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(lnet_nid_t gw_nid, struct list_head *route_list,
 124                                      struct list_head *zombies);
 125
 126 static int
 127 rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 128 {
 129         int rc;
 130         unsigned *sen = (unsigned *)kp->arg;
 131         unsigned long value;
 132
 133         rc = kstrtoul(val, 0, &value);
 134         if (rc) {
 135                 CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
 136                 return rc;
 137         }
 138
 139         if (value < 0 || value > 100) {
 140                 CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
 141                 return -EINVAL;
 142         }
 143
 144         /*
 145          * The purpose of locking the api_mutex here is to ensure that
 146          * the correct value ends up stored properly.
 147          */
 148         mutex_lock(&the_lnet.ln_api_mutex);
 149
 150         *sen = value;
 151
 152         mutex_unlock(&the_lnet.ln_api_mutex);
 153
 154         return 0;
 155 }
 156
 157 void
 158 lnet_move_route(struct lnet_route *route, struct lnet_peer *lp,
 159                 struct list_head *rt_list)
 160 {
 161         struct lnet_remotenet *rnet;
 162         struct list_head zombies;
 163         struct list_head *l;
 164
 165         INIT_LIST_HEAD(&zombies);
 166
 167         if (rt_list)
 168                 l = rt_list;
 169         else
 170                 l = &zombies;
 171
 172         rnet = lnet_find_rnet_locked(route->lr_net);
 173         LASSERT(rnet);
 174
 175         CDEBUG(D_NET, "deleting route %s->%s\n",
 176                libcfs_net2str(route->lr_net),
 177                libcfs_nid2str(route->lr_nid));
 178
 179         /*
 180          * use the gateway's lp_primary_nid to delete the route as the
 181          * lr_nid can be a constituent NID of the peer
 182          */
 183         lnet_del_route_from_rnet(route->lr_gateway->lp_primary_nid,
 184                                  &rnet->lrn_routes, l);
 185
 186         if (lp) {
 187                 route = list_first_entry(l, struct lnet_route,
 188                                         lr_list);
 189                 route->lr_gateway = lp;
 190                 lnet_add_route_to_rnet(rnet, route);
 191         } else {
 192                 while (!list_empty(l) && !rt_list) {
 193                         route = list_first_entry(l, struct lnet_route,
 194                                  lr_list);
 195                         list_del(&route->lr_list);
 196                         LIBCFS_FREE(route, sizeof(*route));
 197                 }
 198         }
 199 }
 200
 201 void
 202 lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target)
 203 {
 204         struct lnet_route *route;
 205         struct lnet_route *tmp, *tmp2;
 206
 207         lnet_net_lock(LNET_LOCK_EX);
 208         CDEBUG(D_NET, "transfering routes from %s -> %s\n",
 209                libcfs_nid2str(src->lp_primary_nid),
 210                libcfs_nid2str(target->lp_primary_nid));
 211         list_for_each_entry(route, &src->lp_routes, lr_gwlist) {
 212                 CDEBUG(D_NET, "%s: %s->%s\n", libcfs_nid2str(src->lp_primary_nid),
 213                        libcfs_net2str(route->lr_net),
 214                        libcfs_nid2str(route->lr_nid));
 215         }
 216         list_splice_init(&src->lp_rtrq, &target->lp_rtrq);
 217         list_for_each_entry_safe(route, tmp, &src->lp_routes, lr_gwlist) {
 218                 struct lnet_route *r2;
 219                 bool present = false;
 220                 list_for_each_entry_safe(r2, tmp2, &target->lp_routes, lr_gwlist) {
 221                         if (route->lr_net == r2->lr_net) {
 222                                 if (route->lr_priority >= r2->lr_priority)
 223                                         present = true;
 224                                 else if (route->lr_hops >= r2->lr_hops)
 225                                         present = true;
 226                                 else
 227                                         lnet_move_route(r2, NULL, NULL);
 228                         }
 229                 }
 230                 if (present)
 231                         lnet_move_route(route, NULL, NULL);
 232                 else
 233                         lnet_move_route(route, target, NULL);
 234         }
 235
 236         if (list_empty(&target->lp_rtr_list)) {
 237                 lnet_peer_addref_locked(target);
 238                 list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers);
 239         }
 240
 241         the_lnet.ln_routers_version++;
 242         lnet_net_unlock(LNET_LOCK_EX);
 243 }
 244
 245 int
 246 lnet_peers_start_down(void)
 247 {
 248         return check_routers_before_use;
 249 }
 250
 251 /*
 252  * The peer_net of a gateway is alive if at least one of the peer_ni's on
 253  * that peer_net is alive.
 254  */
 255 static bool
 256 lnet_is_gateway_net_alive(struct lnet_peer_net *lpn)
 257 {
 258         struct lnet_peer_ni *lpni;
 259
 260         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 261                 if (lnet_is_peer_ni_alive(lpni))
 262                         return true;
 263         }
 264
 265         return false;
 266 }
 267
 268 /*
 269  * a gateway is alive only if all its nets are alive
 270  * called with cpt lock held
 271  */
 272 bool lnet_is_gateway_alive(struct lnet_peer *gw)
 273 {
 274         struct lnet_peer_net *lpn;
 275
 276         if (!gw->lp_alive)
 277                 return false;
 278
 279         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 280                 if (!lnet_is_gateway_net_alive(lpn))
 281                         return false;
 282         }
 283
 284         return true;
 285 }
 286
 287 /*
 288  * lnet_is_route_alive() needs to be called with cpt lock held
 289  * A route is alive if the gateway can route between the local network and
 290  * the remote network of the route.
 291  * This means at least one NI is alive on each of the local and remote
 292  * networks of the gateway.
 293  */
 294 bool lnet_is_route_alive(struct lnet_route *route)
 295 {
 296         struct lnet_peer *gw = route->lr_gateway;
 297         struct lnet_peer_net *llpn;
 298         struct lnet_peer_net *rlpn;
 299
 300         /* If the gateway is down then all routes are considered down */
 301         if (!gw->lp_alive)
 302                 return false;
 303
 304         /*
 305          * if discovery is disabled then rely on the cached aliveness
 306          * information. This is handicapped information which we log when
 307          * we receive the discovery ping response. The most uptodate
 308          * aliveness information can only be obtained when discovery is
 309          * enabled.
