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 void lnet_add_route_to_rnet(struct lnet_remotenet *rnet,
 122                                    struct lnet_route *route);
 123 static void lnet_del_route_from_rnet(struct lnet_nid *gw_nid,
 124                                      struct list_head *route_list,
 125                                      struct list_head *zombies);
 126
 127 static int
 128 rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 129 {
 130         int rc;
 131         unsigned *sen = (unsigned *)kp->arg;
 132         unsigned long value;
 133
 134         rc = kstrtoul(val, 0, &value);
 135         if (rc) {
 136                 CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
 137                 return rc;
 138         }
 139
 140         if (value < 0 || value > 100) {
 141                 CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
 142                 return -EINVAL;
 143         }
 144
 145         /*
 146          * The purpose of locking the api_mutex here is to ensure that
 147          * the correct value ends up stored properly.
 148          */
 149         mutex_lock(&the_lnet.ln_api_mutex);
 150
 151         *sen = value;
 152
 153         mutex_unlock(&the_lnet.ln_api_mutex);
 154
 155         return 0;
 156 }
 157
 158 void
 159 lnet_move_route(struct lnet_route *route, struct lnet_peer *lp,
 160                 struct list_head *rt_list)
 161 {
 162         struct lnet_remotenet *rnet;
 163         struct list_head zombies;
 164         struct list_head *l;
 165
 166         INIT_LIST_HEAD(&zombies);
 167
 168         if (rt_list)
 169                 l = rt_list;
 170         else
 171                 l = &zombies;
 172
 173         rnet = lnet_find_rnet_locked(route->lr_net);
 174         LASSERT(rnet);
 175
 176         CDEBUG(D_NET, "deleting route %s->%s\n",
 177                libcfs_net2str(route->lr_net),
 178                libcfs_nidstr(&route->lr_nid));
 179
 180         /*
 181          * use the gateway's lp_primary_nid to delete the route as the
 182          * lr_nid can be a constituent NID of the peer
 183          */
 184         lnet_del_route_from_rnet(
 185                 &route->lr_gateway->lp_primary_nid,
 186                 &rnet->lrn_routes, l);
 187
 188         if (lp) {
 189                 route = list_first_entry(l, struct lnet_route,
 190                                          lr_list);
 191                 route->lr_gateway = lp;
 192                 lnet_add_route_to_rnet(rnet, route);
 193         } else {
 194                 while (!list_empty(l) && !rt_list) {
 195                         route = list_first_entry(l, struct lnet_route,
 196                                                  lr_list);
 197                         list_del(&route->lr_list);
 198                         LIBCFS_FREE(route, sizeof(*route));
 199                 }
 200         }
 201 }
 202
 203 void
 204 lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target)
 205 {
 206         struct lnet_route *route;
 207         struct lnet_route *tmp, *tmp2;
 208
 209         lnet_net_lock(LNET_LOCK_EX);
 210         CDEBUG(D_NET, "transfering routes from %s -> %s\n",
 211                libcfs_nidstr(&src->lp_primary_nid),
 212                libcfs_nidstr(&target->lp_primary_nid));
 213         list_for_each_entry(route, &src->lp_routes, lr_gwlist) {
 214                 CDEBUG(D_NET, "%s: %s->%s\n",
 215                        libcfs_nidstr(&src->lp_primary_nid),
 216                        libcfs_net2str(route->lr_net),
 217                        libcfs_nidstr(&route->lr_nid));
 218         }
 219         list_splice_init(&src->lp_rtrq, &target->lp_rtrq);
 220         list_for_each_entry_safe(route, tmp, &src->lp_routes, lr_gwlist) {
 221                 struct lnet_route *r2;
 222                 bool present = false;
 223                 list_for_each_entry_safe(r2, tmp2, &target->lp_routes, lr_gwlist) {
 224                         if (route->lr_net == r2->lr_net) {
 225                                 if (route->lr_priority >= r2->lr_priority)
 226                                         present = true;
 227                                 else if (route->lr_hops >= r2->lr_hops)
 228                                         present = true;
 229                                 else
 230                                         lnet_move_route(r2, NULL, NULL);
 231                         }
 232                 }
 233                 if (present)
 234                         lnet_move_route(route, NULL, NULL);
 235                 else
 236                         lnet_move_route(route, target, NULL);
 237         }
 238
 239         if (list_empty(&target->lp_rtr_list)) {
 240                 lnet_peer_addref_locked(target);
 241                 list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers);
 242         }
 243
 244         the_lnet.ln_routers_version++;
 245         lnet_net_unlock(LNET_LOCK_EX);
 246 }
 247
 248 int
 249 lnet_peers_start_down(void)
 250 {
 251         return check_routers_before_use;
 252 }
 253
 254 /*
 255  * The peer_net of a gateway is alive if at least one of the peer_ni's on
 256  * that peer_net is alive.
 257  */
 258 static bool
 259 lnet_is_gateway_net_alive(struct lnet_peer_net *lpn)
 260 {
 261         struct lnet_peer_ni *lpni;
 262
 263         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 264                 if (lnet_is_peer_ni_alive(lpni))
 265                         return true;
 266         }
 267
 268         return false;
 269 }
 270
 271 /*
 272  * a gateway is alive only if all its nets are alive
 273  * called with cpt lock held
 274  */
 275 bool lnet_is_gateway_alive(struct lnet_peer *gw)
 276 {
 277         struct lnet_peer_net *lpn;
 278
 279         if (!gw->lp_alive)
 280                 return false;
 281
 282         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 283                 if (!lnet_is_gateway_net_alive(lpn))
 284                         return false;
 285         }
 286
 287         return true;
 288 }
 289
 290 /*
 291  * lnet_is_route_alive() needs to be called with cpt lock held
 292  * A route is alive if the gateway can route between the local network and
 293  * the remote network of the route.
 294  * This means at least one NI is alive on each of the local and remote
 295  * networks of the gateway.
 296  */
 297 bool lnet_is_route_alive(struct lnet_route *route)
 298 {
 299         struct lnet_peer *gw = route->lr_gateway;
 300         struct lnet_peer_net *llpn;
 301         struct lnet_peer_net *rlpn;
 302
 303         /* If the gateway is down then all routes are considered down */
 304         if (!gw->lp_alive)
 305                 return false;
 306
 307         /*
 308          * if discovery is disabled then rely on the cached aliveness
 309          * information. This is handicapped information which we log when
 310          * we receive the discovery ping response. The most uptodate
 311          * aliveness information can only be obtained when discovery is
 312          * enabled.
 313          */
 314         if (lnet_is_discovery_disabled(gw))
 315                 return atomic_read(&route->lr_alive) == 1;
 316
 317         /*
 318          * check the gateway's interfaces on the local network
 319          */
 320         llpn = lnet_peer_get_net_locked(gw, route->lr_lnet);
 321         if (!llpn)
 322                 return false;
 323
 324         if (!lnet_is_gateway_net_alive(llpn))
 325                 return false;
 326
 327         /*
 328          * For single hop routes avoid_asym_router_failure dictates
 329          * that the remote net must exist on the gateway. For multi-hop
 330          * routes the next-hop will not have the remote net.
