIs, then mark all
1001                  * the NIs down.
1002                  */
1003                 push = lnet_net_set_status_locked(net, LNET_NI_STATUS_DOWN);
1004         }
1005
1006         return push;
1007 }
1008
1009 void lnet_wait_router_start(void)
1010 {
1011         if (check_routers_before_use) {
1012                 /* Note that a helpful side-effect of pinging all known routers
1013                  * at startup is that it makes them drop stale connections they
1014                  * may have to a previous instance of me. */
1015                 lnet_wait_known_routerstate();
1016         }
1017 }
1018
1019 /*
1020  * This function is called from the monitor thread to check if there are
1021  * any active routers that need to be checked.
1022  */
1023 inline bool
1024 lnet_router_checker_active(void)
1025 {
1026         /* Router Checker thread needs to run when routing is enabled in
1027          * order to call lnet_update_ni_status_locked() */
1028         if (the_lnet.ln_routing)
1029                 return true;
1030
1031         return !list_empty(&the_lnet.ln_routers) &&
1032                 alive_router_check_interval > 0;
1033 }
1034
1035 void
1036 lnet_check_routers(void)
1037 {
1038         struct lnet_peer_net *first_lpn = NULL;
1039         struct lnet_peer_net *lpn;
1040         struct lnet_peer_ni *lpni;
1041         struct list_head *entry;
1042         struct lnet_peer *rtr;
1043         bool push = false;
1044         bool found_lpn;
1045         __u64 version;
1046         __u32 net_id;
1047         time64_t now;
1048         int cpt;
1049         int rc;
1050
1051         cpt = lnet_net_lock_current();
1052 rescan:
1053         version = the_lnet.ln_routers_version;
1054
1055         list_for_each(entry, &the_lnet.ln_routers) {
1056                 rtr = list_entry(entry, struct lnet_peer,
1057                                  lp_rtr_list);
1058
1059                 now = ktime_get_real_seconds();
1060
1061                 /*
1062                  * only discover the router if we've passed
1063                  * alive_router_check_interval seconds. Some of the router
1064                  * interfaces could be down and in that case they would be
1065                  * undergoing recovery separately from this discovery.
1066                  */
1067                 /* find next peer net which is also local */
1068                 net_id = rtr->lp_disc_net_id;
1069                 do {
1070                         lpn = lnet_get_next_peer_net_locked(rtr, net_id);
1071                         if (!lpn) {
1072                                 CERROR("gateway %s has no networks\n",
1073                                 libcfs_nid2str(rtr->lp_primary_nid));
1074                                 break;
1075                         }
1076                         if (first_lpn == lpn)
1077                                 break;
1078                         if (!first_lpn)
1079                                 first_lpn = lpn;
1080                         found_lpn = lnet_islocalnet_locked(lpn->lpn_net_id);
1081                         net_id = lpn->lpn_net_id;
1082                 } while (!found_lpn);
1083
1084                 if (!found_lpn || !lpn) {
1085                         CERROR("no local network found for gateway %s\n",
1086                                libcfs_nid2str(rtr->lp_primary_nid));
1087                         continue;
1088                 }
1089
1090                 if (now - lpn->lpn_rtrcheck_timestamp <
1091                     alive_router_check_interval / lnet_current_net_count)
1092                        continue;
1093
1094                 /*
1095                  * If we're currently discovering the peer then don't
1096                  * issue another discovery
1097                  */
1098                 spin_lock(&rtr->lp_lock);
1099                 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY) {
1100                         spin_unlock(&rtr->lp_lock);
1101                         continue;
1102                 }
1103                 /* make sure we actively discover the router */
1104                 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1105                 rtr->lp_state |= LNET_PEER_RTR_DISCOVERY;
1106                 spin_unlock(&rtr->lp_lock);
1107
1108                 /* find the peer_ni associated with the primary NID */
1109                 lpni = lnet_peer_get_ni_locked(rtr, rtr->lp_primary_nid);
1110                 if (!lpni) {
1111                         CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
1112                                libcfs_nid2str(rtr->lp_primary_nid));
