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_SMALL_PAGES    1
  34 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  35 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  36 #define LNET_NRB_LARGE_PAGES    ((LNET_MTU + PAGE_CACHE_SIZE - 1) >> \
  37                                   PAGE_CACHE_SHIFT)
  38
  39 static char *forwarding = "";
  40 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
  41                 "Explicitly enable/disable forwarding between networks");
  42
  43 static int tiny_router_buffers;
  44 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
  45                 "# of 0 payload messages to buffer in the router");
  46 static int small_router_buffers;
  47 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
  48                 "# of small (1 page) messages to buffer in the router");
  49 static int large_router_buffers;
  50 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
  51                 "# of large messages to buffer in the router");
  52 static int peer_buffer_credits = 0;
  53 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  54                 "# router buffer credits per peer");
  55
  56 static int auto_down = 1;
  57 CFS_MODULE_PARM(auto_down, "i", int, 0444,
  58                 "Automatically mark peers down on comms error");
  59
  60 int
  61 lnet_peer_buffer_credits(lnet_ni_t *ni)
  62 {
  63         /* NI option overrides LNet default */
  64         if (ni->ni_peerrtrcredits > 0)
  65                 return ni->ni_peerrtrcredits;
  66         if (peer_buffer_credits > 0)
  67                 return peer_buffer_credits;
  68
  69         /* As an approximation, allow this peer the same number of router
  70          * buffers as it is allowed outstanding sends */
  71         return ni->ni_peertxcredits;
  72 }
  73
  74 /* forward ref's */
  75 static int lnet_router_checker(void *);
  76 #else
  77
  78 int
  79 lnet_peer_buffer_credits(lnet_ni_t *ni)
  80 {
  81         return 0;
  82 }
  83
  84 #endif
  85
  86 static int check_routers_before_use = 0;
  87 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
  88                 "Assume routers are down and ping them before use");
  89
  90 static int avoid_asym_router_failure = 1;
  91 CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644,
  92                 "Avoid asymmetrical router failures (0 to disable)");
  93
  94 static int dead_router_check_interval = 60;
  95 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644,
  96                 "Seconds between dead router health checks (<= 0 to disable)");
  97
  98 static int live_router_check_interval = 60;
  99 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644,
 100                 "Seconds between live router health checks (<= 0 to disable)");
 101
 102 static int router_ping_timeout = 50;
 103 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644,
 104                 "Seconds to wait for the reply to a router health query");
 105
 106 int
 107 lnet_peers_start_down(void)
 108 {
 109         return check_routers_before_use;
 110 }
 111
 112 void
 113 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
 114 {
 115         if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
 116                 CDEBUG(D_NET, "Out of date\n");
 117                 return;
 118         }
 119
 120         lp->lp_timestamp = when;                /* update timestamp */
 121         lp->lp_ping_deadline = 0;               /* disable ping timeout */
 122
 123         if (lp->lp_alive_count != 0 &&          /* got old news */
 124             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
 125                 CDEBUG(D_NET, "Old news\n");
 126                 return;
 127         }
 128
 129         /* Flag that notification is outstanding */
 130
 131         lp->lp_alive_count++;
 132         lp->lp_alive = !(!alive);               /* 1 bit! */
 133         lp->lp_notify = 1;
 134         lp->lp_notifylnd |= notifylnd;
 135         if (lp->lp_alive)
 136                 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 137
 138         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
 139 }
 140
 141 void
 142 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
 143 {
 144         int        alive;
 145         int        notifylnd;
 146
 147         /* Notify only in 1 thread at any time to ensure ordered notification.
 148          * NB individual events can be missed; the only guarantee is that you
 149          * always get the most recent news */
 150
 151         if (lp->lp_notifying)
 152                 return;
 153
 154         lp->lp_notifying = 1;
 155
 156         while (lp->lp_notify) {
 157                 alive     = lp->lp_alive;
 158                 notifylnd = lp->lp_notifylnd;
 159
 160                 lp->lp_notifylnd = 0;
 161                 lp->lp_notify    = 0;
 162
 163                 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
 164                         lnet_net_unlock(lp->lp_cpt);
 165
 166                         /* A new notification could happen now; I'll handle it
 167                          * when control returns to me */
 168
 169                         (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
 170
 171                         lnet_net_lock(lp->lp_cpt);
 172                 }
 173         }
 174
 175         lp->lp_notifying = 0;
 176 }
 177
 178
 179 static void
 180 lnet_rtr_addref_locked(lnet_peer_t *lp)
 181 {
 182         LASSERT(lp->lp_refcount > 0);
 183         LASSERT(lp->lp_rtr_refcount >= 0);
 184
 185         /* lnet_net_lock must be exclusively locked */
 186         lp->lp_rtr_refcount++;
 187         if (lp->lp_rtr_refcount == 1) {
 188                 struct list_head *pos;
 189
 190                 /* a simple insertion sort */
 191                 list_for_each_prev(pos, &the_lnet.ln_routers) {
 192                         lnet_peer_t *rtr = list_entry(pos, lnet_peer_t,
 193                                                       lp_rtr_list);
 194
 195                         if (rtr->lp_nid < lp->lp_nid)
 196                                 break;
 197                 }
 198
 199                 list_add(&lp->lp_rtr_list, pos);
 200                 /* addref for the_lnet.ln_routers */
 201                 lnet_peer_addref_locked(lp);
 202                 the_lnet.ln_routers_version++;
 203         }
 204 }
 205
 206 static void
 207 lnet_rtr_decref_locked(lnet_peer_t *lp)
 208 {
 209         LASSERT(lp->lp_refcount > 0);
 210         LASSERT(lp->lp_rtr_refcount > 0);
 211
 212         /* lnet_net_lock must be exclusively locked */
 213         lp->lp_rtr_refcount--;
 214         if (lp->lp_rtr_refcount == 0) {
 215                 LASSERT(list_empty(&lp->lp_routes));
 216
 217                 if (lp->lp_rcd != NULL) {
 218                         list_add(&lp->lp_rcd->rcd_list,
 219                                  &the_lnet.ln_rcd_deathrow);
 220                         lp->lp_rcd = NULL;
 221                 }
 222
 223                 list_del(&lp->lp_rtr_list);
 224                 /* decref for the_lnet.ln_routers */
 225                 lnet_peer_decref_locked(lp);
 226                 the_lnet.ln_routers_version++;
 227         }
 228 }
 229
 230 lnet_remotenet_t *
 231 lnet_find_net_locked (__u32 net)
 232 {
 233         lnet_remotenet_t *rnet;
 234         struct list_head *tmp;
 235         struct list_head *rn_list;
 236
 237         LASSERT(!the_lnet.ln_shutdown);
 238
 239         rn_list = lnet_net2rnethash(net);
 240         list_for_each(tmp, rn_list) {
 241                 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
 242
 243                 if (rnet->lrn_net == net)
 244                         return rnet;
 245         }
 246         return NULL;
 247 }
 248
 249 static void lnet_shuffle_seed(void)
 250 {
 251         static int seeded = 0;
 252         int lnd_type, seed[2];
 253         struct timeval tv;
 254         lnet_ni_t *ni;
 255         struct list_head *tmp;
 256
 257         if (seeded)
 258                 return;
 259
 260         cfs_get_random_bytes(seed, sizeof(seed));
 261
 262         /* Nodes with small feet have little entropy
 263          * the NID for this node gives the most entropy in the low bits */
 264         list_for_each(tmp, &the_lnet.ln_nis) {
 265                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 266                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 267
 268                 if (lnd_type != LOLND)
 269                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
 270         }
 271
 272         do_gettimeofday(&tv);
 273         cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
 274         seeded = 1;
 275         return;
 276 }
 277
 278 /* NB expects LNET_LOCK held */
 279 void
 280 lnet_add_route_to_rnet (lnet_remotenet_t *rnet, lnet_route_t *route)
 281 {
 282         unsigned int      len = 0;
 283         unsigned int      offset = 0;
 284         struct list_head *e;
 285
 286         lnet_shuffle_seed();
 287
 288         list_for_each(e, &rnet->lrn_routes) {
 289                 len++;
 290         }
 291
 292         /* len+1 positions to add a new entry, also prevents division by 0 */
 293         offset = cfs_rand() % (len + 1);
 294         list_for_each(e, &rnet->lrn_routes) {
 295                 if (offset == 0)
 296                         break;
 297                 offset--;
 298         }
 299         list_add(&route->lr_list, e);
