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                 CWARN("route %s->%s is detected to be multi-hop but hop count is set to %d\n",
 387                         libcfs_net2str(route->lr_net),
 388                         libcfs_nidstr(&route->lr_gateway->lp_primary_nid),
 389                         (int) route->lr_hops);
 390         }
 391 }
 392
 393 static void
 394 lnet_set_route_hop_type(struct lnet_peer *gw, struct lnet_route *route)
 395 {
 396         struct lnet_peer_net *lpn;
 397         bool single_hop = false;
 398
 399         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 400                 if (route->lr_net == lpn->lpn_net_id) {
 401                         single_hop = true;
 402                         break;
 403                 }
 404         }
 405         route->lr_single_hop = single_hop;
 406         lnet_check_route_inconsistency(route);
 407 }
 408
 409 /* Must hold net_lock/EX */
 410 void
 411 lnet_router_discovery_ping_reply(struct lnet_peer *lp)
 412 {
 413         struct lnet_ping_buffer *pbuf = lp->lp_data;
 414         struct lnet_peer_net *llpn;
 415         struct lnet_route *route;
 416         bool single_hop = false;
 417         bool net_up = false;
 418         unsigned lp_state;
 419         __u32 net;
 420         int i;
 421
 422
 423         spin_lock(&lp->lp_lock);
 424         lp_state = lp->lp_state;
 425
 426         /* only handle replies if discovery is disabled. */
 427         if (!lnet_is_discovery_disabled_locked(lp)) {
 428                 spin_unlock(&lp->lp_lock);
 429                 return;
 430         }
 431
 432         spin_unlock(&lp->lp_lock);
 433
 434         if (lp_state & LNET_PEER_PING_FAILED ||
 435             pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED) {
 436                 CDEBUG(D_NET, "Set routes down for gw %s because %s %d\n",
 437                        libcfs_nidstr(&lp->lp_primary_nid),
 438                        lp_state & LNET_PEER_PING_FAILED ? "ping failed" :
 439                        "route feature is disabled", lp->lp_ping_error);
 440                 /* If the ping failed or the peer has routing disabled then
 441                  * mark the routes served by this peer down
 442                  */
 443                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 444                         lnet_set_route_aliveness(route, false);
 445                 return;
 446         }
 447
 448         CDEBUG(D_NET, "Discovery is disabled. Processing reply for gw: %s:%d\n",
 449                libcfs_nidstr(&lp->lp_primary_nid), pbuf->pb_info.pi_nnis);
 450
 451         /*
 452          * examine the ping response to determine if the routes on that
 453          * gateway should be declared alive.
 454          * The route is alive if:
 455          *  1. local network to reach the route is alive and
 456          *  2. route is single hop, avoid_async_router_failure is set and
 457          *     there exists at least one NI on the route's remote net
 458          */
 459         list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
 460                 llpn = lnet_peer_get_net_locked(lp, route->lr_lnet);
 461                 if (!llpn) {
 462                         lnet_set_route_aliveness(route, false);
 463                         continue;
 464                 }
 465
 466                 if (!lnet_is_gateway_net_alive(llpn)) {
 467                         lnet_set_route_aliveness(route, false);
 468                         continue;
 469                 }
 470
 471                 single_hop = net_up = false;
 472                 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
 473                         net = LNET_NIDNET(pbuf->pb_info.pi_ni[i].ns_nid);
 474
 475                         if (route->lr_net == net) {
 476                                 single_hop = true;
 477                                 if (pbuf->pb_info.pi_ni[i].ns_status ==
 478                                     LNET_NI_STATUS_UP) {
 479                                         net_up = true;
 480                                         break;
 481                                 }
 482                         }
 483                 }
 484
 485                 route->lr_single_hop = single_hop;
 486                 if (avoid_asym_router_failure &&
 487                     (route->lr_hops == 1 || route->lr_single_hop))
 488                         lnet_set_route_aliveness(route, net_up);
 489                 else
 490                         lnet_set_route_aliveness(route, true);
 491
 492                 /*
 493                  * warn that the route is configured as single-hop but it
 494                  * really is multi-hop as far as we can tell.
 495                  */
 496                 lnet_check_route_inconsistency(route);
 497         }
 498 }
 499
 500 void
 501 lnet_router_discovery_complete(struct lnet_peer *lp)
 502 {
 503         struct lnet_peer_ni *lpni = NULL;
 504         struct lnet_route *route;
 505
 506         spin_lock(&lp->lp_lock);
 507         lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
 508         lp->lp_state |= LNET_PEER_RTR_DISCOVERED;
 509         lp->lp_alive = lp->lp_dc_error == 0;
 510         spin_unlock(&lp->lp_lock);
 511
 512         if (!lp->lp_dc_error) {
 513                 /* ping replies are being handled when discovery is disabled */
 514                 if (lnet_is_discovery_disabled_locked(lp))
 515                         return;
 516
 517                 /*
 518                 * mark single-hop routes.  If the remote net is not configured on
 519                 * the gateway we assume this is intentional and we mark the
 520                 * gateway as multi-hop
 521                 */
 522                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
 523                         lnet_set_route_aliveness(route, true);
 524                         lnet_set_route_hop_type(lp, route);
 525                 }
 526
 527                 return;
 528         }
 529
 530         /*
 531          * We do not send messages directly to the remote interfaces
 532          * of an LNet router. As such, we rely on the PING response
 533          * to determine the up/down status of these interfaces. If
 534          * a PING response is not receieved, or some other problem with
 535          * discovery occurs that prevents us from getting this status,
 536          * we assume all interfaces are down until we're able to
 537          * determine otherwise.
