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_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                 nnis = pbuf->pb_nnis;
 716
 717         list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 718                 int     down = 0;
 719                 int     up = 0;
 720                 int     i;
 721
 722                 /* If routing disabled then the route is down. */
 723                 if ((gw->lpni_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
 724                         rte->lr_downis = 1;
 725                         continue;
 726                 }
 727
 728                 for (i = 0; i < nnis; i++) {
 729                         lnet_ni_status_t *stat = &pbuf->pb_info.pi_ni[i];
 730                         lnet_nid_t       nid = stat->ns_nid;
 731
 732                         if (nid == LNET_NID_ANY) {
 733                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 734                                        libcfs_nid2str(gw->lpni_nid));
 735                                 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 736                                 goto out;
 737                         }
 738
 739                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 740                                 continue;
 741
 742                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 743                                 down++;
 744                                 continue;
 745                         }
 746
 747                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 748                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 749                                         up = 1;
 750                                         break;
 751                                 }
 752                                 continue;
 753                         }
 754
 755                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 756                                libcfs_nid2str(gw->lpni_nid), stat->ns_status);
 757                         gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 758                         goto out;
 759                 }
 760
 761                 if (up) { /* ignore downed NIs if NI for dest network is up */
 762                         rte->lr_downis = 0;
 763                         continue;
 764                 }
 765                 /* if @down is zero and this route is single-hop, it means
 766                  * we can't find NI for target network */
 767                 if (down == 0 && rte->lr_hops == 1)
 768                         down = 1;
 769
 770                 rte->lr_downis = down;
 771         }
 772 out:
 773         spin_unlock(&gw->lpni_lock);
 774 }
 775
 776 static void
 777 lnet_router_checker_event(struct lnet_event *event)
 778 {
 779         struct lnet_rc_data *rcd = event->md.user_ptr;
 780         struct lnet_peer_ni *lp;
 781
 782         LASSERT(rcd != NULL);
 783
 784         if (event->unlinked) {
 785                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 786                 return;
 787         }
 788
 789         LASSERT(event->type == LNET_EVENT_SEND ||
 790                 event->type == LNET_EVENT_REPLY);
 791
 792         lp = rcd->rcd_gateway;
 793         LASSERT(lp != NULL);
 794
 795          /* NB: it's called with holding lnet_res_lock, we have a few
 796           * places need to hold both locks at the same time, please take
 797           * care of lock ordering */
 798         lnet_net_lock(lp->lpni_cpt);
 799         if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
 800                 /* ignore if no longer a router or rcd is replaced */
 801                 goto out;
 802         }
 803
 804         if (event->type == LNET_EVENT_SEND) {
 805                 lp->lpni_ping_notsent = 0;
 806                 if (event->status == 0)
 807                         goto out;
 808         }
 809
 810         /* LNET_EVENT_REPLY */
 811         /* A successful REPLY means the router is up.  If _any_ comms
 812          * to the router fail I assume it's down (this will happen if
 813          * we ping alive routers to try to detect router death before
 814          * apps get burned). */
 815
 816         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
 817         /* The router checker will wake up very shortly and do the
 818          * actual notification.
 819          * XXX If 'lp' stops being a router before then, it will still
 820          * have the notification pending!!! */
 821
 822         if (avoid_asym_router_failure && event->status == 0)
 823                 lnet_parse_rc_info(rcd);
 824
 825  out:
 826         lnet_net_unlock(lp->lpni_cpt);
 827 }
 828
 829 static void
 830 lnet_wait_known_routerstate(void)
 831 {
 832         struct lnet_peer_ni *rtr;
 833         struct list_head *entry;
 834         int all_known;
 835
 836         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 837
 838         for (;;) {
 839                 int cpt = lnet_net_lock_current();
 840
 841                 all_known = 1;
 842                 list_for_each(entry, &the_lnet.ln_routers) {
