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 extern unsigned int lnet_current_net_count;
  39
  40 static char *forwarding = "";
  41 module_param(forwarding, charp, 0444);
  42 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
  43
  44 static int tiny_router_buffers;
  45 module_param(tiny_router_buffers, int, 0444);
  46 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
  47 static int small_router_buffers;
  48 module_param(small_router_buffers, int, 0444);
  49 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
  50 static int large_router_buffers;
  51 module_param(large_router_buffers, int, 0444);
  52 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
  53 static int peer_buffer_credits;
  54 module_param(peer_buffer_credits, int, 0444);
  55 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
  56
  57 static int auto_down = 1;
  58 module_param(auto_down, int, 0444);
  59 MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
  60
  61 int
  62 lnet_peer_buffer_credits(struct lnet_net *net)
  63 {
  64         /* NI option overrides LNet default */
  65         if (net->net_tunables.lct_peer_rtr_credits > 0)
  66                 return net->net_tunables.lct_peer_rtr_credits;
  67         if (peer_buffer_credits > 0)
  68                 return peer_buffer_credits;
  69
  70         /* As an approximation, allow this peer the same number of router
  71          * buffers as it is allowed outstanding sends */
  72         return net->net_tunables.lct_peer_tx_credits;
  73 }
  74
  75 static int check_routers_before_use;
  76 module_param(check_routers_before_use, int, 0444);
  77 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
  78
  79 int avoid_asym_router_failure = 1;
  80 module_param(avoid_asym_router_failure, int, 0644);
  81 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
  82
  83 int alive_router_check_interval = 60;
  84 module_param(alive_router_check_interval, int, 0644);
  85 MODULE_PARM_DESC(alive_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
  86
  87 static int router_ping_timeout = 50;
  88 module_param(router_ping_timeout, int, 0644);
  89 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
  90
  91 /*
  92  * A value between 0 and 100. 0 meaning that even if router's interfaces
  93  * have the worse health still consider the gateway usable.
  94  * 100 means that at least one interface on the route's remote net is 100%
  95  * healthy to consider the route alive.
  96  * The default is set to 100 to ensure we maintain the original behavior.
  97  */
  98 unsigned int router_sensitivity_percentage = 100;
  99 static int rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
 100 static struct kernel_param_ops param_ops_rtr_sensitivity = {
 101         .set = rtr_sensitivity_set,
 102         .get = param_get_int,
 103 };
 104 #define param_check_rtr_sensitivity(name, p) \
 105                 __param_check(name, p, int)
 106 #ifdef HAVE_KERNEL_PARAM_OPS
 107 module_param(router_sensitivity_percentage, rtr_sensitivity, S_IRUGO|S_IWUSR);
 108 #else
 109 module_param_call(router_sensitivity_percentage, rtr_sensitivity_set, param_get_int,
 110                   &router_sensitivity_percentage, S_IRUGO|S_IWUSR);
 111 #endif
 112 MODULE_PARM_DESC(router_sensitivity_percentage,
 113                 "How healthy a gateway should be to be used in percent");
 114
 115 static int
 116 rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 117 {
 118         int rc;
 119         unsigned *sen = (unsigned *)kp->arg;
 120         unsigned long value;
 121
 122         rc = kstrtoul(val, 0, &value);
 123         if (rc) {
 124                 CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
 125                 return rc;
 126         }
 127
 128         if (value < 0 || value > 100) {
 129                 CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
 130                 return -EINVAL;
 131         }
 132
 133         /*
 134          * The purpose of locking the api_mutex here is to ensure that
 135          * the correct value ends up stored properly.
 136          */
 137         mutex_lock(&the_lnet.ln_api_mutex);
 138
 139         *sen = value;
 140
 141         mutex_unlock(&the_lnet.ln_api_mutex);
 142
 143         return 0;
 144 }
 145
 146 void
 147 lnet_rtr_transfer_to_peer(struct lnet_peer *src, struct lnet_peer *target)
 148 {
 149         struct lnet_route *route;
 150
 151         lnet_net_lock(LNET_LOCK_EX);
 152         target->lp_rtr_refcount += src->lp_rtr_refcount;
 153         /* move the list of queued messages to the new peer */
 154         list_splice_init(&src->lp_rtrq, &target->lp_rtrq);
 155         /* move all the routes that reference the peer */
 156         list_splice_init(&src->lp_routes, &target->lp_routes);
 157         /* update all the routes to point to the new peer */
 158         list_for_each_entry(route, &target->lp_routes, lr_gwlist)
 159                 route->lr_gateway = target;
 160         /* remove the old peer from the ln_routers list */
 161         list_del_init(&src->lp_rtr_list);
 162         /* add the new peer to the ln_routers list */
 163         if (list_empty(&target->lp_rtr_list)) {
 164                 lnet_peer_addref_locked(target);
 165                 list_add_tail(&target->lp_rtr_list, &the_lnet.ln_routers);
 166         }
 167         /* reset the ref count on the old peer and decrement its ref count */
 168         src->lp_rtr_refcount = 0;
 169         lnet_peer_decref_locked(src);
 170         /* update the router version */
 171         the_lnet.ln_routers_version++;
 172         lnet_net_unlock(LNET_LOCK_EX);
 173 }
 174
 175 int
 176 lnet_peers_start_down(void)
 177 {
 178         return check_routers_before_use;
 179 }
 180
 181 /*
 182  * A net is alive if at least one gateway NI on the network is alive.
 183  */
 184 static bool
 185 lnet_is_gateway_net_alive(struct lnet_peer_net *lpn)
 186 {
 187         struct lnet_peer_ni *lpni;
 188
 189         list_for_each_entry(lpni, &lpn->lpn_peer_nis, lpni_peer_nis) {
 190                 if (lnet_is_peer_ni_alive(lpni))
 191                         return true;
 192         }
 193
 194         return false;
 195 }
 196
 197 /*
 198  * a gateway is alive only if all its nets are alive
 199  * called with cpt lock held
 200  */
 201 bool lnet_is_gateway_alive(struct lnet_peer *gw)
 202 {
 203         struct lnet_peer_net *lpn;
 204
 205         list_for_each_entry(lpn, &gw->lp_peer_nets, lpn_peer_nets) {
 206                 if (!lnet_is_gateway_net_alive(lpn))
 207                         return false;
 208         }
 209
 210         return true;
 211 }
 212
 213 /*
 214  * lnet_is_route_alive() needs to be called with cpt lock held
 215  * A route is alive if the gateway can route between the local network and
 216  * the remote network of the route.
