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 alive_router_check_interval = 60;
  82 module_param(alive_router_check_interval, int, 0644);
  83 MODULE_PARM_DESC(alive_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
  84
  85 static int router_ping_timeout = 50;
  86 module_param(router_ping_timeout, int, 0644);
  87 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
  88
  89 /*
  90  * A value between 0 and 100. 0 meaning that even if router's interfaces
  91  * have the worse health still consider the gateway usable.
  92  * 100 means that at least one interface on the route's remote net is 100%
  93  * healthy to consider the route alive.
  94  * The default is set to 100 to ensure we maintain the original behavior.
  95  */
  96 unsigned int router_sensitivity_percentage = 100;
  97 static int rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
  98 static struct kernel_param_ops param_ops_rtr_sensitivity = {
  99         .set = rtr_sensitivity_set,
 100         .get = param_get_int,
 101 };
 102 #define param_check_rtr_sensitivity(name, p) \
 103                 __param_check(name, p, int)
 104 #ifdef HAVE_KERNEL_PARAM_OPS
 105 module_param(router_sensitivity_percentage, rtr_sensitivity, S_IRUGO|S_IWUSR);
 106 #else
 107 module_param_call(router_sensitivity_percentage, rtr_sensitivity_set, param_get_int,
 108                   &router_sensitivity_percentage, S_IRUGO|S_IWUSR);
 109 #endif
 110 MODULE_PARM_DESC(router_sensitivity_percentage,
 111                 "How healthy a gateway should be to be used in percent");
 112
 113 static int
 114 rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 115 {
 116         int rc;
 117         unsigned *sen = (unsigned *)kp->arg;
 118         unsigned long value;
 119
 120         rc = kstrtoul(val, 0, &value);
 121         if (rc) {
 122                 CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
 123                 return rc;
 124         }
 125
 126         if (value < 0 || value > 100) {
 127                 CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
 128                 return -EINVAL;
 129         }
 130
 131         /*
 132          * The purpose of locking the api_mutex here is to ensure that
 133          * the correct value ends up stored properly.
 134          */
 135         mutex_lock(&the_lnet.ln_api_mutex);
 136
 137         *sen = value;
 138
 139         mutex_unlock(&the_lnet.ln_api_mutex);
 140
 141         return 0;
 142 }
 143
 144 void
 145 lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target)
 146 {
 147         struct lnet_route *route;
 148
 149         lnet_net_lock(LNET_LOCK_EX);
 150         target->lp_rtr_refcount += src->lp_rtr_refcount;
 151         /* move the list of queued messages to the new peer */
 152         list_splice_init(&src->lp_rtrq, &target->lp_rtrq);
 153         /* move all the routes that reference the peer */
 154         list_splice_init(&src->lp_routes, &target->lp_routes);
 155         /* update all the routes to point to the new peer */
 156         list_for_each_entry(route, &target->lp_routes, lr_gwlist)
 157                 route->lr_gateway = target;
 158         /* remove the old peer from the ln_routers list */
 159         list_del_init(&src->lp_rtr_list);
 160         /* add the new peer to the ln_routers list */
 161         if (list_empty(&target->lp_rtr_list)) {
 162                 lnet_peer_addref_locked(target);
 163                 list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers);
 164         }
 165         /* reset the ref count on the old peer and decrement its ref count */
 166         src->lp_rtr_refcount = 0;
 167         lnet_peer_decref_locked(src);
 168         /* update the router version */
 169         the_lnet.ln_routers_version++;
 170         lnet_net_unlock(LNET_LOCK_EX);
 171 }
 172
 173 int
 174 lnet_peers_start_down(void)
 175 {
 176         return check_routers_before_use;
 177 }
 178
 179 /*
 180  * A net is alive if at least one gateway NI on the network is alive.
 181  */
 182 static bool
 183 lnet_is_gateway_net_alive(struct lnet_peer_net *lpn)
 184 {
 185         struct lnet_peer_ni *lpni;
 186
 187         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 188                 if (lnet_is_peer_ni_alive(lpni))
 189                         return true;
 190         }
 191
 192         return false;
 193 }
 194
 195 /*
 196  * a gateway is alive only if all its nets are alive
 197  * called with cpt lock held
 198  */
 199 bool lnet_is_gateway_alive(struct lnet_peer *gw)
 200 {
 201         struct lnet_peer_net *lpn;
 202
 203         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 204                 if (!lnet_is_gateway_net_alive(lpn))
 205                         return false;
 206         }
 207
 208         return true;
 209 }
 210
 211 /*
 212  * lnet_is_route_alive() needs to be called with cpt lock held
 213  * A route is alive if the gateway can route between the local network and
 214  * the remote network of the route.
 215  * This means at least one NI is alive on each of the local and remote
 216  * networks of the gateway.
 217  */
 218 bool lnet_is_route_alive(struct lnet_route *route)
 219 {
 220         struct lnet_peer *gw = route->lr_gateway;
 221         struct lnet_peer_net *llpn;
 222         struct lnet_peer_net *rlpn;
 223         bool route_alive;
 224
 225         /*
 226          * check the gateway's interfaces on the route rnet to make sure
 227          * that the gateway is viable.
