drivers/staging/lustre/lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2015, 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  */
  19
  20 #define DEBUG_SUBSYSTEM S_LNET
  21 #include "../../include/linux/lnet/lib-lnet.h"
  22
  23 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
  24 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
  25 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
  26 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
  27 #define LNET_NRB_SMALL_PAGES    1
  28 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  29 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  30 #define LNET_NRB_LARGE_PAGES   ((LNET_MTU + PAGE_CACHE_SIZE - 1) >> \
  31                                  PAGE_CACHE_SHIFT)
  32
  33 static char *forwarding = "";
  34 module_param(forwarding, charp, 0444);
  35 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
  36
  37 static int tiny_router_buffers;
  38 module_param(tiny_router_buffers, int, 0444);
  39 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
  40 static int small_router_buffers;
  41 module_param(small_router_buffers, int, 0444);
  42 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
  43 static int large_router_buffers;
  44 module_param(large_router_buffers, int, 0444);
  45 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
  46 static int peer_buffer_credits;
  47 module_param(peer_buffer_credits, int, 0444);
  48 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
  49
  50 static int auto_down = 1;
  51 module_param(auto_down, int, 0444);
  52 MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
  53
  54 int
  55 lnet_peer_buffer_credits(lnet_ni_t *ni)
  56 {
  57         /* NI option overrides LNet default */
  58         if (ni->ni_peerrtrcredits > 0)
  59                 return ni->ni_peerrtrcredits;
  60         if (peer_buffer_credits > 0)
  61                 return peer_buffer_credits;
  62
  63         /*
  64          * As an approximation, allow this peer the same number of router
  65          * buffers as it is allowed outstanding sends
  66          */
  67         return ni->ni_peertxcredits;
  68 }
  69
  70 /* forward ref's */
  71 static int lnet_router_checker(void *);
  72
  73 static int check_routers_before_use;
  74 module_param(check_routers_before_use, int, 0444);
  75 MODULE_PARM_DESC(check_routers_before_use, "Assume routers are down and ping them before use");
  76
  77 int avoid_asym_router_failure = 1;
  78 module_param(avoid_asym_router_failure, int, 0644);
  79 MODULE_PARM_DESC(avoid_asym_router_failure, "Avoid asymmetrical router failures (0 to disable)");
  80
  81 static int dead_router_check_interval = 60;
  82 module_param(dead_router_check_interval, int, 0644);
  83 MODULE_PARM_DESC(dead_router_check_interval, "Seconds between dead router health checks (<= 0 to disable)");
  84
  85 static int live_router_check_interval = 60;
  86 module_param(live_router_check_interval, int, 0644);
  87 MODULE_PARM_DESC(live_router_check_interval, "Seconds between live router health checks (<= 0 to disable)");
  88
  89 static int router_ping_timeout = 50;
  90 module_param(router_ping_timeout, int, 0644);
  91 MODULE_PARM_DESC(router_ping_timeout, "Seconds to wait for the reply to a router health query");
  92
  93 int
  94 lnet_peers_start_down(void)
  95 {
  96         return check_routers_before_use;
  97 }
  98
  99 void
 100 lnet_notify_locked(lnet_peer_t *lp, int notifylnd, int alive,
 101                    unsigned long when)
 102 {
 103         if (time_before(when, lp->lp_timestamp)) { /* out of date information */
 104                 CDEBUG(D_NET, "Out of date\n");
 105                 return;
 106         }
 107
 108         lp->lp_timestamp = when;                /* update timestamp */
 109         lp->lp_ping_deadline = 0;              /* disable ping timeout */
 110
 111         if (lp->lp_alive_count &&         /* got old news */
 112             (!lp->lp_alive) == (!alive)) {      /* new date for old news */
 113                 CDEBUG(D_NET, "Old news\n");
 114                 return;
 115         }
 116
 117         /* Flag that notification is outstanding */
 118
 119         lp->lp_alive_count++;
 120         lp->lp_alive = !(!alive);              /* 1 bit! */
 121         lp->lp_notify = 1;
 122         lp->lp_notifylnd |= notifylnd;
 123         if (lp->lp_alive)
 124                 lp->lp_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 125
 126         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lp_nid), alive);
 127 }
 128
 129 static void
 130 lnet_ni_notify_locked(lnet_ni_t *ni, lnet_peer_t *lp)
 131 {
 132         int alive;