 843                         rtr = list_entry(entry, struct lnet_peer_ni,
 844                                          lpni_rtr_list);
 845
 846                         spin_lock(&rtr->lpni_lock);
 847
 848                         if (rtr->lpni_alive_count == 0) {
 849                                 all_known = 0;
 850                                 spin_unlock(&rtr->lpni_lock);
 851                                 break;
 852                         }
 853                         spin_unlock(&rtr->lpni_lock);
 854                 }
 855
 856                 lnet_net_unlock(cpt);
 857
 858                 if (all_known)
 859                         return;
 860
 861                 set_current_state(TASK_UNINTERRUPTIBLE);
 862                 schedule_timeout(cfs_time_seconds(1));
 863         }
 864 }
 865
 866 void
 867 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
 868 {
 869         struct lnet_route *rte;
 870
 871         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
 872                 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 873                         if (rte->lr_net == net) {
 874                                 rte->lr_downis = 0;
 875                                 break;
 876                         }
 877                 }
 878         }
 879 }
 880
 881 static void
 882 lnet_update_ni_status_locked(void)
 883 {
 884         struct lnet_ni *ni = NULL;
 885         time64_t now;
 886         int timeout;
 887
 888         LASSERT(the_lnet.ln_routing);
 889
 890         timeout = router_ping_timeout +
 891                   MAX(live_router_check_interval, dead_router_check_interval);
 892
 893         now = ktime_get_real_seconds();
 894         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 895                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
 896                         continue;
 897
 898                 if (now < ni->ni_last_alive + timeout)
 899                         continue;
 900
 901                 lnet_ni_lock(ni);
 902                 /* re-check with lock */
 903                 if (now < ni->ni_last_alive + timeout) {
 904                         lnet_ni_unlock(ni);
 905                         continue;
 906                 }
 907
 908                 LASSERT(ni->ni_status != NULL);
 909
 910                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 911                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 912                                libcfs_nid2str(ni->ni_nid), timeout);
 913                         /* NB: so far, this is the only place to set
 914                          * NI status to "down" */
 915                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 916                 }
 917                 lnet_ni_unlock(ni);
 918         }
 919 }
 920
 921 static void
 922 lnet_destroy_rc_data(struct lnet_rc_data *rcd)
 923 {
 924         LASSERT(list_empty(&rcd->rcd_list));
 925         /* detached from network */
 926         LASSERT(LNetMDHandleIsInvalid(rcd->rcd_mdh));
 927
 928         if (rcd->rcd_gateway != NULL) {
 929                 int cpt = rcd->rcd_gateway->lpni_cpt;
 930
 931                 lnet_net_lock(cpt);
 932                 lnet_peer_ni_decref_locked(rcd->rcd_gateway);
 933                 lnet_net_unlock(cpt);
 934         }
 935
 936         if (rcd->rcd_pingbuffer != NULL)
 937                 lnet_ping_buffer_decref(rcd->rcd_pingbuffer);
 938
 939         LIBCFS_FREE(rcd, sizeof(*rcd));
 940 }
 941
 942 static struct lnet_rc_data *
 943 lnet_create_rc_data_locked(struct lnet_peer_ni *gateway)
 944 {
 945         struct lnet_rc_data             *rcd = NULL;
 946         struct lnet_ping_buffer *pbuf;
 947         int                     rc;
 948         int                     i;
 949
 950         lnet_net_unlock(gateway->lpni_cpt);
 951
 952         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 953         if (rcd == NULL)
 954                 goto out;
 955
 956         LNetInvalidateMDHandle(&rcd->rcd_mdh);
 957         INIT_LIST_HEAD(&rcd->rcd_list);
 958
 959         pbuf = lnet_ping_buffer_alloc(LNET_MAX_RTR_NIS, GFP_NOFS);
 960         if (pbuf == NULL)
 961                 goto out;
 962
 963         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 964                 pbuf->pb_info.pi_ni[i].ns_nid = LNET_NID_ANY;
 965                 pbuf->pb_info.pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 966         }
 967         rcd->rcd_pingbuffer = pbuf;
 968
 969         LASSERT(!LNetEQHandleIsInvalid(the_lnet.ln_rc_eqh));
 970         rc = LNetMDBind((struct lnet_md){.start     = &pbuf->pb_info,
 971                                     .user_ptr  = rcd,
 972                                     .length    = LNET_RTR_PINGINFO_SIZE,
 973                                     .threshold = LNET_MD_THRESH_INF,
 974                                     .options   = LNET_MD_TRUNCATE,
 975                                     .eq_handle = the_lnet.ln_rc_eqh},
 976                         LNET_UNLINK,
 977                         &rcd->rcd_mdh);
 978         if (rc < 0) {