 300         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 301
 302         the_lnet.ln_remote_nets_version++;
 303         lnet_rtr_addref_locked(route->lr_gateway);
 304 }
 305
 306 int
 307 lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway,
 308                unsigned int priority)
 309 {
 310         struct list_head        *e;
 311         lnet_remotenet_t        *rnet;
 312         lnet_remotenet_t        *rnet2;
 313         lnet_route_t            *route;
 314         lnet_ni_t               *ni;
 315         int                     add_route;
 316         int                     rc;
 317
 318         CDEBUG(D_NET, "Add route: net %s hops %u priority %u gw %s\n",
 319                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 320
 321         if (gateway == LNET_NID_ANY ||
 322             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 323             net == LNET_NIDNET(LNET_NID_ANY) ||
 324             LNET_NETTYP(net) == LOLND ||
 325             LNET_NIDNET(gateway) == net ||
 326             hops < 1 || hops > 255)
 327                 return -EINVAL;
 328
 329         if (lnet_islocalnet(net))       /* it's a local network */
 330                 return 0;               /* ignore the route entry */
 331
 332         /* Assume net, route, all new */
 333         LIBCFS_ALLOC(route, sizeof(*route));
 334         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 335         if (route == NULL || rnet == NULL) {
 336                 CERROR("Out of memory creating route %s %d %s\n",
 337                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 338                 if (route != NULL)
 339                         LIBCFS_FREE(route, sizeof(*route));
 340                 if (rnet != NULL)
 341                         LIBCFS_FREE(rnet, sizeof(*rnet));
 342                 return -ENOMEM;
 343         }
 344
 345         INIT_LIST_HEAD(&rnet->lrn_routes);
 346         rnet->lrn_net = net;
 347         route->lr_hops = hops;
 348         route->lr_net = net;
 349         route->lr_priority = priority;
 350
 351         lnet_net_lock(LNET_LOCK_EX);
 352
 353         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
 354         if (rc != 0) {
 355                 lnet_net_unlock(LNET_LOCK_EX);
 356
 357                 LIBCFS_FREE(route, sizeof(*route));
 358                 LIBCFS_FREE(rnet, sizeof(*rnet));
 359
 360                 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
 361                         return 0;        /* ignore the route entry */
 362                 CERROR("Error %d creating route %s %d %s\n", rc,
 363                         libcfs_net2str(net), hops,
 364                         libcfs_nid2str(gateway));
 365                 return rc;
 366         }
 367
 368         LASSERT(!the_lnet.ln_shutdown);
 369
 370         rnet2 = lnet_find_net_locked(net);
 371         if (rnet2 == NULL) {
 372                 /* new network */
 373                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 374                 rnet2 = rnet;
 375         }
 376
 377         /* Search for a duplicate route (it's a NOOP if it is) */
 378         add_route = 1;
 379         list_for_each(e, &rnet2->lrn_routes) {
 380                 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
 381
 382                 if (route2->lr_gateway == route->lr_gateway) {
 383                         add_route = 0;
 384                         break;
 385                 }
 386
 387                 /* our lookups must be true */
 388                 LASSERT(route2->lr_gateway->lp_nid != gateway);
 389         }
 390
 391         if (add_route) {
 392                 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
 393                 lnet_add_route_to_rnet(rnet2, route);
 394
 395                 ni = route->lr_gateway->lp_ni;
 396                 lnet_net_unlock(LNET_LOCK_EX);
 397
 398                 /* XXX Assume alive */
 399                 if (ni->ni_lnd->lnd_notify != NULL)
 400                         (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
 401
 402                 lnet_net_lock(LNET_LOCK_EX);
 403         }
 404
 405         /* -1 for notify or !add_route */
 406         lnet_peer_decref_locked(route->lr_gateway);
 407         lnet_net_unlock(LNET_LOCK_EX);
 408
 409         if (!add_route)
 410                 LIBCFS_FREE(route, sizeof(*route));
 411
 412         if (rnet != rnet2)
 413                 LIBCFS_FREE(rnet, sizeof(*rnet));
 414
 415         return 0;
 416 }
 417
 418 int
 419 lnet_check_routes(void)
 420 {
 421         lnet_remotenet_t *rnet;
 422         lnet_route_t     *route;
 423         lnet_route_t     *route2;
 424         struct list_head *e1;
 425         struct list_head *e2;
 426         int               cpt;
 427         struct list_head *rn_list;
 428         int               i;
 429
 430         cpt = lnet_net_lock_current();
 431
 432         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 433                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 434                 list_for_each(e1, rn_list) {
 435                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 436
 437                         route2 = NULL;
 438                         list_for_each(e2, &rnet->lrn_routes) {
 439                                 lnet_nid_t      nid1;
 440                                 lnet_nid_t      nid2;
 441                                 int             net;
 442
 443                                 route = list_entry(e2, lnet_route_t,
 444                                                    lr_list);
 445
 446                                 if (route2 == NULL) {
 447                                         route2 = route;
 448                                         continue;
 449                                 }
 450
 451                                 if (route->lr_gateway->lp_ni ==
 452                                     route2->lr_gateway->lp_ni)
 453                                         continue;
 454
 455                                 nid1 = route->lr_gateway->lp_nid;
 456                                 nid2 = route2->lr_gateway->lp_nid;
 457                                 net = rnet->lrn_net;
 458
 459                                 lnet_net_unlock(cpt);
 460
 461                                 CERROR("Routes to %s via %s and %s not "
 462                                        "supported\n",
 463                                        libcfs_net2str(net),
 464                                        libcfs_nid2str(nid1),
 465                                        libcfs_nid2str(nid2));
 466                                 return -EINVAL;
 467                         }
 468                 }
 469         }
 470
 471         lnet_net_unlock(cpt);
 472         return 0;
 473 }
 474
 475 int
 476 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 477 {
 478         struct lnet_peer        *gateway;
 479         lnet_remotenet_t        *rnet;
 480         lnet_route_t            *route;
 481         struct list_head        *e1;
 482         struct list_head        *e2;
 483         int                     rc = -ENOENT;
 484         struct list_head        *rn_list;
 485         int                     idx = 0;
 486
 487         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 488                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 489
 490         /* NB Caller may specify either all routes via the given gateway
 491          * or a specific route entry actual NIDs) */
 492
 493         lnet_net_lock(LNET_LOCK_EX);
 494         if (net == LNET_NIDNET(LNET_NID_ANY))
 495                 rn_list = &the_lnet.ln_remote_nets_hash[0];
 496         else
 497                 rn_list = lnet_net2rnethash(net);
 498
 499 again:
 500         list_for_each(e1, rn_list) {
 501                 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 502
 503                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 504                         net == rnet->lrn_net))
 505                         continue;
 506
 507                 list_for_each(e2, &rnet->lrn_routes) {
 508                         route = list_entry(e2, lnet_route_t, lr_list);
 509
 510                         gateway = route->lr_gateway;
 511                         if (!(gw_nid == LNET_NID_ANY ||
 512                               gw_nid == gateway->lp_nid))
 513                                 continue;
 514
 515                         list_del(&route->lr_list);
 516                         list_del(&route->lr_gwlist);
 517                         the_lnet.ln_remote_nets_version++;
 518
 519                         if (list_empty(&rnet->lrn_routes))
 520                                 list_del(&rnet->lrn_list);
 521                         else
 522                                 rnet = NULL;
 523
 524                         lnet_rtr_decref_locked(gateway);
 525                         lnet_peer_decref_locked(gateway);
 526
 527                         lnet_net_unlock(LNET_LOCK_EX);
 528
 529                         LIBCFS_FREE(route, sizeof(*route));
 530
 531                         if (rnet != NULL)
 532                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 533
 534                         rc = 0;
 535                         lnet_net_lock(LNET_LOCK_EX);
 536                         goto again;
 537                 }
 538         }
 539
 540         if (net == LNET_NIDNET(LNET_NID_ANY) &&
