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.hpdd.intel.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 #include <lnet/lib-lnet.h>
  25
  26 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
  27 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
  28 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
  29 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
  30 #define LNET_NRB_SMALL_PAGES    1
  31 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  32 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  33 #define LNET_NRB_LARGE_PAGES    ((LNET_MTU + PAGE_SIZE - 1) >> \
  34                                   PAGE_SHIFT)
  35
  36 static char *forwarding = "";
  37 module_param(forwarding, charp, 0444);
  38 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
  39
  40 static int tiny_router_buffers;
  41 module_param(tiny_router_buffers, int, 0444);
  42 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
  43 static int small_router_buffers;
  44 module_param(small_router_buffers, int, 0444);
  45 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
  46 static int large_router_buffers;
  47 module_param(large_router_buffers, int, 0444);
  48 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
  49 static int peer_buffer_credits;
  50 module_param(peer_buffer_credits, int, 0444);
  51 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
  52
  53 static int auto_down = 1;
  54 module_param(auto_down, int, 0444);
  55 MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
  56
  57 int
  58 lnet_peer_buffer_credits(struct lnet_net *net)
  59 {
  60         /* NI option overrides LNet default */
  61         if (net->net_tunables.lct_peer_rtr_credits > 0)
  62                 return net->net_tunables.lct_peer_rtr_credits;
  63         if (peer_buffer_credits > 0)
  64                 return peer_buffer_credits;
  65
  66         /* As an approximation, allow this peer the same number of router
  67          * buffers as it is allowed outstanding sends */
  68         return net->net_tunables.lct_peer_tx_credits;
  69 }
  70
  71 /* forward ref's */
  72 static int lnet_router_checker(void *);
  73
  74 static int check_routers_before_use;
  75 module_param(check_routers_before_use, int, 0444);
  76 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
  77
  78 int avoid_asym_router_failure = 1;
  79 module_param(avoid_asym_router_failure, int, 0644);
  80 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
  81
  82 static int dead_router_check_interval = 60;
  83 module_param(dead_router_check_interval, int, 0644);
  84 MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)");
  85
  86 static int live_router_check_interval = 60;
  87 module_param(live_router_check_interval, int, 0644);
  88 MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
  89
  90 static int router_ping_timeout = 50;
  91 module_param(router_ping_timeout, int, 0644);
  92 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
  93
  94 int
  95 lnet_peers_start_down(void)
  96 {
  97         return check_routers_before_use;
  98 }
  99
 100 void
 101 lnet_notify_locked(struct lnet_peer_ni *lp, int notifylnd, int alive,
 102                    time64_t when)
 103 {
 104         if (lp->lpni_timestamp > when) { /* out of date information */
 105                 CDEBUG(D_NET, "Out of date\n");
 106                 return;
 107         }
 108
 109         /*
 110          * This function can be called with different cpt locks being
 111          * held. lpni_alive_count modification needs to be properly protected.
 112          * Significant reads to lpni_alive_count are also protected with
 113          * the same lock
 114          */
 115         spin_lock(&lp->lpni_lock);
 116
 117         lp->lpni_timestamp = when; /* update timestamp */
 118         lp->lpni_ping_deadline = 0;               /* disable ping timeout */
 119
 120         if (lp->lpni_alive_count != 0 &&          /* got old news */
 121             (!lp->lpni_alive) == (!alive)) {      /* new date for old news */
 122                 spin_unlock(&lp->lpni_lock);
 123                 CDEBUG(D_NET, "Old news\n");
 124                 return;
 125         }
 126
 127         /* Flag that notification is outstanding */
 128
 129         lp->lpni_alive_count++;
 130         lp->lpni_alive = (alive) ? 1 : 0;
 131         lp->lpni_notify = 1;
 132         lp->lpni_notifylnd = notifylnd;
 133         if (lp->lpni_alive)
 134                 lp->lpni_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 135
 136         spin_unlock(&lp->lpni_lock);
 137
 138         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive);
 139 }
 140
 141 /*
 142  * This function will always be called with lp->lpni_cpt lock held.
 143  */
 144 static void
 145 lnet_ni_notify_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
 146 {
 147         int alive;
 148         int notifylnd;
 149
 150         /* Notify only in 1 thread at any time to ensure ordered notification.
 151          * NB individual events can be missed; the only guarantee is that you
 152          * always get the most recent news */
 153
 154         spin_lock(&lp->lpni_lock);
 155
 156         if (lp->lpni_notifying || ni == NULL) {
 157                 spin_unlock(&lp->lpni_lock);
 158                 return;
 159         }
 160
 161         lp->lpni_notifying = 1;
 162
 163         /*
 164          * lp->lpni_notify needs to be protected because it can be set in
 165          * lnet_notify_locked().
 166          */
 167         while (lp->lpni_notify) {
 168                 alive     = lp->lpni_alive;
 169                 notifylnd = lp->lpni_notifylnd;
 170
 171                 lp->lpni_notifylnd = 0;
 172                 lp->lpni_notify    = 0;
 173
 174                 if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
 175                         spin_unlock(&lp->lpni_lock);
 176                         lnet_net_unlock(lp->lpni_cpt);
 177
 178                         /* A new notification could happen now; I'll handle it
 179                          * when control returns to me */
 180
 181                         (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid,
 182                                                           alive);
 183
 184                         lnet_net_lock(lp->lpni_cpt);
 185                         spin_lock(&lp->lpni_lock);
 186                 }
 187         }
 188
 189         lp->lpni_notifying = 0;
 190         spin_unlock(&lp->lpni_lock);
 191 }
 192
 193 static void
 194 lnet_rtr_addref_locked(struct lnet_peer_ni *lp)
 195 {
 196         LASSERT(atomic_read(&lp->lpni_refcount) > 0);
 197         LASSERT(lp->lpni_rtr_refcount >= 0);
 198
 199         /* lnet_net_lock must be exclusively locked */
 200         lp->lpni_rtr_refcount++;
 201         if (lp->lpni_rtr_refcount == 1) {
 202                 struct list_head *pos;
 203
 204                 /* a simple insertion sort */
 205                 list_for_each_prev(pos, &the_lnet.ln_routers) {
 206                         struct lnet_peer_ni *rtr;
 207
 208                         rtr = list_entry(pos, struct lnet_peer_ni,
 209                                          lpni_rtr_list);
 210                         if (rtr->lpni_nid < lp->lpni_nid)
 211                                 break;
 212                 }
 213
 214                 list_add(&lp->lpni_rtr_list, pos);
 215                 /* addref for the_lnet.ln_routers */
 216                 lnet_peer_ni_addref_locked(lp);
 217                 the_lnet.ln_routers_version++;
 218         }
 219 }
 220
 221 static void
 222 lnet_rtr_decref_locked(struct lnet_peer_ni *lp)
 223 {
 224         LASSERT(atomic_read(&lp->lpni_refcount) > 0);
 225         LASSERT(lp->lpni_rtr_refcount > 0);
 226
 227         /* lnet_net_lock must be exclusively locked */
 228         lp->lpni_rtr_refcount--;
 229         if (lp->lpni_rtr_refcount == 0) {
 230                 LASSERT(list_empty(&lp->lpni_routes));
 231
 232                 if (lp->lpni_rcd != NULL) {
 233                         list_add(&lp->lpni_rcd->rcd_list,
 234                                  &the_lnet.ln_rcd_deathrow);
 235                         lp->lpni_rcd = NULL;
 236                 }
 237
