lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2016, 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                    cfs_time_t when)
 103 {
 104         if (cfs_time_before(when, lp->lpni_timestamp)) { /* 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_info *info)
 654 {
 655         int               i;
 656         struct lnet_ni_status *stat;
 657
 658         __swab32s(&info->pi_magic);
 659         __swab32s(&info->pi_features);
 660         __swab32s(&info->pi_pid);
 661         __swab32s(&info->pi_nnis);
 662         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 663                 stat = &info->pi_ni[i];
 664                 __swab64s(&stat->ns_nid);
 665                 __swab32s(&stat->ns_status);
 666         }
 667         return;
 668 }
 669
 670 /**
 671  * parse router-checker pinginfo, record number of down NIs for remote
 672  * networks on that router.
 673  */
 674 static void
 675 lnet_parse_rc_info(struct lnet_rc_data *rcd)
 676 {
 677         struct lnet_ping_info   *info = rcd->rcd_pinginfo;
 678         struct lnet_peer_ni     *gw   = rcd->rcd_gateway;
 679         struct lnet_route               *rte;
 680
 681         if (!gw->lpni_alive)
 682                 return;
 683
 684         /*
 685          * Protect gw->lpni_ping_feats. This can be set from
 686          * lnet_notify_locked with different locks being held
 687          */
 688         spin_lock(&gw->lpni_lock);
 689
 690         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 691                 lnet_swap_pinginfo(info);
 692
 693         /* NB always racing with network! */
 694         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 695                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 696                        libcfs_nid2str(gw->lpni_nid), info->pi_magic);
 697                 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 698                 spin_unlock(&gw->lpni_lock);
 699                 return;
 700         }
 701
 702         gw->lpni_ping_feats = info->pi_features;
 703         if ((gw->lpni_ping_feats & LNET_PING_FEAT_MASK) == 0) {
 704                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
 705                        libcfs_nid2str(gw->lpni_nid), gw->lpni_ping_feats);
 706                 spin_unlock(&gw->lpni_lock);
 707                 return; /* nothing I can understand */
 708         }
 709
 710         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0) {
 711                 spin_unlock(&gw->lpni_lock);
 712                 return; /* can't carry NI status info */
 713         }
 714
 715         list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 716                 int     down = 0;
 717                 int     up = 0;
 718                 int     i;
 719
 720                 if ((gw->lpni_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
 721                         rte->lr_downis = 1;
 722                         continue;
 723                 }
 724
 725                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 726                         struct lnet_ni_status *stat = &info->pi_ni[i];
 727                         lnet_nid_t       nid = stat->ns_nid;
 728
 729                         if (nid == LNET_NID_ANY) {
 730                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 731                                        libcfs_nid2str(gw->lpni_nid));
 732                                 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 733                                 spin_unlock(&gw->lpni_lock);
 734                                 return;
 735                         }
 736
 737                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 738                                 continue;
 739
 740                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 741                                 down++;
 742                                 continue;
 743                         }
 744
 745                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 746                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 747                                         up = 1;
 748                                         break;
 749                                 }
 750                                 continue;
 751                         }
 752
 753                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 754                                libcfs_nid2str(gw->lpni_nid), stat->ns_status);
 755                         gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 756                         spin_unlock(&gw->lpni_lock);
 757                         return;
 758                 }
 759
 760                 if (up) { /* ignore downed NIs if NI for dest network is up */
 761                         rte->lr_downis = 0;
 762                         continue;
 763                 }
 764                 /* if @down is zero and this route is single-hop, it means
 765                  * we can't find NI for target network */
 766                 if (down == 0 && rte->lr_hops == 1)
 767                         down = 1;
 768
 769                 rte->lr_downis = down;
 770         }
 771
 772         spin_unlock(&gw->lpni_lock);
 773 }
 774
 775 static void
 776 lnet_router_checker_event(struct lnet_event *event)
 777 {
 778         struct lnet_rc_data *rcd = event->md.user_ptr;
 779         struct lnet_peer_ni *lp;
 780
 781         LASSERT(rcd != NULL);
 782
 783         if (event->unlinked) {
 784                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 785                 return;
 786         }
 787
 788         LASSERT(event->type == LNET_EVENT_SEND ||
 789                 event->type == LNET_EVENT_REPLY);
 790
 791         lp = rcd->rcd_gateway;
 792         LASSERT(lp != NULL);
 793
 794          /* NB: it's called with holding lnet_res_lock, we have a few
 795           * places need to hold both locks at the same time, please take
 796           * care of lock ordering */
 797         lnet_net_lock(lp->lpni_cpt);
 798         if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
 799                 /* ignore if no longer a router or rcd is replaced */
 800                 goto out;
 801         }
 802
 803         if (event->type == LNET_EVENT_SEND) {
 804                 lp->lpni_ping_notsent = 0;
 805                 if (event->status == 0)
 806                         goto out;
 807         }
 808
 809         /* LNET_EVENT_REPLY */
 810         /* A successful REPLY means the router is up.  If _any_ comms
 811          * to the router fail I assume it's down (this will happen if
 812          * we ping alive routers to try to detect router death before
 813          * apps get burned). */
 814
 815         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
 816         /* The router checker will wake up very shortly and do the
 817          * actual notification.