 979                 CERROR("Can't bind MD: %d\n", rc);
 980                 goto out;
 981         }
 982         LASSERT(rc == 0);
 983
 984         lnet_net_lock(gateway->lpni_cpt);
 985         /* router table changed or someone has created rcd for this gateway */
 986         if (!lnet_isrouter(gateway) || gateway->lpni_rcd != NULL) {
 987                 lnet_net_unlock(gateway->lpni_cpt);
 988                 goto out;
 989         }
 990
 991         lnet_peer_ni_addref_locked(gateway);
 992         rcd->rcd_gateway = gateway;
 993         gateway->lpni_rcd = rcd;
 994         gateway->lpni_ping_notsent = 0;
 995
 996         return rcd;
 997
 998 out:
 999         if (rcd != NULL) {
1000                 if (!LNetMDHandleIsInvalid(rcd->rcd_mdh)) {
1001                         rc = LNetMDUnlink(rcd->rcd_mdh);
1002                         LASSERT(rc == 0);
1003                 }
1004                 lnet_destroy_rc_data(rcd);
1005         }
1006
1007         lnet_net_lock(gateway->lpni_cpt);
1008         return gateway->lpni_rcd;
1009 }
1010
1011 static int
1012 lnet_router_check_interval(struct lnet_peer_ni *rtr)
1013 {
1014         int secs;
1015
1016         secs = rtr->lpni_alive ? live_router_check_interval :
1017                                dead_router_check_interval;
1018         if (secs < 0)
1019                 secs = 0;
1020
1021         return secs;
1022 }
1023
1024 static void
1025 lnet_ping_router_locked(struct lnet_peer_ni *rtr)
1026 {
1027         struct lnet_rc_data *rcd = NULL;
1028         cfs_time_t      now = cfs_time_current();
1029         int             secs;
1030         struct lnet_ni  *ni;
1031
1032         lnet_peer_ni_addref_locked(rtr);
1033
1034         if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */
1035             cfs_time_after(now, rtr->lpni_ping_deadline))
1036                 lnet_notify_locked(rtr, 1, 0, now);
1037
1038         /* Run any outstanding notifications */
1039         ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL);
1040         lnet_ni_notify_locked(ni, rtr);
1041
1042         if (!lnet_isrouter(rtr) ||
1043             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1044                 /* router table changed or router checker is shutting down */
1045                 lnet_peer_ni_decref_locked(rtr);
1046                 return;
1047         }
1048
1049         rcd = rtr->lpni_rcd != NULL ?
1050               rtr->lpni_rcd : lnet_create_rc_data_locked(rtr);
1051
1052         if (rcd == NULL)
1053                 return;
1054
1055         secs = lnet_router_check_interval(rtr);
1056
1057         CDEBUG(D_NET,
1058                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1059                "alive_count %d lpni_ping_timestamp %lu\n",
1060                libcfs_nid2str(rtr->lpni_nid), secs,
1061                rtr->lpni_ping_deadline, rtr->lpni_ping_notsent,
1062                rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp);
1063
1064         if (secs != 0 && !rtr->lpni_ping_notsent &&
1065             cfs_time_after(now, cfs_time_add(rtr->lpni_ping_timestamp,
1066                                              cfs_time_seconds(secs)))) {
1067                 int               rc;
1068                 struct lnet_process_id id;
1069                 struct lnet_handle_md mdh;
1070
1071                 id.nid = rtr->lpni_nid;
1072                 id.pid = LNET_PID_LUSTRE;
1073                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1074
1075                 rtr->lpni_ping_notsent   = 1;
1076                 rtr->lpni_ping_timestamp = now;
1077
1078                 mdh = rcd->rcd_mdh;
1079
1080                 if (rtr->lpni_ping_deadline == 0) {
1081                         rtr->lpni_ping_deadline =
1082                                 cfs_time_shift(router_ping_timeout);
1083                 }
1084
1085                 lnet_net_unlock(rtr->lpni_cpt);
1086
1087                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1088                              LNET_PROTO_PING_MATCHBITS, 0);
1089
1090                 lnet_net_lock(rtr->lpni_cpt);
1091                 if (rc != 0)
1092                         rtr->lpni_ping_notsent = 0; /* no event pending */
1093         }
1094
1095         lnet_peer_ni_decref_locked(rtr);
1096         return;
1097 }
1098
1099 int
1100 lnet_router_checker_start(void)
1101 {
1102         int                     rc;
1103         int                     eqsz = 0;
1104         struct task_struct     *task;
1105
1106         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1107
1108         if (check_routers_before_use &&
1109             dead_router_check_interval <= 0) {
1110                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1111                                    " set if 'check_routers_before_use' is set"
1112                                    "\n");
1113                 return -EINVAL;
1114         }
1115
1116         sema_init(&the_lnet.ln_rc_signal, 0);
1117
1118         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1119         if (rc != 0) {
1120                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1121                 return -ENOMEM;