 238                 list_del(&lp->lpni_rtr_list);
 239                 /* decref for the_lnet.ln_routers */
 240                 lnet_peer_ni_decref_locked(lp);
 241                 the_lnet.ln_routers_version++;
 242         }
 243 }
 244
 245 struct lnet_remotenet *
 246 lnet_find_rnet_locked(__u32 net)
 247 {
 248         struct lnet_remotenet *rnet;
 249         struct list_head *tmp;
 250         struct list_head *rn_list;
 251
 252         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 253
 254         rn_list = lnet_net2rnethash(net);
 255         list_for_each(tmp, rn_list) {
 256                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 257
 258                 if (rnet->lrn_net == net)
 259                         return rnet;
 260         }
 261         return NULL;
 262 }
 263
 264 static void lnet_shuffle_seed(void)
 265 {
 266         static int seeded;
 267         __u32 lnd_type;
 268         __u32 seed[2];
 269         struct timespec64 ts;
 270         struct lnet_ni *ni = NULL;
 271
 272         if (seeded)
 273                 return;
 274
 275         cfs_get_random_bytes(seed, sizeof(seed));
 276
 277         /* Nodes with small feet have little entropy
 278          * the NID for this node gives the most entropy in the low bits */
 279         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 280                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 281
 282                 if (lnd_type != LOLND)
 283                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
 284         }
 285
 286         ktime_get_ts64(&ts);
 287         cfs_srand(ts.tv_sec ^ seed[0], ts.tv_nsec ^ seed[1]);
 288         seeded = 1;
 289         return;
 290 }
 291
 292 /* NB expects LNET_LOCK held */
 293 static void
 294 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 295 {
 296         unsigned int      len = 0;
 297         unsigned int      offset = 0;
 298         struct list_head *e;
 299
 300         lnet_shuffle_seed();
 301
 302         list_for_each(e, &rnet->lrn_routes) {
 303                 len++;
 304         }
 305
 306         /* len+1 positions to add a new entry, also prevents division by 0 */
 307         offset = cfs_rand() % (len + 1);
 308         list_for_each(e, &rnet->lrn_routes) {
 309                 if (offset == 0)
 310                         break;
 311                 offset--;
 312         }
 313         list_add(&route->lr_list, e);
 314         list_add(&route->lr_gwlist, &route->lr_gateway->lpni_routes);
 315
 316         the_lnet.ln_remote_nets_version++;
 317         lnet_rtr_addref_locked(route->lr_gateway);
 318 }
 319
 320 int
 321 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 322                unsigned int priority)
 323 {
 324         struct list_head        *e;
 325         struct lnet_remotenet   *rnet;
 326         struct lnet_remotenet   *rnet2;
 327         struct lnet_route               *route;
 328         struct lnet_ni          *ni;
 329         struct lnet_peer_ni     *lpni;
 330         int                     add_route;
 331         int                     rc;
 332
 333         CDEBUG(D_NET, "Add route: net %s hops %d priority %u gw %s\n",
 334                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 335
 336         if (gateway == LNET_NID_ANY ||
 337             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 338             net == LNET_NIDNET(LNET_NID_ANY) ||
 339             LNET_NETTYP(net) == LOLND ||
 340             LNET_NIDNET(gateway) == net ||
 341             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 342                 return -EINVAL;
 343
 344         if (lnet_islocalnet(net))       /* it's a local network */
 345                 return -EEXIST;
 346
 347         /* Assume net, route, all new */
 348         LIBCFS_ALLOC(route, sizeof(*route));
 349         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 350         if (route == NULL || rnet == NULL) {
 351                 CERROR("Out of memory creating route %s %d %s\n",
 352                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 353                 if (route != NULL)
 354                         LIBCFS_FREE(route, sizeof(*route));
 355                 if (rnet != NULL)
 356                         LIBCFS_FREE(rnet, sizeof(*rnet));
 357                 return -ENOMEM;
 358         }
 359
 360         INIT_LIST_HEAD(&rnet->lrn_routes);
 361         rnet->lrn_net = net;
 362         route->lr_hops = hops;
 363         route->lr_net = net;
 364         route->lr_priority = priority;
 365
 366         lnet_net_lock(LNET_LOCK_EX);
 367
 368         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 369         if (IS_ERR(lpni)) {
 370                 lnet_net_unlock(LNET_LOCK_EX);
 371
 372                 LIBCFS_FREE(route, sizeof(*route));
 373                 LIBCFS_FREE(rnet, sizeof(*rnet));
 374
 375                 rc = PTR_ERR(lpni);
 376                 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
 377                         return rc;       /* ignore the route entry */
 378                 CERROR("Error %d creating route %s %d %s\n", rc,
 379                         libcfs_net2str(net), hops,
 380                         libcfs_nid2str(gateway));
 381                 return rc;
 382         }
 383         route->lr_gateway = lpni;
 384         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 385
 386         rnet2 = lnet_find_rnet_locked(net);
 387         if (rnet2 == NULL) {
 388                 /* new network */
 389                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 390                 rnet2 = rnet;
 391         }
 392
 393         /* Search for a duplicate route (it's a NOOP if it is) */
 394         add_route = 1;
 395         list_for_each(e, &rnet2->lrn_routes) {
 396                 struct lnet_route *route2;
 397
 398                 route2 = list_entry(e, struct lnet_route, lr_list);
 399                 if (route2->lr_gateway == route->lr_gateway) {
 400                         add_route = 0;
 401                         break;
 402                 }
 403
 404                 /* our lookups must be true */
 405                 LASSERT(route2->lr_gateway->lpni_nid != gateway);
 406         }
 407
 408         if (add_route) {
 409                 lnet_peer_ni_addref_locked(route->lr_gateway); /* +1 for notify */
 410                 lnet_add_route_to_rnet(rnet2, route);
 411
 412                 ni = lnet_get_next_ni_locked(route->lr_gateway->lpni_net, NULL);
 413                 lnet_net_unlock(LNET_LOCK_EX);
 414
 415                 /* XXX Assume alive */
 416                 if (ni->ni_net->net_lnd->lnd_notify != NULL)
 417                         (ni->ni_net->net_lnd->lnd_notify)(ni, gateway, 1);
 418
 419                 lnet_net_lock(LNET_LOCK_EX);
 420         }
 421
 422         /* -1 for notify or !add_route */
 423         lnet_peer_ni_decref_locked(route->lr_gateway);
 424         lnet_net_unlock(LNET_LOCK_EX);
 425
 426         rc = 0;
 427
 428         if (!add_route) {
 429                 rc = -EEXIST;
 430                 LIBCFS_FREE(route, sizeof(*route));
 431         }
 432
 433         if (rnet != rnet2)
 434                 LIBCFS_FREE(rnet, sizeof(*rnet));
 435
 436         /* indicate to startup the router checker if configured */
 437         wake_up(&the_lnet.ln_rc_waitq);
 438
 439         return rc;
 440 }
 441
 442 int
 443 lnet_check_routes(void)
 444 {
 445         struct lnet_remotenet *rnet;
 446         struct lnet_route        *route;
 447         struct lnet_route        *route2;
 448         struct list_head *e1;
 449         struct list_head *e2;
 450         int               cpt;
 451         struct list_head *rn_list;
 452         int               i;
 453
 454         cpt = lnet_net_lock_current();
 455
 456         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 457                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 458                 list_for_each(e1, rn_list) {
 459                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 460
 461                         route2 = NULL;
 462                         list_for_each(e2, &rnet->lrn_routes) {
 463                                 lnet_nid_t      nid1;
 464                                 lnet_nid_t      nid2;