 818          * XXX If 'lp' stops being a router before then, it will still
 819          * have the notification pending!!! */
 820
 821         if (avoid_asym_router_failure && event->status == 0)
 822                 lnet_parse_rc_info(rcd);
 823
 824  out:
 825         lnet_net_unlock(lp->lpni_cpt);
 826 }
 827
 828 static void
 829 lnet_wait_known_routerstate(void)
 830 {
 831         struct lnet_peer_ni *rtr;
 832         struct list_head *entry;
 833         int all_known;
 834
 835         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 836
 837         for (;;) {
 838                 int cpt = lnet_net_lock_current();
 839
 840                 all_known = 1;
 841                 list_for_each(entry, &the_lnet.ln_routers) {
 842                         rtr = list_entry(entry, struct lnet_peer_ni,
 843                                          lpni_rtr_list);
 844
 845                         spin_lock(&rtr->lpni_lock);
 846
 847                         if (rtr->lpni_alive_count == 0) {
 848                                 all_known = 0;
 849                                 spin_unlock(&rtr->lpni_lock);
 850                                 break;
 851                         }
 852                         spin_unlock(&rtr->lpni_lock);
 853                 }
 854
 855                 lnet_net_unlock(cpt);
 856
 857                 if (all_known)
 858                         return;
 859
 860                 set_current_state(TASK_UNINTERRUPTIBLE);
 861                 schedule_timeout(cfs_time_seconds(1));
 862         }
 863 }
 864
 865 void
 866 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
 867 {
 868         struct lnet_route *rte;
 869
 870         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
 871                 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 872                         if (rte->lr_net == net) {
 873                                 rte->lr_downis = 0;
 874                                 break;
 875                         }
 876                 }
 877         }
 878 }
 879
 880 static void
 881 lnet_update_ni_status_locked(void)
 882 {
 883         struct lnet_ni *ni = NULL;
 884         time64_t now;
 885         int timeout;
 886
 887         LASSERT(the_lnet.ln_routing);
 888
 889         timeout = router_ping_timeout +
 890                   MAX(live_router_check_interval, dead_router_check_interval);
 891
 892         now = ktime_get_real_seconds();
 893         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 894                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
 895                         continue;
 896
 897                 if (now < ni->ni_last_alive + timeout)
 898                         continue;
 899
 900                 lnet_ni_lock(ni);
 901                 /* re-check with lock */
 902                 if (now < ni->ni_last_alive + timeout) {
 903                         lnet_ni_unlock(ni);
 904                         continue;
 905                 }
 906
 907                 LASSERT(ni->ni_status != NULL);
 908
 909                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 910                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 911                                libcfs_nid2str(ni->ni_nid), timeout);
 912                         /* NB: so far, this is the only place to set
 913                          * NI status to "down" */
 914                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 915                 }
 916                 lnet_ni_unlock(ni);
 917         }
 918 }
 919
 920 static void
 921 lnet_destroy_rc_data(struct lnet_rc_data *rcd)
 922 {
 923         LASSERT(list_empty(&rcd->rcd_list));
 924         /* detached from network */
 925         LASSERT(LNetMDHandleIsInvalid(rcd->rcd_mdh));
 926
 927         if (rcd->rcd_gateway != NULL) {
 928                 int cpt = rcd->rcd_gateway->lpni_cpt;
 929
 930                 lnet_net_lock(cpt);
 931                 lnet_peer_ni_decref_locked(rcd->rcd_gateway);
 932                 lnet_net_unlock(cpt);
 933         }
 934
 935         if (rcd->rcd_pinginfo != NULL)
 936                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 937
 938         LIBCFS_FREE(rcd, sizeof(*rcd));
 939 }
 940
 941 static struct lnet_rc_data *
 942 lnet_create_rc_data_locked(struct lnet_peer_ni *gateway)
 943 {
 944         struct lnet_rc_data             *rcd = NULL;
 945         struct lnet_ping_info   *pi;
 946         int                     rc;
 947         int                     i;
 948
 949         lnet_net_unlock(gateway->lpni_cpt);
 950
 951         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 952         if (rcd == NULL)
 953                 goto out;
 954
 955         LNetInvalidateMDHandle(&rcd->rcd_mdh);
 956         INIT_LIST_HEAD(&rcd->rcd_list);
 957
 958         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 959         if (pi == NULL)
 960                 goto out;
 961
 962         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 963                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 964                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 965         }
 966         rcd->rcd_pinginfo = pi;
 967
 968         LASSERT(!LNetEQHandleIsInvalid(the_lnet.ln_rc_eqh));
 969         rc = LNetMDBind((struct lnet_md){.start     = pi,