 217  * This means at least one NI is alive on each of the local and remote
 218  * networks of the gateway.
 219  */
 220 bool lnet_is_route_alive(struct lnet_route *route)
 221 {
 222         struct lnet_peer *gw = route->lr_gateway;
 223         struct lnet_peer_net *llpn;
 224         struct lnet_peer_net *rlpn;
 225         bool route_alive;
 226
 227         /*
 228          * check the gateway's interfaces on the route rnet to make sure
 229          * that the gateway is viable.
 230          */
 231         llpn = lnet_peer_get_net_locked(gw, route->lr_lnet);
 232         if (!llpn)
 233                 return false;
 234
 235         route_alive = lnet_is_gateway_net_alive(llpn);
 236
 237         if (avoid_asym_router_failure) {
 238                 rlpn = lnet_peer_get_net_locked(gw, route->lr_net);
 239                 if (!rlpn)
 240                         return false;
 241                 route_alive = route_alive &&
 242                               lnet_is_gateway_net_alive(rlpn);
 243         }
 244
 245         if (!route_alive)
 246                 return route_alive;
 247
 248         spin_lock(&gw->lp_lock);
 249         if (!(gw->lp_state & LNET_PEER_ROUTER_ENABLED)) {
 250                 if (gw->lp_rtr_refcount > 0)
 251                         CERROR("peer %s is being used as a gateway but routing feature is not turned on\n",
 252                                libcfs_nid2str(gw->lp_primary_nid));
 253                 route_alive = false;
 254         }
 255         spin_unlock(&gw->lp_lock);
 256
 257         return route_alive;
 258 }
 259
 260 void
 261 lnet_consolidate_routes_locked(struct lnet_peer *orig_lp,
 262                                struct lnet_peer *new_lp)
 263 {
 264         struct lnet_peer_ni *lpni;
 265         struct lnet_route *route;
 266
 267         /*
 268          * Although a route is correlated with a peer, but when it's added
 269          * a specific NID is used. That NID refers to a peer_ni within
 270          * a peer. There could be other peer_nis on the same net, which
 271          * can be used to send to that gateway. However when we are
 272          * consolidating gateways because of discovery, the nid used to
 273          * add the route might've moved between gateway peers. In this
 274          * case we want to move the route to the new gateway as well. The
 275          * intent here is not to confuse the user who added the route.
 276          */
 277         list_for_each_entry(route, &orig_lp->lp_routes, lr_gwlist) {
 278                 lpni = lnet_peer_get_ni_locked(orig_lp, route->lr_nid);
 279                 if (!lpni) {
 280                         lnet_net_lock(LNET_LOCK_EX);
 281                         list_move(&route->lr_gwlist, &new_lp->lp_routes);
 282                         lnet_net_unlock(LNET_LOCK_EX);
 283                 }
 284         }
 285
 286 }
 287
 288 void
 289 lnet_router_discovery_complete(struct lnet_peer *lp)
 290 {
 291         struct lnet_peer_ni *lpni = NULL;
 292
 293         spin_lock(&lp->lp_lock);
 294         lp->lp_state &= ~LNET_PEER_RTR_DISCOVERY;
 295         spin_unlock(&lp->lp_lock);
 296
 297         /*
 298          * Router discovery successful? All peer information would've been
 299          * updated already. No need to do any more processing
 300          */
 301         if (!lp->lp_dc_error)
 302                 return;
 303         /*
 304          * discovery failed? then we need to set the status of each lpni
 305          * to DOWN. It will be updated the next time we discover the
 306          * router. For router peer NIs not on local networks, we never send
 307          * messages directly to them, so their health will always remain
 308          * at maximum. We can only tell if they are up or down from the
 309          * status returned in the PING response. If we fail to get that
 310          * status in our scheduled router discovery, then we'll assume
 311          * it's down until we're told otherwise.
 312          */
 313         CDEBUG(D_NET, "%s: Router discovery failed %d\n",
 314                libcfs_nid2str(lp->lp_primary_nid), lp->lp_dc_error);
 315         while ((lpni = lnet_get_next_peer_ni_locked(lp, NULL, lpni)) != NULL)
 316                 lpni->lpni_ns_status = LNET_NI_STATUS_DOWN;
 317 }
 318
 319 static void
 320 lnet_rtr_addref_locked(struct lnet_peer *lp)
 321 {
 322         LASSERT(lp->lp_rtr_refcount >= 0);
 323
 324         /* lnet_net_lock must be exclusively locked */
 325         lp->lp_rtr_refcount++;
 326         if (lp->lp_rtr_refcount == 1) {
 327                 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
 328                 /* addref for the_lnet.ln_routers */
 329                 lnet_peer_addref_locked(lp);
 330                 the_lnet.ln_routers_version++;
 331         }
 332 }
 333
 334 static void
 335 lnet_rtr_decref_locked(struct lnet_peer *lp)
 336 {
 337         LASSERT(atomic_read(&lp->lp_refcount) > 0);
 338         LASSERT(lp->lp_rtr_refcount > 0);
 339
 340         /* lnet_net_lock must be exclusively locked */
 341         lp->lp_rtr_refcount--;
 342         if (lp->lp_rtr_refcount == 0) {
 343                 LASSERT(list_empty(&lp->lp_routes));
 344
 345                 list_del(&lp->lp_rtr_list);
 346                 /* decref for the_lnet.ln_routers */
 347                 lnet_peer_decref_locked(lp);
 348                 the_lnet.ln_routers_version++;
 349         }
 350 }
 351
 352 struct lnet_remotenet *
 353 lnet_find_rnet_locked(__u32 net)
 354 {
 355         struct lnet_remotenet *rnet;
 356         struct list_head *tmp;
 357         struct list_head *rn_list;
 358
 359         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 360
 361         rn_list = lnet_net2rnethash(net);
 362         list_for_each(tmp, rn_list) {
 363                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 364
 365                 if (rnet->lrn_net == net)
 366                         return rnet;
 367         }
 368         return NULL;
 369 }
 370
 371 static void lnet_shuffle_seed(void)
 372 {
 373         static int seeded;
 374         struct lnet_ni *ni = NULL;
 375
 376         if (seeded)
 377                 return;
 378
 379         /* Nodes with small feet have little entropy
 380          * the NID for this node gives the most entropy in the low bits */
 381         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 382                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 383
 384         seeded = 1;
 385         return;
 386 }
 387
 388 /* NB expects LNET_LOCK held */
 389 static void
 390 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 391 {
 392         struct lnet_peer_net *lpn;
 393         unsigned int offset = 0;
 394         unsigned int len = 0;
 395         struct list_head *e;
 396
 397         lnet_shuffle_seed();
 398
 399         list_for_each(e, &rnet->lrn_routes)
 400                 len++;
 401
 402         /*
 403          * Randomly adding routes to the list is done to ensure that when
 404          * different nodes are using the same list of routers, they end up
 405          * preferring different routers.
