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 *rn_list;
 586
 587         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 588
 589         rn_list = lnet_net2rnethash(net);
 590         list_for_each_entry(rnet, rn_list, lrn_list) {
 591                 if (rnet->lrn_net == net)
 592                         return rnet;
 593         }
 594         return NULL;
 595 }
 596
 597 static void lnet_shuffle_seed(void)
 598 {
 599         static int seeded;
 600         struct lnet_ni *ni = NULL;
 601
 602         if (seeded)
 603                 return;
 604
 605         /* Nodes with small feet have little entropy
 606          * the NID for this node gives the most entropy in the low bits */
 607         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 608                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 609
 610         seeded = 1;
 611 }
 612
 613 /* NB expects LNET_LOCK held */
 614 static void
 615 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 616 {
 617         struct lnet_peer_net *lpn;
 618         unsigned int offset = 0;
 619         unsigned int len = 0;
 620         struct list_head *e;
 621         time64_t now;
 622
 623         lnet_shuffle_seed();
 624
 625         list_for_each(e, &rnet->lrn_routes)
 626                 len++;
 627
 628         /*
 629          * Randomly adding routes to the list is done to ensure that when
 630          * different nodes are using the same list of routers, they end up
 631          * preferring different routers.
 632          */
 633         offset = prandom_u32_max(len + 1);
 634         list_for_each(e, &rnet->lrn_routes) {
 635                 if (offset == 0)
 636                         break;
 637                 offset--;
 638         }
 639         list_add(&route->lr_list, e);
 640         /*
 641          * force a router check on the gateway to make sure the route is
 642          * alive
 643          */
 644         now = ktime_get_real_seconds();
 645         list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets,
 646                             lpn_peer_nets) {
 647                 lpn->lpn_next_ping = now;
 648         }
 649
 650         the_lnet.ln_remote_nets_version++;
 651
 652         /* add the route on the gateway list */
 653         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 654
 655         /* take a router reference count on the gateway */
 656         lnet_rtr_addref_locked(route->lr_gateway);
 657 }
 658
 659 int
 660 lnet_add_route(__u32 net, __u32 hops, struct lnet_nid *gateway,
 661                __u32 priority, __u32 sensitivity)
 662 {
 663         struct list_head *route_entry;
 664         struct lnet_remotenet *rnet;
 665         struct lnet_remotenet *rnet2;
 666         struct lnet_route *route;
 667         struct lnet_peer_ni *lpni;
 668         struct lnet_peer *gw;
 669         int add_route;
 670         int rc;
 671
 672         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 673                libcfs_net2str(net), hops, priority, libcfs_nidstr(gateway));
 674
 675         if (LNET_NID_IS_ANY(gateway) ||
 676             nid_is_lo0(gateway) ||
 677             net == LNET_NET_ANY ||
 678             LNET_NETTYP(net) == LOLND ||
 679             LNET_NID_NET(gateway) == net ||
 680             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 681                 return -EINVAL;
 682
 683         /* it's a local network */
 684         if (lnet_islocalnet(net))
 685                 return -EEXIST;
 686
 687         if (!lnet_islocalnet(LNET_NID_NET(gateway))) {
 688                 CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n",
 689                        libcfs_nidstr(gateway),
 690                        libcfs_net2str(LNET_NID_NET(gateway)));
 691                 return -EHOSTUNREACH;
 692         }
 693
 694         /* Assume net, route, all new */
 695         LIBCFS_ALLOC(route, sizeof(*route));
 696         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 697         if (route == NULL || rnet == NULL) {
 698                 CERROR("Out of memory creating route %s %d %s\n",
 699                        libcfs_net2str(net), hops, libcfs_nidstr(gateway));
 700                 if (route != NULL)
 701                         LIBCFS_FREE(route, sizeof(*route));
 702                 if (rnet != NULL)
 703                         LIBCFS_FREE(rnet, sizeof(*rnet));
 704                 return -ENOMEM;
 705         }
 706
 707         INIT_LIST_HEAD(&rnet->lrn_routes);
 708         rnet->lrn_net = net;
 709         /* store the local and remote net that the route represents */
 710         route->lr_lnet = LNET_NID_NET(gateway);
 711         route->lr_net = net;
 712         route->lr_nid = *gateway;
 713         route->lr_priority = priority;
 714         route->lr_hops = hops;
 715         if (lnet_peers_start_down())
 716                 atomic_set(&route->lr_alive, 0);
 717         else
 718                 atomic_set(&route->lr_alive, 1);
 719
 720         lnet_net_lock(LNET_LOCK_EX);
 721
 722         /*
 723          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 724          * lose that once we're done
 725          */
 726         lpni = lnet_nid2peerni_ex(gateway);
 727         if (IS_ERR(lpni)) {
 728                 lnet_net_unlock(LNET_LOCK_EX);
 729
 730                 LIBCFS_FREE(route, sizeof(*route));
 731                 LIBCFS_FREE(rnet, sizeof(*rnet));
 732
 733                 rc = PTR_ERR(lpni);
 734                 CERROR("Error %d creating route %s %d %s\n", rc,
 735                         libcfs_net2str(net), hops,
 736                         libcfs_nidstr(gateway));
 737                 return rc;
 738         }
 739
 740         LASSERT(lpni);
 741         LASSERT(lpni->lpni_peer_net);
 742         LASSERT(lpni->lpni_peer_net->lpn_peer);
 743         gw = lpni->lpni_peer_net->lpn_peer;
 744
 745         route->lr_gateway = gw;
 746
 747         rnet2 = lnet_find_rnet_locked(net);
 748         if (rnet2 == NULL) {
 749                 /* new network */
 750                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 751                 rnet2 = rnet;
 752         }
 753
 754         /* Search for a duplicate route (it's a NOOP if it is) */
 755         add_route = 1;
 756         list_for_each(route_entry, &rnet2->lrn_routes) {
 757                 struct lnet_route *route2;
 758
 759                 route2 = list_entry(route_entry, struct lnet_route, lr_list);