 538          */
 539         CDEBUG(D_NET, "%s: Router discovery failed %d\n",
 540                libcfs_nidstr(&lp->lp_primary_nid), lp->lp_dc_error);
 541         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 542                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 543
 544         list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 545                 lnet_set_route_aliveness(route, false);
 546 }
 547
 548 static void
 549 lnet_rtr_addref_locked(struct lnet_peer *lp)
 550 {
 551         LASSERT(lp->lp_rtr_refcount >= 0);
 552
 553         /* lnet_net_lock must be exclusively locked */
 554         lp->lp_rtr_refcount++;
 555         if (lp->lp_rtr_refcount == 1) {
 556                 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
 557                 /* addref for the_lnet.ln_routers */
 558                 lnet_peer_addref_locked(lp);
 559                 the_lnet.ln_routers_version++;
 560         }
 561 }
 562
 563 static void
 564 lnet_rtr_decref_locked(struct lnet_peer *lp)
 565 {
 566         LASSERT(atomic_read(&lp->lp_refcount) > 0);
 567         LASSERT(lp->lp_rtr_refcount > 0);
 568
 569         /* lnet_net_lock must be exclusively locked */
 570         lp->lp_rtr_refcount--;
 571         if (lp->lp_rtr_refcount == 0) {
 572                 LASSERT(list_empty(&lp->lp_routes));
 573
 574                 list_del(&lp->lp_rtr_list);
 575                 /* decref for the_lnet.ln_routers */
 576                 lnet_peer_decref_locked(lp);
 577                 the_lnet.ln_routers_version++;
 578         }
 579 }
 580
 581 struct lnet_remotenet *
 582 lnet_find_rnet_locked(__u32 net)
 583 {
 584         struct lnet_remotenet *rnet;
 585         struct list_head *tmp;
 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(tmp, rn_list) {
 592                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 593
 594                 if (rnet->lrn_net == net)
 595                         return rnet;
 596         }
 597         return NULL;
 598 }
 599
 600 static void lnet_shuffle_seed(void)
 601 {
 602         static int seeded;
 603         struct lnet_ni *ni = NULL;
 604
 605         if (seeded)
 606                 return;
 607
 608         /* Nodes with small feet have little entropy
 609          * the NID for this node gives the most entropy in the low bits */
 610         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 611                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 612
 613         seeded = 1;
 614 }
 615
 616 /* NB expects LNET_LOCK held */
 617 static void
 618 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 619 {
 620         struct lnet_peer_net *lpn;
 621         unsigned int offset = 0;
 622         unsigned int len = 0;
 623         struct list_head *e;
 624         time64_t now;
 625
 626         lnet_shuffle_seed();
 627
 628         list_for_each(e, &rnet->lrn_routes)
 629                 len++;
 630
 631         /*
 632          * Randomly adding routes to the list is done to ensure that when
 633          * different nodes are using the same list of routers, they end up
 634          * preferring different routers.
 635          */
 636         offset = prandom_u32_max(len + 1);
 637         list_for_each(e, &rnet->lrn_routes) {
 638                 if (offset == 0)
 639                         break;
 640                 offset--;
 641         }
 642         list_add(&route->lr_list, e);
 643         /*
 644          * force a router check on the gateway to make sure the route is
 645          * alive
 646          */
 647         now = ktime_get_real_seconds();
 648         list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets,
 649                             lpn_peer_nets) {
 650                 lpn->lpn_next_ping = now;
 651         }
 652
 653         the_lnet.ln_remote_nets_version++;
 654
 655         /* add the route on the gateway list */
 656         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 657
 658         /* take a router reference count on the gateway */
 659         lnet_rtr_addref_locked(route->lr_gateway);
 660 }
 661
 662 int
 663 lnet_add_route(__u32 net, __u32 hops, struct lnet_nid *gateway,
 664                __u32 priority, __u32 sensitivity)
 665 {
 666         struct list_head *route_entry;
 667         struct lnet_remotenet *rnet;
 668         struct lnet_remotenet *rnet2;
 669         struct lnet_route *route;
 670         struct lnet_peer_ni *lpni;
 671         struct lnet_peer *gw;
 672         int add_route;
 673         int rc;
 674
 675         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 676                libcfs_net2str(net), hops, priority, libcfs_nidstr(gateway));
 677
 678         if (LNET_NID_IS_ANY(gateway) ||
 679             nid_is_lo0(gateway) ||
 680             net == LNET_NET_ANY ||
 681             LNET_NETTYP(net) == LOLND ||
 682             LNET_NID_NET(gateway) == net ||
 683             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 684                 return -EINVAL;
 685
 686         /* it's a local network */
 687         if (lnet_islocalnet(net))
 688                 return -EEXIST;
 689
 690         if (!lnet_islocalnet(LNET_NID_NET(gateway))) {
 691                 CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n",
 692                        libcfs_nidstr(gateway),
 693                        libcfs_net2str(LNET_NID_NET(gateway)));
 694                 return -EHOSTUNREACH;
 695         }
 696
 697         /* Assume net, route, all new */
 698         LIBCFS_ALLOC(route, sizeof(*route));
 699         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 700         if (route == NULL || rnet == NULL) {
 701                 CERROR("Out of memory creating route %s %d %s\n",
 702                        libcfs_net2str(net), hops, libcfs_nidstr(gateway));
 703                 if (route != NULL)
 704                         LIBCFS_FREE(route, sizeof(*route));
 705                 if (rnet != NULL)
 706                         LIBCFS_FREE(rnet, sizeof(*rnet));
 707                 return -ENOMEM;
 708         }
 709
 710         INIT_LIST_HEAD(&rnet->lrn_routes);
