lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2014, 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 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
  28 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
  29 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
  30 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
  31 #define LNET_NRB_SMALL_PAGES    1
  32 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  33 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  34 #define LNET_NRB_LARGE_PAGES    ((LNET_MTU + PAGE_CACHE_SIZE - 1) >> \
  35                                   PAGE_CACHE_SHIFT)
  36
  37 static char *forwarding = "";
  38 CFS_MODULE_PARM(forwarding, "s", charp, 0444,
  39                 "Explicitly enable/disable forwarding between networks");
  40
  41 static int tiny_router_buffers;
  42 CFS_MODULE_PARM(tiny_router_buffers, "i", int, 0444,
  43                 "# of 0 payload messages to buffer in the router");
  44 static int small_router_buffers;
  45 CFS_MODULE_PARM(small_router_buffers, "i", int, 0444,
  46                 "# of small (1 page) messages to buffer in the router");
  47 static int large_router_buffers;
  48 CFS_MODULE_PARM(large_router_buffers, "i", int, 0444,
  49                 "# of large messages to buffer in the router");
  50 static int peer_buffer_credits = 0;
  51 CFS_MODULE_PARM(peer_buffer_credits, "i", int, 0444,
  52                 "# router buffer credits per peer");
  53
  54 static int auto_down = 1;
  55 CFS_MODULE_PARM(auto_down, "i", int, 0444,
  56                 "Automatically mark peers down on comms error");
  57
  58 int
  59 lnet_peer_buffer_credits(lnet_ni_t *ni)
  60 {
  61         /* NI option overrides LNet default */
  62         if (ni->ni_peerrtrcredits > 0)
  63                 return ni->ni_peerrtrcredits;
  64         if (peer_buffer_credits > 0)
  65                 return peer_buffer_credits;
  66
  67         /* As an approximation, allow this peer the same number of router
  68          * buffers as it is allowed outstanding sends */
  69         return ni->ni_peertxcredits;
  70 }
  71
  72 /* forward ref's */
  73 static int lnet_router_checker(void *);
  74
  75 static int check_routers_before_use = 0;
  76 CFS_MODULE_PARM(check_routers_before_use, "i", int, 0444,
  77                 "Assume routers are down and ping them before use");
  78
  79 int avoid_asym_router_failure = 1;
  80 CFS_MODULE_PARM(avoid_asym_router_failure, "i", int, 0644,
  81                 "Avoid asymmetrical router failures (0 to disable)");
  82
  83 static int dead_router_check_interval = 60;
  84 CFS_MODULE_PARM(dead_router_check_interval, "i", int, 0644,
  85                 "Seconds between dead router health checks (<= 0 to disable)");
  86
  87 static int live_router_check_interval = 60;
  88 CFS_MODULE_PARM(live_router_check_interval, "i", int, 0644,
  89                 "Seconds between live router health checks (<= 0 to disable)");
  90
  91 static int router_ping_timeout = 50;
  92 CFS_MODULE_PARM(router_ping_timeout, "i", int, 0644,
  93                 "Seconds to wait for the reply to a router health query");
  94
  95 int
  96 lnet_peers_start_down(void)
  97 {
  98         return check_routers_before_use;
  99 }
 100
 101 void
 102 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive, cfs_time_t when)
 103 {
 104         if (cfs_time_before(when, lp->lp_timestamp)) { /* out of date information */
 105                 CDEBUG(D_NET, "Out of date\n");
 106                 return;
 107         }
 108
 109         lp->lp_timestamp = when;                /* update timestamp */
 110         lp->lp_ping_deadline = 0;               /* disable ping timeout */
 111
 112         if (lp->lp_alive_count != 0 &&          /* got old news */
 113             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
 114                 CDEBUG(D_NET, "Old news\n");
 115                 return;
 116         }
 117
 118         /* Flag that notification is outstanding */
 119
 120         lp->lp_alive_count++;
 121         lp->lp_alive = !(!alive);               /* 1 bit! */
 122         lp->lp_notify = 1;
 123         lp->lp_notifylnd |= notifylnd;
 124         if (lp->lp_alive)
 125                 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 126
 127         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
 128 }
 129
 130 static void
 131 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
 132 {
 133         int        alive;
 134         int        notifylnd;
 135
 136         /* Notify only in 1 thread at any time to ensure ordered notification.
 137          * NB individual events can be missed; the only guarantee is that you
 138          * always get the most recent news */
 139
 140         if (lp->lp_notifying)
 141                 return;
 142
 143         lp->lp_notifying = 1;
 144
 145         while (lp->lp_notify) {
 146                 alive     = lp->lp_alive;
 147                 notifylnd = lp->lp_notifylnd;
 148
 149                 lp->lp_notifylnd = 0;
 150                 lp->lp_notify    = 0;
 151
 152                 if (notifylnd && ni->ni_lnd->lnd_notify != NULL) {
 153                         lnet_net_unlock(lp->lp_cpt);
 154
 155                         /* A new notification could happen now; I'll handle it
 156                          * when control returns to me */
 157
 158                         (ni->ni_lnd->lnd_notify)(ni, lp->lp_nid, alive);
 159
 160                         lnet_net_lock(lp->lp_cpt);
 161                 }
 162         }
 163
 164         lp->lp_notifying = 0;
 165 }
 166
 167
 168 static void
 169 lnet_rtr_addref_locked(lnet_peer_t *lp)
 170 {
 171         LASSERT(lp->lp_refcount > 0);
 172         LASSERT(lp->lp_rtr_refcount >= 0);
 173
 174         /* lnet_net_lock must be exclusively locked */
 175         lp->lp_rtr_refcount++;
 176         if (lp->lp_rtr_refcount == 1) {
 177                 struct list_head *pos;
 178
 179                 /* a simple insertion sort */
 180                 list_for_each_prev(pos, &the_lnet.ln_routers) {
 181                         lnet_peer_t *rtr = list_entry(pos, lnet_peer_t,
 182                                                       lp_rtr_list);
 183
 184                         if (rtr->lp_nid < lp->lp_nid)
 185                                 break;
 186                 }
 187
 188                 list_add(&lp->lp_rtr_list, pos);
 189                 /* addref for the_lnet.ln_routers */
 190                 lnet_peer_addref_locked(lp);
 191                 the_lnet.ln_routers_version++;
 192         }
 193 }
 194
 195 static void
 196 lnet_rtr_decref_locked(lnet_peer_t *lp)