 406          */
 407         offset = prandom_u32_max(len + 1);
 408         list_for_each(e, &rnet->lrn_routes) {
 409                 if (offset == 0)
 410                         break;
 411                 offset--;
 412         }
 413         list_add(&route->lr_list, e);
 414         /*
 415          * force a router check on the gateway to make sure the route is
 416          * alive
 417          */
 418         list_for_each_entry(lpn, &route->lr_gateway->lp_peer_nets,
 419                             lpn_peer_nets) {
 420                 lpn->lpn_rtrcheck_timestamp = 0;
 421         }
 422
 423         the_lnet.ln_remote_nets_version++;
 424
 425         /* add the route on the gateway list */
 426         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 427
 428         /* take a router reference count on the gateway */
 429         lnet_rtr_addref_locked(route->lr_gateway);
 430 }
 431
 432 int
 433 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 434                __u32 priority, __u32 sensitivity)
 435 {
 436         struct list_head *route_entry;
 437         struct lnet_remotenet *rnet;
 438         struct lnet_remotenet *rnet2;
 439         struct lnet_route *route;
 440         struct lnet_peer_ni *lpni;
 441         struct lnet_peer *gw;
 442         int add_route;
 443         int rc;
 444
 445         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 446                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 447
 448         if (gateway == LNET_NID_ANY ||
 449             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 450             net == LNET_NIDNET(LNET_NID_ANY) ||
 451             LNET_NETTYP(net) == LOLND ||
 452             LNET_NIDNET(gateway) == net ||
 453             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 454                 return -EINVAL;
 455
 456         /* it's a local network */
 457         if (lnet_islocalnet(net))
 458                 return -EEXIST;
 459
 460         if (!lnet_islocalnet(LNET_NIDNET(gateway))) {
 461                 CERROR("Cannot add route with gateway %s. There is no local interface configured on LNet %s\n",
 462                        libcfs_nid2str(gateway),
 463                        libcfs_net2str(LNET_NIDNET(gateway)));
 464                 return -EINVAL;
 465         }
 466
 467         /* Assume net, route, all new */
 468         LIBCFS_ALLOC(route, sizeof(*route));
 469         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 470         if (route == NULL || rnet == NULL) {
 471                 CERROR("Out of memory creating route %s %d %s\n",
 472                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 473                 if (route != NULL)
 474                         LIBCFS_FREE(route, sizeof(*route));
 475                 if (rnet != NULL)
 476                         LIBCFS_FREE(rnet, sizeof(*rnet));
 477                 return -ENOMEM;
 478         }
 479
 480         INIT_LIST_HEAD(&rnet->lrn_routes);
 481         rnet->lrn_net = net;
 482         /* store the local and remote net that the route represents */
 483         route->lr_lnet = LNET_NIDNET(gateway);
 484         route->lr_net = net;
 485         route->lr_nid = gateway;
 486         route->lr_priority = priority;
 487         route->lr_hops = hops;
 488
 489         lnet_net_lock(LNET_LOCK_EX);
 490
 491         /*
 492          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 493          * lose that once we're done
 494          */
 495         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 496         if (IS_ERR(lpni)) {
 497                 lnet_net_unlock(LNET_LOCK_EX);
 498
 499                 LIBCFS_FREE(route, sizeof(*route));
 500                 LIBCFS_FREE(rnet, sizeof(*rnet));
 501
 502                 rc = PTR_ERR(lpni);
 503                 CERROR("Error %d creating route %s %d %s\n", rc,
 504                         libcfs_net2str(net), hops,
 505                         libcfs_nid2str(gateway));
 506                 return rc;
 507         }
 508
 509         LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
 510         gw = lpni->lpni_peer_net->lpn_peer;
 511
 512         route->lr_gateway = gw;
 513
 514         rnet2 = lnet_find_rnet_locked(net);
 515         if (rnet2 == NULL) {
 516                 /* new network */
 517                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 518                 rnet2 = rnet;
 519         }
 520
 521         /* Search for a duplicate route (it's a NOOP if it is) */
 522         add_route = 1;
 523         list_for_each(route_entry, &rnet2->lrn_routes) {
 524                 struct lnet_route *route2;
 525
 526                 route2 = list_entry(route_entry, struct lnet_route, lr_list);
 527                 if (route2->lr_gateway == route->lr_gateway) {
 528                         add_route = 0;
 529                         break;
 530                 }
 531
 532                 /* our lookups must be true */
 533                 LASSERT(route2->lr_gateway->lp_primary_nid != gateway);
 534         }
 535
 536         /*
 537          * It is possible to add multiple routes through the same peer,
 538          * but it'll be using a different NID of that peer. When the
 539          * gateway is discovered, discovery will consolidate the different
 540          * peers into one peer. In this case the discovery code will have
 541          * to move the routes from the peer that's being deleted to the
 542          * consolidated peer lp_routes list
 543          */
 544         if (add_route) {
 545                 gw->lp_health_sensitivity = sensitivity;
 546                 lnet_add_route_to_rnet(rnet2, route);
 547                 if (lnet_peer_discovery_disabled)
 548                         CWARN("Consider turning discovery on to enable full "
 549                               "Multi-Rail routing functionality\n");
 550         }
 551
 552         /*
 553          * get rid of the reference on the lpni.