 711         rnet->lrn_net = net;
 712         /* store the local and remote net that the route represents */
 713         route->lr_lnet = LNET_NID_NET(gateway);
 714         route->lr_net = net;
 715         route->lr_nid = *gateway;
 716         route->lr_priority = priority;
 717         route->lr_hops = hops;
 718         if (lnet_peers_start_down())
 719                 atomic_set(&route->lr_alive, 0);
 720         else
 721                 atomic_set(&route->lr_alive, 1);
 722
 723         lnet_net_lock(LNET_LOCK_EX);
 724
 725         /*
 726          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 727          * lose that once we're done
 728          */
 729         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 730         if (IS_ERR(lpni)) {
 731                 lnet_net_unlock(LNET_LOCK_EX);
 732
 733                 LIBCFS_FREE(route, sizeof(*route));
 734                 LIBCFS_FREE(rnet, sizeof(*rnet));
 735
 736                 rc = PTR_ERR(lpni);
 737                 CERROR("Error %d creating route %s %d %s\n", rc,
 738                         libcfs_net2str(net), hops,
 739                         libcfs_nidstr(gateway));
 740                 return rc;
 741         }
 742
 743         LASSERT(lpni->lpni_peer_net && 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         struct list_head *e1;
 981         struct list_head *e2;
 982         int cpt;
 983         int i;
 984
 985         cpt = lnet_net_lock_current();
 986
 987         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 988                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 989                 list_for_each(e1, rn_list) {
 990                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 991
 992                         list_for_each(e2, &rnet->lrn_routes) {
 993                                 route = list_entry(e2, struct lnet_route,
 994                                                    lr_list);
 995
 996                                 if (idx-- == 0) {
 997                                         *net      = rnet->lrn_net;
 998                                         *gateway  = lnet_nid_to_nid4(&route->lr_nid);
 999                                         *hops     = route->lr_hops;
1000                                         *priority = route->lr_priority;
1001                                         *sensitivity = route->lr_gateway->
1002                                                 lp_health_sensitivity;
1003                                         if (lnet_is_route_alive(route))
1004                                                 *flags |= LNET_RT_ALIVE;
1005                                         else
1006                                                 *flags &= ~LNET_RT_ALIVE;
1007                                         if (route->lr_single_hop)
1008                                                 *flags &= ~LNET_RT_MULTI_HOP;
1009                                         else
1010                                                 *flags |= LNET_RT_MULTI_HOP;
1011                                         lnet_net_unlock(cpt);
1012                                         return 0;
1013                                 }
1014                         }
1015                 }
1016         }
1017
1018         lnet_net_unlock(cpt);
1019         return -ENOENT;
1020 }
1021
1022 static void
1023 lnet_wait_known_routerstate(void)
1024 {
1025         struct lnet_peer *rtr;
1026         struct list_head *entry;
1027         int all_known;
1028
1029         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
1030
1031         for (;;) {
1032                 int cpt = lnet_net_lock_current();
1033
1034                 all_known = 1;
1035                 list_for_each(entry, &the_lnet.ln_routers) {
1036                         rtr = list_entry(entry, struct lnet_peer,
1037                                          lp_rtr_list);
1038
1039                         spin_lock(&rtr->lp_lock);
1040
1041                         if ((rtr->lp_state & LNET_PEER_RTR_DISCOVERED) == 0) {
1042                                 all_known = 0;
1043                                 spin_unlock(&rtr->lp_lock);
1044                                 break;
1045                         }
1046                         spin_unlock(&rtr->lp_lock);
1047                 }
1048
1049                 lnet_net_unlock(cpt);
1050
1051                 if (all_known)
1052                         return;
1053
1054                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
1055         }
1056 }
1057
1058 static inline bool
1059 lnet_net_set_status_locked(struct lnet_net *net, __u32 status)
1060 {
1061         struct lnet_ni *ni;
1062         bool update = false;
1063
1064         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1065                 if (lnet_ni_set_status(ni, status))
1066                         update = true;
1067
1068         return update;
1069 }
1070
1071 static bool
1072 lnet_update_ni_status_locked(void)
1073 {
1074         struct lnet_net *net;
1075         struct lnet_ni *ni;
1076         bool push = false;
1077         time64_t now;
1078         time64_t timeout;
1079
1080         LASSERT(the_lnet.ln_routing);
1081
1082         timeout = router_ping_timeout + alive_router_check_interval;
1083
1084         now = ktime_get_seconds();
1085         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1086                 if (net->net_lnd->lnd_type == LOLND)
1087                         continue;
1088
1089                 if (now < net->net_last_alive + timeout)
1090                         goto check_ni_fatal;
1091
1092                 spin_lock(&net->net_lock);
1093                 /* re-check with lock */
1094                 if (now < net->net_last_alive + timeout) {
1095                         spin_unlock(&net->net_lock);
1096                         goto check_ni_fatal;
1097                 }
1098                 spin_unlock(&net->net_lock);
1099
1100                 /*
1101                  * if the net didn't receive any traffic for past the
1102                  * timeout on any of its constituent NIs, then mark all
1103                  * the NIs down.