 310          */
 311         if (lnet_is_discovery_disabled(gw))
 312                 return route->lr_alive;
 313
 314         /*
 315          * check the gateway's interfaces on the local network
 316          */
 317         llpn = lnet_peer_get_net_locked(gw, route->lr_lnet);
 318         if (!llpn)
 319                 return false;
 320
 321         if (!lnet_is_gateway_net_alive(llpn))
 322                 return false;
 323
 324         /*
 325          * For single hop routes avoid_asym_router_failure dictates
 326          * that the remote net must exist on the gateway. For multi-hop
 327          * routes the next-hop will not have the remote net.
 328          */
 329         if (avoid_asym_router_failure && route->lr_single_hop) {
 330                 rlpn = lnet_peer_get_net_locked(gw, route->lr_net);
 331                 if (!rlpn)
 332                         return false;
 333                 if (!lnet_is_gateway_net_alive(rlpn))
 334                         return false;
 335         }
 336
 337         spin_lock(&gw->lp_lock);
 338         if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) {
 339                 spin_unlock(&gw->lp_lock);
 340                 if (gw->lp_rtr_refcount > 0)
 341                         CERROR("peer %s is being used as a gateway but routing feature is not turned on\n",
 342                                libcfs_nid2str(gw->lp_primary_nid));
 343                 return false;
 344         }
 345         spin_unlock(&gw->lp_lock);
 346
 347         return true;
 348 }
 349
 350 void
 351 lnet_consolidate_routes_locked(struct lnet_peer *orig_lp,
 352                                struct lnet_peer *new_lp)
 353 {
 354         struct lnet_peer_ni *lpni;
 355         struct lnet_route *route;
 356
 357         /*
 358          * Although a route is correlated with a peer, but when it's added
 359          * a specific NID is used. That NID refers to a peer_ni within
 360          * a peer. There could be other peer_nis on the same net, which
 361          * can be used to send to that gateway. However when we are
 362          * consolidating gateways because of discovery, the nid used to
 363          * add the route might've moved between gateway peers. In this
 364          * case we want to move the route to the new gateway as well. The
 365          * intent here is not to confuse the user who added the route.
 366          */
 367         list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) {
 368                 lpni = lnet_peer_get_ni_locked(orig_lp, route->lr_nid);
 369                 if (!lpni) {
 370                         lnet_net_lock(LNET_LOCK_EX);
 371                         list_move(&route->lr_gwlist, &new_lp->lp_routes);
 372                         lnet_net_unlock(LNET_LOCK_EX);
 373                 }
 374         }
 375 }
 376
 377 static inline void
 378 lnet_check_route_inconsistency(struct lnet_route *route)
 379 {
 380         if (!route->lr_single_hop && (int)route->lr_hops <= 1) {
 381                 CWARN("route %s->%s is detected to be multi-hop but hop count is set to %d\n",
 382                         libcfs_net2str(route->lr_net),
 383                         libcfs_nid2str(route->lr_gateway->lp_primary_nid),
 384                         (int) route->lr_hops);
 385         }
 386 }
 387
 388 static void
 389 lnet_set_route_hop_type(struct lnet_peer *gw, struct lnet_route *route)
 390 {
 391         struct lnet_peer_net *lpn;
 392         bool single_hop = false;
 393
 394         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 395                 if (route->lr_net == lpn->lpn_net_id) {
 396                         single_hop = true;
 397                         break;
 398                 }
 399         }
 400         route->lr_single_hop = single_hop;
 401         lnet_check_route_inconsistency(route);
 402 }
 403
 404 static inline void
 405 lnet_set_route_aliveness(struct lnet_route *route, bool alive)
 406 {
 407         /* Log when there's a state change */
 408         if (route->lr_alive != alive) {
 409                 CERROR("route to %s through %s has gone from %s to %s\n",
 410                        libcfs_net2str(route->lr_net),
 411                        libcfs_nid2str(route->lr_gateway->lp_primary_nid),
 412                        (route->lr_alive) ? "up" : "down",
 413                        alive ? "up" : "down");
 414                 route->lr_alive = alive;
 415         }
 416 }
 417
 418 void
 419 lnet_router_discovery_ping_reply(struct lnet_peer *lp)
 420 {
 421         struct lnet_ping_buffer *pbuf = lp->lp_data;
 422         struct lnet_peer_net *llpn;
 423         struct lnet_route *route;
 424         bool single_hop = false;
 425         bool net_up = false;
 426         unsigned lp_state;
 427         __u32 net;
 428         int i;
 429
 430
 431         spin_lock(&lp->lp_lock);
 432         lp_state = lp->lp_state;
 433
 434         /* only handle replies if discovery is disabled. */
 435         if (!lnet_is_discovery_disabled_locked(lp)) {
 436                 spin_unlock(&lp->lp_lock);
 437                 return;
 438         }
 439
 440         spin_unlock(&lp->lp_lock);
 441
 442         if (lp_state & LNET_PEER_PING_FAILED ||
 443             pbuf->pb_info.pi_features & LNET_PING_FEAT_RTE_DISABLED) {
 444                 CDEBUG(D_NET, "Set routes down for gw %s because %s %d\n",
 445                        libcfs_nid2str(lp->lp_primary_nid),
 446                        lp_state & LNET_PEER_PING_FAILED ? "ping failed" :
 447                        "route feature is disabled", lp->lp_ping_error);
 448                 /* If the ping failed or the peer has routing disabled then
 449                  * mark the routes served by this peer down
 450                  */
 451                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 452                         lnet_set_route_aliveness(route, false);
 453                 return;
 454         }
 455
 456         CDEBUG(D_NET, "Discovery is disabled. Processing reply for gw: %s:%d\n",
 457                libcfs_nid2str(lp->lp_primary_nid), pbuf->pb_info.pi_nnis);
 458
 459         /*
 460          * examine the ping response to determine if the routes on that
 461          * gateway should be declared alive.
 462          * The route is alive if:
 463          *  1. local network to reach the route is alive and
 464          *  2. route is single hop, avoid_async_router_failure is set and
 465          *     there exists at least one NI on the route's remote net
 466          */
 467         list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
 468                 llpn = lnet_peer_get_net_locked(lp, route->lr_lnet);
 469                 if (!llpn) {
 470                         lnet_set_route_aliveness(route, false);
 471                         continue;
 472                 }
 473
 474                 if (!lnet_is_gateway_net_alive(llpn)) {
 475                         lnet_set_route_aliveness(route, false);
 476                         continue;
 477                 }
 478
 479                 single_hop = net_up = false;
 480                 for (i = 1; i < pbuf->pb_info.pi_nnis; i++) {
 481                         net = LNET_NIDNET(pbuf->pb_info.pi_ni[i].ns_nid);
 482
 483                         if (route->lr_net == net) {
 484                                 single_hop = true;
 485                                 if (pbuf->pb_info.pi_ni[i].ns_status ==
 486                                     LNET_NI_STATUS_UP) {
 487                                         net_up = true;
 488                                         break;
 489                                 }
 490                         }
 491                 }
 492
 493                 route->lr_single_hop = single_hop;
 494                 if (avoid_asym_router_failure && single_hop)
 495                         lnet_set_route_aliveness(route, net_up);
 496                 else
 497                         lnet_set_route_aliveness(route, true);
 498
 499                 /*
 500                  * warn that the route is configured as single-hop but it
 501                  * really is multi-hop as far as we can tell.
