lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, Whamcloud, Inc.
   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 static int peer_buffer_credits = 0;
  43 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  44                 "# router buffer credits per peer");
  45
  46 static int auto_down = 1;
  47 CFS_MODULE_PARM(auto_down, "i", int, 0444,
  48                 "Automatically mark peers down on comms error");
  49
  50 int
  51 lnet_peer_buffer_credits(lnet_ni_t *ni)
  52 {
  53         /* NI option overrides LNet default */
  54         if (ni->ni_peerrtrcredits > 0)
  55                 return ni->ni_peerrtrcredits;
  56         if (peer_buffer_credits > 0)
  57                 return peer_buffer_credits;
  58
  59         /* As an approximation, allow this peer the same number of router
  60          * buffers as it is allowed outstanding sends */
  61         return ni->ni_peertxcredits;
  62 }
  63
  64 /* forward ref's */
  65 static int lnet_router_checker(void *);
  66 #else
  67
  68 int
  69 lnet_peer_buffer_credits(lnet_ni_t *ni)
  70 {
  71         return 0;
  72 }
  73
  74 #endif
  75
  76 static int check_routers_before_use = 0;
  77 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
  78                 "Assume routers are down and ping them before use");
  79
  80 static int avoid_asym_router_failure = 0;
  81 CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0444,
  82                 "Avoid asymmetrical failures: reserved, use at your own risk");
  83
  84 static int dead_router_check_interval = 0;
  85 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0444,
  86                 "Seconds between dead router health checks (<= 0 to disable)");
  87
  88 static int live_router_check_interval = 0;
  89 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0444,
  90                 "Seconds between live router health checks (<= 0 to disable)");
  91
  92 static int router_ping_timeout = 50;
  93 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0444,
  94                 "Seconds to wait for the reply to a router health query");
  95
  96 int
  97 lnet_peers_start_down(void)
  98 {
  99         return check_routers_before_use;
 100 }
 101
 102 void
 103 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
 104 {
 105         if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
 106                 CDEBUG(D_NET, "Out of date\n");
 107                 return;
 108         }
 109
 110         lp->lp_timestamp = when;                /* update timestamp */
 111         lp->lp_ping_deadline = 0;               /* disable ping timeout */
 112
 113         if (lp->lp_alive_count != 0 &&          /* got old news */
 114             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
 115                 CDEBUG(D_NET, "Old news\n");
 116                 return;
 117         }
 118
 119         /* Flag that notification is outstanding */
 120
 121         lp->lp_alive_count++;
 122         lp->lp_alive = !(!alive);               /* 1 bit! */
 123         lp->lp_notify = 1;
 124         lp->lp_notifylnd |= notifylnd;
 125
 126         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
 127 }
 128
 129 void
 130 lnet_do_notify (lnet_peer_t *lp)
 131 {
 132         lnet_ni_t *ni = lp->lp_ni;
 133         int        alive;
 134         int        notifylnd;
 135
 136         LNET_LOCK();
 137
 138         /* Notify only in 1 thread at any time to ensure ordered notification.
 139          * NB individual events can be missed; the only guarantee is that you
 140          * always get the most recent news */
 141
 142         if (lp->lp_notifying) {
 143                 LNET_UNLOCK();
 144                 return;
 145         }
 146
 147         lp->lp_notifying = 1;
 148
 149         while (lp->lp_notify) {
 150                 alive     = lp->lp_alive;
 151                 notifylnd = lp->lp_notifylnd;
 152
 153                 lp->lp_notifylnd = 0;
 154                 lp->lp_notify    = 0;
 155
 156                 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
 157                         LNET_UNLOCK();
 158
 159                         /* A new notification could happen now; I'll handle it
 160                          * when control returns to me */
 161
 162                         (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
 163
 164                         LNET_LOCK();
 165                 }
 166         }
 167
 168         lp->lp_notifying = 0;
 169
 170         LNET_UNLOCK();
 171 }
 172
 173
 174 static void
 175 lnet_rtr_addref_locked(lnet_peer_t *lp)
 176 {
 177         LASSERT (lp->lp_refcount > 0);
 178         LASSERT (lp->lp_rtr_refcount >= 0);
 179
 180         lp->lp_rtr_refcount++;
 181         if (lp->lp_rtr_refcount == 1) {
 182                 cfs_list_t *pos;
 183
 184                 /* a simple insertion sort */
 185                 cfs_list_for_each_prev(pos, &the_lnet.ln_routers) {
 186                         lnet_peer_t *rtr = cfs_list_entry(pos, lnet_peer_t,
 187                                                           lp_rtr_list);
 188
 189                         if (rtr->lp_nid < lp->lp_nid)
 190                                 break;
 191                 }
 192
 193                 cfs_list_add(&lp->lp_rtr_list, pos);
 194                 /* addref for the_lnet.ln_routers */
 195                 lnet_peer_addref_locked(lp);
 196                 the_lnet.ln_routers_version++;
 197         }
 198 }
 199
 200 static void
 201 lnet_rtr_decref_locked(lnet_peer_t *lp)
 202 {
 203         LASSERT (lp->lp_refcount > 0);
 204         LASSERT (lp->lp_rtr_refcount > 0);
 205
 206         lp->lp_rtr_refcount--;
 207         if (lp->lp_rtr_refcount == 0) {
 208                 if (lp->lp_rcd != NULL) {
 209                         cfs_list_add(&lp->lp_rcd->rcd_list,
 210                                      &the_lnet.ln_zombie_rcd);
 211                         lp->lp_rcd = NULL;
 212                 }
 213
 214                 cfs_list_del(&lp->lp_rtr_list);
 215                 /* decref for the_lnet.ln_routers */
 216                 lnet_peer_decref_locked(lp);
 217                 the_lnet.ln_routers_version++;
 218         }
 219 }
 220
 221 lnet_remotenet_t *
 222 lnet_find_net_locked (__u32 net)
 223 {
 224         lnet_remotenet_t *rnet;
 225         cfs_list_t       *tmp;
 226
 227         LASSERT (!the_lnet.ln_shutdown);
 228
 229         cfs_list_for_each (tmp, &the_lnet.ln_remote_nets) {
 230                 rnet = cfs_list_entry(tmp, lnet_remotenet_t, lrn_list);
 231
 232                 if (rnet->lrn_net == net)
 233                         return rnet;
 234         }
 235         return NULL;
 236 }
 237
 238 static void lnet_shuffle_seed(void)
 239 {
 240         static int seeded = 0;
 241         int lnd_type, seed[2];
 242         struct timeval tv;
 243         lnet_ni_t *ni;
 244         cfs_list_t *tmp;
 245
 246         if (seeded)
 247                 return;
 248
 249         cfs_get_random_bytes(seed, sizeof(seed));
 250
 251         /* Nodes with small feet have little entropy
 252          * the NID for this node gives the most entropy in the low bits */
 253         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 254                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 255                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 256
 257                 if (lnd_type != LOLND)
 258                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
 259         }
 260
 261         cfs_gettimeofday(&tv);
 262         cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
 263         seeded = 1;
 264         return;
 265 }
 266
 267 /* NB expects LNET_LOCK held */
 268 void
 269 lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route)
 270 {
