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     ptl_status = LNET_NI_STATUS_INVALID;
 650                 int     down = 0;
 651                 int     up = 0;
 652                 int     i;
 653
 654                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 655                         lnet_ni_status_t *stat = &info->pi_ni[i];
 656                         lnet_nid_t       nid = stat->ns_nid;
 657
 658                         if (nid == LNET_NID_ANY) {
 659                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 660                                        libcfs_nid2str(gw->lp_nid));
 661                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 662                                 return;
 663                         }
 664
 665                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 666                                 continue;
 667
 668                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 669                                 if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND)
 670                                         down++;
 671                                 else if (ptl_status != LNET_NI_STATUS_UP)
 672                                         ptl_status = LNET_NI_STATUS_DOWN;
 673                                 continue;
 674                         }
 675
 676                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 677                                 if (LNET_NIDNET(nid) == rtr->lr_net) {
 678                                         up = 1;
 679                                         break;
 680                                 }
 681                                 /* ptl NIs are considered down only when
 682                                  * they're all down */
 683                                 if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
 684                                         ptl_status = LNET_NI_STATUS_UP;
 685                                 continue;
 686                         }
 687
 688                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 689                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
 690                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 691                         return;
 692                 }
 693
 694                 if (up) { /* ignore downed NIs if NI for dest network is up */
 695                         rtr->lr_downis = 0;
 696                         continue;
 697                 }
 698                 rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN);
 699         }
 700 }
 701
 702 static void
 703 lnet_router_checker_event(lnet_event_t *event)
 704 {
 705         lnet_rc_data_t          *rcd = event->md.user_ptr;
 706         struct lnet_peer        *lp;
 707
 708         LASSERT(rcd != NULL);
 709
 710         if (event->unlinked) {
 711                 LNetInvalidateHandle(&rcd->rcd_mdh);
 712                 return;
 713         }
 714
 715         LASSERT(event->type == LNET_EVENT_SEND ||
 716                 event->type == LNET_EVENT_REPLY);
 717
 718         lp = rcd->rcd_gateway;
 719         LASSERT(lp != NULL);
 720
 721          /* NB: it's called with holding lnet_res_lock, we have a few
 722           * places need to hold both locks at the same time, please take
 723           * care of lock ordering */
 724         lnet_net_lock(lp->lp_cpt);
 725         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
 726                 /* ignore if no longer a router or rcd is replaced */
 727                 goto out;
 728         }
 729
 730         if (event->type == LNET_EVENT_SEND) {
 731                 lp->lp_ping_notsent = 0;
 732                 if (event->status == 0)
 733                         goto out;
 734         }
 735
 736         /* LNET_EVENT_REPLY */
 737         /* A successful REPLY means the router is up.  If _any_ comms
 738          * to the router fail I assume it's down (this will happen if
 739          * we ping alive routers to try to detect router death before
 740          * apps get burned). */
 741
 742         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
 743         /* The router checker will wake up very shortly and do the
 744          * actual notification.
 745          * XXX If 'lp' stops being a router before then, it will still
 746          * have the notification pending!!! */
 747
 748         if (avoid_asym_router_failure && event->status == 0)
 749                 lnet_parse_rc_info(rcd);
 750
 751  out:
 752         lnet_net_unlock(lp->lp_cpt);
 753 }
 754
 755 void
 756 lnet_wait_known_routerstate(void)
 757 {
 758         lnet_peer_t         *rtr;
 759         cfs_list_t          *entry;
 760         int                  all_known;
 761
 762         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 763
 764         for (;;) {
 765                 int     cpt = lnet_net_lock_current();
 766
 767                 all_known = 1;
 768                 cfs_list_for_each (entry, &the_lnet.ln_routers) {
 769                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
 770
 771                         if (rtr->lp_alive_count == 0) {
 772                                 all_known = 0;
 773                                 break;
 774                         }
 775                 }
 776
 777                 lnet_net_unlock(cpt);
 778
 779                 if (all_known)
 780                         return;
 781
 782 #ifndef __KERNEL__
 783                 lnet_router_checker();
 784 #endif
 785                 cfs_pause(cfs_time_seconds(1));
 786         }
 787 }
 788
 789 void
 790 lnet_update_ni_status_locked(void)
 791 {
 792         lnet_ni_t       *ni;
 793         long            now;
 794         int             timeout;
 795
 796         LASSERT(the_lnet.ln_routing);