 331          */
 332         if (avoid_asym_router_failure &&
 333             (route->lr_hops == 1 || route->lr_single_hop)) {
 334                 rlpn = lnet_peer_get_net_locked(gw, route->lr_net);
 335                 if (!rlpn)
 336                         return false;
 337                 if (!lnet_is_gateway_net_alive(rlpn))
 338                         return false;
 339         }
 340
 341         spin_lock(&gw->lp_lock);
 342         if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) {
 343                 spin_unlock(&gw->lp_lock);
 344                 if (gw->lp_rtr_refcount > 0)
 345                         CERROR("peer %s is being used as a gateway but routing feature is not turned on\n",
 346                                libcfs_nidstr(&gw->lp_primary_nid));
 347                 return false;
 348         }
 349         spin_unlock(&gw->lp_lock);
 350
 351         return true;
 352 }
 353
 354 void
 355 lnet_consolidate_routes_locked(struct lnet_peer *orig_lp,
 356                                struct lnet_peer *new_lp)
 357 {
 358         struct lnet_peer_ni *lpni;
 359         struct lnet_route *route;
 360
 361         /*
 362          * Although a route is correlated with a peer, but when it's added
 363          * a specific NID is used. That NID refers to a peer_ni within
 364          * a peer. There could be other peer_nis on the same net, which
 365          * can be used to send to that gateway. However when we are
 366          * consolidating gateways because of discovery, the nid used to
 367          * add the route might've moved between gateway peers. In this
 368          * case we want to move the route to the new gateway as well. The
 369          * intent here is not to confuse the user who added the route.
 370          */
 371         list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) {
 372                 lpni = lnet_peer_ni_get_locked(orig_lp, &route->lr_nid);
 373                 if (!lpni) {
 374                         lnet_net_lock(LNET_LOCK_EX);
 375                         list_move(&route->lr_gwlist, &new_lp->lp_routes);
 376                         lnet_net_unlock(LNET_LOCK_EX);
 377                 }
 378         }
 379 }
 380
 381 static inline void
 382 lnet_check_route_inconsistency(struct lnet_route *route)
 383 {
 384         if (!route->lr_single_hop &&
 385             (route->lr_hops == 1 || route->lr_hops == LNET_UNDEFINED_HOPS) &&
 386             avoid_asym_router_failure) {
 387                 CWARN("route %s->%s is detected to be multi-hop but hop count is set to %d\n",
 388                         libcfs_net2str(route->lr_net),
 389                         libcfs_nidstr(&route->lr_gateway->lp_primary_nid),
 390                         (int) route->lr_hops);
 391         }
 392 }
 393
 394 static void
 395 lnet_set_route_hop_type(struct lnet_peer *gw, struct lnet_route *route)
 396 {
 397         struct lnet_peer_net *lpn;
 398         bool single_hop = false;
 399
 400         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 401                 if (route->lr_net == lpn->lpn_net_id) {
 402                         single_hop = true;
 403                         break;
 404                 }
 405         }
 406         route->lr_single_hop = single_hop;
 407         lnet_check_route_inconsistency(route);
 408 }
 409
 410 /* Must hold net_lock/EX */
 411 void
 412 lnet_router_discovery_ping_reply(struct lnet_peer *lp)
 413 {
 414         struct lnet_ping_buffer *pbuf = lp->lp_data;
 415         struct lnet_peer_net *llpn;
 416         struct lnet_route *route;
 417         bool single_hop = false;
 418         bool net_up = false;
 419         unsigned lp_state;
 420         __u32 net;
 421         int i;
 422
 423
 424         spin_lock(&lp->lp_lock);
 425         lp_state = lp->lp_state;
 426
 427         /* only handle replies if discovery is disabled. */
 428         if (!lnet_is_discovery_disabled_locked(lp)) {
 429                 spin_unlock(&lp->lp_lock);
 430                 return;
 431         }
 432
 433         spin_unlock(&lp->lp_lock);
 434
 435         if (lp_state & LNET_PEER_PING_FAILED ||
 436             pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED) {
 437                 CDEBUG(D_NET, "Set routes down for gw %s because %s %d\n",
 438                        libcfs_nidstr(&lp->lp_primary_nid),
 439                        lp_state & LNET_PEER_PING_FAILED ? "ping failed" :
 440                        "route feature is disabled", lp->lp_ping_error);
 441                 /* If the ping failed or the peer has routing disabled then
 442                  * mark the routes served by this peer down
 443                  */
 444                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 445                         lnet_set_route_aliveness(route, false);
 446                 return;
 447         }
 448
 449         CDEBUG(D_NET, "Discovery is disabled. Processing reply for gw: %s:%d\n",
 450                libcfs_nidstr(&lp->lp_primary_nid), pbuf->pb_info.pi_nnis);
 451
 452         /*
 453          * examine the ping response to determine if the routes on that
 454          * gateway should be declared alive.
 455          * The route is alive if:
 456          *  1. local network to reach the route is alive and
 457          *  2. route is single hop, avoid_async_router_failure is set and
 458          *     there exists at least one NI on the route's remote net
 459          */
 460         list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
 461                 llpn = lnet_peer_get_net_locked(lp, route->lr_lnet);
 462                 if (!llpn) {
 463                         lnet_set_route_aliveness(route, false);
 464                         continue;
 465                 }
 466
 467                 if (!lnet_is_gateway_net_alive(llpn)) {
 468                         lnet_set_route_aliveness(route, false);
 469                         continue;
 470                 }
 471
 472                 single_hop = net_up = false;
 473                 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
 474                         net = LNET_NIDNET(pbuf->pb_info.pi_ni[i].ns_nid);
 475
 476                         if (route->lr_net == net) {
 477                                 single_hop = true;
 478                                 if (pbuf->pb_info.pi_ni[i].ns_status ==
 479                                     LNET_NI_STATUS_UP) {
 480                                         net_up = true;
 481                                         break;
 482                                 }
 483                         }
 484                 }
 485
 486                 route->lr_single_hop = single_hop;
 487                 if (avoid_asym_router_failure &&
 488                     (route->lr_hops == 1 || route->lr_single_hop))
 489                         lnet_set_route_aliveness(route, net_up);
 490                 else
 491                         lnet_set_route_aliveness(route, true);
 492
 493                 /*
 494                  * warn that the route is configured as single-hop but it
 495                  * really is multi-hop as far as we can tell.
 496                  */
 497                 lnet_check_route_inconsistency(route);
 498         }
 499 }
 500
 501 void
 502 lnet_router_discovery_complete(struct lnet_peer *lp)
 503 {
 504         struct lnet_peer_ni *lpni = NULL;
 505         struct lnet_route *route;
 506
 507         spin_lock(&lp->lp_lock);
 508         lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
 509         lp->lp_state |= LNET_PEER_RTR_DISCOVERED;
 510         lp->lp_alive = lp->lp_dc_error == 0;
 511         spin_unlock(&lp->lp_lock);
 512
 513         if (!lp->lp_dc_error) {
 514                 /* ping replies are being handled when discovery is disabled */
 515                 if (lnet_is_discovery_disabled_locked(lp))
 516                         return;
 517
 518                 /*
 519                 * mark single-hop routes.  If the remote net is not configured on
 520                 * the gateway we assume this is intentional and we mark the
 521                 * gateway as multi-hop
 522                 */
 523                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
 524                         lnet_set_route_aliveness(route, true);
 525                         lnet_set_route_hop_type(lp, route);
 526                 }
 527
 528                 return;
 529         }
 530
 531         /*
 532          * We do not send messages directly to the remote interfaces
 533          * of an LNet router. As such, we rely on the PING response
 534          * to determine the up/down status of these interfaces. If
 535          * a PING response is not receieved, or some other problem with
 536          * discovery occurs that prevents us from getting this status,
 537          * we assume all interfaces are down until we're able to
 538          * determine otherwise.
