lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2013, Intel Corporation.
   5  *
   6  *   This file is part of Portals
   7  *   http://sourceforge.net/projects/sandiaportals/
   8  *
   9  *   Portals is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Portals is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Portals; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LNET
  25 #include <lnet/lib-lnet.h>
  26
  27 #if defined(__KERNEL__) && defined(LNET_ROUTER)
  28
  29 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
  30 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
  31 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
  32 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
  33 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  34 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  35
  36 static char *forwarding = "";
  37 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
  38                 "Explicitly enable/disable forwarding between networks");
  39
  40 static int tiny_router_buffers;
  41 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
  42                 "# of 0 payload messages to buffer in the router");
  43 static int small_router_buffers;
  44 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
  45                 "# of small (1 page) messages to buffer in the router");
  46 static int large_router_buffers;
  47 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
  48                 "# of large messages to buffer in the router");
  49 static int peer_buffer_credits = 0;
  50 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  51                 "# router buffer credits per peer");
  52
  53 static int auto_down = 1;
  54 CFS_MODULE_PARM(auto_down, "i", int, 0444,
  55                 "Automatically mark peers down on comms error");
  56
  57 int
  58 lnet_peer_buffer_credits(lnet_ni_t *ni)
  59 {
  60         /* NI option overrides LNet default */
  61         if (ni->ni_peerrtrcredits > 0)
  62                 return ni->ni_peerrtrcredits;
  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 ni->ni_peertxcredits;
  69 }
  70
  71 /* forward ref's */
  72 static int lnet_router_checker(void *);
  73 #else
  74
  75 int
  76 lnet_peer_buffer_credits(lnet_ni_t *ni)
  77 {
  78         return 0;
  79 }
  80
  81 #endif
  82
  83 static int check_routers_before_use = 0;
  84 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
  85                 "Assume routers are down and ping them before use");
  86
  87 static int avoid_asym_router_failure = 1;
  88 CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644,
  89                 "Avoid asymmetrical router failures (0 to disable)");
  90
  91 static int dead_router_check_interval = 60;
  92 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644,
  93                 "Seconds between dead router health checks (<= 0 to disable)");
  94
  95 static int live_router_check_interval = 60;
  96 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644,
  97                 "Seconds between live router health checks (<= 0 to disable)");
  98
  99 static int router_ping_timeout = 50;
 100 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644,
 101                 "Seconds to wait for the reply to a router health query");
 102
 103 int
 104 lnet_peers_start_down(void)
 105 {
 106         return check_routers_before_use;
 107 }
 108
 109 void
 110 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
 111 {
 112         if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
 113                 CDEBUG(D_NET, "Out of date\n");
 114                 return;
 115         }
 116
 117         lp->lp_timestamp = when;                /* update timestamp */
 118         lp->lp_ping_deadline = 0;               /* disable ping timeout */
 119
 120         if (lp->lp_alive_count != 0 &&          /* got old news */
 121             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
 122                 CDEBUG(D_NET, "Old news\n");
 123                 return;
 124         }
 125
 126         /* Flag that notification is outstanding */
 127
 128         lp->lp_alive_count++;
 129         lp->lp_alive = !(!alive);               /* 1 bit! */
 130         lp->lp_notify = 1;
 131         lp->lp_notifylnd |= notifylnd;
 132         if (lp->lp_alive)
 133                 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 134
 135         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
 136 }
 137
 138 void
 139 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
 140 {
 141         int        alive;
 142         int        notifylnd;
 143
 144         /* Notify only in 1 thread at any time to ensure ordered notification.
 145          * NB individual events can be missed; the only guarantee is that you
 146          * always get the most recent news */
 147
 148         if (lp->lp_notifying)
 149                 return;
 150
 151         lp->lp_notifying = 1;
 152
 153         while (lp->lp_notify) {
 154                 alive     = lp->lp_alive;
 155                 notifylnd = lp->lp_notifylnd;
 156
 157                 lp->lp_notifylnd = 0;
 158                 lp->lp_notify    = 0;
 159
 160                 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
 161                         lnet_net_unlock(lp->lp_cpt);
 162
 163                         /* A new notification could happen now; I'll handle it
 164                          * when control returns to me */
 165
 166                         (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
 167
 168                         lnet_net_lock(lp->lp_cpt);
 169                 }
 170         }
 171
 172         lp->lp_notifying = 0;
 173 }
 174
 175
 176 static void
 177 lnet_rtr_addref_locked(lnet_peer_t *lp)
 178 {
 179         LASSERT(lp->lp_refcount > 0);
 180         LASSERT(lp->lp_rtr_refcount >= 0);
 181
 182         /* lnet_net_lock must be exclusively locked */
 183         lp->lp_rtr_refcount++;
 184         if (lp->lp_rtr_refcount == 1) {
 185                 cfs_list_t *pos;
 186
 187                 /* a simple insertion sort */
 188                 cfs_list_for_each_prev(pos, &the_lnet.ln_routers) {
 189                         lnet_peer_t *rtr = cfs_list_entry(pos, lnet_peer_t,
 190                                                           lp_rtr_list);
 191
 192                         if (rtr->lp_nid < lp->lp_nid)
 193                                 break;
 194                 }
 195
 196                 cfs_list_add(&lp->lp_rtr_list, pos);
 197                 /* addref for the_lnet.ln_routers */
 198                 lnet_peer_addref_locked(lp);
 199                 the_lnet.ln_routers_version++;
 200         }
 201 }
 202
 203 static void
 204 lnet_rtr_decref_locked(lnet_peer_t *lp)
 205 {
 206         LASSERT(lp->lp_refcount > 0);
 207         LASSERT(lp->lp_rtr_refcount > 0);
 208
 209         /* lnet_net_lock must be exclusively locked */
 210         lp->lp_rtr_refcount--;
 211         if (lp->lp_rtr_refcount == 0) {
 212                 LASSERT(cfs_list_empty(&lp->lp_routes));
 213
 214                 if (lp->lp_rcd != NULL) {
 215                         cfs_list_add(&lp->lp_rcd->rcd_list,
 216                                      &the_lnet.ln_rcd_deathrow);
 217                         lp->lp_rcd = NULL;
 218                 }
 219
 220                 cfs_list_del(&lp->lp_rtr_list);
 221                 /* decref for the_lnet.ln_routers */