 760                 if (route2->lr_gateway == route->lr_gateway) {
 761                         add_route = 0;
 762                         break;
 763                 }
 764
 765                 /* our lookups must be true */
 766                 LASSERT(!nid_same(&route2->lr_gateway->lp_primary_nid,
 767                                   gateway));
 768         }
 769
 770         /*
 771          * It is possible to add multiple routes through the same peer,
 772          * but it'll be using a different NID of that peer. When the
 773          * gateway is discovered, discovery will consolidate the different
 774          * peers into one peer. In this case the discovery code will have
 775          * to move the routes from the peer that's being deleted to the
 776          * consolidated peer lp_routes list
 777          */
 778         if (add_route) {
 779                 gw->lp_health_sensitivity = sensitivity;
 780                 lnet_add_route_to_rnet(rnet2, route);
 781                 if (lnet_peer_discovery_disabled)
 782                         CWARN("Consider turning discovery on to enable full Multi-Rail routing functionality\n");
 783         }
 784
 785         /*
 786          * get rid of the reference on the lpni.
 787          */
 788         lnet_peer_ni_decref_locked(lpni);
 789         lnet_net_unlock(LNET_LOCK_EX);
 790
 791         /* If avoid_asym_router_failure is enabled and hop count is not
 792          * set to 1 for a route that is actually single-hop, then the
 793          * feature will fail to prevent the router from being selected
 794          * if it is missing a NI on the remote network due to misconfiguration.
 795          */
 796         if (avoid_asym_router_failure && hops == LNET_UNDEFINED_HOPS)
 797                 CWARN("Use hops = 1 for a single-hop route when avoid_asym_router_failure feature is enabled\n");
 798
 799         rc = 0;
 800
 801         if (!add_route) {
 802                 rc = -EEXIST;
 803                 LIBCFS_FREE(route, sizeof(*route));
 804         }
 805
 806         if (rnet != rnet2)
 807                 LIBCFS_FREE(rnet, sizeof(*rnet));
 808
 809         /* kick start the monitor thread to handle the added route */
 810         complete(&the_lnet.ln_mt_wait_complete);
 811
 812         return rc;
 813 }
 814
 815 void
 816 lnet_del_route_from_rnet(struct lnet_nid *gw_nid,
 817                          struct list_head *route_list,
 818                          struct list_head *zombies)
 819 {
 820         struct lnet_peer *gateway;
 821         struct lnet_route *route;
 822         struct lnet_route *tmp;
 823
 824         list_for_each_entry_safe(route, tmp, route_list, lr_list) {
 825                 gateway = route->lr_gateway;
 826                 if (gw_nid && !nid_same(gw_nid, &gateway->lp_primary_nid))
 827                         continue;
 828
 829                 /*
 830                  * move to zombie to delete outside the lock
 831                  * Note that this function is called with the
 832                  * ln_api_mutex held as well as the exclusive net
 833                  * lock. Adding to the remote net list happens
 834                  * under the same conditions. Same goes for the
 835                  * gateway router list
 836                  */
 837                 list_move(&route->lr_list, zombies);
 838                 the_lnet.ln_remote_nets_version++;
 839
 840                 list_del(&route->lr_gwlist);
 841                 lnet_rtr_decref_locked(gateway);
 842         }
 843 }
 844
 845 int
 846 lnet_del_route(__u32 net, struct lnet_nid *gw)
 847 {
 848         LIST_HEAD(rnet_zombies);
 849         struct lnet_remotenet *rnet;
 850         struct lnet_remotenet *tmp;
 851         struct list_head *rn_list;
 852         struct lnet_peer_ni *lpni;
 853         struct lnet_route *route;
 854         struct lnet_nid gw_nid;
 855         LIST_HEAD(zombies);
 856         struct lnet_peer *lp = NULL;
 857         int i = 0;
 858
 859         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 860                libcfs_net2str(net), libcfs_nidstr(gw));
 861
 862         /* NB Caller may specify either all routes via the given gateway
 863          * or a specific route entry actual NIDs) */
 864
 865         lnet_net_lock(LNET_LOCK_EX);
 866
 867         if (gw)
 868                 lpni = lnet_peer_ni_find_locked(gw);
 869         else
 870                 lpni = NULL;
 871         if (lpni) {
 872                 lp = lpni->lpni_peer_net->lpn_peer;
 873                 LASSERT(lp);
 874                 gw_nid = lp->lp_primary_nid;
 875                 gw = &gw_nid;
 876                 lnet_peer_ni_decref_locked(lpni);
 877         }
 878
 879         if (net != LNET_NET_ANY) {
 880                 rnet = lnet_find_rnet_locked(net);
 881                 if (!rnet) {
 882                         lnet_net_unlock(LNET_LOCK_EX);
 883                         return -ENOENT;
 884                 }
 885                 lnet_del_route_from_rnet(gw, &rnet->lrn_routes,
 886                                          &zombies);
 887                 if (list_empty(&rnet->lrn_routes))
 888                         list_move(&rnet->lrn_list, &rnet_zombies);
 889                 goto delete_zombies;
 890         }
 891
 892         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 893                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 894
 895                 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
 896                         lnet_del_route_from_rnet(gw, &rnet->lrn_routes,
 897                                                  &zombies);
 898                         if (list_empty(&rnet->lrn_routes))
 899                                 list_move(&rnet->lrn_list, &rnet_zombies);
 900                 }
 901         }
 902
 903 delete_zombies:
 904         /*
 905          * check if there are any routes remaining on the gateway
 906          * If there are no more routes make sure to set the peer's
 907          * lp_disc_net_id to 0 (invalid), in case we add more routes in
 908          * the future on that gateway, then we start our discovery process
 909          * from scratch
 910          */
 911         if (lpni) {
 912                 if (list_empty(&lp->lp_routes))
 913                         lp->lp_disc_net_id = 0;
 914         }
 915
 916         lnet_net_unlock(LNET_LOCK_EX);
 917
 918         while (!list_empty(&zombies)) {