 465                                 int             net;
 466
 467                                 route = list_entry(e2, struct lnet_route,
 468                                                    lr_list);
 469
 470                                 if (route2 == NULL) {
 471                                         route2 = route;
 472                                         continue;
 473                                 }
 474
 475                                 if (route->lr_gateway->lpni_net ==
 476                                     route2->lr_gateway->lpni_net)
 477                                         continue;
 478
 479                                 nid1 = route->lr_gateway->lpni_nid;
 480                                 nid2 = route2->lr_gateway->lpni_nid;
 481                                 net = rnet->lrn_net;
 482
 483                                 lnet_net_unlock(cpt);
 484
 485                                 CERROR("Routes to %s via %s and %s not "
 486                                        "supported\n",
 487                                        libcfs_net2str(net),
 488                                        libcfs_nid2str(nid1),
 489                                        libcfs_nid2str(nid2));
 490                                 return -EINVAL;
 491                         }
 492                 }
 493         }
 494
 495         lnet_net_unlock(cpt);
 496         return 0;
 497 }
 498
 499 int
 500 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 501 {
 502         struct lnet_peer_ni     *gateway;
 503         struct lnet_remotenet   *rnet;
 504         struct lnet_route               *route;
 505         struct list_head        *e1;
 506         struct list_head        *e2;
 507         int                     rc = -ENOENT;
 508         struct list_head        *rn_list;
 509         int                     idx = 0;
 510
 511         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 512                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 513
 514         /* NB Caller may specify either all routes via the given gateway
 515          * or a specific route entry actual NIDs) */
 516
 517         lnet_net_lock(LNET_LOCK_EX);
 518         if (net == LNET_NIDNET(LNET_NID_ANY))
 519                 rn_list = &the_lnet.ln_remote_nets_hash[0];
 520         else
 521                 rn_list = lnet_net2rnethash(net);
 522
 523 again:
 524         list_for_each(e1, rn_list) {
 525                 rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 526
 527                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 528                         net == rnet->lrn_net))
 529                         continue;
 530
 531                 list_for_each(e2, &rnet->lrn_routes) {
 532                         route = list_entry(e2, struct lnet_route, lr_list);
 533
 534                         gateway = route->lr_gateway;
 535                         if (!(gw_nid == LNET_NID_ANY ||
 536                               gw_nid == gateway->lpni_nid))
 537                                 continue;
 538
 539                         list_del(&route->lr_list);
 540                         list_del(&route->lr_gwlist);
 541                         the_lnet.ln_remote_nets_version++;
 542
 543                         if (list_empty(&rnet->lrn_routes))
 544                                 list_del(&rnet->lrn_list);
 545                         else
 546                                 rnet = NULL;
 547
 548                         lnet_rtr_decref_locked(gateway);
 549                         lnet_peer_ni_decref_locked(gateway);
 550
 551                         lnet_net_unlock(LNET_LOCK_EX);
 552
 553                         LIBCFS_FREE(route, sizeof(*route));
 554
 555                         if (rnet != NULL)
 556                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 557
 558                         rc = 0;
 559                         lnet_net_lock(LNET_LOCK_EX);
 560                         goto again;
 561                 }
 562         }
 563
 564         if (net == LNET_NIDNET(LNET_NID_ANY) &&
 565             ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
 566                 rn_list = &the_lnet.ln_remote_nets_hash[idx];
 567                 goto again;
 568         }
 569         lnet_net_unlock(LNET_LOCK_EX);
 570
 571         return rc;
 572 }
 573
 574 void
 575 lnet_destroy_routes (void)
 576 {
 577         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 578 }
 579
 580 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
 581 {
 582         int i, rc = -ENOENT, j;
 583
 584         if (the_lnet.ln_rtrpools == NULL)
 585                 return rc;
 586
 587         for (i = 0; i < LNET_NRBPOOLS; i++) {
 588                 struct lnet_rtrbufpool *rbp;
 589
 590                 lnet_net_lock(LNET_LOCK_EX);
 591                 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
 592                         if (i++ != idx)
 593                                 continue;
 594
 595                         pool_cfg->pl_pools[i].pl_npages = rbp[i].rbp_npages;
 596                         pool_cfg->pl_pools[i].pl_nbuffers = rbp[i].rbp_nbuffers;
 597                         pool_cfg->pl_pools[i].pl_credits = rbp[i].rbp_credits;
 598                         pool_cfg->pl_pools[i].pl_mincredits = rbp[i].rbp_mincredits;
 599                         rc = 0;
 600                         break;
 601                 }
 602                 lnet_net_unlock(LNET_LOCK_EX);
 603         }
 604
 605         lnet_net_lock(LNET_LOCK_EX);
 606         pool_cfg->pl_routing = the_lnet.ln_routing;
 607         lnet_net_unlock(LNET_LOCK_EX);
 608
 609         return rc;
 610 }
 611
 612 int
 613 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 614                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 615 {
 616         struct list_head *e1;
 617         struct list_head *e2;
 618         struct lnet_remotenet *rnet;
 619         struct lnet_route        *route;
 620         int               cpt;
 621         int               i;
 622         struct list_head *rn_list;
 623
 624         cpt = lnet_net_lock_current();
 625
 626         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 627                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 628                 list_for_each(e1, rn_list) {
 629                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 630
 631                         list_for_each(e2, &rnet->lrn_routes) {
 632                                 route = list_entry(e2, struct lnet_route,
 633                                                    lr_list);
 634
 635                                 if (idx-- == 0) {
 636                                         *net      = rnet->lrn_net;
 637                                         *hops     = route->lr_hops;
 638                                         *priority = route->lr_priority;
 639                                         *gateway  = route->lr_gateway->lpni_nid;
 640                                         *alive    = lnet_is_route_alive(route);
 641                                         lnet_net_unlock(cpt);
 642                                         return 0;
 643                                 }
 644                         }
 645                 }
 646         }
 647
 648         lnet_net_unlock(cpt);
 649         return -ENOENT;
 650 }
 651
 652 void
 653 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
 654 {
 655         struct lnet_ni_status *stat;
 656         int nnis;
 657         int i;
 658
 659         __swab32s(&pbuf->pb_info.pi_magic);
 660         __swab32s(&pbuf->pb_info.pi_features);
 661         __swab32s(&pbuf->pb_info.pi_pid);
 662         __swab32s(&pbuf->pb_info.pi_nnis);
 663         nnis = pbuf->pb_info.pi_nnis;
 664         if (nnis > pbuf->pb_nnis)
 665                 nnis = pbuf->pb_nnis;
 666         for (i = 0; i < nnis; i++) {
 667                 stat = &pbuf->pb_info.pi_ni[i];
 668                 __swab64s(&stat->ns_nid);
 669                 __swab32s(&stat->ns_status);
 670         }
 671         return;
 672 }
 673
 674 /**
 675  * parse router-checker pinginfo, record number of down NIs for remote
 676  * networks on that router.