1122         }
1123
1124         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1125         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1126         if (IS_ERR(task)) {
1127                 rc = PTR_ERR(task);
1128                 CERROR("Can't start router checker thread: %d\n", rc);
1129                 /* block until event callback signals exit */
1130                 down(&the_lnet.ln_rc_signal);
1131                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1132                 LASSERT(rc == 0);
1133                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1134                 return -ENOMEM;
1135         }
1136
1137         if (check_routers_before_use) {
1138                 /* Note that a helpful side-effect of pinging all known routers
1139                  * at startup is that it makes them drop stale connections they
1140                  * may have to a previous instance of me. */
1141                 lnet_wait_known_routerstate();
1142         }
1143
1144         return 0;
1145 }
1146
1147 void
1148 lnet_router_checker_stop (void)
1149 {
1150         int rc;
1151
1152         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1153                 return;
1154
1155         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1156         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1157         /* wakeup the RC thread if it's sleeping */
1158         wake_up(&the_lnet.ln_rc_waitq);
1159
1160         /* block until event callback signals exit */
1161         down(&the_lnet.ln_rc_signal);
1162         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1163
1164         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1165         LASSERT(rc == 0);
1166         return;
1167 }
1168
1169 static void
1170 lnet_prune_rc_data(int wait_unlink)
1171 {
1172         struct lnet_rc_data *rcd;
1173         struct lnet_rc_data *tmp;
1174         struct lnet_peer_ni *lp;
1175         struct list_head head;
1176         int i = 2;
1177
1178         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1179                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1180                    list_empty(&the_lnet.ln_rcd_zombie)))
1181                 return;
1182
1183         INIT_LIST_HEAD(&head);
1184
1185         lnet_net_lock(LNET_LOCK_EX);
1186
1187         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1188                 /* router checker is stopping, prune all */
1189                 list_for_each_entry(lp, &the_lnet.ln_routers,
1190                                     lpni_rtr_list) {
1191                         if (lp->lpni_rcd == NULL)
1192                                 continue;
1193
1194                         LASSERT(list_empty(&lp->lpni_rcd->rcd_list));
1195                         list_add(&lp->lpni_rcd->rcd_list,
1196                                  &the_lnet.ln_rcd_deathrow);
1197                         lp->lpni_rcd = NULL;
1198                 }
1199         }
1200
1201         /* unlink all RCDs on deathrow list */
1202         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1203
1204         if (!list_empty(&head)) {
1205                 lnet_net_unlock(LNET_LOCK_EX);
1206
1207                 list_for_each_entry(rcd, &head, rcd_list)
1208                         LNetMDUnlink(rcd->rcd_mdh);
1209
1210                 lnet_net_lock(LNET_LOCK_EX);
1211         }
1212
1213         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1214
1215         /* release all zombie RCDs */
1216         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1217                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1218                                          rcd_list) {
1219                         if (LNetMDHandleIsInvalid(rcd->rcd_mdh))
1220                                 list_move(&rcd->rcd_list, &head);
1221                 }
1222
1223                 wait_unlink = wait_unlink &&
1224                               !list_empty(&the_lnet.ln_rcd_zombie);
1225
1226                 lnet_net_unlock(LNET_LOCK_EX);
1227
1228                 while (!list_empty(&head)) {
1229                         rcd = list_entry(head.next,
1230                                          struct lnet_rc_data, rcd_list);
1231                         list_del_init(&rcd->rcd_list);
1232                         lnet_destroy_rc_data(rcd);
1233                 }
1234
1235                 if (!wait_unlink)
1236                         return;
1237
1238                 i++;
1239                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1240                        "Waiting for rc buffers to unlink\n");
1241                 set_current_state(TASK_UNINTERRUPTIBLE);
1242                 schedule_timeout(cfs_time_seconds(1) / 4);
1243
1244                 lnet_net_lock(LNET_LOCK_EX);
1245         }
1246
1247         lnet_net_unlock(LNET_LOCK_EX);
1248 }
1249
1250 /*
1251  * This function is called to check if the RC should block indefinitely.
1252  * It's called from lnet_router_checker() as well as being passed to
1253  * wait_event_interruptible() to avoid the lost wake_up problem.
