lnet/lnet/router.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
   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 static char *forwarding = "";
  30 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
  31                 "Explicitly enable/disable forwarding between networks");
  32
  33 static int tiny_router_buffers = 1024;
  34 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
  35                 "# of 0 payload messages to buffer in the router");
  36 static int small_router_buffers = 8192;
  37 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
  38                 "# of small (1 page) messages to buffer in the router");
  39 static int large_router_buffers = 512;
  40 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
  41                 "# of large messages to buffer in the router");
  42
  43 static int auto_down = 1;
  44 CFS_MODULE_PARM(auto_down, "i", int, 0444,
  45                 "Automatically mark peers down on comms error");
  46
  47 static int check_routers_before_use = 0;
  48 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
  49                 "Assume routers are down and ping them before use");
  50
  51 static int dead_router_check_interval = 0;
  52 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0444,
  53                 "Seconds between dead router health checks (<= 0 to disable)");
  54
  55 static int live_router_check_interval = 0;
  56 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0444,
  57                 "Seconds between live router health checks (<= 0 to disable)");
  58
  59 static int router_ping_timeout = 50;
  60 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0444,
  61                 "Seconds to wait for the reply to a router health query");
  62
  63 int
  64 lnet_peers_start_down(void)
  65 {
  66         return check_routers_before_use;
  67 }
  68
  69 void
  70 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, time_t when)
  71 {
  72         if (when < lp->lp_timestamp) {          /* out of date information */
  73                 CDEBUG(D_NET, "Out of date\n");
  74                 return;
  75         }
  76
  77         lp->lp_timestamp = when;                /* update timestamp */
  78         lp->lp_ping_deadline = 0;               /* disable ping timeout */
  79
  80         if (lp->lp_alive_count != 0 &&          /* got old news */
  81             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
  82                 CDEBUG(D_NET, "Old news\n");
  83                 return;
  84         }
  85
  86         /* Flag that notification is outstanding */
  87
  88         lp->lp_alive_count++;
  89         lp->lp_alive = !(!alive);               /* 1 bit! */
  90         lp->lp_notify = 1;
  91         lp->lp_notifylnd |= notifylnd;
  92
  93         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
  94 }
  95
  96 void
  97 lnet_do_notify (lnet_peer_t *lp)
  98 {
  99         lnet_ni_t *ni = lp->lp_ni;
 100         int        alive;
 101         int        notifylnd;
 102
 103         LNET_LOCK();
 104
 105         /* Notify only in 1 thread at any time to ensure ordered notification.
 106          * NB individual events can be missed; the only guarantee is that you
 107          * always get the most recent news */
 108
 109         if (lp->lp_notifying) {
 110                 LNET_UNLOCK();
 111                 return;
 112         }
 113
 114         lp->lp_notifying = 1;
 115
 116         while (lp->lp_notify) {
 117                 alive     = lp->lp_alive;
 118                 notifylnd = lp->lp_notifylnd;
 119
 120                 lp->lp_notifylnd = 0;
 121                 lp->lp_notify    = 0;
 122
 123                 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
 124                         LNET_UNLOCK();
 125
 126                         /* A new notification could happen now; I'll handle it
 127                          * when control returns to me */
 128
 129                         (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
 130
 131                         LNET_LOCK();
 132                 }
 133         }
 134
 135         lp->lp_notifying = 0;
 136
 137         LNET_UNLOCK();
 138 }
 139
 140 int
 141 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, time_t when)
 142 {
 143         lnet_peer_t         *lp = NULL;
 144         time_t               now = cfs_time_current_sec();
 145
 146         LASSERT (!in_interrupt ());
 147
 148         CDEBUG (D_NET, "%s notifying %s: %s\n",
 149                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
 150                 libcfs_nid2str(nid),
 151                 alive ? "up" : "down");
 152
 153         if (ni != NULL &&
 154             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
 155                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
 156                         libcfs_nid2str(nid), alive ? "birth" : "death",