 271         unsigned int      len = 0;
 272         unsigned int      offset = 0;
 273         cfs_list_t       *e;
 274
 275         lnet_shuffle_seed();
 276
 277         cfs_list_for_each (e, &rnet->lrn_routes) {
 278                 len++;
 279         }
 280
 281         /* len+1 positions to add a new entry, also prevents division by 0 */
 282         offset = cfs_rand() % (len + 1);
 283         cfs_list_for_each (e, &rnet->lrn_routes) {
 284                 if (offset == 0)
 285                         break;
 286                 offset--;
 287         }
 288         cfs_list_add(&route->lr_list, e);
 289
 290         the_lnet.ln_remote_nets_version++;
 291         lnet_rtr_addref_locked(route->lr_gateway);
 292 }
 293
 294 int
 295 lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway)
 296 {
 297         cfs_list_t          *e;
 298         lnet_remotenet_t    *rnet;
 299         lnet_remotenet_t    *rnet2;
 300         lnet_route_t        *route;
 301         lnet_ni_t           *ni;
 302         int                  add_route;
 303         int                  rc;
 304
 305         CDEBUG(D_NET, "Add route: net %s hops %u gw %s\n",
 306                libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 307
 308         if (gateway == LNET_NID_ANY ||
 309             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 310             net == LNET_NIDNET(LNET_NID_ANY) ||
 311             LNET_NETTYP(net) == LOLND ||
 312             LNET_NIDNET(gateway) == net ||
 313             hops < 1 || hops > 255)
 314                 return (-EINVAL);
 315
 316         if (lnet_islocalnet(net))               /* it's a local network */
 317                 return 0;                       /* ignore the route entry */
 318
 319         /* Assume net, route, all new */
 320         LIBCFS_ALLOC(route, sizeof(*route));
 321         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 322         if (route == NULL || rnet == NULL) {
 323                 CERROR("Out of memory creating route %s %d %s\n",
 324                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 325                 if (route != NULL)
 326                         LIBCFS_FREE(route, sizeof(*route));
 327                 if (rnet != NULL)
 328                         LIBCFS_FREE(rnet, sizeof(*rnet));
 329                 return -ENOMEM;
 330         }
 331
 332         CFS_INIT_LIST_HEAD(&rnet->lrn_routes);
 333         rnet->lrn_net = net;
 334         route->lr_hops = hops;
 335
 336         LNET_LOCK();
 337
 338         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway);
 339         if (rc != 0) {
 340                 LNET_UNLOCK();
 341
 342                 LIBCFS_FREE(route, sizeof(*route));
 343                 LIBCFS_FREE(rnet, sizeof(*rnet));
 344
 345                 if (rc == -EHOSTUNREACH)        /* gateway is not on a local net */
 346                         return 0;               /* ignore the route entry */
 347
 348                 CERROR("Error %d creating route %s %d %s\n", rc,
 349                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 350                 return rc;
 351         }
 352
 353         LASSERT (!the_lnet.ln_shutdown);
 354
 355         rnet2 = lnet_find_net_locked(net);
 356         if (rnet2 == NULL) {
 357                 /* new network */
 358                 cfs_list_add_tail(&rnet->lrn_list, &the_lnet.ln_remote_nets);
 359                 rnet2 = rnet;
 360         }
 361
 362         /* Search for a duplicate route (it's a NOOP if it is) */
 363         add_route = 1;
 364         cfs_list_for_each (e, &rnet2->lrn_routes) {
 365                 lnet_route_t *route2 = cfs_list_entry(e, lnet_route_t, lr_list);
 366
 367                 if (route2->lr_gateway == route->lr_gateway) {
 368                         add_route = 0;
 369                         break;
 370                 }
 371
 372                 /* our lookups must be true */
 373                 LASSERT (route2->lr_gateway->lp_nid != gateway);
 374         }
 375
 376         if (add_route) {
 377                 ni = route->lr_gateway->lp_ni;
 378                 lnet_ni_addref_locked(ni);
 379
 380                 lnet_add_route_to_rnet(rnet2, route);
 381                 LNET_UNLOCK();
 382
 383                 /* XXX Assume alive */
 384                 if (ni->ni_lnd->lnd_notify != NULL)
 385                         (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
 386
 387                 lnet_ni_decref(ni);
 388         } else {
 389                 lnet_peer_decref_locked(route->lr_gateway);
 390                 LNET_UNLOCK();
 391                 LIBCFS_FREE(route, sizeof(*route));
 392         }
 393
 394         if (rnet != rnet2)
 395                 LIBCFS_FREE(rnet, sizeof(*rnet));
 396
 397         return 0;
 398 }
 399
 400 int
 401 lnet_check_routes (void)
 402 {
 403         lnet_remotenet_t    *rnet;
 404         lnet_route_t        *route;
 405         lnet_route_t        *route2;
 406         cfs_list_t          *e1;
 407         cfs_list_t          *e2;
 408
 409         LNET_LOCK();
 410
 411         cfs_list_for_each (e1, &the_lnet.ln_remote_nets) {
 412                 rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 413
 414                 route2 = NULL;
 415                 cfs_list_for_each (e2, &rnet->lrn_routes) {
 416                         route = cfs_list_entry(e2, lnet_route_t, lr_list);
 417
 418                         if (route2 == NULL)
 419                                 route2 = route;
 420                         else if (route->lr_gateway->lp_ni !=
 421                                  route2->lr_gateway->lp_ni) {
 422                                 LNET_UNLOCK();
 423
 424                                 CERROR("Routes to %s via %s and %s not supported\n",
 425                                        libcfs_net2str(rnet->lrn_net),
 426                                        libcfs_nid2str(route->lr_gateway->lp_nid),
 427                                        libcfs_nid2str(route2->lr_gateway->lp_nid));
 428                                 return -EINVAL;
 429                         }
 430                 }
 431         }
 432
 433         LNET_UNLOCK();
 434         return 0;
 435 }
 436
 437 int
 438 lnet_del_route (__u32 net, lnet_nid_t gw_nid)
 439 {
 440         lnet_remotenet_t    *rnet;
 441         lnet_route_t        *route;
 442         cfs_list_t          *e1;
 443         cfs_list_t          *e2;
 444         int                  rc = -ENOENT;
 445
 446         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 447                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 448
 449         /* NB Caller may specify either all routes via the given gateway
 450          * or a specific route entry actual NIDs) */
 451
 452  again:
 453         LNET_LOCK();
 454
 455         cfs_list_for_each (e1, &the_lnet.ln_remote_nets) {
 456                 rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 457
 458                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 459                       net == rnet->lrn_net))
 460                         continue;
 461
 462                 cfs_list_for_each (e2, &rnet->lrn_routes) {
 463                         route = cfs_list_entry(e2, lnet_route_t, lr_list);
 464
 465                         if (!(gw_nid == LNET_NID_ANY ||
 466                               gw_nid == route->lr_gateway->lp_nid))
 467                                 continue;
 468
 469                         cfs_list_del(&route->lr_list);
 470                         the_lnet.ln_remote_nets_version++;
 471
 472                         if (cfs_list_empty(&rnet->lrn_routes))
 473                                 cfs_list_del(&rnet->lrn_list);
 474                         else
 475                                 rnet = NULL;