 228          */
 229         llpn = lnet_peer_get_net_locked(gw, route->lr_lnet);
 230         if (!llpn)
 231                 return false;
 232
 233         route_alive = lnet_is_gateway_net_alive(llpn);
 234
 235         if (avoid_asym_router_failure) {
 236                 rlpn = lnet_peer_get_net_locked(gw, route->lr_net);
 237                 if (!rlpn)
 238                         return false;
 239                 route_alive = route_alive &&
 240                               lnet_is_gateway_net_alive(rlpn);
 241         }
 242
 243         if (!route_alive)
 244                 return route_alive;
 245
 246         spin_lock(&gw->lp_lock);
 247         if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) {
 248                 if (gw->lp_rtr_refcount > 0)
 249                         CERROR("peer %s is being used as a gateway but routing feature is not turned on\n",
 250                                libcfs_nid2str(gw->lp_primary_nid));
 251                 route_alive = false;
 252         }
 253         spin_unlock(&gw->lp_lock);
 254
 255         return route_alive;
 256 }
 257
 258 void
 259 lnet_consolidate_routes_locked(struct lnet_peer *orig_lp,
 260                                struct lnet_peer *new_lp)
 261 {
 262         struct lnet_peer_ni *lpni;
 263         struct lnet_route *route;
 264
 265         /*
 266          * Although a route is correlated with a peer, but when it's added
 267          * a specific NID is used. That NID refers to a peer_ni within
 268          * a peer. There could be other peer_nis on the same net, which
 269          * can be used to send to that gateway. However when we are
 270          * consolidating gateways because of discovery, the nid used to
 271          * add the route might've moved between gateway peers. In this
 272          * case we want to move the route to the new gateway as well. The
 273          * intent here is not to confuse the user who added the route.
 274          */
 275         list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) {
 276                 lpni = lnet_peer_get_ni_locked(orig_lp, route->lr_nid);
 277                 if (!lpni) {
 278                         lnet_net_lock(LNET_LOCK_EX);
 279                         list_move(&route->lr_gwlist, &new_lp->lp_routes);
 280                         lnet_net_unlock(LNET_LOCK_EX);
 281                 }
 282         }
 283
 284 }
 285
 286 void
 287 lnet_router_discovery_complete(struct lnet_peer *lp)
 288 {
 289         struct lnet_peer_ni *lpni = NULL;
 290
 291         spin_lock(&lp->lp_lock);
 292         lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
 293         spin_unlock(&lp->lp_lock);
 294
 295         /*
 296          * Router discovery successful? All peer information would've been
 297          * updated already. No need to do any more processing
 298          */
 299         if (!lp->lp_dc_error)
 300                 return;
 301         /*
 302          * discovery failed? then we need to set the status of each lpni
 303          * to DOWN. It will be updated the next time we discover the
 304          * router. For router peer NIs not on local networks, we never send
 305          * messages directly to them, so their health will always remain
 306          * at maximum. We can only tell if they are up or down from the
 307          * status returned in the PING response. If we fail to get that
 308          * status in our scheduled router discovery, then we'll assume
 309          * it's down until we're told otherwise.
 310          */
 311         CDEBUG(D_NET, "%s: Router discovery failed %d\n",
 312                libcfs_nid2str(lp->lp_primary_nid), lp->lp_dc_error);
 313         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 314                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 315 }
 316
 317 static void
 318 lnet_rtr_addref_locked(struct lnet_peer *lp)
 319 {
 320         LASSERT(lp->lp_rtr_refcount >= 0);
 321
 322         /* lnet_net_lock must be exclusively locked */
 323         lp->lp_rtr_refcount++;
 324         if (lp->lp_rtr_refcount == 1) {
 325                 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
 326                 /* addref for the_lnet.ln_routers */
 327                 lnet_peer_addref_locked(lp);
 328                 the_lnet.ln_routers_version++;
 329         }
 330 }
 331
 332 static void
 333 lnet_rtr_decref_locked(struct lnet_peer *lp)
 334 {
 335         LASSERT(atomic_read(&lp->lp_refcount) > 0);
 336         LASSERT(lp->lp_rtr_refcount > 0);
 337
 338         /* lnet_net_lock must be exclusively locked */
 339         lp->lp_rtr_refcount--;
 340         if (lp->lp_rtr_refcount == 0) {
 341                 LASSERT(list_empty(&lp->lp_routes));
 342
 343                 list_del(&lp->lp_rtr_list);
 344                 /* decref for the_lnet.ln_routers */
 345                 lnet_peer_decref_locked(lp);
 346                 the_lnet.ln_routers_version++;
 347         }
 348 }
 349
 350 struct lnet_remotenet *
 351 lnet_find_rnet_locked(__u32 net)
 352 {
 353         struct lnet_remotenet *rnet;
 354         struct list_head *tmp;
 355         struct list_head *rn_list;
 356
 357         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 358
 359         rn_list = lnet_net2rnethash(net);
 360         list_for_each(tmp, rn_list) {
 361                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 362
 363                 if (rnet->lrn_net == net)
 364                         return rnet;
 365         }
 366         return NULL;
 367 }
 368
 369 static void lnet_shuffle_seed(void)
 370 {
 371         static int seeded;
 372         struct lnet_ni *ni = NULL;
 373
 374         if (seeded)
 375                 return;
 376
 377         /* Nodes with small feet have little entropy
 378          * the NID for this node gives the most entropy in the low bits */
 379         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 380                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 381
 382         seeded = 1;
 383         return;
 384 }
 385
 386 /* NB expects LNET_LOCK held */
 387 static void
 388 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 389 {
 390         unsigned int len = 0;
 391         unsigned int offset = 0;
 392         struct list_head *e;
 393
 394         lnet_shuffle_seed();
 395
 396         list_for_each(e, &rnet->lrn_routes)
 397                 len++;
 398
 399         /*
 400          * Randomly adding routes to the list is done to ensure that when
 401          * different nodes are using the same list of routers, they end up
 402          * preferring different routers.