 157                         libcfs_nid2str(ni->ni_nid));
 158                 return -EINVAL;
 159         }
 160
 161         /* can't do predictions... */
 162         if (when > now) {
 163                 CWARN ("Ignoring prediction from %s of %s %s "
 164                        "%ld seconds in the future\n",
 165                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
 166                        libcfs_nid2str(nid), alive ? "up" : "down",
 167                        when - now);
 168                 return -EINVAL;
 169         }
 170
 171         if (ni != NULL && !alive &&             /* LND telling me she's down */
 172             !auto_down) {                       /* auto-down disabled */
 173                 CDEBUG(D_NET, "Auto-down disabled\n");
 174                 return 0;
 175         }
 176
 177         LNET_LOCK();
 178
 179         lp = lnet_find_peer_locked(nid);
 180         if (lp == NULL) {
 181                 /* nid not found */
 182                 LNET_UNLOCK();
 183                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
 184                 return 0;
 185         }
 186
 187         lnet_notify_locked(lp, ni == NULL, alive, when);
 188
 189         LNET_UNLOCK();
 190
 191         lnet_do_notify(lp);
 192
 193         LNET_LOCK();
 194
 195         lnet_peer_decref_locked(lp);
 196
 197         LNET_UNLOCK();
 198         return 0;
 199 }
 200 EXPORT_SYMBOL(lnet_notify);
 201
 202 #else
 203
 204 int
 205 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, time_t when)
 206 {
 207         return -EOPNOTSUPP;
 208 }
 209
 210 #endif
 211
 212 static void
 213 lnet_rtr_addref_locked(lnet_peer_t *lp)
 214 {
 215         LASSERT (lp->lp_refcount > 0);
 216         LASSERT (lp->lp_rtr_refcount >= 0);
 217
 218         lp->lp_rtr_refcount++;
 219         if (lp->lp_rtr_refcount == 1) {
 220                 struct list_head *pos;
 221
 222                 /* a simple insertion sort */
 223                 list_for_each_prev(pos, &the_lnet.ln_routers) {
 224                         lnet_peer_t *rtr = list_entry(pos, lnet_peer_t,
 225                                                       lp_rtr_list);
 226
 227                         if (rtr->lp_nid < lp->lp_nid)
 228                                 break;
 229                 }
 230
 231                 list_add(&lp->lp_rtr_list, pos);
 232                 /* addref for the_lnet.ln_routers */
 233                 lnet_peer_addref_locked(lp);
 234                 the_lnet.ln_routers_version++;
 235         }
 236 }
 237
 238 static void
 239 lnet_rtr_decref_locked(lnet_peer_t *lp)
 240 {
 241         LASSERT (lp->lp_refcount > 0);
 242         LASSERT (lp->lp_rtr_refcount > 0);
 243
 244         lp->lp_rtr_refcount--;
 245         if (lp->lp_rtr_refcount == 0) {
 246                 list_del(&lp->lp_rtr_list);
 247                 /* decref for the_lnet.ln_routers */
 248                 lnet_peer_decref_locked(lp);
 249                 the_lnet.ln_routers_version++;
 250         }
 251 }
 252
 253 lnet_remotenet_t *
 254 lnet_find_net_locked (__u32 net)
 255 {
 256         lnet_remotenet_t *rnet;
 257         struct list_head *tmp;
 258
 259         LASSERT (!the_lnet.ln_shutdown);
 260
 261         list_for_each (tmp, &the_lnet.ln_remote_nets) {
 262                 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
 263
 264                 if (rnet->lrn_net == net)
 265                         return rnet;
 266         }
 267         return NULL;
 268 }
 269
 270 int
 271 lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway)
 272 {
 273         struct list_head     zombies;
 274         struct list_head    *e;
 275         lnet_remotenet_t    *rnet;
 276         lnet_remotenet_t    *rnet2;
 277         lnet_route_t        *route;
 278         lnet_route_t        *route2;
 279         lnet_ni_t           *ni;
 280         int                  add_route;
 281         int                  rc;
 282
 283         CDEBUG(D_NET, "Add route: net %s hops %u gw %s\n",
 284                libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 285
 286         if (gateway == LNET_NID_ANY ||
 287             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 288             net == LNET_NIDNET(LNET_NID_ANY) ||
 289             LNET_NETTYP(net) == LOLND ||
 290             LNET_NIDNET(gateway) == net ||
 291             hops < 1 || hops > 255)
 292                 return (-EINVAL);
 293
 294         if (lnet_islocalnet(net))               /* it's a local network */
 295                 return 0;                       /* ignore the route entry */
 296
 297         /* Assume net, route, all new */
 298         LIBCFS_ALLOC(route, sizeof(*route));
 299         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 300         if (route == NULL || rnet == NULL) {
 301                 CERROR("Out of memory creating route %s %d %s\n",