 539          */
 540         CDEBUG(D_NET, "%s: Router discovery failed %d\n",
 541                libcfs_nidstr(&lp->lp_primary_nid), lp->lp_dc_error);
 542         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 543                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 544
 545         list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 546                 lnet_set_route_aliveness(route, false);
 547 }
 548
 549 static void
 550 lnet_rtr_addref_locked(struct lnet_peer *lp)
 551 {
 552         LASSERT(lp->lp_rtr_refcount >= 0);
 553
 554         /* lnet_net_lock must be exclusively locked */
 555         lp->lp_rtr_refcount++;
 556         if (lp->lp_rtr_refcount == 1) {
 557                 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
 558                 /* addref for the_lnet.ln_routers */
 559                 lnet_peer_addref_locked(lp);
 560                 the_lnet.ln_routers_version++;
 561         }
 562 }
 563
 564 static void
 565 lnet_rtr_decref_locked(struct lnet_peer *lp)
 566 {
 567         LASSERT(atomic_read(&lp->lp_refcount) > 0);
 568         LASSERT(lp->lp_rtr_refcount > 0);
 569
 570         /* lnet_net_lock must be exclusively locked */
 571         lp->lp_rtr_refcount--;
 572         if (lp->lp_rtr_refcount == 0) {
 573                 LASSERT(list_empty(&lp->lp_routes));
 574
 575                 list_del(&lp->lp_rtr_list);
 576                 /* decref for the_lnet.ln_routers */
 577                 lnet_peer_decref_locked(lp);
 578                 the_lnet.ln_routers_version++;
 579         }
 580 }
 581
 582 struct lnet_remotenet *
 583 lnet_find_rnet_locked(__u32 net)
 584 {
 585         struct lnet_remotenet *rnet;
 586         struct list_head *rn_list;
 587
 588         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 589
 590         rn_list = lnet_net2rnethash(net);
 591         list_for_each_entry(rnet, rn_list, lrn_list) {
 592                 if (rnet->lrn_net == net)
 593                         return rnet;
 594         }
 595         return NULL;
 596 }
 597
 598 static void lnet_shuffle_seed(void)
 599 {
 600         static int seeded;
 601         struct lnet_ni *ni = NULL;
 602
 603         if (seeded)
 604                 return;
 605
 606         /* Nodes with small feet have little entropy
 607          * the NID for this node gives the most entropy in the low bits */
 608         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 609                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 610
 611         seeded = 1;
 612 }
 613
 614 /* NB expects LNET_LOCK held */
 615 static void
 616 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 617 {
 618         struct lnet_peer_net *lpn;
 619         unsigned int offset = 0;
 620         unsigned int len = 0;
 621         struct list_head *e;
 622         time64_t now;
 623
 624         lnet_shuffle_seed();
 625
 626         list_for_each(e, &rnet->lrn_routes)
 627                 len++;
 628
 629         /*
 630          * Randomly adding routes to the list is done to ensure that when
 631          * different nodes are using the same list of routers, they end up
 632          * preferring different routers.
 633          */
 634         offset = prandom_u32_max(len + 1);
 635         list_for_each(e, &rnet->lrn_routes) {
 636                 if (offset == 0)
 637                         break;
 638                 offset--;
 639         }
 640         list_add(&route->lr_list, e);
 641         /*
 642          * force a router check on the gateway to make sure the route is
 643          * alive
 644          */
 645         now = ktime_get_real_seconds();
 646         list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets,
 647                             lpn_peer_nets) {
 648                 lpn->lpn_next_ping = now;
 649         }
 650
 651         the_lnet.ln_remote_nets_version++;
 652
 653         /* add the route on the gateway list */
 654         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 655
 656         /* take a router reference count on the gateway */
 657         lnet_rtr_addref_locked(route->lr_gateway);
 658 }
 659
 660 int
 661 lnet_add_route(__u32 net, __u32 hops, struct lnet_nid *gateway,
 662                __u32 priority, __u32 sensitivity)
 663 {
 664         struct list_head *route_entry;
 665         struct lnet_remotenet *rnet;
 666         struct lnet_remotenet *rnet2;
 667         struct lnet_route *route;
 668         struct lnet_peer_ni *lpni;
 669         struct lnet_peer *gw;
 670         int add_route;
 671         int rc;
 672
 673         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 674                libcfs_net2str(net), hops, priority, libcfs_nidstr(gateway));
 675
 676         if (LNET_NID_IS_ANY(gateway) ||
 677             nid_is_lo0(gateway) ||
 678             net == LNET_NET_ANY ||
 679             LNET_NETTYP(net) == LOLND ||
 680             LNET_NID_NET(gateway) == net ||
 681             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 682                 return -EINVAL;
 683
 684         /* it's a local network */
 685         if (lnet_islocalnet(net))
 686                 return -EEXIST;
 687
 688         if (!lnet_islocalnet(LNET_NID_NET(gateway))) {
 689                 CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n",
 690                        libcfs_nidstr(gateway),
 691                        libcfs_net2str(LNET_NID_NET(gateway)));
 692                 return -EHOSTUNREACH;
 693         }
 694
 695         /* Assume net, route, all new */
 696         LIBCFS_ALLOC(route, sizeof(*route));
 697         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 698         if (route == NULL || rnet == NULL) {
 699                 CERROR("Out of memory creating route %s %d %s\n",
 700                        libcfs_net2str(net), hops, libcfs_nidstr(gateway));
 701                 if (route != NULL)
 702                         LIBCFS_FREE(route, sizeof(*route));
 703                 if (rnet != NULL)
 704                         LIBCFS_FREE(rnet, sizeof(*rnet));
 705                 return -ENOMEM;
 706         }
 707
 708         INIT_LIST_HEAD(&rnet->lrn_routes);
 709         rnet->lrn_net = net;
 710         /* store the local and remote net that the route represents */
 711         route->lr_lnet = LNET_NID_NET(gateway);
 712         route->lr_net = net;
 713         route->lr_nid = *gateway;
 714         route->lr_priority = priority;
 715         route->lr_hops = hops;
 716         if (lnet_peers_start_down())
 717                 atomic_set(&route->lr_alive, 0);
 718         else
 719                 atomic_set(&route->lr_alive, 1);
 720
 721         lnet_net_lock(LNET_LOCK_EX);
 722
 723         /*
 724          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 725          * lose that once we're done
 726          */
 727         lpni = lnet_nid2peerni_ex(gateway);
 728         if (IS_ERR(lpni)) {
 729                 lnet_net_unlock(LNET_LOCK_EX);
 730
 731                 LIBCFS_FREE(route, sizeof(*route));
 732                 LIBCFS_FREE(rnet, sizeof(*rnet));
 733
 734                 rc = PTR_ERR(lpni);
 735                 CERROR("Error %d creating route %s %d %s\n", rc,
 736                         libcfs_net2str(net), hops,
 737                         libcfs_nidstr(gateway));
 738                 return rc;
 739         }
 740
 741         LASSERT(lpni);
 742         LASSERT(lpni->lpni_peer_net);
 743         LASSERT(lpni->lpni_peer_net->lpn_peer);
 744         gw = lpni->lpni_peer_net->lpn_peer;
 745
 746         route->lr_gateway = gw;
 747
 748         rnet2 = lnet_find_rnet_locked(net);
 749         if (rnet2 == NULL) {
 750                 /* new network */
 751                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 752                 rnet2 = rnet;
 753         }
 754
 755         /* Search for a duplicate route (it's a NOOP if it is) */
 756         add_route = 1;
 757         list_for_each(route_entry, &rnet2->lrn_routes) {
 758                 struct lnet_route *route2;
 759
 760                 route2 = list_entry(route_entry, struct lnet_route, lr_list);
 761                 if (route2->lr_gateway == route->lr_gateway) {
 762                         add_route = 0;
 763                         break;
 764                 }
 765
 766                 /* our lookups must be true */
 767                 LASSERT(!nid_same(&route2->lr_gateway->lp_primary_nid,
 768                                   gateway));
 769         }
 770
 771         /*
 772          * It is possible to add multiple routes through the same peer,
 773          * but it'll be using a different NID of that peer. When the
 774          * gateway is discovered, discovery will consolidate the different
 775          * peers into one peer. In this case the discovery code will have
 776          * to move the routes from the peer that's being deleted to the
 777          * consolidated peer lp_routes list
 778          */
 779         if (add_route) {
 780                 gw->lp_health_sensitivity = sensitivity;
 781                 lnet_add_route_to_rnet(rnet2, route);
 782                 if (lnet_peer_discovery_disabled)
 783                         CWARN("Consider turning discovery on to enable full Multi-Rail routing functionality\n");
 784         }
 785
 786         /*
 787          * get rid of the reference on the lpni.