 403          */
 404         offset = cfs_rand() % (len + 1);
 405         list_for_each(e, &rnet->lrn_routes) {
 406                 if (offset == 0)
 407                         break;
 408                 offset--;
 409         }
 410         list_add(&route->lr_list, e);
 411         /*
 412          * force a router check on the gateway to make sure the route is
 413          * alive
 414          */
 415         route->lr_gateway->lp_rtrcheck_timestamp = 0;
 416
 417         the_lnet.ln_remote_nets_version++;
 418
 419         /* add the route on the gateway list */
 420         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 421
 422         /* take a router reference count on the gateway */
 423         lnet_rtr_addref_locked(route->lr_gateway);
 424 }
 425
 426 int
 427 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 428                __u32 priority, __u32 sensitivity)
 429 {
 430         struct list_head *route_entry;
 431         struct lnet_remotenet *rnet;
 432         struct lnet_remotenet *rnet2;
 433         struct lnet_route *route;
 434         struct lnet_peer_ni *lpni;
 435         struct lnet_peer *gw;
 436         int add_route;
 437         int rc;
 438
 439         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 440                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 441
 442         if (gateway == LNET_NID_ANY ||
 443             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 444             net == LNET_NIDNET(LNET_NID_ANY) ||
 445             LNET_NETTYP(net) == LOLND ||
 446             LNET_NIDNET(gateway) == net ||
 447             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 448                 return -EINVAL;
 449
 450         /* it's a local network */
 451         if (lnet_islocalnet(net))
 452                 return -EEXIST;
 453
 454         /* Assume net, route, all new */
 455         LIBCFS_ALLOC(route, sizeof(*route));
 456         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 457         if (route == NULL || rnet == NULL) {
 458                 CERROR("Out of memory creating route %s %d %s\n",
 459                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 460                 if (route != NULL)
 461                         LIBCFS_FREE(route, sizeof(*route));
 462                 if (rnet != NULL)
 463                         LIBCFS_FREE(rnet, sizeof(*rnet));
 464                 return -ENOMEM;
 465         }
 466
 467         INIT_LIST_HEAD(&rnet->lrn_routes);
 468         rnet->lrn_net = net;
 469         /* store the local and remote net that the route represents */
 470         route->lr_lnet = LNET_NIDNET(gateway);
 471         route->lr_net = net;
 472         route->lr_nid = gateway;
 473         route->lr_priority = priority;
 474         route->lr_hops = hops;
 475
 476         lnet_net_lock(LNET_LOCK_EX);
 477
 478         /*
 479          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 480          * lose that once we're done
 481          */
 482         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 483         if (IS_ERR(lpni)) {
 484                 lnet_net_unlock(LNET_LOCK_EX);
 485
 486                 LIBCFS_FREE(route, sizeof(*route));
 487                 LIBCFS_FREE(rnet, sizeof(*rnet));
 488
 489                 rc = PTR_ERR(lpni);
 490                 CERROR("Error %d creating route %s %d %s\n", rc,
 491                         libcfs_net2str(net), hops,
 492                         libcfs_nid2str(gateway));
 493                 return rc;
 494         }
 495
 496         LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
 497         gw = lpni->lpni_peer_net->lpn_peer;
 498
 499         route->lr_gateway = gw;
 500
 501         rnet2 = lnet_find_rnet_locked(net);
 502         if (rnet2 == NULL) {
 503                 /* new network */
 504                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 505                 rnet2 = rnet;
 506         }
 507
 508         /* Search for a duplicate route (it's a NOOP if it is) */
 509         add_route = 1;
 510         list_for_each(route_entry, &rnet2->lrn_routes) {
 511                 struct lnet_route *route2;
 512
 513                 route2 = list_entry(route_entry, struct lnet_route, lr_list);
 514                 if (route2->lr_gateway == route->lr_gateway) {
 515                         add_route = 0;
 516                         break;
 517                 }
 518
 519                 /* our lookups must be true */
 520                 LASSERT(route2->lr_gateway->lp_primary_nid != gateway);
 521         }
 522
 523         /*
 524          * It is possible to add multiple routes through the same peer,
 525          * but it'll be using a different NID of that peer. When the
 526          * gateway is discovered, discovery will consolidate the different
 527          * peers into one peer. In this case the discovery code will have
 528          * to move the routes from the peer that's being deleted to the
 529          * consolidated peer lp_routes list
 530          */
 531         if (add_route) {
 532                 gw->lp_health_sensitivity = sensitivity;
 533                 lnet_add_route_to_rnet(rnet2, route);
 534         }
 535
 536         /*
 537          * get rid of the reference on the lpni.
