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 static int dead_router_check_interval = 60;
  82 module_param(dead_router_check_interval, int, 0644);
  83 MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)");
  84
  85 static int live_router_check_interval = 60;
  86 module_param(live_router_check_interval, int, 0644);
  87 MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
  88
  89 static int router_ping_timeout = 50;
  90 module_param(router_ping_timeout, int, 0644);
  91 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
  92
  93 /*
  94  * A value between 0 and 100. 0 meaning that even if router's interfaces
  95  * have the worse health still consider the gateway usable.
  96  * 100 means that at least one interface on the route's remote net is 100%
  97  * healthy to consider the route alive.
  98  * The default is set to 100 to ensure we maintain the original behavior.
  99  */
 100 unsigned int router_sensitivity_percentage = 100;
 101 static int rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
 102 static struct kernel_param_ops param_ops_rtr_sensitivity = {
 103         .set = rtr_sensitivity_set,
 104         .get = param_get_int,
 105 };
 106 #define param_check_rtr_sensitivity(name, p) \
 107                 __param_check(name, p, int)
 108 #ifdef HAVE_KERNEL_PARAM_OPS
 109 module_param(router_sensitivity_percentage, rtr_sensitivity, S_IRUGO|S_IWUSR);
 110 #else
 111 module_param_call(router_sensitivity_percentage, rtr_sensitivity_set, param_get_int,
 112                   &router_sensitivity_percentage, S_IRUGO|S_IWUSR);
 113 #endif
 114 MODULE_PARM_DESC(router_sensitivity_percentage,
 115                 "How healthy a gateway should be to be used in percent");
 116
 117 static int
 118 rtr_sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 119 {
 120         int rc;
 121         unsigned *sen = (unsigned *)kp->arg;
 122         unsigned long value;
 123
 124         rc = kstrtoul(val, 0, &value);
 125         if (rc) {
 126                 CERROR("Invalid module parameter value for 'router_sensitivity_percentage'\n");
 127                 return rc;
 128         }
 129
 130         if (value < 0 || value > 100) {
 131                 CERROR("Invalid value: %lu for 'router_sensitivity_percentage'\n", value);
 132                 return -EINVAL;
 133         }
 134
 135         /*
 136          * The purpose of locking the api_mutex here is to ensure that
 137          * the correct value ends up stored properly.
 138          */
 139         mutex_lock(&the_lnet.ln_api_mutex);
 140
 141         *sen = value;
 142
 143         mutex_unlock(&the_lnet.ln_api_mutex);
 144
 145         return 0;
 146 }
 147
 148 int
 149 lnet_peers_start_down(void)
 150 {
 151         return check_routers_before_use;
 152 }
 153
 154 void
 155 lnet_notify_locked(struct lnet_peer_ni *lp, int notifylnd, int alive,
 156                    time64_t when)
 157 {
 158         if (lp->lpni_timestamp > when) { /* out of date information */
 159                 CDEBUG(D_NET, "Out of date\n");
 160                 return;
 161         }
 162
 163         /*
 164          * This function can be called with different cpt locks being
 165          * held. lpni_alive_count modification needs to be properly protected.
 166          * Significant reads to lpni_alive_count are also protected with
 167          * the same lock
 168          */
 169         spin_lock(&lp->lpni_lock);
 170
 171         lp->lpni_timestamp = when; /* update timestamp */
 172
 173         /* got old news */
 174         if (lp->lpni_alive_count != 0 &&
 175             /* new date for old news */
 176             (!lnet_is_peer_ni_alive(lp)) == (!alive)) {
 177                 spin_unlock(&lp->lpni_lock);
 178                 CDEBUG(D_NET, "Old news\n");
 179                 return;
 180         }
 181
 182         /* Flag that notification is outstanding */
 183
 184         lp->lpni_alive_count++;
 185         lp->lpni_notify = 1;
 186         lp->lpni_notifylnd = notifylnd;
 187         if (lnet_is_peer_ni_alive(lp))
 188                 lp->lpni_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 189
 190         spin_unlock(&lp->lpni_lock);
 191
 192         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive);
 193 }
 194
 195 /*
 196  * This function will always be called with lp->lpni_cpt lock held.
 197  */
 198 static void
 199 lnet_ni_notify_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
 200 {
 201         int alive;
 202         int notifylnd;
 203
 204         /* Notify only in 1 thread at any time to ensure ordered notification.
 205          * NB individual events can be missed; the only guarantee is that you
 206          * always get the most recent news */
 207
 208         spin_lock(&lp->lpni_lock);
 209
 210         if (lp->lpni_notifying || ni == NULL) {
 211                 spin_unlock(&lp->lpni_lock);
 212                 return;
 213         }
 214
 215         lp->lpni_notifying = 1;
 216
 217         /*
 218          * lp->lpni_notify needs to be protected because it can be set in
 219          * lnet_notify_locked().