 919                 route = list_first_entry(&zombies, struct lnet_route, lr_list);
 920                 list_del(&route->lr_list);
 921                 LIBCFS_FREE(route, sizeof(*route));
 922         }
 923
 924         while (!list_empty(&rnet_zombies)) {
 925                 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
 926                                         lrn_list);
 927                 list_del(&rnet->lrn_list);
 928                 LIBCFS_FREE(rnet, sizeof(*rnet));
 929         }
 930
 931         return 0;
 932 }
 933
 934 void
 935 lnet_destroy_routes(void)
 936 {
 937         lnet_del_route(LNET_NET_ANY, NULL);
 938 }
 939
 940 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 941 {
 942         struct lnet_rtrbufpool *rbp;
 943         int i, rc = -ENOENT, j;
 944
 945         if (the_lnet.ln_rtrpools == NULL)
 946                 return rc;
 947
 948
 949         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 950                 if (i != cpt)
 951                         continue;
 952
 953                 lnet_net_lock(i);
 954                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 955                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 956                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 957                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 958                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 959                 }
 960                 lnet_net_unlock(i);
 961                 rc = 0;
 962                 break;
 963         }
 964
 965         lnet_net_lock(LNET_LOCK_EX);
 966         pool_cfg->pl_routing = the_lnet.ln_routing;
 967         lnet_net_unlock(LNET_LOCK_EX);
 968
 969         return rc;
 970 }
 971
 972 int
 973 lnet_get_route(int idx, __u32 *net, __u32 *hops, lnet_nid_t *gateway,
 974                __u32 *flags, __u32 *priority, __u32 *sensitivity)
 975 {
 976         struct lnet_remotenet *rnet;
 977         struct list_head *rn_list;
 978         struct lnet_route *route;
 979         int cpt;
 980         int i;
 981
 982         cpt = lnet_net_lock_current();
 983
 984         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 985                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 986                 list_for_each_entry(rnet, rn_list, lrn_list) {
 987                         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
 988                                 if (idx-- == 0) {
 989                                         *net      = rnet->lrn_net;
 990                                         *gateway  = lnet_nid_to_nid4(&route->lr_nid);
 991                                         *hops     = route->lr_hops;
 992                                         *priority = route->lr_priority;
 993                                         *sensitivity = route->lr_gateway->
 994                                                 lp_health_sensitivity;
 995                                         if (lnet_is_route_alive(route))
 996                                                 *flags |= LNET_RT_ALIVE;
 997                                         else
 998                                                 *flags &= ~LNET_RT_ALIVE;
 999                                         if (route->lr_single_hop)
1000                                                 *flags &= ~LNET_RT_MULTI_HOP;
1001                                         else
1002                                                 *flags |= LNET_RT_MULTI_HOP;
1003                                         lnet_net_unlock(cpt);
1004                                         return 0;
1005                                 }
1006                         }
1007                 }
1008         }
1009
1010         lnet_net_unlock(cpt);
1011         return -ENOENT;
1012 }
1013
1014 static void
1015 lnet_wait_known_routerstate(void)
1016 {
1017         struct lnet_peer *rtr;
1018         int all_known;
1019
1020         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
1021
1022         for (;;) {
1023                 int cpt = lnet_net_lock_current();
1024
1025                 all_known = 1;
1026                 list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1027                         spin_lock(&rtr->lp_lock);
1028
1029                         if ((rtr->lp_state & LNET_PEER_RTR_DISCOVERED) == 0) {
1030                                 all_known = 0;
1031                                 spin_unlock(&rtr->lp_lock);
1032                                 break;
1033                         }
1034                         spin_unlock(&rtr->lp_lock);
1035                 }
1036
1037                 lnet_net_unlock(cpt);
1038
1039                 if (all_known)
1040                         return;
1041
1042                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
1043         }
1044 }
1045
1046 static inline bool
1047 lnet_net_set_status_locked(struct lnet_net *net, __u32 status)
1048 {
1049         struct lnet_ni *ni;
1050         bool update = false;
1051
1052         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1053                 if (lnet_ni_set_status(ni, status))
1054                         update = true;
1055
1056         return update;
1057 }
1058
1059 static bool
1060 lnet_update_ni_status_locked(void)
1061 {
1062         struct lnet_net *net;
1063         struct lnet_ni *ni;
1064         bool push = false;
1065         time64_t now;
1066         time64_t timeout;
1067
1068         LASSERT(the_lnet.ln_routing);
1069
1070         timeout = router_ping_timeout + alive_router_check_interval;
1071
1072         now = ktime_get_seconds();
1073         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1074                 if (net->net_lnd->lnd_type == LOLND)
1075                         continue;
1076
1077                 if (now < net->net_last_alive + timeout)
1078                         goto check_ni_fatal;
1079
1080                 spin_lock(&net->net_lock);
1081                 /* re-check with lock */
1082                 if (now < net->net_last_alive + timeout) {
1083                         spin_unlock(&net->net_lock);
1084                         goto check_ni_fatal;
1085                 }
1086                 spin_unlock(&net->net_lock);
1087
1088                 /*
1089                  * if the net didn't receive any traffic for past the
1090                  * timeout on any of its constituent NIs, then mark all
1091                  * the NIs down.