 302                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 303                 if (route != NULL)
 304                         LIBCFS_FREE(route, sizeof(*route));
 305                 if (rnet != NULL)
 306                         LIBCFS_FREE(rnet, sizeof(*rnet));
 307                 return -ENOMEM;
 308         }
 309
 310         CFS_INIT_LIST_HEAD(&rnet->lrn_routes);
 311         rnet->lrn_net = net;
 312         rnet->lrn_hops = hops;
 313
 314         LNET_LOCK();
 315
 316         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway);
 317         if (rc != 0) {
 318                 LNET_UNLOCK();
 319
 320                 LIBCFS_FREE(route, sizeof(*route));
 321                 LIBCFS_FREE(rnet, sizeof(*rnet));
 322
 323                 if (rc == -EHOSTUNREACH)        /* gateway is not on a local net */
 324                         return 0;               /* ignore the route entry */
 325
 326                 CERROR("Error %d creating route %s %d %s\n", rc,
 327                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 328                 return rc;
 329         }
 330
 331         LASSERT (!the_lnet.ln_shutdown);
 332         CFS_INIT_LIST_HEAD(&zombies);
 333
 334         rnet2 = lnet_find_net_locked(net);
 335         if (rnet2 == NULL) {
 336                 /* new network */
 337                 list_add_tail(&rnet->lrn_list, &the_lnet.ln_remote_nets);
 338                 rnet2 = rnet;
 339         }
 340
 341         if (hops > rnet2->lrn_hops) {
 342                 /* New route is longer; ignore it */
 343                 add_route = 0;
 344         } else if (hops < rnet2->lrn_hops) {
 345                 /* new route supercedes all currently known routes to this
 346                  * net */
 347                 list_add(&zombies, &rnet2->lrn_routes);
 348                 list_del_init(&rnet2->lrn_routes);
 349                 add_route = 1;
 350         } else {
 351                 add_route = 1;
 352                 /* New route has the same hopcount as existing routes; search
 353                  * for a duplicate route (it's a NOOP if it is) */
 354                 list_for_each (e, &rnet2->lrn_routes) {
 355                         route2 = list_entry(e, lnet_route_t, lr_list);
 356
 357                         if (route2->lr_gateway == route->lr_gateway) {
 358                                 add_route = 0;
 359                                 break;
 360                         }
 361
 362                         /* our loopups must be true */
 363                         LASSERT (route2->lr_gateway->lp_nid != gateway);
 364                 }
 365         }
 366
 367         if (add_route) {
 368                 ni = route->lr_gateway->lp_ni;
 369                 lnet_ni_addref_locked(ni);
 370
 371                 LASSERT (rc == 0);
 372                 list_add_tail(&route->lr_list, &rnet2->lrn_routes);
 373                 the_lnet.ln_remote_nets_version++;
 374
 375                 lnet_rtr_addref_locked(route->lr_gateway);
 376
 377                 LNET_UNLOCK();
 378
 379                 /* XXX Assume alive */
 380                 if (ni->ni_lnd->lnd_notify != NULL)
 381                         (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
 382
 383                 lnet_ni_decref(ni);
 384         } else {
 385                 lnet_peer_decref_locked(route->lr_gateway);
 386                 LNET_UNLOCK();
 387                 LIBCFS_FREE(route, sizeof(*route));
 388         }
 389
 390         if (rnet != rnet2)
 391                 LIBCFS_FREE(rnet, sizeof(*rnet));
 392
 393         while (!list_empty(&zombies)) {
 394                 route = list_entry(zombies.next, lnet_route_t, lr_list);
 395                 list_del(&route->lr_list);
 396
 397                 LNET_LOCK();
 398                 lnet_rtr_decref_locked(route->lr_gateway);
 399                 lnet_peer_decref_locked(route->lr_gateway);
 400                 LNET_UNLOCK();
 401                 LIBCFS_FREE(route, sizeof(*route));
 402         }
 403
 404         return rc;
 405 }
 406
 407 int
 408 lnet_check_routes (void)
 409 {
 410         lnet_remotenet_t    *rnet;
 411         lnet_route_t        *route;
 412         lnet_route_t        *route2;
 413         struct list_head    *e1;
 414         struct list_head    *e2;
 415
 416         LNET_LOCK();
 417
 418         list_for_each (e1, &the_lnet.ln_remote_nets) {
 419                 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 420
 421                 route2 = NULL;
 422                 list_for_each (e2, &rnet->lrn_routes) {
 423                         route = list_entry(e2, lnet_route_t, lr_list);
 424
 425                         if (route2 == NULL)
 426                                 route2 = route;
 427                         else if (route->lr_gateway->lp_ni !=
 428                                  route2->lr_gateway->lp_ni) {
 429                                 LNET_UNLOCK();
 430