 970                                     .user_ptr  = rcd,
 971                                     .length    = LNET_PINGINFO_SIZE,
 972                                     .threshold = LNET_MD_THRESH_INF,
 973                                     .options   = LNET_MD_TRUNCATE,
 974                                     .eq_handle = the_lnet.ln_rc_eqh},
 975                         LNET_UNLINK,
 976                         &rcd->rcd_mdh);
 977         if (rc < 0) {
 978                 CERROR("Can't bind MD: %d\n", rc);
 979                 goto out;
 980         }
 981         LASSERT(rc == 0);
 982
 983         lnet_net_lock(gateway->lpni_cpt);
 984         /* router table changed or someone has created rcd for this gateway */
 985         if (!lnet_isrouter(gateway) || gateway->lpni_rcd != NULL) {
 986                 lnet_net_unlock(gateway->lpni_cpt);
 987                 goto out;
 988         }
 989
 990         lnet_peer_ni_addref_locked(gateway);
 991         rcd->rcd_gateway = gateway;
 992         gateway->lpni_rcd = rcd;
 993         gateway->lpni_ping_notsent = 0;
 994
 995         return rcd;
 996
 997 out:
 998         if (rcd != NULL) {
 999                 if (!LNetMDHandleIsInvalid(rcd->rcd_mdh)) {
1000                         rc = LNetMDUnlink(rcd->rcd_mdh);
1001                         LASSERT(rc == 0);
1002                 }
1003                 lnet_destroy_rc_data(rcd);
1004         }
1005
1006         lnet_net_lock(gateway->lpni_cpt);
1007         return gateway->lpni_rcd;
1008 }
1009
1010 static int
1011 lnet_router_check_interval(struct lnet_peer_ni *rtr)
1012 {
1013         int secs;
1014
1015         secs = rtr->lpni_alive ? live_router_check_interval :
1016                                dead_router_check_interval;
1017         if (secs < 0)
1018                 secs = 0;
1019
1020         return secs;
1021 }
1022
1023 static void
1024 lnet_ping_router_locked(struct lnet_peer_ni *rtr)
1025 {
1026         struct lnet_rc_data *rcd = NULL;
1027         cfs_time_t      now = cfs_time_current();
1028         int             secs;
1029         struct lnet_ni  *ni;
1030
1031         lnet_peer_ni_addref_locked(rtr);
1032
1033         if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */
1034             cfs_time_after(now, rtr->lpni_ping_deadline))
1035                 lnet_notify_locked(rtr, 1, 0, now);
1036
1037         /* Run any outstanding notifications */
1038         ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL);
1039         lnet_ni_notify_locked(ni, rtr);
1040
1041         if (!lnet_isrouter(rtr) ||
1042             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1043                 /* router table changed or router checker is shutting down */
1044                 lnet_peer_ni_decref_locked(rtr);
1045                 return;
1046         }
1047
1048         rcd = rtr->lpni_rcd != NULL ?
1049               rtr->lpni_rcd : lnet_create_rc_data_locked(rtr);
1050
1051         if (rcd == NULL)
1052                 return;
1053
1054         secs = lnet_router_check_interval(rtr);
1055
1056         CDEBUG(D_NET,
1057                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1058                "alive_count %d lpni_ping_timestamp %lu\n",
1059                libcfs_nid2str(rtr->lpni_nid), secs,
1060                rtr->lpni_ping_deadline, rtr->lpni_ping_notsent,
1061                rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp);
1062
1063         if (secs != 0 && !rtr->lpni_ping_notsent &&
1064             cfs_time_after(now, cfs_time_add(rtr->lpni_ping_timestamp,
1065                                              cfs_time_seconds(secs)))) {
1066                 int               rc;
1067                 struct lnet_process_id id;
1068                 struct lnet_handle_md mdh;
1069
1070                 id.nid = rtr->lpni_nid;
1071                 id.pid = LNET_PID_LUSTRE;
1072                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1073
1074                 rtr->lpni_ping_notsent   = 1;
1075                 rtr->lpni_ping_timestamp = now;
1076
1077                 mdh = rcd->rcd_mdh;
1078
1079                 if (rtr->lpni_ping_deadline == 0) {
1080                         rtr->lpni_ping_deadline =
1081                                 cfs_time_shift(router_ping_timeout);
1082                 }
1083
1084                 lnet_net_unlock(rtr->lpni_cpt);
1085
1086                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1087                              LNET_PROTO_PING_MATCHBITS, 0);
1088
1089                 lnet_net_lock(rtr->lpni_cpt);
1090                 if (rc != 0)
1091                         rtr->lpni_ping_notsent = 0; /* no event pending */
1092         }
1093
1094         lnet_peer_ni_decref_locked(rtr);
1095         return;
1096 }
1097
1098 int
1099 lnet_router_checker_start(void)
1100 {
1101         int                     rc;
1102         int                     eqsz = 0;
1103         struct task_struct     *task;
1104
1105         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1106
1107         if (check_routers_before_use &&
1108             dead_router_check_interval <= 0) {
1109                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1110                                    " set if 'check_routers_before_use' is set"