 677  */
 678 static void
 679 lnet_parse_rc_info(struct lnet_rc_data *rcd)
 680 {
 681         struct lnet_ping_buffer *pbuf = rcd->rcd_pingbuffer;
 682         struct lnet_peer_ni     *gw   = rcd->rcd_gateway;
 683         struct lnet_route               *rte;
 684         int                     nnis;
 685
 686         if (!gw->lpni_alive || !pbuf)
 687                 return;
 688
 689         /*
 690          * Protect gw->lpni_ping_feats. This can be set from
 691          * lnet_notify_locked with different locks being held
 692          */
 693         spin_lock(&gw->lpni_lock);
 694
 695         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 696                 lnet_swap_pinginfo(pbuf);
 697
 698         /* NB always racing with network! */
 699         if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
 700                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 701                        libcfs_nid2str(gw->lpni_nid), pbuf->pb_info.pi_magic);
 702                 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 703                 goto out;
 704         }
 705
 706         gw->lpni_ping_feats = pbuf->pb_info.pi_features;
 707
 708         /* Without NI status info there's nothing more to do. */
 709         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
 710                 goto out;
 711
 712         /* Determine the number of NIs for which there is data. */
 713         nnis = pbuf->pb_info.pi_nnis;
 714         if (pbuf->pb_nnis < nnis) {
 715                 if (rcd->rcd_nnis < nnis)
 716                         rcd->rcd_nnis = nnis;
 717                 nnis = pbuf->pb_nnis;
 718         }
 719
 720         list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 721                 int     down = 0;
 722                 int     up = 0;
 723                 int     i;
 724
 725                 /* If routing disabled then the route is down. */
 726                 if ((gw->lpni_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
 727                         rte->lr_downis = 1;
 728                         continue;
 729                 }
 730
 731                 for (i = 0; i < nnis; i++) {
 732                         struct lnet_ni_status *stat = &pbuf->pb_info.pi_ni[i];
 733                         lnet_nid_t       nid = stat->ns_nid;
 734
 735                         if (nid == LNET_NID_ANY) {
 736                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 737                                        libcfs_nid2str(gw->lpni_nid));
 738                                 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 739                                 goto out;
 740                         }
 741
 742                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 743                                 continue;
 744
 745                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 746                                 down++;
 747                                 continue;
 748                         }
 749
 750                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 751                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 752                                         up = 1;
 753                                         break;
 754                                 }
 755                                 continue;
 756                         }
 757
 758                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 759                                libcfs_nid2str(gw->lpni_nid), stat->ns_status);
 760                         gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 761                         goto out;
 762                 }
 763
 764                 if (up) { /* ignore downed NIs if NI for dest network is up */
 765                         rte->lr_downis = 0;
 766                         continue;
 767                 }
 768                 /* if @down is zero and this route is single-hop, it means
 769                  * we can't find NI for target network */
 770                 if (down == 0 && rte->lr_hops == 1)
 771                         down = 1;
 772
 773                 rte->lr_downis = down;
 774         }
 775 out:
 776         spin_unlock(&gw->lpni_lock);
 777 }
 778
 779 static void
 780 lnet_router_checker_event(struct lnet_event *event)
 781 {
 782         struct lnet_rc_data *rcd = event->md.user_ptr;
 783         struct lnet_peer_ni *lp;
 784
 785         LASSERT(rcd != NULL);
 786
 787         if (event->unlinked) {
 788                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 789                 return;
 790         }
 791
 792         LASSERT(event->type == LNET_EVENT_SEND ||
 793                 event->type == LNET_EVENT_REPLY);
 794
 795         lp = rcd->rcd_gateway;
 796         LASSERT(lp != NULL);
 797
 798          /* NB: it's called with holding lnet_res_lock, we have a few
 799           * places need to hold both locks at the same time, please take
 800           * care of lock ordering */
 801         lnet_net_lock(lp->lpni_cpt);
 802         if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
 803                 /* ignore if no longer a router or rcd is replaced */
 804                 goto out;
 805         }
 806
 807         if (event->type == LNET_EVENT_SEND) {
 808                 lp->lpni_ping_notsent = 0;
 809                 if (event->status == 0)
 810                         goto out;
 811         }
 812
 813         /* LNET_EVENT_REPLY */
 814         /* A successful REPLY means the router is up.  If _any_ comms
 815          * to the router fail I assume it's down (this will happen if
 816          * we ping alive routers to try to detect router death before
 817          * apps get burned). */
 818
 819         lnet_notify_locked(lp, 1, !event->status, ktime_get_seconds());
 820         /* The router checker will wake up very shortly and do the
 821          * actual notification.
 822          * XXX If 'lp' stops being a router before then, it will still
 823          * have the notification pending!!! */
 824
 825         if (avoid_asym_router_failure && event->status == 0)
 826                 lnet_parse_rc_info(rcd);
 827
 828  out:
 829         lnet_net_unlock(lp->lpni_cpt);
 830 }
 831
 832 static void
 833 lnet_wait_known_routerstate(void)
 834 {
 835         struct lnet_peer_ni *rtr;
 836         struct list_head *entry;
 837         int all_known;
 838
 839         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 840
 841         for (;;) {
 842                 int cpt = lnet_net_lock_current();
 843
 844                 all_known = 1;
 845                 list_for_each(entry, &the_lnet.ln_routers) {
 846                         rtr = list_entry(entry, struct lnet_peer_ni,
 847                                          lpni_rtr_list);
 848
 849                         spin_lock(&rtr->lpni_lock);
 850
 851                         if (rtr->lpni_alive_count == 0) {
 852                                 all_known = 0;
 853                                 spin_unlock(&rtr->lpni_lock);
 854                                 break;
 855                         }
 856                         spin_unlock(&rtr->lpni_lock);
 857                 }
 858
 859                 lnet_net_unlock(cpt);
 860
 861                 if (all_known)
 862                         return;
 863
 864                 set_current_state(TASK_UNINTERRUPTIBLE);
 865                 schedule_timeout(cfs_time_seconds(1));
 866         }
 867 }
 868
 869 void
 870 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
 871 {
 872         struct lnet_route *rte;
 873
 874         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
 875                 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 876                         if (rte->lr_net == net) {
 877                                 rte->lr_downis = 0;
 878                                 break;
 879                         }
 880                 }
 881         }
 882 }
 883
 884 static void
 885 lnet_update_ni_status_locked(void)
 886 {
 887         struct lnet_ni *ni = NULL;
 888         time64_t now;
 889         time64_t timeout;
 890
 891         LASSERT(the_lnet.ln_routing);
 892
 893         timeout = router_ping_timeout +
 894                   MAX(live_router_check_interval, dead_router_check_interval);
 895
 896         now = ktime_get_real_seconds();
 897         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 898                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
 899                         continue;
 900
 901                 if (now < ni->ni_last_alive + timeout)
 902                         continue;
 903
 904                 lnet_ni_lock(ni);
 905                 /* re-check with lock */
 906                 if (now < ni->ni_last_alive + timeout) {
 907                         lnet_ni_unlock(ni);
 908                         continue;
 909                 }
 910
 911                 LASSERT(ni->ni_status != NULL);
 912
 913                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 914                         CDEBUG(D_NET, "NI(%s:%lld) status changed to down\n",
 915                                libcfs_nid2str(ni->ni_nid), timeout);
 916                         /* NB: so far, this is the only place to set
 917                          * NI status to "down" */
 918                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 919                 }
 920                 lnet_ni_unlock(ni);
 921         }
 922 }
 923
 924 static void
 925 lnet_destroy_rc_data(struct lnet_rc_data *rcd)
 926 {
 927         LASSERT(list_empty(&rcd->rcd_list));
 928         /* detached from network */
 929         LASSERT(LNetMDHandleIsInvalid(rcd->rcd_mdh));
 930
 931         if (rcd->rcd_gateway != NULL) {
 932                 int cpt = rcd->rcd_gateway->lpni_cpt;
 933
 934                 lnet_net_lock(cpt);
 935                 lnet_peer_ni_decref_locked(rcd->rcd_gateway);