 197 {
 198         LASSERT(lp->lp_refcount > 0);
 199         LASSERT(lp->lp_rtr_refcount > 0);
 200
 201         /* lnet_net_lock must be exclusively locked */
 202         lp->lp_rtr_refcount--;
 203         if (lp->lp_rtr_refcount == 0) {
 204                 LASSERT(list_empty(&lp->lp_routes));
 205
 206                 if (lp->lp_rcd != NULL) {
 207                         list_add(&lp->lp_rcd->rcd_list,
 208                                  &the_lnet.ln_rcd_deathrow);
 209                         lp->lp_rcd = NULL;
 210                 }
 211
 212                 list_del(&lp->lp_rtr_list);
 213                 /* decref for the_lnet.ln_routers */
 214                 lnet_peer_decref_locked(lp);
 215                 the_lnet.ln_routers_version++;
 216         }
 217 }
 218
 219 lnet_remotenet_t *
 220 lnet_find_net_locked (__u32 net)
 221 {
 222         lnet_remotenet_t *rnet;
 223         struct list_head *tmp;
 224         struct list_head *rn_list;
 225
 226         LASSERT(!the_lnet.ln_shutdown);
 227
 228         rn_list = lnet_net2rnethash(net);
 229         list_for_each(tmp, rn_list) {
 230                 rnet = list_entry(tmp, lnet_remotenet_t, lrn_list);
 231
 232                 if (rnet->lrn_net == net)
 233                         return rnet;
 234         }
 235         return NULL;
 236 }
 237
 238 static void lnet_shuffle_seed(void)
 239 {
 240         static int seeded;
 241         __u32 lnd_type;
 242         __u32 seed[2];
 243         struct timeval tv;
 244         lnet_ni_t *ni;
 245         struct list_head *tmp;
 246
 247         if (seeded)
 248                 return;
 249
 250         cfs_get_random_bytes(seed, sizeof(seed));
 251
 252         /* Nodes with small feet have little entropy
 253          * the NID for this node gives the most entropy in the low bits */
 254         list_for_each(tmp, &the_lnet.ln_nis) {
 255                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 256                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 257
 258                 if (lnd_type != LOLND)
 259                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
 260         }
 261
 262         do_gettimeofday(&tv);
 263         cfs_srand(tv.tv_sec ^ seed[0], tv.tv_usec ^ seed[1]);
 264         seeded = 1;
 265         return;
 266 }
 267
 268 /* NB expects LNET_LOCK held */
 269 static void
 270 lnet_add_route_to_rnet(lnet_remotenet_t *rnet, lnet_route_t *route)
 271 {
 272         unsigned int      len = 0;
 273         unsigned int      offset = 0;
 274         struct list_head *e;
 275
 276         lnet_shuffle_seed();
 277
 278         list_for_each(e, &rnet->lrn_routes) {
 279                 len++;
 280         }
 281
 282         /* len+1 positions to add a new entry, also prevents division by 0 */
 283         offset = cfs_rand() % (len + 1);
 284         list_for_each(e, &rnet->lrn_routes) {
 285                 if (offset == 0)
 286                         break;
 287                 offset--;
 288         }
 289         list_add(&route->lr_list, e);
 290         list_add(&route->lr_gwlist, &route->lr_gateway->lp_routes);
 291
 292         the_lnet.ln_remote_nets_version++;
 293         lnet_rtr_addref_locked(route->lr_gateway);
 294 }
 295
 296 int
 297 lnet_add_route(__u32 net, unsigned int hops, lnet_nid_t gateway,
 298                unsigned int priority)
 299 {
 300         struct list_head        *e;
 301         lnet_remotenet_t        *rnet;
 302         lnet_remotenet_t        *rnet2;
 303         lnet_route_t            *route;
 304         lnet_ni_t               *ni;
 305         int                     add_route;
 306         int                     rc;
 307
 308         CDEBUG(D_NET, "Add route: net %s hops %u priority %u gw %s\n",
 309                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 310
 311         if (gateway == LNET_NID_ANY ||
 312             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 313             net == LNET_NIDNET(LNET_NID_ANY) ||
 314             LNET_NETTYP(net) == LOLND ||
 315             LNET_NIDNET(gateway) == net ||
 316             hops < 1 || hops > 255)
 317                 return -EINVAL;
 318
 319         if (lnet_islocalnet(net))       /* it's a local network */
 320                 return -EEXIST;
 321
 322         /* Assume net, route, all new */
 323         LIBCFS_ALLOC(route, sizeof(*route));
 324         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 325         if (route == NULL || rnet == NULL) {
 326                 CERROR("Out of memory creating route %s %d %s\n",
 327                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 328                 if (route != NULL)
 329                         LIBCFS_FREE(route, sizeof(*route));
 330                 if (rnet != NULL)
 331                         LIBCFS_FREE(rnet, sizeof(*rnet));
 332                 return -ENOMEM;
 333         }
 334
 335         INIT_LIST_HEAD(&rnet->lrn_routes);
 336         rnet->lrn_net = net;
 337         route->lr_hops = hops;
 338         route->lr_net = net;
 339         route->lr_priority = priority;
 340
 341         lnet_net_lock(LNET_LOCK_EX);
 342
 343         rc = lnet_nid2peer_locked(&route->lr_gateway, gateway, LNET_LOCK_EX);
 344         if (rc != 0) {
 345                 lnet_net_unlock(LNET_LOCK_EX);
 346
 347                 LIBCFS_FREE(route, sizeof(*route));
 348                 LIBCFS_FREE(rnet, sizeof(*rnet));
 349
 350                 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
 351                         return rc;       /* ignore the route entry */
 352                 CERROR("Error %d creating route %s %d %s\n", rc,
 353                         libcfs_net2str(net), hops,
 354                         libcfs_nid2str(gateway));
 355                 return rc;
 356         }
 357
 358         LASSERT(!the_lnet.ln_shutdown);
 359
 360         rnet2 = lnet_find_net_locked(net);
 361         if (rnet2 == NULL) {
 362                 /* new network */
 363                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 364                 rnet2 = rnet;
 365         }
 366
 367         /* Search for a duplicate route (it's a NOOP if it is) */
 368         add_route = 1;
 369         list_for_each(e, &rnet2->lrn_routes) {
 370                 lnet_route_t *route2 = list_entry(e, lnet_route_t, lr_list);
 371
 372                 if (route2->lr_gateway == route->lr_gateway) {
 373                         add_route = 0;
 374                         break;
 375                 }
 376
 377                 /* our lookups must be true */
 378                 LASSERT(route2->lr_gateway->lp_nid != gateway);