 222                 lnet_peer_decref_locked(lp);
 223                 the_lnet.ln_routers_version++;
 224         }
 225 }
 226
 227 lnet_remotenet_t *
 228 lnet_find_net_locked (__u32 net)
 229 {
 230         lnet_remotenet_t        *rnet;
 231         cfs_list_t              *tmp;
 232         cfs_list_t              *rn_list;
 233
 234         LASSERT(!the_lnet.ln_shutdown);
 235
 236         rn_list = lnet_net2rnethash(net);
 237         cfs_list_for_each(tmp, rn_list) {
 238                 rnet = cfs_list_entry(tmp, lnet_remotenet_t, lrn_list);
 239
 240                 if (rnet->lrn_net == net)
 241                         return rnet;
 242         }
 243         return NULL;
 244 }
 245
 246 static void lnet_shuffle_seed(void)
 247 {
 248         static int seeded = 0;
 249         int lnd_type, seed[2];
 250         struct timeval tv;
 251         lnet_ni_t *ni;
 252         cfs_list_t *tmp;
 253
 254         if (seeded)
 255                 return;
 256
 257         cfs_get_random_bytes(seed, sizeof(seed));
 258
 259         /* Nodes with small feet have little entropy
 260          * the NID for this node gives the most entropy in the low bits */
 261         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 262                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 263                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 264
 265                 if (lnd_type != LOLND)
 266                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
 267         }
 268
 269         do_gettimeofday(&tv);
 270         cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
 271         seeded = 1;
 272         return;
 273 }
 274
 275 /* NB expects LNET_LOCK held */
 276 void
 277 lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route)
 278 {
 279         unsigned int      len = 0;
 280         unsigned int      offset = 0;
 281         cfs_list_t       *e;
 282
 283         lnet_shuffle_seed();
 284
 285         cfs_list_for_each (e, &rnet->lrn_routes) {
 286                 len++;
 287         }
 288
 289         /* len+1 positions to add a new entry, also prevents division by 0 */
 290         offset = cfs_rand() % (len + 1);
 291         cfs_list_for_each (e, &rnet->lrn_routes) {
 292                 if (offset == 0)
 293                         break;
 294                 offset--;
 295         }
 296         cfs_list_add(&route->lr_list, e);
 297         cfs_list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 298
 299         the_lnet.ln_remote_nets_version++;
 300         lnet_rtr_addref_locked(route->lr_gateway);
 301 }
 302
 303 int
 304 lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway,
 305                unsigned int priority)
 306 {
 307         cfs_list_t          *e;
 308         lnet_remotenet_t    *rnet;
 309         lnet_remotenet_t    *rnet2;
 310         lnet_route_t        *route;
 311         lnet_ni_t           *ni;
 312         int                  add_route;
 313         int                  rc;
 314
 315         CDEBUG(D_NET, "Add route: net %s hops %u priority %u gw %s\n",
 316                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 317
 318         if (gateway == LNET_NID_ANY ||
 319             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 320             net == LNET_NIDNET(LNET_NID_ANY) ||
 321             LNET_NETTYP(net) == LOLND ||
 322             LNET_NIDNET(gateway) == net ||
 323             hops < 1 || hops > 255)
 324                 return (-EINVAL);
 325
 326         if (lnet_islocalnet(net))               /* it's a local network */
 327                 return 0;                       /* ignore the route entry */
 328
 329         /* Assume net, route, all new */
 330         LIBCFS_ALLOC(route, sizeof(*route));
 331         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 332         if (route == NULL || rnet == NULL) {
 333                 CERROR("Out of memory creating route %s %d %s\n",
 334                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 335                 if (route != NULL)
 336                         LIBCFS_FREE(route, sizeof(*route));
 337                 if (rnet != NULL)
 338                         LIBCFS_FREE(rnet, sizeof(*rnet));
 339                 return -ENOMEM;
 340         }
 341
 342         CFS_INIT_LIST_HEAD(&rnet->lrn_routes);
 343         rnet->lrn_net = net;
 344         route->lr_hops = hops;
 345         route->lr_net = net;
 346         route->lr_priority = priority;
 347
 348         lnet_net_lock(LNET_LOCK_EX);
 349
 350         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
 351         if (rc != 0) {
 352                 lnet_net_unlock(LNET_LOCK_EX);
 353
 354                 LIBCFS_FREE(route, sizeof(*route));
 355                 LIBCFS_FREE(rnet, sizeof(*rnet));
 356
 357                 if (rc == -EHOSTUNREACH) { /* gateway is not on a local net */
 358                         return 0;       /* ignore the route entry */
 359                 } else {
 360                         CERROR("Error %d creating route %s %d %s\n", rc,
 361                                libcfs_net2str(net), hops,
 362                                libcfs_nid2str(gateway));
 363                 }
 364                 return rc;
 365         }
 366
 367         LASSERT (!the_lnet.ln_shutdown);
 368
 369         rnet2 = lnet_find_net_locked(net);
 370         if (rnet2 == NULL) {
 371                 /* new network */
 372                 cfs_list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 373                 rnet2 = rnet;
 374         }
 375
 376         /* Search for a duplicate route (it's a NOOP if it is) */
 377         add_route = 1;
 378         cfs_list_for_each (e, &rnet2->lrn_routes) {
 379                 lnet_route_t *route2 = cfs_list_entry(e, lnet_route_t, lr_list);
 380
 381                 if (route2->lr_gateway == route->lr_gateway) {
 382                         add_route = 0;
 383                         break;
 384                 }
 385
 386                 /* our lookups must be true */
 387                 LASSERT (route2->lr_gateway->lp_nid != gateway);
 388         }
 389
 390         if (add_route) {
 391                 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
 392                 lnet_add_route_to_rnet(rnet2, route);
 393
 394                 ni = route->lr_gateway->lp_ni;
 395                 lnet_net_unlock(LNET_LOCK_EX);
 396
 397                 /* XXX Assume alive */
 398                 if (ni->ni_lnd->lnd_notify != NULL)
 399                         (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
 400
 401                 lnet_net_lock(LNET_LOCK_EX);
 402         }
 403
 404         /* -1 for notify or !add_route */
 405         lnet_peer_decref_locked(route->lr_gateway);
 406         lnet_net_unlock(LNET_LOCK_EX);
 407
 408         if (!add_route)
 409                 LIBCFS_FREE(route, sizeof(*route));
 410
 411         if (rnet != rnet2)
 412                 LIBCFS_FREE(rnet, sizeof(*rnet));
 413
 414         return 0;
 415 }
 416
 417 int
 418 lnet_check_routes(void)
 419 {
 420         lnet_remotenet_t        *rnet;
 421         lnet_route_t            *route;
 422         lnet_route_t            *route2;
 423         cfs_list_t              *e1;
 424         cfs_list_t              *e2;
 425         int                     cpt;
 426         cfs_list_t              *rn_list;
 427         int                     i;
 428
 429         cpt = lnet_net_lock_current();
