lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2013, Intel Corporation.
   5  *
   6  *   This file is part of Portals
   7  *   http://sourceforge.net/projects/sandiaportals/
   8  *
   9  *   Portals is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Portals is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Portals; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  *
  22  */
  23
  24 #define DEBUG_SUBSYSTEM S_LNET
  25 #include <lnet/lib-lnet.h>
  26
  27 #if defined(__KERNEL__) && defined(LNET_ROUTER)
  28
  29 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
  30 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
  31 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
  32 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
  33 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  34 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  35
  36 static char *forwarding = "";
  37 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
  38                 "Explicitly enable/disable forwarding between networks");
  39
  40 static int tiny_router_buffers;
  41 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
  42                 "# of 0 payload messages to buffer in the router");
  43 static int small_router_buffers;
  44 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
  45                 "# of small (1 page) messages to buffer in the router");
  46 static int large_router_buffers;
  47 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
  48                 "# of large messages to buffer in the router");
  49 static int peer_buffer_credits = 0;
  50 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  51                 "# router buffer credits per peer");
  52
  53 static int auto_down = 1;
  54 CFS_MODULE_PARM(auto_down, "i", int, 0444,
  55                 "Automatically mark peers down on comms error");
  56
  57 int
  58 lnet_peer_buffer_credits(lnet_ni_t *ni)
  59 {
  60         /* NI option overrides LNet default */
  61         if (ni->ni_peerrtrcredits > 0)
  62                 return ni->ni_peerrtrcredits;
  63         if (peer_buffer_credits > 0)
  64                 return peer_buffer_credits;
  65
  66         /* As an approximation, allow this peer the same number of router
  67          * buffers as it is allowed outstanding sends */
  68         return ni->ni_peertxcredits;
  69 }
  70
  71 /* forward ref's */
  72 static int lnet_router_checker(void *);
  73 #else
  74
  75 int
  76 lnet_peer_buffer_credits(lnet_ni_t *ni)
  77 {
  78         return 0;
  79 }
  80
  81 #endif
  82
  83 static int check_routers_before_use = 0;
  84 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
  85                 "Assume routers are down and ping them before use");
  86
  87 static int avoid_asym_router_failure = 1;
  88 CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644,
  89                 "Avoid asymmetrical router failures (0 to disable)");
  90
  91 static int dead_router_check_interval = 60;
  92 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644,
  93                 "Seconds between dead router health checks (<= 0 to disable)");
  94
  95 static int live_router_check_interval = 60;
  96 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644,
  97                 "Seconds between live router health checks (<= 0 to disable)");
  98
  99 static int router_ping_timeout = 50;
 100 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644,
 101                 "Seconds to wait for the reply to a router health query");
 102
 103 int
 104 lnet_peers_start_down(void)
 105 {
 106         return check_routers_before_use;
 107 }
 108
 109 void
 110 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
 111 {
 112         if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
 113                 CDEBUG(D_NET, "Out of date\n");
 114                 return;
 115         }
 116
 117         lp->lp_timestamp = when;                /* update timestamp */
 118         lp->lp_ping_deadline = 0;               /* disable ping timeout */
 119
 120         if (lp->lp_alive_count != 0 &&          /* got old news */
 121             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
 122                 CDEBUG(D_NET, "Old news\n");
 123                 return;
 124         }
 125
 126         /* Flag that notification is outstanding */
 127
 128         lp->lp_alive_count++;
 129         lp->lp_alive = !(!alive);               /* 1 bit! */
 130         lp->lp_notify = 1;
 131         lp->lp_notifylnd |= notifylnd;
 132         if (lp->lp_alive)
 133                 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 134
 135         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
 136 }
 137
 138 void
 139 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
 140 {
 141         int        alive;
 142         int        notifylnd;
 143
 144         /* Notify only in 1 thread at any time to ensure ordered notification.
 145          * NB individual events can be missed; the only guarantee is that you
 146          * always get the most recent news */
 147
 148         if (lp->lp_notifying)
 149                 return;
 150
 151         lp->lp_notifying = 1;
 152
 153         while (lp->lp_notify) {
 154                 alive     = lp->lp_alive;
 155                 notifylnd = lp->lp_notifylnd;
 156
 157                 lp->lp_notifylnd = 0;
 158                 lp->lp_notify    = 0;
 159
 160                 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
 161                         lnet_net_unlock(lp->lp_cpt);
 162
 163                         /* A new notification could happen now; I'll handle it
 164                          * when control returns to me */
 165
 166                         (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
 167
 168                         lnet_net_lock(lp->lp_cpt);
 169                 }
 170         }
 171
 172         lp->lp_notifying = 0;
 173 }
 174
 175
 176 static void
 177 lnet_rtr_addref_locked(lnet_peer_t *lp)
 178 {
 179         LASSERT(lp->lp_refcount > 0);
 180         LASSERT(lp->lp_rtr_refcount >= 0);
 181
 182         /* lnet_net_lock must be exclusively locked */
 183         lp->lp_rtr_refcount++;
 184         if (lp->lp_rtr_refcount == 1) {
 185                 cfs_list_t *pos;
 186
 187                 /* a simple insertion sort */
 188                 cfs_list_for_each_prev(pos, &the_lnet.ln_routers) {
 189                         lnet_peer_t *rtr = cfs_list_entry(pos, lnet_peer_t,
 190                                                           lp_rtr_list);
 191
 192                         if (rtr->lp_nid < lp->lp_nid)
 193                                 break;
 194                 }
 195
 196                 cfs_list_add(&lp->lp_rtr_list, pos);
 197                 /* addref for the_lnet.ln_routers */
 198                 lnet_peer_addref_locked(lp);
 199                 the_lnet.ln_routers_version++;
 200         }
 201 }
 202
 203 static void
 204 lnet_rtr_decref_locked(lnet_peer_t *lp)
 205 {
 206         LASSERT(lp->lp_refcount > 0);
 207         LASSERT(lp->lp_rtr_refcount > 0);
 208
 209         /* lnet_net_lock must be exclusively locked */
 210         lp->lp_rtr_refcount--;
 211         if (lp->lp_rtr_refcount == 0) {
 212                 LASSERT(cfs_list_empty(&lp->lp_routes));
 213
 214                 if (lp->lp_rcd != NULL) {
 215                         cfs_list_add(&lp->lp_rcd->rcd_list,
 216                                      &the_lnet.ln_rcd_deathrow);