 541             ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
 542                 rn_list = &the_lnet.ln_remote_nets_hash[idx];
 543                 goto again;
 544         }
 545         lnet_net_unlock(LNET_LOCK_EX);
 546
 547         return rc;
 548 }
 549
 550 void
 551 lnet_destroy_routes (void)
 552 {
 553         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 554 }
 555
 556 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
 557 {
 558         int i, rc = -ENOENT, lidx, j;
 559
 560         if (the_lnet.ln_rtrpools == NULL)
 561                 return rc;
 562
 563         for (i = 0; i < LNET_NRBPOOLS; i++) {
 564                 lnet_rtrbufpool_t *rbp;
 565
 566                 lnet_net_lock(LNET_LOCK_EX);
 567                 lidx = idx;
 568                 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
 569                         if (lidx-- == 0) {
 570                                 rc = 0;
 571                                 pool_cfg->pl_pools[i].pl_npages =
 572                                         rbp[i].rbp_npages;
 573                                 pool_cfg->pl_pools[i].pl_nbuffers =
 574                                         rbp[i].rbp_nbuffers;
 575                                 pool_cfg->pl_pools[i].pl_credits =
 576                                         rbp[i].rbp_credits;
 577                                 pool_cfg->pl_pools[i].pl_mincredits =
 578                                         rbp[i].rbp_mincredits;
 579                                 break;
 580                         }
 581                 }
 582                 lnet_net_unlock(LNET_LOCK_EX);
 583         }
 584
 585         lnet_net_lock(LNET_LOCK_EX);
 586         pool_cfg->pl_routing = the_lnet.ln_routing;
 587         lnet_net_unlock(LNET_LOCK_EX);
 588
 589         return rc;
 590 }
 591
 592 int
 593 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 594                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 595 {
 596         struct list_head *e1;
 597         struct list_head *e2;
 598         lnet_remotenet_t *rnet;
 599         lnet_route_t     *route;
 600         int               cpt;
 601         int               i;
 602         struct list_head *rn_list;
 603
 604         cpt = lnet_net_lock_current();
 605
 606         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 607                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 608                 list_for_each(e1, rn_list) {
 609                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 610
 611                         list_for_each(e2, &rnet->lrn_routes) {
 612                                 route = list_entry(e2, lnet_route_t,
 613                                                    lr_list);
 614
 615                                 if (idx-- == 0) {
 616                                         *net      = rnet->lrn_net;
 617                                         *hops     = route->lr_hops;
 618                                         *priority = route->lr_priority;
 619                                         *gateway  = route->lr_gateway->lp_nid;
 620                                         *alive    =
 621                                                 route->lr_gateway->lp_alive &&
 622                                                         !route->lr_downis;
 623                                         lnet_net_unlock(cpt);
 624                                         return 0;
 625                                 }
 626                         }
 627                 }
 628         }
 629
 630         lnet_net_unlock(cpt);
 631         return -ENOENT;
 632 }
 633
 634 void
 635 lnet_swap_pinginfo(lnet_ping_info_t *info)
 636 {
 637         int               i;
 638         lnet_ni_status_t *stat;
 639
 640         __swab32s(&info->pi_magic);
 641         __swab32s(&info->pi_features);
 642         __swab32s(&info->pi_pid);
 643         __swab32s(&info->pi_nnis);
 644         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 645                 stat = &info->pi_ni[i];
 646                 __swab64s(&stat->ns_nid);
 647                 __swab32s(&stat->ns_status);
 648         }
 649         return;
 650 }
 651
 652 /**
 653  * parse router-checker pinginfo, record number of down NIs for remote
 654  * networks on that router.
 655  */
 656 static void
 657 lnet_parse_rc_info(lnet_rc_data_t *rcd)
 658 {
 659         lnet_ping_info_t        *info = rcd->rcd_pinginfo;
 660         struct lnet_peer        *gw   = rcd->rcd_gateway;
 661         lnet_route_t            *rte;
 662
 663         if (!gw->lp_alive)
 664                 return;
 665
 666         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 667                 lnet_swap_pinginfo(info);
 668
 669         /* NB always racing with network! */
 670         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 671                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 672                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
 673                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 674                 return;
 675         }
 676
 677         gw->lp_ping_feats = info->pi_features;
 678         if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
 679                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
 680                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
 681                 return; /* nothing I can understand */
 682         }
 683
 684         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
 685                 return; /* can't carry NI status info */
 686
 687         list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
 688                 int     down = 0;
 689                 int     up = 0;
 690                 int     i;
 691
 692                 if ((gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
 693                         rte->lr_downis = 1;
 694                         continue;
 695                 }
 696
 697                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 698                         lnet_ni_status_t *stat = &info->pi_ni[i];
 699                         lnet_nid_t       nid = stat->ns_nid;
 700
 701                         if (nid == LNET_NID_ANY) {
 702                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 703                                        libcfs_nid2str(gw->lp_nid));
 704                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 705                                 return;
 706                         }
 707
 708                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 709                                 continue;
 710
 711                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 712                                 down++;
 713                                 continue;
 714                         }
 715
 716                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 717                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 718                                         up = 1;
 719                                         break;
 720                                 }
 721                                 continue;
 722                         }
 723
 724                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 725                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
 726                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 727                         return;
 728                 }
 729
 730                 if (up) { /* ignore downed NIs if NI for dest network is up */
 731                         rte->lr_downis = 0;
 732                         continue;
 733                 }
 734                 rte->lr_downis = down;
 735         }
 736 }
 737
 738 static void
 739 lnet_router_checker_event(lnet_event_t *event)
 740 {
 741         lnet_rc_data_t          *rcd = event->md.user_ptr;
 742         struct lnet_peer        *lp;
 743
 744         LASSERT(rcd != NULL);
 745
 746         if (event->unlinked) {
 747                 LNetInvalidateHandle(&rcd->rcd_mdh);
 748                 return;
 749         }
 750
 751         LASSERT(event->type == LNET_EVENT_SEND ||
 752                 event->type == LNET_EVENT_REPLY);
 753
 754         lp = rcd->rcd_gateway;
 755         LASSERT(lp != NULL);
 756
 757          /* NB: it's called with holding lnet_res_lock, we have a few
 758           * places need to hold both locks at the same time, please take
 759           * care of lock ordering */
 760         lnet_net_lock(lp->lp_cpt);
 761         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
 762                 /* ignore if no longer a router or rcd is replaced */
 763                 goto out;
 764         }
 765
 766         if (event->type == LNET_EVENT_SEND) {
 767                 lp->lp_ping_notsent = 0;
 768                 if (event->status == 0)
 769                         goto out;
 770         }
 771
 772         /* LNET_EVENT_REPLY */
 773         /* A successful REPLY means the router is up.  If _any_ comms
 774          * to the router fail I assume it's down (this will happen if
 775          * we ping alive routers to try to detect router death before
 776          * apps get burned). */
 777
 778         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
 779         /* The router checker will wake up very shortly and do the
 780          * actual notification.