1113                         continue;
1114                 }
1115                 lnet_peer_ni_addref_locked(lpni);
1116
1117                 /* specify the net to use */
1118                 rtr->lp_disc_net_id = lpn->lpn_net_id;
1119
1120                 /* discover the router */
1121                 CDEBUG(D_NET, "discover %s, cpt = %d\n",
1122                        libcfs_nid2str(lpni->lpni_nid), cpt);
1123                 rc = lnet_discover_peer_locked(lpni, cpt, false);
1124
1125                 /* decrement ref count acquired by find_peer_ni_locked() */
1126                 lnet_peer_ni_decref_locked(lpni);
1127
1128                 if (!rc)
1129                         lpn->lpn_rtrcheck_timestamp = now;
1130                 else
1131                         CERROR("Failed to discover router %s\n",
1132                                libcfs_nid2str(rtr->lp_primary_nid));
1133
1134                 /* NB dropped lock */
1135                 if (version != the_lnet.ln_routers_version) {
1136                         /* the routers list has changed */
1137                         goto rescan;
1138                 }
1139         }
1140
1141         if (the_lnet.ln_routing)
1142                 push = lnet_update_ni_status_locked();
1143
1144         lnet_net_unlock(cpt);
1145
1146         /* if the status of the ni changed update the peers */
1147         if (push)
1148                 lnet_push_update_to_peers(1);
1149 }
1150
1151 void
1152 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1153 {
1154         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1155
1156         while (--npages >= 0)
1157                 __free_page(rb->rb_kiov[npages].kiov_page);
1158
1159         LIBCFS_FREE(rb, sz);
1160 }
1161
1162 static struct lnet_rtrbuf *
1163 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1164 {
1165         int            npages = rbp->rbp_npages;
1166         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1167         struct page   *page;
1168         struct lnet_rtrbuf *rb;
1169         int            i;
1170
1171         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1172         if (rb == NULL)
1173                 return NULL;
1174
1175         rb->rb_pool = rbp;
1176
1177         for (i = 0; i < npages; i++) {
1178                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1179                                           GFP_KERNEL | __GFP_ZERO);
1180                 if (page == NULL) {
1181                         while (--i >= 0)
1182                                 __free_page(rb->rb_kiov[i].kiov_page);
1183
1184                         LIBCFS_FREE(rb, sz);
1185                         return NULL;
1186                 }
1187
1188                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1189                 rb->rb_kiov[i].kiov_offset = 0;
1190                 rb->rb_kiov[i].kiov_page = page;
1191         }
1192
1193         return rb;
1194 }
1195
1196 static void
1197 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1198 {
1199         int npages = rbp->rbp_npages;
1200         struct lnet_rtrbuf *rb;
1201         struct list_head tmp;
1202
1203         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1204                 return;
1205
1206         INIT_LIST_HEAD(&tmp);
1207
1208         lnet_net_lock(cpt);
1209         list_splice_init(&rbp->rbp_msgs, &tmp);
1210         lnet_drop_routed_msgs_locked(&tmp, cpt);
1211         list_splice_init(&rbp->rbp_bufs, &tmp);
1212         rbp->rbp_req_nbuffers = 0;
1213         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1214         rbp->rbp_mincredits = 0;
1215         lnet_net_unlock(cpt);
1216
1217         /* Free buffers on the free list. */
1218         while (!list_empty(&tmp)) {
1219                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1220                 list_del(&rb->rb_list);
1221                 lnet_destroy_rtrbuf(rb, npages);
1222         }
1223 }
1224
1225 static int
1226 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1227 {
1228         struct list_head rb_list;
1229         struct lnet_rtrbuf *rb;
1230         int             num_rb;
1231         int             num_buffers = 0;
1232         int             old_req_nbufs;
1233         int             npages = rbp->rbp_npages;
1234
1235         lnet_net_lock(cpt);
1236         /* If we are called for less buffers than already in the pool, we
1237          * just lower the req_nbuffers number and excess buffers will be
1238          * thrown away as they are returned to the free list.  Credits
1239          * then get adjusted as well.