 133         int notifylnd;
 134
 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 || !ni)
 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) {
 153                         lnet_net_unlock(lp->lp_cpt);
 154
 155                         /*
 156                          * A new notification could happen now; I'll handle it
 157                          * when control returns to me
 158                          */
 159                         ni->ni_lnd->lnd_notify(ni, lp->lp_nid, alive);
 160
 161                         lnet_net_lock(lp->lp_cpt);
 162                 }
 163         }
 164
 165         lp->lp_notifying = 0;
 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) {
 204                 LASSERT(list_empty(&lp->lp_routes));
 205
 206                 if (lp->lp_rcd) {
 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, seed[2];
 242         struct timespec64 ts;
 243         lnet_ni_t *ni;
 244         struct list_head *tmp;
 245
 246         if (seeded)
 247                 return;
 248
 249         cfs_get_random_bytes(seed, sizeof(seed));
 250
 251         /*
 252          * Nodes with small feet have little entropy
 253          * the NID for this node gives the most entropy in the low bits
 254          */
 255         list_for_each(tmp, &the_lnet.ln_nis) {
 256                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 257                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
 258
 259                 if (lnd_type != LOLND)
 260                         seed[0] ^= (LNET_NIDADDR(ni->ni_nid) | lnd_type);
 261         }
 262
 263         ktime_get_ts64(&ts);
 264         cfs_srand(ts.tv_sec ^ seed[0], ts.tv_nsec ^ seed[1]);
 265         seeded = 1;
 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)
 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 || !rnet) {
 326                 CERROR("Out of memory creating route %s %d %s\n",
 327                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 328                 if (route)
 329                         LIBCFS_FREE(route, sizeof(*route));
 330                 if (rnet)
 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) {
 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) {
 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)
 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         rc = 0;
 399
 400         if (!add_route) {
 401                 rc = -EEXIST;
 402                 LIBCFS_FREE(route, sizeof(*route));
 403         }
 404
 405         if (rnet != rnet2)
 406                 LIBCFS_FREE(rnet, sizeof(*rnet));
 407
 408         /* indicate to startup the router checker if configured */
 409         wake_up(&the_lnet.ln_rc_waitq);
 410
 411         return rc;
 412 }
 413
 414 int
 415 lnet_check_routes(void)
 416 {
 417         lnet_remotenet_t *rnet;
 418         lnet_route_t *route;
 419         lnet_route_t *route2;
 420         struct list_head *e1;
 421         struct list_head *e2;
 422         int cpt;
 423         struct list_head *rn_list;
 424         int i;
 425
 426         cpt = lnet_net_lock_current();
 427
 428         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 429                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 430                 list_for_each(e1, rn_list) {
 431                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 432
 433                         route2 = NULL;
 434                         list_for_each(e2, &rnet->lrn_routes) {
 435                                 lnet_nid_t nid1;
 436                                 lnet_nid_t nid2;
 437                                 int net;
 438
 439                                 route = list_entry(e2, lnet_route_t, lr_list);
 440
 441                                 if (!route2) {
 442                                         route2 = route;
 443                                         continue;
 444                                 }
 445
 446                                 if (route->lr_gateway->lp_ni ==
 447                                     route2->lr_gateway->lp_ni)
 448                                         continue;
 449
 450                                 nid1 = route->lr_gateway->lp_nid;
 451                                 nid2 = route2->lr_gateway->lp_nid;
 452                                 net = rnet->lrn_net;
 453
 454                                 lnet_net_unlock(cpt);
 455
 456                                 CERROR("Routes to %s via %s and %s not supported\n",
 457                                        libcfs_net2str(net),
 458                                        libcfs_nid2str(nid1),
 459                                        libcfs_nid2str(nid2));