 781          * XXX If 'lp' stops being a router before then, it will still
 782          * have the notification pending!!! */
 783
 784         if (avoid_asym_router_failure && event->status == 0)
 785                 lnet_parse_rc_info(rcd);
 786
 787  out:
 788         lnet_net_unlock(lp->lp_cpt);
 789 }
 790
 791 void
 792 lnet_wait_known_routerstate(void)
 793 {
 794         lnet_peer_t      *rtr;
 795         struct list_head *entry;
 796         int               all_known;
 797
 798         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 799
 800         for (;;) {
 801                 int cpt = lnet_net_lock_current();
 802
 803                 all_known = 1;
 804                 list_for_each(entry, &the_lnet.ln_routers) {
 805                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
 806
 807                         if (rtr->lp_alive_count == 0) {
 808                                 all_known = 0;
 809                                 break;
 810                         }
 811                 }
 812
 813                 lnet_net_unlock(cpt);
 814
 815                 if (all_known)
 816                         return;
 817
 818 #ifndef __KERNEL__
 819                 lnet_router_checker();
 820 #endif
 821                 cfs_pause(cfs_time_seconds(1));
 822         }
 823 }
 824
 825 void
 826 lnet_update_ni_status_locked(void)
 827 {
 828         lnet_ni_t       *ni;
 829         long            now;
 830         int             timeout;
 831
 832         LASSERT(the_lnet.ln_routing);
 833
 834         timeout = router_ping_timeout +
 835                   MAX(live_router_check_interval, dead_router_check_interval);
 836
 837         now = cfs_time_current_sec();
 838         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 839                 if (ni->ni_lnd->lnd_type == LOLND)
 840                         continue;
 841
 842                 if (now < ni->ni_last_alive + timeout)
 843                         continue;
 844
 845                 lnet_ni_lock(ni);
 846                 /* re-check with lock */
 847                 if (now < ni->ni_last_alive + timeout) {
 848                         lnet_ni_unlock(ni);
 849                         continue;
 850                 }
 851
 852                 LASSERT(ni->ni_status != NULL);
 853
 854                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 855                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 856                                libcfs_nid2str(ni->ni_nid), timeout);
 857                         /* NB: so far, this is the only place to set
 858                          * NI status to "down" */
 859                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 860                 }
 861                 lnet_ni_unlock(ni);
 862         }
 863 }
 864
 865 void
 866 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
 867 {
 868         LASSERT(list_empty(&rcd->rcd_list));
 869         /* detached from network */
 870         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 871
 872         if (rcd->rcd_gateway != NULL) {
 873                 int cpt = rcd->rcd_gateway->lp_cpt;
 874
 875                 lnet_net_lock(cpt);
 876                 lnet_peer_decref_locked(rcd->rcd_gateway);
 877                 lnet_net_unlock(cpt);
 878         }
 879
 880         if (rcd->rcd_pinginfo != NULL)
 881                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 882
 883         LIBCFS_FREE(rcd, sizeof(*rcd));
 884 }
 885
 886 lnet_rc_data_t *
 887 lnet_create_rc_data_locked(lnet_peer_t *gateway)
 888 {
 889         lnet_rc_data_t          *rcd = NULL;
 890         lnet_ping_info_t        *pi;
 891         int                     rc;
 892         int                     i;
 893
 894         lnet_net_unlock(gateway->lp_cpt);
 895
 896         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 897         if (rcd == NULL)
 898                 goto out;
 899
 900         LNetInvalidateHandle(&rcd->rcd_mdh);
 901         INIT_LIST_HEAD(&rcd->rcd_list);
 902
 903         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 904         if (pi == NULL)
 905                 goto out;
 906
 907         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 908                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 909                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 910         }
 911         rcd->rcd_pinginfo = pi;
 912
 913         LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 914         rc = LNetMDBind((lnet_md_t){.start     = pi,
 915                                     .user_ptr  = rcd,
 916                                     .length    = LNET_PINGINFO_SIZE,
 917                                     .threshold = LNET_MD_THRESH_INF,
 918                                     .options   = LNET_MD_TRUNCATE,
 919                                     .eq_handle = the_lnet.ln_rc_eqh},
 920                         LNET_UNLINK,
 921                         &rcd->rcd_mdh);
 922         if (rc < 0) {
 923                 CERROR("Can't bind MD: %d\n", rc);
 924                 goto out;
 925         }
 926         LASSERT(rc == 0);
 927
 928         lnet_net_lock(gateway->lp_cpt);
 929         /* router table changed or someone has created rcd for this gateway */
 930         if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
 931                 lnet_net_unlock(gateway->lp_cpt);
 932                 goto out;
 933         }
 934
 935         lnet_peer_addref_locked(gateway);
 936         rcd->rcd_gateway = gateway;
 937         gateway->lp_rcd = rcd;
 938         gateway->lp_ping_notsent = 0;
 939
 940         return rcd;
 941
 942  out:
 943         if (rcd != NULL) {
 944                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
 945                         rc = LNetMDUnlink(rcd->rcd_mdh);
 946                         LASSERT(rc == 0);
 947                 }
 948                 lnet_destroy_rc_data(rcd);
 949         }
 950
 951         lnet_net_lock(gateway->lp_cpt);
 952         return gateway->lp_rcd;
 953 }
 954
 955 static int
 956 lnet_router_check_interval (lnet_peer_t *rtr)
 957 {
 958         int secs;
 959
 960         secs = rtr->lp_alive ? live_router_check_interval :
 961                                dead_router_check_interval;
 962         if (secs < 0)
 963                 secs = 0;
 964
 965         return secs;
 966 }
 967
 968 static void
 969 lnet_ping_router_locked (lnet_peer_t *rtr)
 970 {
 971         lnet_rc_data_t *rcd = NULL;
 972         cfs_time_t      now = cfs_time_current();
 973         int             secs;
 974
 975         lnet_peer_addref_locked(rtr);
 976
 977         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
 978             cfs_time_after(now, rtr->lp_ping_deadline))
 979                 lnet_notify_locked(rtr, 1, 0, now);
 980
 981         /* Run any outstanding notifications */
 982         lnet_ni_notify_locked(rtr->lp_ni, rtr);
 983
 984         if (!lnet_isrouter(rtr) ||
 985             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
 986                 /* router table changed or router checker is shutting down */
 987                 lnet_peer_decref_locked(rtr);
 988                 return;
 989         }
 990
 991         rcd = rtr->lp_rcd != NULL ?
