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