 217                         lp->lp_rcd = NULL;
 218                 }
 219
 220                 cfs_list_del(&lp->lp_rtr_list);
 221                 /* decref for the_lnet.ln_routers */
 222                 lnet_peer_decref_locked(lp);
 223                 the_lnet.ln_routers_version++;
 224         }
 225 }
 226
 227 lnet_remotenet_t *
 228 lnet_find_net_locked (__u32 net)
 229 {
 230         lnet_remotenet_t        *rnet;
 231         cfs_list_t              *tmp;
 232         cfs_list_t              *rn_list;
 233
 234         LASSERT(!the_lnet.ln_shutdown);
 235
 236         rn_list = lnet_net2rnethash(net);
 237         cfs_list_for_each(tmp, rn_list) {
 238                 rnet = cfs_list_entry(tmp, lnet_remotenet_t, lrn_list);
 239
 240                 if (rnet->lrn_net == net)
 241                         return rnet;
 242         }
 243         return NULL;
 244 }
 245
 246 static void lnet_shuffle_seed(void)
 247 {
 248         static int seeded = 0;
 249         int lnd_type, seed[2];
 250         struct timeval tv;
 251         lnet_ni_t *ni;
 252         cfs_list_t *tmp;
 253
 254         if (seeded)
 255                 return;
 256
 257         cfs_get_random_bytes(seed, sizeof(seed));
 258
 259         /* Nodes with small feet have little entropy
 260          * the NID for this node gives the most entropy in the low bits */
 261         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 262                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 263                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 264
 265                 if (lnd_type != LOLND)
 266                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
 267         }
 268
 269         cfs_gettimeofday(&tv);
 270         cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
 271         seeded = 1;
 272         return;
 273 }
 274
 275 /* NB expects LNET_LOCK held */
 276 void
 277 lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route)
 278 {
 279         unsigned int      len = 0;
 280         unsigned int      offset = 0;
 281         cfs_list_t       *e;
 282
 283         lnet_shuffle_seed();
 284
 285         cfs_list_for_each (e, &rnet->lrn_routes) {
 286                 len++;
 287         }
 288
 289         /* len+1 positions to add a new entry, also prevents division by 0 */
 290         offset = cfs_rand() % (len + 1);
 291         cfs_list_for_each (e, &rnet->lrn_routes) {
 292                 if (offset == 0)
 293                         break;
 294                 offset--;
 295         }
 296         cfs_list_add(&route->lr_list, e);
 297         cfs_list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 298
 299         the_lnet.ln_remote_nets_version++;
 300         lnet_rtr_addref_locked(route->lr_gateway);
 301 }
 302
 303 int
 304 lnet_add_route (__u32 net, unsigned int hops, lnet_nid_t gateway)
 305 {
 306         cfs_list_t          *e;
 307         lnet_remotenet_t    *rnet;
 308         lnet_remotenet_t    *rnet2;
 309         lnet_route_t        *route;
 310         lnet_ni_t           *ni;
 311         int                  add_route;
 312         int                  rc;
 313
 314         CDEBUG(D_NET, "Add route: net %s hops %u gw %s\n",
 315                libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 316
 317         if (gateway == LNET_NID_ANY ||
 318             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 319             net == LNET_NIDNET(LNET_NID_ANY) ||
 320             LNET_NETTYP(net) == LOLND ||
 321             LNET_NIDNET(gateway) == net ||
 322             hops < 1 || hops > 255)
 323                 return (-EINVAL);
 324
 325         if (lnet_islocalnet(net))               /* it's a local network */
 326                 return 0;                       /* ignore the route entry */
 327
 328         /* Assume net, route, all new */
 329         LIBCFS_ALLOC(route, sizeof(*route));
 330         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 331         if (route == NULL || rnet == NULL) {
 332                 CERROR("Out of memory creating route %s %d %s\n",
 333                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 334                 if (route != NULL)
 335                         LIBCFS_FREE(route, sizeof(*route));
 336                 if (rnet != NULL)
 337                         LIBCFS_FREE(rnet, sizeof(*rnet));
 338                 return -ENOMEM;
 339         }
 340
 341         CFS_INIT_LIST_HEAD(&rnet->lrn_routes);
 342         rnet->lrn_net = net;
 343         route->lr_hops = hops;
 344         route->lr_net = net;
 345
 346         lnet_net_lock(LNET_LOCK_EX);
 347
 348         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
 349         if (rc != 0) {
 350                 lnet_net_unlock(LNET_LOCK_EX);
 351
 352                 LIBCFS_FREE(route, sizeof(*route));
 353                 LIBCFS_FREE(rnet, sizeof(*rnet));
 354
 355                 if (rc == -EHOSTUNREACH) { /* gateway is not on a local net */
 356                         return 0;       /* ignore the route entry */
 357                 } else {
 358                         CERROR("Error %d creating route %s %d %s\n", rc,
 359                                libcfs_net2str(net), hops,
 360                                libcfs_nid2str(gateway));
 361                 }
 362                 return rc;
 363         }
 364
 365         LASSERT (!the_lnet.ln_shutdown);
 366
 367         rnet2 = lnet_find_net_locked(net);
 368         if (rnet2 == NULL) {
 369                 /* new network */
 370                 cfs_list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 371                 rnet2 = rnet;
 372         }
 373
 374         /* Search for a duplicate route (it's a NOOP if it is) */
 375         add_route = 1;
 376         cfs_list_for_each (e, &rnet2->lrn_routes) {
 377                 lnet_route_t *route2 = cfs_list_entry(e, lnet_route_t, lr_list);
 378
 379                 if (route2->lr_gateway == route->lr_gateway) {
 380                         add_route = 0;
 381                         break;
 382                 }
 383
 384                 /* our lookups must be true */
 385                 LASSERT (route2->lr_gateway->lp_nid != gateway);
 386         }
 387
 388         if (add_route) {
 389                 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
 390                 lnet_add_route_to_rnet(rnet2, route);
 391
 392                 ni = route->lr_gateway->lp_ni;
 393                 lnet_net_unlock(LNET_LOCK_EX);
 394
 395                 /* XXX Assume alive */
 396                 if (ni->ni_lnd->lnd_notify != NULL)
 397                         (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
 398
 399                 lnet_net_lock(LNET_LOCK_EX);
 400         }
 401
 402         /* -1 for notify or !add_route */
 403         lnet_peer_decref_locked(route->lr_gateway);
 404         lnet_net_unlock(LNET_LOCK_EX);
 405
 406         if (!add_route)
 407                 LIBCFS_FREE(route, sizeof(*route));
 408
 409         if (rnet != rnet2)
 410                 LIBCFS_FREE(rnet, sizeof(*rnet));
 411
 412         return 0;
 413 }
 414
 415 int
 416 lnet_check_routes(void)
 417 {
 418         lnet_remotenet_t        *rnet;
 419         lnet_route_t            *route;
 420         lnet_route_t            *route2;
 421         cfs_list_t              *e1;
 422         cfs_list_t              *e2;
 423         int                     cpt;