 430
 431         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 432                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 433                 cfs_list_for_each(e1, rn_list) {
 434                         rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 435
 436                         route2 = NULL;
 437                         cfs_list_for_each(e2, &rnet->lrn_routes) {
 438                                 lnet_nid_t      nid1;
 439                                 lnet_nid_t      nid2;
 440                                 int             net;
 441
 442                                 route = cfs_list_entry(e2, lnet_route_t,
 443                                                        lr_list);
 444
 445                                 if (route2 == NULL) {
 446                                         route2 = route;
 447                                         continue;
 448                                 }
 449
 450                                 if (route->lr_gateway->lp_ni ==
 451                                     route2->lr_gateway->lp_ni)
 452                                         continue;
 453
 454                                 nid1 = route->lr_gateway->lp_nid;
 455                                 nid2 = route2->lr_gateway->lp_nid;
 456                                 net = rnet->lrn_net;
 457
 458                                 lnet_net_unlock(cpt);
 459
 460                                 CERROR("Routes to %s via %s and %s not "
 461                                        "supported\n",
 462                                        libcfs_net2str(net),
 463                                        libcfs_nid2str(nid1),
 464                                        libcfs_nid2str(nid2));
 465                                 return -EINVAL;
 466                         }
 467                 }
 468         }
 469
 470         lnet_net_unlock(cpt);
 471         return 0;
 472 }
 473
 474 int
 475 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 476 {
 477         struct lnet_peer        *gateway;
 478         lnet_remotenet_t        *rnet;
 479         lnet_route_t            *route;
 480         cfs_list_t              *e1;
 481         cfs_list_t              *e2;
 482         int                     rc = -ENOENT;
 483         cfs_list_t              *rn_list;
 484         int                     idx = 0;
 485
 486         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 487                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 488
 489         /* NB Caller may specify either all routes via the given gateway
 490          * or a specific route entry actual NIDs) */
 491
 492         lnet_net_lock(LNET_LOCK_EX);
 493         if (net == LNET_NIDNET(LNET_NID_ANY))
 494                 rn_list = &the_lnet.ln_remote_nets_hash[0];
 495         else
 496                 rn_list = lnet_net2rnethash(net);
 497
 498  again:
 499         cfs_list_for_each(e1, rn_list) {
 500                 rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 501
 502                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 503                         net == rnet->lrn_net))
 504                         continue;
 505
 506                 cfs_list_for_each(e2, &rnet->lrn_routes) {
 507                         route = cfs_list_entry(e2, lnet_route_t, lr_list);
 508
 509                         gateway = route->lr_gateway;
 510                         if (!(gw_nid == LNET_NID_ANY ||
 511                               gw_nid == gateway->lp_nid))
 512                                 continue;
 513
 514                         cfs_list_del(&route->lr_list);
 515                         cfs_list_del(&route->lr_gwlist);
 516                         the_lnet.ln_remote_nets_version++;
 517
 518                         if (cfs_list_empty(&rnet->lrn_routes))
 519                                 cfs_list_del(&rnet->lrn_list);
 520                         else
 521                                 rnet = NULL;
 522
 523                         lnet_rtr_decref_locked(gateway);
 524                         lnet_peer_decref_locked(gateway);
 525
 526                         lnet_net_unlock(LNET_LOCK_EX);
 527
 528                         LIBCFS_FREE(route, sizeof(*route));
 529
 530                         if (rnet != NULL)
 531                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 532
 533                         rc = 0;
 534                         lnet_net_lock(LNET_LOCK_EX);
 535                         goto again;
 536                 }
 537         }
 538
 539         if (net == LNET_NIDNET(LNET_NID_ANY) &&
 540             ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
 541                 rn_list = &the_lnet.ln_remote_nets_hash[idx];
 542                 goto again;
 543         }
 544         lnet_net_unlock(LNET_LOCK_EX);
 545
 546         return rc;
 547 }
 548
 549 void
 550 lnet_destroy_routes (void)
 551 {
 552         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 553 }
 554
 555 int
 556 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 557                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 558 {
 559         cfs_list_t              *e1;
 560         cfs_list_t              *e2;
 561         lnet_remotenet_t        *rnet;
 562         lnet_route_t            *route;
 563         int                     cpt;
 564         int                     i;
 565         cfs_list_t              *rn_list;
 566
 567         cpt = lnet_net_lock_current();
 568
 569         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 570                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 571                 cfs_list_for_each(e1, rn_list) {
 572                         rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 573
 574                         cfs_list_for_each(e2, &rnet->lrn_routes) {
 575                                 route = cfs_list_entry(e2, lnet_route_t,
 576                                                        lr_list);
 577
 578                                 if (idx-- == 0) {
 579                                         *net      = rnet->lrn_net;
 580                                         *hops     = route->lr_hops;
 581                                         *priority = route->lr_priority;
 582                                         *gateway  = route->lr_gateway->lp_nid;
 583                                         *alive    = route->lr_gateway->lp_alive;
 584                                         lnet_net_unlock(cpt);
 585                                         return 0;
 586                                 }
 587                         }
 588                 }
 589         }
 590
 591         lnet_net_unlock(cpt);
 592         return -ENOENT;
 593 }
 594
 595 void
 596 lnet_swap_pinginfo(lnet_ping_info_t *info)
 597 {
 598         int               i;
 599         lnet_ni_status_t *stat;
 600
 601         __swab32s(&info->pi_magic);
 602         __swab32s(&info->pi_features);
 603         __swab32s(&info->pi_pid);
 604         __swab32s(&info->pi_nnis);
 605         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 606                 stat = &info->pi_ni[i];
 607                 __swab64s(&stat->ns_nid);
 608                 __swab32s(&stat->ns_status);
 609         }
 610         return;
 611 }
 612
 613 /**
 614  * parse router-checker pinginfo, record number of down NIs for remote
 615  * networks on that router.