 502                  */
 503                 lnet_check_route_inconsistency(route);
 504         }
 505 }
 506
 507 void
 508 lnet_router_discovery_complete(struct lnet_peer *lp)
 509 {
 510         struct lnet_peer_ni *lpni = NULL;
 511         struct lnet_route *route;
 512
 513         spin_lock(&lp->lp_lock);
 514         lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
 515         lp->lp_state |= LNET_PEER_RTR_DISCOVERED;
 516         lp->lp_alive = lp->lp_dc_error == 0;
 517         spin_unlock(&lp->lp_lock);
 518
 519         /* ping replies are being handled when discovery is disabled */
 520         if (lnet_is_discovery_disabled_locked(lp))
 521                 return;
 522
 523         if (!lp->lp_dc_error) {
 524                 /*
 525                 * mark single-hop routes.  If the remote net is not configured on
 526                 * the gateway we assume this is intentional and we mark the
 527                 * gateway as multi-hop
 528                 */
 529                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
 530                         lnet_set_route_aliveness(route, true);
 531                         lnet_set_route_hop_type(lp, route);
 532                 }
 533
 534                 return;
 535         }
 536
 537         /*
 538          * We do not send messages directly to the remote interfaces
 539          * of an LNet router. As such, we rely on the PING response
 540          * to determine the up/down status of these interfaces. If
 541          * a PING response is not receieved, or some other problem with
 542          * discovery occurs that prevents us from getting this status,
 543          * we assume all interfaces are down until we're able to
 544          * determine otherwise.
 545          */
 546         CDEBUG(D_NET, "%s: Router discovery failed %d\n",
 547                libcfs_nid2str(lp->lp_primary_nid), lp->lp_dc_error);
 548         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 549                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 550
 551         list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
 552                 lnet_set_route_aliveness(route, false);
 553 }
 554
 555 static void
 556 lnet_rtr_addref_locked(struct lnet_peer *lp)
 557 {
 558         LASSERT(lp->lp_rtr_refcount >= 0);
 559
 560         /* lnet_net_lock must be exclusively locked */
 561         lp->lp_rtr_refcount++;
 562         if (lp->lp_rtr_refcount == 1) {
 563                 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
 564                 /* addref for the_lnet.ln_routers */
 565                 lnet_peer_addref_locked(lp);
 566                 the_lnet.ln_routers_version++;
 567         }
 568 }
 569
 570 static void
 571 lnet_rtr_decref_locked(struct lnet_peer *lp)
 572 {
 573         LASSERT(atomic_read(&lp->lp_refcount) > 0);
 574         LASSERT(lp->lp_rtr_refcount > 0);
 575
 576         /* lnet_net_lock must be exclusively locked */
 577         lp->lp_rtr_refcount--;
 578         if (lp->lp_rtr_refcount == 0) {
 579                 LASSERT(list_empty(&lp->lp_routes));
 580
 581                 list_del(&lp->lp_rtr_list);
 582                 /* decref for the_lnet.ln_routers */
 583                 lnet_peer_decref_locked(lp);
 584                 the_lnet.ln_routers_version++;
 585         }
 586 }
 587
 588 struct lnet_remotenet *
 589 lnet_find_rnet_locked(__u32 net)
 590 {
 591         struct lnet_remotenet *rnet;
 592         struct list_head *tmp;
 593         struct list_head *rn_list;
 594
 595         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 596
 597         rn_list = lnet_net2rnethash(net);
 598         list_for_each(tmp, rn_list) {
 599                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 600
 601                 if (rnet->lrn_net == net)
 602                         return rnet;
 603         }
 604         return NULL;
 605 }
 606
 607 static void lnet_shuffle_seed(void)
 608 {
 609         static int seeded;
 610         struct lnet_ni *ni = NULL;
 611
 612         if (seeded)
 613                 return;
 614
 615         /* Nodes with small feet have little entropy
 616          * the NID for this node gives the most entropy in the low bits */
 617         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 618                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 619
 620         seeded = 1;
 621 }
 622
 623 /* NB expects LNET_LOCK held */
 624 static void
 625 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 626 {
 627         struct lnet_peer_net *lpn;
 628         unsigned int offset = 0;
 629         unsigned int len = 0;
 630         struct list_head *e;
 631
 632         lnet_shuffle_seed();
 633
 634         list_for_each(e, &rnet->lrn_routes)
 635                 len++;
 636
 637         /*
 638          * Randomly adding routes to the list is done to ensure that when
 639          * different nodes are using the same list of routers, they end up
 640          * preferring different routers.