1240          * If we already have enough buffers allocated to serve the
1241          * increase requested, then we can treat that the same way as we
1242          * do the decrease. */
1243         num_rb = nbufs - rbp->rbp_nbuffers;
1244         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1245                 rbp->rbp_req_nbuffers = nbufs;
1246                 lnet_net_unlock(cpt);
1247                 return 0;
1248         }
1249         /* store the older value of rbp_req_nbuffers and then set it to
1250          * the new request to prevent lnet_return_rx_credits_locked() from
1251          * freeing buffers that we need to keep around */
1252         old_req_nbufs = rbp->rbp_req_nbuffers;
1253         rbp->rbp_req_nbuffers = nbufs;
1254         lnet_net_unlock(cpt);
1255
1256         INIT_LIST_HEAD(&rb_list);
1257
1258         /* allocate the buffers on a local list first.  If all buffers are
1259          * allocated successfully then join this list to the rbp buffer
1260          * list.  If not then free all allocated buffers. */
1261         while (num_rb-- > 0) {
1262                 rb = lnet_new_rtrbuf(rbp, cpt);
1263                 if (rb == NULL) {
1264                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1265                                nbufs, npages);
1266
1267                         lnet_net_lock(cpt);
1268                         rbp->rbp_req_nbuffers = old_req_nbufs;
1269                         lnet_net_unlock(cpt);
1270
1271                         goto failed;
1272                 }
1273
1274                 list_add(&rb->rb_list, &rb_list);
1275                 num_buffers++;
1276         }
1277
1278         lnet_net_lock(cpt);
1279
1280         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1281         rbp->rbp_nbuffers += num_buffers;
1282         rbp->rbp_credits += num_buffers;
1283         rbp->rbp_mincredits = rbp->rbp_credits;
1284         /* We need to schedule blocked msg using the newly
1285          * added buffers. */
1286         while (!list_empty(&rbp->rbp_bufs) &&
1287                !list_empty(&rbp->rbp_msgs))
1288                 lnet_schedule_blocked_locked(rbp);
1289
1290         lnet_net_unlock(cpt);
1291
1292         return 0;
1293
1294 failed:
1295         while (!list_empty(&rb_list)) {
1296                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1297                 list_del(&rb->rb_list);
1298                 lnet_destroy_rtrbuf(rb, npages);
1299         }
1300
1301         return -ENOMEM;
1302 }
1303
1304 static void
1305 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1306 {
1307         INIT_LIST_HEAD(&rbp->rbp_msgs);
1308         INIT_LIST_HEAD(&rbp->rbp_bufs);
1309
1310         rbp->rbp_npages = npages;
1311         rbp->rbp_credits = 0;
1312         rbp->rbp_mincredits = 0;
1313 }
1314
1315 void
1316 lnet_rtrpools_free(int keep_pools)
1317 {
1318         struct lnet_rtrbufpool *rtrp;
1319         int               i;
1320
1321         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1322                 return;
1323
1324         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1325                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1326                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1327                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1328         }
1329
1330         if (!keep_pools) {
1331                 cfs_percpt_free(the_lnet.ln_rtrpools);
1332                 the_lnet.ln_rtrpools = NULL;
1333         }
1334 }
1335
1336 static int
1337 lnet_nrb_tiny_calculate(void)
1338 {
1339         int     nrbs = LNET_NRB_TINY;
1340
1341         if (tiny_router_buffers < 0) {
1342                 LCONSOLE_ERROR_MSG(0x10c,
1343                                    "tiny_router_buffers=%d invalid when "
1344                                    "routing enabled\n", tiny_router_buffers);
1345                 return -EINVAL;
1346         }
1347
1348         if (tiny_router_buffers > 0)
1349                 nrbs = tiny_router_buffers;
1350
1351         nrbs /= LNET_CPT_NUMBER;
1352         return max(nrbs, LNET_NRB_TINY_MIN);
1353 }
1354
1355 static int
1356 lnet_nrb_small_calculate(void)
1357 {
1358         int     nrbs = LNET_NRB_SMALL;
1359
1360         if (small_router_buffers < 0) {
1361                 LCONSOLE_ERROR_MSG(0x10c,
1362                                    "small_router_buffers=%d invalid when "
1363                                    "routing enabled\n", small_router_buffers);
1364                 return -EINVAL;
1365         }
1366
1367         if (small_router_buffers > 0)
1368                 nrbs = small_router_buffers;
1369
1370         nrbs /= LNET_CPT_NUMBER;
1371         return max(nrbs, LNET_NRB_SMALL_MIN);
1372 }
1373
1374 static int
1375 lnet_nrb_large_calculate(void)
1376 {
1377         int     nrbs = LNET_NRB_LARGE;
1378
1379         if (large_router_buffers < 0) {
1380                 LCONSOLE_ERROR_MSG(0x10c,
1381                                    "large_router_buffers=%d invalid when "
1382                                    "routing enabled\n", large_router_buffers);
1383                 return -EINVAL;
1384         }
1385
1386         if (large_router_buffers > 0)
1387                 nrbs = large_router_buffers;
1388
1389         nrbs /= LNET_CPT_NUMBER;
1390         return max(nrbs, LNET_NRB_LARGE_MIN);
1391 }
1392
1393 int