 616  */
 617 static void
 618 lnet_parse_rc_info(lnet_rc_data_t *rcd)
 619 {
 620         lnet_ping_info_t        *info = rcd->rcd_pinginfo;
 621         struct lnet_peer        *gw   = rcd->rcd_gateway;
 622         lnet_route_t            *rtr;
 623
 624         if (!gw->lp_alive)
 625                 return;
 626
 627         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 628                 lnet_swap_pinginfo(info);
 629
 630         /* NB always racing with network! */
 631         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 632                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 633                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
 634                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 635                 return;
 636         }
 637
 638         gw->lp_ping_feats = info->pi_features;
 639         if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
 640                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
 641                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
 642                 return; /* nothing I can understand */
 643         }
 644
 645         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
 646                 return; /* can't carry NI status info */
 647
 648         cfs_list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) {
 649                 int     down = 0;
 650                 int     up = 0;
 651                 int     i;
 652
 653                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 654                         lnet_ni_status_t *stat = &info->pi_ni[i];
 655                         lnet_nid_t       nid = stat->ns_nid;
 656
 657                         if (nid == LNET_NID_ANY) {
 658                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 659                                        libcfs_nid2str(gw->lp_nid));
 660                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 661                                 return;
 662                         }
 663
 664                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 665                                 continue;
 666
 667                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 668                                 down++;
 669                                 continue;
 670                         }
 671
 672                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 673                                 if (LNET_NIDNET(nid) == rtr->lr_net) {
 674                                         up = 1;
 675                                         break;
 676                                 }
 677                                 continue;
 678                         }
 679
 680                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 681                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
 682                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 683                         return;
 684                 }
 685
 686                 if (up) { /* ignore downed NIs if NI for dest network is up */
 687                         rtr->lr_downis = 0;
 688                         continue;
 689                 }
 690                 rtr->lr_downis = down;
 691         }
 692 }
 693
 694 static void
 695 lnet_router_checker_event(lnet_event_t *event)
 696 {
 697         lnet_rc_data_t          *rcd = event->md.user_ptr;
 698         struct lnet_peer        *lp;
 699
 700         LASSERT(rcd != NULL);
 701
 702         if (event->unlinked) {
 703                 LNetInvalidateHandle(&rcd->rcd_mdh);
 704                 return;
 705         }
 706
 707         LASSERT(event->type == LNET_EVENT_SEND ||
 708                 event->type == LNET_EVENT_REPLY);
 709
 710         lp = rcd->rcd_gateway;
 711         LASSERT(lp != NULL);
 712
 713          /* NB: it's called with holding lnet_res_lock, we have a few
 714           * places need to hold both locks at the same time, please take
 715           * care of lock ordering */
 716         lnet_net_lock(lp->lp_cpt);
 717         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
 718                 /* ignore if no longer a router or rcd is replaced */
 719                 goto out;
 720         }
 721
 722         if (event->type == LNET_EVENT_SEND) {
 723                 lp->lp_ping_notsent = 0;
 724                 if (event->status == 0)
 725                         goto out;
 726         }
 727
 728         /* LNET_EVENT_REPLY */
 729         /* A successful REPLY means the router is up.  If _any_ comms
 730          * to the router fail I assume it's down (this will happen if
 731          * we ping alive routers to try to detect router death before
 732          * apps get burned). */
 733
 734         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
 735         /* The router checker will wake up very shortly and do the
 736          * actual notification.
 737          * XXX If 'lp' stops being a router before then, it will still
 738          * have the notification pending!!! */
 739
 740         if (avoid_asym_router_failure && event->status == 0)
 741                 lnet_parse_rc_info(rcd);
 742
 743  out:
 744         lnet_net_unlock(lp->lp_cpt);
 745 }
 746
 747 void
 748 lnet_wait_known_routerstate(void)
 749 {
 750         lnet_peer_t         *rtr;
 751         cfs_list_t          *entry;
 752         int                  all_known;
 753
 754         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 755
 756         for (;;) {
 757                 int     cpt = lnet_net_lock_current();
 758
 759                 all_known = 1;
 760                 cfs_list_for_each (entry, &the_lnet.ln_routers) {
 761                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
 762
 763                         if (rtr->lp_alive_count == 0) {
 764                                 all_known = 0;
 765                                 break;
 766                         }
 767                 }
 768
 769                 lnet_net_unlock(cpt);
 770
 771                 if (all_known)
 772                         return;
 773
 774 #ifndef __KERNEL__
 775                 lnet_router_checker();
 776 #endif
 777                 cfs_pause(cfs_time_seconds(1));
 778         }
 779 }
 780
 781 void
 782 lnet_update_ni_status_locked(void)
 783 {
 784         lnet_ni_t       *ni;
 785         long            now;
 786         int             timeout;
 787
 788         LASSERT(the_lnet.ln_routing);
 789
 790         timeout = router_ping_timeout +
 791                   MAX(live_router_check_interval, dead_router_check_interval);
 792
 793         now = cfs_time_current_sec();
 794         cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 795                 if (ni->ni_lnd->lnd_type == LOLND)
 796                         continue;
 797
 798                 if (now < ni->ni_last_alive + timeout)
 799                         continue;
 800
 801                 lnet_ni_lock(ni);
 802                 /* re-check with lock */
 803                 if (now < ni->ni_last_alive + timeout) {
 804                         lnet_ni_unlock(ni);
 805                         continue;
 806                 }
 807
 808                 LASSERT(ni->ni_status != NULL);
 809
 810                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 811                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 812                                libcfs_nid2str(ni->ni_nid), timeout);
 813                         /* NB: so far, this is the only place to set
 814                          * NI status to "down" */
 815                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 816                 }
 817                 lnet_ni_unlock(ni);
 818         }
 819 }
 820
 821 void
 822 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
 823 {
 824         LASSERT(cfs_list_empty(&rcd->rcd_list));
 825         /* detached from network */
 826         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 827
 828         if (rcd->rcd_gateway != NULL) {
 829                 int cpt = rcd->rcd_gateway->lp_cpt;
 830
 831                 lnet_net_lock(cpt);
 832                 lnet_peer_decref_locked(rcd->rcd_gateway);
 833                 lnet_net_unlock(cpt);
 834         }
 835
 836         if (rcd->rcd_pinginfo != NULL)
 837                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 838
 839         LIBCFS_FREE(rcd, sizeof(*rcd));
 840 }
 841
 842 lnet_rc_data_t *
 843 lnet_create_rc_data_locked(lnet_peer_t *gateway)
 844 {
 845         lnet_rc_data_t          *rcd = NULL;
 846         lnet_ping_info_t        *pi;
 847         int                     rc;
 848         int                     i;
 849
 850         lnet_net_unlock(gateway->lp_cpt);
 851
 852         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 853         if (rcd == NULL)
 854                 goto out;
 855
 856         LNetInvalidateHandle(&rcd->rcd_mdh);
 857         CFS_INIT_LIST_HEAD(&rcd->rcd_list);
 858
 859         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 860         if (pi == NULL)
 861                 goto out;
 862
 863         memset(pi, 0, LNET_PINGINFO_SIZE);
 864         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 865                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 866                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 867         }
 868         rcd->rcd_pinginfo = pi;
 869
 870         LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 871         rc = LNetMDBind((lnet_md_t){.start     = pi,
 872                                     .user_ptr  = rcd,
 873                                     .length    = LNET_PINGINFO_SIZE,
 874                                     .threshold = LNET_MD_THRESH_INF,
 875                                     .options   = LNET_MD_TRUNCATE,
 876                                     .eq_handle = the_lnet.ln_rc_eqh},
 877                         LNET_UNLINK,
 878                         &rcd->rcd_mdh);
 879         if (rc < 0) {
 880                 CERROR("Can't bind MD: %d\n", rc);
 881                 goto out;
 882         }
 883         LASSERT(rc == 0);
 884
 885         lnet_net_lock(gateway->lp_cpt);
 886         /* router table changed or someone has created rcd for this gateway */
 887         if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
 888                 lnet_net_unlock(gateway->lp_cpt);
 889                 goto out;
 890         }
 891
 892         lnet_peer_addref_locked(gateway);
 893         rcd->rcd_gateway = gateway;
 894         gateway->lp_rcd = rcd;
 895         gateway->lp_ping_notsent = 0;
 896
 897         return rcd;
 898
 899  out:
 900         if (rcd != NULL) {
 901                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
 902                         rc = LNetMDUnlink(rcd->rcd_mdh);
 903                         LASSERT(rc == 0);
 904                 }
 905                 lnet_destroy_rc_data(rcd);
 906         }
 907
 908         lnet_net_lock(gateway->lp_cpt);
 909         return gateway->lp_rcd;
 910 }
 911
 912 static int
 913 lnet_router_check_interval (lnet_peer_t *rtr)
 914 {
 915         int secs;
 916
 917         secs = rtr->lp_alive ? live_router_check_interval :
 918                                dead_router_check_interval;
 919         if (secs < 0)
 920                 secs = 0;
 921
 922         return secs;
 923 }
 924
 925 static void
 926 lnet_ping_router_locked (lnet_peer_t *rtr)
 927 {
 928         lnet_rc_data_t *rcd = NULL;
 929         cfs_time_t      now = cfs_time_current();
 930         int             secs;
 931
 932         lnet_peer_addref_locked(rtr);
 933
 934         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
 935             cfs_time_after(now, rtr->lp_ping_deadline))
 936                 lnet_notify_locked(rtr, 1, 0, now);
 937
 938         /* Run any outstanding notifications */
 939         lnet_ni_notify_locked(rtr->lp_ni, rtr);
 940
 941         if (!lnet_isrouter(rtr) ||
 942             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
 943                 /* router table changed or router checker is shutting down */
 944                 lnet_peer_decref_locked(rtr);
 945                 return;
 946         }
 947
 948         rcd = rtr->lp_rcd != NULL ?