1092                  */
1093                 if (lnet_net_set_status_locked(net, LNET_NI_STATUS_DOWN)) {
1094                         push = true;
1095                         continue;
1096                 }
1097
1098 check_ni_fatal:
1099                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1100                         /* lnet_ni_set_status() will perform the same check of
1101                          * ni_status while holding the ni lock. We can safely
1102                          * check ni_status without that lock because it is only
1103                          * written to under net_lock/EX and our caller is
1104                          * holding a net lock.
1105                          */
1106                         if (atomic_read(&ni->ni_fatal_error_on) &&
1107                             ni->ni_status &&
1108                             ni->ni_status->ns_status != LNET_NI_STATUS_DOWN &&
1109                             lnet_ni_set_status(ni, LNET_NI_STATUS_DOWN))
1110                                 push = true;
1111                 }
1112         }
1113
1114         return push;
1115 }
1116
1117 void lnet_wait_router_start(void)
1118 {
1119         if (check_routers_before_use) {
1120                 /* Note that a helpful side-effect of pinging all known routers
1121                  * at startup is that it makes them drop stale connections they
1122                  * may have to a previous instance of me. */
1123                 lnet_wait_known_routerstate();
1124         }
1125 }
1126
1127 /*
1128  * This function is called from the monitor thread to check if there are
1129  * any active routers that need to be checked.
1130  */
1131 bool lnet_router_checker_active(void)
1132 {
1133         /* Router Checker thread needs to run when routing is enabled in
1134          * order to call lnet_update_ni_status_locked() */
1135         if (the_lnet.ln_routing)
1136                 return true;
1137
1138         return !list_empty(&the_lnet.ln_routers) &&
1139                 alive_router_check_interval > 0;
1140 }
1141
1142 void
1143 lnet_check_routers(void)
1144 {
1145         struct lnet_peer_net *first_lpn;
1146         struct lnet_peer_net *lpn;
1147         struct lnet_peer_ni *lpni;
1148         struct lnet_peer *rtr;
1149         bool push = false;
1150         bool needs_ping;
1151         bool found_lpn;
1152         __u64 version;
1153         __u32 net_id;
1154         time64_t now;
1155         int cpt;
1156         int rc;
1157
1158         cpt = lnet_net_lock_current();
1159 rescan:
1160         version = the_lnet.ln_routers_version;
1161
1162         list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1163                 /* If we're currently discovering the peer then don't
1164                  * issue another discovery
1165                  */
1166                 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY)
1167                         continue;
1168
1169                 now = ktime_get_real_seconds();
1170
1171                 /* find the next local peer net which needs to be ping'd */
1172                 needs_ping = false;
1173                 first_lpn = NULL;
1174                 found_lpn = false;
1175                 net_id = rtr->lp_disc_net_id;
1176                 do {
1177                         lpn = lnet_get_next_peer_net_locked(rtr, net_id);
1178                         if (!lpn) {
1179                                 CERROR("gateway %s has no networks\n",
1180                                 libcfs_nidstr(&rtr->lp_primary_nid));
1181                                 break;
1182                         }
1183
1184                         /* We looped back to the first peer net */
1185                         if (first_lpn == lpn)
1186                                 break;
1187                         if (!first_lpn)
1188                                 first_lpn = lpn;
1189
1190                         net_id = lpn->lpn_net_id;
1191                         if (!lnet_islocalnet_locked(net_id))
1192                                 continue;
1193
1194                         found_lpn = true;
1195
1196                         CDEBUG(D_NET, "rtr %s(%p) %s(%p) next ping %lld\n",
1197                                libcfs_nidstr(&rtr->lp_primary_nid), rtr,
1198                                libcfs_net2str(net_id), lpn,
1199                                lpn->lpn_next_ping);
1200
1201                         needs_ping = now >= lpn->lpn_next_ping;
1202
1203                 } while (!needs_ping);
1204
1205                 if (!found_lpn || !lpn) {
1206                         CERROR("no local network found for gateway %s\n",
1207                                libcfs_nidstr(&rtr->lp_primary_nid));
1208                         continue;
1209                 }
1210
1211                 if (!needs_ping)
1212                         continue;
1213
1214                 spin_lock(&rtr->lp_lock);
1215                 /* make sure we fully discover the router */
1216                 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1217                 rtr->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH |
1218                         LNET_PEER_RTR_DISCOVERY;
1219                 spin_unlock(&rtr->lp_lock);
1220
1221                 /* find the peer_ni associated with the primary NID */
1222                 lpni = lnet_peer_ni_get_locked(rtr, &rtr->lp_primary_nid);
1223                 if (!lpni) {
1224                         CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
1225                                libcfs_nidstr(&rtr->lp_primary_nid));
1226                         continue;
1227                 }
1228                 lnet_peer_ni_addref_locked(lpni);
1229
1230                 /* specify the net to use */
1231                 rtr->lp_disc_net_id = lpn->lpn_net_id;
1232
1233                 /* discover the router */
1234                 CDEBUG(D_NET, "discover %s, cpt = %d\n",
1235                        libcfs_nidstr(&lpni->lpni_nid), cpt);
1236                 rc = lnet_discover_peer_locked(lpni, cpt, false);
1237
1238                 /* drop ref taken above */
1239                 lnet_peer_ni_decref_locked(lpni);
1240
1241                 if (!rc)
1242                         lpn->lpn_next_ping = now + alive_router_check_interval;
1243                 else
1244                         CERROR("Failed to discover router %s\n",
1245                                libcfs_nidstr(&rtr->lp_primary_nid));
1246
1247                 /* NB cpt lock was dropped in lnet_discover_peer_locked() */
1248                 if (version != the_lnet.ln_routers_version) {
1249                         /* the routers list has changed */
1250                         goto rescan;
1251                 }
1252         }
1253
1254         if (the_lnet.ln_routing)
1255                 push = lnet_update_ni_status_locked();
1256
1257         lnet_net_unlock(cpt);
1258
1259         /* if the status of the ni changed update the peers */
1260         if (push)
1261                 lnet_push_update_to_peers(1);
1262 }
1263
1264 void
1265 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1266 {
1267         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1268
1269         while (--npages >= 0)
1270                 __free_page(rb->rb_kiov[npages].bv_page);
1271
1272         LIBCFS_FREE(rb, sz);
1273 }
1274
1275 static struct lnet_rtrbuf *
1276 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1277 {
1278         int            npages = rbp->rbp_npages;
1279         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1280         struct page   *page;
1281         struct lnet_rtrbuf *rb;
1282         int            i;
1283
1284         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1285         if (rb == NULL)
1286                 return NULL;
1287
1288         rb->rb_pool = rbp;
1289
1290         for (i = 0; i < npages; i++) {
1291                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt, GFP_KERNEL |
1292                                           __GFP_ZERO | __GFP_NORETRY);
1293                 if (page == NULL) {
1294                         while (--i >= 0)
1295                                 __free_page(rb->rb_kiov[i].bv_page);
1296
1297                         LIBCFS_FREE(rb, sz);
1298                         return NULL;
1299                 }
1300
1301                 rb->rb_kiov[i].bv_len = PAGE_SIZE;
1302                 rb->rb_kiov[i].bv_offset = 0;
1303                 rb->rb_kiov[i].bv_page = page;
1304         }
1305
1306         return rb;
1307 }
1308
1309 static void
1310 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1311 {
1312         int npages = rbp->rbp_npages;
1313         struct lnet_rtrbuf *rb;
1314         LIST_HEAD(tmp);
1315
1316         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1317                 return;
1318
1319         lnet_net_lock(cpt);
1320         list_splice_init(&rbp->rbp_msgs, &tmp);
1321         lnet_drop_routed_msgs_locked(&tmp, cpt);
1322         list_splice_init(&rbp->rbp_bufs, &tmp);
1323         rbp->rbp_req_nbuffers = 0;
1324         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1325         rbp->rbp_mincredits = 0;
1326         lnet_net_unlock(cpt);
1327
1328         /* Free buffers on the free list. */
1329         while (!list_empty(&tmp)) {
1330                 rb = list_first_entry(&tmp, struct lnet_rtrbuf, rb_list);
1331                 list_del(&rb->rb_list);
1332                 lnet_destroy_rtrbuf(rb, npages);
1333         }
1334 }
1335
1336 static int
1337 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1338 {
1339         LIST_HEAD(rb_list);
1340         struct lnet_rtrbuf *rb;
1341         int             num_rb;
1342         int             num_buffers = 0;
1343         int             old_req_nbufs;
1344         int             npages = rbp->rbp_npages;
1345
1346         lnet_net_lock(cpt);
1347         /* If we are called for less buffers than already in the pool, we
1348          * just lower the req_nbuffers number and excess buffers will be
1349          * thrown away as they are returned to the free list.  Credits
1350          * then get adjusted as well.