1104                  */
1105                 if (lnet_net_set_status_locked(net, LNET_NI_STATUS_DOWN)) {
1106                         push = true;
1107                         continue;
1108                 }
1109
1110 check_ni_fatal:
1111                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1112                         /* lnet_ni_set_status() will perform the same check of
1113                          * ni_status while holding the ni lock. We can safely
1114                          * check ni_status without that lock because it is only
1115                          * written to under net_lock/EX and our caller is
1116                          * holding a net lock.
1117                          */
1118                         if (atomic_read(&ni->ni_fatal_error_on) &&
1119                             ni->ni_status &&
1120                             ni->ni_status->ns_status != LNET_NI_STATUS_DOWN &&
1121                             lnet_ni_set_status(ni, LNET_NI_STATUS_DOWN))
1122                                 push = true;
1123                 }
1124         }
1125
1126         return push;
1127 }
1128
1129 void lnet_wait_router_start(void)
1130 {
1131         if (check_routers_before_use) {
1132                 /* Note that a helpful side-effect of pinging all known routers
1133                  * at startup is that it makes them drop stale connections they
1134                  * may have to a previous instance of me. */
1135                 lnet_wait_known_routerstate();
1136         }
1137 }
1138
1139 /*
1140  * This function is called from the monitor thread to check if there are
1141  * any active routers that need to be checked.
1142  */
1143 bool lnet_router_checker_active(void)
1144 {
1145         /* Router Checker thread needs to run when routing is enabled in
1146          * order to call lnet_update_ni_status_locked() */
1147         if (the_lnet.ln_routing)
1148                 return true;
1149
1150         return !list_empty(&the_lnet.ln_routers) &&
1151                 alive_router_check_interval > 0;
1152 }
1153
1154 void
1155 lnet_check_routers(void)
1156 {
1157         struct lnet_peer_net *first_lpn;
1158         struct lnet_peer_net *lpn;
1159         struct lnet_peer_ni *lpni;
1160         struct list_head *entry;
1161         struct lnet_peer *rtr;
1162         bool push = false;
1163         bool needs_ping;
1164         bool found_lpn;
1165         __u64 version;
1166         __u32 net_id;
1167         time64_t now;
1168         int cpt;
1169         int rc;
1170
1171         cpt = lnet_net_lock_current();
1172 rescan:
1173         version = the_lnet.ln_routers_version;
1174
1175         list_for_each(entry, &the_lnet.ln_routers) {
1176                 rtr = list_entry(entry, struct lnet_peer,
1177                                  lp_rtr_list);
1178
1179                 /* If we're currently discovering the peer then don't
1180                  * issue another discovery
1181                  */
1182                 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY)
1183                         continue;
1184
1185                 now = ktime_get_real_seconds();
1186
1187                 /* find the next local peer net which needs to be ping'd */
1188                 needs_ping = false;
1189                 first_lpn = NULL;
1190                 found_lpn = false;
1191                 net_id = rtr->lp_disc_net_id;
1192                 do {
1193                         lpn = lnet_get_next_peer_net_locked(rtr, net_id);
1194                         if (!lpn) {
1195                                 CERROR("gateway %s has no networks\n",
1196                                 libcfs_nidstr(&rtr->lp_primary_nid));
1197                                 break;
1198                         }
1199
1200                         /* We looped back to the first peer net */
1201                         if (first_lpn == lpn)
1202                                 break;
1203                         if (!first_lpn)
1204                                 first_lpn = lpn;
1205
1206                         net_id = lpn->lpn_net_id;
1207                         if (!lnet_islocalnet_locked(net_id))
1208                                 continue;
1209
1210                         found_lpn = true;
1211
1212                         CDEBUG(D_NET, "rtr %s(%p) %s(%p) next ping %lld\n",
1213                                libcfs_nidstr(&rtr->lp_primary_nid), rtr,
1214                                libcfs_net2str(net_id), lpn,
1215                                lpn->lpn_next_ping);
1216
1217                         needs_ping = now >= lpn->lpn_next_ping;
1218
1219                 } while (!needs_ping);
1220
1221                 if (!found_lpn || !lpn) {
1222                         CERROR("no local network found for gateway %s\n",
1223                                libcfs_nidstr(&rtr->lp_primary_nid));
1224                         continue;
1225                 }
1226
1227                 if (!needs_ping)
1228                         continue;
1229
1230                 spin_lock(&rtr->lp_lock);
1231                 /* make sure we fully discover the router */