 460                                 return -EINVAL;
 461                         }
 462                 }
 463         }
 464
 465         lnet_net_unlock(cpt);
 466         return 0;
 467 }
 468
 469 int
 470 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 471 {
 472         struct lnet_peer *gateway;
 473         lnet_remotenet_t *rnet;
 474         lnet_route_t *route;
 475         struct list_head *e1;
 476         struct list_head *e2;
 477         int rc = -ENOENT;
 478         struct list_head *rn_list;
 479         int idx = 0;
 480
 481         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 482                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 483
 484         /*
 485          * NB Caller may specify either all routes via the given gateway
 486          * or a specific route entry actual NIDs)
 487          */
 488         lnet_net_lock(LNET_LOCK_EX);
 489         if (net == LNET_NIDNET(LNET_NID_ANY))
 490                 rn_list = &the_lnet.ln_remote_nets_hash[0];
 491         else
 492                 rn_list = lnet_net2rnethash(net);
 493
 494  again:
 495         list_for_each(e1, rn_list) {
 496                 rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 497
 498                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 499                       net == rnet->lrn_net))
 500                         continue;
 501
 502                 list_for_each(e2, &rnet->lrn_routes) {
 503                         route = list_entry(e2, lnet_route_t, lr_list);
 504
 505                         gateway = route->lr_gateway;
 506                         if (!(gw_nid == LNET_NID_ANY ||
 507                               gw_nid == gateway->lp_nid))
 508                                 continue;
 509
 510                         list_del(&route->lr_list);
 511                         list_del(&route->lr_gwlist);
 512                         the_lnet.ln_remote_nets_version++;
 513
 514                         if (list_empty(&rnet->lrn_routes))
 515                                 list_del(&rnet->lrn_list);
 516                         else
 517                                 rnet = NULL;
 518
 519                         lnet_rtr_decref_locked(gateway);
 520                         lnet_peer_decref_locked(gateway);
 521
 522                         lnet_net_unlock(LNET_LOCK_EX);
 523
 524                         LIBCFS_FREE(route, sizeof(*route));
 525
 526                         if (rnet)
 527                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 528
 529                         rc = 0;
 530                         lnet_net_lock(LNET_LOCK_EX);
 531                         goto again;
 532                 }
 533         }
 534
 535         if (net == LNET_NIDNET(LNET_NID_ANY) &&
 536             ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
 537                 rn_list = &the_lnet.ln_remote_nets_hash[idx];
 538                 goto again;
 539         }
 540         lnet_net_unlock(LNET_LOCK_EX);
 541
 542         return rc;
 543 }
 544
 545 void
 546 lnet_destroy_routes(void)
 547 {
 548         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 549 }
 550
 551 int lnet_get_rtr_pool_cfg(int idx, struct lnet_ioctl_pool_cfg *pool_cfg)
 552 {
 553         int i, rc = -ENOENT, j;
 554
 555         if (!the_lnet.ln_rtrpools)
 556                 return rc;
 557
 558         for (i = 0; i < LNET_NRBPOOLS; i++) {
 559                 lnet_rtrbufpool_t *rbp;
 560
 561                 lnet_net_lock(LNET_LOCK_EX);
 562                 cfs_percpt_for_each(rbp, j, the_lnet.ln_rtrpools) {
 563                         if (i++ != idx)
 564                                 continue;
 565
 566                         pool_cfg->pl_pools[i].pl_npages = rbp[i].rbp_npages;
 567                         pool_cfg->pl_pools[i].pl_nbuffers = rbp[i].rbp_nbuffers;
 568                         pool_cfg->pl_pools[i].pl_credits = rbp[i].rbp_credits;
 569                         pool_cfg->pl_pools[i].pl_mincredits = rbp[i].rbp_mincredits;
 570                         rc = 0;
 571                         break;
 572                 }
 573                 lnet_net_unlock(LNET_LOCK_EX);
 574         }
 575
 576         lnet_net_lock(LNET_LOCK_EX);
 577         pool_cfg->pl_routing = the_lnet.ln_routing;
 578         lnet_net_unlock(LNET_LOCK_EX);
 579
 580         return rc;
 581 }
 582
 583 int
 584 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 585                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 586 {
 587         struct list_head *e1;
 588         struct list_head *e2;
 589         lnet_remotenet_t *rnet;
 590         lnet_route_t *route;