 538          */
 539         lnet_peer_ni_decref_locked(lpni);
 540         lnet_net_unlock(LNET_LOCK_EX);
 541
 542         rc = 0;
 543
 544         if (!add_route) {
 545                 rc = -EEXIST;
 546                 LIBCFS_FREE(route, sizeof(*route));
 547         }
 548
 549         if (rnet != rnet2)
 550                 LIBCFS_FREE(rnet, sizeof(*rnet));
 551
 552         /* kick start the monitor thread to handle the added route */
 553         wake_up(&the_lnet.ln_mt_waitq);
 554
 555         return rc;
 556 }
 557
 558 static void
 559 lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list,
 560                          struct list_head *zombies)
 561 {
 562         struct lnet_peer *gateway;
 563         struct lnet_route *route;
 564         struct lnet_route *tmp;
 565
 566         list_for_each_entry_safe(route, tmp, route_list, lr_list) {
 567                 gateway = route->lr_gateway;
 568                 if (gw_nid != LNET_NID_ANY &&
 569                     gw_nid != gateway->lp_primary_nid)
 570                         continue;
 571
 572                 /*
 573                  * move to zombie to delete outside the lock
 574                  * Note that this function is called with the
 575                  * ln_api_mutex held as well as the exclusive net
 576                  * lock. Adding to the remote net list happens
 577                  * under the same conditions. Same goes for the
 578                  * gateway router list
 579                  */
 580                 list_move(&route->lr_list, zombies);
 581                 the_lnet.ln_remote_nets_version++;
 582
 583                 list_del(&route->lr_gwlist);
 584                 lnet_rtr_decref_locked(gateway);
 585         }
 586 }
 587
 588 int
 589 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 590 {
 591         struct list_head rnet_zombies;
 592         struct lnet_remotenet *rnet;
 593         struct lnet_remotenet *tmp;
 594         struct list_head *rn_list;
 595         struct lnet_peer_ni *lpni;
 596         struct lnet_route *route;
 597         struct list_head zombies;
 598         struct lnet_peer *lp;
 599         int i = 0;
 600
 601         INIT_LIST_HEAD(&rnet_zombies);
 602         INIT_LIST_HEAD(&zombies);
 603
 604         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 605                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 606
 607         /* NB Caller may specify either all routes via the given gateway
 608          * or a specific route entry actual NIDs) */
 609
 610         lnet_net_lock(LNET_LOCK_EX);
 611
 612         lpni = lnet_find_peer_ni_locked(gw_nid);
 613         if (lpni) {
 614                 lp = lpni->lpni_peer_net->lpn_peer;
 615                 LASSERT(lp);
 616                 gw_nid = lp->lp_primary_nid;
 617                 lnet_peer_ni_decref_locked(lpni);
 618         }
 619
 620         if (net != LNET_NIDNET(LNET_NID_ANY)) {
 621                 rnet = lnet_find_rnet_locked(net);
 622                 if (!rnet) {
 623                         lnet_net_unlock(LNET_LOCK_EX);
 624                         return -ENOENT;
 625                 }
 626                 lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 627                                          &zombies);
 628                 if (list_empty(&rnet->lrn_routes))
 629                         list_move(&rnet->lrn_list, &rnet_zombies);
 630                 goto delete_zombies;
 631         }
 632
 633         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 634                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 635
 636                 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
 637                         lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 638                                                  &zombies);
 639                         if (list_empty(&rnet->lrn_routes))
 640                                 list_move(&rnet->lrn_list, &rnet_zombies);
 641                 }
 642         }
 643
 644 delete_zombies:
 645         lnet_net_unlock(LNET_LOCK_EX);
 646
 647         while (!list_empty(&zombies)) {
 648                 route = list_first_entry(&zombies, struct lnet_route, lr_list);
 649                 list_del(&route->lr_list);
 650                 LIBCFS_FREE(route, sizeof(*route));
 651         }
 652
 653         while (!list_empty(&rnet_zombies)) {
 654                 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
 655                                         lrn_list);
 656                 list_del(&rnet->lrn_list);
 657                 LIBCFS_FREE(rnet, sizeof(*rnet));
 658         }
 659
 660         return 0;
 661 }
 662
 663 void
 664 lnet_destroy_routes (void)
 665 {
 666         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 667 }
 668
 669 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 670 {
 671         struct lnet_rtrbufpool *rbp;
 672         int i, rc = -ENOENT, j;
 673
 674         if (the_lnet.ln_rtrpools == NULL)
 675                 return rc;
 676
 677
 678         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 679                 if (i != cpt)
 680                         continue;
 681
 682                 lnet_net_lock(i);
 683                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 684                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 685                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 686                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 687                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 688                 }
 689                 lnet_net_unlock(i);
 690                 rc = 0;
 691                 break;
 692         }
 693
 694         lnet_net_lock(LNET_LOCK_EX);
 695         pool_cfg->pl_routing = the_lnet.ln_routing;
 696         lnet_net_unlock(LNET_LOCK_EX);
 697
 698         return rc;
 699 }
 700
 701 int
 702 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 703                lnet_nid_t *gateway, __u32 *alive, __u32 *priority, __u32 *sensitivity)
 704 {
 705         struct lnet_remotenet *rnet;
 706         struct list_head *rn_list;
 707         struct lnet_route *route;
 708         struct list_head *e1;
 709         struct list_head *e2;
 710         int cpt;
 711         int i;
 712
 713         cpt = lnet_net_lock_current();
 714
 715         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 716                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 717                 list_for_each(e1, rn_list) {
 718                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 719
 720                         list_for_each(e2, &rnet->lrn_routes) {
 721                                 route = list_entry(e2, struct lnet_route,