 431                                 CERROR("Routes to %s via %s and %s not supported\n",
 432                                        libcfs_net2str(rnet->lrn_net),
 433                                        libcfs_nid2str(route->lr_gateway->lp_nid),
 434                                        libcfs_nid2str(route2->lr_gateway->lp_nid));
 435                                 return -EINVAL;
 436                         }
 437                 }
 438         }
 439
 440         LNET_UNLOCK();
 441         return 0;
 442 }
 443
 444 int
 445 lnet_del_route (__u32 net, lnet_nid_t gw_nid)
 446 {
 447         lnet_remotenet_t    *rnet;
 448         lnet_route_t        *route;
 449         struct list_head    *e1;
 450         struct list_head    *e2;
 451         int                  rc = -ENOENT;
 452
 453         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 454                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 455
 456         /* NB Caller may specify either all routes via the given gateway
 457          * or a specific route entry actual NIDs) */
 458
 459  again:
 460         LNET_LOCK();
 461
 462         list_for_each (e1, &the_lnet.ln_remote_nets) {
 463                 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 464
 465                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 466                       net == rnet->lrn_net))
 467                         continue;
 468
 469                 list_for_each (e2, &rnet->lrn_routes) {
 470                         route = list_entry(e2, lnet_route_t, lr_list);
 471
 472                         if (!(gw_nid == LNET_NID_ANY ||
 473                               gw_nid == route->lr_gateway->lp_nid))
 474                                 continue;
 475
 476                         list_del(&route->lr_list);
 477                         the_lnet.ln_remote_nets_version++;
 478
 479                         if (list_empty(&rnet->lrn_routes))
 480                                 list_del(&rnet->lrn_list);
 481                         else
 482                                 rnet = NULL;
 483
 484                         lnet_rtr_decref_locked(route->lr_gateway);
 485                         lnet_peer_decref_locked(route->lr_gateway);
 486                         LNET_UNLOCK();
 487
 488                         LIBCFS_FREE(route, sizeof (*route));
 489
 490                         if (rnet != NULL)
 491                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 492
 493                         rc = 0;
 494                         goto again;
 495                 }
 496         }
 497
 498         LNET_UNLOCK();
 499         return rc;
 500 }
 501
 502 void
 503 lnet_destroy_routes (void)
 504 {
 505         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 506 }
 507
 508 int
 509 lnet_get_route (int idx, __u32 *net, __u32 *hops,
 510                lnet_nid_t *gateway, __u32 *alive)
 511 {
 512         struct list_head    *e1;
 513         struct list_head    *e2;
 514         lnet_remotenet_t    *rnet;
 515         lnet_route_t        *route;
 516
 517         LNET_LOCK();
 518
 519         list_for_each (e1, &the_lnet.ln_remote_nets) {
 520                 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 521
 522                 list_for_each (e2, &rnet->lrn_routes) {
 523                         route = list_entry(e2, lnet_route_t, lr_list);
 524
 525                         if (idx-- == 0) {
 526                                 *net     = rnet->lrn_net;
 527                                 *hops    = rnet->lrn_hops;
 528                                 *gateway = route->lr_gateway->lp_nid;
 529                                 *alive   = route->lr_gateway->lp_alive;
 530                                 LNET_UNLOCK();
 531                                 return 0;
 532                         }
 533                 }
 534         }
 535
 536         LNET_UNLOCK();
 537         return -ENOENT;
 538 }
 539
 540 #if defined(__KERNEL__) && defined(LNET_ROUTER)
 541 static void
 542 lnet_router_checker_event (lnet_event_t *event)
 543 {
 544         /* CAVEAT EMPTOR: I'm called with LNET_LOCKed and I'm not allowed to
 545          * drop it (that's how come I see _every_ event, even ones that would
 546          * overflow my EQ) */
 547         lnet_peer_t   *lp;
 548         lnet_nid_t     nid;
 549
 550         if (event->unlinked) {
 551                 /* The router checker thread has unlinked the rc_md
 552                  * and exited. */
 553                 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING);
 554                 the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKED;
 555                 mutex_up(&the_lnet.ln_rc_signal);
 556                 return;
 557         }
 558
 559         LASSERT (event->type == LNET_EVENT_SEND ||
 560                  event->type == LNET_EVENT_REPLY);
 561
 562         nid = (event->type == LNET_EVENT_SEND) ?