 797
 798         timeout = router_ping_timeout +
 799                   MAX(live_router_check_interval, dead_router_check_interval);
 800
 801         now = cfs_time_current_sec();
 802         cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 803                 if (ni->ni_lnd->lnd_type == LOLND)
 804                         continue;
 805
 806                 if (now < ni->ni_last_alive + timeout)
 807                         continue;
 808
 809                 lnet_ni_lock(ni);
 810                 /* re-check with lock */
 811                 if (now < ni->ni_last_alive + timeout) {
 812                         lnet_ni_unlock(ni);
 813                         continue;
 814                 }
 815
 816                 LASSERT(ni->ni_status != NULL);
 817
 818                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 819                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 820                                libcfs_nid2str(ni->ni_nid), timeout);
 821                         /* NB: so far, this is the only place to set
 822                          * NI status to "down" */
 823                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 824                 }
 825                 lnet_ni_unlock(ni);
 826         }
 827 }
 828
 829 void
 830 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
 831 {
 832         LASSERT(cfs_list_empty(&rcd->rcd_list));
 833         /* detached from network */
 834         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 835
 836         if (rcd->rcd_gateway != NULL) {
 837                 int cpt = rcd->rcd_gateway->lp_cpt;
 838
 839                 lnet_net_lock(cpt);
 840                 lnet_peer_decref_locked(rcd->rcd_gateway);
 841                 lnet_net_unlock(cpt);
 842         }
 843
 844         if (rcd->rcd_pinginfo != NULL)
 845                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 846
 847         LIBCFS_FREE(rcd, sizeof(*rcd));
 848 }
 849
 850 lnet_rc_data_t *
 851 lnet_create_rc_data_locked(lnet_peer_t *gateway)
 852 {
 853         lnet_rc_data_t          *rcd = NULL;
 854         lnet_ping_info_t        *pi;
 855         int                     rc;
 856         int                     i;
 857
 858         lnet_net_unlock(gateway->lp_cpt);
 859
 860         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 861         if (rcd == NULL)
 862                 goto out;
 863
 864         LNetInvalidateHandle(&rcd->rcd_mdh);
 865         CFS_INIT_LIST_HEAD(&rcd->rcd_list);
 866
 867         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 868         if (pi == NULL)
 869                 goto out;
 870
 871         memset(pi, 0, LNET_PINGINFO_SIZE);
 872         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 873                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 874                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 875         }
 876         rcd->rcd_pinginfo = pi;
 877
 878         LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 879         rc = LNetMDBind((lnet_md_t){.start     = pi,
 880                                     .user_ptr  = rcd,
 881                                     .length    = LNET_PINGINFO_SIZE,
 882                                     .threshold = LNET_MD_THRESH_INF,
 883                                     .options   = LNET_MD_TRUNCATE,
 884                                     .eq_handle = the_lnet.ln_rc_eqh},
 885                         LNET_UNLINK,
 886                         &rcd->rcd_mdh);
 887         if (rc < 0) {
 888                 CERROR("Can't bind MD: %d\n", rc);
 889                 goto out;
 890         }
 891         LASSERT(rc == 0);
 892
 893         lnet_net_lock(gateway->lp_cpt);
 894         /* router table changed or someone has created rcd for this gateway */
 895         if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
 896                 lnet_net_unlock(gateway->lp_cpt);
 897                 goto out;
 898         }
 899
 900         lnet_peer_addref_locked(gateway);
 901         rcd->rcd_gateway = gateway;
 902         gateway->lp_rcd = rcd;
 903         gateway->lp_ping_notsent = 0;
 904
 905         return rcd;
 906
 907  out:
 908         if (rcd != NULL) {
 909                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
 910                         rc = LNetMDUnlink(rcd->rcd_mdh);
 911                         LASSERT(rc == 0);
 912                 }
 913                 lnet_destroy_rc_data(rcd);
 914         }
 915
 916         lnet_net_lock(gateway->lp_cpt);
 917         return gateway->lp_rcd;
 918 }
 919
 920 static int
 921 lnet_router_check_interval (lnet_peer_t *rtr)
 922 {
 923         int secs;
 924
 925         secs = rtr->lp_alive ? live_router_check_interval :
 926                                dead_router_check_interval;
 927         if (secs < 0)
 928                 secs = 0;
 929
 930         return secs;
 931 }
 932
 933 static void
 934 lnet_ping_router_locked (lnet_peer_t *rtr)
 935 {
 936         lnet_rc_data_t *rcd = NULL;
 937         cfs_time_t      now = cfs_time_current();
 938         int             secs;
 939
 940         lnet_peer_addref_locked(rtr);
 941
 942         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
 943             cfs_time_after(now, rtr->lp_ping_deadline))
 944                 lnet_notify_locked(rtr, 1, 0, now);
 945
 946         /* Run any outstanding notifications */
 947         lnet_ni_notify_locked(rtr->lp_ni, rtr);
 948
 949         if (!lnet_isrouter(rtr) ||
 950             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
 951                 /* router table changed or router checker is shutting down */
 952                 lnet_peer_decref_locked(rtr);
 953                 return;
 954         }
 955
 956         rcd = rtr->lp_rcd != NULL ?