 424         cfs_list_t              *rn_list;
 425         int                     i;
 426
 427         cpt = lnet_net_lock_current();
 428
 429         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 430                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 431                 cfs_list_for_each(e1, rn_list) {
 432                         rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 433
 434                         route2 = NULL;
 435                         cfs_list_for_each(e2, &rnet->lrn_routes) {
 436                                 lnet_nid_t      nid1;
 437                                 lnet_nid_t      nid2;
 438                                 int             net;
 439
 440                                 route = cfs_list_entry(e2, lnet_route_t,
 441                                                        lr_list);
 442
 443                                 if (route2 == NULL) {
 444                                         route2 = route;
 445                                         continue;
 446                                 }
 447
 448                                 if (route->lr_gateway->lp_ni ==
 449                                     route2->lr_gateway->lp_ni)
 450                                         continue;
 451
 452                                 nid1 = route->lr_gateway->lp_nid;
 453                                 nid2 = route2->lr_gateway->lp_nid;
 454                                 net = rnet->lrn_net;
 455
 456                                 lnet_net_unlock(cpt);
 457
 458                                 CERROR("Routes to %s via %s and %s not "
 459                                        "supported\n",
 460                                        libcfs_net2str(net),
 461                                        libcfs_nid2str(nid1),
 462                                        libcfs_nid2str(nid2));
 463                                 return -EINVAL;
 464                         }
 465                 }
 466         }
 467
 468         lnet_net_unlock(cpt);
 469         return 0;
 470 }
 471
 472 int
 473 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 474 {
 475         struct lnet_peer        *gateway;
 476         lnet_remotenet_t        *rnet;
 477         lnet_route_t            *route;
 478         cfs_list_t              *e1;
 479         cfs_list_t              *e2;
 480         int                     rc = -ENOENT;
 481         cfs_list_t              *rn_list;
 482         int                     idx = 0;
 483
 484         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 485                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 486
 487         /* NB Caller may specify either all routes via the given gateway
 488          * or a specific route entry actual NIDs) */
 489
 490         lnet_net_lock(LNET_LOCK_EX);
 491         if (net == LNET_NIDNET(LNET_NID_ANY))
 492                 rn_list = &the_lnet.ln_remote_nets_hash[0];
 493         else
 494                 rn_list = lnet_net2rnethash(net);
 495
 496  again:
 497         cfs_list_for_each(e1, rn_list) {
 498                 rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 499
 500                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 501                         net == rnet->lrn_net))
 502                         continue;
 503
 504                 cfs_list_for_each(e2, &rnet->lrn_routes) {
 505                         route = cfs_list_entry(e2, lnet_route_t, lr_list);
 506
 507                         gateway = route->lr_gateway;
 508                         if (!(gw_nid == LNET_NID_ANY ||
 509                               gw_nid == gateway->lp_nid))
 510                                 continue;
 511
 512                         cfs_list_del(&route->lr_list);
 513                         cfs_list_del(&route->lr_gwlist);
 514                         the_lnet.ln_remote_nets_version++;
 515
 516                         if (cfs_list_empty(&rnet->lrn_routes))
 517                                 cfs_list_del(&rnet->lrn_list);
 518                         else
 519                                 rnet = NULL;
 520
 521                         lnet_rtr_decref_locked(gateway);
 522                         lnet_peer_decref_locked(gateway);
 523
 524                         lnet_net_unlock(LNET_LOCK_EX);
 525
 526                         LIBCFS_FREE(route, sizeof(*route));
 527
 528                         if (rnet != NULL)
 529                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 530
 531                         rc = 0;
 532                         lnet_net_lock(LNET_LOCK_EX);
 533                         goto again;
 534                 }
 535         }
 536
 537         if (net == LNET_NIDNET(LNET_NID_ANY) &&
 538             ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
 539                 rn_list = &the_lnet.ln_remote_nets_hash[idx];
 540                 goto again;
 541         }
 542         lnet_net_unlock(LNET_LOCK_EX);
 543
 544         return rc;
 545 }
 546
 547 void
 548 lnet_destroy_routes (void)
 549 {
 550         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 551 }
 552
 553 int
 554 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 555                lnet_nid_t *gateway, __u32 *alive)
 556 {
 557         cfs_list_t              *e1;
 558         cfs_list_t              *e2;
 559         lnet_remotenet_t        *rnet;
 560         lnet_route_t            *route;
 561         int                     cpt;
 562         int                     i;
 563         cfs_list_t              *rn_list;
 564
 565         cpt = lnet_net_lock_current();
 566
 567         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 568                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 569                 cfs_list_for_each(e1, rn_list) {
 570                         rnet = cfs_list_entry(e1, lnet_remotenet_t, lrn_list);
 571
 572                         cfs_list_for_each(e2, &rnet->lrn_routes) {
 573                                 route = cfs_list_entry(e2, lnet_route_t,
 574                                                        lr_list);
 575
 576                                 if (idx-- == 0) {
 577                                         *net     = rnet->lrn_net;
 578                                         *hops    = route->lr_hops;
 579                                         *gateway = route->lr_gateway->lp_nid;
 580                                         *alive   = route->lr_gateway->lp_alive;
 581                                         lnet_net_unlock(cpt);
 582                                         return 0;
 583                                 }
 584                         }
 585                 }
 586         }
 587
 588         lnet_net_unlock(cpt);
 589         return -ENOENT;
 590 }
 591
 592 void
 593 lnet_swap_pinginfo(lnet_ping_info_t *info)
 594 {
 595         int               i;
 596         lnet_ni_status_t *stat;
 597
 598         __swab32s(&info->pi_magic);
 599         __swab32s(&info->pi_features);
 600         __swab32s(&info->pi_pid);
 601         __swab32s(&info->pi_nnis);
 602         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 603                 stat = &info->pi_ni[i];
 604                 __swab64s(&stat->ns_nid);
 605                 __swab32s(&stat->ns_status);
 606         }
 607         return;
 608 }
 609
 610 /**
 611  * parse router-checker pinginfo, record number of down NIs for remote
 612  * networks on that router.