 476
 477                         lnet_rtr_decref_locked(route->lr_gateway);
 478                         lnet_peer_decref_locked(route->lr_gateway);
 479                         LNET_UNLOCK();
 480
 481                         LIBCFS_FREE(route, sizeof (*route));
 482
 483                         if (rnet != NULL)
 484                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 485
 486                         rc = 0;
 487                         goto again;
 488                 }
 489         }
 490
 491         LNET_UNLOCK();
 492         return rc;
 493 }
 494
 495 void
 496 lnet_destroy_routes (void)
 497 {
 498         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 499 }
 500
 501 int
 502 lnet_get_route (int idx, __u32 *net, __u32 *hops,
 503                lnet_nid_t *gateway, __u32 *alive)
 504 {
 505         cfs_list_t          *e1;
 506         cfs_list_t          *e2;
 507         lnet_remotenet_t    *rnet;
 508         lnet_route_t        *route;
 509
 510         LNET_LOCK();
 511
 512         cfs_list_for_each (e1, &the_lnet.ln_remote_nets) {
 513                 rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 514
 515                 cfs_list_for_each (e2, &rnet->lrn_routes) {
 516                         route = cfs_list_entry(e2, lnet_route_t, lr_list);
 517
 518                         if (idx-- == 0) {
 519                                 *net     = rnet->lrn_net;
 520                                 *hops    = route->lr_hops;
 521                                 *gateway = route->lr_gateway->lp_nid;
 522                                 *alive   = route->lr_gateway->lp_alive;
 523                                 LNET_UNLOCK();
 524                                 return 0;
 525                         }
 526                 }
 527         }
 528
 529         LNET_UNLOCK();
 530         return -ENOENT;
 531 }
 532
 533 void
 534 lnet_swap_pinginfo(lnet_ping_info_t *info)
 535 {
 536         int               i;
 537         lnet_ni_status_t *stat;
 538
 539         __swab32s(&info->pi_version);
 540         __swab32s(&info->pi_pid);
 541         __swab32s(&info->pi_nnis);
 542         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 543                 stat = &info->pi_ni[i];
 544                 __swab64s(&stat->ns_nid);
 545                 __swab32s(&stat->ns_status);
 546         }
 547         return;
 548 }
 549
 550 /* Returns # of down NIs, or negative error codes; ignore downed NIs
 551  * if a NI in 'net' is up */
 552 int
 553 lnet_router_down_ni(lnet_peer_t *rtr, __u32 net)
 554 {
 555         int               i;
 556         int               down = 0;
 557         int               ptl_up = 0;
 558         int               ptl_down = 0;
 559         lnet_ping_info_t *info;
 560
 561         if (!avoid_asym_router_failure)
 562                 return -ENOENT;
 563
 564         if (rtr->lp_rcd == NULL)
 565                 return -EINVAL;
 566
 567         if (!rtr->lp_alive)
 568                 return -EINVAL;  /* stale lp_rcd */
 569
 570         info = rtr->lp_rcd->rcd_pinginfo;
 571         LASSERT (info != NULL);
 572
 573         /* NB always racing with network! */
 574         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
 575                 lnet_swap_pinginfo(info);
 576         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 577                 CNETERR("%s: Unexpected magic %08x\n",
 578                         libcfs_nid2str(rtr->lp_nid), info->pi_magic);
 579                 return -EPROTO;
 580         }
 581
 582         if (info->pi_version == LNET_PROTO_PING_VERSION1)
 583                 return -ENOENT;  /* v1 doesn't carry NI status info */
 584
 585         if (info->pi_version != LNET_PROTO_PING_VERSION) {
 586                 CNETERR("%s: Unexpected version 0x%x\n",
 587                         libcfs_nid2str(rtr->lp_nid), info->pi_version);
 588                 return -EPROTO;
 589         }
 590
 591         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 592                 lnet_ni_status_t *stat = &info->pi_ni[i];
 593                 lnet_nid_t        nid = stat->ns_nid;
 594
 595                 if (nid == LNET_NID_ANY) {
 596                         CNETERR("%s: unexpected LNET_NID_ANY\n",
 597                                 libcfs_nid2str(rtr->lp_nid));
 598                         return -EPROTO;
 599                 }
 600
 601                 if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 602                         continue;
 603
 604                 if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 605                         if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
 606                                 ptl_down = 1;
 607                         else
 608                                 down++;
 609                         continue;
 610                 }
 611
 612                 if (stat->ns_status != LNET_NI_STATUS_UP) {
 613                         CNETERR("%s: Unexpected status 0x%x\n",
 614                                 libcfs_nid2str(rtr->lp_nid), stat->ns_status);
 615                         return -EPROTO;
 616                 }
 617
 618                 /* ignore downed NIs if there's a NI up for dest network */
 619                 if (LNET_NIDNET(nid) == net)
 620                         return 0;
 621
 622                 if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
 623                         ptl_up = 1;
 624         }
 625
 626         /* ptl NIs are considered down only when they're all down */
 627         return down + (ptl_up ? 0 : ptl_down);
 628 }
 629
 630 void
 631 lnet_wait_known_routerstate(void)
 632 {
 633         lnet_peer_t         *rtr;
 634         cfs_list_t          *entry;
 635         int                  all_known;
 636
 637         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 638
 639         for (;;) {
 640                 LNET_LOCK();
 641
 642                 all_known = 1;
 643                 cfs_list_for_each (entry, &the_lnet.ln_routers) {
 644                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
 645
 646                         if (rtr->lp_alive_count == 0) {
 647                                 all_known = 0;
 648                                 break;
 649                         }
 650                 }
 651
 652                 LNET_UNLOCK();
 653
 654                 if (all_known)
 655                         return;
 656
 657 #ifndef __KERNEL__
 658                 lnet_router_checker();
 659 #endif
 660                 cfs_pause(cfs_time_seconds(1));
 661         }
 662 }
 663
 664 static void
 665 lnet_router_checker_event (lnet_event_t *event)
 666 {
 667         /* CAVEAT EMPTOR: I'm called with LNET_LOCKed and I'm not allowed to
 668          * drop it (that's how come I see _every_ event, even ones that would
 669          * overflow my EQ) */
 670         lnet_rc_data_t *rcd = event->md.user_ptr;
 671         lnet_peer_t    *lp;
 672         lnet_nid_t      nid;
 673
 674         if (event->unlinked) {
 675                 if (rcd != NULL) {
 676                         LNetInvalidateHandle(&rcd->rcd_mdh);
 677                         return;
 678                 }
 679
 680                 /* The router checker thread has unlinked the default rc_md
 681                  * and exited. */
 682                 LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING);
 683                 the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKED;
 684 #ifdef __KERNEL__
 685                 cfs_up(&the_lnet.ln_rc_signal);
 686 #endif
 687                 return;
 688         }
 689
 690         LASSERT (event->type == LNET_EVENT_SEND ||
 691                  event->type == LNET_EVENT_REPLY);
 692
 693         nid = (event->type == LNET_EVENT_SEND) ?