 957               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
 958
 959         if (rcd == NULL)
 960                 return;
 961
 962         secs = lnet_router_check_interval(rtr);
 963
 964         CDEBUG(D_NET,
 965                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
 966                "alive_count %d lp_ping_timestamp %lu\n",
 967                libcfs_nid2str(rtr->lp_nid), secs,
 968                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
 969                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
 970
 971         if (secs != 0 && !rtr->lp_ping_notsent &&
 972             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
 973                                              cfs_time_seconds(secs)))) {
 974                 int               rc;
 975                 lnet_process_id_t id;
 976                 lnet_handle_md_t  mdh;
 977
 978                 id.nid = rtr->lp_nid;
 979                 id.pid = LUSTRE_SRV_LNET_PID;
 980                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
 981
 982                 rtr->lp_ping_notsent   = 1;
 983                 rtr->lp_ping_timestamp = now;
 984
 985                 mdh = rcd->rcd_mdh;
 986
 987                 if (rtr->lp_ping_deadline == 0) {
 988                         rtr->lp_ping_deadline =
 989                                 cfs_time_shift(router_ping_timeout);
 990                 }
 991
 992                 lnet_net_unlock(rtr->lp_cpt);
 993
 994                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
 995                              LNET_PROTO_PING_MATCHBITS, 0);
 996
 997                 lnet_net_lock(rtr->lp_cpt);
 998                 if (rc != 0)
 999                         rtr->lp_ping_notsent = 0; /* no event pending */
1000         }
1001
1002         lnet_peer_decref_locked(rtr);
1003         return;
1004 }
1005
1006 int
1007 lnet_router_checker_start(void)
1008 {
1009         int          rc;
1010         int          eqsz;
1011 #ifndef __KERNEL__
1012         lnet_peer_t *rtr;
1013         __u64        version;
1014         int          nrtr = 0;
1015         int          router_checker_max_eqsize = 10240;
1016
1017         LASSERT (check_routers_before_use);
1018         LASSERT (dead_router_check_interval > 0);
1019
1020         lnet_net_lock(0);
1021
1022         /* As an approximation, allow each router the same number of
1023          * outstanding events as it is allowed outstanding sends */
1024         eqsz = 0;
1025         version = the_lnet.ln_routers_version;
1026         cfs_list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1027                 lnet_ni_t         *ni = rtr->lp_ni;
1028                 lnet_process_id_t  id;
1029
1030                 nrtr++;
1031                 eqsz += ni->ni_peertxcredits;
1032
1033                 /* one async ping reply per router */
1034                 id.nid = rtr->lp_nid;
1035                 id.pid = LUSTRE_SRV_LNET_PID;
1036
1037                 lnet_net_unlock(0);
1038
1039                 rc = LNetSetAsync(id, 1);
1040                 if (rc != 0) {
1041                         CWARN("LNetSetAsync %s failed: %d\n",
1042                               libcfs_id2str(id), rc);
1043                         return rc;
1044                 }
1045
1046                 lnet_net_lock(0);
1047                 /* NB router list doesn't change in userspace */
1048                 LASSERT(version == the_lnet.ln_routers_version);
1049         }
1050
1051         lnet_net_unlock(0);
1052
1053         if (nrtr == 0) {
1054                 CDEBUG(D_NET,
1055                        "No router found, not starting router checker\n");
1056                 return 0;
1057         }
1058
1059         /* at least allow a SENT and a REPLY per router */
1060         if (router_checker_max_eqsize < 2 * nrtr)
1061                 router_checker_max_eqsize = 2 * nrtr;
1062
1063         LASSERT (eqsz > 0);
1064         if (eqsz > router_checker_max_eqsize)
1065                 eqsz = router_checker_max_eqsize;
1066 #endif
1067
1068         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1069
1070         if (check_routers_before_use &&
1071             dead_router_check_interval <= 0) {
1072                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1073                                    " set if 'check_routers_before_use' is set"
1074                                    "\n");
1075                 return -EINVAL;
1076         }
1077
1078         if (!the_lnet.ln_routing &&
1079             live_router_check_interval <= 0 &&
1080             dead_router_check_interval <= 0)
1081                 return 0;
1082
1083 #ifdef __KERNEL__
1084         sema_init(&the_lnet.ln_rc_signal, 0);
1085         /* EQ size doesn't matter; the callback is guaranteed to get every