 613  */
 614 static void
 615 lnet_parse_rc_info(lnet_rc_data_t *rcd)
 616 {
 617         lnet_ping_info_t        *info = rcd->rcd_pinginfo;
 618         struct lnet_peer        *gw   = rcd->rcd_gateway;
 619         lnet_route_t            *rtr;
 620
 621         if (!gw->lp_alive)
 622                 return;
 623
 624         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 625                 lnet_swap_pinginfo(info);
 626
 627         /* NB always racing with network! */
 628         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 629                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 630                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
 631                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 632                 return;
 633         }
 634
 635         gw->lp_ping_feats = info->pi_features;
 636         if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
 637                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
 638                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
 639                 return; /* nothing I can understand */
 640         }
 641
 642         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
 643                 return; /* can't carry NI status info */
 644
 645         cfs_list_for_each_entry(rtr, &gw->lp_routes, lr_gwlist) {
 646                 int     ptl_status = LNET_NI_STATUS_INVALID;
 647                 int     down = 0;
 648                 int     up = 0;
 649                 int     i;
 650
 651                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 652                         lnet_ni_status_t *stat = &info->pi_ni[i];
 653                         lnet_nid_t       nid = stat->ns_nid;
 654
 655                         if (nid == LNET_NID_ANY) {
 656                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 657                                        libcfs_nid2str(gw->lp_nid));
 658                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 659                                 return;
 660                         }
 661
 662                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 663                                 continue;
 664
 665                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 666                                 if (LNET_NETTYP(LNET_NIDNET(nid)) != PTLLND)
 667                                         down++;
 668                                 else if (ptl_status != LNET_NI_STATUS_UP)
 669                                         ptl_status = LNET_NI_STATUS_DOWN;
 670                                 continue;
 671                         }
 672
 673                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 674                                 if (LNET_NIDNET(nid) == rtr->lr_net) {
 675                                         up = 1;
 676                                         break;
 677                                 }
 678                                 /* ptl NIs are considered down only when
 679                                  * they're all down */
 680                                 if (LNET_NETTYP(LNET_NIDNET(nid)) == PTLLND)
 681                                         ptl_status = LNET_NI_STATUS_UP;
 682                                 continue;
 683                         }
 684
 685                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 686                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
 687                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 688                         return;
 689                 }
 690
 691                 if (up) { /* ignore downed NIs if NI for dest network is up */
 692                         rtr->lr_downis = 0;
 693                         continue;
 694                 }
 695                 rtr->lr_downis = down + (ptl_status == LNET_NI_STATUS_DOWN);
 696         }
 697 }
 698
 699 static void
 700 lnet_router_checker_event(lnet_event_t *event)
 701 {
 702         lnet_rc_data_t          *rcd = event->md.user_ptr;
 703         struct lnet_peer        *lp;
 704
 705         LASSERT(rcd != NULL);
 706
 707         if (event->unlinked) {
 708                 LNetInvalidateHandle(&rcd->rcd_mdh);
 709                 return;
 710         }
 711
 712         LASSERT(event->type == LNET_EVENT_SEND ||
 713                 event->type == LNET_EVENT_REPLY);
 714
 715         lp = rcd->rcd_gateway;
 716         LASSERT(lp != NULL);
 717
 718          /* NB: it's called with holding lnet_res_lock, we have a few
 719           * places need to hold both locks at the same time, please take
 720           * care of lock ordering */
 721         lnet_net_lock(lp->lp_cpt);
 722         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
 723                 /* ignore if no longer a router or rcd is replaced */
 724                 goto out;
 725         }
 726
 727         if (event->type == LNET_EVENT_SEND) {
 728                 lp->lp_ping_notsent = 0;
 729                 if (event->status == 0)
 730                         goto out;
 731         }
 732
 733         /* LNET_EVENT_REPLY */
 734         /* A successful REPLY means the router is up.  If _any_ comms
 735          * to the router fail I assume it's down (this will happen if
 736          * we ping alive routers to try to detect router death before
 737          * apps get burned). */
 738
 739         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
 740         /* The router checker will wake up very shortly and do the
 741          * actual notification.
 742          * XXX If 'lp' stops being a router before then, it will still
 743          * have the notification pending!!! */
 744
 745         if (avoid_asym_router_failure && event->status == 0)
 746                 lnet_parse_rc_info(rcd);
 747
 748  out:
 749         lnet_net_unlock(lp->lp_cpt);
 750 }
 751
 752 void
 753 lnet_wait_known_routerstate(void)
 754 {
 755         lnet_peer_t         *rtr;
 756         cfs_list_t          *entry;
 757         int                  all_known;
 758
 759         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 760
 761         for (;;) {
 762                 int     cpt = lnet_net_lock_current();
 763
 764                 all_known = 1;
 765                 cfs_list_for_each (entry, &the_lnet.ln_routers) {
 766                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
 767
 768                         if (rtr->lp_alive_count == 0) {
 769                                 all_known = 0;
 770                                 break;
 771                         }
 772                 }
 773
 774                 lnet_net_unlock(cpt);
 775
 776                 if (all_known)
 777                         return;
 778
 779 #ifndef __KERNEL__
 780                 lnet_router_checker();
 781 #endif
 782                 cfs_pause(cfs_time_seconds(1));
 783         }
 784 }
 785
 786 void
 787 lnet_update_ni_status_locked(void)
 788 {
 789         lnet_ni_t       *ni;
 790         long            now;
 791         int             timeout;
 792
 793         LASSERT(the_lnet.ln_routing);
 794
 795         timeout = router_ping_timeout +
 796                   MAX(live_router_check_interval, dead_router_check_interval);
 797
 798         now = cfs_time_current_sec();
 799         cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 800                 if (ni->ni_lnd->lnd_type == LOLND)
 801                         continue;
 802
 803                 if (now < ni->ni_last_alive + timeout)
 804                         continue;
 805
 806                 lnet_ni_lock(ni);
 807                 /* re-check with lock */
 808                 if (now < ni->ni_last_alive + timeout) {
 809                         lnet_ni_unlock(ni);
 810                         continue;
 811                 }
 812
 813                 LASSERT(ni->ni_status != NULL);
 814
 815                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 816                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 817                                libcfs_nid2str(ni->ni_nid), timeout);
 818                         /* NB: so far, this is the only place to set
 819                          * NI status to "down" */
 820                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 821                 }
 822                 lnet_ni_unlock(ni);
 823         }
 824 }
 825
 826 void
 827 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
 828 {
 829         LASSERT(cfs_list_empty(&rcd->rcd_list));
 830         /* detached from network */
 831         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 832
 833         if (rcd->rcd_gateway != NULL) {
 834                 int cpt = rcd->rcd_gateway->lp_cpt;
 835
 836                 lnet_net_lock(cpt);
 837                 lnet_peer_decref_locked(rcd->rcd_gateway);
 838                 lnet_net_unlock(cpt);
 839         }
 840
 841         if (rcd->rcd_pinginfo != NULL)
 842                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 843
 844         LIBCFS_FREE(rcd, sizeof(*rcd));
 845 }
 846
 847 lnet_rc_data_t *
 848 lnet_create_rc_data_locked(lnet_peer_t *gateway)
 849 {
 850         lnet_rc_data_t          *rcd = NULL;
 851         lnet_ping_info_t        *pi;
 852         int                     rc;
 853         int                     i;
 854
 855         lnet_net_unlock(gateway->lp_cpt);
 856
 857         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 858         if (rcd == NULL)
 859                 goto out;
 860
 861         LNetInvalidateHandle(&rcd->rcd_mdh);
 862         CFS_INIT_LIST_HEAD(&rcd->rcd_list);
 863
 864         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 865         if (pi == NULL)
 866                 goto out;
 867
 868         memset(pi, 0, LNET_PINGINFO_SIZE);
 869         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 870                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 871                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 872         }
 873         rcd->rcd_pinginfo = pi;
 874
 875         LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 876         rc = LNetMDBind((lnet_md_t){.start     = pi,
 877                                     .user_ptr  = rcd,
 878                                     .length    = LNET_PINGINFO_SIZE,
 879                                     .threshold = LNET_MD_THRESH_INF,
 880                                     .options   = LNET_MD_TRUNCATE,
 881                                     .eq_handle = the_lnet.ln_rc_eqh},
 882                         LNET_UNLINK,
 883                         &rcd->rcd_mdh);
 884         if (rc < 0) {
 885                 CERROR("Can't bind MD: %d\n", rc);
 886                 goto out;
 887         }
 888         LASSERT(rc == 0);
 889
 890         lnet_net_lock(gateway->lp_cpt);
 891         /* router table changed or someone has created rcd for this gateway */
 892         if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
 893                 lnet_net_unlock(gateway->lp_cpt);
 894                 goto out;
 895         }
 896
 897         lnet_peer_addref_locked(gateway);
 898         rcd->rcd_gateway = gateway;
 899         gateway->lp_rcd = rcd;
 900         gateway->lp_ping_notsent = 0;
 901
 902         return rcd;
 903
 904  out:
 905         if (rcd != NULL) {
 906                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
 907                         rc = LNetMDUnlink(rcd->rcd_mdh);
 908                         LASSERT(rc == 0);
 909                 }
 910                 lnet_destroy_rc_data(rcd);
 911         }
 912
 913         lnet_net_lock(gateway->lp_cpt);
 914         return gateway->lp_rcd;
 915 }
 916
 917 static int
 918 lnet_router_check_interval (lnet_peer_t *rtr)
 919 {
 920         int secs;
 921
 922         secs = rtr->lp_alive ? live_router_check_interval :
 923                                dead_router_check_interval;
 924         if (secs < 0)
 925                 secs = 0;
 926
 927         return secs;
 928 }
 929
 930 static void
 931 lnet_ping_router_locked (lnet_peer_t *rtr)
 932 {
 933         lnet_rc_data_t *rcd = NULL;
 934         cfs_time_t      now = cfs_time_current();
 935         int             secs;
 936
 937         lnet_peer_addref_locked(rtr);
 938
 939         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
 940             cfs_time_after(now, rtr->lp_ping_deadline))
 941                 lnet_notify_locked(rtr, 1, 0, now);
 942
 943         /* Run any outstanding notifications */
 944         lnet_ni_notify_locked(rtr->lp_ni, rtr);
 945
 946         if (!lnet_isrouter(rtr) ||
 947             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
 948                 /* router table changed or router checker is shutting down */
 949                 lnet_peer_decref_locked(rtr);
 950                 return;
 951         }
 952
 953         rcd = rtr->lp_rcd != NULL ?