 788          */
 789         lnet_peer_ni_decref_locked(lpni);
 790         lnet_net_unlock(LNET_LOCK_EX);
 791
 792         /* If avoid_asym_router_failure is enabled and hop count is not
 793          * set to 1 for a route that is actually single-hop, then the
 794          * feature will fail to prevent the router from being selected
 795          * if it is missing a NI on the remote network due to misconfiguration.
 796          */
 797         if (avoid_asym_router_failure && hops == LNET_UNDEFINED_HOPS)
 798                 CWARN("Use hops = 1 for a single-hop route when avoid_asym_router_failure feature is enabled\n");
 799
 800         rc = 0;
 801
 802         if (!add_route) {
 803                 rc = -EEXIST;
 804                 LIBCFS_FREE(route, sizeof(*route));
 805         }
 806
 807         if (rnet != rnet2)
 808                 LIBCFS_FREE(rnet, sizeof(*rnet));
 809
 810         /* kick start the monitor thread to handle the added route */
 811         complete(&the_lnet.ln_mt_wait_complete);
 812
 813         return rc;
 814 }
 815
 816 void
 817 lnet_del_route_from_rnet(struct lnet_nid *gw_nid,
 818                          struct list_head *route_list,
 819                          struct list_head *zombies)
 820 {
 821         struct lnet_peer *gateway;
 822         struct lnet_route *route;
 823         struct lnet_route *tmp;
 824
 825         list_for_each_entry_safe(route, tmp, route_list, lr_list) {
 826                 gateway = route->lr_gateway;
 827                 if (gw_nid && !nid_same(gw_nid, &gateway->lp_primary_nid))
 828                         continue;
 829
 830                 /*
 831                  * move to zombie to delete outside the lock
 832                  * Note that this function is called with the
 833                  * ln_api_mutex held as well as the exclusive net
 834                  * lock. Adding to the remote net list happens
 835                  * under the same conditions. Same goes for the
 836                  * gateway router list
 837                  */
 838                 list_move(&route->lr_list, zombies);
 839                 the_lnet.ln_remote_nets_version++;
 840
 841                 list_del(&route->lr_gwlist);
 842                 lnet_rtr_decref_locked(gateway);
 843         }
 844 }
 845
 846 int
 847 lnet_del_route(__u32 net, struct lnet_nid *gw)
 848 {
 849         LIST_HEAD(rnet_zombies);
 850         struct lnet_remotenet *rnet;
 851         struct lnet_remotenet *tmp;
 852         struct list_head *rn_list;
 853         struct lnet_peer_ni *lpni;
 854         struct lnet_route *route;
 855         struct lnet_nid gw_nid;
 856         LIST_HEAD(zombies);
 857         struct lnet_peer *lp = NULL;
 858         int i = 0;
 859
 860         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 861                libcfs_net2str(net), libcfs_nidstr(gw));
 862
 863         /* NB Caller may specify either all routes via the given gateway
 864          * or a specific route entry actual NIDs) */
 865
 866         lnet_net_lock(LNET_LOCK_EX);
 867
 868         if (gw)
 869                 lpni = lnet_peer_ni_find_locked(gw);
 870         else
 871                 lpni = NULL;
 872         if (lpni) {
 873                 lp = lpni->lpni_peer_net->lpn_peer;
 874                 LASSERT(lp);
 875                 gw_nid = lp->lp_primary_nid;
 876                 gw = &gw_nid;
 877                 lnet_peer_ni_decref_locked(lpni);
 878         }
 879
 880         if (net != LNET_NET_ANY) {
 881                 rnet = lnet_find_rnet_locked(net);
 882                 if (!rnet) {
 883                         lnet_net_unlock(LNET_LOCK_EX);
 884                         return -ENOENT;
 885                 }
 886                 lnet_del_route_from_rnet(gw, &rnet->lrn_routes,
 887                                          &zombies);
 888                 if (list_empty(&rnet->lrn_routes))
 889                         list_move(&rnet->lrn_list, &rnet_zombies);
 890                 goto delete_zombies;
 891         }
 892
 893         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 894                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 895
 896                 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
 897                         lnet_del_route_from_rnet(gw, &rnet->lrn_routes,
 898                                                  &zombies);
 899                         if (list_empty(&rnet->lrn_routes))
 900                                 list_move(&rnet->lrn_list, &rnet_zombies);
 901                 }
 902         }
 903
 904 delete_zombies:
 905         /*
 906          * check if there are any routes remaining on the gateway
 907          * If there are no more routes make sure to set the peer's
 908          * lp_disc_net_id to 0 (invalid), in case we add more routes in
 909          * the future on that gateway, then we start our discovery process
 910          * from scratch
 911          */
 912         if (lpni) {
 913                 if (list_empty(&lp->lp_routes))
 914                         lp->lp_disc_net_id = 0;
 915         }
 916
 917         lnet_net_unlock(LNET_LOCK_EX);
 918
 919         while (!list_empty(&zombies)) {
 920                 route = list_first_entry(&zombies, struct lnet_route, lr_list);
 921                 list_del(&route->lr_list);
 922                 LIBCFS_FREE(route, sizeof(*route));
 923         }
 924
 925         while (!list_empty(&rnet_zombies)) {
 926                 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
 927                                         lrn_list);
 928                 list_del(&rnet->lrn_list);
 929                 LIBCFS_FREE(rnet, sizeof(*rnet));
 930         }
 931
 932         return 0;
 933 }
 934
 935 void
 936 lnet_destroy_routes(void)
 937 {
 938         lnet_del_route(LNET_NET_ANY, NULL);
 939 }
 940
 941 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 942 {
 943         struct lnet_rtrbufpool *rbp;
 944         int i, rc = -ENOENT, j;
 945
 946         if (the_lnet.ln_rtrpools == NULL)
 947                 return rc;
 948
 949
 950         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 951                 if (i != cpt)
 952                         continue;
 953
 954                 lnet_net_lock(i);
 955                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 956                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 957                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 958                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 959                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 960                 }