 992               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
 993
 994         if (rcd == NULL)
 995                 return;
 996
 997         secs = lnet_router_check_interval(rtr);
 998
 999         CDEBUG(D_NET,
1000                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1001                "alive_count %d lp_ping_timestamp %lu\n",
1002                libcfs_nid2str(rtr->lp_nid), secs,
1003                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
1004                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
1005
1006         if (secs != 0 && !rtr->lp_ping_notsent &&
1007             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
1008                                              cfs_time_seconds(secs)))) {
1009                 int               rc;
1010                 lnet_process_id_t id;
1011                 lnet_handle_md_t  mdh;
1012
1013                 id.nid = rtr->lp_nid;
1014                 id.pid = LUSTRE_SRV_LNET_PID;
1015                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1016
1017                 rtr->lp_ping_notsent   = 1;
1018                 rtr->lp_ping_timestamp = now;
1019
1020                 mdh = rcd->rcd_mdh;
1021
1022                 if (rtr->lp_ping_deadline == 0) {
1023                         rtr->lp_ping_deadline =
1024                                 cfs_time_shift(router_ping_timeout);
1025                 }
1026
1027                 lnet_net_unlock(rtr->lp_cpt);
1028
1029                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1030                              LNET_PROTO_PING_MATCHBITS, 0);
1031
1032                 lnet_net_lock(rtr->lp_cpt);
1033                 if (rc != 0)
1034                         rtr->lp_ping_notsent = 0; /* no event pending */
1035         }
1036
1037         lnet_peer_decref_locked(rtr);
1038         return;
1039 }
1040
1041 int
1042 lnet_router_checker_start(void)
1043 {
1044         int                     rc;
1045         int                     eqsz;
1046 #ifdef __KERNEL__
1047         struct task_struct     *task;
1048 #else /* __KERNEL__ */
1049         lnet_peer_t            *rtr;
1050         __u64                   version;
1051         int                     nrtr = 0;
1052         int                     router_checker_max_eqsize = 10240;
1053
1054         LASSERT (check_routers_before_use);
1055         LASSERT (dead_router_check_interval > 0);
1056
1057         lnet_net_lock(0);
1058
1059         /* As an approximation, allow each router the same number of
1060          * outstanding events as it is allowed outstanding sends */
1061         eqsz = 0;
1062         version = the_lnet.ln_routers_version;
1063         list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1064                 lnet_ni_t         *ni = rtr->lp_ni;
1065                 lnet_process_id_t  id;
1066
1067                 nrtr++;
1068                 eqsz += ni->ni_peertxcredits;
1069
1070                 /* one async ping reply per router */
1071                 id.nid = rtr->lp_nid;
1072                 id.pid = LUSTRE_SRV_LNET_PID;
1073
1074                 lnet_net_unlock(0);
1075
1076                 rc = LNetSetAsync(id, 1);
1077                 if (rc != 0) {
1078                         CWARN("LNetSetAsync %s failed: %d\n",
1079                               libcfs_id2str(id), rc);
1080                         return rc;
1081                 }
1082
1083                 lnet_net_lock(0);
1084                 /* NB router list doesn't change in userspace */
1085                 LASSERT(version == the_lnet.ln_routers_version);
1086         }
1087
1088         lnet_net_unlock(0);
1089
1090         if (nrtr == 0) {
1091                 CDEBUG(D_NET,
1092                        "No router found, not starting router checker\n");
1093                 return 0;
1094         }
1095
1096         /* at least allow a SENT and a REPLY per router */
1097         if (router_checker_max_eqsize < 2 * nrtr)
1098                 router_checker_max_eqsize = 2 * nrtr;
1099
1100         LASSERT (eqsz > 0);
1101         if (eqsz > router_checker_max_eqsize)
1102                 eqsz = router_checker_max_eqsize;
1103 #endif
1104
1105         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1106
1107         if (check_routers_before_use &&
1108             dead_router_check_interval <= 0) {
1109                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1110                                    " set if 'check_routers_before_use' is set"
1111                                    "\n");
1112                 return -EINVAL;
1113         }
1114
1115         if (!the_lnet.ln_routing &&
1116             live_router_check_interval <= 0 &&
1117             dead_router_check_interval <= 0)
1118                 return 0;
1119
1120 #ifdef __KERNEL__
1121         sema_init(&the_lnet.ln_rc_signal, 0);
1122         /* EQ size doesn't matter; the callback is guaranteed to get every
1123          * event */
1124         eqsz = 0;
1125         rc = LNetEQAlloc(eqsz, lnet_router_checker_event,
1126                          &the_lnet.ln_rc_eqh);
1127 #else
1128         rc = LNetEQAlloc(eqsz, LNET_EQ_HANDLER_NONE,
1129                          &the_lnet.ln_rc_eqh);
1130 #endif
1131         if (rc != 0) {
1132                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1133                 return -ENOMEM;
1134         }
1135
1136         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1137 #ifdef __KERNEL__
1138         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1139         if (IS_ERR(task)) {
1140                 rc = PTR_ERR(task);
1141                 CERROR("Can't start router checker thread: %d\n", rc);
1142                 /* block until event callback signals exit */
1143                 down(&the_lnet.ln_rc_signal);
1144                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1145                 LASSERT(rc == 0);
1146                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1147                 return -ENOMEM;
1148         }
1149 #endif
1150
1151         if (check_routers_before_use) {
1152                 /* Note that a helpful side-effect of pinging all known routers
1153                  * at startup is that it makes them drop stale connections they
1154                  * may have to a previous instance of me. */
1155                 lnet_wait_known_routerstate();
1156         }
1157
1158         return 0;
1159 }
1160
1161 void
1162 lnet_router_checker_stop (void)
1163 {
1164         int rc;
1165
1166         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1167                 return;
1168
1169         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1170         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1171
1172 #ifdef __KERNEL__
1173         /* block until event callback signals exit */
1174         down(&the_lnet.ln_rc_signal);
1175 #else
1176         lnet_router_checker();
1177 #endif
1178         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1179
1180         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1181         LASSERT (rc == 0);
1182         return;
1183 }
1184
1185 static void
1186 lnet_prune_rc_data(int wait_unlink)
1187 {
1188         lnet_rc_data_t          *rcd;
1189         lnet_rc_data_t          *tmp;
1190         lnet_peer_t             *lp;
1191         struct list_head         head;
1192         int                      i = 2;
1193
1194         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1195                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1196                    list_empty(&the_lnet.ln_rcd_zombie)))
1197                 return;
1198
1199         INIT_LIST_HEAD(&head);
1200
1201         lnet_net_lock(LNET_LOCK_EX);
1202
1203         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1204                 /* router checker is stopping, prune all */
1205                 list_for_each_entry(lp, &the_lnet.ln_routers,
1206                                     lp_rtr_list) {
1207                         if (lp->lp_rcd == NULL)
1208                                 continue;
1209
1210                         LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1211                         list_add(&lp->lp_rcd->rcd_list,
1212                                  &the_lnet.ln_rcd_deathrow);
1213                         lp->lp_rcd = NULL;
1214                 }
1215         }
1216
1217         /* unlink all RCDs on deathrow list */
1218         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1219