 563               event->target.nid : event->initiator.nid;
 564
 565         lp = lnet_find_peer_locked(nid);
 566         if (lp == NULL) {
 567                 /* router may have been removed */
 568                 CDEBUG(D_NET, "Router %s not found\n", libcfs_nid2str(nid));
 569                 return;
 570         }
 571
 572         if (event->type == LNET_EVENT_SEND)     /* re-enable another ping */
 573                 lp->lp_ping_notsent = 0;
 574
 575         if (lnet_isrouter(lp) &&                /* ignore if no longer a router */
 576             (event->status != 0 ||
 577              event->type == LNET_EVENT_REPLY)) {
 578
 579                 /* A successful REPLY means the router is up.  If _any_ comms
 580                  * to the router fail I assume it's down (this will happen if
 581                  * we ping alive routers to try to detect router death before
 582                  * apps get burned). */
 583
 584                 lnet_notify_locked(lp, 1, (event->status == 0),
 585                                    cfs_time_current_sec());
 586
 587                 /* The router checker will wake up very shortly and do the
 588                  * actual notification.
 589                  * XXX If 'lp' stops being a router before then, it will still
 590                  * have the notification pending!!! */
 591         }
 592
 593         /* This decref will NOT drop LNET_LOCK (it had to have 1 ref when it
 594          * was in the peer table and I've not dropped the lock, so no-one else
 595          * can have reduced the refcount) */
 596         LASSERT(lp->lp_refcount > 1);
 597
 598         lnet_peer_decref_locked(lp);
 599 }
 600
 601 static int
 602 lnet_router_checker(void *arg)
 603 {
 604         static lnet_ping_info_t   pinginfo;
 605
 606         int                  rc;
 607         lnet_handle_md_t     mdh;
 608         lnet_peer_t         *rtr;
 609         lnet_md_t            md = {0};
 610         struct list_head    *entry;
 611         time_t               now;
 612         lnet_process_id_t    rtr_id;
 613         int                  secs;
 614
 615         cfs_daemonize("router_checker");
 616         cfs_block_allsigs();
 617
 618         rtr_id.pid = LUSTRE_SRV_LNET_PID;
 619
 620         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
 621
 622         /* initialize md content */
 623         md.start     = &pinginfo;
 624         md.length    = sizeof(pinginfo);
 625         md.threshold = LNET_MD_THRESH_INF;
 626         md.max_size  = 0;
 627         md.options   = LNET_MD_TRUNCATE,
 628         md.user_ptr  = NULL;
 629         md.eq_handle = the_lnet.ln_rc_eqh;
 630
 631         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
 632
 633         if (rc < 0) {
 634                 CERROR("Can't bind MD: %d\n", rc);
 635                 the_lnet.ln_rc_state = rc;
 636                 mutex_up(&the_lnet.ln_rc_signal);
 637                 return rc;
 638         }
 639
 640         LASSERT (rc == 0);
 641
 642         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
 643         mutex_up(&the_lnet.ln_rc_signal);       /* let my parent go */
 644
 645         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
 646                 __u64 version;
 647
 648                 LNET_LOCK();
 649 rescan:
 650                 version = the_lnet.ln_routers_version;
 651
 652                 list_for_each (entry, &the_lnet.ln_routers) {
 653                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
 654
 655                         lnet_peer_addref_locked(rtr);
 656
 657                         now = cfs_time_current_sec();
 658
 659                         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
 660                             now > rtr->lp_ping_deadline)
 661                                 lnet_notify_locked(rtr, 1, 0, now);
 662
 663                         LNET_UNLOCK();
 664
 665                         /* Run any outstanding notificiations */
 666                         lnet_do_notify(rtr);
 667
 668                         if (rtr->lp_alive) {
 669                                 secs = live_router_check_interval;
 670                         } else {
 671                                 secs = dead_router_check_interval;
 672                         }
 673                         if (secs <= 0)
 674                                 secs = 0;
 675