1111                                    "\n");
1112                 return -EINVAL;
1113         }
1114
1115         sema_init(&the_lnet.ln_rc_signal, 0);
1116
1117         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1118         if (rc != 0) {
1119                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1120                 return -ENOMEM;
1121         }
1122
1123         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1124         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1125         if (IS_ERR(task)) {
1126                 rc = PTR_ERR(task);
1127                 CERROR("Can't start router checker thread: %d\n", rc);
1128                 /* block until event callback signals exit */
1129                 down(&the_lnet.ln_rc_signal);
1130                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1131                 LASSERT(rc == 0);
1132                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1133                 return -ENOMEM;
1134         }
1135
1136         if (check_routers_before_use) {
1137                 /* Note that a helpful side-effect of pinging all known routers
1138                  * at startup is that it makes them drop stale connections they
1139                  * may have to a previous instance of me. */
1140                 lnet_wait_known_routerstate();
1141         }
1142
1143         return 0;
1144 }
1145
1146 void
1147 lnet_router_checker_stop (void)
1148 {
1149         int rc;
1150
1151         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1152                 return;
1153
1154         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1155         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1156         /* wakeup the RC thread if it's sleeping */
1157         wake_up(&the_lnet.ln_rc_waitq);
1158
1159         /* block until event callback signals exit */
1160         down(&the_lnet.ln_rc_signal);
1161         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1162
1163         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1164         LASSERT(rc == 0);
1165         return;
1166 }
1167
1168 static void
1169 lnet_prune_rc_data(int wait_unlink)
1170 {
1171         struct lnet_rc_data *rcd;
1172         struct lnet_rc_data *tmp;
1173         struct lnet_peer_ni *lp;
1174         struct list_head head;
1175         int i = 2;
1176
1177         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1178                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1179                    list_empty(&the_lnet.ln_rcd_zombie)))
1180                 return;
1181
1182         INIT_LIST_HEAD(&head);
1183
1184         lnet_net_lock(LNET_LOCK_EX);
1185
1186         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1187                 /* router checker is stopping, prune all */
1188                 list_for_each_entry(lp, &the_lnet.ln_routers,
1189                                     lpni_rtr_list) {
1190                         if (lp->lpni_rcd == NULL)
1191                                 continue;
1192
1193                         LASSERT(list_empty(&lp->lpni_rcd->rcd_list));
1194                         list_add(&lp->lpni_rcd->rcd_list,
1195                                  &the_lnet.ln_rcd_deathrow);
1196                         lp->lpni_rcd = NULL;
1197                 }
1198         }
1199
1200         /* unlink all RCDs on deathrow list */
1201         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1202
1203         if (!list_empty(&head)) {
1204                 lnet_net_unlock(LNET_LOCK_EX);
1205
1206                 list_for_each_entry(rcd, &head, rcd_list)
1207                         LNetMDUnlink(rcd->rcd_mdh);
1208
1209                 lnet_net_lock(LNET_LOCK_EX);
1210         }
1211
1212         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1213
1214         /* release all zombie RCDs */
1215         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1216                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1217                                          rcd_list) {
1218                         if (LNetMDHandleIsInvalid(rcd->rcd_mdh))
1219                                 list_move(&rcd->rcd_list, &head);
1220                 }
1221
1222                 wait_unlink = wait_unlink &&
1223                               !list_empty(&the_lnet.ln_rcd_zombie);
1224
1225                 lnet_net_unlock(LNET_LOCK_EX);
1226
1227                 while (!list_empty(&head)) {
1228                         rcd = list_entry(head.next,
1229                                          struct lnet_rc_data, rcd_list);
1230                         list_del_init(&rcd->rcd_list);
1231                         lnet_destroy_rc_data(rcd);
1232                 }
1233
1234                 if (!wait_unlink)
1235                         return;
1236
1237                 i++;
1238                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1239                        "Waiting for rc buffers to unlink\n");
1240                 set_current_state(TASK_UNINTERRUPTIBLE);
1241                 schedule_timeout(cfs_time_seconds(1) / 4);
1242
1243                 lnet_net_lock(LNET_LOCK_EX);
1244         }
1245
1246         lnet_net_unlock(LNET_LOCK_EX);
1247 }
1248
1249 /*
1250  * This function is called to check if the RC should block indefinitely.