1232                 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1233                 rtr->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH |
1234                         LNET_PEER_RTR_DISCOVERY;
1235                 spin_unlock(&rtr->lp_lock);
1236
1237                 /* find the peer_ni associated with the primary NID */
1238                 lpni = lnet_peer_get_ni_locked(
1239                         rtr, lnet_nid_to_nid4(&rtr->lp_primary_nid));
1240                 if (!lpni) {
1241                         CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
1242                                libcfs_nidstr(&rtr->lp_primary_nid));
1243                         continue;
1244                 }
1245                 lnet_peer_ni_addref_locked(lpni);
1246
1247                 /* specify the net to use */
1248                 rtr->lp_disc_net_id = lpn->lpn_net_id;
1249
1250                 /* discover the router */
1251                 CDEBUG(D_NET, "discover %s, cpt = %d\n",
1252                        libcfs_nidstr(&lpni->lpni_nid), cpt);
1253                 rc = lnet_discover_peer_locked(lpni, cpt, false);
1254
1255                 /* drop ref taken above */
1256                 lnet_peer_ni_decref_locked(lpni);
1257
1258                 if (!rc)
1259                         lpn->lpn_next_ping = now + alive_router_check_interval;
1260                 else
1261                         CERROR("Failed to discover router %s\n",
1262                                libcfs_nidstr(&rtr->lp_primary_nid));
1263
1264                 /* NB cpt lock was dropped in lnet_discover_peer_locked() */
1265                 if (version != the_lnet.ln_routers_version) {
1266                         /* the routers list has changed */
1267                         goto rescan;
1268                 }
1269         }
1270
1271         if (the_lnet.ln_routing)
1272                 push = lnet_update_ni_status_locked();
1273
1274         lnet_net_unlock(cpt);
1275
1276         /* if the status of the ni changed update the peers */
1277         if (push)
1278                 lnet_push_update_to_peers(1);
1279 }
1280
1281 void
1282 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1283 {
1284         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1285
1286         while (--npages >= 0)
1287                 __free_page(rb->rb_kiov[npages].bv_page);
1288
1289         LIBCFS_FREE(rb, sz);
1290 }
1291
1292 static struct lnet_rtrbuf *
1293 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1294 {
1295         int            npages = rbp->rbp_npages;
1296         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1297         struct page   *page;
1298         struct lnet_rtrbuf *rb;
1299         int            i;
1300
1301         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1302         if (rb == NULL)
1303                 return NULL;
1304
1305         rb->rb_pool = rbp;
1306
1307         for (i = 0; i < npages; i++) {
1308                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL |
1309                                           __GFP_ZERO | __GFP_NORETRY);
1310                 if (page == NULL) {
1311                         while (--i >= 0)
1312                                 __free_page(rb->rb_kiov[i].bv_page);
1313
1314                         LIBCFS_FREE(rb, sz);
1315                         return NULL;
1316                 }
1317
1318                 rb->rb_kiov[i].bv_len = PAGE_SIZE;
1319                 rb->rb_kiov[i].bv_offset = 0;
1320                 rb->rb_kiov[i].bv_page = page;
1321         }
1322
1323         return rb;
1324 }
1325
1326 static void
1327 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1328 {
1329         int npages = rbp->rbp_npages;
1330         struct lnet_rtrbuf *rb;
1331         LIST_HEAD(tmp);
1332
1333         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1334                 return;
1335
1336         lnet_net_lock(cpt);
1337         list_splice_init(&rbp->rbp_msgs, &tmp);
1338         lnet_drop_routed_msgs_locked(&tmp, cpt);
1339         list_splice_init(&rbp->rbp_bufs, &tmp);
1340         rbp->rbp_req_nbuffers = 0;
1341         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1342         rbp->rbp_mincredits = 0;
1343         lnet_net_unlock(cpt);
1344
1345         /* Free buffers on the free list. */
1346         while (!list_empty(&tmp)) {
1347                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1348                 list_del(&rb->rb_list);
1349                 lnet_destroy_rtrbuf(rb, npages);
1350         }
1351 }
1352
1353 static int
1354 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1355 {
1356         LIST_HEAD(rb_list);
1357         struct lnet_rtrbuf *rb;
1358         int             num_rb;
1359         int             num_buffers = 0;
1360         int             old_req_nbufs;
1361         int             npages = rbp->rbp_npages;
1362
1363         lnet_net_lock(cpt);
1364         /* If we are called for less buffers than already in the pool, we
1365          * just lower the req_nbuffers number and excess buffers will be
1366          * thrown away as they are returned to the free list.  Credits
1367          * then get adjusted as well.