 694               event->target.nid : event->initiator.nid;
 695
 696         lp = lnet_find_peer_locked(nid);
 697         if (lp == NULL) {
 698                 /* router may have been removed */
 699                 CDEBUG(D_NET, "Router %s not found\n", libcfs_nid2str(nid));
 700                 return;
 701         }
 702
 703         if (event->type == LNET_EVENT_SEND)     /* re-enable another ping */
 704                 lp->lp_ping_notsent = 0;
 705
 706         if (lnet_isrouter(lp) &&                /* ignore if no longer a router */
 707             (event->status != 0 ||
 708              event->type == LNET_EVENT_REPLY)) {
 709
 710                 /* A successful REPLY means the router is up.  If _any_ comms
 711                  * to the router fail I assume it's down (this will happen if
 712                  * we ping alive routers to try to detect router death before
 713                  * apps get burned). */
 714
 715                 lnet_notify_locked(lp, 1, (event->status == 0),
 716                                    cfs_time_current());
 717
 718                 /* The router checker will wake up very shortly and do the
 719                  * actual notification.
 720                  * XXX If 'lp' stops being a router before then, it will still
 721                  * have the notification pending!!! */
 722         }
 723
 724         /* This decref will NOT drop LNET_LOCK (it had to have 1 ref when it
 725          * was in the peer table and I've not dropped the lock, so no-one else
 726          * can have reduced the refcount) */
 727         LASSERT(lp->lp_refcount > 1);
 728
 729         lnet_peer_decref_locked(lp);
 730 }
 731
 732 void
 733 lnet_update_ni_status(void)
 734 {
 735         cfs_time_t now = cfs_time_current();
 736         lnet_ni_t *ni;
 737         int        status;
 738         int        timeout;
 739
 740         LASSERT (the_lnet.ln_routing);
 741
 742         timeout = router_ping_timeout +
 743                   MAX(live_router_check_interval, dead_router_check_interval);
 744
 745         LNET_LOCK();
 746
 747         cfs_list_for_each_entry (ni, &the_lnet.ln_nis, ni_list) {
 748                 lnet_ni_status_t *ns = ni->ni_status;
 749
 750                 LASSERT (ns != NULL);
 751
 752                 status = LNET_NI_STATUS_UP;
 753                 if (ni->ni_lnd->lnd_type != LOLND &&  /* @lo forever alive */
 754                     cfs_time_after(now, cfs_time_add(ni->ni_last_alive,
 755                                                      cfs_time_seconds(timeout))))
 756                         status = LNET_NI_STATUS_DOWN;
 757
 758                 if (ns->ns_status != status) {
 759                         ns->ns_status = status;
 760                         CDEBUG(D_NET, "NI(%s:%d) status changed to %s\n",
 761                                libcfs_nid2str(ni->ni_nid), timeout,
 762                                status == LNET_NI_STATUS_UP ? "up" : "down");
 763                 }
 764         }
 765
 766         LNET_UNLOCK();
 767 }
 768
 769 void
 770 lnet_destroy_rc_data (lnet_rc_data_t *rcd)
 771 {
 772         LASSERT (cfs_list_empty(&rcd->rcd_list));
 773         /* detached from network */
 774         LASSERT (LNetHandleIsInvalid(rcd->rcd_mdh));
 775
 776         LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 777         LIBCFS_FREE(rcd, sizeof(*rcd));
 778         return;
 779 }
 780
 781 lnet_rc_data_t *
 782 lnet_create_rc_data (void)
 783 {
 784         int               i;
 785         int               rc;
 786         lnet_ping_info_t *pi;
 787         lnet_rc_data_t   *rcd;
 788
 789         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 790         if (rcd == NULL)
 791                 return NULL;
 792
 793         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 794         if (pi == NULL) {
 795                 LIBCFS_FREE(rcd, sizeof(*rcd));
 796                 return NULL;
 797         }
 798
 799         memset(pi, 0, LNET_PINGINFO_SIZE);
 800         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 801                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 802                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 803         }
 804         rcd->rcd_pinginfo = pi;
 805         LNetInvalidateHandle(&rcd->rcd_mdh);
 806         CFS_INIT_LIST_HEAD(&rcd->rcd_list);
 807
 808         LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 809         rc = LNetMDBind((lnet_md_t){.start     = pi,
 810                                     .user_ptr  = rcd,
 811                                     .length    = LNET_PINGINFO_SIZE,
 812                                     .threshold = LNET_MD_THRESH_INF,
 813                                     .options   = LNET_MD_TRUNCATE,
 814                                     .eq_handle = the_lnet.ln_rc_eqh},
 815                         LNET_UNLINK,
 816                         &rcd->rcd_mdh);
 817         if (rc < 0) {
 818                 CERROR("Can't bind MD: %d\n", rc);
 819                 lnet_destroy_rc_data(rcd);
 820                 return NULL;
 821         }
 822         LASSERT (rc == 0);
 823         return rcd;
 824 }
 825
 826 static int
 827 lnet_router_check_interval (lnet_peer_t *rtr)
 828 {
 829         int secs;
 830
 831         secs = rtr->lp_alive ? live_router_check_interval :
 832                                dead_router_check_interval;
 833         if (secs < 0)
 834                 secs = 0;
 835
 836         return secs;
 837 }
 838
 839 static void
 840 lnet_ping_router_locked (lnet_peer_t *rtr)
 841 {
 842         int             newrcd = 0;
 843         lnet_rc_data_t *rcd = NULL;
 844         cfs_time_t      now = cfs_time_current();
 845         int             secs;
 846
 847         lnet_peer_addref_locked(rtr);
 848
 849         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
 850             cfs_time_after(now, rtr->lp_ping_deadline))
 851                 lnet_notify_locked(rtr, 1, 0, now);
 852
 853         if (avoid_asym_router_failure && rtr->lp_rcd == NULL)
 854                 newrcd = 1;
 855
 856         LNET_UNLOCK();
 857
 858         /* Run any outstanding notifications */
 859         lnet_do_notify(rtr);
 860
 861         if (newrcd)
 862                 rcd = lnet_create_rc_data();
 863
 864         LNET_LOCK();
 865
 866         if (!lnet_isrouter(rtr)) {
 867                 lnet_peer_decref_locked(rtr);
 868                 if (rcd != NULL)
 869                         cfs_list_add(&rcd->rcd_list, &the_lnet.ln_zombie_rcd);
 870                 return; /* router table changed! */
 871         }
 872
 873         if (rcd != NULL) {