 961                 lnet_net_unlock(i);
 962                 rc = 0;
 963                 break;
 964         }
 965
 966         lnet_net_lock(LNET_LOCK_EX);
 967         pool_cfg->pl_routing = the_lnet.ln_routing;
 968         lnet_net_unlock(LNET_LOCK_EX);
 969
 970         return rc;
 971 }
 972
 973 int
 974 lnet_get_route(int idx, __u32 *net, __u32 *hops, lnet_nid_t *gateway,
 975                __u32 *flags, __u32 *priority, __u32 *sensitivity)
 976 {
 977         struct lnet_remotenet *rnet;
 978         struct list_head *rn_list;
 979         struct lnet_route *route;
 980         int cpt;
 981         int i;
 982
 983         cpt = lnet_net_lock_current();
 984
 985         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 986                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 987                 list_for_each_entry(rnet, rn_list, lrn_list) {
 988                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
 989                                 if (idx-- == 0) {
 990                                         *net      = rnet->lrn_net;
 991                                         *gateway  = lnet_nid_to_nid4(&route->lr_nid);
 992                                         *hops     = route->lr_hops;
 993                                         *priority = route->lr_priority;
 994                                         *sensitivity = route->lr_gateway->
 995                                                 lp_health_sensitivity;
 996                                         if (lnet_is_route_alive(route))
 997                                                 *flags |= LNET_RT_ALIVE;
 998                                         else
 999                                                 *flags &= ~LNET_RT_ALIVE;
1000                                         if (route->lr_single_hop)
1001                                                 *flags &= ~LNET_RT_MULTI_HOP;
1002                                         else
1003                                                 *flags |= LNET_RT_MULTI_HOP;
1004                                         lnet_net_unlock(cpt);
1005                                         return 0;
1006                                 }
1007                         }
1008                 }
1009         }
1010
1011         lnet_net_unlock(cpt);
1012         return -ENOENT;
1013 }
1014
1015 static void
1016 lnet_wait_known_routerstate(void)
1017 {
1018         struct lnet_peer *rtr;
1019         int all_known;
1020
1021         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
1022
1023         for (;;) {
1024                 int cpt = lnet_net_lock_current();
1025
1026                 all_known = 1;
1027                 list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1028                         spin_lock(&rtr->lp_lock);
1029
1030                         if ((rtr->lp_state & LNET_PEER_RTR_DISCOVERED) == 0) {
1031                                 all_known = 0;
1032                                 spin_unlock(&rtr->lp_lock);
1033                                 break;
1034                         }
1035                         spin_unlock(&rtr->lp_lock);
1036                 }
1037
1038                 lnet_net_unlock(cpt);
1039
1040                 if (all_known)
1041                         return;
1042
1043                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
1044         }
1045 }
1046
1047 static inline bool
1048 lnet_net_set_status_locked(struct lnet_net *net, __u32 status)
1049 {
1050         struct lnet_ni *ni;
1051         bool update = false;
1052
1053         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1054                 if (lnet_ni_set_status(ni, status))
1055                         update = true;
1056
1057         return update;
1058 }
1059
1060 static bool
1061 lnet_update_ni_status_locked(void)
1062 {
1063         struct lnet_net *net;
1064         struct lnet_ni *ni;
1065         bool push = false;
1066         time64_t now;
1067         time64_t timeout;
1068
1069         LASSERT(the_lnet.ln_routing);
1070
1071         timeout = router_ping_timeout + alive_router_check_interval;
1072
1073         now = ktime_get_seconds();
1074         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1075                 if (net->net_lnd->lnd_type == LOLND)
1076                         continue;
1077
1078                 if (now < net->net_last_alive + timeout)
1079                         goto check_ni_fatal;
1080
1081                 spin_lock(&net->net_lock);
1082                 /* re-check with lock */
1083                 if (now < net->net_last_alive + timeout) {
1084                         spin_unlock(&net->net_lock);
1085                         goto check_ni_fatal;
1086                 }
1087                 spin_unlock(&net->net_lock);
1088
1089                 /*
1090                  * if the net didn't receive any traffic for past the
1091                  * timeout on any of its constituent NIs, then mark all
1092                  * the NIs down.
1093                  */
1094                 if (lnet_net_set_status_locked(net, LNET_NI_STATUS_DOWN)) {
1095                         push = true;
1096                         continue;
1097                 }
1098
1099 check_ni_fatal:
1100                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1101                         /* lnet_ni_set_status() will perform the same check of
1102                          * ni_status while holding the ni lock. We can safely
1103                          * check ni_status without that lock because it is only
1104                          * written to under net_lock/EX and our caller is
1105                          * holding a net lock.
1106                          */
1107                         if (atomic_read(&ni->ni_fatal_error_on) &&
1108                             ni->ni_status &&
1109                             *ni->ni_status != LNET_NI_STATUS_DOWN &&
1110                             lnet_ni_set_status(ni, LNET_NI_STATUS_DOWN))
1111                                 push = true;
1112                 }
1113         }
1114
1115         return push;
1116 }
1117
1118 void lnet_wait_router_start(void)
1119 {
1120         if (check_routers_before_use) {
1121                 /* Note that a helpful side-effect of pinging all known routers
1122                  * at startup is that it makes them drop stale connections they
1123                  * may have to a previous instance of me. */
1124                 lnet_wait_known_routerstate();
1125         }
1126 }
1127
1128 /*
1129  * This function is called from the monitor thread to check if there are
1130  * any active routers that need to be checked.