 641          */
 642         offset = prandom_u32_max(len + 1);
 643         list_for_each(e, &rnet->lrn_routes) {
 644                 if (offset == 0)
 645                         break;
 646                 offset--;
 647         }
 648         list_add(&route->lr_list, e);
 649         /*
 650          * force a router check on the gateway to make sure the route is
 651          * alive
 652          */
 653         list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets,
 654                             lpn_peer_nets) {
 655                 lpn->lpn_rtrcheck_timestamp = 0;
 656         }
 657
 658         the_lnet.ln_remote_nets_version++;
 659
 660         /* add the route on the gateway list */
 661         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 662
 663         /* take a router reference count on the gateway */
 664         lnet_rtr_addref_locked(route->lr_gateway);
 665 }
 666
 667 int
 668 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 669                __u32 priority, __u32 sensitivity)
 670 {
 671         struct list_head *route_entry;
 672         struct lnet_remotenet *rnet;
 673         struct lnet_remotenet *rnet2;
 674         struct lnet_route *route;
 675         struct lnet_peer_ni *lpni;
 676         struct lnet_peer *gw;
 677         int add_route;
 678         int rc;
 679
 680         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 681                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 682
 683         if (gateway == LNET_NID_ANY ||
 684             gateway == LNET_NID_LO_0 ||
 685             net == LNET_NIDNET(LNET_NID_ANY) ||
 686             LNET_NETTYP(net) == LOLND ||
 687             LNET_NIDNET(gateway) == net ||
 688             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 689                 return -EINVAL;
 690
 691         /* it's a local network */
 692         if (lnet_islocalnet(net))
 693                 return -EEXIST;
 694
 695         if (!lnet_islocalnet(LNET_NIDNET(gateway))) {
 696                 CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n",
 697                        libcfs_nid2str(gateway),
 698                        libcfs_net2str(LNET_NIDNET(gateway)));
 699                 return -EHOSTUNREACH;
 700         }
 701
 702         /* Assume net, route, all new */
 703         LIBCFS_ALLOC(route, sizeof(*route));
 704         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 705         if (route == NULL || rnet == NULL) {
 706                 CERROR("Out of memory creating route %s %d %s\n",
 707                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 708                 if (route != NULL)
 709                         LIBCFS_FREE(route, sizeof(*route));
 710                 if (rnet != NULL)
 711                         LIBCFS_FREE(rnet, sizeof(*rnet));
 712                 return -ENOMEM;
 713         }
 714
 715         INIT_LIST_HEAD(&rnet->lrn_routes);
 716         rnet->lrn_net = net;
 717         /* store the local and remote net that the route represents */
 718         route->lr_lnet = LNET_NIDNET(gateway);
 719         route->lr_net = net;
 720         route->lr_nid = gateway;
 721         route->lr_priority = priority;
 722         route->lr_hops = hops;
 723
 724         lnet_net_lock(LNET_LOCK_EX);
 725
 726         /*
 727          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 728          * lose that once we're done
 729          */
 730         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 731         if (IS_ERR(lpni)) {
 732                 lnet_net_unlock(LNET_LOCK_EX);
 733
 734                 LIBCFS_FREE(route, sizeof(*route));
 735                 LIBCFS_FREE(rnet, sizeof(*rnet));
 736
 737                 rc = PTR_ERR(lpni);
 738                 CERROR("Error %d creating route %s %d %s\n", rc,
 739                         libcfs_net2str(net), hops,
 740                         libcfs_nid2str(gateway));
 741                 return rc;
 742         }
 743
 744         LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
 745         gw = lpni->lpni_peer_net->lpn_peer;
 746
 747         route->lr_gateway = gw;
 748
 749         rnet2 = lnet_find_rnet_locked(net);
 750         if (rnet2 == NULL) {
 751                 /* new network */
 752                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 753                 rnet2 = rnet;
 754         }
 755
 756         /* Search for a duplicate route (it's a NOOP if it is) */
 757         add_route = 1;
 758         list_for_each(route_entry, &rnet2->lrn_routes) {
 759                 struct lnet_route *route2;
 760
 761                 route2 = list_entry(route_entry, struct lnet_route, lr_list);
 762                 if (route2->lr_gateway == route->lr_gateway) {
 763                         add_route = 0;
 764                         break;
 765                 }
 766
 767                 /* our lookups must be true */
 768                 LASSERT(route2->lr_gateway->lp_primary_nid != 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 "
 784                               "Multi-Rail routing functionality\n");
 785         }
 786
 787         /*
 788          * get rid of the reference on the lpni.
 789          */
 790         lnet_peer_ni_decref_locked(lpni);
 791         lnet_net_unlock(LNET_LOCK_EX);
 792
 793         rc = 0;
 794
 795         if (!add_route) {
 796                 rc = -EEXIST;
 797                 LIBCFS_FREE(route, sizeof(*route));
 798         }
 799
 800         if (rnet != rnet2)
 801                 LIBCFS_FREE(rnet, sizeof(*rnet));
 802
 803         /* kick start the monitor thread to handle the added route */
 804         complete(&the_lnet.ln_mt_wait_complete);
 805
 806         return rc;
 807 }
 808
 809 void
 810 lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list,
 811                          struct list_head *zombies)
 812 {
 813         struct lnet_peer *gateway;
 814         struct lnet_route *route;
 815         struct lnet_route *tmp;
 816
 817         list_for_each_entry_safe(route, tmp, route_list, lr_list) {
 818                 gateway = route->lr_gateway;
 819                 if (gw_nid != LNET_NID_ANY &&
 820                     gw_nid != gateway->lp_primary_nid)
 821                         continue;
 822
 823                 /*
 824                  * move to zombie to delete outside the lock
 825                  * Note that this function is called with the
 826                  * ln_api_mutex held as well as the exclusive net
 827                  * lock. Adding to the remote net list happens
 828                  * under the same conditions. Same goes for the
 829                  * gateway router list
 830                  */
 831                 list_move(&route->lr_list, zombies);
 832                 the_lnet.ln_remote_nets_version++;
 833
 834                 list_del(&route->lr_gwlist);
 835                 lnet_rtr_decref_locked(gateway);
 836         }
 837 }
 838
 839 int
 840 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 841 {
 842         LIST_HEAD(rnet_zombies);
 843         struct lnet_remotenet *rnet;
 844         struct lnet_remotenet *tmp;
 845         struct list_head *rn_list;
 846         struct lnet_peer_ni *lpni;
 847         struct lnet_route *route;
 848         LIST_HEAD(zombies);
 849         struct lnet_peer *lp = NULL;
 850         int i = 0;
 851
 852         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 853                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 854
 855         /* NB Caller may specify either all routes via the given gateway
 856          * or a specific route entry actual NIDs) */
 857
 858         lnet_net_lock(LNET_LOCK_EX);
 859
 860         lpni = lnet_find_peer_ni_locked(gw_nid);
 861         if (lpni) {
 862                 lp = lpni->lpni_peer_net->lpn_peer;
 863                 LASSERT(lp);
 864                 gw_nid = lp->lp_primary_nid;
 865                 lnet_peer_ni_decref_locked(lpni);
 866         }
 867