1086          * event */
1087         eqsz = 0;
1088         rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1089                          &the_lnet.ln_rc_eqh);
1090 #else
1091         rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE,
1092                          &the_lnet.ln_rc_eqh);
1093 #endif
1094         if (rc != 0) {
1095                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1096                 return -ENOMEM;
1097         }
1098
1099         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1100 #ifdef __KERNEL__
1101         rc = PTR_ERR(kthread_run(lnet_router_checker,
1102                                  NULL, "router_checker"));
1103         if (IS_ERR_VALUE(rc)) {
1104                 CERROR("Can't start router checker thread: %d\n", rc);
1105                 /* block until event callback signals exit */
1106                 down(&the_lnet.ln_rc_signal);
1107                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1108                 LASSERT(rc == 0);
1109                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1110                 return -ENOMEM;
1111         }
1112 #endif
1113
1114         if (check_routers_before_use) {
1115                 /* Note that a helpful side-effect of pinging all known routers
1116                  * at startup is that it makes them drop stale connections they
1117                  * may have to a previous instance of me. */
1118                 lnet_wait_known_routerstate();
1119         }
1120
1121         return 0;
1122 }
1123
1124 void
1125 lnet_router_checker_stop (void)
1126 {
1127         int rc;
1128
1129         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1130                 return;
1131
1132         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1133         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1134
1135 #ifdef __KERNEL__
1136         /* block until event callback signals exit */
1137         down(&the_lnet.ln_rc_signal);
1138 #else
1139         lnet_router_checker();
1140 #endif
1141         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1142
1143         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1144         LASSERT (rc == 0);
1145         return;
1146 }
1147
1148 static void
1149 lnet_prune_rc_data(int wait_unlink)
1150 {
1151         lnet_rc_data_t          *rcd;
1152         lnet_rc_data_t          *tmp;
1153         lnet_peer_t             *lp;
1154         cfs_list_t              head;
1155         int                     i = 2;
1156
1157         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1158                    cfs_list_empty(&the_lnet.ln_rcd_deathrow) &&
1159                    cfs_list_empty(&the_lnet.ln_rcd_zombie)))
1160                 return;
1161
1162         CFS_INIT_LIST_HEAD(&head);
1163
1164         lnet_net_lock(LNET_LOCK_EX);
1165
1166         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1167                 /* router checker is stopping, prune all */
1168                 cfs_list_for_each_entry(lp, &the_lnet.ln_routers,
1169                                         lp_rtr_list) {
1170                         if (lp->lp_rcd == NULL)
1171                                 continue;
1172
1173                         LASSERT(cfs_list_empty(&lp->lp_rcd->rcd_list));
1174                         cfs_list_add(&lp->lp_rcd->rcd_list,
1175                                      &the_lnet.ln_rcd_deathrow);
1176                         lp->lp_rcd = NULL;
1177                 }
1178         }
1179
1180         /* unlink all RCDs on deathrow list */
1181         cfs_list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1182
1183         if (!cfs_list_empty(&head)) {
1184                 lnet_net_unlock(LNET_LOCK_EX);
1185
1186                 cfs_list_for_each_entry(rcd, &head, rcd_list)
1187                         LNetMDUnlink(rcd->rcd_mdh);
1188
1189                 lnet_net_lock(LNET_LOCK_EX);
1190         }
1191
1192         cfs_list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1193
1194         /* release all zombie RCDs */
1195         while (!cfs_list_empty(&the_lnet.ln_rcd_zombie)) {
1196                 cfs_list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1197                                              rcd_list) {
1198                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1199                                 cfs_list_move(&rcd->rcd_list, &head);
1200                 }
1201
1202                 wait_unlink = wait_unlink &&
1203                               !cfs_list_empty(&the_lnet.ln_rcd_zombie);
1204
1205                 lnet_net_unlock(LNET_LOCK_EX);
1206
1207                 while (!cfs_list_empty(&head)) {
1208                         rcd = cfs_list_entry(head.next,
1209                                              lnet_rc_data_t, rcd_list);