 554          */
 555         lnet_peer_ni_decref_locked(lpni);
 556         lnet_net_unlock(LNET_LOCK_EX);
 557
 558         rc = 0;
 559
 560         if (!add_route) {
 561                 rc = -EEXIST;
 562                 LIBCFS_FREE(route, sizeof(*route));
 563         }
 564
 565         if (rnet != rnet2)
 566                 LIBCFS_FREE(rnet, sizeof(*rnet));
 567
 568         /* kick start the monitor thread to handle the added route */
 569         wake_up(&the_lnet.ln_mt_waitq);
 570
 571         return rc;
 572 }
 573
 574 static void
 575 lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list,
 576                          struct list_head *zombies)
 577 {
 578         struct lnet_peer *gateway;
 579         struct lnet_route *route;
 580         struct lnet_route *tmp;
 581
 582         list_for_each_entry_safe(route, tmp, route_list, lr_list) {
 583                 gateway = route->lr_gateway;
 584                 if (gw_nid != LNET_NID_ANY &&
 585                     gw_nid != gateway->lp_primary_nid)
 586                         continue;
 587
 588                 /*
 589                  * move to zombie to delete outside the lock
 590                  * Note that this function is called with the
 591                  * ln_api_mutex held as well as the exclusive net
 592                  * lock. Adding to the remote net list happens
 593                  * under the same conditions. Same goes for the
 594                  * gateway router list
 595                  */
 596                 list_move(&route->lr_list, zombies);
 597                 the_lnet.ln_remote_nets_version++;
 598
 599                 list_del(&route->lr_gwlist);
 600                 lnet_rtr_decref_locked(gateway);
 601         }
 602 }
 603
 604 int
 605 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 606 {
 607         struct list_head rnet_zombies;
 608         struct lnet_remotenet *rnet;
 609         struct lnet_remotenet *tmp;
 610         struct list_head *rn_list;
 611         struct lnet_peer_ni *lpni;
 612         struct lnet_route *route;
 613         struct list_head zombies;
 614         struct lnet_peer *lp;
 615         int i = 0;
 616
 617         INIT_LIST_HEAD(&rnet_zombies);
 618         INIT_LIST_HEAD(&zombies);
 619
 620         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 621                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 622
 623         /* NB Caller may specify either all routes via the given gateway
 624          * or a specific route entry actual NIDs) */
 625
 626         lnet_net_lock(LNET_LOCK_EX);
 627
 628         lpni = lnet_find_peer_ni_locked(gw_nid);
 629         if (lpni) {
 630                 lp = lpni->lpni_peer_net->lpn_peer;
 631                 LASSERT(lp);
 632                 gw_nid = lp->lp_primary_nid;
 633                 lnet_peer_ni_decref_locked(lpni);
 634         }
 635
 636         if (net != LNET_NIDNET(LNET_NID_ANY)) {
 637                 rnet = lnet_find_rnet_locked(net);
 638                 if (!rnet) {
 639                         lnet_net_unlock(LNET_LOCK_EX);
 640                         return -ENOENT;
 641                 }
 642                 lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 643                                          &zombies);
 644                 if (list_empty(&rnet->lrn_routes))
 645                         list_move(&rnet->lrn_list, &rnet_zombies);
 646                 goto delete_zombies;
 647         }
 648
 649         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 650                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 651
 652                 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
 653                         lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 654                                                  &zombies);
 655                         if (list_empty(&rnet->lrn_routes))
 656                                 list_move(&rnet->lrn_list, &rnet_zombies);
 657                 }
 658         }
 659
 660 delete_zombies:
 661         /*
 662          * check if there are any routes remaining on the gateway
 663          * If there are no more routes make sure to set the peer's
 664          * lp_disc_net_id to 0 (invalid), in case we add more routes in
 665          * the future on that gateway, then we start our discovery process
 666          * from scratch
 667          */
 668         if (lpni) {
 669                 if (list_empty(&lp->lp_routes))
 670                         lp->lp_disc_net_id = 0;
 671         }
 672
 673         lnet_net_unlock(LNET_LOCK_EX);
 674
 675         while (!list_empty(&zombies)) {
 676                 route = list_first_entry(&zombies, struct lnet_route, lr_list);
 677                 list_del(&route->lr_list);
 678                 LIBCFS_FREE(route, sizeof(*route));
 679         }
 680
 681         while (!list_empty(&rnet_zombies)) {
 682                 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
 683                                         lrn_list);
 684                 list_del(&rnet->lrn_list);
 685                 LIBCFS_FREE(rnet, sizeof(*rnet));
 686         }
 687
 688         return 0;
 689 }
 690
 691 void
 692 lnet_destroy_routes (void)
 693 {
 694         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 695 }
 696
 697 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 698 {
 699         struct lnet_rtrbufpool *rbp;
 700         int i, rc = -ENOENT, j;
 701
 702         if (the_lnet.ln_rtrpools == NULL)
 703                 return rc;
 704
 705
 706         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 707                 if (i != cpt)
 708                         continue;
 709
 710                 lnet_net_lock(i);
 711                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 712                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 713                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 714                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 715                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 716                 }
 717                 lnet_net_unlock(i);
 718                 rc = 0;
 719                 break;
 720         }
 721
 722         lnet_net_lock(LNET_LOCK_EX);
 723         pool_cfg->pl_routing = the_lnet.ln_routing;
 724         lnet_net_unlock(LNET_LOCK_EX);
 725
 726         return rc;
 727 }
 728
 729 int
 730 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 731                lnet_nid_t *gateway, __u32 *alive, __u32 *priority, __u32 *sensitivity)
 732 {
 733         struct lnet_remotenet *rnet;
 734         struct list_head *rn_list;
 735         struct lnet_route *route;
 736         struct list_head *e1;