1351          * If we already have enough buffers allocated to serve the
1352          * increase requested, then we can treat that the same way as we
1353          * do the decrease. */
1354         num_rb = nbufs - rbp->rbp_nbuffers;
1355         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1356                 rbp->rbp_req_nbuffers = nbufs;
1357                 lnet_net_unlock(cpt);
1358                 return 0;
1359         }
1360         /* store the older value of rbp_req_nbuffers and then set it to
1361          * the new request to prevent lnet_return_rx_credits_locked() from
1362          * freeing buffers that we need to keep around */
1363         old_req_nbufs = rbp->rbp_req_nbuffers;
1364         rbp->rbp_req_nbuffers = nbufs;
1365         lnet_net_unlock(cpt);
1366
1367         /* allocate the buffers on a local list first.  If all buffers are
1368          * allocated successfully then join this list to the rbp buffer
1369          * list.  If not then free all allocated buffers. */
1370         while (num_rb-- > 0) {
1371                 rb = lnet_new_rtrbuf(rbp, cpt);
1372                 if (rb == NULL) {
1373                         CERROR("lnet: error allocating %ux%u page router buffers on CPT %u: rc = %d\n",
1374                                nbufs, npages, cpt, -ENOMEM);
1375
1376                         lnet_net_lock(cpt);
1377                         rbp->rbp_req_nbuffers = old_req_nbufs;
1378                         lnet_net_unlock(cpt);
1379
1380                         goto failed;
1381                 }
1382
1383                 list_add(&rb->rb_list, &rb_list);
1384                 num_buffers++;
1385         }
1386
1387         lnet_net_lock(cpt);
1388
1389         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1390         rbp->rbp_nbuffers += num_buffers;
1391         rbp->rbp_credits += num_buffers;
1392         rbp->rbp_mincredits = rbp->rbp_credits;
1393         /* We need to schedule blocked msg using the newly
1394          * added buffers. */
1395         while (!list_empty(&rbp->rbp_bufs) &&
1396                !list_empty(&rbp->rbp_msgs))
1397                 lnet_schedule_blocked_locked(rbp);
1398
1399         lnet_net_unlock(cpt);
1400
1401         return 0;
1402
1403 failed:
1404         while ((rb = list_first_entry_or_null(&rb_list,
1405                                               struct lnet_rtrbuf,
1406                                               rb_list)) != NULL) {
1407                 list_del(&rb->rb_list);
1408                 lnet_destroy_rtrbuf(rb, npages);
1409         }
1410
1411         return -ENOMEM;
1412 }
1413
1414 static void
1415 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1416 {
1417         INIT_LIST_HEAD(&rbp->rbp_msgs);
1418         INIT_LIST_HEAD(&rbp->rbp_bufs);
1419
1420         rbp->rbp_npages = npages;
1421         rbp->rbp_credits = 0;
1422         rbp->rbp_mincredits = 0;
1423 }
1424
1425 void
1426 lnet_rtrpools_free(int keep_pools)
1427 {
1428         struct lnet_rtrbufpool *rtrp;
1429         int               i;
1430
1431         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1432                 return;
1433
1434         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1435                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1436                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1437                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1438         }
1439
1440         if (!keep_pools) {
1441                 cfs_percpt_free(the_lnet.ln_rtrpools);
1442                 the_lnet.ln_rtrpools = NULL;
1443         }
1444 }
1445
1446 static int
1447 lnet_nrb_tiny_calculate(void)
1448 {
1449         int     nrbs = LNET_NRB_TINY;
1450
1451         if (tiny_router_buffers < 0) {
1452                 LCONSOLE_ERROR_MSG(0x10c,
1453                                    "tiny_router_buffers=%d invalid when "
1454                                    "routing enabled\n", tiny_router_buffers);
1455                 return -EINVAL;
1456         }
1457
1458         if (tiny_router_buffers > 0)
1459                 nrbs = tiny_router_buffers;
1460
1461         nrbs /= LNET_CPT_NUMBER;
1462         return max(nrbs, LNET_NRB_TINY_MIN);
1463 }
1464
1465 static int
1466 lnet_nrb_small_calculate(void)
1467 {
1468         int     nrbs = LNET_NRB_SMALL;
1469
1470         if (small_router_buffers < 0) {
1471                 LCONSOLE_ERROR_MSG(0x10c,
1472                                    "small_router_buffers=%d invalid when "
1473                                    "routing enabled\n", small_router_buffers);
1474                 return -EINVAL;
1475         }
1476
1477         if (small_router_buffers > 0)
1478                 nrbs = small_router_buffers;
1479
1480         nrbs /= LNET_CPT_NUMBER;
1481         return max(nrbs, LNET_NRB_SMALL_MIN);
1482 }
1483
1484 static int
1485 lnet_nrb_large_calculate(void)
1486 {
1487         int     nrbs = LNET_NRB_LARGE;
1488
1489         if (large_router_buffers < 0) {
1490                 LCONSOLE_ERROR_MSG(0x10c,
1491                                    "large_router_buffers=%d invalid when "
1492                                    "routing enabled\n", large_router_buffers);
1493                 return -EINVAL;
1494         }
1495
1496         if (large_router_buffers > 0)
1497                 nrbs = large_router_buffers;
1498
1499         nrbs /= LNET_CPT_NUMBER;
1500         return max(nrbs, LNET_NRB_LARGE_MIN);
1501 }
1502
1503 int
1504 lnet_rtrpools_alloc(int im_a_router)
1505 {
1506         struct lnet_rtrbufpool *rtrp;
1507         int     nrb_tiny;
1508         int     nrb_small;
1509         int     nrb_large;
1510         int     rc;
1511         int     i;
1512
1513         if (!strcmp(forwarding, "")) {
1514                 /* not set either way */
1515                 if (!im_a_router)
1516                         return 0;
1517         } else if (!strcmp(forwarding, "disabled")) {
1518                 /* explicitly disabled */
1519                 return 0;
1520         } else if (!strcmp(forwarding, "enabled")) {
1521                 /* explicitly enabled */
1522         } else {
1523                 rc = -EINVAL;
1524                 LCONSOLE_ERROR_MSG(0x10b,
1525                                    "lnet: forwarding='%s' not set to either 'enabled' or 'disabled': rc = %d\n",
1526                                    forwarding, rc);
1527                 return rc;
1528         }
1529
1530         nrb_tiny = lnet_nrb_tiny_calculate();
1531         if (nrb_tiny < 0)