 868         if (net != LNET_NIDNET(LNET_NID_ANY)) {
 869                 rnet = lnet_find_rnet_locked(net);
 870                 if (!rnet) {
 871                         lnet_net_unlock(LNET_LOCK_EX);
 872                         return -ENOENT;
 873                 }
 874                 lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 875                                          &zombies);
 876                 if (list_empty(&rnet->lrn_routes))
 877                         list_move(&rnet->lrn_list, &rnet_zombies);
 878                 goto delete_zombies;
 879         }
 880
 881         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 882                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 883
 884                 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
 885                         lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 886                                                  &zombies);
 887                         if (list_empty(&rnet->lrn_routes))
 888                                 list_move(&rnet->lrn_list, &rnet_zombies);
 889                 }
 890         }
 891
 892 delete_zombies:
 893         /*
 894          * check if there are any routes remaining on the gateway
 895          * If there are no more routes make sure to set the peer's
 896          * lp_disc_net_id to 0 (invalid), in case we add more routes in
 897          * the future on that gateway, then we start our discovery process
 898          * from scratch
 899          */
 900         if (lpni) {
 901                 if (list_empty(&lp->lp_routes))
 902                         lp->lp_disc_net_id = 0;
 903         }
 904
 905         lnet_net_unlock(LNET_LOCK_EX);
 906
 907         while (!list_empty(&zombies)) {
 908                 route = list_first_entry(&zombies, struct lnet_route, lr_list);
 909                 list_del(&route->lr_list);
 910                 LIBCFS_FREE(route, sizeof(*route));
 911         }
 912
 913         while (!list_empty(&rnet_zombies)) {
 914                 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
 915                                         lrn_list);
 916                 list_del(&rnet->lrn_list);
 917                 LIBCFS_FREE(rnet, sizeof(*rnet));
 918         }
 919
 920         return 0;
 921 }
 922
 923 void
 924 lnet_destroy_routes (void)
 925 {
 926         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 927 }
 928
 929 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 930 {
 931         struct lnet_rtrbufpool *rbp;
 932         int i, rc = -ENOENT, j;
 933
 934         if (the_lnet.ln_rtrpools == NULL)
 935                 return rc;
 936
 937
 938         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 939                 if (i != cpt)
 940                         continue;
 941
 942                 lnet_net_lock(i);
 943                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 944                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 945                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 946                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 947                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 948                 }
 949                 lnet_net_unlock(i);
 950                 rc = 0;
 951                 break;
 952         }
 953
 954         lnet_net_lock(LNET_LOCK_EX);
 955         pool_cfg->pl_routing = the_lnet.ln_routing;
 956         lnet_net_unlock(LNET_LOCK_EX);
 957
 958         return rc;
 959 }
 960
 961 int
 962 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 963                lnet_nid_t *gateway, __u32 *flags, __u32 *priority, __u32 *sensitivity)
 964 {
 965         struct lnet_remotenet *rnet;
 966         struct list_head *rn_list;
 967         struct lnet_route *route;
 968         struct list_head *e1;
 969         struct list_head *e2;
 970         int cpt;
 971         int i;
 972
 973         cpt = lnet_net_lock_current();
 974
 975         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 976                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 977                 list_for_each(e1, rn_list) {
 978                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 979
 980                         list_for_each(e2, &rnet->lrn_routes) {
 981                                 route = list_entry(e2, struct lnet_route,
 982                                                    lr_list);
 983
 984                                 if (idx-- == 0) {
 985                                         *net      = rnet->lrn_net;
 986                                         *gateway  = route->lr_nid;
 987                                         *hops     = route->lr_hops;
 988                                         *priority = route->lr_priority;
 989                                         *sensitivity = route->lr_gateway->
 990                                                 lp_health_sensitivity;
 991                                         if (lnet_is_route_alive(route))
 992                                                 *flags |= LNET_RT_ALIVE;
 993                                         else
 994                                                 *flags &= ~LNET_RT_ALIVE;
 995                                         if (route->lr_single_hop)
 996                                                 *flags &= ~LNET_RT_MULTI_HOP;
 997                                         else
 998                                                 *flags |= LNET_RT_MULTI_HOP;
 999                                         lnet_net_unlock(cpt);
1000                                         return 0;
1001                                 }
1002                         }
1003                 }
1004         }
1005
1006         lnet_net_unlock(cpt);
1007         return -ENOENT;
1008 }
1009
1010 static void
1011 lnet_wait_known_routerstate(void)
1012 {
1013         struct lnet_peer *rtr;
1014         struct list_head *entry;
1015         int all_known;
1016
1017         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
1018
1019         for (;;) {
1020                 int cpt = lnet_net_lock_current();
1021
1022                 all_known = 1;
1023                 list_for_each(entry, &the_lnet.ln_routers) {
1024                         rtr = list_entry(entry, struct lnet_peer,
1025                                          lp_rtr_list);
1026
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_real_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 inline bool
1132 lnet_router_checker_active(void)
1133 {
1134         /* Router Checker thread needs to run when routing is enabled in
1135          * order to call lnet_update_ni_status_locked() */
1136         if (the_lnet.ln_routing)
1137                 return true;
1138
1139         return !list_empty(&the_lnet.ln_routers) &&
1140                 alive_router_check_interval > 0;
1141 }
1142
1143 void
1144 lnet_check_routers(void)
1145 {
1146         struct lnet_peer_net *first_lpn = NULL;
1147         struct lnet_peer_net *lpn;
1148         struct lnet_peer_ni *lpni;
1149         struct list_head *entry;
1150         struct lnet_peer *rtr;
1151         bool push = false;
1152         bool found_lpn;
1153         __u64 version;
1154         __u32 net_id;
1155         time64_t now;
1156         int cpt;
1157         int rc;
1158
1159         cpt = lnet_net_lock_current();
1160 rescan:
1161         version = the_lnet.ln_routers_version;
1162
1163         list_for_each(entry, &the_lnet.ln_routers) {
1164                 rtr = list_entry(entry, struct lnet_peer,
1165                                  lp_rtr_list);
1166
1167                 now = ktime_get_real_seconds();
1168
1169                 /*
1170                  * only discover the router if we've passed
1171                  * alive_router_check_interval seconds. Some of the router
1172                  * interfaces could be down and in that case they would be
1173                  * undergoing recovery separately from this discovery.