1254  *
1255  * When it's called from wait_event_interruptible() it is necessary to
1256  * also not sleep if the rc state is not running to avoid a deadlock
1257  * when the system is shutting down
1258  */
1259 static inline bool
1260 lnet_router_checker_active(void)
1261 {
1262         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1263                 return true;
1264
1265         /* Router Checker thread needs to run when routing is enabled in
1266          * order to call lnet_update_ni_status_locked() */
1267         if (the_lnet.ln_routing)
1268                 return true;
1269
1270         return !list_empty(&the_lnet.ln_routers) &&
1271                 (live_router_check_interval > 0 ||
1272                  dead_router_check_interval > 0);
1273 }
1274
1275 static int
1276 lnet_router_checker(void *arg)
1277 {
1278         struct lnet_peer_ni *rtr;
1279         struct list_head *entry;
1280
1281         cfs_block_allsigs();
1282
1283         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1284                 __u64   version;
1285                 int     cpt;
1286                 int     cpt2;
1287
1288                 cpt = lnet_net_lock_current();
1289 rescan:
1290                 version = the_lnet.ln_routers_version;
1291
1292                 list_for_each(entry, &the_lnet.ln_routers) {
1293                         rtr = list_entry(entry, struct lnet_peer_ni,
1294                                          lpni_rtr_list);
1295
1296                         cpt2 = rtr->lpni_cpt;
1297                         if (cpt != cpt2) {
1298                                 lnet_net_unlock(cpt);
1299                                 cpt = cpt2;
1300                                 lnet_net_lock(cpt);
1301                                 /* the routers list has changed */
1302                                 if (version != the_lnet.ln_routers_version)
1303                                         goto rescan;
1304                         }
1305
1306                         lnet_ping_router_locked(rtr);
1307
1308                         /* NB dropped lock */
1309                         if (version != the_lnet.ln_routers_version) {
1310                                 /* the routers list has changed */
1311                                 goto rescan;
1312                         }
1313                 }
1314
1315                 if (the_lnet.ln_routing)
1316                         lnet_update_ni_status_locked();
1317
1318                 lnet_net_unlock(cpt);
1319
1320                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1321
1322                 /* Call schedule_timeout() here always adds 1 to load average
1323                  * because kernel counts # active tasks as nr_running
1324                  * + nr_uninterruptible. */
1325                 /* if there are any routes then wakeup every second.  If
1326                  * there are no routes then sleep indefinitely until woken
1327                  * up by a user adding a route */
1328                 if (!lnet_router_checker_active())
1329                         wait_event_interruptible(the_lnet.ln_rc_waitq,
1330                                                  lnet_router_checker_active());
1331                 else
1332                         wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1333                                                          false,
1334                                                          cfs_time_seconds(1));
1335         }
1336
1337         lnet_prune_rc_data(1); /* wait for UNLINK */
1338
1339         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1340         up(&the_lnet.ln_rc_signal);
1341         /* The unlink event callback will signal final completion */
1342         return 0;
1343 }
1344
1345 void
1346 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1347 {
1348         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1349
1350         while (--npages >= 0)
1351                 __free_page(rb->rb_kiov[npages].kiov_page);
1352
1353         LIBCFS_FREE(rb, sz);
1354 }
1355
1356 static struct lnet_rtrbuf *
1357 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1358 {
1359         int            npages = rbp->rbp_npages;
1360         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1361         struct page   *page;
1362         struct lnet_rtrbuf *rb;
1363         int            i;
1364
1365         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1366         if (rb == NULL)
1367                 return NULL;
1368
1369         rb->rb_pool = rbp;
1370
1371         for (i = 0; i < npages; i++) {
1372                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1373                                           GFP_KERNEL | __GFP_ZERO);
1374                 if (page == NULL) {
1375                         while (--i >= 0)
1376                                 __free_page(rb->rb_kiov[i].kiov_page);
1377
1378                         LIBCFS_FREE(rb, sz);
1379                         return NULL;
1380                 }
1381
1382                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1383                 rb->rb_kiov[i].kiov_offset = 0;
1384                 rb->rb_kiov[i].kiov_page = page;
1385         }
1386
1387         return rb;
1388 }
1389
1390 static void
1391 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1392 {
1393         int npages = rbp->rbp_npages;
1394         struct lnet_rtrbuf *rb;
1395         struct list_head tmp;
1396
1397         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1398                 return;
1399
1400         INIT_LIST_HEAD(&tmp);
1401
1402         lnet_net_lock(cpt);
1403         list_splice_init(&rbp->rbp_msgs, &tmp);
1404         lnet_drop_routed_msgs_locked(&tmp, cpt);
1405         list_splice_init(&rbp->rbp_bufs, &tmp);
1406         rbp->rbp_req_nbuffers = 0;
1407         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1408         rbp->rbp_mincredits = 0;
1409         lnet_net_unlock(cpt);
1410
1411         /* Free buffers on the free list. */
1412         while (!list_empty(&tmp)) {
1413                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1414                 list_del(&rb->rb_list);
1415                 lnet_destroy_rtrbuf(rb, npages);
1416         }
1417 }
1418
1419 static int
1420 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1421 {
1422         struct list_head rb_list;
1423         struct lnet_rtrbuf *rb;
1424         int             num_rb;
1425         int             num_buffers = 0;
1426         int             old_req_nbufs;
1427         int             npages = rbp->rbp_npages;
1428
1429         lnet_net_lock(cpt);
1430         /* If we are called for less buffers than already in the pool, we
1431          * just lower the req_nbuffers number and excess buffers will be
1432          * thrown away as they are returned to the free list.  Credits
1433          * then get adjusted as well.