1251  * It's called from lnet_router_checker() as well as being passed to
1252  * wait_event_interruptible() to avoid the lost wake_up problem.
1253  *
1254  * When it's called from wait_event_interruptible() it is necessary to
1255  * also not sleep if the rc state is not running to avoid a deadlock
1256  * when the system is shutting down
1257  */
1258 static inline bool
1259 lnet_router_checker_active(void)
1260 {
1261         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1262                 return true;
1263
1264         /* Router Checker thread needs to run when routing is enabled in
1265          * order to call lnet_update_ni_status_locked() */
1266         if (the_lnet.ln_routing)
1267                 return true;
1268
1269         return !list_empty(&the_lnet.ln_routers) &&
1270                 (live_router_check_interval > 0 ||
1271                  dead_router_check_interval > 0);
1272 }
1273
1274 static int
1275 lnet_router_checker(void *arg)
1276 {
1277         struct lnet_peer_ni *rtr;
1278         struct list_head *entry;
1279
1280         cfs_block_allsigs();
1281
1282         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1283                 __u64   version;
1284                 int     cpt;
1285                 int     cpt2;
1286
1287                 cpt = lnet_net_lock_current();
1288 rescan:
1289                 version = the_lnet.ln_routers_version;
1290
1291                 list_for_each(entry, &the_lnet.ln_routers) {
1292                         rtr = list_entry(entry, struct lnet_peer_ni,
1293                                          lpni_rtr_list);
1294
1295                         cpt2 = rtr->lpni_cpt;
1296                         if (cpt != cpt2) {
1297                                 lnet_net_unlock(cpt);
1298                                 cpt = cpt2;
1299                                 lnet_net_lock(cpt);
1300                                 /* the routers list has changed */
1301                                 if (version != the_lnet.ln_routers_version)
1302                                         goto rescan;
1303                         }
1304
1305                         lnet_ping_router_locked(rtr);
1306
1307                         /* NB dropped lock */
1308                         if (version != the_lnet.ln_routers_version) {
1309                                 /* the routers list has changed */
1310                                 goto rescan;
1311                         }
1312                 }
1313
1314                 if (the_lnet.ln_routing)
1315                         lnet_update_ni_status_locked();
1316
1317                 lnet_net_unlock(cpt);
1318
1319                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1320
1321                 /* Call schedule_timeout() here always adds 1 to load average
1322                  * because kernel counts # active tasks as nr_running
1323                  * + nr_uninterruptible. */
1324                 /* if there are any routes then wakeup every second.  If
1325                  * there are no routes then sleep indefinitely until woken
1326                  * up by a user adding a route */
1327                 if (!lnet_router_checker_active())
1328                         wait_event_interruptible(the_lnet.ln_rc_waitq,
1329                                                  lnet_router_checker_active());
1330                 else
1331                         wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1332                                                          false,
1333                                                          cfs_time_seconds(1));
1334         }
1335
1336         lnet_prune_rc_data(1); /* wait for UNLINK */
1337
1338         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1339         up(&the_lnet.ln_rc_signal);
1340         /* The unlink event callback will signal final completion */
1341         return 0;
1342 }
1343
1344 void
1345 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1346 {
1347         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1348
1349         while (--npages >= 0)
1350                 __free_page(rb->rb_kiov[npages].kiov_page);
1351
1352         LIBCFS_FREE(rb, sz);
1353 }
1354
1355 static struct lnet_rtrbuf *
1356 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1357 {
1358         int            npages = rbp->rbp_npages;
1359         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1360         struct page   *page;
1361         struct lnet_rtrbuf *rb;
1362         int            i;
1363
1364         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1365         if (rb == NULL)
1366                 return NULL;
1367
1368         rb->rb_pool = rbp;
1369
1370         for (i = 0; i < npages; i++) {
1371                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1372                                           GFP_KERNEL | __GFP_ZERO);
1373                 if (page == NULL) {
1374                         while (--i >= 0)
1375                                 __free_page(rb->rb_kiov[i].kiov_page);
1376
1377                         LIBCFS_FREE(rb, sz);
1378                         return NULL;
1379                 }
1380
1381                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1382                 rb->rb_kiov[i].kiov_offset = 0;
1383                 rb->rb_kiov[i].kiov_page = page;
1384         }
1385
1386         return rb;
1387 }
1388
1389 static void
1390 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1391 {
1392         int npages = rbp->rbp_npages;
1393         struct lnet_rtrbuf *rb;
1394         struct list_head tmp;
1395
1396         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1397                 return;
1398
1399         INIT_LIST_HEAD(&tmp);
1400
1401         lnet_net_lock(cpt);
1402         list_splice_init(&rbp->rbp_msgs, &tmp);
1403         lnet_drop_routed_msgs_locked(&tmp, cpt);
1404         list_splice_init(&rbp->rbp_bufs, &tmp);
1405         rbp->rbp_req_nbuffers = 0;
1406         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1407         rbp->rbp_mincredits = 0;
1408         lnet_net_unlock(cpt);
1409
1410         /* Free buffers on the free list. */
1411         while (!list_empty(&tmp)) {
1412                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1413                 list_del(&rb->rb_list);
1414                 lnet_destroy_rtrbuf(rb, npages);
1415         }
1416 }
1417
1418 static int
1419 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1420 {
1421         struct list_head rb_list;
1422         struct lnet_rtrbuf *rb;
1423         int             num_rb;
1424         int             num_buffers = 0;
1425         int             old_req_nbufs;
1426         int             npages = rbp->rbp_npages;
1427
1428         lnet_net_lock(cpt);
1429         /* If we are called for less buffers than already in the pool, we
1430          * just lower the req_nbuffers number and excess buffers will be
1431          * thrown away as they are returned to the free list.  Credits
1432          * then get adjusted as well.