1174                  */
1175                 /* find next peer net which is also local */
1176                 net_id = rtr->lp_disc_net_id;
1177                 do {
1178                         lpn = lnet_get_next_peer_net_locked(rtr, net_id);
1179                         if (!lpn) {
1180                                 CERROR("gateway %s has no networks\n",
1181                                 libcfs_nid2str(rtr->lp_primary_nid));
1182                                 break;
1183                         }
1184                         if (first_lpn == lpn)
1185                                 break;
1186                         if (!first_lpn)
1187                                 first_lpn = lpn;
1188                         found_lpn = lnet_islocalnet_locked(lpn->lpn_net_id);
1189                         net_id = lpn->lpn_net_id;
1190                 } while (!found_lpn);
1191
1192                 if (!found_lpn || !lpn) {
1193                         CERROR("no local network found for gateway %s\n",
1194                                libcfs_nid2str(rtr->lp_primary_nid));
1195                         continue;
1196                 }
1197
1198                 if (now - lpn->lpn_rtrcheck_timestamp <
1199                     alive_router_check_interval / lnet_current_net_count)
1200                        continue;
1201
1202                 /*
1203                  * If we're currently discovering the peer then don't
1204                  * issue another discovery
1205                  */
1206                 spin_lock(&rtr->lp_lock);
1207                 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY) {
1208                         spin_unlock(&rtr->lp_lock);
1209                         continue;
1210                 }
1211                 /* make sure we fully discover the router */
1212                 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
1213                 rtr->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH |
1214                         LNET_PEER_RTR_DISCOVERY;
1215                 spin_unlock(&rtr->lp_lock);
1216
1217                 /* find the peer_ni associated with the primary NID */
1218                 lpni = lnet_peer_get_ni_locked(rtr, rtr->lp_primary_nid);
1219                 if (!lpni) {
1220                         CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
1221                                libcfs_nid2str(rtr->lp_primary_nid));
1222                         continue;
1223                 }
1224                 lnet_peer_ni_addref_locked(lpni);
1225
1226                 /* specify the net to use */
1227                 rtr->lp_disc_net_id = lpn->lpn_net_id;
1228
1229                 /* discover the router */
1230                 CDEBUG(D_NET, "discover %s, cpt = %d\n",
1231                        libcfs_nid2str(lpni->lpni_nid), cpt);
1232                 rc = lnet_discover_peer_locked(lpni, cpt, false);
1233
1234                 /* decrement ref count acquired by find_peer_ni_locked() */
1235                 lnet_peer_ni_decref_locked(lpni);
1236
1237                 if (!rc)
1238                         lpn->lpn_rtrcheck_timestamp = now;
1239                 else
1240                         CERROR("Failed to discover router %s\n",
1241                                libcfs_nid2str(rtr->lp_primary_nid));
1242
1243                 /* NB dropped lock */
1244                 if (version != the_lnet.ln_routers_version) {
1245                         /* the routers list has changed */
1246                         goto rescan;
1247                 }
1248         }
1249
1250         if (the_lnet.ln_routing)
1251                 push = lnet_update_ni_status_locked();
1252
1253         lnet_net_unlock(cpt);
1254
1255         /* if the status of the ni changed update the peers */
1256         if (push)
1257                 lnet_push_update_to_peers(1);
1258 }
1259
1260 void
1261 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1262 {
1263         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1264
1265         while (--npages >= 0)
1266                 __free_page(rb->rb_kiov[npages].bv_page);
1267
1268         LIBCFS_FREE(rb, sz);
1269 }
1270
1271 static struct lnet_rtrbuf *
1272 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1273 {
1274         int            npages = rbp->rbp_npages;
1275         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1276         struct page   *page;
1277         struct lnet_rtrbuf *rb;
1278         int            i;
1279
1280         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1281         if (rb == NULL)
1282                 return NULL;
1283
1284         rb->rb_pool = rbp;
1285
1286         for (i = 0; i < npages; i++) {
1287                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1288                                           GFP_KERNEL | __GFP_ZERO);
1289                 if (page == NULL) {
1290                         while (--i >= 0)
1291                                 __free_page(rb->rb_kiov[i].bv_page);
1292
1293                         LIBCFS_FREE(rb, sz);
1294                         return NULL;
1295                 }
1296
1297                 rb->rb_kiov[i].bv_len = PAGE_SIZE;
1298                 rb->rb_kiov[i].bv_offset = 0;
1299                 rb->rb_kiov[i].bv_page = page;
1300         }
1301
1302         return rb;
1303 }
1304
1305 static void
1306 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1307 {
1308         int npages = rbp->rbp_npages;
1309         struct lnet_rtrbuf *rb;
1310         LIST_HEAD(tmp);
1311
1312         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1313                 return;
1314
1315         lnet_net_lock(cpt);
1316         list_splice_init(&rbp->rbp_msgs, &tmp);
1317         lnet_drop_routed_msgs_locked(&tmp, cpt);
1318         list_splice_init(&rbp->rbp_bufs, &tmp);
1319         rbp->rbp_req_nbuffers = 0;
1320         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1321         rbp->rbp_mincredits = 0;
1322         lnet_net_unlock(cpt);
1323
1324         /* Free buffers on the free list. */
1325         while (!list_empty(&tmp)) {
1326                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1327                 list_del(&rb->rb_list);
1328                 lnet_destroy_rtrbuf(rb, npages);
1329         }
1330 }
1331
1332 static int
1333 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1334 {
1335         LIST_HEAD(rb_list);
1336         struct lnet_rtrbuf *rb;
1337         int             num_rb;
1338         int             num_buffers = 0;
1339         int             old_req_nbufs;
1340         int             npages = rbp->rbp_npages;
1341
1342         lnet_net_lock(cpt);
1343         /* If we are called for less buffers than already in the pool, we
1344          * just lower the req_nbuffers number and excess buffers will be
1345          * thrown away as they are returned to the free list.  Credits
1346          * then get adjusted as well.