 220          */
 221         while (lp->lpni_notify) {
 222                 alive     = lnet_is_peer_ni_alive(lp);
 223                 notifylnd = lp->lpni_notifylnd;
 224
 225                 lp->lpni_notifylnd = 0;
 226                 lp->lpni_notify    = 0;
 227
 228                 if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
 229                         spin_unlock(&lp->lpni_lock);
 230                         lnet_net_unlock(lp->lpni_cpt);
 231
 232                         /* A new notification could happen now; I'll handle it
 233                          * when control returns to me */
 234
 235                         (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid,
 236                                                           alive);
 237
 238                         lnet_net_lock(lp->lpni_cpt);
 239                         spin_lock(&lp->lpni_lock);
 240                 }
 241         }
 242
 243         lp->lpni_notifying = 0;
 244         spin_unlock(&lp->lpni_lock);
 245 }
 246
 247 static void
 248 lnet_rtr_addref_locked(struct lnet_peer *lp)
 249 {
 250         LASSERT(lp->lp_rtr_refcount >= 0);
 251
 252         /* lnet_net_lock must be exclusively locked */
 253         lp->lp_rtr_refcount++;
 254         if (lp->lp_rtr_refcount == 1) {
 255                 list_add_tail(&lp->lp_rtr_list, &the_lnet.ln_routers);
 256                 /* addref for the_lnet.ln_routers */
 257                 lnet_peer_addref_locked(lp);
 258                 the_lnet.ln_routers_version++;
 259         }
 260 }
 261
 262 static void
 263 lnet_rtr_decref_locked(struct lnet_peer *lp)
 264 {
 265         LASSERT(atomic_read(&lp->lp_refcount) > 0);
 266         LASSERT(lp->lp_rtr_refcount > 0);
 267
 268         /* lnet_net_lock must be exclusively locked */
 269         lp->lp_rtr_refcount--;
 270         if (lp->lp_rtr_refcount == 0) {
 271                 LASSERT(list_empty(&lp->lp_routes));
 272
 273                 list_del(&lp->lp_rtr_list);
 274                 /* decref for the_lnet.ln_routers */
 275                 lnet_peer_decref_locked(lp);
 276                 the_lnet.ln_routers_version++;
 277         }
 278 }
 279
 280 struct lnet_remotenet *
 281 lnet_find_rnet_locked(__u32 net)
 282 {
 283         struct lnet_remotenet *rnet;
 284         struct list_head *tmp;
 285         struct list_head *rn_list;
 286
 287         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 288
 289         rn_list = lnet_net2rnethash(net);
 290         list_for_each(tmp, rn_list) {
 291                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 292
 293                 if (rnet->lrn_net == net)
 294                         return rnet;
 295         }
 296         return NULL;
 297 }
 298
 299 static void lnet_shuffle_seed(void)
 300 {
 301         static int seeded;
 302         struct lnet_ni *ni = NULL;
 303
 304         if (seeded)
 305                 return;
 306
 307         /* Nodes with small feet have little entropy
 308          * the NID for this node gives the most entropy in the low bits */
 309         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 310                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 311
 312         seeded = 1;
 313         return;
 314 }
 315
 316 /* NB expects LNET_LOCK held */
 317 static void
 318 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 319 {
 320         unsigned int len = 0;
 321         unsigned int offset = 0;
 322         struct list_head *e;
 323
 324         lnet_shuffle_seed();
 325
 326         list_for_each(e, &rnet->lrn_routes)
 327                 len++;
 328
 329         /*
 330          * Randomly adding routes to the list is done to ensure that when
 331          * different nodes are using the same list of routers, they end up
 332          * preferring different routers.
 333          */
 334         offset = cfs_rand() % (len + 1);
 335         list_for_each(e, &rnet->lrn_routes) {
 336                 if (offset == 0)
 337                         break;
 338                 offset--;
 339         }
 340         list_add(&route->lr_list, e);
 341         /*
 342          * force a router check on the gateway to make sure the route is
 343          * alive
 344          */
 345         route->lr_gateway->lp_rtrcheck_timestamp = 0;
 346
 347         the_lnet.ln_remote_nets_version++;
 348
 349         /* add the route on the gateway list */
 350         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 351
 352         /* take a router reference count on the gateway */
 353         lnet_rtr_addref_locked(route->lr_gateway);
 354 }
 355
 356 int
 357 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 358                unsigned int priority)
 359 {
 360         struct list_head *route_entry;
 361         struct lnet_remotenet *rnet;
 362         struct lnet_remotenet *rnet2;
 363         struct lnet_route *route;
 364         struct lnet_peer_ni *lpni;
 365         struct lnet_peer *gw;
 366         int add_route;
 367         int rc;
 368
 369         CDEBUG(D_NET, "Add route: remote net %s hops %d priority %u gw %s\n",
 370                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 371
 372         if (gateway == LNET_NID_ANY ||
 373             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 374             net == LNET_NIDNET(LNET_NID_ANY) ||
 375             LNET_NETTYP(net) == LOLND ||
 376             LNET_NIDNET(gateway) == net ||
 377             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 378                 return -EINVAL;
 379
 380         /* it's a local network */
 381         if (lnet_islocalnet(net))
 382                 return -EEXIST;
 383
 384         /* Assume net, route, all new */
 385         LIBCFS_ALLOC(route, sizeof(*route));
 386         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 387         if (route == NULL || rnet == NULL) {
 388                 CERROR("Out of memory creating route %s %d %s\n",
 389                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 390                 if (route != NULL)
 391                         LIBCFS_FREE(route, sizeof(*route));
 392                 if (rnet != NULL)
 393                         LIBCFS_FREE(rnet, sizeof(*rnet));
 394                 return -ENOMEM;
 395         }
 396
 397         INIT_LIST_HEAD(&rnet->lrn_routes);
 398         rnet->lrn_net = net;