 379         }
 380
 381         if (add_route) {
 382                 lnet_peer_addref_locked(route->lr_gateway); /* +1 for notify */
 383                 lnet_add_route_to_rnet(rnet2, route);
 384
 385                 ni = route->lr_gateway->lp_ni;
 386                 lnet_net_unlock(LNET_LOCK_EX);
 387
 388                 /* XXX Assume alive */
 389                 if (ni->ni_lnd->lnd_notify != NULL)
 390                         (ni->ni_lnd->lnd_notify)(ni, gateway, 1);
 391
 392                 lnet_net_lock(LNET_LOCK_EX);
 393         }
 394
 395         /* -1 for notify or !add_route */
 396         lnet_peer_decref_locked(route->lr_gateway);
 397         lnet_net_unlock(LNET_LOCK_EX);
 398
 399         rc = 0;
 400
 401         if (!add_route) {
 402                 rc = -EEXIST;
 403                 LIBCFS_FREE(route, sizeof(*route));
 404         }
 405
 406         if (rnet != rnet2)
 407                 LIBCFS_FREE(rnet, sizeof(*rnet));
 408
 409         /* indicate to startup the router checker if configured */
 410         wake_up(&the_lnet.ln_rc_waitq);
 411
 412         return rc;
 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         struct list_head *e1;
 422         struct list_head *e2;
 423         int               cpt;
 424         struct list_head *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                 list_for_each(e1, rn_list) {
 432                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 433
 434                         route2 = NULL;
 435                         list_for_each(e2, &rnet->lrn_routes) {
 436                                 lnet_nid_t      nid1;
 437                                 lnet_nid_t      nid2;
 438                                 int             net;
 439
 440                                 route = 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         struct list_head        *e1;
 479         struct list_head        *e2;
 480         int                     rc = -ENOENT;
 481         struct list_head        *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         list_for_each(e1, rn_list) {
 498                 rnet = 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                 list_for_each(e2, &rnet->lrn_routes) {
 505                         route = 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                         list_del(&route->lr_list);
 513                         list_del(&route->lr_gwlist);
 514                         the_lnet.ln_remote_nets_version++;
 515
 516                         if (list_empty(&rnet->lrn_routes))
 517                                 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 lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
 554 {
 555         int i, rc = -ENOENT, lidx, j;
 556
 557         if (the_lnet.ln_rtrpools == NULL)
 558                 return rc;
 559
 560         for (i = 0; i < LNET_NRBPOOLS; i++) {
 561                 lnet_rtrbufpool_t *rbp;
 562
 563                 lnet_net_lock(LNET_LOCK_EX);
 564                 lidx = idx;
 565                 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
 566                         if (lidx-- == 0) {
 567                                 rc = 0;
 568                                 pool_cfg->pl_pools[i].pl_npages =
 569                                         rbp[i].rbp_npages;
 570                                 pool_cfg->pl_pools[i].pl_nbuffers =
 571                                         rbp[i].rbp_nbuffers;
 572                                 pool_cfg->pl_pools[i].pl_credits =
 573                                         rbp[i].rbp_credits;
 574                                 pool_cfg->pl_pools[i].pl_mincredits =
 575                                         rbp[i].rbp_mincredits;
 576                                 break;
 577                         }
 578                 }
 579                 lnet_net_unlock(LNET_LOCK_EX);
 580         }
 581
 582         lnet_net_lock(LNET_LOCK_EX);
 583         pool_cfg->pl_routing = the_lnet.ln_routing;
 584         lnet_net_unlock(LNET_LOCK_EX);
 585
 586         return rc;
 587 }
 588
 589 int
 590 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 591                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 592 {
 593         struct list_head *e1;
 594         struct list_head *e2;
 595         lnet_remotenet_t *rnet;
 596         lnet_route_t     *route;
 597         int               cpt;
 598         int               i;
 599         struct list_head *rn_list;
 600
 601         cpt = lnet_net_lock_current();
 602
 603         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 604                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 605                 list_for_each(e1, rn_list) {
 606                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 607
 608                         list_for_each(e2, &rnet->lrn_routes) {
 609                                 route = list_entry(e2, lnet_route_t,
 610                                                    lr_list);
 611
 612                                 if (idx-- == 0) {
 613                                         *net      = rnet->lrn_net;
 614                                         *hops     = route->lr_hops;
 615                                         *priority = route->lr_priority;
 616                                         *gateway  = route->lr_gateway->lp_nid;
 617                                         *alive    =
 618                                                 route->lr_gateway->lp_alive &&
 619                                                         !route->lr_downis;
 620                                         lnet_net_unlock(cpt);
 621                                         return 0;
 622                                 }
 623                         }
 624                 }
 625         }
 626
 627         lnet_net_unlock(cpt);
 628         return -ENOENT;
 629 }
 630
 631 void
 632 lnet_swap_pinginfo(lnet_ping_info_t *info)
 633 {
 634         int               i;
 635         lnet_ni_status_t *stat;
 636
 637         __swab32s(&info->pi_magic);
 638         __swab32s(&info->pi_features);
 639         __swab32s(&info->pi_pid);
 640         __swab32s(&info->pi_nnis);
 641         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 642                 stat = &info->pi_ni[i];
 643                 __swab64s(&stat->ns_nid);
 644                 __swab32s(&stat->ns_status);
 645         }
 646         return;
 647 }
 648
 649 /**
 650  * parse router-checker pinginfo, record number of down NIs for remote
 651  * networks on that router.