 874                 LASSERT (rtr->lp_rcd == NULL);
 875                 rtr->lp_rcd = rcd;
 876         }
 877
 878         secs = lnet_router_check_interval(rtr);
 879
 880         CDEBUG(D_NET,
 881                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
 882                "alive_count %d lp_ping_timestamp %lu\n",
 883                libcfs_nid2str(rtr->lp_nid), secs,
 884                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
 885                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
 886
 887         if (secs != 0 && !rtr->lp_ping_notsent &&
 888             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
 889                                              cfs_time_seconds(secs)))) {
 890                 int               rc;
 891                 lnet_process_id_t id;
 892                 lnet_handle_md_t  mdh;
 893
 894                 id.nid = rtr->lp_nid;
 895                 id.pid = LUSTRE_SRV_LNET_PID;
 896                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
 897
 898                 rtr->lp_ping_notsent   = 1;
 899                 rtr->lp_ping_timestamp = now;
 900                 mdh = (rtr->lp_rcd == NULL) ? the_lnet.ln_rc_mdh :
 901                                               rtr->lp_rcd->rcd_mdh;
 902
 903                 if (rtr->lp_ping_deadline == 0)
 904                         rtr->lp_ping_deadline = cfs_time_shift(router_ping_timeout);
 905
 906                 LNET_UNLOCK();
 907
 908                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
 909                              LNET_PROTO_PING_MATCHBITS, 0);
 910
 911                 LNET_LOCK();
 912                 if (rc != 0)
 913                         rtr->lp_ping_notsent = 0; /* no event pending */
 914         }
 915
 916         lnet_peer_decref_locked(rtr);
 917         return;
 918 }
 919
 920 int
 921 lnet_router_checker_start(void)
 922 {
 923         static lnet_ping_info_t pinginfo;
 924
 925         lnet_md_t    md;
 926         int          rc;
 927         int          eqsz;
 928 #ifndef __KERNEL__
 929         lnet_peer_t *rtr;
 930         __u64        version;
 931         int          nrtr = 0;
 932         int          router_checker_max_eqsize = 10240;
 933
 934         LASSERT (check_routers_before_use);
 935         LASSERT (dead_router_check_interval > 0);
 936
 937         LNET_LOCK();
 938
 939         /* As an approximation, allow each router the same number of
 940          * outstanding events as it is allowed outstanding sends */
 941         eqsz = 0;
 942         version = the_lnet.ln_routers_version;
 943         cfs_list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
 944                 lnet_ni_t         *ni = rtr->lp_ni;
 945                 lnet_process_id_t  id;
 946
 947                 nrtr++;
 948                 eqsz += ni->ni_peertxcredits;
 949
 950                 /* one async ping reply per router */
 951                 id.nid = rtr->lp_nid;
 952                 id.pid = LUSTRE_SRV_LNET_PID;
 953
 954                 LNET_UNLOCK();
 955
 956                 rc = LNetSetAsync(id, 1);
 957                 if (rc != 0) {
 958                         CWARN("LNetSetAsync %s failed: %d\n",
 959                               libcfs_id2str(id), rc);
 960                         return rc;
 961                 }
 962
 963                 LNET_LOCK();
 964                 /* NB router list doesn't change in userspace */
 965                 LASSERT (version == the_lnet.ln_routers_version);
 966         }
 967
 968         LNET_UNLOCK();
 969
 970         if (nrtr == 0) {
 971                 CDEBUG(D_NET,
 972                        "No router found, not starting router checker\n");
 973                 return 0;
 974         }
 975
 976         /* at least allow a SENT and a REPLY per router */
 977         if (router_checker_max_eqsize < 2 * nrtr)
 978                 router_checker_max_eqsize = 2 * nrtr;
 979
 980         LASSERT (eqsz > 0);
 981         if (eqsz > router_checker_max_eqsize)
 982                 eqsz = router_checker_max_eqsize;
 983 #endif
 984
 985         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
 986
 987         if (check_routers_before_use &&
 988             dead_router_check_interval <= 0) {
 989                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
 990                                    " set if 'check_routers_before_use' is set"
 991                                    "\n");
 992                 return -EINVAL;
 993         }
 994
 995         if (!the_lnet.ln_routing &&
 996             live_router_check_interval <= 0 &&
 997             dead_router_check_interval <= 0)
 998                 return 0;
 999
1000 #ifdef __KERNEL__
1001         cfs_sema_init(&the_lnet.ln_rc_signal, 0);
1002         /* EQ size doesn't matter; the callback is guaranteed to get every
1003          * event */
1004         eqsz = 1;
1005         rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1006                          &the_lnet.ln_rc_eqh);
1007 #else
1008         rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE,
1009                          &the_lnet.ln_rc_eqh);
1010 #endif
1011         if (rc != 0) {
1012                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1013                 return -ENOMEM;
1014         }
1015
1016         memset(&md, 0, sizeof(md));
1017         md.user_ptr  = NULL;
1018         md.start     = &pinginfo;
1019         md.length    = sizeof(pinginfo);
1020         md.options   = LNET_MD_TRUNCATE;
1021         md.threshold = LNET_MD_THRESH_INF;
1022         md.eq_handle = the_lnet.ln_rc_eqh;
1023         rc = LNetMDBind(md, LNET_UNLINK, &the_lnet.ln_rc_mdh);
1024         if (rc < 0) {
1025                 CERROR("Can't bind MD: %d\n", rc);
1026                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1027                 LASSERT (rc == 0);
1028                 return -ENOMEM;
1029         }
1030         LASSERT (rc == 0);
1031
1032         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1033 #ifdef __KERNEL__
1034         rc = cfs_create_thread(lnet_router_checker, NULL, 0);
1035         if (rc < 0) {
1036                 CERROR("Can't start router checker thread: %d\n", rc);
1037                 the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
1038                 rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
1039                 LASSERT (rc == 0);
1040                 /* block until event callback signals exit */
1041                 cfs_down(&the_lnet.ln_rc_signal);
1042                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1043                 LASSERT (rc == 0);