 591         int cpt;
 592         int i;
 593         struct list_head *rn_list;
 594
 595         cpt = lnet_net_lock_current();
 596
 597         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 598                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 599                 list_for_each(e1, rn_list) {
 600                         rnet = list_entry(e1, lnet_remotenet_t, lrn_list);
 601
 602                         list_for_each(e2, &rnet->lrn_routes) {
 603                                 route = list_entry(e2, lnet_route_t, lr_list);
 604
 605                                 if (!idx--) {
 606                                         *net      = rnet->lrn_net;
 607                                         *hops     = route->lr_hops;
 608                                         *priority = route->lr_priority;
 609                                         *gateway  = route->lr_gateway->lp_nid;
 610                                         *alive = route->lr_gateway->lp_alive &&
 611                                                  !route->lr_downis;
 612                                         lnet_net_unlock(cpt);
 613                                         return 0;
 614                                 }
 615                         }
 616                 }
 617         }
 618
 619         lnet_net_unlock(cpt);
 620         return -ENOENT;
 621 }
 622
 623 void
 624 lnet_swap_pinginfo(lnet_ping_info_t *info)
 625 {
 626         int i;
 627         lnet_ni_status_t *stat;
 628
 629         __swab32s(&info->pi_magic);
 630         __swab32s(&info->pi_features);
 631         __swab32s(&info->pi_pid);
 632         __swab32s(&info->pi_nnis);
 633         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 634                 stat = &info->pi_ni[i];
 635                 __swab64s(&stat->ns_nid);
 636                 __swab32s(&stat->ns_status);
 637         }
 638 }
 639
 640 /**
 641  * parse router-checker pinginfo, record number of down NIs for remote
 642  * networks on that router.
 643  */
 644 static void
 645 lnet_parse_rc_info(lnet_rc_data_t *rcd)
 646 {
 647         lnet_ping_info_t *info = rcd->rcd_pinginfo;
 648         struct lnet_peer *gw = rcd->rcd_gateway;
 649         lnet_route_t *rte;
 650
 651         if (!gw->lp_alive)
 652                 return;
 653
 654         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 655                 lnet_swap_pinginfo(info);
 656
 657         /* NB always racing with network! */
 658         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 659                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 660                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
 661                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 662                 return;
 663         }
 664
 665         gw->lp_ping_feats = info->pi_features;
 666         if (!(gw->lp_ping_feats & LNET_PING_FEAT_MASK)) {
 667                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
 668                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
 669                 return; /* nothing I can understand */
 670         }
 671
 672         if (!(gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS))
 673                 return; /* can't carry NI status info */
 674
 675         list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
 676                 int down = 0;
 677                 int up = 0;
 678                 int i;
 679
 680                 if (gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) {
 681                         rte->lr_downis = 1;
 682                         continue;
 683                 }
 684
 685                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 686                         lnet_ni_status_t *stat = &info->pi_ni[i];
 687                         lnet_nid_t nid = stat->ns_nid;
 688
 689                         if (nid == LNET_NID_ANY) {
 690                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 691                                        libcfs_nid2str(gw->lp_nid));
 692                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 693                                 return;
 694                         }
 695
 696                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 697                                 continue;
 698
 699                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 700                                 down++;
 701                                 continue;
 702                         }
 703
 704                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 705                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 706                                         up = 1;