 399         /* store the local and remote net that the route represents */
 400         route->lr_lnet = LNET_NIDNET(gateway);
 401         route->lr_net = net;
 402         route->lr_priority = priority;
 403         route->lr_hops = hops;
 404
 405         lnet_net_lock(LNET_LOCK_EX);
 406
 407         /*
 408          * lnet_nid2peerni_ex() grabs a ref on the lpni. We will need to
 409          * lose that once we're done
 410          */
 411         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 412         if (IS_ERR(lpni)) {
 413                 lnet_net_unlock(LNET_LOCK_EX);
 414
 415                 LIBCFS_FREE(route, sizeof(*route));
 416                 LIBCFS_FREE(rnet, sizeof(*rnet));
 417
 418                 rc = PTR_ERR(lpni);
 419                 CERROR("Error %d creating route %s %d %s\n", rc,
 420                         libcfs_net2str(net), hops,
 421                         libcfs_nid2str(gateway));
 422                 return rc;
 423         }
 424
 425         LASSERT(lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer);
 426         gw = lpni->lpni_peer_net->lpn_peer;
 427
 428         route->lr_gateway = gw;
 429
 430         rnet2 = lnet_find_rnet_locked(net);
 431         if (rnet2 == NULL) {
 432                 /* new network */
 433                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 434                 rnet2 = rnet;
 435         }
 436
 437         /* Search for a duplicate route (it's a NOOP if it is) */
 438         add_route = 1;
 439         list_for_each(route_entry, &rnet2->lrn_routes) {
 440                 struct lnet_route *route2;
 441
 442                 route2 = list_entry(route_entry, struct lnet_route, lr_list);
 443                 if (route2->lr_gateway == route->lr_gateway) {
 444                         add_route = 0;
 445                         break;
 446                 }
 447
 448                 /* our lookups must be true */
 449                 LASSERT(route2->lr_gateway->lp_primary_nid != gateway);
 450         }
 451
 452         /*
 453          * It is possible to add multiple routes through the same peer,
 454          * but it'll be using a different NID of that peer. When the
 455          * gateway is discovered, discovery will consolidate the different
 456          * peers into one peer. In this case the discovery code will have
 457          * to move the routes from the peer that's being deleted to the
 458          * consolidated peer lp_routes list
 459          */
 460         if (add_route)
 461                 lnet_add_route_to_rnet(rnet2, route);
 462
 463         /*
 464          * get rid of the reference on the lpni.
 465          */
 466         lnet_peer_ni_decref_locked(lpni);
 467         lnet_net_unlock(LNET_LOCK_EX);
 468
 469         rc = 0;
 470
 471         if (!add_route) {
 472                 rc = -EEXIST;
 473                 LIBCFS_FREE(route, sizeof(*route));
 474         }
 475
 476         if (rnet != rnet2)
 477                 LIBCFS_FREE(rnet, sizeof(*rnet));
 478
 479         /* kick start the monitor thread to handle the added route */
 480         wake_up(&the_lnet.ln_mt_waitq);
 481
 482         return rc;
 483 }
 484
 485 static void
 486 lnet_del_route_from_rnet(lnet_nid_t gw_nid, struct list_head *route_list,
 487                          struct list_head *zombies)
 488 {
 489         struct lnet_peer *gateway;
 490         struct lnet_route *route;
 491         struct lnet_route *tmp;
 492
 493         list_for_each_entry_safe(route, tmp, route_list, lr_list) {
 494                 gateway = route->lr_gateway;
 495                 if (gw_nid != LNET_NID_ANY &&
 496                     gw_nid != gateway->lp_primary_nid)
 497                         continue;
 498
 499                 /*
 500                  * move to zombie to delete outside the lock
 501                  * Note that this function is called with the
 502                  * ln_api_mutex held as well as the exclusive net
 503                  * lock. Adding to the remote net list happens
 504                  * under the same conditions. Same goes for the
 505                  * gateway router list
 506                  */
 507                 list_move(&route->lr_list, zombies);
 508                 the_lnet.ln_remote_nets_version++;
 509
 510                 list_del(&route->lr_gwlist);
 511                 lnet_rtr_decref_locked(gateway);
 512         }
 513 }
 514
 515 int
 516 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 517 {
 518         struct list_head rnet_zombies;
 519         struct lnet_remotenet *rnet;
 520         struct lnet_remotenet *tmp;
 521         struct list_head *rn_list;
 522         struct lnet_peer_ni *lpni;
 523         struct lnet_route *route;
 524         struct list_head zombies;
 525         struct lnet_peer *lp;
 526         int i = 0;
 527
 528         INIT_LIST_HEAD(&rnet_zombies);
 529         INIT_LIST_HEAD(&zombies);
 530
 531         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 532                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 533
 534         /* NB Caller may specify either all routes via the given gateway
 535          * or a specific route entry actual NIDs) */
 536
 537         lnet_net_lock(LNET_LOCK_EX);
 538
 539         lpni = lnet_find_peer_ni_locked(gw_nid);
 540         if (lpni) {
 541                 lp = lpni->lpni_peer_net->lpn_peer;
 542                 LASSERT(lp);
 543                 gw_nid = lp->lp_primary_nid;
 544                 lnet_peer_ni_decref_locked(lpni);
 545         }
 546
 547         if (net != LNET_NIDNET(LNET_NID_ANY)) {
 548                 rnet = lnet_find_rnet_locked(net);
 549                 if (!rnet) {
 550                         lnet_net_unlock(LNET_LOCK_EX);
 551                         return -ENOENT;
 552                 }
 553                 lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 554                                          &zombies);
 555                 if (list_empty(&rnet->lrn_routes))
 556                         list_move(&rnet->lrn_list, &rnet_zombies);
 557                 goto delete_zombies;
 558         }
 559
 560         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 561                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 562
 563                 list_for_each_entry_safe(rnet, tmp, rn_list, lrn_list) {
 564                         lnet_del_route_from_rnet(gw_nid, &rnet->lrn_routes,