1434          * If we already have enough buffers allocated to serve the
1435          * increase requested, then we can treat that the same way as we
1436          * do the decrease. */
1437         num_rb = nbufs - rbp->rbp_nbuffers;
1438         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1439                 rbp->rbp_req_nbuffers = nbufs;
1440                 lnet_net_unlock(cpt);
1441                 return 0;
1442         }
1443         /* store the older value of rbp_req_nbuffers and then set it to
1444          * the new request to prevent lnet_return_rx_credits_locked() from
1445          * freeing buffers that we need to keep around */
1446         old_req_nbufs = rbp->rbp_req_nbuffers;
1447         rbp->rbp_req_nbuffers = nbufs;
1448         lnet_net_unlock(cpt);
1449
1450         INIT_LIST_HEAD(&rb_list);
1451
1452         /* allocate the buffers on a local list first.  If all buffers are
1453          * allocated successfully then join this list to the rbp buffer
1454          * list.  If not then free all allocated buffers. */
1455         while (num_rb-- > 0) {
1456                 rb = lnet_new_rtrbuf(rbp, cpt);
1457                 if (rb == NULL) {
1458                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1459                                nbufs, npages);
1460
1461                         lnet_net_lock(cpt);
1462                         rbp->rbp_req_nbuffers = old_req_nbufs;
1463                         lnet_net_unlock(cpt);
1464
1465                         goto failed;
1466                 }
1467
1468                 list_add(&rb->rb_list, &rb_list);
1469                 num_buffers++;
1470         }
1471
1472         lnet_net_lock(cpt);
1473
1474         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1475         rbp->rbp_nbuffers += num_buffers;
1476         rbp->rbp_credits += num_buffers;
1477         rbp->rbp_mincredits = rbp->rbp_credits;
1478         /* We need to schedule blocked msg using the newly
1479          * added buffers. */
1480         while (!list_empty(&rbp->rbp_bufs) &&
1481                !list_empty(&rbp->rbp_msgs))
1482                 lnet_schedule_blocked_locked(rbp);
1483
1484         lnet_net_unlock(cpt);
1485
1486         return 0;
1487
1488 failed:
1489         while (!list_empty(&rb_list)) {
1490                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1491                 list_del(&rb->rb_list);
1492                 lnet_destroy_rtrbuf(rb, npages);
1493         }
1494
1495         return -ENOMEM;
1496 }
1497
1498 static void
1499 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1500 {
1501         INIT_LIST_HEAD(&rbp->rbp_msgs);
1502         INIT_LIST_HEAD(&rbp->rbp_bufs);
1503
1504         rbp->rbp_npages = npages;
1505         rbp->rbp_credits = 0;
1506         rbp->rbp_mincredits = 0;
1507 }
1508
1509 void
1510 lnet_rtrpools_free(int keep_pools)
1511 {
1512         struct lnet_rtrbufpool *rtrp;
1513         int               i;
1514
1515         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1516                 return;
1517
1518         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1519                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1520                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1521                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1522         }
1523
1524         if (!keep_pools) {
1525                 cfs_percpt_free(the_lnet.ln_rtrpools);
1526                 the_lnet.ln_rtrpools = NULL;
1527         }
1528 }
1529
1530 static int
1531 lnet_nrb_tiny_calculate(void)
1532 {
1533         int     nrbs = LNET_NRB_TINY;
1534
1535         if (tiny_router_buffers < 0) {
1536                 LCONSOLE_ERROR_MSG(0x10c,
1537                                    "tiny_router_buffers=%d invalid when "
1538                                    "routing enabled\n", tiny_router_buffers);
1539                 return -EINVAL;
1540         }
1541
1542         if (tiny_router_buffers > 0)
1543                 nrbs = tiny_router_buffers;
1544
1545         nrbs /= LNET_CPT_NUMBER;
1546         return max(nrbs, LNET_NRB_TINY_MIN);
1547 }
1548
1549 static int
1550 lnet_nrb_small_calculate(void)
1551 {
1552         int     nrbs = LNET_NRB_SMALL;
1553
1554         if (small_router_buffers < 0) {
1555                 LCONSOLE_ERROR_MSG(0x10c,
1556                                    "small_router_buffers=%d invalid when "
1557                                    "routing enabled\n", small_router_buffers);
1558                 return -EINVAL;
1559         }
1560
1561         if (small_router_buffers > 0)
1562                 nrbs = small_router_buffers;
1563
1564         nrbs /= LNET_CPT_NUMBER;
1565         return max(nrbs, LNET_NRB_SMALL_MIN);
1566 }
1567
1568 static int
1569 lnet_nrb_large_calculate(void)
1570 {
1571         int     nrbs = LNET_NRB_LARGE;
1572
1573         if (large_router_buffers < 0) {
1574                 LCONSOLE_ERROR_MSG(0x10c,
1575                                    "large_router_buffers=%d invalid when "
1576                                    "routing enabled\n", large_router_buffers);