 676                         if (secs != 0 &&
 677                             !rtr->lp_ping_notsent &&
 678                             now > rtr->lp_ping_timestamp + secs) {
 679                                 CDEBUG(D_NET, "Check: %s\n",
 680                                        libcfs_nid2str(rtr->lp_nid));
 681
 682                                 LNET_LOCK();
 683                                 rtr_id.nid = rtr->lp_nid;
 684                                 rtr->lp_ping_notsent = 1;
 685                                 rtr->lp_ping_timestamp = now;
 686
 687                                 if (rtr->lp_ping_deadline == 0)
 688                                         rtr->lp_ping_deadline =
 689                                                 now + router_ping_timeout;
 690
 691                                 LNET_UNLOCK();
 692
 693                                 LNetGet(LNET_NID_ANY, mdh, rtr_id,
 694                                         LNET_RESERVED_PORTAL,
 695                                         LNET_PROTO_PING_MATCHBITS, 0);
 696                         }
 697
 698                         LNET_LOCK();
 699                         lnet_peer_decref_locked(rtr);
 700
 701                         if (version != the_lnet.ln_routers_version) {
 702                                 /* the routers list has changed */
 703                                 goto rescan;
 704                         }
 705                 }
 706
 707                 LNET_UNLOCK();
 708
 709                 /* Call cfs_pause() here always adds 1 to load average
 710                  * because kernel counts # active tasks as nr_running
 711                  * + nr_uninterruptible. */
 712                 cfs_schedule_timeout(CFS_TASK_INTERRUPTIBLE,
 713                                      cfs_time_seconds(1));
 714         }
 715
 716         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD);
 717         the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
 718
 719         rc = LNetMDUnlink(mdh);
 720         LASSERT (rc == 0);
 721
 722         /* The unlink event callback will signal final completion */
 723         return 0;
 724 }
 725
 726
 727 void
 728 lnet_wait_known_routerstate(void)
 729 {
 730         lnet_peer_t         *rtr;
 731         struct list_head    *entry;
 732         int                  all_known;
 733
 734         for (;;) {
 735                 LNET_LOCK();
 736
 737                 all_known = 1;
 738                 list_for_each (entry, &the_lnet.ln_routers) {
 739                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
 740
 741                         if (rtr->lp_alive_count == 0) {
 742                                 all_known = 0;
 743                                 break;
 744                         }
 745                 }
 746
 747                 LNET_UNLOCK();
 748
 749                 if (all_known)
 750                         return;
 751
 752                 cfs_pause(cfs_time_seconds(1));
 753         }
 754 }
 755
 756 void
 757 lnet_router_checker_stop(void)
 758 {
 759         int       rc;
 760
 761         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING ||
 762                  the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
 763
 764         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
 765                 return;
 766
 767         the_lnet.ln_rc_state = LNET_RC_STATE_STOPTHREAD;
 768         /* block until event callback signals exit */
 769         mutex_down(&the_lnet.ln_rc_signal);
 770
 771         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED);
 772
 773         rc = LNetEQFree(the_lnet.ln_rc_eqh);
 774         LASSERT (rc == 0);
 775
 776         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
 777 }
 778
 779 int
 780 lnet_router_checker_start(void)
 781 {
 782         int  rc;
 783
 784         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
 785
 786         if (check_routers_before_use &&
 787             dead_router_check_interval <= 0) {
 788                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
 789                                    " set if 'check_routers_before_use' is set"
 790                                    "\n");
 791                 return -EINVAL;
 792         }
 793
 794         if (live_router_check_interval <= 0 &&
 795             dead_router_check_interval <= 0)
 796                 return 0;