 722                                                    lr_list);
 723
 724                                 if (idx-- == 0) {
 725                                         *net      = rnet->lrn_net;
 726                                         *gateway  = route->lr_nid;
 727                                         *hops     = route->lr_hops;
 728                                         *priority = route->lr_priority;
 729                                         *sensitivity = route->lr_gateway->
 730                                                 lp_health_sensitivity;
 731                                         *alive    = lnet_is_route_alive(route);
 732                                         lnet_net_unlock(cpt);
 733                                         return 0;
 734                                 }
 735                         }
 736                 }
 737         }
 738
 739         lnet_net_unlock(cpt);
 740         return -ENOENT;
 741 }
 742
 743 static void
 744 lnet_wait_known_routerstate(void)
 745 {
 746         struct lnet_peer *rtr;
 747         struct list_head *entry;
 748         int all_known;
 749
 750         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
 751
 752         for (;;) {
 753                 int cpt = lnet_net_lock_current();
 754
 755                 all_known = 1;
 756                 list_for_each(entry, &the_lnet.ln_routers) {
 757                         rtr = list_entry(entry, struct lnet_peer,
 758                                          lp_rtr_list);
 759
 760                         spin_lock(&rtr->lp_lock);
 761
 762                         if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
 763                                 all_known = 0;
 764                                 spin_unlock(&rtr->lp_lock);
 765                                 break;
 766                         }
 767                         spin_unlock(&rtr->lp_lock);
 768                 }
 769
 770                 lnet_net_unlock(cpt);
 771
 772                 if (all_known)
 773                         return;
 774
 775                 set_current_state(TASK_UNINTERRUPTIBLE);
 776                 schedule_timeout(cfs_time_seconds(1));
 777         }
 778 }
 779
 780 static bool
 781 lnet_update_ni_status_locked(void)
 782 {
 783         struct lnet_ni *ni = NULL;
 784         bool push = false;
 785         time64_t now;
 786         time64_t timeout;
 787
 788         LASSERT(the_lnet.ln_routing);
 789
 790         timeout = router_ping_timeout + alive_router_check_interval;
 791
 792         now = ktime_get_real_seconds();
 793         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 794                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
 795                         continue;
 796
 797                 if (now < ni->ni_last_alive + timeout)
 798                         continue;
 799
 800                 lnet_ni_lock(ni);
 801                 /* re-check with lock */
 802                 if (now < ni->ni_last_alive + timeout) {
 803                         lnet_ni_unlock(ni);
 804                         continue;
 805                 }
 806
 807                 LASSERT(ni->ni_status != NULL);
 808
 809                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 810                         CDEBUG(D_NET, "NI(%s:%lld) status changed to down\n",
 811                                libcfs_nid2str(ni->ni_nid), timeout);
 812                         /* NB: so far, this is the only place to set
 813                          * NI status to "down" */
 814                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 815                         push = true;
 816                 }
 817                 lnet_ni_unlock(ni);
 818         }
 819
 820         return push;
 821 }
 822
 823 void lnet_wait_router_start(void)
 824 {
 825         if (check_routers_before_use) {
 826                 /* Note that a helpful side-effect of pinging all known routers
 827                  * at startup is that it makes them drop stale connections they
 828                  * may have to a previous instance of me. */
 829                 lnet_wait_known_routerstate();
 830         }
 831 }
 832
 833 /*
 834  * This function is called from the monitor thread to check if there are
 835  * any active routers that need to be checked.
 836  */
 837 inline bool
 838 lnet_router_checker_active(void)
 839 {
 840         /* Router Checker thread needs to run when routing is enabled in
 841          * order to call lnet_update_ni_status_locked() */
 842         if (the_lnet.ln_routing)
 843                 return true;
 844
 845         return !list_empty(&the_lnet.ln_routers) &&
 846                 alive_router_check_interval > 0;
 847 }
 848
 849 void
 850 lnet_check_routers(void)
 851 {
 852         struct lnet_peer_ni *lpni;
 853         struct list_head *entry;
 854         struct lnet_peer *rtr;
 855         bool push = false;
 856         __u64 version;
 857         time64_t now;
 858         int cpt;
 859         int rc;
 860
 861         cpt = lnet_net_lock_current();
 862 rescan:
 863         version = the_lnet.ln_routers_version;
 864
 865         list_for_each(entry, &the_lnet.ln_routers) {
 866                 rtr = list_entry(entry, struct lnet_peer,
 867                                  lp_rtr_list);
 868
 869                 now = ktime_get_real_seconds();
 870
 871                 /*
 872                  * only discover the router if we've passed
 873                  * alive_router_check_interval seconds. Some of the router
 874                  * interfaces could be down and in that case they would be
 875                  * undergoing recovery separately from this discovery.
 876                  */
 877                 if (now - rtr->lp_rtrcheck_timestamp <
 878                     alive_router_check_interval)
 879                        continue;
 880
 881                 /*
 882                  * If we're currently discovering the peer then don't
 883                  * issue another discovery
 884                  */
 885                 spin_lock(&rtr->lp_lock);
 886                 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY) {
 887                         spin_unlock(&rtr->lp_lock);
 888                         continue;
 889                 }
 890                 /* make sure we actively discover the router */
 891                 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
 892                 rtr->lp_state |= LNET_PEER_RTR_DISCOVERY;
 893                 spin_unlock(&rtr->lp_lock);
 894
 895                 /* find the peer_ni associated with the primary NID */
 896                 lpni = lnet_peer_get_ni_locked(rtr, rtr->lp_primary_nid);