 565                                                  &zombies);
 566                         if (list_empty(&rnet->lrn_routes))
 567                                 list_move(&rnet->lrn_list, &rnet_zombies);
 568                 }
 569         }
 570
 571 delete_zombies:
 572         lnet_net_unlock(LNET_LOCK_EX);
 573
 574         while (!list_empty(&zombies)) {
 575                 route = list_first_entry(&zombies, struct lnet_route, lr_list);
 576                 list_del(&route->lr_list);
 577                 LIBCFS_FREE(route, sizeof(*route));
 578         }
 579
 580         while (!list_empty(&rnet_zombies)) {
 581                 rnet = list_first_entry(&rnet_zombies, struct lnet_remotenet,
 582                                         lrn_list);
 583                 list_del(&rnet->lrn_list);
 584                 LIBCFS_FREE(rnet, sizeof(*rnet));
 585         }
 586
 587         return 0;
 588 }
 589
 590 void
 591 lnet_destroy_routes (void)
 592 {
 593         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 594 }
 595
 596 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 597 {
 598         struct lnet_rtrbufpool *rbp;
 599         int i, rc = -ENOENT, j;
 600
 601         if (the_lnet.ln_rtrpools == NULL)
 602                 return rc;
 603
 604
 605         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 606                 if (i != cpt)
 607                         continue;
 608
 609                 lnet_net_lock(i);
 610                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 611                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 612                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 613                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 614                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 615                 }
 616                 lnet_net_unlock(i);
 617                 rc = 0;
 618                 break;
 619         }
 620
 621         lnet_net_lock(LNET_LOCK_EX);
 622         pool_cfg->pl_routing = the_lnet.ln_routing;
 623         lnet_net_unlock(LNET_LOCK_EX);
 624
 625         return rc;
 626 }
 627
 628 int
 629 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 630                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 631 {
 632         struct list_head *e1;
 633         struct list_head *e2;
 634         struct lnet_remotenet *rnet;
 635         struct lnet_route        *route;
 636         int               cpt;
 637         int               i;
 638         struct list_head *rn_list;
 639
 640         cpt = lnet_net_lock_current();
 641
 642         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 643                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 644                 list_for_each(e1, rn_list) {
 645                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 646
 647                         list_for_each(e2, &rnet->lrn_routes) {
 648                                 route = list_entry(e2, struct lnet_route,
 649                                                    lr_list);
 650
 651                                 if (idx-- == 0) {
 652                                         *net      = rnet->lrn_net;
 653                                         *hops     = route->lr_hops;
 654                                         *priority = route->lr_priority;
 655                                         *gateway  = route->lr_gateway->lp_primary_nid;
 656                                         *alive    = lnet_is_route_alive(route);
 657                                         lnet_net_unlock(cpt);
 658                                         return 0;
 659                                 }
 660                         }
 661                 }
 662         }
 663
 664         lnet_net_unlock(cpt);
 665         return -ENOENT;
 666 }
 667
 668 void
 669 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
 670 {
 671         struct lnet_ni_status *stat;
 672         int nnis;
 673         int i;
 674
 675         __swab32s(&pbuf->pb_info.pi_magic);
 676         __swab32s(&pbuf->pb_info.pi_features);
 677         __swab32s(&pbuf->pb_info.pi_pid);
 678         __swab32s(&pbuf->pb_info.pi_nnis);
 679         nnis = pbuf->pb_info.pi_nnis;
 680         if (nnis > pbuf->pb_nnis)
 681                 nnis = pbuf->pb_nnis;
 682         for (i = 0; i < nnis; i++) {
 683                 stat = &pbuf->pb_info.pi_ni[i];
 684                 __swab64s(&stat->ns_nid);
 685                 __swab32s(&stat->ns_status);
 686         }
 687         return;
 688 }
 689
 690 /**
 691  * TODO: re-implement
 692  */
 693 static void
 694 lnet_parse_rc_info(struct lnet_rc_data *rcd)
 695 {
 696         rcd = rcd;
 697 }
 698
 699 static void
 700 lnet_router_checker_event(struct lnet_event *event)
 701 {
 702         struct lnet_rc_data *rcd = event->md.user_ptr;
 703         struct lnet_peer_ni *lp;
 704
 705         LASSERT(rcd != NULL);
 706
 707         if (event->unlinked) {
 708                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 709                 return;
 710         }
 711
 712         LASSERT(event->type == LNET_EVENT_SEND ||
 713                 event->type == LNET_EVENT_REPLY);
 714
 715         lp = rcd->rcd_gateway;
 716         LASSERT(lp != NULL);
 717
 718          /* NB: it's called with holding lnet_res_lock, we have a few
 719           * places need to hold both locks at the same time, please take
 720           * care of lock ordering */
 721         lnet_net_lock(lp->lpni_cpt);
 722         if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
 723                 /* ignore if no longer a router or rcd is replaced */
 724                 goto out;
 725         }
 726
 727         if (event->type == LNET_EVENT_SEND) {
 728                 if (event->status == 0)
 729                         goto out;
 730         }
 731
 732         /* LNET_EVENT_REPLY */
 733         /* A successful REPLY means the router is up.  If _any_ comms
 734          * to the router fail I assume it's down (this will happen if
 735          * we ping alive routers to try to detect router death before
 736          * apps get burned). */
 737
 738         lnet_notify_locked(lp, 1, !event->status, ktime_get_seconds());
 739         /* The router checker will wake up very shortly and do the
 740          * actual notification.