 936                 lnet_net_unlock(cpt);
 937         }
 938
 939         if (rcd->rcd_pingbuffer != NULL)
 940                 lnet_ping_buffer_decref(rcd->rcd_pingbuffer);
 941
 942         LIBCFS_FREE(rcd, sizeof(*rcd));
 943 }
 944
 945 static struct lnet_rc_data *
 946 lnet_update_rc_data_locked(struct lnet_peer_ni *gateway)
 947 {
 948         struct lnet_handle_md mdh;
 949         struct lnet_rc_data *rcd;
 950         struct lnet_ping_buffer *pbuf = NULL;
 951         int nnis = LNET_INTERFACES_MIN;
 952         int rc;
 953         int i;
 954
 955         rcd = gateway->lpni_rcd;
 956         if (rcd) {
 957                 nnis = rcd->rcd_nnis;
 958                 mdh = rcd->rcd_mdh;
 959                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 960                 pbuf = rcd->rcd_pingbuffer;
 961                 rcd->rcd_pingbuffer = NULL;
 962         } else {
 963                 LNetInvalidateMDHandle(&mdh);
 964         }
 965
 966         lnet_net_unlock(gateway->lpni_cpt);
 967
 968         if (rcd) {
 969                 LNetMDUnlink(mdh);
 970                 lnet_ping_buffer_decref(pbuf);
 971         } else {
 972                 LIBCFS_ALLOC(rcd, sizeof(*rcd));
 973                 if (rcd == NULL)
 974                         goto out;
 975
 976                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 977                 INIT_LIST_HEAD(&rcd->rcd_list);
 978                 rcd->rcd_nnis = nnis;
 979         }
 980
 981         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
 982         if (pbuf == NULL)
 983                 goto out;
 984
 985         for (i = 0; i < nnis; i++) {
 986                 pbuf->pb_info.pi_ni[i].ns_nid = LNET_NID_ANY;
 987                 pbuf->pb_info.pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 988         }
 989         rcd->rcd_pingbuffer = pbuf;
 990
 991         LASSERT(!LNetEQHandleIsInvalid(the_lnet.ln_rc_eqh));
 992         rc = LNetMDBind((struct lnet_md){.start     = &pbuf->pb_info,
 993                                     .user_ptr  = rcd,
 994                                     .length    = LNET_PING_INFO_SIZE(nnis),
 995                                     .threshold = LNET_MD_THRESH_INF,
 996                                     .options   = LNET_MD_TRUNCATE,
 997                                     .eq_handle = the_lnet.ln_rc_eqh},
 998                         LNET_UNLINK,
 999                         &rcd->rcd_mdh);
1000         if (rc < 0) {
1001                 CERROR("Can't bind MD: %d\n", rc);
1002                 goto out_ping_buffer_decref;
1003         }
1004         LASSERT(rc == 0);
1005
1006         lnet_net_lock(gateway->lpni_cpt);
1007         /* Check if this is still a router. */
1008         if (!lnet_isrouter(gateway))
1009                 goto out_unlock;
1010         /* Check if someone else installed router data. */
1011         if (gateway->lpni_rcd && gateway->lpni_rcd != rcd)
1012                 goto out_unlock;
1013
1014         /* Install and/or update the router data. */
1015         if (!gateway->lpni_rcd) {
1016                 lnet_peer_ni_addref_locked(gateway);
1017                 rcd->rcd_gateway = gateway;
1018                 gateway->lpni_rcd = rcd;
1019         }
1020         gateway->lpni_ping_notsent = 0;
1021
1022         return rcd;
1023
1024 out_unlock:
1025         lnet_net_unlock(gateway->lpni_cpt);
1026         rc = LNetMDUnlink(mdh);
1027         LASSERT(rc == 0);
1028 out_ping_buffer_decref:
1029         lnet_ping_buffer_decref(pbuf);
1030 out:
1031         if (rcd && rcd != gateway->lpni_rcd)
1032                 lnet_destroy_rc_data(rcd);
1033         lnet_net_lock(gateway->lpni_cpt);
1034         return gateway->lpni_rcd;
1035 }
1036
1037 static int
1038 lnet_router_check_interval(struct lnet_peer_ni *rtr)
1039 {
1040         int secs;
1041
1042         secs = rtr->lpni_alive ? live_router_check_interval :
1043                                dead_router_check_interval;
1044         if (secs < 0)
1045                 secs = 0;
1046
1047         return secs;
1048 }
1049
1050 static void
1051 lnet_ping_router_locked(struct lnet_peer_ni *rtr)
1052 {
1053         struct lnet_rc_data *rcd = NULL;
1054         time64_t now = ktime_get_seconds();
1055         time64_t secs;
1056         struct lnet_ni *ni;
1057
1058         lnet_peer_ni_addref_locked(rtr);
1059
1060         if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */
1061             now >  rtr->lpni_ping_deadline)
1062                 lnet_notify_locked(rtr, 1, 0, now);
1063
1064         /* Run any outstanding notifications */
1065         ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL);
1066         lnet_ni_notify_locked(ni, rtr);
1067
1068         if (!lnet_isrouter(rtr) ||
1069             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1070                 /* router table changed or router checker is shutting down */
1071                 lnet_peer_ni_decref_locked(rtr);
1072                 return;
1073         }
1074
1075         rcd = rtr->lpni_rcd;
1076         if (!rcd || rcd->rcd_nnis > rcd->rcd_pingbuffer->pb_nnis)
1077                 rcd = lnet_update_rc_data_locked(rtr);
1078         if (rcd == NULL)
1079                 return;
1080
1081         secs = lnet_router_check_interval(rtr);
1082
1083         CDEBUG(D_NET,
1084                "rtr %s %lld: deadline %lld ping_notsent %d alive %d "
1085                "alive_count %d lpni_ping_timestamp %lld\n",
1086                libcfs_nid2str(rtr->lpni_nid), secs,
1087                rtr->lpni_ping_deadline, rtr->lpni_ping_notsent,
1088                rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp);
1089
1090         if (secs != 0 && !rtr->lpni_ping_notsent &&
1091             now > rtr->lpni_ping_timestamp + secs) {
1092                 int               rc;
1093                 struct lnet_process_id id;
1094                 struct lnet_handle_md mdh;
1095
1096                 id.nid = rtr->lpni_nid;
1097                 id.pid = LNET_PID_LUSTRE;
1098                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1099
1100                 rtr->lpni_ping_notsent   = 1;
1101                 rtr->lpni_ping_timestamp = now;
1102
1103                 mdh = rcd->rcd_mdh;
1104
1105                 if (rtr->lpni_ping_deadline == 0) {
1106                         rtr->lpni_ping_deadline = ktime_get_seconds() +
1107                                                   router_ping_timeout;
1108                 }
1109
1110                 lnet_net_unlock(rtr->lpni_cpt);
1111
1112                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1113                              LNET_PROTO_PING_MATCHBITS, 0);
1114
1115                 lnet_net_lock(rtr->lpni_cpt);
1116                 if (rc != 0)
1117                         rtr->lpni_ping_notsent = 0; /* no event pending */
1118         }
1119
1120         lnet_peer_ni_decref_locked(rtr);
1121         return;
1122 }
1123
1124 int
1125 lnet_router_checker_start(void)
1126 {
1127         int                     rc;
1128         int                     eqsz = 0;
1129         struct task_struct     *task;
1130
1131         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1132
1133         if (check_routers_before_use &&
1134             dead_router_check_interval <= 0) {
1135                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1136                                    " set if 'check_routers_before_use' is set"
1137                                    "\n");
1138                 return -EINVAL;
1139         }
1140
1141         sema_init(&the_lnet.ln_rc_signal, 0);
1142
1143         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1144         if (rc != 0) {
1145                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1146                 return -ENOMEM;
1147         }
1148
1149         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1150         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1151         if (IS_ERR(task)) {
1152                 rc = PTR_ERR(task);
1153                 CERROR("Can't start router checker thread: %d\n", rc);
1154                 /* block until event callback signals exit */
1155                 down(&the_lnet.ln_rc_signal);
1156                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1157                 LASSERT(rc == 0);
1158                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1159                 return -ENOMEM;
1160         }
1161
1162         if (check_routers_before_use) {
1163                 /* Note that a helpful side-effect of pinging all known routers
1164                  * at startup is that it makes them drop stale connections they
1165                  * may have to a previous instance of me. */
1166                 lnet_wait_known_routerstate();
1167         }
1168
1169         return 0;
1170 }
1171
1172 void
1173 lnet_router_checker_stop (void)
1174 {
1175         int rc;
1176
1177         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1178                 return;
1179
1180         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1181         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1182         /* wakeup the RC thread if it's sleeping */
1183         wake_up(&the_lnet.ln_rc_waitq);
1184
1185         /* block until event callback signals exit */
1186         down(&the_lnet.ln_rc_signal);
1187         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1188
1189         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1190         LASSERT(rc == 0);
1191         return;
1192 }
1193
1194 static void
1195 lnet_prune_rc_data(int wait_unlink)
1196 {
1197         struct lnet_rc_data *rcd;
1198         struct lnet_rc_data *tmp;
1199         struct lnet_peer_ni *lp;
1200         struct list_head head;
1201         int i = 2;
1202