 897                 if (!lpni) {
 898                         CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
 899                                libcfs_nid2str(rtr->lp_primary_nid));
 900                         continue;
 901                 }
 902                 lnet_peer_ni_addref_locked(lpni);
 903
 904                 /* discover the router */
 905                 CDEBUG(D_NET, "discover %s, cpt = %d\n",
 906                        libcfs_nid2str(lpni->lpni_nid), cpt);
 907                 rc = lnet_discover_peer_locked(lpni, cpt, false);
 908
 909                 /* decrement ref count acquired by find_peer_ni_locked() */
 910                 lnet_peer_ni_decref_locked(lpni);
 911
 912                 if (!rc)
 913                         rtr->lp_rtrcheck_timestamp = now;
 914                 else
 915                         CERROR("Failed to discover router %s\n",
 916                                libcfs_nid2str(rtr->lp_primary_nid));
 917
 918                 /* NB dropped lock */
 919                 if (version != the_lnet.ln_routers_version) {
 920                         /* the routers list has changed */
 921                         goto rescan;
 922                 }
 923         }
 924
 925         if (the_lnet.ln_routing)
 926                 push = lnet_update_ni_status_locked();
 927
 928         lnet_net_unlock(cpt);
 929
 930         /* if the status of the ni changed update the peers */
 931         if (push)
 932                 lnet_push_update_to_peers(1);
 933 }
 934
 935 void
 936 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
 937 {
 938         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
 939
 940         while (--npages >= 0)
 941                 __free_page(rb->rb_kiov[npages].kiov_page);
 942
 943         LIBCFS_FREE(rb, sz);
 944 }
 945
 946 static struct lnet_rtrbuf *
 947 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
 948 {
 949         int            npages = rbp->rbp_npages;
 950         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
 951         struct page   *page;
 952         struct lnet_rtrbuf *rb;
 953         int            i;
 954
 955         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
 956         if (rb == NULL)
 957                 return NULL;
 958
 959         rb->rb_pool = rbp;
 960
 961         for (i = 0; i < npages; i++) {
 962                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
 963                                           GFP_KERNEL | __GFP_ZERO);
 964                 if (page == NULL) {
 965                         while (--i >= 0)
 966                                 __free_page(rb->rb_kiov[i].kiov_page);
 967
 968                         LIBCFS_FREE(rb, sz);
 969                         return NULL;
 970                 }
 971
 972                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
 973                 rb->rb_kiov[i].kiov_offset = 0;
 974                 rb->rb_kiov[i].kiov_page = page;
 975         }
 976
 977         return rb;
 978 }
 979
 980 static void
 981 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
 982 {
 983         int npages = rbp->rbp_npages;
 984         struct lnet_rtrbuf *rb;
 985         struct list_head tmp;
 986
 987         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
 988                 return;
 989
 990         INIT_LIST_HEAD(&tmp);
 991
 992         lnet_net_lock(cpt);
 993         list_splice_init(&rbp->rbp_msgs, &tmp);
 994         lnet_drop_routed_msgs_locked(&tmp, cpt);
 995         list_splice_init(&rbp->rbp_bufs, &tmp);
 996         rbp->rbp_req_nbuffers = 0;
 997         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
 998         rbp->rbp_mincredits = 0;
 999         lnet_net_unlock(cpt);
1000
1001         /* Free buffers on the free list. */
1002         while (!list_empty(&tmp)) {
1003                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1004                 list_del(&rb->rb_list);
1005                 lnet_destroy_rtrbuf(rb, npages);
1006         }
1007 }
1008
1009 static int
1010 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1011 {
1012         struct list_head rb_list;
1013         struct lnet_rtrbuf *rb;
1014         int             num_rb;
1015         int             num_buffers = 0;
1016         int             old_req_nbufs;
1017         int             npages = rbp->rbp_npages;
1018
1019         lnet_net_lock(cpt);
1020         /* If we are called for less buffers than already in the pool, we
1021          * just lower the req_nbuffers number and excess buffers will be
1022          * thrown away as they are returned to the free list.  Credits
1023          * then get adjusted as well.
1024          * If we already have enough buffers allocated to serve the
1025          * increase requested, then we can treat that the same way as we
1026          * do the decrease. */
1027         num_rb = nbufs - rbp->rbp_nbuffers;
1028         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1029                 rbp->rbp_req_nbuffers = nbufs;
1030                 lnet_net_unlock(cpt);
1031                 return 0;
1032         }
1033         /* store the older value of rbp_req_nbuffers and then set it to
1034          * the new request to prevent lnet_return_rx_credits_locked() from
1035          * freeing buffers that we need to keep around */
1036         old_req_nbufs = rbp->rbp_req_nbuffers;
1037         rbp->rbp_req_nbuffers = nbufs;
1038         lnet_net_unlock(cpt);
1039
1040         INIT_LIST_HEAD(&rb_list);
1041
1042         /* allocate the buffers on a local list first.  If all buffers are
1043          * allocated successfully then join this list to the rbp buffer
1044          * list.  If not then free all allocated buffers. */
1045         while (num_rb-- > 0) {
1046                 rb = lnet_new_rtrbuf(rbp, cpt);
1047                 if (rb == NULL) {
1048                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1049                                nbufs, npages);
1050
1051                         lnet_net_lock(cpt);
1052                         rbp->rbp_req_nbuffers = old_req_nbufs;
1053                         lnet_net_unlock(cpt);
1054
1055                         goto failed;
1056                 }
1057
1058                 list_add(&rb->rb_list, &rb_list);
1059                 num_buffers++;
1060         }
1061
1062         lnet_net_lock(cpt);
1063
1064         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1065         rbp->rbp_nbuffers += num_buffers;
1066         rbp->rbp_credits += num_buffers;
1067         rbp->rbp_mincredits = rbp->rbp_credits;
1068         /* We need to schedule blocked msg using the newly
1069          * added buffers. */
1070         while (!list_empty(&rbp->rbp_bufs) &&