1220         if (!list_empty(&head)) {
1221                 lnet_net_unlock(LNET_LOCK_EX);
1222
1223                 list_for_each_entry(rcd, &head, rcd_list)
1224                         LNetMDUnlink(rcd->rcd_mdh);
1225
1226                 lnet_net_lock(LNET_LOCK_EX);
1227         }
1228
1229         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1230
1231         /* release all zombie RCDs */
1232         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1233                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1234                                          rcd_list) {
1235                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1236                                 list_move(&rcd->rcd_list, &head);
1237                 }
1238
1239                 wait_unlink = wait_unlink &&
1240                               !list_empty(&the_lnet.ln_rcd_zombie);
1241
1242                 lnet_net_unlock(LNET_LOCK_EX);
1243
1244                 while (!list_empty(&head)) {
1245                         rcd = list_entry(head.next,
1246                                          lnet_rc_data_t, rcd_list);
1247                         list_del_init(&rcd->rcd_list);
1248                         lnet_destroy_rc_data(rcd);
1249                 }
1250
1251                 if (!wait_unlink)
1252                         return;
1253
1254                 i++;
1255                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1256                        "Waiting for rc buffers to unlink\n");
1257                 cfs_pause(cfs_time_seconds(1) / 4);
1258
1259                 lnet_net_lock(LNET_LOCK_EX);
1260         }
1261
1262         lnet_net_unlock(LNET_LOCK_EX);
1263 }
1264
1265
1266 #if defined(__KERNEL__) && defined(LNET_ROUTER)
1267
1268 static int
1269 lnet_router_checker(void *arg)
1270 {
1271         lnet_peer_t       *rtr;
1272         struct list_head  *entry;
1273
1274         cfs_block_allsigs();
1275
1276         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1277
1278         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1279                 __u64   version;
1280                 int     cpt;
1281                 int     cpt2;
1282
1283                 cpt = lnet_net_lock_current();
1284 rescan:
1285                 version = the_lnet.ln_routers_version;
1286
1287                 list_for_each(entry, &the_lnet.ln_routers) {
1288                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1289
1290                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1291                         if (cpt != cpt2) {
1292                                 lnet_net_unlock(cpt);
1293                                 cpt = cpt2;
1294                                 lnet_net_lock(cpt);
1295                                 /* the routers list has changed */
1296                                 if (version != the_lnet.ln_routers_version)
1297                                         goto rescan;
1298                         }
1299
1300                         lnet_ping_router_locked(rtr);
1301
1302                         /* NB dropped lock */
1303                         if (version != the_lnet.ln_routers_version) {
1304                                 /* the routers list has changed */
1305                                 goto rescan;
1306                         }
1307                 }
1308
1309                 if (the_lnet.ln_routing)
1310                         lnet_update_ni_status_locked();
1311
1312                 lnet_net_unlock(cpt);
1313
1314                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1315
1316                 /* Call cfs_pause() here always adds 1 to load average
1317                  * because kernel counts # active tasks as nr_running
1318                  * + nr_uninterruptible. */
1319                 schedule_timeout_and_set_state(TASK_INTERRUPTIBLE,
1320                                                    cfs_time_seconds(1));
1321         }
1322
1323         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1324
1325         lnet_prune_rc_data(1); /* wait for UNLINK */
1326
1327         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1328         up(&the_lnet.ln_rc_signal);
1329         /* The unlink event callback will signal final completion */
1330         return 0;
1331 }
1332
1333 void
1334 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1335 {
1336         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1337
1338         while (--npages >= 0)
1339                 __free_page(rb->rb_kiov[npages].kiov_page);
1340
1341         LIBCFS_FREE(rb, sz);
1342 }
1343
1344 lnet_rtrbuf_t *
1345 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1346 {
1347         int            npages = rbp->rbp_npages;
1348         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1349         struct page   *page;
1350         lnet_rtrbuf_t *rb;
1351         int            i;
1352
1353         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1354         if (rb == NULL)
1355                 return NULL;
1356
1357         rb->rb_pool = rbp;
1358
1359         for (i = 0; i < npages; i++) {
1360                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1361                                           __GFP_ZERO | GFP_IOFS);
1362                 if (page == NULL) {
1363                         while (--i >= 0)
1364                                 __free_page(rb->rb_kiov[i].kiov_page);
1365
1366                         LIBCFS_FREE(rb, sz);
1367                         return NULL;
1368                 }
1369
1370                 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1371                 rb->rb_kiov[i].kiov_offset = 0;
1372                 rb->rb_kiov[i].kiov_page = page;
1373         }
1374
1375         return rb;
1376 }
1377
1378 void
1379 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1380 {
1381         int              npages = rbp->rbp_npages;
1382         lnet_rtrbuf_t    *rb;
1383         struct list_head tmp;
1384
1385         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1386                 return;
1387
1388         INIT_LIST_HEAD(&tmp);
1389
1390         lnet_net_lock(cpt);
1391         lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, cpt);
1392         list_splice_init(&rbp->rbp_bufs, &tmp);
1393         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1394         rbp->rbp_mincredits = 0;
1395         lnet_net_unlock(cpt);
1396
1397         /* Free buffers on the free list. */
1398         while (!list_empty(&tmp)) {
1399                 rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
1400                 list_del(&rb->rb_list);
1401                 lnet_destroy_rtrbuf(rb, npages);
1402         }
1403 }
1404
1405 static int
1406 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1407 {
1408         struct list_head rb_list;
1409         lnet_rtrbuf_t   *rb;
1410         int             num_rb;
1411         int             num_buffers = 0;
1412         int             npages = rbp->rbp_npages;
1413
1414         /* If we are called for less buffers than already in the pool, we
1415          * just lower the nbuffers number and excess buffers will be
1416          * thrown away as they are returned to the free list.  Credits
1417          * then get adjusted as well. */
1418         if (nbufs <= rbp->rbp_nbuffers) {
1419                 lnet_net_lock(cpt);
1420                 rbp->rbp_nbuffers = nbufs;
1421                 lnet_net_unlock(cpt);
1422                 return 0;
1423         }
1424
1425         INIT_LIST_HEAD(&rb_list);
1426
1427         /* allocate the buffers on a local list first.  If all buffers are
1428          * allocated successfully then join this list to the rbp buffer
1429          * list.  If not then free all allocated buffers. */
1430         num_rb = rbp->rbp_nbuffers;
1431
1432         while (num_rb < nbufs) {
1433                 rb = lnet_new_rtrbuf(rbp, cpt);
1434                 if (rb == NULL) {
1435                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1436                                nbufs, npages);
1437                         goto failed;
1438                 }
1439
1440                 list_add(&rb->rb_list, &rb_list);
1441                 num_buffers++;
1442                 num_rb++;
1443         }
1444
1445         lnet_net_lock(cpt);
1446
1447         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1448         rbp->rbp_nbuffers += num_buffers;
1449         rbp->rbp_credits += num_buffers;