 652  */
 653 static void
 654 lnet_parse_rc_info(lnet_rc_data_t *rcd)
 655 {
 656         lnet_ping_info_t        *info = rcd->rcd_pinginfo;
 657         struct lnet_peer        *gw   = rcd->rcd_gateway;
 658         lnet_route_t            *rte;
 659
 660         if (!gw->lp_alive)
 661                 return;
 662
 663         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 664                 lnet_swap_pinginfo(info);
 665
 666         /* NB always racing with network! */
 667         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 668                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 669                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
 670                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 671                 return;
 672         }
 673
 674         gw->lp_ping_feats = info->pi_features;
 675         if ((gw->lp_ping_feats & LNET_PING_FEAT_MASK) == 0) {
 676                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
 677                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
 678                 return; /* nothing I can understand */
 679         }
 680
 681         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
 682                 return; /* can't carry NI status info */
 683
 684         list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
 685                 int     down = 0;
 686                 int     up = 0;
 687                 int     i;
 688
 689                 if ((gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
 690                         rte->lr_downis = 1;
 691                         continue;
 692                 }
 693
 694                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 695                         lnet_ni_status_t *stat = &info->pi_ni[i];
 696                         lnet_nid_t       nid = stat->ns_nid;
 697
 698                         if (nid == LNET_NID_ANY) {
 699                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 700                                        libcfs_nid2str(gw->lp_nid));
 701                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 702                                 return;
 703                         }
 704
 705                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 706                                 continue;
 707
 708                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 709                                 down++;
 710                                 continue;
 711                         }
 712
 713                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 714                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 715                                         up = 1;
 716                                         break;
 717                                 }
 718                                 continue;
 719                         }
 720
 721                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 722                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
 723                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 724                         return;
 725                 }
 726
 727                 if (up) { /* ignore downed NIs if NI for dest network is up */
 728                         rte->lr_downis = 0;
 729                         continue;
 730                 }
 731                 /* if @down is zero and this route is single-hop, it means
 732                  * we can't find NI for target network */
 733                 if (down == 0 && rte->lr_hops == 1)
 734                         down = 1;
 735
 736                 rte->lr_downis = down;
 737         }
 738 }
 739
 740 static void
 741 lnet_router_checker_event(lnet_event_t *event)
 742 {
 743         lnet_rc_data_t          *rcd = event->md.user_ptr;
 744         struct lnet_peer        *lp;
 745
 746         LASSERT(rcd != NULL);
 747
 748         if (event->unlinked) {
 749                 LNetInvalidateHandle(&rcd->rcd_mdh);
 750                 return;
 751         }
 752
 753         LASSERT(event->type == LNET_EVENT_SEND ||
 754                 event->type == LNET_EVENT_REPLY);
 755
 756         lp = rcd->rcd_gateway;
 757         LASSERT(lp != NULL);
 758
 759          /* NB: it's called with holding lnet_res_lock, we have a few
 760           * places need to hold both locks at the same time, please take
 761           * care of lock ordering */
 762         lnet_net_lock(lp->lp_cpt);
 763         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
 764                 /* ignore if no longer a router or rcd is replaced */
 765                 goto out;
 766         }
 767
 768         if (event->type == LNET_EVENT_SEND) {
 769                 lp->lp_ping_notsent = 0;
 770                 if (event->status == 0)
 771                         goto out;
 772         }
 773
 774         /* LNET_EVENT_REPLY */
 775         /* A successful REPLY means the router is up.  If _any_ comms
 776          * to the router fail I assume it's down (this will happen if
 777          * we ping alive routers to try to detect router death before
 778          * apps get burned). */
 779
 780         lnet_notify_locked(lp, 1, (event->status == 0), cfs_time_current());
 781         /* The router checker will wake up very shortly and do the
 782          * actual notification.
 783          * XXX If 'lp' stops being a router before then, it will still
 784          * have the notification pending!!! */
 785
 786         if (avoid_asym_router_failure && event->status == 0)
 787                 lnet_parse_rc_info(rcd);
 788
 789  out:
 790         lnet_net_unlock(lp->lp_cpt);
 791 }
 792
 793 static void
 794 lnet_wait_known_routerstate(void)
 795 {
 796         lnet_peer_t      *rtr;
 797         struct list_head *entry;
 798         int               all_known;
 799
 800         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 801
 802         for (;;) {
 803                 int cpt = lnet_net_lock_current();
 804
 805                 all_known = 1;
 806                 list_for_each(entry, &the_lnet.ln_routers) {
 807                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
 808
 809                         if (rtr->lp_alive_count == 0) {
 810                                 all_known = 0;
 811                                 break;
 812                         }
 813                 }
 814
 815                 lnet_net_unlock(cpt);
 816
 817                 if (all_known)
 818                         return;
 819
 820                 set_current_state(TASK_UNINTERRUPTIBLE);
 821                 schedule_timeout(cfs_time_seconds(1));
 822         }
 823 }
 824
 825 void
 826 lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net)
 827 {
 828         lnet_route_t *rte;
 829
 830         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
 831                 list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
 832                         if (rte->lr_net == net) {
 833                                 rte->lr_downis = 0;
 834                                 break;
 835                         }
 836                 }
 837         }
 838 }
 839
 840 static void
 841 lnet_update_ni_status_locked(void)
 842 {
 843         lnet_ni_t       *ni;
 844         long            now;
 845         int             timeout;
 846
 847         LASSERT(the_lnet.ln_routing);
 848
 849         timeout = router_ping_timeout +
 850                   MAX(live_router_check_interval, dead_router_check_interval);
 851
 852         now = cfs_time_current_sec();
 853         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 854                 if (ni->ni_lnd->lnd_type == LOLND)
 855                         continue;
 856
 857                 if (now < ni->ni_last_alive + timeout)
 858                         continue;
 859
 860                 lnet_ni_lock(ni);
 861                 /* re-check with lock */
 862                 if (now < ni->ni_last_alive + timeout) {
 863                         lnet_ni_unlock(ni);
 864                         continue;
 865                 }
 866
 867                 LASSERT(ni->ni_status != NULL);
 868
 869                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 870                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 871                                libcfs_nid2str(ni->ni_nid), timeout);