 949               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
 950
 951         if (rcd == NULL)
 952                 return;
 953
 954         secs = lnet_router_check_interval(rtr);
 955
 956         CDEBUG(D_NET,
 957                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
 958                "alive_count %d lp_ping_timestamp %lu\n",
 959                libcfs_nid2str(rtr->lp_nid), secs,
 960                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
 961                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
 962
 963         if (secs != 0 && !rtr->lp_ping_notsent &&
 964             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
 965                                              cfs_time_seconds(secs)))) {
 966                 int               rc;
 967                 lnet_process_id_t id;
 968                 lnet_handle_md_t  mdh;
 969
 970                 id.nid = rtr->lp_nid;
 971                 id.pid = LUSTRE_SRV_LNET_PID;
 972                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
 973
 974                 rtr->lp_ping_notsent   = 1;
 975                 rtr->lp_ping_timestamp = now;
 976
 977                 mdh = rcd->rcd_mdh;
 978
 979                 if (rtr->lp_ping_deadline == 0) {
 980                         rtr->lp_ping_deadline =
 981                                 cfs_time_shift(router_ping_timeout);
 982                 }
 983
 984                 lnet_net_unlock(rtr->lp_cpt);
 985
 986                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
 987                              LNET_PROTO_PING_MATCHBITS, 0);
 988
 989                 lnet_net_lock(rtr->lp_cpt);
 990                 if (rc != 0)
 991                         rtr->lp_ping_notsent = 0; /* no event pending */
 992         }
 993
 994         lnet_peer_decref_locked(rtr);
 995         return;
 996 }
 997
 998 int
 999 lnet_router_checker_start(void)
1000 {
1001         int          rc;
1002         int          eqsz;
1003 #ifndef __KERNEL__
1004         lnet_peer_t *rtr;
1005         __u64        version;
1006         int          nrtr = 0;
1007         int          router_checker_max_eqsize = 10240;
1008
1009         LASSERT (check_routers_before_use);
1010         LASSERT (dead_router_check_interval > 0);
1011
1012         lnet_net_lock(0);
1013
1014         /* As an approximation, allow each router the same number of
1015          * outstanding events as it is allowed outstanding sends */
1016         eqsz = 0;
1017         version = the_lnet.ln_routers_version;
1018         cfs_list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1019                 lnet_ni_t         *ni = rtr->lp_ni;
1020                 lnet_process_id_t  id;
1021
1022                 nrtr++;
1023                 eqsz += ni->ni_peertxcredits;
1024
1025                 /* one async ping reply per router */
1026                 id.nid = rtr->lp_nid;
1027                 id.pid = LUSTRE_SRV_LNET_PID;
1028
1029                 lnet_net_unlock(0);
1030
1031                 rc = LNetSetAsync(id, 1);
1032                 if (rc != 0) {
1033                         CWARN("LNetSetAsync %s failed: %d\n",
1034                               libcfs_id2str(id), rc);
1035                         return rc;
1036                 }
1037
1038                 lnet_net_lock(0);
1039                 /* NB router list doesn't change in userspace */
1040                 LASSERT(version == the_lnet.ln_routers_version);
1041         }
1042
1043         lnet_net_unlock(0);
1044
1045         if (nrtr == 0) {
1046                 CDEBUG(D_NET,
1047                        "No router found, not starting router checker\n");
1048                 return 0;
1049         }
1050
1051         /* at least allow a SENT and a REPLY per router */
1052         if (router_checker_max_eqsize < 2 * nrtr)
1053                 router_checker_max_eqsize = 2 * nrtr;
1054
1055         LASSERT (eqsz > 0);
1056         if (eqsz > router_checker_max_eqsize)
1057                 eqsz = router_checker_max_eqsize;
1058 #endif
1059
1060         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1061
1062         if (check_routers_before_use &&
1063             dead_router_check_interval <= 0) {
1064                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1065                                    " set if 'check_routers_before_use' is set"
1066                                    "\n");
1067                 return -EINVAL;
1068         }
1069
1070         if (!the_lnet.ln_routing &&
1071             live_router_check_interval <= 0 &&
1072             dead_router_check_interval <= 0)
1073                 return 0;
1074
1075 #ifdef __KERNEL__
1076         sema_init(&the_lnet.ln_rc_signal, 0);
1077         /* EQ size doesn't matter; the callback is guaranteed to get every
1078          * event */
1079         eqsz = 0;
1080         rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1081                          &the_lnet.ln_rc_eqh);
1082 #else
1083         rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE,
1084                          &the_lnet.ln_rc_eqh);
1085 #endif
1086         if (rc != 0) {
1087                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1088                 return -ENOMEM;
1089         }
1090
1091         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1092 #ifdef __KERNEL__
1093         rc = PTR_ERR(kthread_run(lnet_router_checker,
1094                                  NULL, "router_checker"));
1095         if (IS_ERR_VALUE(rc)) {
1096                 CERROR("Can't start router checker thread: %d\n", rc);
1097                 /* block until event callback signals exit */
1098                 down(&the_lnet.ln_rc_signal);
1099                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1100                 LASSERT(rc == 0);
1101                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1102                 return -ENOMEM;
1103         }
1104 #endif
1105
1106         if (check_routers_before_use) {
1107                 /* Note that a helpful side-effect of pinging all known routers
1108                  * at startup is that it makes them drop stale connections they
1109                  * may have to a previous instance of me. */
1110                 lnet_wait_known_routerstate();
1111         }
1112
1113         return 0;
1114 }
1115
1116 void
1117 lnet_router_checker_stop (void)
1118 {
1119         int rc;
1120
1121         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1122                 return;
1123
1124         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1125         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1126
1127 #ifdef __KERNEL__
1128         /* block until event callback signals exit */
1129         down(&the_lnet.ln_rc_signal);
1130 #else
1131         lnet_router_checker();
1132 #endif
1133         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1134
1135         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1136         LASSERT (rc == 0);
1137         return;
1138 }
1139
1140 static void
1141 lnet_prune_rc_data(int wait_unlink)
1142 {
1143         lnet_rc_data_t          *rcd;
1144         lnet_rc_data_t          *tmp;
1145         lnet_peer_t             *lp;