1368          * If we already have enough buffers allocated to serve the
1369          * increase requested, then we can treat that the same way as we
1370          * do the decrease. */
1371         num_rb = nbufs - rbp->rbp_nbuffers;
1372         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1373                 rbp->rbp_req_nbuffers = nbufs;
1374                 lnet_net_unlock(cpt);
1375                 return 0;
1376         }
1377         /* store the older value of rbp_req_nbuffers and then set it to
1378          * the new request to prevent lnet_return_rx_credits_locked() from
1379          * freeing buffers that we need to keep around */
1380         old_req_nbufs = rbp->rbp_req_nbuffers;
1381         rbp->rbp_req_nbuffers = nbufs;
1382         lnet_net_unlock(cpt);
1383
1384         /* allocate the buffers on a local list first.  If all buffers are
1385          * allocated successfully then join this list to the rbp buffer
1386          * list.  If not then free all allocated buffers. */
1387         while (num_rb-- > 0) {
1388                 rb = lnet_new_rtrbuf(rbp, cpt);
1389                 if (rb == NULL) {
1390                         CERROR("lnet: error allocating %ux%u page router buffers on CPT %u: rc = %d\n",
1391                                nbufs, npages, cpt, -ENOMEM);
1392
1393                         lnet_net_lock(cpt);
1394                         rbp->rbp_req_nbuffers = old_req_nbufs;
1395                         lnet_net_unlock(cpt);
1396
1397                         goto failed;
1398                 }
1399
1400                 list_add(&rb->rb_list, &rb_list);
1401                 num_buffers++;
1402         }
1403
1404         lnet_net_lock(cpt);
1405
1406         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1407         rbp->rbp_nbuffers += num_buffers;
1408         rbp->rbp_credits += num_buffers;
1409         rbp->rbp_mincredits = rbp->rbp_credits;
1410         /* We need to schedule blocked msg using the newly
1411          * added buffers. */
1412         while (!list_empty(&rbp->rbp_bufs) &&
1413                !list_empty(&rbp->rbp_msgs))
1414                 lnet_schedule_blocked_locked(rbp);
1415
1416         lnet_net_unlock(cpt);
1417
1418         return 0;
1419
1420 failed:
1421         while (!list_empty(&rb_list)) {
1422                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1423                 list_del(&rb->rb_list);
1424                 lnet_destroy_rtrbuf(rb, npages);
1425         }
1426
1427         return -ENOMEM;
1428 }
1429
1430 static void
1431 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1432 {
1433         INIT_LIST_HEAD(&rbp->rbp_msgs);
1434         INIT_LIST_HEAD(&rbp->rbp_bufs);
1435
1436         rbp->rbp_npages = npages;
1437         rbp->rbp_credits = 0;
1438         rbp->rbp_mincredits = 0;
1439 }
1440
1441 void
1442 lnet_rtrpools_free(int keep_pools)
1443 {
1444         struct lnet_rtrbufpool *rtrp;
1445         int               i;
1446
1447         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1448                 return;
1449
1450         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1451                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1452                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1453                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1454         }
1455
1456         if (!keep_pools) {
1457                 cfs_percpt_free(the_lnet.ln_rtrpools);
1458                 the_lnet.ln_rtrpools = NULL;
1459         }
1460 }
1461
1462 static int
1463 lnet_nrb_tiny_calculate(void)
1464 {
1465         int     nrbs = LNET_NRB_TINY;
1466
1467         if (tiny_router_buffers < 0) {
1468                 LCONSOLE_ERROR_MSG(0x10c,
1469                                    "tiny_router_buffers=%d invalid when "
1470                                    "routing enabled\n", tiny_router_buffers);
1471                 return -EINVAL;
1472         }
1473
1474         if (tiny_router_buffers > 0)
1475                 nrbs = tiny_router_buffers;
1476
1477         nrbs /= LNET_CPT_NUMBER;
1478         return max(nrbs, LNET_NRB_TINY_MIN);
1479 }
1480
1481 static int
1482 lnet_nrb_small_calculate(void)
1483 {
1484         int     nrbs = LNET_NRB_SMALL;
1485
1486         if (small_router_buffers < 0) {
1487                 LCONSOLE_ERROR_MSG(0x10c,
1488                                    "small_router_buffers=%d invalid when "
1489                                    "routing enabled\n", small_router_buffers);
1490                 return -EINVAL;
1491         }
1492
1493         if (small_router_buffers > 0)
1494                 nrbs = small_router_buffers;
1495
1496         nrbs /= LNET_CPT_NUMBER;
1497         return max(nrbs, LNET_NRB_SMALL_MIN);
1498 }
1499
1500 static int
1501 lnet_nrb_large_calculate(void)
1502 {
1503         int     nrbs = LNET_NRB_LARGE;
1504
1505         if (large_router_buffers < 0) {
1506                 LCONSOLE_ERROR_MSG(0x10c,
1507                                    "large_router_buffers=%d invalid when "
1508                                    "routing enabled\n", large_router_buffers);
1509                 return -EINVAL;
1510         }
1511
1512         if (large_router_buffers > 0)
1513                 nrbs = large_router_buffers;
1514
1515         nrbs /= LNET_CPT_NUMBER;
1516         return max(nrbs, LNET_NRB_LARGE_MIN);
1517 }
1518
1519 int
1520 lnet_rtrpools_alloc(int im_a_router)
1521 {
1522         struct lnet_rtrbufpool *rtrp;
1523         int     nrb_tiny;
1524         int     nrb_small;
1525         int     nrb_large;
1526         int     rc;
1527         int     i;
1528
1529         if (!strcmp(forwarding, "")) {
1530                 /* not set either way */
1531                 if (!im_a_router)
1532                         return 0;
1533         } else if (!strcmp(forwarding, "disabled")) {
1534                 /* explicitly disabled */
1535                 return 0;
1536         } else if (!strcmp(forwarding, "enabled")) {
1537                 /* explicitly enabled */
1538         } else {
1539                 rc = -EINVAL;
1540                 LCONSOLE_ERROR_MSG(0x10b,
1541                                    "lnet: forwarding='%s' not set to either 'enabled' or 'disabled': rc = %d\n",
1542                                    forwarding, rc);
1543                 return rc;
1544         }
1545
1546         nrb_tiny = lnet_nrb_tiny_calculate();
1547         if (nrb_tiny < 0)
1548                 return -EINVAL;
1549