1532                 return -EINVAL;
1533
1534         nrb_small = lnet_nrb_small_calculate();
1535         if (nrb_small < 0)
1536                 return -EINVAL;
1537
1538         nrb_large = lnet_nrb_large_calculate();
1539         if (nrb_large < 0)
1540                 return -EINVAL;
1541
1542         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1543                                                 LNET_NRBPOOLS *
1544                                                 sizeof(struct lnet_rtrbufpool));
1545         if (the_lnet.ln_rtrpools == NULL) {
1546                 rc = -ENOMEM;
1547                 LCONSOLE_ERROR_MSG(0x10c,
1548                         "lnet: error allocating router buffer pool: rc = %d\n",
1549                         rc);
1550                 return rc;
1551         }
1552
1553         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1554                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1555                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1556                                               nrb_tiny, i);
1557                 if (rc)
1558                         goto failed;
1559
1560                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1561                                   LNET_NRB_SMALL_PAGES);
1562                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1563                                               nrb_small, i);
1564                 if (rc)
1565                         goto failed;
1566
1567                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1568                                   LNET_NRB_LARGE_PAGES);
1569                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1570                                               nrb_large, i);
1571                 if (rc)
1572                         goto failed;
1573         }
1574
1575         lnet_net_lock(LNET_LOCK_EX);
1576         the_lnet.ln_routing = 1;
1577         lnet_net_unlock(LNET_LOCK_EX);
1578         complete(&the_lnet.ln_mt_wait_complete);
1579         return 0;
1580
1581  failed:
1582         lnet_rtrpools_free(0);
1583         return rc;
1584 }
1585
1586 static int
1587 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1588 {
1589         int nrb = 0;
1590         int rc = 0;
1591         int i;
1592         struct lnet_rtrbufpool *rtrp;
1593
1594         /* If the provided values for each buffer pool are different than the
1595          * configured values, we need to take action. */
1596         if (tiny >= 0) {
1597                 tiny_router_buffers = tiny;
1598                 nrb = lnet_nrb_tiny_calculate();
1599                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1600                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1601                                                       nrb, i);
1602                         if (rc != 0)
1603                                 return rc;
1604                 }
1605         }
1606         if (small >= 0) {
1607                 small_router_buffers = small;
1608                 nrb = lnet_nrb_small_calculate();
1609                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1610                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1611                                                       nrb, i);
1612                         if (rc != 0)
1613                                 return rc;
1614                 }
1615         }
1616         if (large >= 0) {
1617                 large_router_buffers = large;
1618                 nrb = lnet_nrb_large_calculate();
1619                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1620                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1621                                                       nrb, i);
1622                         if (rc != 0)
1623                                 return rc;
1624                 }
1625         }
1626
1627         return 0;
1628 }
1629
1630 int
1631 lnet_rtrpools_adjust(int tiny, int small, int large)
1632 {
1633         /* this function doesn't revert the changes if adding new buffers
1634          * failed.  It's up to the user space caller to revert the
1635          * changes. */
1636
1637         if (!the_lnet.ln_routing)
1638                 return 0;
1639
1640         return lnet_rtrpools_adjust_helper(tiny, small, large);
1641 }
1642
1643 int
1644 lnet_rtrpools_enable(void)
1645 {
1646         int rc = 0;
1647
1648         if (the_lnet.ln_routing)
1649                 return 0;
1650
1651         if (the_lnet.ln_rtrpools == NULL)
1652                 /* If routing is turned off, and we have never
1653                  * initialized the pools before, just call the
1654                  * standard buffer pool allocation routine as
1655                  * if we are just configuring this for the first
1656                  * time. */
1657                 rc = lnet_rtrpools_alloc(1);
1658         else
1659                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1660         if (rc != 0)
1661                 return rc;
1662
1663         lnet_net_lock(LNET_LOCK_EX);
1664         the_lnet.ln_routing = 1;
1665
1666         the_lnet.ln_ping_target->pb_info.pi_features &=
1667                 ~LNET_PING_FEAT_RTE_DISABLED;
1668         lnet_net_unlock(LNET_LOCK_EX);
1669
1670         if (lnet_peer_discovery_disabled)
1671                 CWARN("Consider turning discovery on to enable full "
1672                       "Multi-Rail routing functionality\n");
1673
1674         return rc;
1675 }
1676
1677 void
1678 lnet_rtrpools_disable(void)
1679 {
1680         if (!the_lnet.ln_routing)
1681                 return;
1682
1683         lnet_net_lock(LNET_LOCK_EX);
1684         the_lnet.ln_routing = 0;
1685         the_lnet.ln_ping_target->pb_info.pi_features |=
1686                 LNET_PING_FEAT_RTE_DISABLED;
1687
1688         tiny_router_buffers = 0;
1689         small_router_buffers = 0;
1690         large_router_buffers = 0;
1691         lnet_net_unlock(LNET_LOCK_EX);
1692         lnet_rtrpools_free(1);
1693 }
1694
1695 static inline void
1696 lnet_notify_peer_down(struct lnet_ni *ni, struct lnet_nid *nid)
1697 {
1698         if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1699                 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1700 }
1701
1702 /*
1703  * ni: local NI used to communicate with the peer
1704  * nid: peer NID
1705  * alive: true if peer is alive, false otherwise
1706  * reset: reset health value. This is requested by the LND.