 737         struct list_head *e2;
 738         int cpt;
 739         int i;
 740
 741         cpt = lnet_net_lock_current();
 742
 743         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 744                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 745                 list_for_each(e1, rn_list) {
 746                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 747
 748                         list_for_each(e2, &rnet->lrn_routes) {
 749                                 route = list_entry(e2, struct lnet_route,
 750                                                    lr_list);
 751
 752                                 if (idx-- == 0) {
 753                                         *net      = rnet->lrn_net;
 754                                         *gateway  = route->lr_nid;
 755                                         *hops     = route->lr_hops;
 756                                         *priority = route->lr_priority;
 757                                         *sensitivity = route->lr_gateway->
 758                                                 lp_health_sensitivity;
 759                                         *alive    = lnet_is_route_alive(route);
 760                                         lnet_net_unlock(cpt);
 761                                         return 0;
 762                                 }
 763                         }
 764                 }
 765         }
 766
 767         lnet_net_unlock(cpt);
 768         return -ENOENT;
 769 }
 770
 771 static void
 772 lnet_wait_known_routerstate(void)
 773 {
 774         struct lnet_peer *rtr;
 775         struct list_head *entry;
 776         int all_known;
 777
 778         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
 779
 780         for (;;) {
 781                 int cpt = lnet_net_lock_current();
 782
 783                 all_known = 1;
 784                 list_for_each(entry, &the_lnet.ln_routers) {
 785                         rtr = list_entry(entry, struct lnet_peer,
 786                                          lp_rtr_list);
 787
 788                         spin_lock(&rtr->lp_lock);
 789
 790                         if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
 791                                 all_known = 0;
 792                                 spin_unlock(&rtr->lp_lock);
 793                                 break;
 794                         }
 795                         spin_unlock(&rtr->lp_lock);
 796                 }
 797
 798                 lnet_net_unlock(cpt);
 799
 800                 if (all_known)
 801                         return;
 802
 803                 set_current_state(TASK_UNINTERRUPTIBLE);
 804                 schedule_timeout(cfs_time_seconds(1));
 805         }
 806 }
 807
 808 static inline bool
 809 lnet_net_set_status_locked(struct lnet_net *net, __u32 status)
 810 {
 811         struct lnet_ni *ni;
 812         bool update = false;
 813
 814         list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
 815                 lnet_ni_lock(ni);
 816                 if (ni->ni_status &&
 817                     ni->ni_status->ns_status != status) {
 818                     ni->ni_status->ns_status = status;
 819                     update = true;
 820                 }
 821                 lnet_ni_unlock(ni);
 822         }
 823
 824         return update;
 825 }
 826
 827 static bool
 828 lnet_update_ni_status_locked(void)
 829 {
 830         struct lnet_net *net;
 831         bool push = false;
 832         time64_t now;
 833         time64_t timeout;
 834
 835         LASSERT(the_lnet.ln_routing);
 836
 837         timeout = router_ping_timeout + alive_router_check_interval;
 838
 839         now = ktime_get_real_seconds();
 840         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 841                 if (net->net_lnd->lnd_type == LOLND)
 842                         continue;
 843
 844                 if (now < net->net_last_alive + timeout)
 845                         continue;
 846
 847                 spin_lock(&net->net_lock);
 848                 /* re-check with lock */
 849                 if (now < net->net_last_alive + timeout) {
 850                         spin_unlock(&net->net_lock);
 851                         continue;
 852                 }
 853                 spin_unlock(&net->net_lock);
 854
 855                 /*
 856                  * if the net didn't receive any traffic for past the
 857                  * timeout on any of its constituent NIs, then mark all
 858                  * the NIs down.
 859                  */
 860                 push = lnet_net_set_status_locked(net, LNET_NI_STATUS_DOWN);
 861         }
 862
 863         return push;
 864 }
 865
 866 void lnet_wait_router_start(void)
 867 {
 868         if (check_routers_before_use) {
 869                 /* Note that a helpful side-effect of pinging all known routers
 870                  * at startup is that it makes them drop stale connections they
 871                  * may have to a previous instance of me. */
 872                 lnet_wait_known_routerstate();
 873         }
 874 }
 875
 876 /*
 877  * This function is called from the monitor thread to check if there are
 878  * any active routers that need to be checked.
 879  */
 880 inline bool
 881 lnet_router_checker_active(void)
 882 {
 883         /* Router Checker thread needs to run when routing is enabled in
 884          * order to call lnet_update_ni_status_locked() */
 885         if (the_lnet.ln_routing)
 886                 return true;
 887
 888         return !list_empty(&the_lnet.ln_routers) &&
 889                 alive_router_check_interval > 0;
 890 }
 891
 892 void
 893 lnet_check_routers(void)
 894 {
 895         struct lnet_peer_net *first_lpn = NULL;
 896         struct lnet_peer_net *lpn;
 897         struct lnet_peer_ni *lpni;
 898         struct list_head *entry;
 899         struct lnet_peer *rtr;
 900         bool push = false;
 901         bool found_lpn;
 902         __u64 version;
 903         __u32 net_id;
 904         time64_t now;
 905         int cpt;
 906         int rc;
 907
 908         cpt = lnet_net_lock_current();
 909 rescan:
 910         version = the_lnet.ln_routers_version;
 911
 912         list_for_each(entry, &the_lnet.ln_routers) {
 913                 rtr = list_entry(entry, struct lnet_peer,
 914                                  lp_rtr_list);
 915
 916                 now = ktime_get_real_seconds();
 917
 918                 /*
 919                  * only discover the router if we've passed
 920                  * alive_router_check_interval seconds. Some of the router
 921                  * interfaces could be down and in that case they would be
 922                  * undergoing recovery separately from this discovery.