1071                !list_empty(&rbp->rbp_msgs))
1072                 lnet_schedule_blocked_locked(rbp);
1073
1074         lnet_net_unlock(cpt);
1075
1076         return 0;
1077
1078 failed:
1079         while (!list_empty(&rb_list)) {
1080                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1081                 list_del(&rb->rb_list);
1082                 lnet_destroy_rtrbuf(rb, npages);
1083         }
1084
1085         return -ENOMEM;
1086 }
1087
1088 static void
1089 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1090 {
1091         INIT_LIST_HEAD(&rbp->rbp_msgs);
1092         INIT_LIST_HEAD(&rbp->rbp_bufs);
1093
1094         rbp->rbp_npages = npages;
1095         rbp->rbp_credits = 0;
1096         rbp->rbp_mincredits = 0;
1097 }
1098
1099 void
1100 lnet_rtrpools_free(int keep_pools)
1101 {
1102         struct lnet_rtrbufpool *rtrp;
1103         int               i;
1104
1105         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1106                 return;
1107
1108         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1109                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1110                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1111                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1112         }
1113
1114         if (!keep_pools) {
1115                 cfs_percpt_free(the_lnet.ln_rtrpools);
1116                 the_lnet.ln_rtrpools = NULL;
1117         }
1118 }
1119
1120 static int
1121 lnet_nrb_tiny_calculate(void)
1122 {
1123         int     nrbs = LNET_NRB_TINY;
1124
1125         if (tiny_router_buffers < 0) {
1126                 LCONSOLE_ERROR_MSG(0x10c,
1127                                    "tiny_router_buffers=%d invalid when "
1128                                    "routing enabled\n", tiny_router_buffers);
1129                 return -EINVAL;
1130         }
1131
1132         if (tiny_router_buffers > 0)
1133                 nrbs = tiny_router_buffers;
1134
1135         nrbs /= LNET_CPT_NUMBER;
1136         return max(nrbs, LNET_NRB_TINY_MIN);
1137 }
1138
1139 static int
1140 lnet_nrb_small_calculate(void)
1141 {
1142         int     nrbs = LNET_NRB_SMALL;
1143
1144         if (small_router_buffers < 0) {
1145                 LCONSOLE_ERROR_MSG(0x10c,
1146                                    "small_router_buffers=%d invalid when "
1147                                    "routing enabled\n", small_router_buffers);
1148                 return -EINVAL;
1149         }
1150
1151         if (small_router_buffers > 0)
1152                 nrbs = small_router_buffers;
1153
1154         nrbs /= LNET_CPT_NUMBER;
1155         return max(nrbs, LNET_NRB_SMALL_MIN);
1156 }
1157
1158 static int
1159 lnet_nrb_large_calculate(void)
1160 {
1161         int     nrbs = LNET_NRB_LARGE;
1162
1163         if (large_router_buffers < 0) {
1164                 LCONSOLE_ERROR_MSG(0x10c,
1165                                    "large_router_buffers=%d invalid when "
1166                                    "routing enabled\n", large_router_buffers);
1167                 return -EINVAL;
1168         }
1169
1170         if (large_router_buffers > 0)
1171                 nrbs = large_router_buffers;
1172
1173         nrbs /= LNET_CPT_NUMBER;
1174         return max(nrbs, LNET_NRB_LARGE_MIN);
1175 }
1176
1177 int
1178 lnet_rtrpools_alloc(int im_a_router)
1179 {
1180         struct lnet_rtrbufpool *rtrp;
1181         int     nrb_tiny;
1182         int     nrb_small;
1183         int     nrb_large;
1184         int     rc;
1185         int     i;
1186
1187         if (!strcmp(forwarding, "")) {
1188                 /* not set either way */
1189                 if (!im_a_router)
1190                         return 0;
1191         } else if (!strcmp(forwarding, "disabled")) {
1192                 /* explicitly disabled */
1193                 return 0;
1194         } else if (!strcmp(forwarding, "enabled")) {
1195                 /* explicitly enabled */
1196         } else {
1197                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1198                                    "'enabled' or 'disabled'\n");
1199                 return -EINVAL;
1200         }
1201
1202         nrb_tiny = lnet_nrb_tiny_calculate();
1203         if (nrb_tiny < 0)
1204                 return -EINVAL;
1205
1206         nrb_small = lnet_nrb_small_calculate();
1207         if (nrb_small < 0)
1208                 return -EINVAL;
1209
1210         nrb_large = lnet_nrb_large_calculate();
1211         if (nrb_large < 0)
1212                 return -EINVAL;
1213
1214         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1215                                                 LNET_NRBPOOLS *
1216                                                 sizeof(struct lnet_rtrbufpool));
1217         if (the_lnet.ln_rtrpools == NULL) {
1218                 LCONSOLE_ERROR_MSG(0x10c,
1219                                    "Failed to initialize router buffe pool\n");
1220                 return -ENOMEM;
1221         }
1222
1223         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1224                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1225                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1226                                               nrb_tiny, i);
1227                 if (rc != 0)
1228                         goto failed;
1229
1230                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1231                                   LNET_NRB_SMALL_PAGES);
1232                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1233                                               nrb_small, i);
1234                 if (rc != 0)
1235                         goto failed;
1236
1237                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1238                                   LNET_NRB_LARGE_PAGES);
1239                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1240                                               nrb_large, i);
1241                 if (rc != 0)
1242                         goto failed;
1243         }
1244
1245         lnet_net_lock(LNET_LOCK_EX);
1246         the_lnet.ln_routing = 1;
1247         lnet_net_unlock(LNET_LOCK_EX);
1248         wake_up(&the_lnet.ln_mt_waitq);
1249         return 0;
1250
1251  failed:
1252         lnet_rtrpools_free(0);
1253         return rc;
1254 }
1255
1256 static int
1257 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1258 {
1259         int nrb = 0;
1260         int rc = 0;
1261         int i;
1262         struct lnet_rtrbufpool *rtrp;
1263
1264         /* If the provided values for each buffer pool are different than the
1265          * configured values, we need to take action. */
1266         if (tiny >= 0) {