1433          * If we already have enough buffers allocated to serve the
1434          * increase requested, then we can treat that the same way as we
1435          * do the decrease. */
1436         num_rb = nbufs - rbp->rbp_nbuffers;
1437         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1438                 rbp->rbp_req_nbuffers = nbufs;
1439                 lnet_net_unlock(cpt);
1440                 return 0;
1441         }
1442         /* store the older value of rbp_req_nbuffers and then set it to
1443          * the new request to prevent lnet_return_rx_credits_locked() from
1444          * freeing buffers that we need to keep around */
1445         old_req_nbufs = rbp->rbp_req_nbuffers;
1446         rbp->rbp_req_nbuffers = nbufs;
1447         lnet_net_unlock(cpt);
1448
1449         INIT_LIST_HEAD(&rb_list);
1450
1451         /* allocate the buffers on a local list first.  If all buffers are
1452          * allocated successfully then join this list to the rbp buffer
1453          * list.  If not then free all allocated buffers. */
1454         while (num_rb-- > 0) {
1455                 rb = lnet_new_rtrbuf(rbp, cpt);
1456                 if (rb == NULL) {
1457                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1458                                nbufs, npages);
1459
1460                         lnet_net_lock(cpt);
1461                         rbp->rbp_req_nbuffers = old_req_nbufs;
1462                         lnet_net_unlock(cpt);
1463
1464                         goto failed;
1465                 }
1466
1467                 list_add(&rb->rb_list, &rb_list);
1468                 num_buffers++;
1469         }
1470
1471         lnet_net_lock(cpt);
1472
1473         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1474         rbp->rbp_nbuffers += num_buffers;
1475         rbp->rbp_credits += num_buffers;
1476         rbp->rbp_mincredits = rbp->rbp_credits;
1477         /* We need to schedule blocked msg using the newly
1478          * added buffers. */
1479         while (!list_empty(&rbp->rbp_bufs) &&
1480                !list_empty(&rbp->rbp_msgs))
1481                 lnet_schedule_blocked_locked(rbp);
1482
1483         lnet_net_unlock(cpt);
1484
1485         return 0;
1486
1487 failed:
1488         while (!list_empty(&rb_list)) {
1489                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1490                 list_del(&rb->rb_list);
1491                 lnet_destroy_rtrbuf(rb, npages);
1492         }
1493
1494         return -ENOMEM;
1495 }
1496
1497 static void
1498 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1499 {
1500         INIT_LIST_HEAD(&rbp->rbp_msgs);
1501         INIT_LIST_HEAD(&rbp->rbp_bufs);
1502
1503         rbp->rbp_npages = npages;
1504         rbp->rbp_credits = 0;
1505         rbp->rbp_mincredits = 0;
1506 }
1507
1508 void
1509 lnet_rtrpools_free(int keep_pools)
1510 {
1511         struct lnet_rtrbufpool *rtrp;
1512         int               i;
1513
1514         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1515                 return;
1516
1517         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1518                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1519                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1520                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1521         }
1522
1523         if (!keep_pools) {
1524                 cfs_percpt_free(the_lnet.ln_rtrpools);
1525                 the_lnet.ln_rtrpools = NULL;
1526         }
1527 }
1528
1529 static int
1530 lnet_nrb_tiny_calculate(void)
1531 {
1532         int     nrbs = LNET_NRB_TINY;
1533
1534         if (tiny_router_buffers < 0) {
1535                 LCONSOLE_ERROR_MSG(0x10c,
1536                                    "tiny_router_buffers=%d invalid when "
1537                                    "routing enabled\n", tiny_router_buffers);
1538                 return -EINVAL;
1539         }
1540
1541         if (tiny_router_buffers > 0)
1542                 nrbs = tiny_router_buffers;
1543
1544         nrbs /= LNET_CPT_NUMBER;
1545         return max(nrbs, LNET_NRB_TINY_MIN);
1546 }
1547
1548 static int
1549 lnet_nrb_small_calculate(void)
1550 {
1551         int     nrbs = LNET_NRB_SMALL;
1552
1553         if (small_router_buffers < 0) {
1554                 LCONSOLE_ERROR_MSG(0x10c,
1555                                    "small_router_buffers=%d invalid when "
1556                                    "routing enabled\n", small_router_buffers);
1557                 return -EINVAL;
1558         }
1559
1560         if (small_router_buffers > 0)
1561                 nrbs = small_router_buffers;
1562
1563         nrbs /= LNET_CPT_NUMBER;
1564         return max(nrbs, LNET_NRB_SMALL_MIN);
1565 }
1566
1567 static int
1568 lnet_nrb_large_calculate(void)
1569 {
1570         int     nrbs = LNET_NRB_LARGE;
1571
1572         if (large_router_buffers < 0) {
1573                 LCONSOLE_ERROR_MSG(0x10c,
1574                                    "large_router_buffers=%d invalid when "
1575                                    "routing enabled\n", large_router_buffers);