 741          * XXX If 'lp' stops being a router before then, it will still
 742          * have the notification pending!!! */
 743
 744         if (avoid_asym_router_failure && event->status == 0)
 745                 lnet_parse_rc_info(rcd);
 746
 747  out:
 748         lnet_net_unlock(lp->lpni_cpt);
 749 }
 750
 751 static void
 752 lnet_wait_known_routerstate(void)
 753 {
 754         struct lnet_peer *rtr;
 755         struct list_head *entry;
 756         int all_known;
 757
 758         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
 759
 760         for (;;) {
 761                 int cpt = lnet_net_lock_current();
 762
 763                 all_known = 1;
 764                 list_for_each(entry, &the_lnet.ln_routers) {
 765                         rtr = list_entry(entry, struct lnet_peer,
 766                                          lp_rtr_list);
 767
 768                         spin_lock(&rtr->lp_lock);
 769
 770                         if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
 771                                 all_known = 0;
 772                                 spin_unlock(&rtr->lp_lock);
 773                                 break;
 774                         }
 775                         spin_unlock(&rtr->lp_lock);
 776                 }
 777
 778                 lnet_net_unlock(cpt);
 779
 780                 if (all_known)
 781                         return;
 782
 783                 set_current_state(TASK_UNINTERRUPTIBLE);
 784                 schedule_timeout(cfs_time_seconds(1));
 785         }
 786 }
 787
 788 /* TODO: reimplement */
 789 void
 790 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
 791 {
 792         struct lnet_route *rte;
 793         struct lnet_peer *lp;
 794
 795         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0)
 796                 lp = gw->lpni_peer_net->lpn_peer;
 797         else
 798                 return;
 799
 800         list_for_each_entry(rte, &lp->lp_routes, lr_gwlist) {
 801                 if (rte->lr_net == net) {
 802                         rte->lr_downis = 0;
 803                         break;
 804                 }
 805         }
 806 }
 807
 808 static void
 809 lnet_update_ni_status_locked(void)
 810 {
 811         struct lnet_ni *ni = NULL;
 812         time64_t now;
 813         time64_t timeout;
 814
 815         LASSERT(the_lnet.ln_routing);
 816
 817         timeout = router_ping_timeout +
 818                   MAX(live_router_check_interval, dead_router_check_interval);
 819
 820         now = ktime_get_real_seconds();
 821         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 822                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
 823                         continue;
 824
 825                 if (now < ni->ni_last_alive + timeout)
 826                         continue;
 827
 828                 lnet_ni_lock(ni);
 829                 /* re-check with lock */
 830                 if (now < ni->ni_last_alive + timeout) {
 831                         lnet_ni_unlock(ni);
 832                         continue;
 833                 }
 834
 835                 LASSERT(ni->ni_status != NULL);
 836
 837                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 838                         CDEBUG(D_NET, "NI(%s:%lld) status changed to down\n",
 839                                libcfs_nid2str(ni->ni_nid), timeout);
 840                         /* NB: so far, this is the only place to set
 841                          * NI status to "down" */
 842                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 843                 }
 844                 lnet_ni_unlock(ni);
 845         }
 846 }
 847
 848 int lnet_router_pre_mt_start(void)
 849 {
 850         int rc;
 851
 852         if (check_routers_before_use &&
 853             dead_router_check_interval <= 0) {
 854                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
 855                                    " set if 'check_routers_before_use' is set"
 856                                    "\n");
 857                 return -EINVAL;
 858         }
 859
 860         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
 861         if (rc != 0) {
 862                 CERROR("Can't allocate EQ(0): %d\n", rc);
 863                 return -ENOMEM;
 864         }
 865
 866         return 0;
 867 }
 868
 869 void lnet_router_post_mt_start(void)
 870 {
 871         if (check_routers_before_use) {
 872                 /* Note that a helpful side-effect of pinging all known routers
 873                  * at startup is that it makes them drop stale connections they
 874                  * may have to a previous instance of me. */
 875                 lnet_wait_known_routerstate();
 876         }
 877 }
 878
 879 void
 880 lnet_router_cleanup(void)
 881 {
 882         int rc;
 883
 884         rc = LNetEQFree(the_lnet.ln_rc_eqh);
 885         LASSERT(rc == 0);
 886         return;
 887 }
 888
 889 void
 890 lnet_prune_rc_data(int wait_unlink)
 891 {
 892         wait_unlink = wait_unlink;
 893 }
 894
 895 /*
 896  * This function is called from the monitor thread to check if there are
 897  * any active routers that need to be checked.