1347          * If we already have enough buffers allocated to serve the
1348          * increase requested, then we can treat that the same way as we
1349          * do the decrease. */
1350         num_rb = nbufs - rbp->rbp_nbuffers;
1351         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1352                 rbp->rbp_req_nbuffers = nbufs;
1353                 lnet_net_unlock(cpt);
1354                 return 0;
1355         }
1356         /* store the older value of rbp_req_nbuffers and then set it to
1357          * the new request to prevent lnet_return_rx_credits_locked() from
1358          * freeing buffers that we need to keep around */
1359         old_req_nbufs = rbp->rbp_req_nbuffers;
1360         rbp->rbp_req_nbuffers = nbufs;
1361         lnet_net_unlock(cpt);
1362
1363         /* allocate the buffers on a local list first.  If all buffers are
1364          * allocated successfully then join this list to the rbp buffer
1365          * list.  If not then free all allocated buffers. */
1366         while (num_rb-- > 0) {
1367                 rb = lnet_new_rtrbuf(rbp, cpt);
1368                 if (rb == NULL) {
1369                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1370                                nbufs, npages);
1371
1372                         lnet_net_lock(cpt);
1373                         rbp->rbp_req_nbuffers = old_req_nbufs;
1374                         lnet_net_unlock(cpt);
1375
1376                         goto failed;
1377                 }
1378
1379                 list_add(&rb->rb_list, &rb_list);
1380                 num_buffers++;
1381         }
1382
1383         lnet_net_lock(cpt);
1384
1385         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1386         rbp->rbp_nbuffers += num_buffers;
1387         rbp->rbp_credits += num_buffers;
1388         rbp->rbp_mincredits = rbp->rbp_credits;
1389         /* We need to schedule blocked msg using the newly
1390          * added buffers. */
1391         while (!list_empty(&rbp->rbp_bufs) &&
1392                !list_empty(&rbp->rbp_msgs))
1393                 lnet_schedule_blocked_locked(rbp);
1394
1395         lnet_net_unlock(cpt);
1396
1397         return 0;
1398
1399 failed:
1400         while (!list_empty(&rb_list)) {
1401                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1402                 list_del(&rb->rb_list);
1403                 lnet_destroy_rtrbuf(rb, npages);
1404         }
1405
1406         return -ENOMEM;
1407 }
1408
1409 static void
1410 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1411 {
1412         INIT_LIST_HEAD(&rbp->rbp_msgs);
1413         INIT_LIST_HEAD(&rbp->rbp_bufs);
1414
1415         rbp->rbp_npages = npages;
1416         rbp->rbp_credits = 0;
1417         rbp->rbp_mincredits = 0;
1418 }
1419
1420 void
1421 lnet_rtrpools_free(int keep_pools)
1422 {
1423         struct lnet_rtrbufpool *rtrp;
1424         int               i;
1425
1426         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1427                 return;
1428
1429         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1430                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1431                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1432                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1433         }
1434
1435         if (!keep_pools) {
1436                 cfs_percpt_free(the_lnet.ln_rtrpools);
1437                 the_lnet.ln_rtrpools = NULL;
1438         }
1439 }
1440
1441 static int
1442 lnet_nrb_tiny_calculate(void)
1443 {
1444         int     nrbs = LNET_NRB_TINY;
1445
1446         if (tiny_router_buffers < 0) {
1447                 LCONSOLE_ERROR_MSG(0x10c,
1448                                    "tiny_router_buffers=%d invalid when "
1449                                    "routing enabled\n", tiny_router_buffers);
1450                 return -EINVAL;
1451         }
1452
1453         if (tiny_router_buffers > 0)
1454                 nrbs = tiny_router_buffers;
1455
1456         nrbs /= LNET_CPT_NUMBER;
1457         return max(nrbs, LNET_NRB_TINY_MIN);
1458 }
1459
1460 static int
1461 lnet_nrb_small_calculate(void)
1462 {
1463         int     nrbs = LNET_NRB_SMALL;
1464
1465         if (small_router_buffers < 0) {
1466                 LCONSOLE_ERROR_MSG(0x10c,
1467                                    "small_router_buffers=%d invalid when "
1468                                    "routing enabled\n", small_router_buffers);
1469                 return -EINVAL;
1470         }
1471
1472         if (small_router_buffers > 0)
1473                 nrbs = small_router_buffers;
1474
1475         nrbs /= LNET_CPT_NUMBER;
1476         return max(nrbs, LNET_NRB_SMALL_MIN);
1477 }
1478
1479 static int
1480 lnet_nrb_large_calculate(void)
1481 {
1482         int     nrbs = LNET_NRB_LARGE;
1483
1484         if (large_router_buffers < 0) {
1485                 LCONSOLE_ERROR_MSG(0x10c,
1486                                    "large_router_buffers=%d invalid when "
1487                                    "routing enabled\n", large_router_buffers);
1488                 return -EINVAL;
1489         }
1490
1491         if (large_router_buffers > 0)
1492                 nrbs = large_router_buffers;
1493
1494         nrbs /= LNET_CPT_NUMBER;
1495         return max(nrbs, LNET_NRB_LARGE_MIN);
1496 }
1497
1498 int
1499 lnet_rtrpools_alloc(int im_a_router)
1500 {
1501         struct lnet_rtrbufpool *rtrp;
1502         int     nrb_tiny;
1503         int     nrb_small;
1504         int     nrb_large;
1505         int     rc;
1506         int     i;
1507
1508         if (!strcmp(forwarding, "")) {
1509                 /* not set either way */
1510                 if (!im_a_router)
1511                         return 0;
1512         } else if (!strcmp(forwarding, "disabled")) {
1513                 /* explicitly disabled */
1514                 return 0;
1515         } else if (!strcmp(forwarding, "enabled")) {
1516                 /* explicitly enabled */
1517         } else {
1518                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1519                                    "'enabled' or 'disabled'\n");
1520                 return -EINVAL;
1521         }
1522
1523         nrb_tiny = lnet_nrb_tiny_calculate();
1524         if (nrb_tiny < 0)
1525                 return -EINVAL;
1526
1527         nrb_small = lnet_nrb_small_calculate();
1528         if (nrb_small < 0)
1529                 return -EINVAL;
1530
1531         nrb_large = lnet_nrb_large_calculate();
1532         if (nrb_large < 0)
1533                 return -EINVAL;
1534
1535         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1536                                                 LNET_NRBPOOLS *
1537                                                 sizeof(struct lnet_rtrbufpool));
1538         if (the_lnet.ln_rtrpools == NULL) {
1539                 LCONSOLE_ERROR_MSG(0x10c,
1540                                    "Failed to initialize router buffe pool\n");
1541                 return -ENOMEM;
1542         }
1543
1544         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1545                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1546                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1547                                               nrb_tiny, i);
1548                 if (rc != 0)
1549                         goto failed;
1550
1551                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1552                                   LNET_NRB_SMALL_PAGES);
1553                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1554                                               nrb_small, i);
1555                 if (rc != 0)
1556                         goto failed;
1557
1558                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1559                                   LNET_NRB_LARGE_PAGES);
1560                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1561                                               nrb_large, i);
1562                 if (rc != 0)
1563                         goto failed;
1564         }
1565
1566         lnet_net_lock(LNET_LOCK_EX);
1567         the_lnet.ln_routing = 1;
1568         lnet_net_unlock(LNET_LOCK_EX);
1569         complete(&the_lnet.ln_mt_wait_complete);