1394 lnet_rtrpools_alloc(int im_a_router)
1395 {
1396         struct lnet_rtrbufpool *rtrp;
1397         int     nrb_tiny;
1398         int     nrb_small;
1399         int     nrb_large;
1400         int     rc;
1401         int     i;
1402
1403         if (!strcmp(forwarding, "")) {
1404                 /* not set either way */
1405                 if (!im_a_router)
1406                         return 0;
1407         } else if (!strcmp(forwarding, "disabled")) {
1408                 /* explicitly disabled */
1409                 return 0;
1410         } else if (!strcmp(forwarding, "enabled")) {
1411                 /* explicitly enabled */
1412         } else {
1413                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1414                                    "'enabled' or 'disabled'\n");
1415                 return -EINVAL;
1416         }
1417
1418         nrb_tiny = lnet_nrb_tiny_calculate();
1419         if (nrb_tiny < 0)
1420                 return -EINVAL;
1421
1422         nrb_small = lnet_nrb_small_calculate();
1423         if (nrb_small < 0)
1424                 return -EINVAL;
1425
1426         nrb_large = lnet_nrb_large_calculate();
1427         if (nrb_large < 0)
1428                 return -EINVAL;
1429
1430         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1431                                                 LNET_NRBPOOLS *
1432                                                 sizeof(struct lnet_rtrbufpool));
1433         if (the_lnet.ln_rtrpools == NULL) {
1434                 LCONSOLE_ERROR_MSG(0x10c,
1435                                    "Failed to initialize router buffe pool\n");
1436                 return -ENOMEM;
1437         }
1438
1439         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1440                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1441                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1442                                               nrb_tiny, i);
1443                 if (rc != 0)
1444                         goto failed;
1445
1446                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1447                                   LNET_NRB_SMALL_PAGES);
1448                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1449                                               nrb_small, i);
1450                 if (rc != 0)
1451                         goto failed;
1452
1453                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1454                                   LNET_NRB_LARGE_PAGES);
1455                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1456                                               nrb_large, i);
1457                 if (rc != 0)
1458                         goto failed;
1459         }
1460
1461         lnet_net_lock(LNET_LOCK_EX);
1462         the_lnet.ln_routing = 1;
1463         lnet_net_unlock(LNET_LOCK_EX);
1464         complete(&the_lnet.ln_mt_wait_complete);
1465         return 0;
1466
1467  failed:
1468         lnet_rtrpools_free(0);
1469         return rc;
1470 }
1471
1472 static int
1473 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1474 {
1475         int nrb = 0;
1476         int rc = 0;
1477         int i;
1478         struct lnet_rtrbufpool *rtrp;
1479
1480         /* If the provided values for each buffer pool are different than the
1481          * configured values, we need to take action. */
1482         if (tiny >= 0) {
1483                 tiny_router_buffers = tiny;
1484                 nrb = lnet_nrb_tiny_calculate();
1485                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1486                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1487                                                       nrb, i);
1488                         if (rc != 0)
1489                                 return rc;
1490                 }
1491         }
1492         if (small >= 0) {
1493                 small_router_buffers = small;
1494                 nrb = lnet_nrb_small_calculate();
1495                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1496                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1497                                                       nrb, i);
1498                         if (rc != 0)
1499                                 return rc;
1500                 }
1501         }
1502         if (large >= 0) {
1503                 large_router_buffers = large;
1504                 nrb = lnet_nrb_large_calculate();
1505                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1506                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1507                                                       nrb, i);
1508                         if (rc != 0)
1509                                 return rc;
1510                 }
1511         }
1512
1513         return 0;
1514 }
1515
1516 int
1517 lnet_rtrpools_adjust(int tiny, int small, int large)
1518 {
1519         /* this function doesn't revert the changes if adding new buffers
1520          * failed.  It's up to the user space caller to revert the
1521          * changes. */
1522
1523         if (!the_lnet.ln_routing)
1524                 return 0;
1525
1526         return lnet_rtrpools_adjust_helper(tiny, small, large);
1527 }
1528
1529 int
1530 lnet_rtrpools_enable(void)
1531 {
1532         int rc = 0;
1533
1534         if (the_lnet.ln_routing)