1576                 return -EINVAL;
1577         }
1578
1579         if (large_router_buffers > 0)
1580                 nrbs = large_router_buffers;
1581
1582         nrbs /= LNET_CPT_NUMBER;
1583         return max(nrbs, LNET_NRB_LARGE_MIN);
1584 }
1585
1586 int
1587 lnet_rtrpools_alloc(int im_a_router)
1588 {
1589         struct lnet_rtrbufpool *rtrp;
1590         int     nrb_tiny;
1591         int     nrb_small;
1592         int     nrb_large;
1593         int     rc;
1594         int     i;
1595
1596         if (!strcmp(forwarding, "")) {
1597                 /* not set either way */
1598                 if (!im_a_router)
1599                         return 0;
1600         } else if (!strcmp(forwarding, "disabled")) {
1601                 /* explicitly disabled */
1602                 return 0;
1603         } else if (!strcmp(forwarding, "enabled")) {
1604                 /* explicitly enabled */
1605         } else {
1606                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1607                                    "'enabled' or 'disabled'\n");
1608                 return -EINVAL;
1609         }
1610
1611         nrb_tiny = lnet_nrb_tiny_calculate();
1612         if (nrb_tiny < 0)
1613                 return -EINVAL;
1614
1615         nrb_small = lnet_nrb_small_calculate();
1616         if (nrb_small < 0)
1617                 return -EINVAL;
1618
1619         nrb_large = lnet_nrb_large_calculate();
1620         if (nrb_large < 0)
1621                 return -EINVAL;
1622
1623         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1624                                                 LNET_NRBPOOLS *
1625                                                 sizeof(struct lnet_rtrbufpool));
1626         if (the_lnet.ln_rtrpools == NULL) {
1627                 LCONSOLE_ERROR_MSG(0x10c,
1628                                    "Failed to initialize router buffe pool\n");
1629                 return -ENOMEM;
1630         }
1631
1632         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1633                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1634                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1635                                               nrb_tiny, i);
1636                 if (rc != 0)
1637                         goto failed;
1638
1639                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1640                                   LNET_NRB_SMALL_PAGES);
1641                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1642                                               nrb_small, i);
1643                 if (rc != 0)
1644                         goto failed;
1645
1646                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1647                                   LNET_NRB_LARGE_PAGES);
1648                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1649                                               nrb_large, i);
1650                 if (rc != 0)
1651                         goto failed;
1652         }
1653
1654         lnet_net_lock(LNET_LOCK_EX);
1655         the_lnet.ln_routing = 1;
1656         lnet_net_unlock(LNET_LOCK_EX);
1657         return 0;
1658
1659  failed:
1660         lnet_rtrpools_free(0);
1661         return rc;
1662 }
1663
1664 static int
1665 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1666 {
1667         int nrb = 0;
1668         int rc = 0;
1669         int i;
1670         struct lnet_rtrbufpool *rtrp;
1671
1672         /* If the provided values for each buffer pool are different than the
1673          * configured values, we need to take action. */
1674         if (tiny >= 0) {
1675                 tiny_router_buffers = tiny;
1676                 nrb = lnet_nrb_tiny_calculate();
1677                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1678                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1679                                                       nrb, i);
1680                         if (rc != 0)
1681                                 return rc;
1682                 }
1683         }
1684         if (small >= 0) {
1685                 small_router_buffers = small;
1686                 nrb = lnet_nrb_small_calculate();
1687                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1688                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1689                                                       nrb, i);
1690                         if (rc != 0)
1691                                 return rc;
1692                 }
1693         }
1694         if (large >= 0) {
1695                 large_router_buffers = large;
1696                 nrb = lnet_nrb_large_calculate();
1697                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1698                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1699                                                       nrb, i);
1700                         if (rc != 0)
1701                                 return rc;
1702                 }
1703         }
1704
1705         return 0;
1706 }
1707
1708 int