 954               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
 955
 956         if (rcd == NULL)
 957                 return;
 958
 959         secs = lnet_router_check_interval(rtr);
 960
 961         CDEBUG(D_NET,
 962                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
 963                "alive_count %d lp_ping_timestamp %lu\n",
 964                libcfs_nid2str(rtr->lp_nid), secs,
 965                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
 966                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
 967
 968         if (secs != 0 && !rtr->lp_ping_notsent &&
 969             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
 970                                              cfs_time_seconds(secs)))) {
 971                 int               rc;
 972                 lnet_process_id_t id;
 973                 lnet_handle_md_t  mdh;
 974
 975                 id.nid = rtr->lp_nid;
 976                 id.pid = LUSTRE_SRV_LNET_PID;
 977                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
 978
 979                 rtr->lp_ping_notsent   = 1;
 980                 rtr->lp_ping_timestamp = now;
 981
 982                 mdh = rcd->rcd_mdh;
 983
 984                 if (rtr->lp_ping_deadline == 0) {
 985                         rtr->lp_ping_deadline =
 986                                 cfs_time_shift(router_ping_timeout);
 987                 }
 988
 989                 lnet_net_unlock(rtr->lp_cpt);
 990
 991                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
 992                              LNET_PROTO_PING_MATCHBITS, 0);
 993
 994                 lnet_net_lock(rtr->lp_cpt);
 995                 if (rc != 0)
 996                         rtr->lp_ping_notsent = 0; /* no event pending */
 997         }
 998
 999         lnet_peer_decref_locked(rtr);
1000         return;
1001 }
1002
1003 int
1004 lnet_router_checker_start(void)
1005 {
1006         int          rc;
1007         int          eqsz;
1008 #ifndef __KERNEL__
1009         lnet_peer_t *rtr;
1010         __u64        version;
1011         int          nrtr = 0;
1012         int          router_checker_max_eqsize = 10240;
1013
1014         LASSERT (check_routers_before_use);
1015         LASSERT (dead_router_check_interval > 0);
1016
1017         lnet_net_lock(0);
1018
1019         /* As an approximation, allow each router the same number of
1020          * outstanding events as it is allowed outstanding sends */
1021         eqsz = 0;
1022         version = the_lnet.ln_routers_version;
1023         cfs_list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1024                 lnet_ni_t         *ni = rtr->lp_ni;
1025                 lnet_process_id_t  id;
1026
1027                 nrtr++;
1028                 eqsz += ni->ni_peertxcredits;
1029
1030                 /* one async ping reply per router */
1031                 id.nid = rtr->lp_nid;
1032                 id.pid = LUSTRE_SRV_LNET_PID;
1033
1034                 lnet_net_unlock(0);
1035
1036                 rc = LNetSetAsync(id, 1);
1037                 if (rc != 0) {
1038                         CWARN("LNetSetAsync %s failed: %d\n",
1039                               libcfs_id2str(id), rc);
1040                         return rc;
1041                 }
1042
1043                 lnet_net_lock(0);
1044                 /* NB router list doesn't change in userspace */
1045                 LASSERT(version == the_lnet.ln_routers_version);
1046         }
1047
1048         lnet_net_unlock(0);
1049
1050         if (nrtr == 0) {
1051                 CDEBUG(D_NET,
1052                        "No router found, not starting router checker\n");
1053                 return 0;
1054         }
1055
1056         /* at least allow a SENT and a REPLY per router */
1057         if (router_checker_max_eqsize < 2 * nrtr)
1058                 router_checker_max_eqsize = 2 * nrtr;
1059
1060         LASSERT (eqsz > 0);
1061         if (eqsz > router_checker_max_eqsize)
1062                 eqsz = router_checker_max_eqsize;
1063 #endif
1064
1065         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1066
1067         if (check_routers_before_use &&
1068             dead_router_check_interval <= 0) {
1069                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1070                                    " set if 'check_routers_before_use' is set"
1071                                    "\n");
1072                 return -EINVAL;
1073         }
1074
1075         if (!the_lnet.ln_routing &&
1076             live_router_check_interval <= 0 &&
1077             dead_router_check_interval <= 0)
1078                 return 0;
1079
1080 #ifdef __KERNEL__
1081         sema_init(&the_lnet.ln_rc_signal, 0);
1082         /* EQ size doesn't matter; the callback is guaranteed to get every
1083          * event */
1084         eqsz = 0;
1085         rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1086                          &the_lnet.ln_rc_eqh);
1087 #else
1088         rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE,
1089                          &the_lnet.ln_rc_eqh);
1090 #endif
1091         if (rc != 0) {
1092                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1093                 return -ENOMEM;
1094         }
1095
1096         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1097 #ifdef __KERNEL__
1098         rc = cfs_create_thread(lnet_router_checker, NULL, 0);
1099         if (rc < 0) {
1100                 CERROR("Can't start router checker thread: %d\n", rc);
1101                 /* block until event callback signals exit */
1102                 down(&the_lnet.ln_rc_signal);
1103                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1104                 LASSERT (rc == 0);
1105                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1106                 return -ENOMEM;
1107         }
1108 #endif
1109
1110         if (check_routers_before_use) {
1111                 /* Note that a helpful side-effect of pinging all known routers
1112                  * at startup is that it makes them drop stale connections they
1113                  * may have to a previous instance of me. */
1114                 lnet_wait_known_routerstate();
1115         }
1116
1117         return 0;
1118 }
1119
1120 void
1121 lnet_router_checker_stop (void)
1122 {
1123         int rc;
1124
1125         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1126                 return;
1127
1128         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1129         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1130
1131 #ifdef __KERNEL__
1132         /* block until event callback signals exit */
1133         down(&the_lnet.ln_rc_signal);
1134 #else
1135         lnet_router_checker();
1136 #endif
1137         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1138
1139         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1140         LASSERT (rc == 0);
1141         return;
1142 }
1143
1144 static void
1145 lnet_prune_rc_data(int wait_unlink)
1146 {
1147         lnet_rc_data_t          *rcd;
1148         lnet_rc_data_t          *tmp;
1149         lnet_peer_t             *lp;
1150         cfs_list_t              head;