 872                         /* NB: so far, this is the only place to set
 873                          * NI status to "down" */
 874                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 875                 }
 876                 lnet_ni_unlock(ni);
 877         }
 878 }
 879
 880 static void
 881 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
 882 {
 883         LASSERT(list_empty(&rcd->rcd_list));
 884         /* detached from network */
 885         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 886
 887         if (rcd->rcd_gateway != NULL) {
 888                 int cpt = rcd->rcd_gateway->lp_cpt;
 889
 890                 lnet_net_lock(cpt);
 891                 lnet_peer_decref_locked(rcd->rcd_gateway);
 892                 lnet_net_unlock(cpt);
 893         }
 894
 895         if (rcd->rcd_pinginfo != NULL)
 896                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 897
 898         LIBCFS_FREE(rcd, sizeof(*rcd));
 899 }
 900
 901 static lnet_rc_data_t *
 902 lnet_create_rc_data_locked(lnet_peer_t *gateway)
 903 {
 904         lnet_rc_data_t          *rcd = NULL;
 905         lnet_ping_info_t        *pi;
 906         int                     rc;
 907         int                     i;
 908
 909         lnet_net_unlock(gateway->lp_cpt);
 910
 911         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 912         if (rcd == NULL)
 913                 goto out;
 914
 915         LNetInvalidateHandle(&rcd->rcd_mdh);
 916         INIT_LIST_HEAD(&rcd->rcd_list);
 917
 918         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 919         if (pi == NULL)
 920                 goto out;
 921
 922         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 923                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 924                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 925         }
 926         rcd->rcd_pinginfo = pi;
 927
 928         LASSERT (!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 929         rc = LNetMDBind((lnet_md_t){.start     = pi,
 930                                     .user_ptr  = rcd,
 931                                     .length    = LNET_PINGINFO_SIZE,
 932                                     .threshold = LNET_MD_THRESH_INF,
 933                                     .options   = LNET_MD_TRUNCATE,
 934                                     .eq_handle = the_lnet.ln_rc_eqh},
 935                         LNET_UNLINK,
 936                         &rcd->rcd_mdh);
 937         if (rc < 0) {
 938                 CERROR("Can't bind MD: %d\n", rc);
 939                 goto out;
 940         }
 941         LASSERT(rc == 0);
 942
 943         lnet_net_lock(gateway->lp_cpt);
 944         /* router table changed or someone has created rcd for this gateway */
 945         if (!lnet_isrouter(gateway) || gateway->lp_rcd != NULL) {
 946                 lnet_net_unlock(gateway->lp_cpt);
 947                 goto out;
 948         }
 949
 950         lnet_peer_addref_locked(gateway);
 951         rcd->rcd_gateway = gateway;
 952         gateway->lp_rcd = rcd;
 953         gateway->lp_ping_notsent = 0;
 954
 955         return rcd;
 956
 957  out:
 958         if (rcd != NULL) {
 959                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
 960                         rc = LNetMDUnlink(rcd->rcd_mdh);
 961                         LASSERT(rc == 0);
 962                 }
 963                 lnet_destroy_rc_data(rcd);
 964         }
 965
 966         lnet_net_lock(gateway->lp_cpt);
 967         return gateway->lp_rcd;
 968 }
 969
 970 static int
 971 lnet_router_check_interval (lnet_peer_t *rtr)
 972 {
 973         int secs;
 974
 975         secs = rtr->lp_alive ? live_router_check_interval :
 976                                dead_router_check_interval;
 977         if (secs < 0)
 978                 secs = 0;
 979
 980         return secs;
 981 }
 982
 983 static void
 984 lnet_ping_router_locked (lnet_peer_t *rtr)
 985 {
 986         lnet_rc_data_t *rcd = NULL;
 987         cfs_time_t      now = cfs_time_current();
 988         int             secs;
 989
 990         lnet_peer_addref_locked(rtr);
 991
 992         if (rtr->lp_ping_deadline != 0 && /* ping timed out? */
 993             cfs_time_after(now, rtr->lp_ping_deadline))
 994                 lnet_notify_locked(rtr, 1, 0, now);
 995
 996         /* Run any outstanding notifications */
 997         lnet_ni_notify_locked(rtr->lp_ni, rtr);
 998
 999         if (!lnet_isrouter(rtr) ||
1000             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1001                 /* router table changed or router checker is shutting down */
1002                 lnet_peer_decref_locked(rtr);
1003                 return;
1004         }
1005
1006         rcd = rtr->lp_rcd != NULL ?
1007               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
1008
1009         if (rcd == NULL)
1010                 return;
1011
1012         secs = lnet_router_check_interval(rtr);
1013
1014         CDEBUG(D_NET,
1015                "rtr %s %d: deadline %lu ping_notsent %d alive %d "
1016                "alive_count %d lp_ping_timestamp %lu\n",
1017                libcfs_nid2str(rtr->lp_nid), secs,
1018                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
1019                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
1020
1021         if (secs != 0 && !rtr->lp_ping_notsent &&
1022             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
1023                                              cfs_time_seconds(secs)))) {
1024                 int               rc;
1025                 lnet_process_id_t id;
1026                 lnet_handle_md_t  mdh;
1027
1028                 id.nid = rtr->lp_nid;
1029                 id.pid = LNET_PID_LUSTRE;
1030                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1031
1032                 rtr->lp_ping_notsent   = 1;
1033                 rtr->lp_ping_timestamp = now;
1034
1035                 mdh = rcd->rcd_mdh;
1036
1037                 if (rtr->lp_ping_deadline == 0) {
1038                         rtr->lp_ping_deadline =
1039                                 cfs_time_shift(router_ping_timeout);
1040                 }
1041
1042                 lnet_net_unlock(rtr->lp_cpt);
1043
1044                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1045                              LNET_PROTO_PING_MATCHBITS, 0);
1046
1047                 lnet_net_lock(rtr->lp_cpt);
1048                 if (rc != 0)
1049                         rtr->lp_ping_notsent = 0; /* no event pending */
1050         }
1051
1052         lnet_peer_decref_locked(rtr);
1053         return;
1054 }
1055
1056 int
1057 lnet_router_checker_start(void)
1058 {
1059         int                     rc;
1060         int                     eqsz = 0;
1061         struct task_struct     *task;
1062
1063         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1064
1065         if (check_routers_before_use &&
1066             dead_router_check_interval <= 0) {
1067                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1068                                    " set if 'check_routers_before_use' is set"
1069                                    "\n");
1070                 return -EINVAL;
1071         }
1072
1073         sema_init(&the_lnet.ln_rc_signal, 0);
1074
1075         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1076         if (rc != 0) {
1077                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1078                 return -ENOMEM;
1079         }
1080
1081         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1082         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1083         if (IS_ERR(task)) {
1084                 rc = PTR_ERR(task);
1085                 CERROR("Can't start router checker thread: %d\n", rc);
1086                 /* block until event callback signals exit */
1087                 down(&the_lnet.ln_rc_signal);
1088                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1089                 LASSERT(rc == 0);
1090                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1091                 return -ENOMEM;
1092         }
1093
1094         if (check_routers_before_use) {
1095                 /* Note that a helpful side-effect of pinging all known routers
1096                  * at startup is that it makes them drop stale connections they
1097                  * may have to a previous instance of me. */