 923                  */
 924                 /* find next peer net which is also local */
 925                 net_id = rtr->lp_disc_net_id;
 926                 do {
 927                         lpn = lnet_get_next_peer_net_locked(rtr, net_id);
 928                         if (!lpn) {
 929                                 CERROR("gateway %s has no networks\n",
 930                                 libcfs_nid2str(rtr->lp_primary_nid));
 931                                 break;
 932                         }
 933                         if (first_lpn == lpn)
 934                                 break;
 935                         if (!first_lpn)
 936                                 first_lpn = lpn;
 937                         found_lpn = lnet_islocalnet_locked(lpn->lpn_net_id);
 938                         net_id = lpn->lpn_net_id;
 939                 } while (!found_lpn);
 940
 941                 if (!found_lpn || !lpn) {
 942                         CERROR("no local network found for gateway %s\n",
 943                                libcfs_nid2str(rtr->lp_primary_nid));
 944                         continue;
 945                 }
 946
 947                 if (now - lpn->lpn_rtrcheck_timestamp <
 948                     alive_router_check_interval / lnet_current_net_count)
 949                        continue;
 950
 951                 /*
 952                  * If we're currently discovering the peer then don't
 953                  * issue another discovery
 954                  */
 955                 spin_lock(&rtr->lp_lock);
 956                 if (rtr->lp_state & LNET_PEER_RTR_DISCOVERY) {
 957                         spin_unlock(&rtr->lp_lock);
 958                         continue;
 959                 }
 960                 /* make sure we actively discover the router */
 961                 rtr->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
 962                 rtr->lp_state |= LNET_PEER_RTR_DISCOVERY;
 963                 spin_unlock(&rtr->lp_lock);
 964
 965                 /* find the peer_ni associated with the primary NID */
 966                 lpni = lnet_peer_get_ni_locked(rtr, rtr->lp_primary_nid);
 967                 if (!lpni) {
 968                         CDEBUG(D_NET, "Expected to find an lpni for %s, but non found\n",
 969                                libcfs_nid2str(rtr->lp_primary_nid));
 970                         continue;
 971                 }
 972                 lnet_peer_ni_addref_locked(lpni);
 973
 974                 /* specify the net to use */
 975                 rtr->lp_disc_net_id = lpn->lpn_net_id;
 976
 977                 /* discover the router */
 978                 CDEBUG(D_NET, "discover %s, cpt = %d\n",
 979                        libcfs_nid2str(lpni->lpni_nid), cpt);
 980                 rc = lnet_discover_peer_locked(lpni, cpt, false);
 981
 982                 /* decrement ref count acquired by find_peer_ni_locked() */
 983                 lnet_peer_ni_decref_locked(lpni);
 984
 985                 if (!rc)
 986                         lpn->lpn_rtrcheck_timestamp = now;
 987                 else
 988                         CERROR("Failed to discover router %s\n",
 989                                libcfs_nid2str(rtr->lp_primary_nid));
 990
 991                 /* NB dropped lock */
 992                 if (version != the_lnet.ln_routers_version) {
 993                         /* the routers list has changed */
 994                         goto rescan;
 995                 }
 996         }
 997
 998         if (the_lnet.ln_routing)
 999                 push = lnet_update_ni_status_locked();
1000
1001         lnet_net_unlock(cpt);
1002
1003         /* if the status of the ni changed update the peers */
1004         if (push)
1005                 lnet_push_update_to_peers(1);
1006 }
1007
1008 void
1009 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1010 {
1011         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1012
1013         while (--npages >= 0)
1014                 __free_page(rb->rb_kiov[npages].kiov_page);
1015
1016         LIBCFS_FREE(rb, sz);
1017 }
1018
1019 static struct lnet_rtrbuf *
1020 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1021 {
1022         int            npages = rbp->rbp_npages;
1023         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1024         struct page   *page;
1025         struct lnet_rtrbuf *rb;
1026         int            i;
1027
1028         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1029         if (rb == NULL)
1030                 return NULL;
1031
1032         rb->rb_pool = rbp;
1033
1034         for (i = 0; i < npages; i++) {
1035                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1036                                           GFP_KERNEL | __GFP_ZERO);
1037                 if (page == NULL) {
1038                         while (--i >= 0)
1039                                 __free_page(rb->rb_kiov[i].kiov_page);
1040
1041                         LIBCFS_FREE(rb, sz);
1042                         return NULL;
1043                 }
1044
1045                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1046                 rb->rb_kiov[i].kiov_offset = 0;
1047                 rb->rb_kiov[i].kiov_page = page;
1048         }
1049
1050         return rb;
1051 }
1052
1053 static void
1054 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1055 {
1056         int npages = rbp->rbp_npages;
1057         struct lnet_rtrbuf *rb;
1058         struct list_head tmp;
1059
1060         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1061                 return;
1062
1063         INIT_LIST_HEAD(&tmp);
1064
1065         lnet_net_lock(cpt);
1066         list_splice_init(&rbp->rbp_msgs, &tmp);
1067         lnet_drop_routed_msgs_locked(&tmp, cpt);
1068         list_splice_init(&rbp->rbp_bufs, &tmp);
1069         rbp->rbp_req_nbuffers = 0;
1070         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1071         rbp->rbp_mincredits = 0;
1072         lnet_net_unlock(cpt);
1073
1074         /* Free buffers on the free list. */
1075         while (!list_empty(&tmp)) {
1076                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1077                 list_del(&rb->rb_list);
1078                 lnet_destroy_rtrbuf(rb, npages);
1079         }
1080 }
1081
1082 static int
1083 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1084 {
1085         struct list_head rb_list;
1086         struct lnet_rtrbuf *rb;
1087         int             num_rb;
1088         int             num_buffers = 0;
1089         int             old_req_nbufs;
1090         int             npages = rbp->rbp_npages;
1091
1092         lnet_net_lock(cpt);
1093         /* If we are called for less buffers than already in the pool, we
1094          * just lower the req_nbuffers number and excess buffers will be
1095          * thrown away as they are returned to the free list.  Credits
1096          * then get adjusted as well.