1044                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1045                 return -ENOMEM;
1046         }
1047 #endif
1048
1049         if (check_routers_before_use) {
1050                 /* Note that a helpful side-effect of pinging all known routers
1051                  * at startup is that it makes them drop stale connections they
1052                  * may have to a previous instance of me. */
1053                 lnet_wait_known_routerstate();
1054         }
1055
1056         return 0;
1057 }
1058
1059 void
1060 lnet_router_checker_stop (void)
1061 {
1062         int rc;
1063
1064         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1065                 return;
1066
1067         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1068         the_lnet.ln_rc_state = LNET_RC_STATE_STOPTHREAD;
1069
1070 #ifdef __KERNEL__
1071         /* block until event callback signals exit */
1072         cfs_down(&the_lnet.ln_rc_signal);
1073 #else
1074         while (the_lnet.ln_rc_state != LNET_RC_STATE_UNLINKED) {
1075                 lnet_router_checker();
1076                 cfs_pause(cfs_time_seconds(1));
1077         }
1078 #endif
1079         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED);
1080
1081         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1082         LASSERT (rc == 0);
1083         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1084         return;
1085 }
1086
1087 #if defined(__KERNEL__) && defined(LNET_ROUTER)
1088
1089 static void
1090 lnet_prune_zombie_rcd (int wait_unlink)
1091 {
1092         lnet_rc_data_t   *rcd;
1093         lnet_rc_data_t   *tmp;
1094         cfs_list_t        free_rcd;
1095         int               i;
1096         __u64             version;
1097
1098         CFS_INIT_LIST_HEAD(&free_rcd);
1099
1100         LNET_LOCK();
1101 rescan:
1102         version = the_lnet.ln_routers_version;
1103         cfs_list_for_each_entry_safe (rcd, tmp, &the_lnet.ln_zombie_rcd,
1104                                       rcd_list) {
1105                 if (LNetHandleIsInvalid(rcd->rcd_mdh)) {
1106                         cfs_list_del(&rcd->rcd_list);
1107                         cfs_list_add(&rcd->rcd_list, &free_rcd);
1108                         continue;
1109                 }
1110
1111                 LNET_UNLOCK();
1112
1113                 LNetMDUnlink(rcd->rcd_mdh);
1114
1115                 LNET_LOCK();
1116                 if (version != the_lnet.ln_routers_version)
1117                         goto rescan;
1118         }
1119
1120         i = 2;
1121         while (wait_unlink && !cfs_list_empty(&the_lnet.ln_zombie_rcd)) {
1122                 rcd = cfs_list_entry(the_lnet.ln_zombie_rcd.next,
1123                                      lnet_rc_data_t, rcd_list);
1124                 if (LNetHandleIsInvalid(rcd->rcd_mdh)) {
1125                         cfs_list_del(&rcd->rcd_list);
1126                         cfs_list_add(&rcd->rcd_list, &free_rcd);
1127                         continue;
1128                 }
1129
1130                 LNET_UNLOCK();
1131
1132                 LNetMDUnlink(rcd->rcd_mdh);
1133
1134                 i++;
1135                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1136                        "Waiting for rc buffers to unlink\n");
1137                 cfs_pause(cfs_time_seconds(1));
1138
1139                 LNET_LOCK();
1140         }
1141
1142         LNET_UNLOCK();
1143
1144         while (!cfs_list_empty(&free_rcd)) {
1145                 rcd = cfs_list_entry(free_rcd.next, lnet_rc_data_t, rcd_list);
1146                 cfs_list_del_init(&rcd->rcd_list);
1147                 lnet_destroy_rc_data(rcd);
1148         }
1149         return;
1150 }
1151
1152 static int
1153 lnet_router_checker(void *arg)
1154 {
1155         int                rc;
1156         lnet_peer_t       *rtr;
1157         cfs_list_t        *entry;
1158
1159         cfs_daemonize("router_checker");
1160         cfs_block_allsigs();
1161
1162         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1163
1164         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1165                 __u64 version;
1166
1167                 LNET_LOCK();
1168 rescan:
1169                 version = the_lnet.ln_routers_version;
1170
1171                 cfs_list_for_each (entry, &the_lnet.ln_routers) {
1172                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
1173                         lnet_ping_router_locked(rtr);
1174
1175                         /* NB dropped lock */
1176                         if (version != the_lnet.ln_routers_version) {
1177                                 /* the routers list has changed */
1178                                 goto rescan;
1179                         }
1180                 }
1181
1182                 LNET_UNLOCK();
1183
1184                 if (the_lnet.ln_routing)
1185                         lnet_update_ni_status();
1186
1187                 lnet_prune_zombie_rcd(0); /* don't wait for UNLINK */
1188
1189                 /* Call cfs_pause() here always adds 1 to load average
1190                  * because kernel counts # active tasks as nr_running
1191                  * + nr_uninterruptible. */
1192                 cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE,
1193                                                    cfs_time_seconds(1));
1194         }
1195
1196         LNET_LOCK();
1197
1198         cfs_list_for_each (entry, &the_lnet.ln_routers) {
1199                 rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
1200
1201                 if (rtr->lp_rcd == NULL)
1202                         continue;
1203
1204                 LASSERT (cfs_list_empty(&rtr->lp_rcd->rcd_list));
1205                 cfs_list_add(&rtr->lp_rcd->rcd_list, &the_lnet.ln_zombie_rcd);
1206                 rtr->lp_rcd = NULL;
1207         }
1208
1209         LNET_UNLOCK();
1210
1211         lnet_prune_zombie_rcd(1); /* wait for UNLINK */
1212
1213         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD);
1214         the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
1215
1216         rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
1217         LASSERT (rc == 0);
1218
1219         /* The unlink event callback will signal final completion */
1220         return 0;
1221 }
1222
1223 void
1224 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1225 {
1226         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1227
1228         while (--npages >= 0)
1229                 cfs_free_page(rb->rb_kiov[npages].kiov_page);
1230
1231         LIBCFS_FREE(rb, sz);
1232 }
1233
1234 lnet_rtrbuf_t *