 707                                         break;
 708                                 }
 709                                 continue;
 710                         }
 711
 712                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 713                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
 714                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 715                         return;
 716                 }
 717
 718                 if (up) { /* ignore downed NIs if NI for dest network is up */
 719                         rte->lr_downis = 0;
 720                         continue;
 721                 }
 722                 rte->lr_downis = down;
 723         }
 724 }
 725
 726 static void
 727 lnet_router_checker_event(lnet_event_t *event)
 728 {
 729         lnet_rc_data_t *rcd = event->md.user_ptr;
 730         struct lnet_peer *lp;
 731
 732         LASSERT(rcd);
 733
 734         if (event->unlinked) {
 735                 LNetInvalidateHandle(&rcd->rcd_mdh);
 736                 return;
 737         }
 738
 739         LASSERT(event->type == LNET_EVENT_SEND ||
 740                 event->type == LNET_EVENT_REPLY);
 741
 742         lp = rcd->rcd_gateway;
 743         LASSERT(lp);
 744
 745         /*
 746          * NB: it's called with holding lnet_res_lock, we have a few
 747          * places need to hold both locks at the same time, please take
 748          * care of lock ordering
 749          */
 750         lnet_net_lock(lp->lp_cpt);
 751         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
 752                 /* ignore if no longer a router or rcd is replaced */
 753                 goto out;
 754         }
 755
 756         if (event->type == LNET_EVENT_SEND) {
 757                 lp->lp_ping_notsent = 0;
 758                 if (!event->status)
 759                         goto out;
 760         }
 761
 762         /* LNET_EVENT_REPLY */
 763         /*
 764          * A successful REPLY means the router is up.  If _any_ comms
 765          * to the router fail I assume it's down (this will happen if
 766          * we ping alive routers to try to detect router death before
 767          * apps get burned).
 768          */
 769         lnet_notify_locked(lp, 1, !event->status, cfs_time_current());
 770
 771         /*
 772          * The router checker will wake up very shortly and do the
 773          * actual notification.
 774          * XXX If 'lp' stops being a router before then, it will still
 775          * have the notification pending!!!
 776          */
 777         if (avoid_asym_router_failure && !event->status)
 778                 lnet_parse_rc_info(rcd);
 779
 780  out:
 781         lnet_net_unlock(lp->lp_cpt);
 782 }
 783
 784 static void
 785 lnet_wait_known_routerstate(void)
 786 {
 787         lnet_peer_t *rtr;
 788         struct list_head *entry;
 789         int all_known;
 790
 791         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 792
 793         for (;;) {
 794                 int cpt = lnet_net_lock_current();
 795
 796                 all_known = 1;
 797                 list_for_each(entry, &the_lnet.ln_routers) {
 798                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
 799
 800                         if (!rtr->lp_alive_count) {
 801                                 all_known = 0;
 802                                 break;
 803                         }
 804                 }
 805
 806                 lnet_net_unlock(cpt);
 807
 808                 if (all_known)
 809                         return;
 810
 811                 set_current_state(TASK_UNINTERRUPTIBLE);
 812                 schedule_timeout(cfs_time_seconds(1));
 813         }
 814 }
 815
 816 void
 817 lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net)
 818 {
 819         lnet_route_t *rte;
 820
 821         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS)) {
 822                 list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
 823                         if (rte->lr_net == net) {
 824                                 rte->lr_downis = 0;
 825                                 break;
 826                         }
 827                 }
 828         }
 829 }
 830
 831 static void
 832 lnet_update_ni_status_locked(void)
 833 {
 834         lnet_ni_t *ni;
 835         time64_t now;
 836         int timeout;
 837
 838         LASSERT(the_lnet.ln_routing);
 839
 840         timeout = router_ping_timeout +
 841                   max(live_router_check_interval, dead_router_check_interval);
 842