1210                         cfs_list_del_init(&rcd->rcd_list);
1211                         lnet_destroy_rc_data(rcd);
1212                 }
1213
1214                 if (!wait_unlink)
1215                         return;
1216
1217                 i++;
1218                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1219                        "Waiting for rc buffers to unlink\n");
1220                 cfs_pause(cfs_time_seconds(1) / 4);
1221
1222                 lnet_net_lock(LNET_LOCK_EX);
1223         }
1224
1225         lnet_net_unlock(LNET_LOCK_EX);
1226 }
1227
1228
1229 #if defined(__KERNEL__) && defined(LNET_ROUTER)
1230
1231 static int
1232 lnet_router_checker(void *arg)
1233 {
1234         lnet_peer_t       *rtr;
1235         cfs_list_t        *entry;
1236
1237         cfs_block_allsigs();
1238
1239         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1240
1241         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1242                 __u64   version;
1243                 int     cpt;
1244                 int     cpt2;
1245
1246                 cpt = lnet_net_lock_current();
1247 rescan:
1248                 version = the_lnet.ln_routers_version;
1249
1250                 cfs_list_for_each(entry, &the_lnet.ln_routers) {
1251                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
1252
1253                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1254                         if (cpt != cpt2) {
1255                                 lnet_net_unlock(cpt);
1256                                 cpt = cpt2;
1257                                 lnet_net_lock(cpt);
1258                                 /* the routers list has changed */
1259                                 if (version != the_lnet.ln_routers_version)
1260                                         goto rescan;
1261                         }
1262
1263                         lnet_ping_router_locked(rtr);
1264
1265                         /* NB dropped lock */
1266                         if (version != the_lnet.ln_routers_version) {
1267                                 /* the routers list has changed */
1268                                 goto rescan;
1269                         }
1270                 }
1271
1272                 if (the_lnet.ln_routing)
1273                         lnet_update_ni_status_locked();
1274
1275                 lnet_net_unlock(cpt);
1276
1277                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1278
1279                 /* Call cfs_pause() here always adds 1 to load average
1280                  * because kernel counts # active tasks as nr_running
1281                  * + nr_uninterruptible. */
1282                 schedule_timeout_and_set_state(TASK_INTERRUPTIBLE,
1283                                                    cfs_time_seconds(1));
1284         }
1285
1286         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1287
1288         lnet_prune_rc_data(1); /* wait for UNLINK */
1289
1290         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1291         up(&the_lnet.ln_rc_signal);
1292         /* The unlink event callback will signal final completion */
1293         return 0;
1294 }
1295
1296 void
1297 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1298 {
1299         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1300
1301         while (--npages >= 0)
1302                 __free_page(rb->rb_kiov[npages].kiov_page);
1303
1304         LIBCFS_FREE(rb, sz);
1305 }
1306
1307 lnet_rtrbuf_t *
1308 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1309 {
1310         int            npages = rbp->rbp_npages;
1311         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1312         struct page   *page;
1313         lnet_rtrbuf_t *rb;
1314         int            i;
1315
1316         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1317         if (rb == NULL)
1318                 return NULL;
1319
1320         rb->rb_pool = rbp;
1321
1322         for (i = 0; i < npages; i++) {
1323                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1324                                           __GFP_ZERO | GFP_IOFS);
1325                 if (page == NULL) {
1326                         while (--i >= 0)
1327                                 __free_page(rb->rb_kiov[i].kiov_page);
1328
1329                         LIBCFS_FREE(rb, sz);
1330                         return NULL;
1331                 }
1332
1333                 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1334                 rb->rb_kiov[i].kiov_offset = 0;
1335                 rb->rb_kiov[i].kiov_page = page;
1336         }
1337
1338         return rb;
1339 }
1340
1341 void
1342 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
1343 {
1344         int             npages = rbp->rbp_npages;