1146         cfs_list_t              head;
1147         int                     i = 2;
1148
1149         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1150                    cfs_list_empty(&the_lnet.ln_rcd_deathrow) &&
1151                    cfs_list_empty(&the_lnet.ln_rcd_zombie)))
1152                 return;
1153
1154         CFS_INIT_LIST_HEAD(&head);
1155
1156         lnet_net_lock(LNET_LOCK_EX);
1157
1158         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1159                 /* router checker is stopping, prune all */
1160                 cfs_list_for_each_entry(lp, &the_lnet.ln_routers,
1161                                         lp_rtr_list) {
1162                         if (lp->lp_rcd == NULL)
1163                                 continue;
1164
1165                         LASSERT(cfs_list_empty(&lp->lp_rcd->rcd_list));
1166                         cfs_list_add(&lp->lp_rcd->rcd_list,
1167                                      &the_lnet.ln_rcd_deathrow);
1168                         lp->lp_rcd = NULL;
1169                 }
1170         }
1171
1172         /* unlink all RCDs on deathrow list */
1173         cfs_list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1174
1175         if (!cfs_list_empty(&head)) {
1176                 lnet_net_unlock(LNET_LOCK_EX);
1177
1178                 cfs_list_for_each_entry(rcd, &head, rcd_list)
1179                         LNetMDUnlink(rcd->rcd_mdh);
1180
1181                 lnet_net_lock(LNET_LOCK_EX);
1182         }
1183
1184         cfs_list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1185
1186         /* release all zombie RCDs */
1187         while (!cfs_list_empty(&the_lnet.ln_rcd_zombie)) {
1188                 cfs_list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1189                                              rcd_list) {
1190                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1191                                 cfs_list_move(&rcd->rcd_list, &head);
1192                 }
1193
1194                 wait_unlink = wait_unlink &&
1195                               !cfs_list_empty(&the_lnet.ln_rcd_zombie);
1196
1197                 lnet_net_unlock(LNET_LOCK_EX);
1198
1199                 while (!cfs_list_empty(&head)) {
1200                         rcd = cfs_list_entry(head.next,
1201                                              lnet_rc_data_t, rcd_list);
1202                         cfs_list_del_init(&rcd->rcd_list);
1203                         lnet_destroy_rc_data(rcd);
1204                 }
1205
1206                 if (!wait_unlink)
1207                         return;
1208
1209                 i++;
1210                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1211                        "Waiting for rc buffers to unlink\n");
1212                 cfs_pause(cfs_time_seconds(1) / 4);
1213
1214                 lnet_net_lock(LNET_LOCK_EX);
1215         }
1216
1217         lnet_net_unlock(LNET_LOCK_EX);
1218 }
1219
1220
1221 #if defined(__KERNEL__) && defined(LNET_ROUTER)
1222
1223 static int
1224 lnet_router_checker(void *arg)
1225 {
1226         lnet_peer_t       *rtr;
1227         cfs_list_t        *entry;
1228
1229         cfs_block_allsigs();
1230
1231         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1232
1233         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1234                 __u64   version;
1235                 int     cpt;
1236                 int     cpt2;
1237
1238                 cpt = lnet_net_lock_current();
1239 rescan:
1240                 version = the_lnet.ln_routers_version;
1241
1242                 cfs_list_for_each(entry, &the_lnet.ln_routers) {
1243                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
1244
1245                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1246                         if (cpt != cpt2) {
1247                                 lnet_net_unlock(cpt);
1248                                 cpt = cpt2;
1249                                 lnet_net_lock(cpt);
1250                                 /* the routers list has changed */
1251                                 if (version != the_lnet.ln_routers_version)
1252                                         goto rescan;
1253                         }
1254
1255                         lnet_ping_router_locked(rtr);
1256
1257                         /* NB dropped lock */
1258                         if (version != the_lnet.ln_routers_version) {
1259                                 /* the routers list has changed */
1260                                 goto rescan;
1261                         }
1262                 }
1263
1264                 if (the_lnet.ln_routing)
1265                         lnet_update_ni_status_locked();
1266
1267                 lnet_net_unlock(cpt);
1268
1269                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1270
1271                 /* Call cfs_pause() here always adds 1 to load average
1272                  * because kernel counts # active tasks as nr_running
1273                  * + nr_uninterruptible. */
1274                 schedule_timeout_and_set_state(TASK_INTERRUPTIBLE,
1275                                                    cfs_time_seconds(1));
1276         }
1277
1278         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1279
1280         lnet_prune_rc_data(1); /* wait for UNLINK */
1281
1282         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1283         up(&the_lnet.ln_rc_signal);
1284         /* The unlink event callback will signal final completion */
1285         return 0;
1286 }
1287
1288 void
1289 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1290 {
1291         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1292
1293         while (--npages >= 0)
1294                 __free_page(rb->rb_kiov[npages].kiov_page);
1295
1296         LIBCFS_FREE(rb, sz);
1297 }
1298
1299 lnet_rtrbuf_t *
1300 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1301 {
1302         int            npages = rbp->rbp_npages;
1303         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1304         struct page   *page;
1305         lnet_rtrbuf_t *rb;
1306         int            i;
1307
1308         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1309         if (rb == NULL)
1310                 return NULL;
1311
1312         rb->rb_pool = rbp;
1313
1314         for (i = 0; i < npages; i++) {
1315                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1316                                           __GFP_ZERO | GFP_IOFS);
1317                 if (page == NULL) {
1318                         while (--i >= 0)
1319                                 __free_page(rb->rb_kiov[i].kiov_page);
1320
1321                         LIBCFS_FREE(rb, sz);
1322                         return NULL;
1323                 }
1324
1325                 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1326                 rb->rb_kiov[i].kiov_offset = 0;
1327                 rb->rb_kiov[i].kiov_page = page;
1328         }
1329
1330         return rb;
1331 }
1332
1333 void
1334 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
1335 {
1336         int             npages = rbp->rbp_npages;
1337         int             nbuffers = 0;