1203         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1204                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1205                    list_empty(&the_lnet.ln_rcd_zombie)))
1206                 return;
1207
1208         INIT_LIST_HEAD(&head);
1209
1210         lnet_net_lock(LNET_LOCK_EX);
1211
1212         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1213                 /* router checker is stopping, prune all */
1214                 list_for_each_entry(lp, &the_lnet.ln_routers,
1215                                     lpni_rtr_list) {
1216                         if (lp->lpni_rcd == NULL)
1217                                 continue;
1218
1219                         LASSERT(list_empty(&lp->lpni_rcd->rcd_list));
1220                         list_add(&lp->lpni_rcd->rcd_list,
1221                                  &the_lnet.ln_rcd_deathrow);
1222                         lp->lpni_rcd = NULL;
1223                 }
1224         }
1225
1226         /* unlink all RCDs on deathrow list */
1227         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1228
1229         if (!list_empty(&head)) {
1230                 lnet_net_unlock(LNET_LOCK_EX);
1231
1232                 list_for_each_entry(rcd, &head, rcd_list)
1233                         LNetMDUnlink(rcd->rcd_mdh);
1234
1235                 lnet_net_lock(LNET_LOCK_EX);
1236         }
1237
1238         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1239
1240         /* release all zombie RCDs */
1241         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1242                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1243                                          rcd_list) {
1244                         if (LNetMDHandleIsInvalid(rcd->rcd_mdh))
1245                                 list_move(&rcd->rcd_list, &head);
1246                 }
1247
1248                 wait_unlink = wait_unlink &&
1249                               !list_empty(&the_lnet.ln_rcd_zombie);
1250
1251                 lnet_net_unlock(LNET_LOCK_EX);
1252
1253                 while (!list_empty(&head)) {
1254                         rcd = list_entry(head.next,
1255                                          struct lnet_rc_data, rcd_list);
1256                         list_del_init(&rcd->rcd_list);
1257                         lnet_destroy_rc_data(rcd);
1258                 }
1259
1260                 if (!wait_unlink)
1261                         return;
1262
1263                 i++;
1264                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1265                        "Waiting for rc buffers to unlink\n");
1266                 set_current_state(TASK_UNINTERRUPTIBLE);
1267                 schedule_timeout(cfs_time_seconds(1) / 4);
1268
1269                 lnet_net_lock(LNET_LOCK_EX);
1270         }
1271
1272         lnet_net_unlock(LNET_LOCK_EX);
1273 }
1274
1275 /*
1276  * This function is called to check if the RC should block indefinitely.
1277  * It's called from lnet_router_checker() as well as being passed to
1278  * wait_event_interruptible() to avoid the lost wake_up problem.
1279  *
1280  * When it's called from wait_event_interruptible() it is necessary to
1281  * also not sleep if the rc state is not running to avoid a deadlock
1282  * when the system is shutting down
1283  */
1284 static inline bool
1285 lnet_router_checker_active(void)
1286 {
1287         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1288                 return true;
1289
1290         /* Router Checker thread needs to run when routing is enabled in
1291          * order to call lnet_update_ni_status_locked() */
1292         if (the_lnet.ln_routing)
1293                 return true;
1294
1295         return !list_empty(&the_lnet.ln_routers) &&
1296                 (live_router_check_interval > 0 ||
1297                  dead_router_check_interval > 0);
1298 }
1299
1300 static int
1301 lnet_router_checker(void *arg)
1302 {
1303         struct lnet_peer_ni *rtr;
1304         struct list_head *entry;
1305
1306         cfs_block_allsigs();
1307
1308         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1309                 __u64   version;
1310                 int     cpt;
1311                 int     cpt2;
1312
1313                 cpt = lnet_net_lock_current();
1314 rescan:
1315                 version = the_lnet.ln_routers_version;
1316
1317                 list_for_each(entry, &the_lnet.ln_routers) {
1318                         rtr = list_entry(entry, struct lnet_peer_ni,
1319                                          lpni_rtr_list);
1320
1321                         cpt2 = rtr->lpni_cpt;
1322                         if (cpt != cpt2) {
1323                                 lnet_net_unlock(cpt);
1324                                 cpt = cpt2;
1325                                 lnet_net_lock(cpt);
1326                                 /* the routers list has changed */
1327                                 if (version != the_lnet.ln_routers_version)
1328                                         goto rescan;
1329                         }
1330
1331                         lnet_ping_router_locked(rtr);
1332
1333                         /* NB dropped lock */
1334                         if (version != the_lnet.ln_routers_version) {
1335                                 /* the routers list has changed */
1336                                 goto rescan;
1337                         }
1338                 }
1339
1340                 if (the_lnet.ln_routing)
1341                         lnet_update_ni_status_locked();
1342
1343                 lnet_net_unlock(cpt);
1344
1345                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1346
1347                 /* Call schedule_timeout() here always adds 1 to load average
1348                  * because kernel counts # active tasks as nr_running
1349                  * + nr_uninterruptible. */
1350                 /* if there are any routes then wakeup every second.  If
1351                  * there are no routes then sleep indefinitely until woken
1352                  * up by a user adding a route */
1353                 if (!lnet_router_checker_active())
1354                         wait_event_interruptible(the_lnet.ln_rc_waitq,
1355                                                  lnet_router_checker_active());
1356                 else
1357                         wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1358                                                          false,
1359                                                          cfs_time_seconds(1));
1360         }
1361
1362         lnet_prune_rc_data(1); /* wait for UNLINK */
1363
1364         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1365         up(&the_lnet.ln_rc_signal);
1366         /* The unlink event callback will signal final completion */
1367         return 0;
1368 }
1369
1370 void
1371 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1372 {
1373         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1374
1375         while (--npages >= 0)
1376                 __free_page(rb->rb_kiov[npages].kiov_page);
1377
1378         LIBCFS_FREE(rb, sz);
1379 }
1380
1381 static struct lnet_rtrbuf *
1382 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1383 {
1384         int            npages = rbp->rbp_npages;
1385         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1386         struct page   *page;
1387         struct lnet_rtrbuf *rb;
1388         int            i;
1389
1390         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1391         if (rb == NULL)
1392                 return NULL;
1393
1394         rb->rb_pool = rbp;
1395
1396         for (i = 0; i < npages; i++) {
1397                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1398                                           GFP_KERNEL | __GFP_ZERO);
1399                 if (page == NULL) {
1400                         while (--i >= 0)
1401                                 __free_page(rb->rb_kiov[i].kiov_page);
1402
1403                         LIBCFS_FREE(rb, sz);
1404                         return NULL;
1405                 }
1406
1407                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1408                 rb->rb_kiov[i].kiov_offset = 0;
1409                 rb->rb_kiov[i].kiov_page = page;
1410         }
1411
1412         return rb;
1413 }
1414
1415 static void
1416 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1417 {
1418         int npages = rbp->rbp_npages;
1419         struct lnet_rtrbuf *rb;
1420         struct list_head tmp;
1421
1422         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1423                 return;
1424
1425         INIT_LIST_HEAD(&tmp);
1426
1427         lnet_net_lock(cpt);
1428         list_splice_init(&rbp->rbp_msgs, &tmp);
1429         lnet_drop_routed_msgs_locked(&tmp, cpt);
1430         list_splice_init(&rbp->rbp_bufs, &tmp);
1431         rbp->rbp_req_nbuffers = 0;
1432         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1433         rbp->rbp_mincredits = 0;
1434         lnet_net_unlock(cpt);
1435
1436         /* Free buffers on the free list. */
1437         while (!list_empty(&tmp)) {
1438                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1439                 list_del(&rb->rb_list);
1440                 lnet_destroy_rtrbuf(rb, npages);
1441         }
1442 }
1443
1444 static int
1445 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1446 {
1447         struct list_head rb_list;
1448         struct lnet_rtrbuf *rb;
1449         int             num_rb;
1450         int             num_buffers = 0;
1451         int             old_req_nbufs;
1452         int             npages = rbp->rbp_npages;
1453
1454         lnet_net_lock(cpt);
1455         /* If we are called for less buffers than already in the pool, we
1456          * just lower the req_nbuffers number and excess buffers will be
1457          * thrown away as they are returned to the free list.  Credits
1458          * then get adjusted as well.