1709 lnet_rtrpools_adjust(int tiny, int small, int large)
1710 {
1711         /* this function doesn't revert the changes if adding new buffers
1712          * failed.  It's up to the user space caller to revert the
1713          * changes. */
1714
1715         if (!the_lnet.ln_routing)
1716                 return 0;
1717
1718         return lnet_rtrpools_adjust_helper(tiny, small, large);
1719 }
1720
1721 int
1722 lnet_rtrpools_enable(void)
1723 {
1724         int rc = 0;
1725
1726         if (the_lnet.ln_routing)
1727                 return 0;
1728
1729         if (the_lnet.ln_rtrpools == NULL)
1730                 /* If routing is turned off, and we have never
1731                  * initialized the pools before, just call the
1732                  * standard buffer pool allocation routine as
1733                  * if we are just configuring this for the first
1734                  * time. */
1735                 rc = lnet_rtrpools_alloc(1);
1736         else
1737                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1738         if (rc != 0)
1739                 return rc;
1740
1741         lnet_net_lock(LNET_LOCK_EX);
1742         the_lnet.ln_routing = 1;
1743
1744         the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1745         lnet_net_unlock(LNET_LOCK_EX);
1746
1747         return rc;
1748 }
1749
1750 void
1751 lnet_rtrpools_disable(void)
1752 {
1753         if (!the_lnet.ln_routing)
1754                 return;
1755
1756         lnet_net_lock(LNET_LOCK_EX);
1757         the_lnet.ln_routing = 0;
1758         the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1759
1760         tiny_router_buffers = 0;
1761         small_router_buffers = 0;
1762         large_router_buffers = 0;
1763         lnet_net_unlock(LNET_LOCK_EX);
1764         lnet_rtrpools_free(1);
1765 }
1766
1767 int
1768 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1769 {
1770         struct lnet_peer_ni *lp = NULL;
1771         cfs_time_t now = cfs_time_current();
1772         int cpt = lnet_cpt_of_nid(nid, ni);
1773
1774         LASSERT (!in_interrupt ());
1775
1776         CDEBUG (D_NET, "%s notifying %s: %s\n",
1777                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1778                 libcfs_nid2str(nid),
1779                 alive ? "up" : "down");
1780
1781         if (ni != NULL &&
1782             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1783                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1784                       libcfs_nid2str(nid), alive ? "birth" : "death",
1785                       libcfs_nid2str(ni->ni_nid));
1786                 return -EINVAL;
1787         }
1788
1789         /* can't do predictions... */
1790         if (cfs_time_after(when, now)) {
1791                 CWARN("Ignoring prediction from %s of %s %s "
1792                       "%ld seconds in the future\n",
1793                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1794                       libcfs_nid2str(nid), alive ? "up" : "down",
1795                       cfs_duration_sec(cfs_time_sub(when, now)));
1796                 return -EINVAL;
1797         }
1798
1799         if (ni != NULL && !alive &&             /* LND telling me she's down */
1800             !auto_down) {                       /* auto-down disabled */
1801                 CDEBUG(D_NET, "Auto-down disabled\n");
1802                 return 0;
1803         }
1804
1805         lnet_net_lock(cpt);
1806
1807         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1808                 lnet_net_unlock(cpt);
1809                 return -ESHUTDOWN;
1810         }
1811
1812         lp = lnet_find_peer_ni_locked(nid);
1813         if (lp == NULL) {
1814                 /* nid not found */
1815                 lnet_net_unlock(cpt);
1816                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1817                 return 0;
1818         }
1819
1820         /*
1821          * It is possible for this function to be called for the same peer
1822          * but with different NIs. We want to synchronize the notification
1823          * between the different calls. So we will use the lpni_cpt to
1824          * grab the net lock.
1825          */
1826         if (lp->lpni_cpt != cpt) {
1827                 lnet_net_unlock(cpt);
1828                 cpt = lp->lpni_cpt;
1829                 lnet_net_lock(cpt);
1830         }
1831
1832         /* We can't fully trust LND on reporting exact peer last_alive
1833          * if he notifies us about dead peer. For example ksocklnd can
1834          * call us with when == _time_when_the_node_was_booted_ if
1835          * no connections were successfully established */
1836         if (ni != NULL && !alive && when < lp->lpni_last_alive)
1837                 when = lp->lpni_last_alive;
1838
1839         lnet_notify_locked(lp, ni == NULL, alive, when);
1840
1841         if (ni != NULL)
1842                 lnet_ni_notify_locked(ni, lp);
1843
1844         lnet_peer_ni_decref_locked(lp);
1845
1846         lnet_net_unlock(cpt);
1847         return 0;
1848 }
1849 EXPORT_SYMBOL(lnet_notify);