1151         int                     i = 2;
1152
1153         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1154                    cfs_list_empty(&the_lnet.ln_rcd_deathrow) &&
1155                    cfs_list_empty(&the_lnet.ln_rcd_zombie)))
1156                 return;
1157
1158         CFS_INIT_LIST_HEAD(&head);
1159
1160         lnet_net_lock(LNET_LOCK_EX);
1161
1162         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1163                 /* router checker is stopping, prune all */
1164                 cfs_list_for_each_entry(lp, &the_lnet.ln_routers,
1165                                         lp_rtr_list) {
1166                         if (lp->lp_rcd == NULL)
1167                                 continue;
1168
1169                         LASSERT(cfs_list_empty(&lp->lp_rcd->rcd_list));
1170                         cfs_list_add(&lp->lp_rcd->rcd_list,
1171                                      &the_lnet.ln_rcd_deathrow);
1172                         lp->lp_rcd = NULL;
1173                 }
1174         }
1175
1176         /* unlink all RCDs on deathrow list */
1177         cfs_list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1178
1179         if (!cfs_list_empty(&head)) {
1180                 lnet_net_unlock(LNET_LOCK_EX);
1181
1182                 cfs_list_for_each_entry(rcd, &head, rcd_list)
1183                         LNetMDUnlink(rcd->rcd_mdh);
1184
1185                 lnet_net_lock(LNET_LOCK_EX);
1186         }
1187
1188         cfs_list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1189
1190         /* release all zombie RCDs */
1191         while (!cfs_list_empty(&the_lnet.ln_rcd_zombie)) {
1192                 cfs_list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1193                                              rcd_list) {
1194                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1195                                 cfs_list_move(&rcd->rcd_list, &head);
1196                 }
1197
1198                 wait_unlink = wait_unlink &&
1199                               !cfs_list_empty(&the_lnet.ln_rcd_zombie);
1200
1201                 lnet_net_unlock(LNET_LOCK_EX);
1202
1203                 while (!cfs_list_empty(&head)) {
1204                         rcd = cfs_list_entry(head.next,
1205                                              lnet_rc_data_t, rcd_list);
1206                         cfs_list_del_init(&rcd->rcd_list);
1207                         lnet_destroy_rc_data(rcd);
1208                 }
1209
1210                 if (!wait_unlink)
1211                         return;
1212
1213                 i++;
1214                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1215                        "Waiting for rc buffers to unlink\n");
1216                 cfs_pause(cfs_time_seconds(1) / 4);
1217
1218                 lnet_net_lock(LNET_LOCK_EX);
1219         }
1220
1221         lnet_net_unlock(LNET_LOCK_EX);
1222 }
1223
1224
1225 #if defined(__KERNEL__) && defined(LNET_ROUTER)
1226
1227 static int
1228 lnet_router_checker(void *arg)
1229 {
1230         lnet_peer_t       *rtr;
1231         cfs_list_t        *entry;
1232
1233         cfs_daemonize("router_checker");
1234         cfs_block_allsigs();
1235
1236         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1237
1238         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1239                 __u64   version;
1240                 int     cpt;
1241                 int     cpt2;
1242
1243                 cpt = lnet_net_lock_current();
1244 rescan:
1245                 version = the_lnet.ln_routers_version;
1246
1247                 cfs_list_for_each(entry, &the_lnet.ln_routers) {
1248                         rtr = cfs_list_entry(entry, lnet_peer_t, lp_rtr_list);
1249
1250                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1251                         if (cpt != cpt2) {
1252                                 lnet_net_unlock(cpt);
1253                                 cpt = cpt2;
1254                                 lnet_net_lock(cpt);
1255                                 /* the routers list has changed */
1256                                 if (version != the_lnet.ln_routers_version)
1257                                         goto rescan;
1258                         }
1259
1260                         lnet_ping_router_locked(rtr);
1261
1262                         /* NB dropped lock */
1263                         if (version != the_lnet.ln_routers_version) {
1264                                 /* the routers list has changed */
1265                                 goto rescan;
1266                         }
1267                 }
1268
1269                 if (the_lnet.ln_routing)
1270                         lnet_update_ni_status_locked();
1271
1272                 lnet_net_unlock(cpt);
1273
1274                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1275
1276                 /* Call cfs_pause() here always adds 1 to load average
1277                  * because kernel counts # active tasks as nr_running
1278                  * + nr_uninterruptible. */
1279                 cfs_schedule_timeout_and_set_state(CFS_TASK_INTERRUPTIBLE,
1280                                                    cfs_time_seconds(1));
1281         }
1282
1283         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1284
1285         lnet_prune_rc_data(1); /* wait for UNLINK */
1286
1287         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1288         up(&the_lnet.ln_rc_signal);
1289         /* The unlink event callback will signal final completion */
1290         return 0;
1291 }
1292
1293 void
1294 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1295 {
1296         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1297
1298         while (--npages >= 0)
1299                 cfs_free_page(rb->rb_kiov[npages].kiov_page);
1300
1301         LIBCFS_FREE(rb, sz);
1302 }
1303
1304 lnet_rtrbuf_t *
1305 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1306 {
1307         int            npages = rbp->rbp_npages;
1308         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1309         struct page   *page;
1310         lnet_rtrbuf_t *rb;
1311         int            i;
1312
1313         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1314         if (rb == NULL)
1315                 return NULL;
1316
1317         rb->rb_pool = rbp;
1318
1319         for (i = 0; i < npages; i++) {
1320                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1321                                           CFS_ALLOC_ZERO | CFS_ALLOC_STD);
1322                 if (page == NULL) {
1323                         while (--i >= 0)
1324                                 cfs_free_page(rb->rb_kiov[i].kiov_page);
1325
1326                         LIBCFS_FREE(rb, sz);
1327                         return NULL;
1328                 }
1329
1330                 rb->rb_kiov[i].kiov_len = CFS_PAGE_SIZE;
1331                 rb->rb_kiov[i].kiov_offset = 0;
1332                 rb->rb_kiov[i].kiov_page = page;
1333         }
1334
1335         return rb;
1336 }
1337
1338 void
1339 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp)
1340 {
1341         int             npages = rbp->rbp_npages;