1131  */
1132 bool lnet_router_checker_active(void)
1133 {
1134         /* Router Checker thread needs to run when routing is enabled in
1135          * order to call lnet_update_ni_status_locked() */
1136         if (the_lnet.ln_routing)
1137                 return true;
1138
1139         return !list_empty(&the_lnet.ln_routers) &&
1140                 alive_router_check_interval > 0;
1141 }
1142
1143 void
1144 lnet_check_routers(void)
1145 {
1146         struct lnet_peer_net *first_lpn;
1147         struct lnet_peer_net *lpn;
1148         struct lnet_peer_ni *lpni;
1149         struct lnet_peer *rtr;
1150         bool push = false;
1151         bool needs_ping;
1152         bool found_lpn;
1153         __u64 version;
1154         __u32 net_id;
1155         time64_t now;
1156         int cpt;
1157         int rc;
1158
1159         cpt = lnet_net_lock_current();
1160 rescan:
1161         version = the_lnet.ln_routers_version;
1162
1163         list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1164                 /* If we're currently discovering the peer then don't
1165                  * issue another discovery
1166                  */
1167                 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY)
1168                         continue;
1169
1170                 now = ktime_get_real_seconds();
1171
1172                 /* find the next local peer net which needs to be ping'd */
1173                 needs_ping = false;
1174                 first_lpn = NULL;
1175                 found_lpn = false;
1176                 net_id = rtr->lp_disc_net_id;
1177                 do {
1178                         lpn = lnet_get_next_peer_net_locked(rtr, net_id);
1179                         if (!lpn) {
1180                                 CERROR("gateway %s has no networks\n",
1181                                 libcfs_nidstr(&rtr->lp_primary_nid));
1182                                 break;
1183                         }
1184
1185                         /* We looped back to the first peer net */
1186                         if (first_lpn == lpn)
1187                                 break;
1188                         if (!first_lpn)
1189                                 first_lpn = lpn;
1190
1191                         net_id = lpn->lpn_net_id;
1192                         if (!lnet_islocalnet_locked(net_id))
1193                                 continue;
1194
1195                         found_lpn = true;
1196
1197                         CDEBUG(D_NET, "rtr %s(%p) %s(%p) next ping %lld\n",
1198                                libcfs_nidstr(&rtr->lp_primary_nid), rtr,
1199                                libcfs_net2str(net_id), lpn,
1200                                lpn->lpn_next_ping);
1201
1202                         needs_ping = now >= lpn->lpn_next_ping;
1203
1204                 } while (!needs_ping);
1205
1206                 if (!found_lpn || !lpn) {
1207                         CERROR("no local network found for gateway %s\n",
1208                                libcfs_nidstr(&rtr->lp_primary_nid));
1209                         continue;
1210                 }
1211
1212                 if (!needs_ping)
1213                         continue;
1214
1215                 spin_lock(&rtr->lp_lock);
1216                 /* make sure we fully discover the router */
1217                 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1218                 rtr->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH |
1219                         LNET_PEER_RTR_DISCOVERY;
1220                 spin_unlock(&rtr->lp_lock);
1221
1222                 /* find the peer_ni associated with the primary NID */
1223                 lpni = lnet_peer_ni_get_locked(rtr, &rtr->lp_primary_nid);
1224                 if (!lpni) {
1225                         CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
1226                                libcfs_nidstr(&rtr->lp_primary_nid));
1227                         continue;
1228                 }
1229                 lnet_peer_ni_addref_locked(lpni);
1230
1231                 /* specify the net to use */
1232                 rtr->lp_disc_net_id = lpn->lpn_net_id;
1233
1234                 /* discover the router */
1235                 CDEBUG(D_NET, "discover %s, cpt = %d\n",
1236                        libcfs_nidstr(&lpni->lpni_nid), cpt);
1237                 rc = lnet_discover_peer_locked(lpni, cpt, false);
1238
1239                 /* drop ref taken above */
1240                 lnet_peer_ni_decref_locked(lpni);
1241
1242                 if (!rc)
1243                         lpn->lpn_next_ping = now + alive_router_check_interval;
1244                 else
1245                         CERROR("Failed to discover router %s\n",
1246                                libcfs_nidstr(&rtr->lp_primary_nid));
1247
1248                 /* NB cpt lock was dropped in lnet_discover_peer_locked() */
1249                 if (version != the_lnet.ln_routers_version) {
1250                         /* the routers list has changed */
1251                         goto rescan;
1252                 }
1253         }
1254
1255         if (the_lnet.ln_routing)
1256                 push = lnet_update_ni_status_locked();
1257
1258         lnet_net_unlock(cpt);
1259
1260         /* if the status of the ni changed update the peers */
1261         if (push)
1262                 lnet_push_update_to_peers(1);
1263 }
1264
1265 void
1266 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1267 {
1268         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1269
1270         while (--npages >= 0)
1271                 __free_page(rb->rb_kiov[npages].bv_page);
1272
1273         LIBCFS_FREE(rb, sz);
1274 }
1275
1276 static struct lnet_rtrbuf *
1277 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1278 {
1279         int            npages = rbp->rbp_npages;
1280         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1281         struct page   *page;
1282         struct lnet_rtrbuf *rb;
1283         int            i;
1284
1285         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1286         if (rb == NULL)
1287                 return NULL;
1288
1289         rb->rb_pool = rbp;
1290
1291         for (i = 0; i < npages; i++) {
1292                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL |
1293                                           __GFP_ZERO | __GFP_NORETRY);
1294                 if (page == NULL) {
1295                         while (--i >= 0)
1296                                 __free_page(rb->rb_kiov[i].bv_page);
1297
1298                         LIBCFS_FREE(rb, sz);
1299                         return NULL;
1300                 }
1301
1302                 rb->rb_kiov[i].bv_len = PAGE_SIZE;
1303                 rb->rb_kiov[i].bv_offset = 0;
1304                 rb->rb_kiov[i].bv_page = page;
1305         }
1306
1307         return rb;
1308 }
1309
1310 static void
1311 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1312 {
1313         int npages = rbp->rbp_npages;
1314         struct lnet_rtrbuf *rb;
1315         LIST_HEAD(tmp);
1316
1317         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1318                 return;
1319
1320         lnet_net_lock(cpt);
1321         list_splice_init(&rbp->rbp_msgs, &tmp);
1322         lnet_drop_routed_msgs_locked(&tmp, cpt);
1323         list_splice_init(&rbp->rbp_bufs, &tmp);
1324         rbp->rbp_req_nbuffers = 0;
1325         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1326         rbp->rbp_mincredits = 0;
1327         lnet_net_unlock(cpt);
1328
1329         /* Free buffers on the free list. */
1330         while (!list_empty(&tmp)) {
1331                 rb = list_first_entry(&tmp, struct lnet_rtrbuf, rb_list);
1332                 list_del(&rb->rb_list);
1333                 lnet_destroy_rtrbuf(rb, npages);
1334         }
1335 }
1336
1337 static int
1338 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1339 {
1340         LIST_HEAD(rb_list);
1341         struct lnet_rtrbuf *rb;
1342         int             num_rb;
1343         int             num_buffers = 0;
1344         int             old_req_nbufs;
1345         int             npages = rbp->rbp_npages;
1346
1347         lnet_net_lock(cpt);
1348         /* If we are called for less buffers than already in the pool, we
1349          * just lower the req_nbuffers number and excess buffers will be
1350          * thrown away as they are returned to the free list.  Credits
1351          * then get adjusted as well.