1459          * If we already have enough buffers allocated to serve the
1460          * increase requested, then we can treat that the same way as we
1461          * do the decrease. */
1462         num_rb = nbufs - rbp->rbp_nbuffers;
1463         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1464                 rbp->rbp_req_nbuffers = nbufs;
1465                 lnet_net_unlock(cpt);
1466                 return 0;
1467         }
1468         /* store the older value of rbp_req_nbuffers and then set it to
1469          * the new request to prevent lnet_return_rx_credits_locked() from
1470          * freeing buffers that we need to keep around */
1471         old_req_nbufs = rbp->rbp_req_nbuffers;
1472         rbp->rbp_req_nbuffers = nbufs;
1473         lnet_net_unlock(cpt);
1474
1475         INIT_LIST_HEAD(&rb_list);
1476
1477         /* allocate the buffers on a local list first.  If all buffers are
1478          * allocated successfully then join this list to the rbp buffer
1479          * list.  If not then free all allocated buffers. */
1480         while (num_rb-- > 0) {
1481                 rb = lnet_new_rtrbuf(rbp, cpt);
1482                 if (rb == NULL) {
1483                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1484                                nbufs, npages);
1485
1486                         lnet_net_lock(cpt);
1487                         rbp->rbp_req_nbuffers = old_req_nbufs;
1488                         lnet_net_unlock(cpt);
1489
1490                         goto failed;
1491                 }
1492
1493                 list_add(&rb->rb_list, &rb_list);
1494                 num_buffers++;
1495         }
1496
1497         lnet_net_lock(cpt);
1498
1499         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1500         rbp->rbp_nbuffers += num_buffers;
1501         rbp->rbp_credits += num_buffers;
1502         rbp->rbp_mincredits = rbp->rbp_credits;
1503         /* We need to schedule blocked msg using the newly
1504          * added buffers. */
1505         while (!list_empty(&rbp->rbp_bufs) &&
1506                !list_empty(&rbp->rbp_msgs))
1507                 lnet_schedule_blocked_locked(rbp);
1508
1509         lnet_net_unlock(cpt);
1510
1511         return 0;
1512
1513 failed:
1514         while (!list_empty(&rb_list)) {
1515                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1516                 list_del(&rb->rb_list);
1517                 lnet_destroy_rtrbuf(rb, npages);
1518         }
1519
1520         return -ENOMEM;
1521 }
1522
1523 static void
1524 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1525 {
1526         INIT_LIST_HEAD(&rbp->rbp_msgs);
1527         INIT_LIST_HEAD(&rbp->rbp_bufs);
1528
1529         rbp->rbp_npages = npages;
1530         rbp->rbp_credits = 0;
1531         rbp->rbp_mincredits = 0;
1532 }
1533
1534 void
1535 lnet_rtrpools_free(int keep_pools)
1536 {
1537         struct lnet_rtrbufpool *rtrp;
1538         int               i;
1539
1540         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1541                 return;
1542
1543         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1544                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1545                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1546                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1547         }
1548
1549         if (!keep_pools) {
1550                 cfs_percpt_free(the_lnet.ln_rtrpools);
1551                 the_lnet.ln_rtrpools = NULL;
1552         }
1553 }
1554
1555 static int
1556 lnet_nrb_tiny_calculate(void)
1557 {
1558         int     nrbs = LNET_NRB_TINY;
1559
1560         if (tiny_router_buffers < 0) {
1561                 LCONSOLE_ERROR_MSG(0x10c,
1562                                    "tiny_router_buffers=%d invalid when "
1563                                    "routing enabled\n", tiny_router_buffers);
1564                 return -EINVAL;
1565         }
1566
1567         if (tiny_router_buffers > 0)
1568                 nrbs = tiny_router_buffers;
1569
1570         nrbs /= LNET_CPT_NUMBER;
1571         return max(nrbs, LNET_NRB_TINY_MIN);
1572 }
1573
1574 static int
1575 lnet_nrb_small_calculate(void)
1576 {
1577         int     nrbs = LNET_NRB_SMALL;
1578
1579         if (small_router_buffers < 0) {
1580                 LCONSOLE_ERROR_MSG(0x10c,
1581                                    "small_router_buffers=%d invalid when "
1582                                    "routing enabled\n", small_router_buffers);
1583                 return -EINVAL;
1584         }
1585
1586         if (small_router_buffers > 0)
1587                 nrbs = small_router_buffers;
1588
1589         nrbs /= LNET_CPT_NUMBER;
1590         return max(nrbs, LNET_NRB_SMALL_MIN);
1591 }
1592
1593 static int
1594 lnet_nrb_large_calculate(void)
1595 {
1596         int     nrbs = LNET_NRB_LARGE;
1597
1598         if (large_router_buffers < 0) {
1599                 LCONSOLE_ERROR_MSG(0x10c,
1600                                    "large_router_buffers=%d invalid when "
1601                                    "routing enabled\n", large_router_buffers);
1602                 return -EINVAL;
1603         }
1604
1605         if (large_router_buffers > 0)
1606                 nrbs = large_router_buffers;
1607
1608         nrbs /= LNET_CPT_NUMBER;
1609         return max(nrbs, LNET_NRB_LARGE_MIN);
1610 }
1611
1612 int
1613 lnet_rtrpools_alloc(int im_a_router)
1614 {
1615         struct lnet_rtrbufpool *rtrp;
1616         int     nrb_tiny;
1617         int     nrb_small;
1618         int     nrb_large;
1619         int     rc;
1620         int     i;
1621
1622         if (!strcmp(forwarding, "")) {
1623                 /* not set either way */
1624                 if (!im_a_router)
1625                         return 0;
1626         } else if (!strcmp(forwarding, "disabled")) {
1627                 /* explicitly disabled */
1628                 return 0;
1629         } else if (!strcmp(forwarding, "enabled")) {
1630                 /* explicitly enabled */
1631         } else {
1632                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1633                                    "'enabled' or 'disabled'\n");
1634                 return -EINVAL;
1635         }
1636
1637         nrb_tiny = lnet_nrb_tiny_calculate();
1638         if (nrb_tiny < 0)
1639                 return -EINVAL;
1640
1641         nrb_small = lnet_nrb_small_calculate();
1642         if (nrb_small < 0)
1643                 return -EINVAL;
1644
1645         nrb_large = lnet_nrb_large_calculate();
1646         if (nrb_large < 0)
1647                 return -EINVAL;
1648
1649         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1650                                                 LNET_NRBPOOLS *
1651                                                 sizeof(struct lnet_rtrbufpool));
1652         if (the_lnet.ln_rtrpools == NULL) {
1653                 LCONSOLE_ERROR_MSG(0x10c,
1654                                    "Failed to initialize router buffe pool\n");
1655                 return -ENOMEM;
1656         }
1657
1658         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1659                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1660                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1661                                               nrb_tiny, i);
1662                 if (rc != 0)
1663                         goto failed;
1664
1665                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1666                                   LNET_NRB_SMALL_PAGES);
1667                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1668                                               nrb_small, i);