1550         nrb_small = lnet_nrb_small_calculate();
1551         if (nrb_small < 0)
1552                 return -EINVAL;
1553
1554         nrb_large = lnet_nrb_large_calculate();
1555         if (nrb_large < 0)
1556                 return -EINVAL;
1557
1558         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1559                                                 LNET_NRBPOOLS *
1560                                                 sizeof(struct lnet_rtrbufpool));
1561         if (the_lnet.ln_rtrpools == NULL) {
1562                 rc = -ENOMEM;
1563                 LCONSOLE_ERROR_MSG(0x10c,
1564                         "lnet: error allocating router buffer pool: rc = %d\n",
1565                         rc);
1566                 return rc;
1567         }
1568
1569         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1570                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1571                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1572                                               nrb_tiny, i);
1573                 if (rc)
1574                         goto failed;
1575
1576                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1577                                   LNET_NRB_SMALL_PAGES);
1578                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1579                                               nrb_small, i);
1580                 if (rc)
1581                         goto failed;
1582
1583                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1584                                   LNET_NRB_LARGE_PAGES);
1585                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1586                                               nrb_large, i);
1587                 if (rc)
1588                         goto failed;
1589         }
1590
1591         lnet_net_lock(LNET_LOCK_EX);
1592         the_lnet.ln_routing = 1;
1593         lnet_net_unlock(LNET_LOCK_EX);
1594         complete(&the_lnet.ln_mt_wait_complete);
1595         return 0;
1596
1597  failed:
1598         lnet_rtrpools_free(0);
1599         return rc;
1600 }
1601
1602 static int
1603 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1604 {
1605         int nrb = 0;
1606         int rc = 0;
1607         int i;
1608         struct lnet_rtrbufpool *rtrp;
1609
1610         /* If the provided values for each buffer pool are different than the
1611          * configured values, we need to take action. */
1612         if (tiny >= 0) {
1613                 tiny_router_buffers = tiny;
1614                 nrb = lnet_nrb_tiny_calculate();
1615                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1616                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1617                                                       nrb, i);
1618                         if (rc != 0)
1619                                 return rc;
1620                 }
1621         }
1622         if (small >= 0) {
1623                 small_router_buffers = small;
1624                 nrb = lnet_nrb_small_calculate();
1625                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1626                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1627                                                       nrb, i);
1628                         if (rc != 0)
1629                                 return rc;
1630                 }
1631         }
1632         if (large >= 0) {
1633                 large_router_buffers = large;
1634                 nrb = lnet_nrb_large_calculate();
1635                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1636                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1637                                                       nrb, i);
1638                         if (rc != 0)
1639                                 return rc;
1640                 }
1641         }
1642
1643         return 0;
1644 }
1645
1646 int
1647 lnet_rtrpools_adjust(int tiny, int small, int large)
1648 {
1649         /* this function doesn't revert the changes if adding new buffers
1650          * failed.  It's up to the user space caller to revert the
1651          * changes. */
1652
1653         if (!the_lnet.ln_routing)
1654                 return 0;
1655
1656         return lnet_rtrpools_adjust_helper(tiny, small, large);
1657 }
1658
1659 int
1660 lnet_rtrpools_enable(void)
1661 {
1662         int rc = 0;
1663
1664         if (the_lnet.ln_routing)
1665                 return 0;
1666
1667         if (the_lnet.ln_rtrpools == NULL)
1668                 /* If routing is turned off, and we have never
1669                  * initialized the pools before, just call the
1670                  * standard buffer pool allocation routine as
1671                  * if we are just configuring this for the first
1672                  * time. */
1673                 rc = lnet_rtrpools_alloc(1);
1674         else
1675                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1676         if (rc != 0)
1677                 return rc;
1678
1679         lnet_net_lock(LNET_LOCK_EX);
1680         the_lnet.ln_routing = 1;
1681
1682         the_lnet.ln_ping_target->pb_info.pi_features &=
1683                 ~LNET_PING_FEAT_RTE_DISABLED;
1684         lnet_net_unlock(LNET_LOCK_EX);
1685
1686         if (lnet_peer_discovery_disabled)
1687                 CWARN("Consider turning discovery on to enable full "
1688                       "Multi-Rail routing functionality\n");
1689
1690         return rc;
1691 }
1692
1693 void
1694 lnet_rtrpools_disable(void)
1695 {
1696         if (!the_lnet.ln_routing)
1697                 return;
1698
1699         lnet_net_lock(LNET_LOCK_EX);
1700         the_lnet.ln_routing = 0;
1701         the_lnet.ln_ping_target->pb_info.pi_features |=
1702                 LNET_PING_FEAT_RTE_DISABLED;
1703
1704         tiny_router_buffers = 0;
1705         small_router_buffers = 0;
1706         large_router_buffers = 0;
1707         lnet_net_unlock(LNET_LOCK_EX);
1708         lnet_rtrpools_free(1);
1709 }
1710
1711 static inline void
1712 lnet_notify_peer_down(struct lnet_ni *ni, struct lnet_nid *nid)
1713 {
1714         if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1715                 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1716 }
1717
1718 /*
1719  * ni: local NI used to communicate with the peer
1720  * nid: peer NID
1721  * alive: true if peer is alive, false otherwise
1722  * reset: reset health value. This is requested by the LND.