1267                 tiny_router_buffers = tiny;
1268                 nrb = lnet_nrb_tiny_calculate();
1269                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1270                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1271                                                       nrb, i);
1272                         if (rc != 0)
1273                                 return rc;
1274                 }
1275         }
1276         if (small >= 0) {
1277                 small_router_buffers = small;
1278                 nrb = lnet_nrb_small_calculate();
1279                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1280                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1281                                                       nrb, i);
1282                         if (rc != 0)
1283                                 return rc;
1284                 }
1285         }
1286         if (large >= 0) {
1287                 large_router_buffers = large;
1288                 nrb = lnet_nrb_large_calculate();
1289                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1290                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1291                                                       nrb, i);
1292                         if (rc != 0)
1293                                 return rc;
1294                 }
1295         }
1296
1297         return 0;
1298 }
1299
1300 int
1301 lnet_rtrpools_adjust(int tiny, int small, int large)
1302 {
1303         /* this function doesn't revert the changes if adding new buffers
1304          * failed.  It's up to the user space caller to revert the
1305          * changes. */
1306
1307         if (!the_lnet.ln_routing)
1308                 return 0;
1309
1310         return lnet_rtrpools_adjust_helper(tiny, small, large);
1311 }
1312
1313 int
1314 lnet_rtrpools_enable(void)
1315 {
1316         int rc = 0;
1317
1318         if (the_lnet.ln_routing)
1319                 return 0;
1320
1321         if (the_lnet.ln_rtrpools == NULL)
1322                 /* If routing is turned off, and we have never
1323                  * initialized the pools before, just call the
1324                  * standard buffer pool allocation routine as
1325                  * if we are just configuring this for the first
1326                  * time. */
1327                 rc = lnet_rtrpools_alloc(1);
1328         else
1329                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1330         if (rc != 0)
1331                 return rc;
1332
1333         lnet_net_lock(LNET_LOCK_EX);
1334         the_lnet.ln_routing = 1;
1335
1336         the_lnet.ln_ping_target->pb_info.pi_features &=
1337                 ~LNET_PING_FEAT_RTE_DISABLED;
1338         lnet_net_unlock(LNET_LOCK_EX);
1339
1340         return rc;
1341 }
1342
1343 void
1344 lnet_rtrpools_disable(void)
1345 {
1346         if (!the_lnet.ln_routing)
1347                 return;
1348
1349         lnet_net_lock(LNET_LOCK_EX);
1350         the_lnet.ln_routing = 0;
1351         the_lnet.ln_ping_target->pb_info.pi_features |=
1352                 LNET_PING_FEAT_RTE_DISABLED;
1353
1354         tiny_router_buffers = 0;
1355         small_router_buffers = 0;
1356         large_router_buffers = 0;
1357         lnet_net_unlock(LNET_LOCK_EX);
1358         lnet_rtrpools_free(1);
1359 }
1360
1361 static inline void
1362 lnet_notify_peer_down(struct lnet_ni *ni, lnet_nid_t nid)
1363 {
1364         if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1365                 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1366 }
1367
1368 /*
1369  * ni: local NI used to communicate with the peer
1370  * nid: peer NID
1371  * alive: true if peer is alive, false otherwise
1372  * reset: reset health value. This is requested by the LND.
1373  * when: notificaiton time.
1374  */
1375 int
1376 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset,
1377             time64_t when)
1378 {
1379         struct lnet_peer_ni *lpni = NULL;
1380         time64_t now = ktime_get_seconds();
1381         int cpt;
1382
1383         LASSERT (!in_interrupt ());
1384
1385         CDEBUG (D_NET, "%s notifying %s: %s\n",
1386                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1387                 libcfs_nid2str(nid),
1388                 alive ? "up" : "down");
1389
1390         if (ni != NULL &&
1391             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1392                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1393                       libcfs_nid2str(nid), alive ? "birth" : "death",
1394                       libcfs_nid2str(ni->ni_nid));
1395                 return -EINVAL;
1396         }
1397
1398         /* can't do predictions... */
1399         if (when > now) {
1400                 CWARN("Ignoring prediction from %s of %s %s "
1401                       "%lld seconds in the future\n",
1402                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1403                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1404                 return -EINVAL;
1405         }
1406
1407         if (ni != NULL && !alive &&             /* LND telling me she's down */
1408             !auto_down) {                       /* auto-down disabled */
1409                 CDEBUG(D_NET, "Auto-down disabled\n");
1410                 return 0;
1411         }
1412
1413         /* must lock 0 since this is used for synchronization */
1414         lnet_net_lock(0);
1415
1416         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1417                 lnet_net_unlock(0);
1418                 return -ESHUTDOWN;
1419         }
1420
1421         lpni = lnet_find_peer_ni_locked(nid);
1422         if (lpni == NULL) {
1423                 /* nid not found */
1424                 lnet_net_unlock(0);
1425                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1426                 return 0;
1427         }
1428
1429         if (alive) {
1430                 if (reset)
1431                         lnet_set_healthv(&lpni->lpni_healthv,
1432                                          LNET_MAX_HEALTH_VALUE);
1433                 else
1434                         lnet_inc_healthv(&lpni->lpni_healthv);
1435         } else {
1436                 lnet_handle_remote_failure_locked(lpni);
1437         }
1438
1439         /* recalculate aliveness */
1440         alive = lnet_is_peer_ni_alive(lpni);
1441         lnet_net_unlock(0);
1442
1443         if (ni != NULL && !alive)
1444                 lnet_notify_peer_down(ni, lpni->lpni_nid);
1445
1446         cpt = lpni->lpni_cpt;
1447         lnet_net_lock(cpt);
1448         lnet_peer_ni_decref_locked(lpni);
1449         lnet_net_unlock(cpt);
1450
1451         return 0;
1452 }
1453 EXPORT_SYMBOL(lnet_notify);