1342         int             nbuffers = 0;
1343         lnet_rtrbuf_t   *rb;
1344
1345         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1346                 return;
1347
1348         LASSERT (cfs_list_empty(&rbp->rbp_msgs));
1349         LASSERT (rbp->rbp_credits == rbp->rbp_nbuffers);
1350
1351         while (!cfs_list_empty(&rbp->rbp_bufs)) {
1352                 LASSERT (rbp->rbp_credits > 0);
1353
1354                 rb = cfs_list_entry(rbp->rbp_bufs.next,
1355                                     lnet_rtrbuf_t, rb_list);
1356                 cfs_list_del(&rb->rb_list);
1357                 lnet_destroy_rtrbuf(rb, npages);
1358                 nbuffers++;
1359         }
1360
1361         LASSERT (rbp->rbp_nbuffers == nbuffers);
1362         LASSERT (rbp->rbp_credits == nbuffers);
1363
1364         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1365 }
1366
1367 int
1368 lnet_rtrpool_alloc_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1369 {
1370         lnet_rtrbuf_t *rb;
1371         int            i;
1372
1373         if (rbp->rbp_nbuffers != 0) {
1374                 LASSERT (rbp->rbp_nbuffers == nbufs);
1375                 return 0;
1376         }
1377
1378         for (i = 0; i < nbufs; i++) {
1379                 rb = lnet_new_rtrbuf(rbp, cpt);
1380
1381                 if (rb == NULL) {
1382                         CERROR("Failed to allocate %d router bufs of %d pages\n",
1383                                nbufs, rbp->rbp_npages);
1384                         return -ENOMEM;
1385                 }
1386
1387                 rbp->rbp_nbuffers++;
1388                 rbp->rbp_credits++;
1389                 rbp->rbp_mincredits++;
1390                 cfs_list_add(&rb->rb_list, &rbp->rbp_bufs);
1391
1392                 /* No allocation "under fire" */
1393                 /* Otherwise we'd need code to schedule blocked msgs etc */
1394                 LASSERT (!the_lnet.ln_routing);
1395         }
1396
1397         LASSERT (rbp->rbp_credits == nbufs);
1398         return 0;
1399 }
1400
1401 void
1402 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1403 {
1404         CFS_INIT_LIST_HEAD(&rbp->rbp_msgs);
1405         CFS_INIT_LIST_HEAD(&rbp->rbp_bufs);
1406
1407         rbp->rbp_npages = npages;
1408         rbp->rbp_credits = 0;
1409         rbp->rbp_mincredits = 0;
1410 }
1411
1412 void
1413 lnet_rtrpools_free(void)
1414 {
1415         lnet_rtrbufpool_t *rtrp;
1416         int               i;
1417
1418         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1419                 return;
1420
1421         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1422                 lnet_rtrpool_free_bufs(&rtrp[0]);
1423                 lnet_rtrpool_free_bufs(&rtrp[1]);
1424                 lnet_rtrpool_free_bufs(&rtrp[2]);
1425         }
1426
1427         cfs_percpt_free(the_lnet.ln_rtrpools);
1428         the_lnet.ln_rtrpools = NULL;
1429 }
1430
1431 static int
1432 lnet_nrb_tiny_calculate(int npages)
1433 {
1434         int     nrbs = LNET_NRB_TINY;
1435
1436         if (tiny_router_buffers < 0) {
1437                 LCONSOLE_ERROR_MSG(0x10c,
1438                                    "tiny_router_buffers=%d invalid when "
1439                                    "routing enabled\n", tiny_router_buffers);
1440                 return -1;
1441         }
1442
1443         if (tiny_router_buffers > 0)
1444                 nrbs = tiny_router_buffers;
1445
1446         nrbs /= LNET_CPT_NUMBER;
1447         return max(nrbs, LNET_NRB_TINY_MIN);
1448 }
1449
1450 static int
1451 lnet_nrb_small_calculate(int npages)
1452 {
1453         int     nrbs = LNET_NRB_SMALL;
1454
1455         if (small_router_buffers < 0) {
1456                 LCONSOLE_ERROR_MSG(0x10c,
1457                                    "small_router_buffers=%d invalid when "
1458                                    "routing enabled\n", small_router_buffers);
1459                 return -1;
1460         }
1461
1462         if (small_router_buffers > 0)
1463                 nrbs = small_router_buffers;
1464
1465         nrbs /= LNET_CPT_NUMBER;
1466         return max(nrbs, LNET_NRB_SMALL_MIN);
1467 }
1468
1469 static int
1470 lnet_nrb_large_calculate(int npages)
1471 {
1472         int     nrbs = LNET_NRB_LARGE;
1473
1474         if (large_router_buffers < 0) {
1475                 LCONSOLE_ERROR_MSG(0x10c,
1476                                    "large_router_buffers=%d invalid when "
1477                                    "routing enabled\n", large_router_buffers);
1478                 return -1;
1479         }
1480
1481         if (large_router_buffers > 0)
1482                 nrbs = large_router_buffers;
1483
1484         nrbs /= LNET_CPT_NUMBER;
1485         return max(nrbs, LNET_NRB_LARGE_MIN);
1486 }
1487
1488 int
1489 lnet_rtrpools_alloc(int im_a_router)
1490 {
1491         lnet_rtrbufpool_t *rtrp;
1492         int     large_pages = (LNET_MTU + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1493         int     small_pages = 1;
1494         int     nrb_tiny;
1495         int     nrb_small;
1496         int     nrb_large;
1497         int     rc;
1498         int     i;
1499
1500         if (!strcmp(forwarding, "")) {
1501                 /* not set either way */
1502                 if (!im_a_router)
1503                         return 0;
1504         } else if (!strcmp(forwarding, "disabled")) {
1505                 /* explicitly disabled */
1506                 return 0;
1507         } else if (!strcmp(forwarding, "enabled")) {
1508                 /* explicitly enabled */
1509         } else {
1510                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1511                                    "'enabled' or 'disabled'\n");
1512                 return -EINVAL;
1513         }
1514
1515         nrb_tiny = lnet_nrb_tiny_calculate(0);
1516         if (nrb_tiny < 0)
1517                 return -EINVAL;
1518
1519         nrb_small = lnet_nrb_small_calculate(small_pages);
1520         if (nrb_small < 0)
1521                 return -EINVAL;
1522
1523         nrb_large = lnet_nrb_large_calculate(large_pages);
1524         if (nrb_large < 0)
1525                 return -EINVAL;
1526
1527         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1528                                                 LNET_NRBPOOLS *
1529                                                 sizeof(lnet_rtrbufpool_t));
1530         if (the_lnet.ln_rtrpools == NULL) {
1531                 LCONSOLE_ERROR_MSG(0x10c,
1532                                    "Failed to initialize router buffe pool\n");
1533                 return -ENOMEM;
1534         }
1535
1536         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1537                 lnet_rtrpool_init(&rtrp[0], 0);
1538                 rc = lnet_rtrpool_alloc_bufs(&rtrp[0], nrb_tiny, i);
1539                 if (rc != 0)
1540                         goto failed;
1541
1542                 lnet_rtrpool_init(&rtrp[1], small_pages);
1543                 rc = lnet_rtrpool_alloc_bufs(&rtrp[1], nrb_small, i);
1544                 if (rc != 0)
1545                         goto failed;
1546
1547                 lnet_rtrpool_init(&rtrp[2], large_pages);
1548                 rc = lnet_rtrpool_alloc_bufs(&rtrp[2], nrb_large, i);