 797
 798         init_mutex_locked(&the_lnet.ln_rc_signal);
 799
 800         /* EQ size doesn't matter; the callback is guaranteed to get every
 801          * event */
 802         rc = LNetEQAlloc(1, lnet_router_checker_event,
 803                          &the_lnet.ln_rc_eqh);
 804         if (rc != 0) {
 805                 CERROR("Can't allocate EQ: %d\n", rc);
 806                 return -ENOMEM;
 807         }
 808
 809         rc = (int)cfs_kernel_thread(lnet_router_checker, NULL, 0);
 810         if (rc < 0) {
 811                 CERROR("Can't start router checker thread: %d\n", rc);
 812                 goto failed;
 813         }
 814
 815         mutex_down(&the_lnet.ln_rc_signal);     /* wait for checker to startup */
 816
 817         rc = the_lnet.ln_rc_state;
 818         if (rc < 0) {
 819                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
 820                 goto failed;
 821         }
 822
 823         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 824
 825         if (check_routers_before_use) {
 826                 /* Note that a helpful side-effect of pinging all known routers
 827                  * at startup is that it makes them drop stale connections they
 828                  * may have to a previous instance of me. */
 829                 lnet_wait_known_routerstate();
 830         }
 831
 832         return 0;
 833
 834  failed:
 835         rc = LNetEQFree(the_lnet.ln_rc_eqh);
 836         LASSERT (rc == 0);
 837         return rc;
 838 }
 839
 840 void
 841 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
 842 {
 843         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
 844
 845         while (--npages >= 0)
 846                 cfs_free_page(rb->rb_kiov[npages].kiov_page);
 847
 848         LIBCFS_FREE(rb, sz);
 849 }
 850
 851 lnet_rtrbuf_t *
 852 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp)
 853 {
 854         int            npages = rbp->rbp_npages;
 855         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
 856         struct page   *page;
 857         lnet_rtrbuf_t *rb;
 858         int            i;
 859
 860         LIBCFS_ALLOC(rb, sz);
 861         if (rb == NULL)
 862                 return NULL;
 863
 864         rb->rb_pool = rbp;
 865
 866         for (i = 0; i < npages; i++) {
 867                 page = cfs_alloc_page(CFS_ALLOC_ZERO | CFS_ALLOC_STD);
 868                 if (page == NULL) {
 869                         while (--i >= 0)
 870                                 cfs_free_page(rb->rb_kiov[i].kiov_page);
 871
 872                         LIBCFS_FREE(rb, sz);
 873                         return NULL;
 874                 }
 875
 876                 rb->rb_kiov[i].kiov_len = CFS_PAGE_SIZE;
 877                 rb->rb_kiov[i].kiov_offset = 0;
 878                 rb->rb_kiov[i].kiov_page = page;
 879         }
 880
 881         return rb;
 882 }
 883
 884 void
 885 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
 886 {
 887         int            npages = rbp->rbp_npages;
 888         int            nbuffers = 0;
 889         lnet_rtrbuf_t *rb;
 890
 891         LASSERT (list_empty(&rbp->rbp_msgs));
 892         LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
 893
 894         while (!list_empty(&rbp->rbp_bufs)) {
 895                 LASSERT (rbp->rbp_credits > 0);
 896
 897                 rb = list_entry(rbp->rbp_bufs.next,
 898                                 lnet_rtrbuf_t, rb_list);
 899                 list_del(&rb->rb_list);
 900                 lnet_destroy_rtrbuf(rb, npages);
 901                 nbuffers++;
 902         }
 903
 904         LASSERT (rbp->rbp_nbuffers == nbuffers);
 905         LASSERT (rbp->rbp_credits == nbuffers);
 906
 907         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
 908 }
 909
 910 int
 911 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs)
 912 {
 913         lnet_rtrbuf_t *rb;
 914         int            i;
 915
 916         if (rbp->rbp_nbuffers != 0) {
 917                 LASSERT (rbp->rbp_nbuffers == nbufs);
 918                 return 0;
 919         }
 920
 921         for (i = 0; i < nbufs; i++) {
 922                 rb = lnet_new_rtrbuf(rbp);
 923
 924                 if (rb == NULL) {
 925                         CERROR("Failed to allocate %d router bufs of %d pages\n",
 926                                nbufs, rbp->rbp_npages);
 927                         return -ENOMEM;