1338         lnet_rtrbuf_t   *rb;
1339
1340         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1341                 return;
1342
1343         LASSERT (cfs_list_empty(&rbp->rbp_msgs));
1344         LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
1345
1346         while (!cfs_list_empty(&rbp->rbp_bufs)) {
1347                 LASSERT (rbp->rbp_credits > 0);
1348
1349                 rb = cfs_list_entry(rbp->rbp_bufs.next,
1350                                     lnet_rtrbuf_t, rb_list);
1351                 cfs_list_del(&rb->rb_list);
1352                 lnet_destroy_rtrbuf(rb, npages);
1353                 nbuffers++;
1354         }
1355
1356         LASSERT (rbp->rbp_nbuffers == nbuffers);
1357         LASSERT (rbp->rbp_credits == nbuffers);
1358
1359         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1360 }
1361
1362 int
1363 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1364 {
1365         lnet_rtrbuf_t *rb;
1366         int            i;
1367
1368         if (rbp->rbp_nbuffers != 0) {
1369                 LASSERT (rbp->rbp_nbuffers == nbufs);
1370                 return 0;
1371         }
1372
1373         for (i = 0; i < nbufs; i++) {
1374                 rb = lnet_new_rtrbuf(rbp, cpt);
1375
1376                 if (rb == NULL) {
1377                         CERROR("Failed to allocate %d router bufs of %d pages\n",
1378                                nbufs, rbp->rbp_npages);
1379                         return -ENOMEM;
1380                 }
1381
1382                 rbp->rbp_nbuffers++;
1383                 rbp->rbp_credits++;
1384                 rbp->rbp_mincredits++;
1385                 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1386
1387                 /* No allocation "under fire" */
1388                 /* Otherwise we'd need code to schedule blocked msgs etc */
1389                 LASSERT (!the_lnet.ln_routing);
1390         }
1391
1392         LASSERT (rbp->rbp_credits == nbufs);
1393         return 0;
1394 }
1395
1396 void
1397 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1398 {
1399         CFS_INIT_LIST_HEAD(&rbp->rbp_msgs);
1400         CFS_INIT_LIST_HEAD(&rbp->rbp_bufs);
1401
1402         rbp->rbp_npages = npages;
1403         rbp->rbp_credits = 0;
1404         rbp->rbp_mincredits = 0;
1405 }
1406
1407 void
1408 lnet_rtrpools_free(void)
1409 {
1410         lnet_rtrbufpool_t *rtrp;
1411         int               i;
1412
1413         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1414                 return;
1415
1416         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1417                 lnet_rtrpool_free_bufs(&rtrp[0]);
1418                 lnet_rtrpool_free_bufs(&rtrp[1]);
1419                 lnet_rtrpool_free_bufs(&rtrp[2]);
1420         }
1421
1422         cfs_percpt_free(the_lnet.ln_rtrpools);
1423         the_lnet.ln_rtrpools = NULL;
1424 }
1425
1426 static int
1427 lnet_nrb_tiny_calculate(int npages)
1428 {
1429         int     nrbs = LNET_NRB_TINY;
1430
1431         if (tiny_router_buffers < 0) {
1432                 LCONSOLE_ERROR_MSG(0x10c,
1433                                    "tiny_router_buffers=%d invalid when "
1434                                    "routing enabled\n", tiny_router_buffers);
1435                 return -1;
1436         }
1437
1438         if (tiny_router_buffers > 0)
1439                 nrbs = tiny_router_buffers;
1440
1441         nrbs /= LNET_CPT_NUMBER;
1442         return max(nrbs, LNET_NRB_TINY_MIN);
1443 }
1444
1445 static int
1446 lnet_nrb_small_calculate(int npages)
1447 {
1448         int     nrbs = LNET_NRB_SMALL;
1449
1450         if (small_router_buffers < 0) {
1451                 LCONSOLE_ERROR_MSG(0x10c,
1452                                    "small_router_buffers=%d invalid when "
1453                                    "routing enabled\n", small_router_buffers);
1454                 return -1;
1455         }
1456
1457         if (small_router_buffers > 0)
1458                 nrbs = small_router_buffers;
1459
1460         nrbs /= LNET_CPT_NUMBER;
1461         return max(nrbs, LNET_NRB_SMALL_MIN);
1462 }
1463
1464 static int
1465 lnet_nrb_large_calculate(int npages)
1466 {
1467         int     nrbs = LNET_NRB_LARGE;
1468
1469         if (large_router_buffers < 0) {
1470                 LCONSOLE_ERROR_MSG(0x10c,
1471                                    "large_router_buffers=%d invalid when "
1472                                    "routing enabled\n", large_router_buffers);
1473                 return -1;
1474         }
1475
1476         if (large_router_buffers > 0)
1477                 nrbs = large_router_buffers;
1478
1479         nrbs /= LNET_CPT_NUMBER;
1480         return max(nrbs, LNET_NRB_LARGE_MIN);
1481 }
1482
1483 int
1484 lnet_rtrpools_alloc(int im_a_router)
1485 {
1486         lnet_rtrbufpool_t *rtrp;
1487         int     large_pages = (LNET_MTU + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1488         int     small_pages = 1;
1489         int     nrb_tiny;
1490         int     nrb_small;
1491         int     nrb_large;
1492         int     rc;
1493         int     i;
1494
1495         if (!strcmp(forwarding, "")) {
1496                 /* not set either way */
1497                 if (!im_a_router)
1498                         return 0;
1499         } else if (!strcmp(forwarding, "disabled")) {
1500                 /* explicitly disabled */
1501                 return 0;
1502         } else if (!strcmp(forwarding, "enabled")) {
1503                 /* explicitly enabled */
1504         } else {
1505                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1506                                    "'enabled' or 'disabled'\n");
1507                 return -EINVAL;
1508         }
1509
1510         nrb_tiny = lnet_nrb_tiny_calculate(0);
1511         if (nrb_tiny < 0)
1512                 return -EINVAL;
1513
1514         nrb_small = lnet_nrb_small_calculate(small_pages);
1515         if (nrb_small < 0)
1516                 return -EINVAL;
1517
1518         nrb_large = lnet_nrb_large_calculate(large_pages);
1519         if (nrb_large < 0)
1520                 return -EINVAL;
1521
1522         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1523                                                 LNET_NRBPOOLS *
1524                                                 sizeof(lnet_rtrbufpool_t));
1525         if (the_lnet.ln_rtrpools == NULL) {
1526                 LCONSOLE_ERROR_MSG(0x10c,
1527                                    "Failed to initialize router buffe pool\n");
1528                 return -ENOMEM;
1529         }
1530
1531         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1532                 lnet_rtrpool_init(&rtrp[0], 0);
1533                 rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i);
1534                 if (rc != 0)
1535                         goto failed;
1536
1537                 lnet_rtrpool_init(&rtrp[1], small_pages);
1538                 rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i);
1539                 if (rc != 0)
1540                         goto failed;
1541
1542                 lnet_rtrpool_init(&rtrp[2], large_pages);
1543                 rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i);