1570         return 0;
1571
1572  failed:
1573         lnet_rtrpools_free(0);
1574         return rc;
1575 }
1576
1577 static int
1578 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1579 {
1580         int nrb = 0;
1581         int rc = 0;
1582         int i;
1583         struct lnet_rtrbufpool *rtrp;
1584
1585         /* If the provided values for each buffer pool are different than the
1586          * configured values, we need to take action. */
1587         if (tiny >= 0) {
1588                 tiny_router_buffers = tiny;
1589                 nrb = lnet_nrb_tiny_calculate();
1590                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1591                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1592                                                       nrb, i);
1593                         if (rc != 0)
1594                                 return rc;
1595                 }
1596         }
1597         if (small >= 0) {
1598                 small_router_buffers = small;
1599                 nrb = lnet_nrb_small_calculate();
1600                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1601                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1602                                                       nrb, i);
1603                         if (rc != 0)
1604                                 return rc;
1605                 }
1606         }
1607         if (large >= 0) {
1608                 large_router_buffers = large;
1609                 nrb = lnet_nrb_large_calculate();
1610                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1611                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1612                                                       nrb, i);
1613                         if (rc != 0)
1614                                 return rc;
1615                 }
1616         }
1617
1618         return 0;
1619 }
1620
1621 int
1622 lnet_rtrpools_adjust(int tiny, int small, int large)
1623 {
1624         /* this function doesn't revert the changes if adding new buffers
1625          * failed.  It's up to the user space caller to revert the
1626          * changes. */
1627
1628         if (!the_lnet.ln_routing)
1629                 return 0;
1630
1631         return lnet_rtrpools_adjust_helper(tiny, small, large);
1632 }
1633
1634 int
1635 lnet_rtrpools_enable(void)
1636 {
1637         int rc = 0;
1638
1639         if (the_lnet.ln_routing)
1640                 return 0;
1641
1642         if (the_lnet.ln_rtrpools == NULL)
1643                 /* If routing is turned off, and we have never
1644                  * initialized the pools before, just call the
1645                  * standard buffer pool allocation routine as
1646                  * if we are just configuring this for the first
1647                  * time. */
1648                 rc = lnet_rtrpools_alloc(1);
1649         else
1650                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1651         if (rc != 0)
1652                 return rc;
1653
1654         lnet_net_lock(LNET_LOCK_EX);
1655         the_lnet.ln_routing = 1;
1656
1657         the_lnet.ln_ping_target->pb_info.pi_features &=
1658                 ~LNET_PING_FEAT_RTE_DISABLED;
1659         lnet_net_unlock(LNET_LOCK_EX);
1660
1661         if (lnet_peer_discovery_disabled)
1662                 CWARN("Consider turning discovery on to enable full "
1663                       "Multi-Rail routing functionality\n");
1664
1665         return rc;
1666 }
1667
1668 void
1669 lnet_rtrpools_disable(void)
1670 {
1671         if (!the_lnet.ln_routing)
1672                 return;
1673
1674         lnet_net_lock(LNET_LOCK_EX);
1675         the_lnet.ln_routing = 0;
1676         the_lnet.ln_ping_target->pb_info.pi_features |=
1677                 LNET_PING_FEAT_RTE_DISABLED;
1678
1679         tiny_router_buffers = 0;
1680         small_router_buffers = 0;
1681         large_router_buffers = 0;
1682         lnet_net_unlock(LNET_LOCK_EX);
1683         lnet_rtrpools_free(1);
1684 }
1685
1686 static inline void
1687 lnet_notify_peer_down(struct lnet_ni *ni, lnet_nid_t nid)
1688 {
1689         if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1690                 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1691 }
1692
1693 /*
1694  * ni: local NI used to communicate with the peer
1695  * nid: peer NID
1696  * alive: true if peer is alive, false otherwise
1697  * reset: reset health value. This is requested by the LND.
1698  * when: notificaiton time.
1699  */
1700 int
1701 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset,
1702             time64_t when)
1703 {
1704         struct lnet_peer_ni *lpni = NULL;
1705         struct lnet_route *route;
1706         struct lnet_peer *lp;
1707         time64_t now = ktime_get_seconds();
1708         int cpt;
1709
1710         LASSERT(!in_interrupt());
1711
1712         CDEBUG(D_NET, "%s notifying %s: %s\n",
1713                (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1714                libcfs_nid2str(nid), alive ? "up" : "down");
1715
1716         if (ni != NULL &&
1717             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1718                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1719                       libcfs_nid2str(nid), alive ? "birth" : "death",
1720                       libcfs_nid2str(ni->ni_nid));
1721                 return -EINVAL;
1722         }
1723
1724         /* can't do predictions... */
1725         if (when > now) {
1726                 CWARN("Ignoring prediction from %s of %s %s "
1727                       "%lld seconds in the future\n",
1728                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1729                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1730                 return -EINVAL;
1731         }
1732
1733         if (ni != NULL && !alive &&             /* LND telling me she's down */
1734             !auto_down) {                       /* auto-down disabled */
1735                 CDEBUG(D_NET, "Auto-down disabled\n");
1736                 return 0;
1737         }
1738
1739         /* must lock 0 since this is used for synchronization */
1740         lnet_net_lock(0);
1741
1742         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1743                 lnet_net_unlock(0);
1744                 return -ESHUTDOWN;
1745         }
1746
1747         lpni = lnet_find_peer_ni_locked(nid);
1748         if (lpni == NULL) {
1749                 /* nid not found */
1750                 lnet_net_unlock(0);
1751                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1752                 return 0;
1753         }
1754
1755         if (alive) {
1756                 if (reset) {
1757                         lpni->lpni_ns_status = LNET_NI_STATUS_UP;
1758                         lnet_set_lpni_healthv_locked(lpni,
1759                                                      LNET_MAX_HEALTH_VALUE);
1760                 } else {
1761                         __u32 sensitivity = lpni->lpni_peer_net->
1762                                         lpn_peer->lp_health_sensitivity;
1763
1764                         lnet_inc_lpni_healthv_locked(lpni,
1765                                         (sensitivity) ? sensitivity :
1766                                         lnet_health_sensitivity);
1767                 }
1768         } else if (reset) {
1769                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
1770         }
1771
1772         /* recalculate aliveness */
1773         alive = lnet_is_peer_ni_alive(lpni);
1774         lnet_net_unlock(0);
1775
1776         if (ni != NULL && !alive)
1777                 lnet_notify_peer_down(ni, lpni->lpni_nid);
1778
1779         cpt = lpni->lpni_cpt;
1780         lnet_net_lock(cpt);
1781         lnet_peer_ni_decref_locked(lpni);
1782         if (lpni && lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer) {
1783                 lp = lpni->lpni_peer_net->lpn_peer;
1784                 lp->lp_alive = alive;
1785                 list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
1786                         lnet_set_route_aliveness(route, alive);
1787         }
1788         lnet_net_unlock(cpt);
1789
1790         return 0;
1791 }
1792 EXPORT_SYMBOL(lnet_notify);