1345         int             nbuffers = 0;
1346         lnet_rtrbuf_t   *rb;
1347
1348         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1349                 return;
1350
1351         LASSERT (cfs_list_empty(&rbp->rbp_msgs));
1352         LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
1353
1354         while (!cfs_list_empty(&rbp->rbp_bufs)) {
1355                 LASSERT (rbp->rbp_credits > 0);
1356
1357                 rb = cfs_list_entry(rbp->rbp_bufs.next,
1358                                     lnet_rtrbuf_t, rb_list);
1359                 cfs_list_del(&rb->rb_list);
1360                 lnet_destroy_rtrbuf(rb, npages);
1361                 nbuffers++;
1362         }
1363
1364         LASSERT (rbp->rbp_nbuffers == nbuffers);
1365         LASSERT (rbp->rbp_credits == nbuffers);
1366
1367         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1368 }
1369
1370 int
1371 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1372 {
1373         lnet_rtrbuf_t *rb;
1374         int            i;
1375
1376         if (rbp->rbp_nbuffers != 0) {
1377                 LASSERT (rbp->rbp_nbuffers == nbufs);
1378                 return 0;
1379         }
1380
1381         for (i = 0; i < nbufs; i++) {
1382                 rb = lnet_new_rtrbuf(rbp, cpt);
1383
1384                 if (rb == NULL) {
1385                         CERROR("Failed to allocate %d router bufs of %d pages\n",
1386                                nbufs, rbp->rbp_npages);
1387                         return -ENOMEM;
1388                 }
1389
1390                 rbp->rbp_nbuffers++;
1391                 rbp->rbp_credits++;
1392                 rbp->rbp_mincredits++;
1393                 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1394
1395                 /* No allocation "under fire" */
1396                 /* Otherwise we'd need code to schedule blocked msgs etc */
1397                 LASSERT (!the_lnet.ln_routing);
1398         }
1399
1400         LASSERT (rbp->rbp_credits == nbufs);
1401         return 0;
1402 }
1403
1404 void
1405 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1406 {
1407         CFS_INIT_LIST_HEAD(&rbp->rbp_msgs);
1408         CFS_INIT_LIST_HEAD(&rbp->rbp_bufs);
1409
1410         rbp->rbp_npages = npages;
1411         rbp->rbp_credits = 0;
1412         rbp->rbp_mincredits = 0;
1413 }
1414
1415 void
1416 lnet_rtrpools_free(void)
1417 {
1418         lnet_rtrbufpool_t *rtrp;
1419         int               i;
1420
1421         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1422                 return;
1423
1424         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1425                 lnet_rtrpool_free_bufs(&rtrp[0]);
1426                 lnet_rtrpool_free_bufs(&rtrp[1]);
1427                 lnet_rtrpool_free_bufs(&rtrp[2]);
1428         }
1429
1430         cfs_percpt_free(the_lnet.ln_rtrpools);
1431         the_lnet.ln_rtrpools = NULL;
1432 }
1433
1434 static int
1435 lnet_nrb_tiny_calculate(int npages)
1436 {
1437         int     nrbs = LNET_NRB_TINY;
1438
1439         if (tiny_router_buffers < 0) {
1440                 LCONSOLE_ERROR_MSG(0x10c,
1441                                    "tiny_router_buffers=%d invalid when "
1442                                    "routing enabled\n", tiny_router_buffers);
1443                 return -1;
1444         }
1445
1446         if (tiny_router_buffers > 0)
1447                 nrbs = tiny_router_buffers;
1448
1449         nrbs /= LNET_CPT_NUMBER;
1450         return max(nrbs, LNET_NRB_TINY_MIN);
1451 }
1452
1453 static int
1454 lnet_nrb_small_calculate(int npages)
1455 {
1456         int     nrbs = LNET_NRB_SMALL;
1457
1458         if (small_router_buffers < 0) {
1459                 LCONSOLE_ERROR_MSG(0x10c,
1460                                    "small_router_buffers=%d invalid when "
1461                                    "routing enabled\n", small_router_buffers);
1462                 return -1;
1463         }
1464
1465         if (small_router_buffers > 0)
1466                 nrbs = small_router_buffers;
1467
1468         nrbs /= LNET_CPT_NUMBER;
1469         return max(nrbs, LNET_NRB_SMALL_MIN);
1470 }
1471
1472 static int
1473 lnet_nrb_large_calculate(int npages)
1474 {
1475         int     nrbs = LNET_NRB_LARGE;
1476
1477         if (large_router_buffers < 0) {
1478                 LCONSOLE_ERROR_MSG(0x10c,
1479                                    "large_router_buffers=%d invalid when "
1480                                    "routing enabled\n", large_router_buffers);
1481                 return -1;
1482         }
1483
1484         if (large_router_buffers > 0)
1485                 nrbs = large_router_buffers;
1486
1487         nrbs /= LNET_CPT_NUMBER;
1488         return max(nrbs, LNET_NRB_LARGE_MIN);
1489 }
1490
1491 int
1492 lnet_rtrpools_alloc(int im_a_router)
1493 {