1097          * If we already have enough buffers allocated to serve the
1098          * increase requested, then we can treat that the same way as we
1099          * do the decrease. */
1100         num_rb = nbufs - rbp->rbp_nbuffers;
1101         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1102                 rbp->rbp_req_nbuffers = nbufs;
1103                 lnet_net_unlock(cpt);
1104                 return 0;
1105         }
1106         /* store the older value of rbp_req_nbuffers and then set it to
1107          * the new request to prevent lnet_return_rx_credits_locked() from
1108          * freeing buffers that we need to keep around */
1109         old_req_nbufs = rbp->rbp_req_nbuffers;
1110         rbp->rbp_req_nbuffers = nbufs;
1111         lnet_net_unlock(cpt);
1112
1113         INIT_LIST_HEAD(&rb_list);
1114
1115         /* allocate the buffers on a local list first.  If all buffers are
1116          * allocated successfully then join this list to the rbp buffer
1117          * list.  If not then free all allocated buffers. */
1118         while (num_rb-- > 0) {
1119                 rb = lnet_new_rtrbuf(rbp, cpt);
1120                 if (rb == NULL) {
1121                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1122                                nbufs, npages);
1123
1124                         lnet_net_lock(cpt);
1125                         rbp->rbp_req_nbuffers = old_req_nbufs;
1126                         lnet_net_unlock(cpt);
1127
1128                         goto failed;
1129                 }
1130
1131                 list_add(&rb->rb_list, &rb_list);
1132                 num_buffers++;
1133         }
1134
1135         lnet_net_lock(cpt);
1136
1137         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1138         rbp->rbp_nbuffers += num_buffers;
1139         rbp->rbp_credits += num_buffers;
1140         rbp->rbp_mincredits = rbp->rbp_credits;
1141         /* We need to schedule blocked msg using the newly
1142          * added buffers. */
1143         while (!list_empty(&rbp->rbp_bufs) &&
1144                !list_empty(&rbp->rbp_msgs))
1145                 lnet_schedule_blocked_locked(rbp);
1146
1147         lnet_net_unlock(cpt);
1148
1149         return 0;
1150
1151 failed:
1152         while (!list_empty(&rb_list)) {
1153                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1154                 list_del(&rb->rb_list);
1155                 lnet_destroy_rtrbuf(rb, npages);
1156         }
1157
1158         return -ENOMEM;
1159 }
1160
1161 static void
1162 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1163 {
1164         INIT_LIST_HEAD(&rbp->rbp_msgs);
1165         INIT_LIST_HEAD(&rbp->rbp_bufs);
1166
1167         rbp->rbp_npages = npages;
1168         rbp->rbp_credits = 0;
1169         rbp->rbp_mincredits = 0;
1170 }
1171
1172 void
1173 lnet_rtrpools_free(int keep_pools)
1174 {
1175         struct lnet_rtrbufpool *rtrp;
1176         int               i;
1177
1178         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1179                 return;
1180
1181         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1182                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1183                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1184                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1185         }
1186
1187         if (!keep_pools) {
1188                 cfs_percpt_free(the_lnet.ln_rtrpools);
1189                 the_lnet.ln_rtrpools = NULL;
1190         }
1191 }
1192
1193 static int
1194 lnet_nrb_tiny_calculate(void)
1195 {
1196         int     nrbs = LNET_NRB_TINY;
1197
1198         if (tiny_router_buffers < 0) {
1199                 LCONSOLE_ERROR_MSG(0x10c,
1200                                    "tiny_router_buffers=%d invalid when "
1201                                    "routing enabled\n", tiny_router_buffers);
1202                 return -EINVAL;
1203         }
1204
1205         if (tiny_router_buffers > 0)
1206                 nrbs = tiny_router_buffers;
1207
1208         nrbs /= LNET_CPT_NUMBER;
1209         return max(nrbs, LNET_NRB_TINY_MIN);
1210 }
1211
1212 static int
1213 lnet_nrb_small_calculate(void)
1214 {
1215         int     nrbs = LNET_NRB_SMALL;
1216
1217         if (small_router_buffers < 0) {
1218                 LCONSOLE_ERROR_MSG(0x10c,
1219                                    "small_router_buffers=%d invalid when "
1220                                    "routing enabled\n", small_router_buffers);
1221                 return -EINVAL;
1222         }
1223
1224         if (small_router_buffers > 0)
1225                 nrbs = small_router_buffers;
1226
1227         nrbs /= LNET_CPT_NUMBER;
1228         return max(nrbs, LNET_NRB_SMALL_MIN);
1229 }
1230
1231 static int
1232 lnet_nrb_large_calculate(void)
1233 {
1234         int     nrbs = LNET_NRB_LARGE;
1235
1236         if (large_router_buffers < 0) {
1237                 LCONSOLE_ERROR_MSG(0x10c,
1238                                    "large_router_buffers=%d invalid when "
1239                                    "routing enabled\n", large_router_buffers);
1240                 return -EINVAL;
1241         }
1242
1243         if (large_router_buffers > 0)
1244                 nrbs = large_router_buffers;
1245
1246         nrbs /= LNET_CPT_NUMBER;
1247         return max(nrbs, LNET_NRB_LARGE_MIN);
1248 }
1249
1250 int
1251 lnet_rtrpools_alloc(int im_a_router)
1252 {
1253         struct lnet_rtrbufpool *rtrp;
1254         int     nrb_tiny;
1255         int     nrb_small;
1256         int     nrb_large;
1257         int     rc;
1258         int     i;
1259
1260         if (!strcmp(forwarding, "")) {
1261                 /* not set either way */
1262                 if (!im_a_router)
1263                         return 0;
1264         } else if (!strcmp(forwarding, "disabled")) {
1265                 /* explicitly disabled */
1266                 return 0;
1267         } else if (!strcmp(forwarding, "enabled")) {
1268                 /* explicitly enabled */
1269         } else {
1270                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1271                                    "'enabled' or 'disabled'\n");
1272                 return -EINVAL;
1273         }
1274
1275         nrb_tiny = lnet_nrb_tiny_calculate();
1276         if (nrb_tiny < 0)
1277                 return -EINVAL;
1278
1279         nrb_small = lnet_nrb_small_calculate();
1280         if (nrb_small < 0)
1281                 return -EINVAL;
1282
1283         nrb_large = lnet_nrb_large_calculate();
1284         if (nrb_large < 0)
1285                 return -EINVAL;
1286
1287         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1288                                                 LNET_NRBPOOLS *
1289                                                 sizeof(struct lnet_rtrbufpool));
1290         if (the_lnet.ln_rtrpools == NULL) {
1291                 LCONSOLE_ERROR_MSG(0x10c,
1292                                    "Failed to initialize router buffe pool\n");
1293                 return -ENOMEM;
1294         }
1295
1296         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1297                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1298                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1299                                               nrb_tiny, i);
1300                 if (rc != 0)
1301                         goto failed;
1302
1303                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1304                                   LNET_NRB_SMALL_PAGES);
1305                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1306                                               nrb_small, i);
1307                 if (rc != 0)
1308                         goto failed;
1309
1310                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1311                                   LNET_NRB_LARGE_PAGES);