1577                 return -EINVAL;
1578         }
1579
1580         if (large_router_buffers > 0)
1581                 nrbs = large_router_buffers;
1582
1583         nrbs /= LNET_CPT_NUMBER;
1584         return max(nrbs, LNET_NRB_LARGE_MIN);
1585 }
1586
1587 int
1588 lnet_rtrpools_alloc(int im_a_router)
1589 {
1590         struct lnet_rtrbufpool *rtrp;
1591         int     nrb_tiny;
1592         int     nrb_small;
1593         int     nrb_large;
1594         int     rc;
1595         int     i;
1596
1597         if (!strcmp(forwarding, "")) {
1598                 /* not set either way */
1599                 if (!im_a_router)
1600                         return 0;
1601         } else if (!strcmp(forwarding, "disabled")) {
1602                 /* explicitly disabled */
1603                 return 0;
1604         } else if (!strcmp(forwarding, "enabled")) {
1605                 /* explicitly enabled */
1606         } else {
1607                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1608                                    "'enabled' or 'disabled'\n");
1609                 return -EINVAL;
1610         }
1611
1612         nrb_tiny = lnet_nrb_tiny_calculate();
1613         if (nrb_tiny < 0)
1614                 return -EINVAL;
1615
1616         nrb_small = lnet_nrb_small_calculate();
1617         if (nrb_small < 0)
1618                 return -EINVAL;
1619
1620         nrb_large = lnet_nrb_large_calculate();
1621         if (nrb_large < 0)
1622                 return -EINVAL;
1623
1624         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1625                                                 LNET_NRBPOOLS *
1626                                                 sizeof(struct lnet_rtrbufpool));
1627         if (the_lnet.ln_rtrpools == NULL) {
1628                 LCONSOLE_ERROR_MSG(0x10c,
1629                                    "Failed to initialize router buffe pool\n");
1630                 return -ENOMEM;
1631         }
1632
1633         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1634                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1635                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1636                                               nrb_tiny, i);
1637                 if (rc != 0)
1638                         goto failed;
1639
1640                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1641                                   LNET_NRB_SMALL_PAGES);
1642                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1643                                               nrb_small, i);
1644                 if (rc != 0)
1645                         goto failed;
1646
1647                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1648                                   LNET_NRB_LARGE_PAGES);
1649                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1650                                               nrb_large, i);
1651                 if (rc != 0)
1652                         goto failed;
1653         }
1654
1655         lnet_net_lock(LNET_LOCK_EX);
1656         the_lnet.ln_routing = 1;
1657         lnet_net_unlock(LNET_LOCK_EX);
1658         return 0;
1659
1660  failed:
1661         lnet_rtrpools_free(0);
1662         return rc;
1663 }
1664
1665 static int
1666 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1667 {
1668         int nrb = 0;
1669         int rc = 0;
1670         int i;
1671         struct lnet_rtrbufpool *rtrp;
1672
1673         /* If the provided values for each buffer pool are different than the
1674          * configured values, we need to take action. */
1675         if (tiny >= 0) {
1676                 tiny_router_buffers = tiny;
1677                 nrb = lnet_nrb_tiny_calculate();
1678                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1679                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1680                                                       nrb, i);
1681                         if (rc != 0)
1682                                 return rc;
1683                 }
1684         }
1685         if (small >= 0) {
1686                 small_router_buffers = small;
1687                 nrb = lnet_nrb_small_calculate();
1688                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1689                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1690                                                       nrb, i);
1691                         if (rc != 0)
1692                                 return rc;
1693                 }
1694         }
1695         if (large >= 0) {
1696                 large_router_buffers = large;
1697                 nrb = lnet_nrb_large_calculate();
1698                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1699                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1700                                                       nrb, i);
1701                         if (rc != 0)
1702                                 return rc;
1703                 }
1704         }
1705
1706         return 0;
1707 }
1708
1709 int
1710 lnet_rtrpools_adjust(int tiny, int small, int large)