1494         lnet_rtrbufpool_t *rtrp;
1495         int     large_pages = (LNET_MTU + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1496         int     small_pages = 1;
1497         int     nrb_tiny;
1498         int     nrb_small;
1499         int     nrb_large;
1500         int     rc;
1501         int     i;
1502
1503         if (!strcmp(forwarding, "")) {
1504                 /* not set either way */
1505                 if (!im_a_router)
1506                         return 0;
1507         } else if (!strcmp(forwarding, "disabled")) {
1508                 /* explicitly disabled */
1509                 return 0;
1510         } else if (!strcmp(forwarding, "enabled")) {
1511                 /* explicitly enabled */
1512         } else {
1513                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1514                                    "'enabled' or 'disabled'\n");
1515                 return -EINVAL;
1516         }
1517
1518         nrb_tiny = lnet_nrb_tiny_calculate(0);
1519         if (nrb_tiny < 0)
1520                 return -EINVAL;
1521
1522         nrb_small = lnet_nrb_small_calculate(small_pages);
1523         if (nrb_small < 0)
1524                 return -EINVAL;
1525
1526         nrb_large = lnet_nrb_large_calculate(large_pages);
1527         if (nrb_large < 0)
1528                 return -EINVAL;
1529
1530         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1531                                                 LNET_NRBPOOLS *
1532                                                 sizeof(lnet_rtrbufpool_t));
1533         if (the_lnet.ln_rtrpools == NULL) {
1534                 LCONSOLE_ERROR_MSG(0x10c,
1535                                    "Failed to initialize router buffe pool\n");
1536                 return -ENOMEM;
1537         }
1538
1539         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1540                 lnet_rtrpool_init(&rtrp[0], 0);
1541                 rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i);
1542                 if (rc != 0)
1543                         goto failed;
1544
1545                 lnet_rtrpool_init(&rtrp[1], small_pages);
1546                 rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i);
1547                 if (rc != 0)
1548                         goto failed;
1549
1550                 lnet_rtrpool_init(&rtrp[2], large_pages);
1551                 rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i);
1552                 if (rc != 0)
1553                         goto failed;
1554         }
1555
1556         lnet_net_lock(LNET_LOCK_EX);
1557         the_lnet.ln_routing = 1;
1558         lnet_net_unlock(LNET_LOCK_EX);
1559
1560         return 0;
1561
1562  failed:
1563         lnet_rtrpools_free();
1564         return rc;
1565 }
1566
1567 int
1568 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1569 {
1570         struct lnet_peer        *lp = NULL;
1571         cfs_time_t              now = cfs_time_current();
1572         int                     cpt = lnet_cpt_of_nid(nid);
1573
1574         LASSERT (!cfs_in_interrupt ());
1575
1576         CDEBUG (D_NET, "%s notifying %s: %s\n",
1577                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1578                 libcfs_nid2str(nid),
1579                 alive ? "up" : "down");
1580
1581         if (ni != NULL &&
1582             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1583                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1584                         libcfs_nid2str(nid), alive ? "birth" : "death",
1585                         libcfs_nid2str(ni->ni_nid));
1586                 return -EINVAL;
1587         }
1588
1589         /* can't do predictions... */
1590         if (cfs_time_after(when, now)) {
1591                 CWARN ("Ignoring prediction from %s of %s %s "
1592                        "%ld seconds in the future\n",
1593                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1594                        libcfs_nid2str(nid), alive ? "up" : "down",
1595                        cfs_duration_sec(cfs_time_sub(when, now)));
1596                 return -EINVAL;
1597         }
1598
1599         if (ni != NULL && !alive &&             /* LND telling me she's down */
1600             !auto_down) {                       /* auto-down disabled */
1601                 CDEBUG(D_NET, "Auto-down disabled\n");
1602                 return 0;
1603         }
1604
1605         lnet_net_lock(cpt);
1606
1607         if (the_lnet.ln_shutdown) {
1608                 lnet_net_unlock(cpt);
1609                 return -ESHUTDOWN;
1610         }
1611
1612         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1613         if (lp == NULL) {
1614                 /* nid not found */
1615                 lnet_net_unlock(cpt);
1616                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1617                 return 0;