1235 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp)
1236 {
1237         int            npages = rbp->rbp_npages;
1238         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1239         struct page   *page;
1240         lnet_rtrbuf_t *rb;
1241         int            i;
1242
1243         LIBCFS_ALLOC(rb, sz);
1244         if (rb == NULL)
1245                 return NULL;
1246
1247         rb->rb_pool = rbp;
1248
1249         for (i = 0; i < npages; i++) {
1250                 page = cfs_alloc_page(CFS_ALLOC_ZERO | CFS_ALLOC_STD);
1251                 if (page == NULL) {
1252                         while (--i >= 0)
1253                                 cfs_free_page(rb->rb_kiov[i].kiov_page);
1254
1255                         LIBCFS_FREE(rb, sz);
1256                         return NULL;
1257                 }
1258
1259                 rb->rb_kiov[i].kiov_len = CFS_PAGE_SIZE;
1260                 rb->rb_kiov[i].kiov_offset = 0;
1261                 rb->rb_kiov[i].kiov_page = page;
1262         }
1263
1264         return rb;
1265 }
1266
1267 void
1268 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
1269 {
1270         int            npages = rbp->rbp_npages;
1271         int            nbuffers = 0;
1272         lnet_rtrbuf_t *rb;
1273
1274         LASSERT (cfs_list_empty(&rbp->rbp_msgs));
1275         LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
1276
1277         while (!cfs_list_empty(&rbp->rbp_bufs)) {
1278                 LASSERT (rbp->rbp_credits > 0);
1279
1280                 rb = cfs_list_entry(rbp->rbp_bufs.next,
1281                                     lnet_rtrbuf_t, rb_list);
1282                 cfs_list_del(&rb->rb_list);
1283                 lnet_destroy_rtrbuf(rb, npages);
1284                 nbuffers++;
1285         }
1286
1287         LASSERT (rbp->rbp_nbuffers == nbuffers);
1288         LASSERT (rbp->rbp_credits == nbuffers);
1289
1290         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1291 }
1292
1293 int
1294 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs)
1295 {
1296         lnet_rtrbuf_t *rb;
1297         int            i;
1298
1299         if (rbp->rbp_nbuffers != 0) {
1300                 LASSERT (rbp->rbp_nbuffers == nbufs);
1301                 return 0;
1302         }
1303
1304         for (i = 0; i < nbufs; i++) {
1305                 rb = lnet_new_rtrbuf(rbp);
1306
1307                 if (rb == NULL) {
1308                         CERROR("Failed to allocate %d router bufs of %d pages\n",
1309                                nbufs, rbp->rbp_npages);
1310                         return -ENOMEM;
1311                 }
1312
1313                 rbp->rbp_nbuffers++;
1314                 rbp->rbp_credits++;
1315                 rbp->rbp_mincredits++;
1316                 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1317
1318                 /* No allocation "under fire" */
1319                 /* Otherwise we'd need code to schedule blocked msgs etc */
1320                 LASSERT (!the_lnet.ln_routing);
1321         }
1322
1323         LASSERT (rbp->rbp_credits == nbufs);
1324         return 0;
1325 }
1326
1327 void
1328 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1329 {
1330         CFS_INIT_LIST_HEAD(&rbp->rbp_msgs);
1331         CFS_INIT_LIST_HEAD(&rbp->rbp_bufs);
1332
1333         rbp->rbp_npages = npages;
1334         rbp->rbp_credits = 0;
1335         rbp->rbp_mincredits = 0;
1336 }
1337
1338 void
1339 lnet_free_rtrpools(void)
1340 {
1341         lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[0]);
1342         lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[1]);
1343         lnet_rtrpool_free_bufs(&the_lnet.ln_rtrpools[2]);
1344 }
1345
1346 void
1347 lnet_init_rtrpools(void)
1348 {
1349         int small_pages = 1;
1350         int large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1351
1352         lnet_rtrpool_init(&the_lnet.ln_rtrpools[0], 0);
1353         lnet_rtrpool_init(&the_lnet.ln_rtrpools[1], small_pages);
1354         lnet_rtrpool_init(&the_lnet.ln_rtrpools[2], large_pages);
1355 }
1356
1357
1358 int
1359 lnet_alloc_rtrpools(int im_a_router)
1360 {
1361         int       rc;
1362
1363         if (!strcmp(forwarding, "")) {
1364                 /* not set either way */
1365                 if (!im_a_router)
1366                         return 0;
1367         } else if (!strcmp(forwarding, "disabled")) {
1368                 /* explicitly disabled */
1369                 return 0;
1370         } else if (!strcmp(forwarding, "enabled")) {
1371                 /* explicitly enabled */
1372         } else {
1373                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1374                                    "'enabled' or 'disabled'\n");
1375                 return -EINVAL;
1376         }
1377
1378         if (tiny_router_buffers <= 0) {
1379                 LCONSOLE_ERROR_MSG(0x10c, "tiny_router_buffers=%d invalid when "
1380                                    "routing enabled\n", tiny_router_buffers);
1381                 rc = -EINVAL;
1382                 goto failed;
1383         }
1384
1385         rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[0],
1386                                      tiny_router_buffers);
1387         if (rc != 0)
1388                 goto failed;
1389
1390         if (small_router_buffers <= 0) {
1391                 LCONSOLE_ERROR_MSG(0x10d, "small_router_buffers=%d invalid when"
1392                                    " routing enabled\n", small_router_buffers);
1393                 rc = -EINVAL;
1394                 goto failed;
1395         }
1396
1397         rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[1],
1398                                      small_router_buffers);
1399         if (rc != 0)
1400                 goto failed;
1401
1402         if (large_router_buffers <= 0) {
1403                 LCONSOLE_ERROR_MSG(0x10e, "large_router_buffers=%d invalid when"
1404                                    " routing enabled\n", large_router_buffers);
1405                 rc = -EINVAL;
1406                 goto failed;
1407         }
1408
1409         rc = lnet_rtrpool_alloc_bufs(&the_lnet.ln_rtrpools[2],
1410                                      large_router_buffers);
1411         if (rc != 0)
1412                 goto failed;
1413
1414         LNET_LOCK();
1415         the_lnet.ln_routing = 1;
1416         LNET_UNLOCK();
1417
1418         return 0;
1419
1420  failed:
1421         lnet_free_rtrpools();
1422         return rc;
1423 }
1424
1425 int
1426 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1427 {
1428         lnet_peer_t *lp = NULL;
1429         cfs_time_t   now = cfs_time_current();