1312                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1313                                               nrb_large, i);
1314                 if (rc != 0)
1315                         goto failed;
1316         }
1317
1318         lnet_net_lock(LNET_LOCK_EX);
1319         the_lnet.ln_routing = 1;
1320         lnet_net_unlock(LNET_LOCK_EX);
1321         wake_up(&the_lnet.ln_mt_waitq);
1322         return 0;
1323
1324  failed:
1325         lnet_rtrpools_free(0);
1326         return rc;
1327 }
1328
1329 static int
1330 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1331 {
1332         int nrb = 0;
1333         int rc = 0;
1334         int i;
1335         struct lnet_rtrbufpool *rtrp;
1336
1337         /* If the provided values for each buffer pool are different than the
1338          * configured values, we need to take action. */
1339         if (tiny >= 0) {
1340                 tiny_router_buffers = tiny;
1341                 nrb = lnet_nrb_tiny_calculate();
1342                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1343                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1344                                                       nrb, i);
1345                         if (rc != 0)
1346                                 return rc;
1347                 }
1348         }
1349         if (small >= 0) {
1350                 small_router_buffers = small;
1351                 nrb = lnet_nrb_small_calculate();
1352                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1353                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1354                                                       nrb, i);
1355                         if (rc != 0)
1356                                 return rc;
1357                 }
1358         }
1359         if (large >= 0) {
1360                 large_router_buffers = large;
1361                 nrb = lnet_nrb_large_calculate();
1362                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1363                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1364                                                       nrb, i);
1365                         if (rc != 0)
1366                                 return rc;
1367                 }
1368         }
1369
1370         return 0;
1371 }
1372
1373 int
1374 lnet_rtrpools_adjust(int tiny, int small, int large)
1375 {
1376         /* this function doesn't revert the changes if adding new buffers
1377          * failed.  It's up to the user space caller to revert the
1378          * changes. */
1379
1380         if (!the_lnet.ln_routing)
1381                 return 0;
1382
1383         return lnet_rtrpools_adjust_helper(tiny, small, large);
1384 }
1385
1386 int
1387 lnet_rtrpools_enable(void)
1388 {
1389         int rc = 0;
1390
1391         if (the_lnet.ln_routing)
1392                 return 0;
1393
1394         if (the_lnet.ln_rtrpools == NULL)
1395                 /* If routing is turned off, and we have never
1396                  * initialized the pools before, just call the
1397                  * standard buffer pool allocation routine as
1398                  * if we are just configuring this for the first
1399                  * time. */
1400                 rc = lnet_rtrpools_alloc(1);
1401         else
1402                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1403         if (rc != 0)
1404                 return rc;
1405
1406         lnet_net_lock(LNET_LOCK_EX);
1407         the_lnet.ln_routing = 1;
1408
1409         the_lnet.ln_ping_target->pb_info.pi_features &=
1410                 ~LNET_PING_FEAT_RTE_DISABLED;
1411         lnet_net_unlock(LNET_LOCK_EX);
1412
1413         if (lnet_peer_discovery_disabled)
1414                 CWARN("Consider turning discovery on to enable full "
1415                       "Multi-Rail routing functionality\n");
1416
1417         return rc;
1418 }
1419
1420 void
1421 lnet_rtrpools_disable(void)
1422 {
1423         if (!the_lnet.ln_routing)
1424                 return;
1425
1426         lnet_net_lock(LNET_LOCK_EX);
1427         the_lnet.ln_routing = 0;
1428         the_lnet.ln_ping_target->pb_info.pi_features |=
1429                 LNET_PING_FEAT_RTE_DISABLED;
1430
1431         tiny_router_buffers = 0;
1432         small_router_buffers = 0;
1433         large_router_buffers = 0;
1434         lnet_net_unlock(LNET_LOCK_EX);
1435         lnet_rtrpools_free(1);
1436 }
1437
1438 static inline void
1439 lnet_notify_peer_down(struct lnet_ni *ni, lnet_nid_t nid)
1440 {
1441         if (ni->ni_net->net_lnd->lnd_notify_peer_down != NULL)
1442                 (ni->ni_net->net_lnd->lnd_notify_peer_down)(nid);
1443 }
1444
1445 /*
1446  * ni: local NI used to communicate with the peer
1447  * nid: peer NID
1448  * alive: true if peer is alive, false otherwise
1449  * reset: reset health value. This is requested by the LND.
1450  * when: notificaiton time.
1451  */
1452 int
1453 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, bool alive, bool reset,
1454             time64_t when)
1455 {
1456         struct lnet_peer_ni *lpni = NULL;
1457         time64_t now = ktime_get_seconds();
1458         int cpt;
1459
1460         LASSERT (!in_interrupt ());
1461
1462         CDEBUG (D_NET, "%s notifying %s: %s\n",
1463                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1464                 libcfs_nid2str(nid),
1465                 alive ? "up" : "down");
1466
1467         if (ni != NULL &&
1468             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1469                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1470                       libcfs_nid2str(nid), alive ? "birth" : "death",
1471                       libcfs_nid2str(ni->ni_nid));
1472                 return -EINVAL;
1473         }
1474
1475         /* can't do predictions... */
1476         if (when > now) {
1477                 CWARN("Ignoring prediction from %s of %s %s "
1478                       "%lld seconds in the future\n",
1479                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1480                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1481                 return -EINVAL;
1482         }
1483
1484         if (ni != NULL && !alive &&             /* LND telling me she's down */
1485             !auto_down) {                       /* auto-down disabled */
1486                 CDEBUG(D_NET, "Auto-down disabled\n");
1487                 return 0;
1488         }
1489
1490         /* must lock 0 since this is used for synchronization */
1491         lnet_net_lock(0);
1492
1493         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1494                 lnet_net_unlock(0);
1495                 return -ESHUTDOWN;
1496         }
1497
1498         lpni = lnet_find_peer_ni_locked(nid);
1499         if (lpni == NULL) {
1500                 /* nid not found */
1501                 lnet_net_unlock(0);
1502                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1503                 return 0;
1504         }
1505
1506         if (alive) {
1507                 if (reset)
1508                         lnet_set_healthv(&lpni->lpni_healthv,
1509                                          LNET_MAX_HEALTH_VALUE);
1510                 else
1511                         lnet_inc_healthv(&lpni->lpni_healthv);
1512         } else {
1513                 lnet_handle_remote_failure_locked(lpni);
1514         }
1515
1516         /* recalculate aliveness */
1517         alive = lnet_is_peer_ni_alive(lpni);
1518         lnet_net_unlock(0);
1519
1520         if (ni != NULL && !alive)
1521                 lnet_notify_peer_down(ni, lpni->lpni_nid);
1522
1523         cpt = lpni->lpni_cpt;
1524         lnet_net_lock(cpt);
1525         lnet_peer_ni_decref_locked(lpni);
1526         lnet_net_unlock(cpt);
1527
1528         return 0;
1529 }
1530 EXPORT_SYMBOL(lnet_notify);