1618         }
1619
1620         /* We can't fully trust LND on reporting exact peer last_alive
1621          * if he notifies us about dead peer. For example ksocklnd can
1622          * call us with when == _time_when_the_node_was_booted_ if
1623          * no connections were successfully established */
1624         if (ni != NULL && !alive && when < lp->lp_last_alive)
1625                 when = lp->lp_last_alive;
1626
1627         lnet_notify_locked(lp, ni == NULL, alive, when);
1628
1629         lnet_ni_notify_locked(ni, lp);
1630
1631         lnet_peer_decref_locked(lp);
1632
1633         lnet_net_unlock(cpt);
1634         return 0;
1635 }
1636 EXPORT_SYMBOL(lnet_notify);
1637
1638 void
1639 lnet_get_tunables (void)
1640 {
1641         return;
1642 }
1643
1644 #else
1645
1646 int
1647 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1648 {
1649         return -EOPNOTSUPP;
1650 }
1651
1652 void
1653 lnet_router_checker (void)
1654 {
1655         static time_t last = 0;
1656         static int    running = 0;
1657
1658         time_t            now = cfs_time_current_sec();
1659         int               interval = now - last;
1660         int               rc;
1661         __u64             version;
1662         lnet_peer_t      *rtr;
1663
1664         /* It's no use to call me again within a sec - all intervals and
1665          * timeouts are measured in seconds */
1666         if (last != 0 && interval < 2)
1667                 return;
1668
1669         if (last != 0 &&
1670             interval > MAX(live_router_check_interval,
1671                            dead_router_check_interval))
1672                 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1673                         live_router_check_interval, dead_router_check_interval,
1674                         interval);
1675
1676         LASSERT(LNET_CPT_NUMBER == 1);
1677
1678         lnet_net_lock(0);
1679         LASSERT(!running); /* recursion check */
1680         running = 1;
1681         lnet_net_unlock(0);
1682
1683         last = now;
1684
1685         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1686                 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1687
1688         /* consume all pending events */
1689         while (1) {
1690                 int          i;
1691                 lnet_event_t ev;
1692
1693                 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1694                  * recursion breaker in LNetEQPoll would fail */
1695                 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1696                 if (rc == 0)   /* no event pending */
1697                         break;
1698
1699                 /* NB a lost SENT prevents me from pinging a router again */
1700                 if (rc == -EOVERFLOW) {
1701                         CERROR("Dropped an event!!!\n");
1702                         abort();
1703                 }
1704
1705                 LASSERT (rc == 1);
1706
1707                 lnet_router_checker_event(&ev);
1708         }
1709
1710         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1711                 lnet_prune_rc_data(1); /* release rcd */
1712                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1713                 running = 0;
1714                 return;
1715         }
1716
1717         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1718
1719         lnet_net_lock(0);
1720
1721         version = the_lnet.ln_routers_version;
1722         cfs_list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) {
1723                 lnet_ping_router_locked(rtr);
1724                 LASSERT (version == the_lnet.ln_routers_version);
1725         }
1726
1727         lnet_net_unlock(0);
1728
1729         running = 0; /* lock only needed for the recursion check */
1730         return;
1731 }
1732
1733 /* NB lnet_peers_start_down depends on me,
1734  * so must be called before any peer creation */
1735 void
1736 lnet_get_tunables (void)
1737 {
1738         char *s;
1739
1740         s = getenv("LNET_ROUTER_PING_TIMEOUT");
1741         if (s != NULL) router_ping_timeout = atoi(s);
1742
1743         s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1744         if (s != NULL) live_router_check_interval = atoi(s);
1745
1746         s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1747         if (s != NULL) dead_router_check_interval = atoi(s);
1748
1749         /* This replaces old lnd_notify mechanism */
1750         check_routers_before_use = 1;
1751         if (dead_router_check_interval <= 0)
1752                 dead_router_check_interval = 30;
1753 }
1754
1755 void
1756 lnet_rtrpools_free(void)
1757 {
1758 }
1759
1760 int
1761 lnet_rtrpools_alloc(int im_a_arouter)
1762 {
1763         return 0;
1764 }
1765
1766 #endif