1430
1431         LASSERT (!cfs_in_interrupt ());
1432
1433         CDEBUG (D_NET, "%s notifying %s: %s\n",
1434                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1435                 libcfs_nid2str(nid),
1436                 alive ? "up" : "down");
1437
1438         if (ni != NULL &&
1439             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1440                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1441                         libcfs_nid2str(nid), alive ? "birth" : "death",
1442                         libcfs_nid2str(ni->ni_nid));
1443                 return -EINVAL;
1444         }
1445
1446         /* can't do predictions... */
1447         if (cfs_time_after(when, now)) {
1448                 CWARN ("Ignoring prediction from %s of %s %s "
1449                        "%ld seconds in the future\n",
1450                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1451                        libcfs_nid2str(nid), alive ? "up" : "down",
1452                        cfs_duration_sec(cfs_time_sub(when, now)));
1453                 return -EINVAL;
1454         }
1455
1456         if (ni != NULL && !alive &&             /* LND telling me she's down */
1457             !auto_down) {                       /* auto-down disabled */
1458                 CDEBUG(D_NET, "Auto-down disabled\n");
1459                 return 0;
1460         }
1461
1462         LNET_LOCK();
1463
1464         lp = lnet_find_peer_locked(nid);
1465         if (lp == NULL) {
1466                 /* nid not found */
1467                 LNET_UNLOCK();
1468                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1469                 return 0;
1470         }
1471
1472         /* We can't fully trust LND on reporting exact peer last_alive
1473          * if he notifies us about dead peer. For example ksocklnd can
1474          * call us with when == _time_when_the_node_was_booted_ if
1475          * no connections were successfully established */
1476         if (ni != NULL && !alive && when < lp->lp_last_alive)
1477                 when = lp->lp_last_alive;
1478
1479         lnet_notify_locked(lp, ni == NULL, alive, when);
1480
1481         LNET_UNLOCK();
1482
1483         lnet_do_notify(lp);
1484
1485         LNET_LOCK();
1486
1487         lnet_peer_decref_locked(lp);
1488
1489         LNET_UNLOCK();
1490         return 0;
1491 }
1492 EXPORT_SYMBOL(lnet_notify);
1493
1494 void
1495 lnet_get_tunables (void)
1496 {
1497         return;
1498 }
1499
1500 #else
1501
1502 int
1503 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1504 {
1505         return -EOPNOTSUPP;
1506 }
1507
1508 void
1509 lnet_router_checker (void)
1510 {
1511         static time_t last = 0;
1512         static int    running = 0;
1513
1514         time_t            now = cfs_time_current_sec();
1515         int               interval = now - last;
1516         int               rc;
1517         __u64             version;
1518         lnet_peer_t      *rtr;
1519
1520         /* It's no use to call me again within a sec - all intervals and
1521          * timeouts are measured in seconds */
1522         if (last != 0 && interval < 2)
1523                 return;
1524
1525         if (last != 0 &&
1526             interval > MAX(live_router_check_interval,
1527                            dead_router_check_interval))
1528                 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1529                         live_router_check_interval, dead_router_check_interval,
1530                         interval);
1531
1532         LNET_LOCK();
1533         LASSERT (!running); /* recursion check */
1534         running = 1;
1535         LNET_UNLOCK();
1536
1537         last = now;
1538
1539         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPTHREAD) {
1540                 the_lnet.ln_rc_state = LNET_RC_STATE_UNLINKING;
1541                 rc = LNetMDUnlink(the_lnet.ln_rc_mdh);
1542                 LASSERT (rc == 0);
1543         }
1544
1545         /* consume all pending events */
1546         while (1) {
1547                 int          i;
1548                 lnet_event_t ev;
1549
1550                 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1551                  * recursion breaker in LNetEQPoll would fail */
1552                 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1553                 if (rc == 0)   /* no event pending */
1554                         break;
1555
1556                 /* NB a lost SENT prevents me from pinging a router again */
1557                 if (rc == -EOVERFLOW) {
1558                         CERROR("Dropped an event!!!\n");
1559                         abort();
1560                 }
1561
1562                 LASSERT (rc == 1);
1563
1564                 LNET_LOCK();
1565                 lnet_router_checker_event(&ev);
1566                 LNET_UNLOCK();
1567         }
1568
1569         if (the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKED ||
1570             the_lnet.ln_rc_state == LNET_RC_STATE_UNLINKING) {
1571                 running = 0;
1572                 return;
1573         }
1574
1575         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1576
1577         LNET_LOCK();
1578
1579         version = the_lnet.ln_routers_version;
1580         cfs_list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) {
1581                 lnet_ping_router_locked(rtr);
1582                 LASSERT (version == the_lnet.ln_routers_version);
1583         }
1584
1585         LNET_UNLOCK();
1586
1587         running = 0; /* lock only needed for the recursion check */
1588         return;
1589 }
1590
1591 /* NB lnet_peers_start_down depends on me,
1592  * so must be called before any peer creation */
1593 void
1594 lnet_get_tunables (void)
1595 {
1596         char *s;
1597
1598         s = getenv("LNET_ROUTER_PING_TIMEOUT");
1599         if (s != NULL) router_ping_timeout = atoi(s);
1600
1601         s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1602         if (s != NULL) live_router_check_interval = atoi(s);
1603
1604         s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1605         if (s != NULL) dead_router_check_interval = atoi(s);
1606
1607         /* This replaces old lnd_notify mechanism */
1608         check_routers_before_use = 1;
1609         if (dead_router_check_interval <= 0)
1610                 dead_router_check_interval = 30;
1611 }
1612
1613 void
1614 lnet_free_rtrpools (void)
1615 {
1616 }
1617
1618 void
1619 lnet_init_rtrpools (void)
1620 {
1621 }
1622
1623 int
1624 lnet_alloc_rtrpools (int im_a_arouter)
1625 {
1626         return 0;
1627 }
1628
1629 #endif