1098                 lnet_wait_known_routerstate();
1099         }
1100
1101         return 0;
1102 }
1103
1104 void
1105 lnet_router_checker_stop (void)
1106 {
1107         int rc;
1108
1109         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1110                 return;
1111
1112         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1113         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1114         /* wakeup the RC thread if it's sleeping */
1115         wake_up(&the_lnet.ln_rc_waitq);
1116
1117         /* block until event callback signals exit */
1118         down(&the_lnet.ln_rc_signal);
1119         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1120
1121         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1122         LASSERT (rc == 0);
1123         return;
1124 }
1125
1126 static void
1127 lnet_prune_rc_data(int wait_unlink)
1128 {
1129         lnet_rc_data_t          *rcd;
1130         lnet_rc_data_t          *tmp;
1131         lnet_peer_t             *lp;
1132         struct list_head         head;
1133         int                      i = 2;
1134
1135         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1136                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1137                    list_empty(&the_lnet.ln_rcd_zombie)))
1138                 return;
1139
1140         INIT_LIST_HEAD(&head);
1141
1142         lnet_net_lock(LNET_LOCK_EX);
1143
1144         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1145                 /* router checker is stopping, prune all */
1146                 list_for_each_entry(lp, &the_lnet.ln_routers,
1147                                     lp_rtr_list) {
1148                         if (lp->lp_rcd == NULL)
1149                                 continue;
1150
1151                         LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1152                         list_add(&lp->lp_rcd->rcd_list,
1153                                  &the_lnet.ln_rcd_deathrow);
1154                         lp->lp_rcd = NULL;
1155                 }
1156         }
1157
1158         /* unlink all RCDs on deathrow list */
1159         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1160
1161         if (!list_empty(&head)) {
1162                 lnet_net_unlock(LNET_LOCK_EX);
1163
1164                 list_for_each_entry(rcd, &head, rcd_list)
1165                         LNetMDUnlink(rcd->rcd_mdh);
1166
1167                 lnet_net_lock(LNET_LOCK_EX);
1168         }
1169
1170         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1171
1172         /* release all zombie RCDs */
1173         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1174                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1175                                          rcd_list) {
1176                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1177                                 list_move(&rcd->rcd_list, &head);
1178                 }
1179
1180                 wait_unlink = wait_unlink &&
1181                               !list_empty(&the_lnet.ln_rcd_zombie);
1182
1183                 lnet_net_unlock(LNET_LOCK_EX);
1184
1185                 while (!list_empty(&head)) {
1186                         rcd = list_entry(head.next,
1187                                          lnet_rc_data_t, rcd_list);
1188                         list_del_init(&rcd->rcd_list);
1189                         lnet_destroy_rc_data(rcd);
1190                 }
1191
1192                 if (!wait_unlink)
1193                         return;
1194
1195                 i++;
1196                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1197                        "Waiting for rc buffers to unlink\n");
1198                 set_current_state(TASK_UNINTERRUPTIBLE);
1199                 schedule_timeout(cfs_time_seconds(1) / 4);
1200
1201                 lnet_net_lock(LNET_LOCK_EX);
1202         }
1203
1204         lnet_net_unlock(LNET_LOCK_EX);
1205 }
1206
1207 /*
1208  * This function is called to check if the RC should block indefinitely.
1209  * It's called from lnet_router_checker() as well as being passed to
1210  * wait_event_interruptible() to avoid the lost wake_up problem.
1211  *
1212  * When it's called from wait_event_interruptible() it is necessary to
1213  * also not sleep if the rc state is not running to avoid a deadlock
1214  * when the system is shutting down
1215  */
1216 static inline bool
1217 lnet_router_checker_active(void)
1218 {
1219         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1220                 return true;
1221
1222         /* Router Checker thread needs to run when routing is enabled in
1223          * order to call lnet_update_ni_status_locked() */
1224         if (the_lnet.ln_routing)
1225                 return true;
1226
1227         return !list_empty(&the_lnet.ln_routers) &&
1228                 (live_router_check_interval > 0 ||
1229                  dead_router_check_interval > 0);
1230 }
1231
1232 static int
1233 lnet_router_checker(void *arg)
1234 {
1235         lnet_peer_t       *rtr;
1236         struct list_head  *entry;
1237
1238         cfs_block_allsigs();
1239
1240         LASSERT (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1241
1242         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1243                 __u64   version;
1244                 int     cpt;
1245                 int     cpt2;
1246
1247                 cpt = lnet_net_lock_current();
1248 rescan:
1249                 version = the_lnet.ln_routers_version;
1250
1251                 list_for_each(entry, &the_lnet.ln_routers) {
1252                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1253
1254                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1255                         if (cpt != cpt2) {
1256                                 lnet_net_unlock(cpt);
1257                                 cpt = cpt2;
1258                                 lnet_net_lock(cpt);
1259                                 /* the routers list has changed */
1260                                 if (version != the_lnet.ln_routers_version)
1261                                         goto rescan;
1262                         }
1263
1264                         lnet_ping_router_locked(rtr);
1265
1266                         /* NB dropped lock */
1267                         if (version != the_lnet.ln_routers_version) {
1268                                 /* the routers list has changed */
1269                                 goto rescan;
1270                         }
1271                 }
1272
1273                 if (the_lnet.ln_routing)
1274                         lnet_update_ni_status_locked();
1275
1276                 lnet_net_unlock(cpt);
1277
1278                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1279
1280                 /* Call schedule_timeout() here always adds 1 to load average
1281                  * because kernel counts # active tasks as nr_running
1282                  * + nr_uninterruptible. */
1283                 /* if there are any routes then wakeup every second.  If
1284                  * there are no routes then sleep indefinitely until woken
1285                  * up by a user adding a route */
1286                 if (!lnet_router_checker_active())
1287                         wait_event_interruptible(the_lnet.ln_rc_waitq,
1288                                                  lnet_router_checker_active());
1289                 else
1290                         wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1291                                                          false,
1292                                                          cfs_time_seconds(1));
1293         }
1294
1295         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1296
1297         lnet_prune_rc_data(1); /* wait for UNLINK */
1298
1299         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1300         up(&the_lnet.ln_rc_signal);
1301         /* The unlink event callback will signal final completion */
1302         return 0;
1303 }
1304
1305 void
1306 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1307 {
1308         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1309
1310         while (--npages >= 0)
1311                 __free_page(rb->rb_kiov[npages].kiov_page);
1312
1313         LIBCFS_FREE(rb, sz);
1314 }
1315
1316 static lnet_rtrbuf_t *