1723  * when: notificaiton time.
1724  */
1725 int
1726 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset,
1727             time64_t when)
1728 {
1729         struct lnet_peer_ni *lpni = NULL;
1730         struct lnet_route *route;
1731         struct lnet_peer *lp;
1732         time64_t now = ktime_get_seconds();
1733         int cpt;
1734
1735         LASSERT(!in_interrupt());
1736
1737         CDEBUG(D_NET, "%s notifying %s: %s\n",
1738                (ni == NULL) ? "userspace" : libcfs_nidstr(&ni->ni_nid),
1739                libcfs_nid2str(nid), alive ? "up" : "down");
1740
1741         if (ni != NULL &&
1742             LNET_NID_NET(&ni->ni_nid) != LNET_NIDNET(nid)) {
1743                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1744                       libcfs_nid2str(nid), alive ? "birth" : "death",
1745                       libcfs_nidstr(&ni->ni_nid));
1746                 return -EINVAL;
1747         }
1748
1749         /* can't do predictions... */
1750         if (when > now) {
1751                 CWARN("Ignoring prediction from %s of %s %s %lld seconds in the future\n",
1752                         ni ? libcfs_nidstr(&ni->ni_nid) :  "userspace",
1753                         libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1754                 return -EINVAL;
1755         }
1756
1757         if (ni != NULL && !alive &&             /* LND telling me she's down */
1758             !auto_down) {                       /* auto-down disabled */
1759                 CDEBUG(D_NET, "Auto-down disabled\n");
1760                 return 0;
1761         }
1762
1763         /* must lock 0 since this is used for synchronization */
1764         lnet_net_lock(0);
1765
1766         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1767                 lnet_net_unlock(0);
1768                 return -ESHUTDOWN;
1769         }
1770
1771         lpni = lnet_find_peer_ni_locked(nid);
1772         if (lpni == NULL) {
1773                 /* nid not found */
1774                 lnet_net_unlock(0);
1775                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1776                 return 0;
1777         }
1778
1779         if (alive) {
1780                 if (reset) {
1781                         lpni->lpni_ns_status = LNET_NI_STATUS_UP;
1782                         lnet_set_lpni_healthv_locked(lpni,
1783                                                      LNET_MAX_HEALTH_VALUE);
1784                 } else {
1785                         __u32 sensitivity = lpni->lpni_peer_net->
1786                                         lpn_peer->lp_health_sensitivity;
1787
1788                         lnet_inc_lpni_healthv_locked(lpni,
1789                                         (sensitivity) ? sensitivity :
1790                                         lnet_health_sensitivity);
1791                 }
1792         } else if (reset) {
1793                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
1794         }
1795
1796         /* recalculate aliveness */
1797         alive = lnet_is_peer_ni_alive(lpni);
1798
1799         lp = lpni->lpni_peer_net->lpn_peer;
1800         /* If this is an LNet router then update route aliveness */
1801         if (lp->lp_rtr_refcount) {
1802                 if (reset)
1803                         /* reset flag indicates gateway peer went up or down */
1804                         lp->lp_alive = alive;
1805
1806                 /* If discovery is disabled, locally or on the gateway, then
1807                  * any routes using lpni as next-hop need to be updated
1808                  *
1809                  * NB: We can get many notifications while a route is down, so
1810                  * we try and avoid the expensive net_lock/EX here for the
1811                  * common case of receiving duplicate lnet_notify() calls (i.e.
1812                  * only grab EX lock when we actually need to update the route
1813                  * aliveness).
1814                  */
1815                 if (lnet_is_discovery_disabled(lp)) {
1816                         list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
1817                                 if (nid_same(&route->lr_nid, &lpni->lpni_nid))
1818                                         lnet_set_route_aliveness(route, alive);
1819                         }
1820                 }
1821         }
1822
1823         lnet_net_unlock(0);
1824
1825         if (ni != NULL && !alive)
1826                 lnet_notify_peer_down(ni, &lpni->lpni_nid);
1827
1828         cpt = lpni->lpni_cpt;
1829         lnet_net_lock(cpt);
1830         lnet_peer_ni_decref_locked(lpni);
1831         lnet_net_unlock(cpt);
1832
1833         return 0;
1834 }
1835 EXPORT_SYMBOL(lnet_notify);