1711 {
1712         /* this function doesn't revert the changes if adding new buffers
1713          * failed.  It's up to the user space caller to revert the
1714          * changes. */
1715
1716         if (!the_lnet.ln_routing)
1717                 return 0;
1718
1719         return lnet_rtrpools_adjust_helper(tiny, small, large);
1720 }
1721
1722 int
1723 lnet_rtrpools_enable(void)
1724 {
1725         int rc = 0;
1726
1727         if (the_lnet.ln_routing)
1728                 return 0;
1729
1730         if (the_lnet.ln_rtrpools == NULL)
1731                 /* If routing is turned off, and we have never
1732                  * initialized the pools before, just call the
1733                  * standard buffer pool allocation routine as
1734                  * if we are just configuring this for the first
1735                  * time. */
1736                 rc = lnet_rtrpools_alloc(1);
1737         else
1738                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1739         if (rc != 0)
1740                 return rc;
1741
1742         lnet_net_lock(LNET_LOCK_EX);
1743         the_lnet.ln_routing = 1;
1744
1745         the_lnet.ln_ping_target->pb_info.pi_features &=
1746                 ~LNET_PING_FEAT_RTE_DISABLED;
1747         lnet_net_unlock(LNET_LOCK_EX);
1748
1749         return rc;
1750 }
1751
1752 void
1753 lnet_rtrpools_disable(void)
1754 {
1755         if (!the_lnet.ln_routing)
1756                 return;
1757
1758         lnet_net_lock(LNET_LOCK_EX);
1759         the_lnet.ln_routing = 0;
1760         the_lnet.ln_ping_target->pb_info.pi_features |=
1761                 LNET_PING_FEAT_RTE_DISABLED;
1762
1763         tiny_router_buffers = 0;
1764         small_router_buffers = 0;
1765         large_router_buffers = 0;
1766         lnet_net_unlock(LNET_LOCK_EX);
1767         lnet_rtrpools_free(1);
1768 }
1769
1770 int
1771 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1772 {
1773         struct lnet_peer_ni *lp = NULL;
1774         cfs_time_t now = cfs_time_current();
1775         int cpt = lnet_cpt_of_nid(nid, ni);
1776
1777         LASSERT (!in_interrupt ());
1778
1779         CDEBUG (D_NET, "%s notifying %s: %s\n",
1780                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1781                 libcfs_nid2str(nid),
1782                 alive ? "up" : "down");
1783
1784         if (ni != NULL &&
1785             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1786                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1787                       libcfs_nid2str(nid), alive ? "birth" : "death",
1788                       libcfs_nid2str(ni->ni_nid));
1789                 return -EINVAL;
1790         }
1791
1792         /* can't do predictions... */
1793         if (cfs_time_after(when, now)) {
1794                 CWARN("Ignoring prediction from %s of %s %s "
1795                       "%ld seconds in the future\n",
1796                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1797                       libcfs_nid2str(nid), alive ? "up" : "down",
1798                       cfs_duration_sec(cfs_time_sub(when, now)));
1799                 return -EINVAL;
1800         }
1801
1802         if (ni != NULL && !alive &&             /* LND telling me she's down */
1803             !auto_down) {                       /* auto-down disabled */
1804                 CDEBUG(D_NET, "Auto-down disabled\n");
1805                 return 0;
1806         }
1807
1808         lnet_net_lock(cpt);
1809
1810         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1811                 lnet_net_unlock(cpt);
1812                 return -ESHUTDOWN;
1813         }
1814
1815         lp = lnet_find_peer_ni_locked(nid);
1816         if (lp == NULL) {
1817                 /* nid not found */
1818                 lnet_net_unlock(cpt);
1819                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1820                 return 0;
1821         }
1822
1823         /*
1824          * It is possible for this function to be called for the same peer
1825          * but with different NIs. We want to synchronize the notification
1826          * between the different calls. So we will use the lpni_cpt to
1827          * grab the net lock.
1828          */
1829         if (lp->lpni_cpt != cpt) {
1830                 lnet_net_unlock(cpt);
1831                 cpt = lp->lpni_cpt;
1832                 lnet_net_lock(cpt);
1833         }
1834
1835         /* We can't fully trust LND on reporting exact peer last_alive
1836          * if he notifies us about dead peer. For example ksocklnd can
1837          * call us with when == _time_when_the_node_was_booted_ if
1838          * no connections were successfully established */
1839         if (ni != NULL && !alive && when < lp->lpni_last_alive)
1840                 when = lp->lpni_last_alive;
1841
1842         lnet_notify_locked(lp, ni == NULL, alive, when);
1843
1844         if (ni != NULL)
1845                 lnet_ni_notify_locked(ni, lp);
1846
1847         lnet_peer_ni_decref_locked(lp);
1848
1849         lnet_net_unlock(cpt);
1850         return 0;
1851 }
1852 EXPORT_SYMBOL(lnet_notify);