1707  * when: notificaiton time.
1708  */
1709 int
1710 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid4, bool alive, bool reset,
1711             time64_t when)
1712 {
1713         struct lnet_peer_ni *lpni = NULL;
1714         struct lnet_route *route;
1715         struct lnet_peer *lp;
1716         time64_t now = ktime_get_seconds();
1717         struct lnet_nid nid;
1718         int cpt;
1719
1720         lnet_nid4_to_nid(nid4, &nid);
1721         LASSERT(!in_interrupt());
1722
1723         CDEBUG(D_NET, "%s notifying %s: %s\n",
1724                (ni == NULL) ? "userspace" : libcfs_nidstr(&ni->ni_nid),
1725                libcfs_nidstr(&nid), alive ? "up" : "down");
1726
1727         if (ni != NULL &&
1728             LNET_NID_NET(&ni->ni_nid) != LNET_NID_NET(&nid)) {
1729                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1730                       libcfs_nidstr(&nid), alive ? "birth" : "death",
1731                       libcfs_nidstr(&ni->ni_nid));
1732                 return -EINVAL;
1733         }
1734
1735         /* can't do predictions... */
1736         if (when > now) {
1737                 CWARN("Ignoring prediction from %s of %s %s %lld seconds in the future\n",
1738                         ni ? libcfs_nidstr(&ni->ni_nid) :  "userspace",
1739                         libcfs_nidstr(&nid), alive ? "up" : "down", when - now);
1740                 return -EINVAL;
1741         }
1742
1743         if (ni != NULL && !alive &&             /* LND telling me she's down */
1744             !auto_down) {                       /* auto-down disabled */
1745                 CDEBUG(D_NET, "Auto-down disabled\n");
1746                 return 0;
1747         }
1748
1749         /* must lock 0 since this is used for synchronization */
1750         lnet_net_lock(0);
1751
1752         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1753                 lnet_net_unlock(0);
1754                 return -ESHUTDOWN;
1755         }
1756
1757         lpni = lnet_peer_ni_find_locked(&nid);
1758         if (lpni == NULL) {
1759                 /* nid not found */
1760                 lnet_net_unlock(0);
1761                 CDEBUG(D_NET, "%s not found\n", libcfs_nidstr(&nid));
1762                 return 0;
1763         }
1764
1765         if (alive) {
1766                 if (reset) {
1767                         lpni->lpni_ns_status = LNET_NI_STATUS_UP;
1768                         lnet_set_lpni_healthv_locked(lpni,
1769                                                      LNET_MAX_HEALTH_VALUE);
1770                 } else {
1771                         __u32 sensitivity = lpni->lpni_peer_net->
1772                                         lpn_peer->lp_health_sensitivity;
1773
1774                         lnet_inc_lpni_healthv_locked(lpni,
1775                                         (sensitivity) ? sensitivity :
1776                                         lnet_health_sensitivity);
1777                 }
1778         } else if (reset) {
1779                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
1780         }
1781
1782         /* recalculate aliveness */
1783         alive = lnet_is_peer_ni_alive(lpni);
1784
1785         lp = lpni->lpni_peer_net->lpn_peer;
1786         /* If this is an LNet router then update route aliveness */
1787         if (lp->lp_rtr_refcount) {
1788                 if (reset)
1789                         /* reset flag indicates gateway peer went up or down */
1790                         lp->lp_alive = alive;
1791
1792                 /* If discovery is disabled, locally or on the gateway, then
1793                  * any routes using lpni as next-hop need to be updated
1794                  *
1795                  * NB: We can get many notifications while a route is down, so
1796                  * we try and avoid the expensive net_lock/EX here for the
1797                  * common case of receiving duplicate lnet_notify() calls (i.e.
1798                  * only grab EX lock when we actually need to update the route
1799                  * aliveness).
1800                  */
1801                 if (lnet_is_discovery_disabled(lp)) {
1802                         list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
1803                                 if (nid_same(&route->lr_nid, &lpni->lpni_nid))
1804                                         lnet_set_route_aliveness(route, alive);
1805                         }
1806                 }
1807         }
1808
1809         lnet_net_unlock(0);
1810
1811         if (ni != NULL && !alive)
1812                 lnet_notify_peer_down(ni, &lpni->lpni_nid);
1813
1814         cpt = lpni->lpni_cpt;
1815         lnet_net_lock(cpt);
1816         lnet_peer_ni_decref_locked(lpni);
1817         lnet_net_unlock(cpt);
1818
1819         return 0;
1820 }
1821 EXPORT_SYMBOL(lnet_notify);