1669                 if (rc != 0)
1670                         goto failed;
1671
1672                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1673                                   LNET_NRB_LARGE_PAGES);
1674                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1675                                               nrb_large, i);
1676                 if (rc != 0)
1677                         goto failed;
1678         }
1679
1680         lnet_net_lock(LNET_LOCK_EX);
1681         the_lnet.ln_routing = 1;
1682         lnet_net_unlock(LNET_LOCK_EX);
1683         return 0;
1684
1685  failed:
1686         lnet_rtrpools_free(0);
1687         return rc;
1688 }
1689
1690 static int
1691 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1692 {
1693         int nrb = 0;
1694         int rc = 0;
1695         int i;
1696         struct lnet_rtrbufpool *rtrp;
1697
1698         /* If the provided values for each buffer pool are different than the
1699          * configured values, we need to take action. */
1700         if (tiny >= 0) {
1701                 tiny_router_buffers = tiny;
1702                 nrb = lnet_nrb_tiny_calculate();
1703                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1704                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1705                                                       nrb, i);
1706                         if (rc != 0)
1707                                 return rc;
1708                 }
1709         }
1710         if (small >= 0) {
1711                 small_router_buffers = small;
1712                 nrb = lnet_nrb_small_calculate();
1713                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1714                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1715                                                       nrb, i);
1716                         if (rc != 0)
1717                                 return rc;
1718                 }
1719         }
1720         if (large >= 0) {
1721                 large_router_buffers = large;
1722                 nrb = lnet_nrb_large_calculate();
1723                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1724                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1725                                                       nrb, i);
1726                         if (rc != 0)
1727                                 return rc;
1728                 }
1729         }
1730
1731         return 0;
1732 }
1733
1734 int
1735 lnet_rtrpools_adjust(int tiny, int small, int large)
1736 {
1737         /* this function doesn't revert the changes if adding new buffers
1738          * failed.  It's up to the user space caller to revert the
1739          * changes. */
1740
1741         if (!the_lnet.ln_routing)
1742                 return 0;
1743
1744         return lnet_rtrpools_adjust_helper(tiny, small, large);
1745 }
1746
1747 int
1748 lnet_rtrpools_enable(void)
1749 {
1750         int rc = 0;
1751
1752         if (the_lnet.ln_routing)
1753                 return 0;
1754
1755         if (the_lnet.ln_rtrpools == NULL)
1756                 /* If routing is turned off, and we have never
1757                  * initialized the pools before, just call the
1758                  * standard buffer pool allocation routine as
1759                  * if we are just configuring this for the first
1760                  * time. */
1761                 rc = lnet_rtrpools_alloc(1);
1762         else
1763                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1764         if (rc != 0)
1765                 return rc;
1766
1767         lnet_net_lock(LNET_LOCK_EX);
1768         the_lnet.ln_routing = 1;
1769
1770         the_lnet.ln_ping_target->pb_info.pi_features &=
1771                 ~LNET_PING_FEAT_RTE_DISABLED;
1772         lnet_net_unlock(LNET_LOCK_EX);
1773
1774         return rc;
1775 }
1776
1777 void
1778 lnet_rtrpools_disable(void)
1779 {
1780         if (!the_lnet.ln_routing)
1781                 return;
1782
1783         lnet_net_lock(LNET_LOCK_EX);
1784         the_lnet.ln_routing = 0;
1785         the_lnet.ln_ping_target->pb_info.pi_features |=
1786                 LNET_PING_FEAT_RTE_DISABLED;
1787
1788         tiny_router_buffers = 0;
1789         small_router_buffers = 0;
1790         large_router_buffers = 0;
1791         lnet_net_unlock(LNET_LOCK_EX);
1792         lnet_rtrpools_free(1);
1793 }
1794
1795 int
1796 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, time64_t when)
1797 {
1798         struct lnet_peer_ni *lp = NULL;
1799         time64_t now = ktime_get_seconds();
1800         int cpt = lnet_cpt_of_nid(nid, ni);
1801
1802         LASSERT (!in_interrupt ());
1803
1804         CDEBUG (D_NET, "%s notifying %s: %s\n",
1805                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1806                 libcfs_nid2str(nid),
1807                 alive ? "up" : "down");
1808
1809         if (ni != NULL &&
1810             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1811                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1812                       libcfs_nid2str(nid), alive ? "birth" : "death",
1813                       libcfs_nid2str(ni->ni_nid));
1814                 return -EINVAL;
1815         }
1816
1817         /* can't do predictions... */
1818         if (when > now) {
1819                 CWARN("Ignoring prediction from %s of %s %s "
1820                       "%lld seconds in the future\n",
1821                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1822                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1823                 return -EINVAL;
1824         }
1825
1826         if (ni != NULL && !alive &&             /* LND telling me she's down */
1827             !auto_down) {                       /* auto-down disabled */
1828                 CDEBUG(D_NET, "Auto-down disabled\n");
1829                 return 0;
1830         }
1831
1832         lnet_net_lock(cpt);
1833
1834         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1835                 lnet_net_unlock(cpt);
1836                 return -ESHUTDOWN;
1837         }
1838
1839         lp = lnet_find_peer_ni_locked(nid);
1840         if (lp == NULL) {
1841                 /* nid not found */
1842                 lnet_net_unlock(cpt);
1843                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1844                 return 0;
1845         }
1846
1847         /*
1848          * It is possible for this function to be called for the same peer
1849          * but with different NIs. We want to synchronize the notification
1850          * between the different calls. So we will use the lpni_cpt to
1851          * grab the net lock.
1852          */
1853         if (lp->lpni_cpt != cpt) {
1854                 lnet_net_unlock(cpt);
1855                 cpt = lp->lpni_cpt;
1856                 lnet_net_lock(cpt);
1857         }
1858
1859         /* We can't fully trust LND on reporting exact peer last_alive
1860          * if he notifies us about dead peer. For example ksocklnd can
1861          * call us with when == _time_when_the_node_was_booted_ if
1862          * no connections were successfully established */
1863         if (ni != NULL && !alive && when < lp->lpni_last_alive)
1864                 when = lp->lpni_last_alive;
1865
1866         lnet_notify_locked(lp, ni == NULL, alive, when);
1867
1868         if (ni != NULL)
1869                 lnet_ni_notify_locked(ni, lp);
1870
1871         lnet_peer_ni_decref_locked(lp);
1872
1873         lnet_net_unlock(cpt);
1874         return 0;
1875 }
1876 EXPORT_SYMBOL(lnet_notify);