1317 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1318 {
1319         int            npages = rbp->rbp_npages;
1320         int            sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1321         struct page   *page;
1322         lnet_rtrbuf_t *rb;
1323         int            i;
1324
1325         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1326         if (rb == NULL)
1327                 return NULL;
1328
1329         rb->rb_pool = rbp;
1330
1331         for (i = 0; i < npages; i++) {
1332                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1333                                           __GFP_ZERO | GFP_IOFS);
1334                 if (page == NULL) {
1335                         while (--i >= 0)
1336                                 __free_page(rb->rb_kiov[i].kiov_page);
1337
1338                         LIBCFS_FREE(rb, sz);
1339                         return NULL;
1340                 }
1341
1342                 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1343                 rb->rb_kiov[i].kiov_offset = 0;
1344                 rb->rb_kiov[i].kiov_page = page;
1345         }
1346
1347         return rb;
1348 }
1349
1350 static void
1351 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1352 {
1353         int              npages = rbp->rbp_npages;
1354         lnet_rtrbuf_t    *rb;
1355         struct list_head tmp;
1356
1357         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1358                 return;
1359
1360         INIT_LIST_HEAD(&tmp);
1361
1362         lnet_net_lock(cpt);
1363         lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, cpt);
1364         list_splice_init(&rbp->rbp_bufs, &tmp);
1365         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1366         rbp->rbp_mincredits = 0;
1367         lnet_net_unlock(cpt);
1368
1369         /* Free buffers on the free list. */
1370         while (!list_empty(&tmp)) {
1371                 rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
1372                 list_del(&rb->rb_list);
1373                 lnet_destroy_rtrbuf(rb, npages);
1374         }
1375 }
1376
1377 static int
1378 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1379 {
1380         struct list_head rb_list;
1381         lnet_rtrbuf_t   *rb;
1382         int             num_rb;
1383         int             num_buffers = 0;
1384         int             npages = rbp->rbp_npages;
1385
1386         /* If we are called for less buffers than already in the pool, we
1387          * just lower the nbuffers number and excess buffers will be
1388          * thrown away as they are returned to the free list.  Credits
1389          * then get adjusted as well. */
1390         if (nbufs <= rbp->rbp_nbuffers) {
1391                 lnet_net_lock(cpt);
1392                 rbp->rbp_nbuffers = nbufs;
1393                 lnet_net_unlock(cpt);
1394                 return 0;
1395         }
1396
1397         INIT_LIST_HEAD(&rb_list);
1398
1399         /* allocate the buffers on a local list first.  If all buffers are
1400          * allocated successfully then join this list to the rbp buffer
1401          * list.  If not then free all allocated buffers. */
1402         num_rb = rbp->rbp_nbuffers;
1403
1404         while (num_rb < nbufs) {
1405                 rb = lnet_new_rtrbuf(rbp, cpt);
1406                 if (rb == NULL) {
1407                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1408                                nbufs, npages);
1409                         goto failed;
1410                 }
1411
1412                 list_add(&rb->rb_list, &rb_list);
1413                 num_buffers++;
1414                 num_rb++;
1415         }
1416
1417         lnet_net_lock(cpt);
1418
1419         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1420         rbp->rbp_nbuffers += num_buffers;
1421         rbp->rbp_credits += num_buffers;
1422         rbp->rbp_mincredits = rbp->rbp_credits;
1423         /* We need to schedule blocked msg using the newly
1424          * added buffers. */
1425         while (!list_empty(&rbp->rbp_bufs) &&
1426                !list_empty(&rbp->rbp_msgs))
1427                 lnet_schedule_blocked_locked(rbp);
1428
1429         lnet_net_unlock(cpt);
1430
1431         return 0;
1432
1433 failed:
1434         while (!list_empty(&rb_list)) {
1435                 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1436                 list_del(&rb->rb_list);
1437                 lnet_destroy_rtrbuf(rb, npages);
1438         }
1439
1440         return -ENOMEM;
1441 }
1442
1443 static void
1444 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1445 {
1446         INIT_LIST_HEAD(&rbp->rbp_msgs);
1447         INIT_LIST_HEAD(&rbp->rbp_bufs);
1448
1449         rbp->rbp_npages = npages;
1450         rbp->rbp_credits = 0;
1451         rbp->rbp_mincredits = 0;
1452 }
1453
1454 void
1455 lnet_rtrpools_free(int keep_pools)
1456 {
1457         lnet_rtrbufpool_t *rtrp;
1458         int               i;
1459
1460         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1461                 return;
1462
1463         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1464                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1465                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1466                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1467         }
1468
1469         if (!keep_pools) {
1470                 cfs_percpt_free(the_lnet.ln_rtrpools);
1471                 the_lnet.ln_rtrpools = NULL;
1472         }
1473 }
1474
1475 static int
1476 lnet_nrb_tiny_calculate(void)
1477 {
1478         int     nrbs = LNET_NRB_TINY;
1479
1480         if (tiny_router_buffers < 0) {
1481                 LCONSOLE_ERROR_MSG(0x10c,
1482                                    "tiny_router_buffers=%d invalid when "
1483                                    "routing enabled\n", tiny_router_buffers);
1484                 return -1;
1485         }
1486
1487         if (tiny_router_buffers > 0)
1488                 nrbs = tiny_router_buffers;
1489
1490         nrbs /= LNET_CPT_NUMBER;
1491         return max(nrbs, LNET_NRB_TINY_MIN);
1492 }
1493
1494 static int
1495 lnet_nrb_small_calculate(void)
1496 {
1497         int     nrbs = LNET_NRB_SMALL;
1498
1499         if (small_router_buffers < 0) {
1500                 LCONSOLE_ERROR_MSG(0x10c,
1501                                    "small_router_buffers=%d invalid when "
1502                                    "routing enabled\n", small_router_buffers);
1503                 return -1;
1504         }
1505
1506         if (small_router_buffers > 0)
1507                 nrbs = small_router_buffers;
1508
1509         nrbs /= LNET_CPT_NUMBER;
1510         return max(nrbs, LNET_NRB_SMALL_MIN);
1511 }
1512
1513 static int
1514 lnet_nrb_large_calculate(void)
1515 {
1516         int     nrbs = LNET_NRB_LARGE;
1517
1518         if (large_router_buffers < 0) {
1519                 LCONSOLE_ERROR_MSG(0x10c,
1520                                    "large_router_buffers=%d invalid when "
1521                                    "routing enabled\n", large_router_buffers);
1522                 return -1;
1523         }
1524
1525         if (large_router_buffers > 0)
1526                 nrbs = large_router_buffers;
1527
1528         nrbs /= LNET_CPT_NUMBER;
1529         return max(nrbs, LNET_NRB_LARGE_MIN);
1530 }
1531
1532 int
1533 lnet_rtrpools_alloc(int im_a_router)
1534 {
1535         lnet_rtrbufpool_t *rtrp;
1536         int     nrb_tiny;
1537         int     nrb_small;
1538         int     nrb_large;
1539         int     rc;
1540         int     i;
1541
1542         if (!strcmp(forwarding, "")) {
1543                 /* not set either way */
1544                 if (!im_a_router)
1545                         return 0;
1546         } else if (!strcmp(forwarding, "disabled")) {
1547                 /* explicitly disabled */
1548                 return 0;
1549         } else if (!strcmp(forwarding, "enabled")) {
1550                 /* explicitly enabled */
1551         } else {
1552                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1553                                    "'enabled' or 'disabled'\n");
1554                 return -EINVAL;
1555         }
1556
1557         nrb_tiny = lnet_nrb_tiny_calculate();
1558         if (nrb_tiny < 0)
1559                 return -EINVAL;
1560
1561         nrb_small = lnet_nrb_small_calculate();