 898  */
 899 inline bool
 900 lnet_router_checker_active(void)
 901 {
 902         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 903                 return true;
 904
 905         /* Router Checker thread needs to run when routing is enabled in
 906          * order to call lnet_update_ni_status_locked() */
 907         if (the_lnet.ln_routing)
 908                 return true;
 909
 910         /* if there are routers that need to be cleaned up then do so */
 911         if (!list_empty(&the_lnet.ln_rcd_deathrow) ||
 912             !list_empty(&the_lnet.ln_rcd_zombie))
 913                 return true;
 914
 915         return !list_empty(&the_lnet.ln_routers) &&
 916                 (live_router_check_interval > 0 ||
 917                  dead_router_check_interval > 0);
 918 }
 919
 920 void
 921 lnet_check_routers(void)
 922 {
 923         struct lnet_peer *rtr;
 924         struct list_head *entry;
 925         __u64   version;
 926         int     cpt;
 927
 928         cpt = lnet_net_lock_current();
 929 rescan:
 930         version = the_lnet.ln_routers_version;
 931
 932         list_for_each(entry, &the_lnet.ln_routers) {
 933                 rtr = list_entry(entry, struct lnet_peer,
 934                                  lp_rtr_list);
 935
 936                 /* TODO use discovery to determine if router is alive */
 937
 938                 /* NB dropped lock */
 939                 if (version != the_lnet.ln_routers_version) {
 940                         /* the routers list has changed */
 941                         goto rescan;
 942                 }
 943         }
 944
 945         if (the_lnet.ln_routing)
 946                 lnet_update_ni_status_locked();
 947
 948         lnet_net_unlock(cpt);
 949
 950         lnet_prune_rc_data(0); /* don't wait for UNLINK */
 951 }
 952
 953 void
 954 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
 955 {
 956         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
 957
 958         while (--npages >= 0)
 959                 __free_page(rb->rb_kiov[npages].kiov_page);
 960
 961         LIBCFS_FREE(rb, sz);
 962 }
 963
 964 static struct lnet_rtrbuf *
 965 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
 966 {
 967         int            npages = rbp->rbp_npages;
 968         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
 969         struct page   *page;
 970         struct lnet_rtrbuf *rb;
 971         int            i;
 972
 973         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
 974         if (rb == NULL)
 975                 return NULL;
 976
 977         rb->rb_pool = rbp;
 978
 979         for (i = 0; i < npages; i++) {
 980                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
 981                                           GFP_KERNEL | __GFP_ZERO);
 982                 if (page == NULL) {
 983                         while (--i >= 0)
 984                                 __free_page(rb->rb_kiov[i].kiov_page);
 985
 986                         LIBCFS_FREE(rb, sz);
 987                         return NULL;
 988                 }
 989
 990                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
 991                 rb->rb_kiov[i].kiov_offset = 0;
 992                 rb->rb_kiov[i].kiov_page = page;
 993         }
 994
 995         return rb;
 996 }
 997
 998 static void
 999 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1000 {
1001         int npages = rbp->rbp_npages;
1002         struct lnet_rtrbuf *rb;
1003         struct list_head tmp;
1004
1005         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1006                 return;
1007
1008         INIT_LIST_HEAD(&tmp);
1009
1010         lnet_net_lock(cpt);
1011         list_splice_init(&rbp->rbp_msgs, &tmp);
1012         lnet_drop_routed_msgs_locked(&tmp, cpt);
1013         list_splice_init(&rbp->rbp_bufs, &tmp);
1014         rbp->rbp_req_nbuffers = 0;
1015         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1016         rbp->rbp_mincredits = 0;
1017         lnet_net_unlock(cpt);
1018
1019         /* Free buffers on the free list. */
1020         while (!list_empty(&tmp)) {
1021                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1022                 list_del(&rb->rb_list);
1023                 lnet_destroy_rtrbuf(rb, npages);
1024         }
1025 }
1026
1027 static int
1028 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1029 {
1030         struct list_head rb_list;
1031         struct lnet_rtrbuf *rb;
1032         int             num_rb;
1033         int             num_buffers = 0;
1034         int             old_req_nbufs;
1035         int             npages = rbp->rbp_npages;
1036
1037         lnet_net_lock(cpt);
1038         /* If we are called for less buffers than already in the pool, we
1039          * just lower the req_nbuffers number and excess buffers will be
1040          * thrown away as they are returned to the free list.  Credits
1041          * then get adjusted as well.
1042          * If we already have enough buffers allocated to serve the
1043          * increase requested, then we can treat that the same way as we
1044          * do the decrease. */
1045         num_rb = nbufs - rbp->rbp_nbuffers;
1046         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1047                 rbp->rbp_req_nbuffers = nbufs;
1048                 lnet_net_unlock(cpt);
1049                 return 0;
1050         }
1051         /* store the older value of rbp_req_nbuffers and then set it to
1052          * the new request to prevent lnet_return_rx_credits_locked() from
1053          * freeing buffers that we need to keep around */
1054         old_req_nbufs = rbp->rbp_req_nbuffers;
1055         rbp->rbp_req_nbuffers = nbufs;
1056         lnet_net_unlock(cpt);
1057
1058         INIT_LIST_HEAD(&rb_list);
1059
1060         /* allocate the buffers on a local list first.  If all buffers are
1061          * allocated successfully then join this list to the rbp buffer
1062          * list.  If not then free all allocated buffers. */
1063         while (num_rb-- > 0) {
1064                 rb = lnet_new_rtrbuf(rbp, cpt);
1065                 if (rb == NULL) {
1066                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1067                                nbufs, npages);
1068
1069                         lnet_net_lock(cpt);
1070                         rbp->rbp_req_nbuffers = old_req_nbufs;
1071                         lnet_net_unlock(cpt);
1072
1073                         goto failed;
1074                 }
1075
1076                 list_add(&rb->rb_list, &rb_list);
1077                 num_buffers++;
1078         }
1079
1080         lnet_net_lock(cpt);
1081
1082         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1083         rbp->rbp_nbuffers += num_buffers;