1544                 if (rc != 0)
1545                         goto failed;
1546         }
1547
1548         lnet_net_lock(LNET_LOCK_EX);
1549         the_lnet.ln_routing = 1;
1550         lnet_net_unlock(LNET_LOCK_EX);
1551
1552         return 0;
1553
1554  failed:
1555         lnet_rtrpools_free();
1556         return rc;
1557 }
1558
1559 int
1560 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1561 {
1562         struct lnet_peer        *lp = NULL;
1563         cfs_time_t              now = cfs_time_current();
1564         int                     cpt = lnet_cpt_of_nid(nid);
1565
1566         LASSERT (!in_interrupt ());
1567
1568         CDEBUG (D_NET, "%s notifying %s: %s\n",
1569                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1570                 libcfs_nid2str(nid),
1571                 alive ? "up" : "down");
1572
1573         if (ni != NULL &&
1574             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1575                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1576                         libcfs_nid2str(nid), alive ? "birth" : "death",
1577                         libcfs_nid2str(ni->ni_nid));
1578                 return -EINVAL;
1579         }
1580
1581         /* can't do predictions... */
1582         if (cfs_time_after(when, now)) {
1583                 CWARN ("Ignoring prediction from %s of %s %s "
1584                        "%ld seconds in the future\n",
1585                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1586                        libcfs_nid2str(nid), alive ? "up" : "down",
1587                        cfs_duration_sec(cfs_time_sub(when, now)));
1588                 return -EINVAL;
1589         }
1590
1591         if (ni != NULL && !alive &&             /* LND telling me she's down */
1592             !auto_down) {                       /* auto-down disabled */
1593                 CDEBUG(D_NET, "Auto-down disabled\n");
1594                 return 0;
1595         }
1596
1597         lnet_net_lock(cpt);
1598
1599         if (the_lnet.ln_shutdown) {
1600                 lnet_net_unlock(cpt);
1601                 return -ESHUTDOWN;
1602         }
1603
1604         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1605         if (lp == NULL) {
1606                 /* nid not found */
1607                 lnet_net_unlock(cpt);
1608                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1609                 return 0;
1610         }
1611
1612         /* We can't fully trust LND on reporting exact peer last_alive
1613          * if he notifies us about dead peer. For example ksocklnd can
1614          * call us with when == _time_when_the_node_was_booted_ if
1615          * no connections were successfully established */
1616         if (ni != NULL && !alive && when < lp->lp_last_alive)
1617                 when = lp->lp_last_alive;
1618
1619         lnet_notify_locked(lp, ni == NULL, alive, when);
1620
1621         if (ni != NULL)
1622                 lnet_ni_notify_locked(ni, lp);
1623
1624         lnet_peer_decref_locked(lp);
1625
1626         lnet_net_unlock(cpt);
1627         return 0;
1628 }
1629 EXPORT_SYMBOL(lnet_notify);
1630
1631 void
1632 lnet_get_tunables (void)
1633 {
1634         return;
1635 }
1636
1637 #else
1638
1639 int
1640 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1641 {
1642         return -EOPNOTSUPP;
1643 }
1644
1645 void
1646 lnet_router_checker (void)
1647 {
1648         static time_t last = 0;
1649         static int    running = 0;
1650
1651         time_t            now = cfs_time_current_sec();
1652         int               interval = now - last;
1653         int               rc;
1654         __u64             version;
1655         lnet_peer_t      *rtr;
1656
1657         /* It's no use to call me again within a sec - all intervals and
1658          * timeouts are measured in seconds */
1659         if (last != 0 && interval < 2)
1660                 return;
1661
1662         if (last != 0 &&
1663             interval > MAX(live_router_check_interval,
1664                            dead_router_check_interval))
1665                 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1666                         live_router_check_interval, dead_router_check_interval,
1667                         interval);
1668
1669         LASSERT(LNET_CPT_NUMBER == 1);
1670
1671         lnet_net_lock(0);
1672         LASSERT(!running); /* recursion check */
1673         running = 1;
1674         lnet_net_unlock(0);
1675
1676         last = now;
1677
1678         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1679                 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1680
1681         /* consume all pending events */
1682         while (1) {
1683                 int          i;
1684                 lnet_event_t ev;
1685
1686                 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1687                  * recursion breaker in LNetEQPoll would fail */
1688                 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1689                 if (rc == 0)   /* no event pending */
1690                         break;
1691
1692                 /* NB a lost SENT prevents me from pinging a router again */
1693                 if (rc == -EOVERFLOW) {
1694                         CERROR("Dropped an event!!!\n");
1695                         abort();
1696                 }
1697
1698                 LASSERT (rc == 1);
1699
1700                 lnet_router_checker_event(&ev);
1701         }
1702
1703         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1704                 lnet_prune_rc_data(1); /* release rcd */
1705                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1706                 running = 0;
1707                 return;
1708         }
1709
1710         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1711
1712         lnet_net_lock(0);
1713
1714         version = the_lnet.ln_routers_version;
1715         cfs_list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) {
1716                 lnet_ping_router_locked(rtr);
1717                 LASSERT (version == the_lnet.ln_routers_version);
1718         }
1719
1720         lnet_net_unlock(0);
1721
1722         running = 0; /* lock only needed for the recursion check */
1723         return;
1724 }
1725
1726 /* NB lnet_peers_start_down depends on me,
1727  * so must be called before any peer creation */
1728 void
1729 lnet_get_tunables (void)
1730 {
1731         char *s;
1732
1733         s = getenv("LNET_ROUTER_PING_TIMEOUT");
1734         if (s != NULL) router_ping_timeout = atoi(s);
1735
1736         s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1737         if (s != NULL) live_router_check_interval = atoi(s);
1738
1739         s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1740         if (s != NULL) dead_router_check_interval = atoi(s);
1741
1742         /* This replaces old lnd_notify mechanism */
1743         check_routers_before_use = 1;
1744         if (dead_router_check_interval <= 0)
1745                 dead_router_check_interval = 30;
1746 }
1747
1748 void
1749 lnet_rtrpools_free(void)
1750 {
1751 }
1752
1753 int
1754 lnet_rtrpools_alloc(int im_a_arouter)
1755 {
1756         return 0;
1757 }
1758
1759 #endif