1352          * If we already have enough buffers allocated to serve the
1353          * increase requested, then we can treat that the same way as we
1354          * do the decrease. */
1355         num_rb = nbufs - rbp->rbp_nbuffers;
1356         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1357                 rbp->rbp_req_nbuffers = nbufs;
1358                 lnet_net_unlock(cpt);
1359                 return 0;
1360         }
1361         /* store the older value of rbp_req_nbuffers and then set it to
1362          * the new request to prevent lnet_return_rx_credits_locked() from
1363          * freeing buffers that we need to keep around */
1364         old_req_nbufs = rbp->rbp_req_nbuffers;
1365         rbp->rbp_req_nbuffers = nbufs;
1366         lnet_net_unlock(cpt);
1367
1368         /* allocate the buffers on a local list first.  If all buffers are
1369          * allocated successfully then join this list to the rbp buffer
1370          * list.  If not then free all allocated buffers. */
1371         while (num_rb-- > 0) {
1372                 rb = lnet_new_rtrbuf(rbp, cpt);
1373                 if (rb == NULL) {
1374                         CERROR("lnet: error allocating %ux%u page router buffers on CPT %u: rc = %d\n",
1375                                nbufs, npages, cpt, -ENOMEM);
1376
1377                         lnet_net_lock(cpt);
1378                         rbp->rbp_req_nbuffers = old_req_nbufs;
1379                         lnet_net_unlock(cpt);
1380
1381                         goto failed;
1382                 }
1383
1384                 list_add(&rb->rb_list, &rb_list);
1385                 num_buffers++;
1386         }
1387
1388         lnet_net_lock(cpt);
1389
1390         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1391         rbp->rbp_nbuffers += num_buffers;
1392         rbp->rbp_credits += num_buffers;
1393         rbp->rbp_mincredits = rbp->rbp_credits;
1394         /* We need to schedule blocked msg using the newly
1395          * added buffers. */
1396         while (!list_empty(&rbp->rbp_bufs) &&
1397                !list_empty(&rbp->rbp_msgs))
1398                 lnet_schedule_blocked_locked(rbp);
1399
1400         lnet_net_unlock(cpt);
1401
1402         return 0;
1403
1404 failed:
1405         while ((rb = list_first_entry_or_null(&rb_list,
1406                                               struct lnet_rtrbuf,
1407                                               rb_list)) != NULL) {
1408                 list_del(&rb->rb_list);
1409                 lnet_destroy_rtrbuf(rb, npages);
1410         }
1411
1412         return -ENOMEM;
1413 }
1414
1415 static void
1416 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1417 {
1418         INIT_LIST_HEAD(&rbp->rbp_msgs);
1419         INIT_LIST_HEAD(&rbp->rbp_bufs);
1420
1421         rbp->rbp_npages = npages;
1422         rbp->rbp_credits = 0;
1423         rbp->rbp_mincredits = 0;
1424 }
1425
1426 void
1427 lnet_rtrpools_free(int keep_pools)
1428 {
1429         struct lnet_rtrbufpool *rtrp;
1430         int               i;
1431
1432         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1433                 return;
1434
1435         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1436                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1437                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1438                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1439         }
1440
1441         if (!keep_pools) {
1442                 cfs_percpt_free(the_lnet.ln_rtrpools);
1443                 the_lnet.ln_rtrpools = NULL;
1444         }
1445 }
1446
1447 static int
1448 lnet_nrb_tiny_calculate(void)
1449 {
1450         int     nrbs = LNET_NRB_TINY;
1451
1452         if (tiny_router_buffers < 0) {
1453                 LCONSOLE_ERROR_MSG(0x10c,
1454                                    "tiny_router_buffers=%d invalid when "
1455                                    "routing enabled\n", tiny_router_buffers);
1456                 return -EINVAL;
1457         }
1458
1459         if (tiny_router_buffers > 0)
1460                 nrbs = tiny_router_buffers;
1461
1462         nrbs /= LNET_CPT_NUMBER;
1463         return max(nrbs, LNET_NRB_TINY_MIN);
1464 }
1465
1466 static int
1467 lnet_nrb_small_calculate(void)
1468 {
1469         int     nrbs = LNET_NRB_SMALL;
1470
1471         if (small_router_buffers < 0) {
1472                 LCONSOLE_ERROR_MSG(0x10c,
1473                                    "small_router_buffers=%d invalid when "
1474                                    "routing enabled\n", small_router_buffers);
1475                 return -EINVAL;
1476         }
1477
1478         if (small_router_buffers > 0)
1479                 nrbs = small_router_buffers;
1480
1481         nrbs /= LNET_CPT_NUMBER;
1482         return max(nrbs, LNET_NRB_SMALL_MIN);
1483 }
1484
1485 static int
1486 lnet_nrb_large_calculate(void)
1487 {
1488         int     nrbs = LNET_NRB_LARGE;
1489
1490         if (large_router_buffers < 0) {
1491                 LCONSOLE_ERROR_MSG(0x10c,
1492                                    "large_router_buffers=%d invalid when "
1493                                    "routing enabled\n", large_router_buffers);
1494                 return -EINVAL;
1495         }
1496
1497         if (large_router_buffers > 0)
1498                 nrbs = large_router_buffers;
1499
1500         nrbs /= LNET_CPT_NUMBER;
1501         return max(nrbs, LNET_NRB_LARGE_MIN);
1502 }
1503
1504 int
1505 lnet_rtrpools_alloc(int im_a_router)
1506 {
1507         struct lnet_rtrbufpool *rtrp;
1508         int     nrb_tiny;
1509         int     nrb_small;
1510         int     nrb_large;
1511         int     rc;
1512         int     i;
1513
1514         if (!strcmp(forwarding, "")) {
1515                 /* not set either way */
1516                 if (!im_a_router)
1517                         return 0;
1518         } else if (!strcmp(forwarding, "disabled")) {
1519                 /* explicitly disabled */
1520                 return 0;
1521         } else if (!strcmp(forwarding, "enabled")) {
1522                 /* explicitly enabled */
1523         } else {
1524                 rc = -EINVAL;
1525                 LCONSOLE_ERROR_MSG(0x10b,
1526                                    "lnet: forwarding='%s' not set to either 'enabled' or 'disabled': rc = %d\n",
1527                                    forwarding, rc);
1528                 return rc;
1529         }
1530
1531         nrb_tiny = lnet_nrb_tiny_calculate();
1532         if (nrb_tiny < 0)
1533                 return -EINVAL;
1534
1535         nrb_small = lnet_nrb_small_calculate();
1536         if (nrb_small < 0)
1537                 return -EINVAL;
1538
1539         nrb_large = lnet_nrb_large_calculate();
1540         if (nrb_large < 0)
1541                 return -EINVAL;
1542
1543         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1544                                                 LNET_NRBPOOLS *
1545                                                 sizeof(struct lnet_rtrbufpool));
1546         if (the_lnet.ln_rtrpools == NULL) {
1547                 rc = -ENOMEM;
1548                 LCONSOLE_ERROR_MSG(0x10c,
1549                         "lnet: error allocating router buffer pool: rc = %d\n",
1550                         rc);
1551                 return rc;
1552         }
1553
1554         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1555                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1556                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1557                                               nrb_tiny, i);
1558                 if (rc)
1559                         goto failed;
1560
1561                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1562                                   LNET_NRB_SMALL_PAGES);
1563                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1564                                               nrb_small, i);
1565                 if (rc)
1566                         goto failed;
1567
1568                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1569                                   LNET_NRB_LARGE_PAGES);
1570                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1571                                               nrb_large, i);
1572                 if (rc)
1573                         goto failed;
1574         }
1575
1576         lnet_net_lock(LNET_LOCK_EX);
1577         the_lnet.ln_routing = 1;
1578         lnet_net_unlock(LNET_LOCK_EX);
1579         complete(&the_lnet.ln_mt_wait_complete);
1580         return 0;
1581
1582  failed:
1583         lnet_rtrpools_free(0);
1584         return rc;
1585 }
1586
1587 static int
1588 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1589 {
1590         int nrb = 0;
1591         int rc = 0;
1592         int i;
1593         struct lnet_rtrbufpool *rtrp;
1594
1595         /* If the provided values for each buffer pool are different than the