1084         rbp->rbp_credits += num_buffers;
1085         rbp->rbp_mincredits = rbp->rbp_credits;
1086         /* We need to schedule blocked msg using the newly
1087          * added buffers. */
1088         while (!list_empty(&rbp->rbp_bufs) &&
1089                !list_empty(&rbp->rbp_msgs))
1090                 lnet_schedule_blocked_locked(rbp);
1091
1092         lnet_net_unlock(cpt);
1093
1094         return 0;
1095
1096 failed:
1097         while (!list_empty(&rb_list)) {
1098                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1099                 list_del(&rb->rb_list);
1100                 lnet_destroy_rtrbuf(rb, npages);
1101         }
1102
1103         return -ENOMEM;
1104 }
1105
1106 static void
1107 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1108 {
1109         INIT_LIST_HEAD(&rbp->rbp_msgs);
1110         INIT_LIST_HEAD(&rbp->rbp_bufs);
1111
1112         rbp->rbp_npages = npages;
1113         rbp->rbp_credits = 0;
1114         rbp->rbp_mincredits = 0;
1115 }
1116
1117 void
1118 lnet_rtrpools_free(int keep_pools)
1119 {
1120         struct lnet_rtrbufpool *rtrp;
1121         int               i;
1122
1123         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1124                 return;
1125
1126         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1127                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1128                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1129                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1130         }
1131
1132         if (!keep_pools) {
1133                 cfs_percpt_free(the_lnet.ln_rtrpools);
1134                 the_lnet.ln_rtrpools = NULL;
1135         }
1136 }
1137
1138 static int
1139 lnet_nrb_tiny_calculate(void)
1140 {
1141         int     nrbs = LNET_NRB_TINY;
1142
1143         if (tiny_router_buffers < 0) {
1144                 LCONSOLE_ERROR_MSG(0x10c,
1145                                    "tiny_router_buffers=%d invalid when "
1146                                    "routing enabled\n", tiny_router_buffers);
1147                 return -EINVAL;
1148         }
1149
1150         if (tiny_router_buffers > 0)
1151                 nrbs = tiny_router_buffers;
1152
1153         nrbs /= LNET_CPT_NUMBER;
1154         return max(nrbs, LNET_NRB_TINY_MIN);
1155 }
1156
1157 static int
1158 lnet_nrb_small_calculate(void)
1159 {
1160         int     nrbs = LNET_NRB_SMALL;
1161
1162         if (small_router_buffers < 0) {
1163                 LCONSOLE_ERROR_MSG(0x10c,
1164                                    "small_router_buffers=%d invalid when "
1165                                    "routing enabled\n", small_router_buffers);
1166                 return -EINVAL;
1167         }
1168
1169         if (small_router_buffers > 0)
1170                 nrbs = small_router_buffers;
1171
1172         nrbs /= LNET_CPT_NUMBER;
1173         return max(nrbs, LNET_NRB_SMALL_MIN);
1174 }
1175
1176 static int
1177 lnet_nrb_large_calculate(void)
1178 {
1179         int     nrbs = LNET_NRB_LARGE;
1180
1181         if (large_router_buffers < 0) {
1182                 LCONSOLE_ERROR_MSG(0x10c,
1183                                    "large_router_buffers=%d invalid when "
1184                                    "routing enabled\n", large_router_buffers);
1185                 return -EINVAL;
1186         }
1187
1188         if (large_router_buffers > 0)
1189                 nrbs = large_router_buffers;
1190
1191         nrbs /= LNET_CPT_NUMBER;
1192         return max(nrbs, LNET_NRB_LARGE_MIN);
1193 }
1194
1195 int
1196 lnet_rtrpools_alloc(int im_a_router)
1197 {
1198         struct lnet_rtrbufpool *rtrp;
1199         int     nrb_tiny;
1200         int     nrb_small;
1201         int     nrb_large;
1202         int     rc;
1203         int     i;
1204
1205         if (!strcmp(forwarding, "")) {
1206                 /* not set either way */
1207                 if (!im_a_router)
1208                         return 0;
1209         } else if (!strcmp(forwarding, "disabled")) {
1210                 /* explicitly disabled */
1211                 return 0;
1212         } else if (!strcmp(forwarding, "enabled")) {
1213                 /* explicitly enabled */
1214         } else {
1215                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1216                                    "'enabled' or 'disabled'\n");
1217                 return -EINVAL;
1218         }
1219
1220         nrb_tiny = lnet_nrb_tiny_calculate();
1221         if (nrb_tiny < 0)
1222                 return -EINVAL;
1223
1224         nrb_small = lnet_nrb_small_calculate();
1225         if (nrb_small < 0)
1226                 return -EINVAL;
1227
1228         nrb_large = lnet_nrb_large_calculate();
1229         if (nrb_large < 0)
1230                 return -EINVAL;
1231
1232         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1233                                                 LNET_NRBPOOLS *
1234                                                 sizeof(struct lnet_rtrbufpool));
1235         if (the_lnet.ln_rtrpools == NULL) {
1236                 LCONSOLE_ERROR_MSG(0x10c,
1237                                    "Failed to initialize router buffe pool\n");
1238                 return -ENOMEM;
1239         }
1240
1241         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1242                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1243                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1244                                               nrb_tiny, i);
1245                 if (rc != 0)
1246                         goto failed;
1247
1248                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1249                                   LNET_NRB_SMALL_PAGES);
1250                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1251                                               nrb_small, i);
1252                 if (rc != 0)
1253                         goto failed;
1254
1255                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1256                                   LNET_NRB_LARGE_PAGES);
1257                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1258                                               nrb_large, i);
1259                 if (rc != 0)
1260                         goto failed;
1261         }
1262
1263         lnet_net_lock(LNET_LOCK_EX);
1264         the_lnet.ln_routing = 1;
1265         lnet_net_unlock(LNET_LOCK_EX);
1266         wake_up(&the_lnet.ln_mt_waitq);
1267         return 0;
1268
1269  failed:
1270         lnet_rtrpools_free(0);
1271         return rc;
1272 }
1273
1274 static int
1275 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1276 {
1277         int nrb = 0;
1278         int rc = 0;
1279         int i;