 928                 }
 929
 930                 rbp->rbp_nbuffers++;
 931                 rbp->rbp_credits++;
 932                 rbp->rbp_mincredits++;
 933                 list_add(&rb->rb_list, &rbp->rbp_bufs);
 934
 935                 /* No allocation "under fire" */
 936                 /* Otherwise we'd need code to schedule blocked msgs etc */
 937                 LASSERT (!the_lnet.ln_routing);
 938         }
 939
 940         LASSERT (rbp->rbp_credits == nbufs);
 941         return 0;
 942 }
 943
 944 void
 945 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
 946 {
 947         CFS_INIT_LIST_HEAD(&rbp->rbp_msgs);
 948         CFS_INIT_LIST_HEAD(&rbp->rbp_bufs);
 949
 950         rbp->rbp_npages = npages;
 951         rbp->rbp_credits = 0;
 952         rbp->rbp_mincredits = 0;
 953 }
 954
 955 void
 956 lnet_free_rtrpools(void)
 957 {
 958         lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[0]);
 959         lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[1]);
 960         lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[2]);
 961 }
 962
 963 void
 964 lnet_init_rtrpools(void)
 965 {
 966         int small_pages = 1;
 967         int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 968
 969         lnet_rtrpool_init(&the_lnet.ln_rtrpools[0], 0);
 970         lnet_rtrpool_init(&the_lnet.ln_rtrpools[1], small_pages);
 971         lnet_rtrpool_init(&the_lnet.ln_rtrpools[2], large_pages);
 972 }
 973
 974
 975 int
 976 lnet_alloc_rtrpools(int im_a_router)
 977 {
 978         int       rc;
 979
 980         if (!strcmp(forwarding, "")) {
 981                 /* not set either way */
 982                 if (!im_a_router)
 983                         return 0;
 984         } else if (!strcmp(forwarding, "disabled")) {
 985                 /* explicitly disabled */
 986                 return 0;
 987         } else if (!strcmp(forwarding, "enabled")) {
 988                 /* explicitly enabled */
 989         } else {
 990                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
 991                                    "'enabled' or 'disabled'\n");
 992                 return -EINVAL;
 993         }
 994
 995         if (tiny_router_buffers <= 0) {
 996                 LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when "
 997                                    "routing enabled\n", tiny_router_buffers);
 998                 rc = -EINVAL;
 999                 goto failed;
1000         }
1001
1002         rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[0],
1003                                      tiny_router_buffers);
1004         if (rc != 0)
1005                 goto failed;
1006
1007         if (small_router_buffers <= 0) {
1008                 LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when"
1009                                    " routing enabled\n", small_router_buffers);
1010                 rc = -EINVAL;
1011                 goto failed;
1012         }
1013
1014         rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[1],
1015                                      small_router_buffers);
1016         if (rc != 0)
1017                 goto failed;
1018
1019         if (large_router_buffers <= 0) {
1020                 LCONSOLE_ERROR_MSG(0x10e, "large_router_buffers=%d invalid when"
1021                                    " routing enabled\n", large_router_buffers);
1022                 rc = -EINVAL;
1023                 goto failed;
1024         }
1025
1026         rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[2],
1027                                      large_router_buffers);
1028         if (rc != 0)
1029                 goto failed;
1030
1031         LNET_LOCK();
1032         the_lnet.ln_routing = 1;
1033         LNET_UNLOCK();
1034
1035         return 0;
1036
1037  failed:
1038         lnet_free_rtrpools();
1039         return rc;
1040 }
1041
1042 #else
1043
1044 int
1045 lnet_peers_start_down(void)
1046 {
1047         return 0;
1048 }
1049
1050 void
1051 lnet_router_checker_stop(void)
1052 {
1053         return;
1054 }
1055
1056 int
1057 lnet_router_checker_start(void)
1058 {
1059         return 0;
1060 }
1061
1062 void
1063 lnet_free_rtrpools (void)
1064 {
1065 }
1066
1067 void
1068 lnet_init_rtrpools (void)
1069 {
1070 }
1071
1072 int
1073 lnet_alloc_rtrpools (int im_a_arouter)
1074 {
1075         return 0;
1076 }
1077
1078 #endif