1562         if (nrb_small < 0)
1563                 return -EINVAL;
1564
1565         nrb_large = lnet_nrb_large_calculate();
1566         if (nrb_large < 0)
1567                 return -EINVAL;
1568
1569         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1570                                                 LNET_NRBPOOLS *
1571                                                 sizeof(lnet_rtrbufpool_t));
1572         if (the_lnet.ln_rtrpools == NULL) {
1573                 LCONSOLE_ERROR_MSG(0x10c,
1574                                    "Failed to initialize router buffe pool\n");
1575                 return -ENOMEM;
1576         }
1577
1578         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1579                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1580                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1581                                               nrb_tiny, i);
1582                 if (rc != 0)
1583                         goto failed;
1584
1585                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1586                                   LNET_NRB_SMALL_PAGES);
1587                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1588                                               nrb_small, i);
1589                 if (rc != 0)
1590                         goto failed;
1591
1592                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1593                                   LNET_NRB_LARGE_PAGES);
1594                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1595                                               nrb_large, i);
1596                 if (rc != 0)
1597                         goto failed;
1598         }
1599
1600         lnet_net_lock(LNET_LOCK_EX);
1601         the_lnet.ln_routing = 1;
1602         lnet_net_unlock(LNET_LOCK_EX);
1603         return 0;
1604
1605  failed:
1606         lnet_rtrpools_free(0);
1607         return rc;
1608 }
1609
1610 static int
1611 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1612 {
1613         int nrb = 0;
1614         int rc = 0;
1615         int i;
1616         lnet_rtrbufpool_t *rtrp;
1617
1618         /* If the provided values for each buffer pool are different than the
1619          * configured values, we need to take action. */
1620         if (tiny >= 0) {
1621                 tiny_router_buffers = tiny;
1622                 nrb = lnet_nrb_tiny_calculate();
1623                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1624                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1625                                                       nrb, i);
1626                         if (rc != 0)
1627                                 return rc;
1628                 }
1629         }
1630         if (small >= 0) {
1631                 small_router_buffers = small;
1632                 nrb = lnet_nrb_small_calculate();
1633                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1634                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1635                                                       nrb, i);
1636                         if (rc != 0)
1637                                 return rc;
1638                 }
1639         }
1640         if (large >= 0) {
1641                 large_router_buffers = large;
1642                 nrb = lnet_nrb_large_calculate();
1643                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1644                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1645                                                       nrb, i);
1646                         if (rc != 0)
1647                                 return rc;
1648                 }
1649         }
1650
1651         return 0;
1652 }
1653
1654 int
1655 lnet_rtrpools_adjust(int tiny, int small, int large)
1656 {
1657         /* this function doesn't revert the changes if adding new buffers
1658          * failed.  It's up to the user space caller to revert the
1659          * changes. */
1660
1661         if (!the_lnet.ln_routing)
1662                 return 0;
1663
1664         return lnet_rtrpools_adjust_helper(tiny, small, large);
1665 }
1666
1667 int
1668 lnet_rtrpools_enable(void)
1669 {
1670         int rc;
1671
1672         if (the_lnet.ln_routing)
1673                 return 0;
1674
1675         if (the_lnet.ln_rtrpools == NULL)
1676                 /* If routing is turned off, and we have never
1677                  * initialized the pools before, just call the
1678                  * standard buffer pool allocation routine as
1679                  * if we are just configuring this for the first
1680                  * time. */
1681                 return lnet_rtrpools_alloc(1);
1682
1683         rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1684         if (rc != 0)
1685                 return rc;
1686
1687         lnet_net_lock(LNET_LOCK_EX);
1688         the_lnet.ln_routing = 1;
1689
1690         the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1691         lnet_net_unlock(LNET_LOCK_EX);
1692
1693         return 0;
1694 }
1695
1696 void
1697 lnet_rtrpools_disable(void)
1698 {
1699         if (!the_lnet.ln_routing)
1700                 return;
1701
1702         lnet_net_lock(LNET_LOCK_EX);
1703         the_lnet.ln_routing = 0;
1704         the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1705
1706         tiny_router_buffers = 0;
1707         small_router_buffers = 0;
1708         large_router_buffers = 0;
1709         lnet_net_unlock(LNET_LOCK_EX);
1710         lnet_rtrpools_free(1);
1711 }
1712
1713 int
1714 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, cfs_time_t when)
1715 {
1716         struct lnet_peer        *lp = NULL;
1717         cfs_time_t              now = cfs_time_current();
1718         int                     cpt = lnet_cpt_of_nid(nid);
1719
1720         LASSERT (!in_interrupt ());
1721
1722         CDEBUG (D_NET, "%s notifying %s: %s\n",
1723                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1724                 libcfs_nid2str(nid),
1725                 alive ? "up" : "down");
1726
1727         if (ni != NULL &&
1728             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1729                 CWARN ("Ignoring notification of %s %s by %s (different net)\n",
1730                         libcfs_nid2str(nid), alive ? "birth" : "death",
1731                         libcfs_nid2str(ni->ni_nid));
1732                 return -EINVAL;
1733         }
1734
1735         /* can't do predictions... */
1736         if (cfs_time_after(when, now)) {
1737                 CWARN ("Ignoring prediction from %s of %s %s "
1738                        "%ld seconds in the future\n",
1739                        (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1740                        libcfs_nid2str(nid), alive ? "up" : "down",
1741                        cfs_duration_sec(cfs_time_sub(when, now)));
1742                 return -EINVAL;
1743         }
1744
1745         if (ni != NULL && !alive &&             /* LND telling me she's down */
1746             !auto_down) {                       /* auto-down disabled */
1747                 CDEBUG(D_NET, "Auto-down disabled\n");
1748                 return 0;
1749         }
1750
1751         lnet_net_lock(cpt);
1752
1753         if (the_lnet.ln_shutdown) {
1754                 lnet_net_unlock(cpt);
1755                 return -ESHUTDOWN;
1756         }
1757
1758         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1759         if (lp == NULL) {
1760                 /* nid not found */
1761                 lnet_net_unlock(cpt);
1762                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1763                 return 0;
1764         }
1765
1766         /* We can't fully trust LND on reporting exact peer last_alive
1767          * if he notifies us about dead peer. For example ksocklnd can
1768          * call us with when == _time_when_the_node_was_booted_ if
1769          * no connections were successfully established */
1770         if (ni != NULL && !alive && when < lp->lp_last_alive)
1771                 when = lp->lp_last_alive;
1772
1773         lnet_notify_locked(lp, ni == NULL, alive, when);
1774
1775         if (ni != NULL)
1776                 lnet_ni_notify_locked(ni, lp);
1777
1778         lnet_peer_decref_locked(lp);
1779
1780         lnet_net_unlock(cpt);
1781         return 0;
1782 }
1783 EXPORT_SYMBOL(lnet_notify);