1280         struct lnet_rtrbufpool *rtrp;
1281
1282         /* If the provided values for each buffer pool are different than the
1283          * configured values, we need to take action. */
1284         if (tiny >= 0) {
1285                 tiny_router_buffers = tiny;
1286                 nrb = lnet_nrb_tiny_calculate();
1287                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1288                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1289                                                       nrb, i);
1290                         if (rc != 0)
1291                                 return rc;
1292                 }
1293         }
1294         if (small >= 0) {
1295                 small_router_buffers = small;
1296                 nrb = lnet_nrb_small_calculate();
1297                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1298                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1299                                                       nrb, i);
1300                         if (rc != 0)
1301                                 return rc;
1302                 }
1303         }
1304         if (large >= 0) {
1305                 large_router_buffers = large;
1306                 nrb = lnet_nrb_large_calculate();
1307                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1308                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1309                                                       nrb, i);
1310                         if (rc != 0)
1311                                 return rc;
1312                 }
1313         }
1314
1315         return 0;
1316 }
1317
1318 int
1319 lnet_rtrpools_adjust(int tiny, int small, int large)
1320 {
1321         /* this function doesn't revert the changes if adding new buffers
1322          * failed.  It's up to the user space caller to revert the
1323          * changes. */
1324
1325         if (!the_lnet.ln_routing)
1326                 return 0;
1327
1328         return lnet_rtrpools_adjust_helper(tiny, small, large);
1329 }
1330
1331 int
1332 lnet_rtrpools_enable(void)
1333 {
1334         int rc = 0;
1335
1336         if (the_lnet.ln_routing)
1337                 return 0;
1338
1339         if (the_lnet.ln_rtrpools == NULL)
1340                 /* If routing is turned off, and we have never
1341                  * initialized the pools before, just call the
1342                  * standard buffer pool allocation routine as
1343                  * if we are just configuring this for the first
1344                  * time. */
1345                 rc = lnet_rtrpools_alloc(1);
1346         else
1347                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1348         if (rc != 0)
1349                 return rc;
1350
1351         lnet_net_lock(LNET_LOCK_EX);
1352         the_lnet.ln_routing = 1;
1353
1354         the_lnet.ln_ping_target->pb_info.pi_features &=
1355                 ~LNET_PING_FEAT_RTE_DISABLED;
1356         lnet_net_unlock(LNET_LOCK_EX);
1357
1358         return rc;
1359 }
1360
1361 void
1362 lnet_rtrpools_disable(void)
1363 {
1364         if (!the_lnet.ln_routing)
1365                 return;
1366
1367         lnet_net_lock(LNET_LOCK_EX);
1368         the_lnet.ln_routing = 0;
1369         the_lnet.ln_ping_target->pb_info.pi_features |=
1370                 LNET_PING_FEAT_RTE_DISABLED;
1371
1372         tiny_router_buffers = 0;
1373         small_router_buffers = 0;
1374         large_router_buffers = 0;
1375         lnet_net_unlock(LNET_LOCK_EX);
1376         lnet_rtrpools_free(1);
1377 }
1378
1379 int
1380 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, time64_t when)
1381 {
1382         struct lnet_peer_ni *lp = NULL;
1383         time64_t now = ktime_get_seconds();
1384         int cpt = lnet_cpt_of_nid(nid, ni);
1385
1386         LASSERT (!in_interrupt ());
1387
1388         CDEBUG (D_NET, "%s notifying %s: %s\n",
1389                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1390                 libcfs_nid2str(nid),
1391                 alive ? "up" : "down");
1392
1393         if (ni != NULL &&
1394             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1395                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1396                       libcfs_nid2str(nid), alive ? "birth" : "death",
1397                       libcfs_nid2str(ni->ni_nid));
1398                 return -EINVAL;
1399         }
1400
1401         /* can't do predictions... */
1402         if (when > now) {
1403                 CWARN("Ignoring prediction from %s of %s %s "
1404                       "%lld seconds in the future\n",
1405                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1406                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1407                 return -EINVAL;
1408         }
1409
1410         if (ni != NULL && !alive &&             /* LND telling me she's down */
1411             !auto_down) {                       /* auto-down disabled */
1412                 CDEBUG(D_NET, "Auto-down disabled\n");
1413                 return 0;
1414         }
1415
1416         lnet_net_lock(cpt);
1417
1418         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1419                 lnet_net_unlock(cpt);
1420                 return -ESHUTDOWN;
1421         }
1422
1423         lp = lnet_find_peer_ni_locked(nid);
1424         if (lp == NULL) {
1425                 /* nid not found */
1426                 lnet_net_unlock(cpt);
1427                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1428                 return 0;
1429         }
1430
1431         /*
1432          * It is possible for this function to be called for the same peer
1433          * but with different NIs. We want to synchronize the notification
1434          * between the different calls. So we will use the lpni_cpt to
1435          * grab the net lock.
1436          */
1437         if (lp->lpni_cpt != cpt) {
1438                 lnet_net_unlock(cpt);
1439                 cpt = lp->lpni_cpt;
1440                 lnet_net_lock(cpt);
1441         }
1442
1443         /* We can't fully trust LND on reporting exact peer last_alive
1444          * if he notifies us about dead peer. For example ksocklnd can
1445          * call us with when == _time_when_the_node_was_booted_ if
1446          * no connections were successfully established */
1447         if (ni != NULL && !alive && when < lp->lpni_last_alive)
1448                 when = lp->lpni_last_alive;
1449
1450         lnet_notify_locked(lp, ni == NULL, alive, when);
1451
1452         if (ni != NULL)
1453                 lnet_ni_notify_locked(ni, lp);
1454
1455         lnet_peer_ni_decref_locked(lp);
1456
1457         lnet_net_unlock(cpt);
1458         return 0;
1459 }
1460 EXPORT_SYMBOL(lnet_notify);