1535                 return 0;
1536
1537         if (the_lnet.ln_rtrpools == NULL)
1538                 /* If routing is turned off, and we have never
1539                  * initialized the pools before, just call the
1540                  * standard buffer pool allocation routine as
1541                  * if we are just configuring this for the first
1542                  * time. */
1543                 rc = lnet_rtrpools_alloc(1);
1544         else
1545                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1546         if (rc != 0)
1547                 return rc;
1548
1549         lnet_net_lock(LNET_LOCK_EX);
1550         the_lnet.ln_routing = 1;
1551
1552         the_lnet.ln_ping_target->pb_info.pi_features &=
1553                 ~LNET_PING_FEAT_RTE_DISABLED;
1554         lnet_net_unlock(LNET_LOCK_EX);
1555
1556         if (lnet_peer_discovery_disabled)
1557                 CWARN("Consider turning discovery on to enable full "
1558                       "Multi-Rail routing functionality\n");
1559
1560         return rc;
1561 }
1562
1563 void
1564 lnet_rtrpools_disable(void)
1565 {
1566         if (!the_lnet.ln_routing)
1567                 return;
1568
1569         lnet_net_lock(LNET_LOCK_EX);
1570         the_lnet.ln_routing = 0;
1571         the_lnet.ln_ping_target->pb_info.pi_features |=
1572                 LNET_PING_FEAT_RTE_DISABLED;
1573
1574         tiny_router_buffers = 0;
1575         small_router_buffers = 0;
1576         large_router_buffers = 0;
1577         lnet_net_unlock(LNET_LOCK_EX);
1578         lnet_rtrpools_free(1);
1579 }
1580
1581 static inline void
1582 lnet_notify_peer_down(struct lnet_ni *ni, lnet_nid_t nid)
1583 {
1584         if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1585                 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1586 }
1587
1588 /*
1589  * ni: local NI used to communicate with the peer
1590  * nid: peer NID
1591  * alive: true if peer is alive, false otherwise
1592  * reset: reset health value. This is requested by the LND.
1593  * when: notificaiton time.
1594  */
1595 int
1596 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset,
1597             time64_t when)
1598 {
1599         struct lnet_peer_ni *lpni = NULL;
1600         struct lnet_route *route;
1601         struct lnet_peer *lp;
1602         time64_t now = ktime_get_seconds();
1603         int cpt;
1604
1605         LASSERT (!in_interrupt ());
1606
1607         CDEBUG (D_NET, "%s notifying %s: %s\n",
1608                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1609                 libcfs_nid2str(nid),
1610                 alive ? "up" : "down");
1611
1612         if (ni != NULL &&
1613             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1614                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1615                       libcfs_nid2str(nid), alive ? "birth" : "death",
1616                       libcfs_nid2str(ni->ni_nid));
1617                 return -EINVAL;
1618         }
1619
1620         /* can't do predictions... */
1621         if (when > now) {
1622                 CWARN("Ignoring prediction from %s of %s %s "
1623                       "%lld seconds in the future\n",
1624                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1625                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1626                 return -EINVAL;
1627         }
1628
1629         if (ni != NULL && !alive &&             /* LND telling me she's down */
1630             !auto_down) {                       /* auto-down disabled */
1631                 CDEBUG(D_NET, "Auto-down disabled\n");
1632                 return 0;
1633         }
1634
1635         /* must lock 0 since this is used for synchronization */
1636         lnet_net_lock(0);
1637
1638         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1639                 lnet_net_unlock(0);
1640                 return -ESHUTDOWN;
1641         }
1642
1643         lpni = lnet_find_peer_ni_locked(nid);
1644         if (lpni == NULL) {
1645                 /* nid not found */
1646                 lnet_net_unlock(0);
1647                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1648                 return 0;
1649         }
1650
1651         if (alive) {
1652                 if (reset)
1653                         lnet_set_healthv(&lpni->lpni_healthv,
1654                                          LNET_MAX_HEALTH_VALUE);
1655                 else
1656                         lnet_inc_healthv(&lpni->lpni_healthv);
1657         } else {
1658                 lnet_handle_remote_failure_locked(lpni);
1659         }
1660
1661         /* recalculate aliveness */
1662         alive = lnet_is_peer_ni_alive(lpni);
1663         lnet_net_unlock(0);
1664
1665         if (ni != NULL && !alive)
1666                 lnet_notify_peer_down(ni, lpni->lpni_nid);
1667
1668         cpt = lpni->lpni_cpt;
1669         lnet_net_lock(cpt);
1670         lnet_peer_ni_decref_locked(lpni);
1671         if (lpni && lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer) {
1672                 lp = lpni->lpni_peer_net->lpn_peer;
1673                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
1674                         lnet_set_route_aliveness(route, alive);
1675         }
1676         lnet_net_unlock(cpt);
1677
1678         return 0;
1679 }
1680 EXPORT_SYMBOL(lnet_notify);