1450         rbp->rbp_mincredits = rbp->rbp_credits;
1451         /* We need to schedule blocked msg using the newly
1452          * added buffers. */
1453         while (!list_empty(&rbp->rbp_bufs) &&
1454                !list_empty(&rbp->rbp_msgs))
1455                 lnet_schedule_blocked_locked(rbp);
1456
1457         lnet_net_unlock(cpt);
1458
1459         return 0;
1460
1461 failed:
1462         while (!list_empty(&rb_list)) {
1463                 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1464                 list_del(&rb->rb_list);
1465                 lnet_destroy_rtrbuf(rb, npages);
1466         }
1467
1468         return -ENOMEM;
1469 }
1470
1471 void
1472 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1473 {
1474         INIT_LIST_HEAD(&rbp->rbp_msgs);
1475         INIT_LIST_HEAD(&rbp->rbp_bufs);
1476
1477         rbp->rbp_npages = npages;
1478         rbp->rbp_credits = 0;
1479         rbp->rbp_mincredits = 0;
1480 }
1481
1482 void
1483 lnet_rtrpools_free(int keep_pools)
1484 {
1485         lnet_rtrbufpool_t *rtrp;
1486         int               i;
1487
1488         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1489                 return;
1490
1491         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1492                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1493                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1494                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1495         }
1496
1497         if (!keep_pools) {
1498                 cfs_percpt_free(the_lnet.ln_rtrpools);
1499                 the_lnet.ln_rtrpools = NULL;
1500         }
1501 }
1502
1503 static int
1504 lnet_nrb_tiny_calculate(void)
1505 {
1506         int     nrbs = LNET_NRB_TINY;
1507
1508         if (tiny_router_buffers < 0) {
1509                 LCONSOLE_ERROR_MSG(0x10c,
1510                                    "tiny_router_buffers=%d invalid when "
1511                                    "routing enabled\n", tiny_router_buffers);
1512                 return -1;
1513         }
1514
1515         if (tiny_router_buffers > 0)
1516                 nrbs = tiny_router_buffers;
1517
1518         nrbs /= LNET_CPT_NUMBER;
1519         return max(nrbs, LNET_NRB_TINY_MIN);
1520 }
1521
1522 static int
1523 lnet_nrb_small_calculate(void)
1524 {
1525         int     nrbs = LNET_NRB_SMALL;
1526
1527         if (small_router_buffers < 0) {
1528                 LCONSOLE_ERROR_MSG(0x10c,
1529                                    "small_router_buffers=%d invalid when "
1530                                    "routing enabled\n", small_router_buffers);
1531                 return -1;
1532         }
1533
1534         if (small_router_buffers > 0)
1535                 nrbs = small_router_buffers;
1536
1537         nrbs /= LNET_CPT_NUMBER;
1538         return max(nrbs, LNET_NRB_SMALL_MIN);
1539 }
1540
1541 static int
1542 lnet_nrb_large_calculate(void)
1543 {
1544         int     nrbs = LNET_NRB_LARGE;
1545
1546         if (large_router_buffers < 0) {
1547                 LCONSOLE_ERROR_MSG(0x10c,
1548                                    "large_router_buffers=%d invalid when "
1549                                    "routing enabled\n", large_router_buffers);
1550                 return -1;
1551         }
1552
1553         if (large_router_buffers > 0)
1554                 nrbs = large_router_buffers;
1555
1556         nrbs /= LNET_CPT_NUMBER;
1557         return max(nrbs, LNET_NRB_LARGE_MIN);
1558 }
1559
1560 int
1561 lnet_rtrpools_alloc(int im_a_router)
1562 {
1563         lnet_rtrbufpool_t *rtrp;
1564         int     nrb_tiny;
1565         int     nrb_small;
1566         int     nrb_large;
1567         int     rc;
1568         int     i;
1569
1570         if (!strcmp(forwarding, "")) {
1571                 /* not set either way */
1572                 if (!im_a_router)
1573                         return 0;
1574         } else if (!strcmp(forwarding, "disabled")) {
1575                 /* explicitly disabled */
1576                 return 0;
1577         } else if (!strcmp(forwarding, "enabled")) {
1578                 /* explicitly enabled */
1579         } else {
1580                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1581                                    "'enabled' or 'disabled'\n");
1582                 return -EINVAL;
1583         }
1584
1585         nrb_tiny = lnet_nrb_tiny_calculate();
1586         if (nrb_tiny < 0)
1587                 return -EINVAL;
1588
1589         nrb_small = lnet_nrb_small_calculate();
1590         if (nrb_small < 0)
1591                 return -EINVAL;
1592
1593         nrb_large = lnet_nrb_large_calculate();
1594         if (nrb_large < 0)
1595                 return -EINVAL;
1596
1597         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1598                                                 LNET_NRBPOOLS *
1599                                                 sizeof(lnet_rtrbufpool_t));
1600         if (the_lnet.ln_rtrpools == NULL) {
1601                 LCONSOLE_ERROR_MSG(0x10c,
1602                                    "Failed to initialize router buffe pool\n");
1603                 return -ENOMEM;
1604         }
1605
1606         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1607                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1608                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1609                                               nrb_tiny, i);
1610                 if (rc != 0)
1611                         goto failed;
1612
1613                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1614                                   LNET_NRB_SMALL_PAGES);
1615                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1616                                               nrb_small, i);
1617                 if (rc != 0)
1618                         goto failed;
1619
1620                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1621                                   LNET_NRB_LARGE_PAGES);
1622                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1623                                               nrb_large, i);
1624                 if (rc != 0)
1625                         goto failed;
1626         }
1627
1628         lnet_net_lock(LNET_LOCK_EX);
1629         the_lnet.ln_routing = 1;
1630         lnet_net_unlock(LNET_LOCK_EX);
1631         return 0;
1632
1633  failed:
1634         lnet_rtrpools_free(0);
1635         return rc;
1636 }
1637
1638 static int
1639 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1640 {
1641         int nrb = 0;
1642         int rc = 0;
1643         int i;
1644         lnet_rtrbufpool_t *rtrp;
1645
1646         /* If the provided values for each buffer pool are different than the
1647          * configured values, we need to take action. */
1648         if (tiny >= 0) {
1649                 tiny_router_buffers = tiny;
1650                 nrb = lnet_nrb_tiny_calculate();
1651                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1652                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1653                                                       nrb, i);
1654                         if (rc != 0)
1655                                 return rc;
1656                 }
1657         }
1658         if (small >= 0) {
1659                 small_router_buffers = small;
1660                 nrb = lnet_nrb_small_calculate();
1661                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1662                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1663                                                       nrb, i);
1664                         if (rc != 0)
1665                                 return rc;
1666                 }
1667         }
1668         if (large >= 0) {
1669                 large_router_buffers = large;
1670                 nrb = lnet_nrb_large_calculate();
1671                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1672                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1673                                                       nrb, i);
1674                         if (rc != 0)
1675                                 return rc;
1676                 }
1677         }
1678
1679         return 0;
1680 }
1681
1682 int
1683 lnet_rtrpools_adjust(int tiny, int small, int large)
1684 {
1685         /* this function doesn't revert the changes if adding new buffers
1686          * failed.  It's up to the user space caller to revert the