1596          * configured values, we need to take action. */
1597         if (tiny >= 0) {
1598                 tiny_router_buffers = tiny;
1599                 nrb = lnet_nrb_tiny_calculate();
1600                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1601                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1602                                                       nrb, i);
1603                         if (rc != 0)
1604                                 return rc;
1605                 }
1606         }
1607         if (small >= 0) {
1608                 small_router_buffers = small;
1609                 nrb = lnet_nrb_small_calculate();
1610                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1611                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1612                                                       nrb, i);
1613                         if (rc != 0)
1614                                 return rc;
1615                 }
1616         }
1617         if (large >= 0) {
1618                 large_router_buffers = large;
1619                 nrb = lnet_nrb_large_calculate();
1620                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1621                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1622                                                       nrb, i);
1623                         if (rc != 0)
1624                                 return rc;
1625                 }
1626         }
1627
1628         return 0;
1629 }
1630
1631 int
1632 lnet_rtrpools_adjust(int tiny, int small, int large)
1633 {
1634         /* this function doesn't revert the changes if adding new buffers
1635          * failed.  It's up to the user space caller to revert the
1636          * changes. */
1637
1638         if (!the_lnet.ln_routing)
1639                 return 0;
1640
1641         return lnet_rtrpools_adjust_helper(tiny, small, large);
1642 }
1643
1644 int
1645 lnet_rtrpools_enable(void)
1646 {
1647         int rc = 0;
1648
1649         if (the_lnet.ln_routing)
1650                 return 0;
1651
1652         if (the_lnet.ln_rtrpools == NULL)
1653                 /* If routing is turned off, and we have never
1654                  * initialized the pools before, just call the
1655                  * standard buffer pool allocation routine as
1656                  * if we are just configuring this for the first
1657                  * time. */
1658                 rc = lnet_rtrpools_alloc(1);
1659         else
1660                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1661         if (rc != 0)
1662                 return rc;
1663
1664         lnet_net_lock(LNET_LOCK_EX);
1665         the_lnet.ln_routing = 1;
1666
1667         the_lnet.ln_ping_target->pb_info.pi_features &=
1668                 ~LNET_PING_FEAT_RTE_DISABLED;
1669         lnet_net_unlock(LNET_LOCK_EX);
1670
1671         if (lnet_peer_discovery_disabled)
1672                 CWARN("Consider turning discovery on to enable full "
1673                       "Multi-Rail routing functionality\n");
1674
1675         return rc;
1676 }
1677
1678 void
1679 lnet_rtrpools_disable(void)
1680 {
1681         if (!the_lnet.ln_routing)
1682                 return;
1683
1684         lnet_net_lock(LNET_LOCK_EX);
1685         the_lnet.ln_routing = 0;
1686         the_lnet.ln_ping_target->pb_info.pi_features |=
1687                 LNET_PING_FEAT_RTE_DISABLED;
1688
1689         tiny_router_buffers = 0;
1690         small_router_buffers = 0;
1691         large_router_buffers = 0;
1692         lnet_net_unlock(LNET_LOCK_EX);
1693         lnet_rtrpools_free(1);
1694 }
1695
1696 static inline void
1697 lnet_notify_peer_down(struct lnet_ni *ni, struct lnet_nid *nid)
1698 {
1699         if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1700                 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1701 }
1702
1703 /*
1704  * ni: local NI used to communicate with the peer
1705  * nid: peer NID
1706  * alive: true if peer is alive, false otherwise
1707  * reset: reset health value. This is requested by the LND.
1708  * when: notificaiton time.
1709  */
1710 int
1711 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid4, bool alive, bool reset,
1712             time64_t when)
1713 {
1714         struct lnet_peer_ni *lpni = NULL;
1715         struct lnet_route *route;
1716         struct lnet_peer *lp;
1717         time64_t now = ktime_get_seconds();
1718         struct lnet_nid nid;
1719         int cpt;
1720
1721         lnet_nid4_to_nid(nid4, &nid);
1722         LASSERT(!in_interrupt());
1723
1724         CDEBUG(D_NET, "%s notifying %s: %s\n",
1725                (ni == NULL) ? "userspace" : libcfs_nidstr(&ni->ni_nid),
1726                libcfs_nidstr(&nid), alive ? "up" : "down");
1727
1728         if (ni != NULL &&
1729             LNET_NID_NET(&ni->ni_nid) != LNET_NID_NET(&nid)) {
1730                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1731                       libcfs_nidstr(&nid), alive ? "birth" : "death",
1732                       libcfs_nidstr(&ni->ni_nid));
1733                 return -EINVAL;
1734         }
1735
1736         /* can't do predictions... */
1737         if (when > now) {
1738                 CWARN("Ignoring prediction from %s of %s %s %lld seconds in the future\n",
1739                         ni ? libcfs_nidstr(&ni->ni_nid) :  "userspace",
1740                         libcfs_nidstr(&nid), alive ? "up" : "down", when - now);
1741                 return -EINVAL;
1742         }
1743
1744         if (ni != NULL && !alive &&             /* LND telling me she's down */
1745             !auto_down) {                       /* auto-down disabled */
1746                 CDEBUG(D_NET, "Auto-down disabled\n");
1747                 return 0;
1748         }
1749
1750         /* must lock 0 since this is used for synchronization */
1751         lnet_net_lock(0);
1752
1753         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1754                 lnet_net_unlock(0);
1755                 return -ESHUTDOWN;
1756         }
1757
1758         lpni = lnet_peer_ni_find_locked(&nid);
1759         if (lpni == NULL) {
1760                 /* nid not found */
1761                 lnet_net_unlock(0);
1762                 CDEBUG(D_NET, "%s not found\n", libcfs_nidstr(&nid));
1763                 return 0;
1764         }
1765
1766         if (alive) {
1767                 if (reset) {
1768                         lpni->lpni_ns_status = LNET_NI_STATUS_UP;
1769                         lnet_set_lpni_healthv_locked(lpni,
1770                                                      LNET_MAX_HEALTH_VALUE);
1771                 } else {
1772                         __u32 sensitivity = lpni->lpni_peer_net->
1773                                         lpn_peer->lp_health_sensitivity;
1774
1775                         lnet_inc_lpni_healthv_locked(lpni,
1776                                         (sensitivity) ? sensitivity :
1777                                         lnet_health_sensitivity);
1778                 }
1779         } else if (reset) {
1780                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
1781         }
1782
1783         /* recalculate aliveness */
1784         alive = lnet_is_peer_ni_alive(lpni);
1785
1786         lp = lpni->lpni_peer_net->lpn_peer;
1787         /* If this is an LNet router then update route aliveness */
1788         if (lp->lp_rtr_refcount) {
1789                 if (reset)
1790                         /* reset flag indicates gateway peer went up or down */
1791                         lp->lp_alive = alive;
1792
1793                 /* If discovery is disabled, locally or on the gateway, then
1794                  * any routes using lpni as next-hop need to be updated
1795                  *
1796                  * NB: We can get many notifications while a route is down, so
1797                  * we try and avoid the expensive net_lock/EX here for the
1798                  * common case of receiving duplicate lnet_notify() calls (i.e.
1799                  * only grab EX lock when we actually need to update the route
1800                  * aliveness).
1801                  */
1802                 if (lnet_is_discovery_disabled(lp)) {
1803                         list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
1804                                 if (nid_same(&route->lr_nid, &lpni->lpni_nid))
1805                                         lnet_set_route_aliveness(route, alive);
1806                         }
1807                 }
1808         }
1809
1810         lnet_net_unlock(0);
1811
1812         if (ni != NULL && !alive)
1813                 lnet_notify_peer_down(ni, &lpni->lpni_nid);
1814
1815         cpt = lpni->lpni_cpt;
1816         lnet_net_lock(cpt);
1817         lnet_peer_ni_decref_locked(lpni);
1818         lnet_net_unlock(cpt);
1819
1820         return 0;
1821 }
1822 EXPORT_SYMBOL(lnet_notify);