1549                 if (rc != 0)
1550                         goto failed;
1551         }
1552
1553         lnet_net_lock(LNET_LOCK_EX);
1554         the_lnet.ln_routing = 1;
1555         lnet_net_unlock(LNET_LOCK_EX);
1556
1557         return 0;
1558
1559  failed:
1560         lnet_rtrpools_free();
1561         return rc;
1562 }
1563
1564 int
1565 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1566 {
1567         struct lnet_peer        *lp = NULL;
1568         cfs_time_t              now = cfs_time_current();
1569         int                     cpt = lnet_cpt_of_nid(nid);
1570
1571         LASSERT (!cfs_in_interrupt ());
1572
1573         CDEBUG (D_NET, "%s notifying %s: %s\n",
1574                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1575                 libcfs_nid2str(nid),
1576                 alive ? "up" : "down");
1577
1578         if (ni != NULL &&
1579             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1580                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1581                         libcfs_nid2str(nid), alive ? "birth" : "death",
1582                         libcfs_nid2str(ni->ni_nid));
1583                 return -EINVAL;
1584         }
1585
1586         /* can't do predictions... */
1587         if (cfs_time_after(when, now)) {
1588                 CWARN ("Ignoring prediction from %s of %s %s "
1589                        "%ld seconds in the future\n",
1590                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1591                        libcfs_nid2str(nid), alive ? "up" : "down",
1592                        cfs_duration_sec(cfs_time_sub(when, now)));
1593                 return -EINVAL;
1594         }
1595
1596         if (ni != NULL && !alive &&             /* LND telling me she's down */
1597             !auto_down) {                       /* auto-down disabled */
1598                 CDEBUG(D_NET, "Auto-down disabled\n");
1599                 return 0;
1600         }
1601
1602         lnet_net_lock(cpt);
1603
1604         if (the_lnet.ln_shutdown) {
1605                 lnet_net_unlock(cpt);
1606                 return -ESHUTDOWN;
1607         }
1608
1609         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1610         if (lp == NULL) {
1611                 /* nid not found */
1612                 lnet_net_unlock(cpt);
1613                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1614                 return 0;
1615         }
1616
1617         /* We can't fully trust LND on reporting exact peer last_alive
1618          * if he notifies us about dead peer. For example ksocklnd can
1619          * call us with when == _time_when_the_node_was_booted_ if
1620          * no connections were successfully established */
1621         if (ni != NULL && !alive && when < lp->lp_last_alive)
1622                 when = lp->lp_last_alive;
1623
1624         lnet_notify_locked(lp, ni == NULL, alive, when);
1625
1626         lnet_ni_notify_locked(ni, lp);
1627
1628         lnet_peer_decref_locked(lp);
1629
1630         lnet_net_unlock(cpt);
1631         return 0;
1632 }
1633 EXPORT_SYMBOL(lnet_notify);
1634
1635 void
1636 lnet_get_tunables (void)
1637 {
1638         return;
1639 }
1640
1641 #else
1642
1643 int
1644 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1645 {
1646         return -EOPNOTSUPP;
1647 }
1648
1649 void
1650 lnet_router_checker (void)
1651 {
1652         static time_t last = 0;
1653         static int    running = 0;
1654
1655         time_t            now = cfs_time_current_sec();
1656         int               interval = now - last;
1657         int               rc;
1658         __u64             version;
1659         lnet_peer_t      *rtr;
1660
1661         /* It's no use to call me again within a sec - all intervals and
1662          * timeouts are measured in seconds */
1663         if (last != 0 && interval < 2)
1664                 return;
1665
1666         if (last != 0 &&
1667             interval > MAX(live_router_check_interval,
1668                            dead_router_check_interval))
1669                 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1670                         live_router_check_interval, dead_router_check_interval,
1671                         interval);
1672
1673         LASSERT(LNET_CPT_NUMBER == 1);
1674
1675         lnet_net_lock(0);
1676         LASSERT(!running); /* recursion check */
1677         running = 1;
1678         lnet_net_unlock(0);
1679
1680         last = now;
1681
1682         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1683                 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1684
1685         /* consume all pending events */
1686         while (1) {
1687                 int          i;
1688                 lnet_event_t ev;
1689
1690                 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1691                  * recursion breaker in LNetEQPoll would fail */
1692                 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1693                 if (rc == 0)   /* no event pending */
1694                         break;
1695
1696                 /* NB a lost SENT prevents me from pinging a router again */
1697                 if (rc == -EOVERFLOW) {
1698                         CERROR("Dropped an event!!!\n");
1699                         abort();
1700                 }
1701
1702                 LASSERT (rc == 1);
1703
1704                 lnet_router_checker_event(&ev);
1705         }
1706
1707         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1708                 lnet_prune_rc_data(1); /* release rcd */
1709                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1710                 running = 0;
1711                 return;
1712         }
1713
1714         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1715
1716         lnet_net_lock(0);
1717
1718         version = the_lnet.ln_routers_version;
1719         cfs_list_for_each_entry (rtr, &the_lnet.ln_routers, lp_rtr_list) {
1720                 lnet_ping_router_locked(rtr);
1721                 LASSERT (version == the_lnet.ln_routers_version);
1722         }
1723
1724         lnet_net_unlock(0);
1725
1726         running = 0; /* lock only needed for the recursion check */
1727         return;
1728 }
1729
1730 /* NB lnet_peers_start_down depends on me,
1731  * so must be called before any peer creation */
1732 void
1733 lnet_get_tunables (void)
1734 {
1735         char *s;
1736
1737         s = getenv("LNET_ROUTER_PING_TIMEOUT");
1738         if (s != NULL) router_ping_timeout = atoi(s);
1739
1740         s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1741         if (s != NULL) live_router_check_interval = atoi(s);
1742
1743         s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1744         if (s != NULL) dead_router_check_interval = atoi(s);
1745
1746         /* This replaces old lnd_notify mechanism */
1747         check_routers_before_use = 1;
1748         if (dead_router_check_interval <= 0)
1749                 dead_router_check_interval = 30;
1750 }
1751
1752 void
1753 lnet_rtrpools_free(void)
1754 {
1755 }
1756
1757 int
1758 lnet_rtrpools_alloc(int im_a_arouter)
1759 {
1760         return 0;
1761 }
1762
1763 #endif