1687          * changes. */
1688
1689         if (!the_lnet.ln_routing)
1690                 return 0;
1691
1692         return lnet_rtrpools_adjust_helper(tiny, small, large);
1693 }
1694
1695 int
1696 lnet_rtrpools_enable(void)
1697 {
1698         int rc;
1699
1700         if (the_lnet.ln_routing)
1701                 return 0;
1702
1703         if (the_lnet.ln_rtrpools == NULL)
1704                 /* If routing is turned off, and we have never
1705                  * initialized the pools before, just call the
1706                  * standard buffer pool allocation routine as
1707                  * if we are just configuring this for the first
1708                  * time. */
1709                 return lnet_rtrpools_alloc(1);
1710
1711         rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1712         if (rc != 0)
1713                 return rc;
1714
1715         lnet_net_lock(LNET_LOCK_EX);
1716         the_lnet.ln_routing = 1;
1717
1718         the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1719         lnet_net_unlock(LNET_LOCK_EX);
1720
1721         return 0;
1722 }
1723
1724 void
1725 lnet_rtrpools_disable(void)
1726 {
1727         if (!the_lnet.ln_routing)
1728                 return;
1729
1730         lnet_net_lock(LNET_LOCK_EX);
1731         the_lnet.ln_routing = 0;
1732         the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1733
1734         tiny_router_buffers = 0;
1735         small_router_buffers = 0;
1736         large_router_buffers = 0;
1737         lnet_net_unlock(LNET_LOCK_EX);
1738         lnet_rtrpools_free(1);
1739 }
1740
1741 int
1742 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1743 {
1744         struct lnet_peer        *lp = NULL;
1745         cfs_time_t              now = cfs_time_current();
1746         int                     cpt = lnet_cpt_of_nid(nid);
1747
1748         LASSERT (!in_interrupt ());
1749
1750         CDEBUG (D_NET, "%s notifying %s: %s\n",
1751                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1752                 libcfs_nid2str(nid),
1753                 alive ? "up" : "down");
1754
1755         if (ni != NULL &&
1756             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1757                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1758                         libcfs_nid2str(nid), alive ? "birth" : "death",
1759                         libcfs_nid2str(ni->ni_nid));
1760                 return -EINVAL;
1761         }
1762
1763         /* can't do predictions... */
1764         if (cfs_time_after(when, now)) {
1765                 CWARN ("Ignoring prediction from %s of %s %s "
1766                        "%ld seconds in the future\n",
1767                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1768                        libcfs_nid2str(nid), alive ? "up" : "down",
1769                        cfs_duration_sec(cfs_time_sub(when, now)));
1770                 return -EINVAL;
1771         }
1772
1773         if (ni != NULL && !alive &&             /* LND telling me she's down */
1774             !auto_down) {                       /* auto-down disabled */
1775                 CDEBUG(D_NET, "Auto-down disabled\n");
1776                 return 0;
1777         }
1778
1779         lnet_net_lock(cpt);
1780
1781         if (the_lnet.ln_shutdown) {
1782                 lnet_net_unlock(cpt);
1783                 return -ESHUTDOWN;
1784         }
1785
1786         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1787         if (lp == NULL) {
1788                 /* nid not found */
1789                 lnet_net_unlock(cpt);
1790                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1791                 return 0;
1792         }
1793
1794         /* We can't fully trust LND on reporting exact peer last_alive
1795          * if he notifies us about dead peer. For example ksocklnd can
1796          * call us with when == _time_when_the_node_was_booted_ if
1797          * no connections were successfully established */
1798         if (ni != NULL && !alive && when < lp->lp_last_alive)
1799                 when = lp->lp_last_alive;
1800
1801         lnet_notify_locked(lp, ni == NULL, alive, when);
1802
1803         if (ni != NULL)
1804                 lnet_ni_notify_locked(ni, lp);
1805
1806         lnet_peer_decref_locked(lp);
1807
1808         lnet_net_unlock(cpt);
1809         return 0;
1810 }
1811 EXPORT_SYMBOL(lnet_notify);
1812
1813 void
1814 lnet_get_tunables (void)
1815 {
1816         return;
1817 }
1818
1819 #else
1820
1821 int
1822 lnet_notify (lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1823 {
1824         return -EOPNOTSUPP;
1825 }
1826
1827 void
1828 lnet_router_checker (void)
1829 {
1830         static time_t last = 0;
1831         static int    running = 0;
1832
1833         time_t            now = cfs_time_current_sec();
1834         int               interval = now - last;
1835         int               rc;
1836         __u64             version;
1837         lnet_peer_t      *rtr;
1838
1839         /* It's no use to call me again within a sec - all intervals and
1840          * timeouts are measured in seconds */
1841         if (last != 0 && interval < 2)
1842                 return;
1843
1844         if (last != 0 &&
1845             interval > MAX(live_router_check_interval,
1846                            dead_router_check_interval))
1847                 CNETERR("Checker(%d/%d) not called for %d seconds\n",
1848                         live_router_check_interval, dead_router_check_interval,
1849                         interval);
1850
1851         LASSERT(LNET_CPT_NUMBER == 1);
1852
1853         lnet_net_lock(0);
1854         LASSERT(!running); /* recursion check */
1855         running = 1;
1856         lnet_net_unlock(0);
1857
1858         last = now;
1859
1860         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING)
1861                 lnet_prune_rc_data(0); /* unlink all rcd and nowait */
1862
1863         /* consume all pending events */
1864         while (1) {
1865                 int          i;
1866                 lnet_event_t ev;
1867
1868                 /* NB ln_rc_eqh must be the 1st in 'eventqs' otherwise the
1869                  * recursion breaker in LNetEQPoll would fail */
1870                 rc = LNetEQPoll(&the_lnet.ln_rc_eqh, 1, 0, &ev, &i);
1871                 if (rc == 0)   /* no event pending */
1872                         break;
1873
1874                 /* NB a lost SENT prevents me from pinging a router again */
1875                 if (rc == -EOVERFLOW) {
1876                         CERROR("Dropped an event!!!\n");
1877                         abort();
1878                 }
1879
1880                 LASSERT (rc == 1);
1881
1882                 lnet_router_checker_event(&ev);
1883         }
1884
1885         if (the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING) {
1886                 lnet_prune_rc_data(1); /* release rcd */
1887                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1888                 running = 0;
1889                 return;
1890         }
1891
1892         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1893
1894         lnet_net_lock(0);
1895
1896         version = the_lnet.ln_routers_version;
1897         list_for_each_entry(rtr, &the_lnet.ln_routers, lp_rtr_list) {
1898                 lnet_ping_router_locked(rtr);
1899                 LASSERT(version == the_lnet.ln_routers_version);
1900         }
1901
1902         lnet_net_unlock(0);
1903
1904         running = 0; /* lock only needed for the recursion check */
1905         return;
1906 }
1907
1908 /* NB lnet_peers_start_down depends on me,
1909  * so must be called before any peer creation */
1910 void
1911 lnet_get_tunables (void)
1912 {
1913         char *s;
1914
1915         s = getenv("LNET_ROUTER_PING_TIMEOUT");
1916         if (s != NULL) router_ping_timeout = atoi(s);
1917
1918         s = getenv("LNET_LIVE_ROUTER_CHECK_INTERVAL");
1919         if (s != NULL) live_router_check_interval = atoi(s);
1920
1921         s = getenv("LNET_DEAD_ROUTER_CHECK_INTERVAL");
1922         if (s != NULL) dead_router_check_interval = atoi(s);
1923
1924         /* This replaces old lnd_notify mechanism */
1925         check_routers_before_use = 1;
1926         if (dead_router_check_interval <= 0)
1927                 dead_router_check_interval = 30;
1928 }
1929
1930 void
1931 lnet_rtrpools_free(int keep_pools)
1932 {
1933 }
1934
1935 int
1936 lnet_rtrpools_alloc(int im_a_arouter)
1937 {
1938         return 0;
1939 }
1940
1941 #endif