 843         now = ktime_get_real_seconds();
 844         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 845                 if (ni->ni_lnd->lnd_type == LOLND)
 846                         continue;
 847
 848                 if (now < ni->ni_last_alive + timeout)
 849                         continue;
 850
 851                 lnet_ni_lock(ni);
 852                 /* re-check with lock */
 853                 if (now < ni->ni_last_alive + timeout) {
 854                         lnet_ni_unlock(ni);
 855                         continue;
 856                 }
 857
 858                 LASSERT(ni->ni_status);
 859
 860                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 861                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 862                                libcfs_nid2str(ni->ni_nid), timeout);
 863                         /*
 864                          * NB: so far, this is the only place to set
 865                          * NI status to "down"
 866                          */
 867                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 868                 }
 869                 lnet_ni_unlock(ni);
 870         }
 871 }
 872
 873 static void
 874 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
 875 {
 876         LASSERT(list_empty(&rcd->rcd_list));
 877         /* detached from network */
 878         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 879
 880         if (rcd->rcd_gateway) {
 881                 int cpt = rcd->rcd_gateway->lp_cpt;
 882
 883                 lnet_net_lock(cpt);
 884                 lnet_peer_decref_locked(rcd->rcd_gateway);
 885                 lnet_net_unlock(cpt);
 886         }
 887
 888         if (rcd->rcd_pinginfo)
 889                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 890
 891         LIBCFS_FREE(rcd, sizeof(*rcd));
 892 }
 893
 894 static lnet_rc_data_t *
 895 lnet_create_rc_data_locked(lnet_peer_t *gateway)
 896 {
 897         lnet_rc_data_t *rcd = NULL;
 898         lnet_ping_info_t *pi;
 899         int rc;
 900         int i;
 901
 902         lnet_net_unlock(gateway->lp_cpt);
 903
 904         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 905         if (!rcd)
 906                 goto out;
 907
 908         LNetInvalidateHandle(&rcd->rcd_mdh);
 909         INIT_LIST_HEAD(&rcd->rcd_list);
 910
 911         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 912         if (!pi)
 913                 goto out;
 914
 915         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 916                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 917                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 918         }
 919         rcd->rcd_pinginfo = pi;
 920
 921         LASSERT(!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 922         rc = LNetMDBind((lnet_md_t){.start     = pi,
 923                                     .user_ptr  = rcd,
 924                                     .length    = LNET_PINGINFO_SIZE,
 925                                     .threshold = LNET_MD_THRESH_INF,
 926                                     .options   = LNET_MD_TRUNCATE,
 927                                     .eq_handle = the_lnet.ln_rc_eqh},
 928                         LNET_UNLINK,
 929                         &rcd->rcd_mdh);
 930         if (rc < 0) {
 931                 CERROR("Can't bind MD: %d\n", rc);
 932                 goto out;
 933         }
 934         LASSERT(!rc);
 935
 936         lnet_net_lock(gateway->lp_cpt);
 937         /* router table changed or someone has created rcd for this gateway */
 938         if (!lnet_isrouter(gateway) || gateway->lp_rcd) {
 939                 lnet_net_unlock(gateway->lp_cpt);
 940                 goto out;
 941         }
 942
 943         lnet_peer_addref_locked(gateway);
 944         rcd->rcd_gateway = gateway;
 945         gateway->lp_rcd = rcd;
 946         gateway->lp_ping_notsent = 0;
 947
 948         return rcd;
 949
 950  out:
 951         if (rcd) {
 952                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
 953                         rc = LNetMDUnlink(rcd->rcd_mdh);
 954                         LASSERT(!rc);
 955                 }
 956                 lnet_destroy_rc_data(rcd);
 957         }
 958
 959         lnet_net_lock(gateway->lp_cpt);
 960         return gateway->lp_rcd;
 961 }
 962
 963 static int
 964 lnet_router_check_interval(lnet_peer_t *rtr)
 965 {
 966         int secs;
 967
 968         secs = rtr->lp_alive ? live_router_check_interval :
 969                                dead_router_check_interval;
 970         if (secs < 0)
 971                 secs = 0;
 972
 973         return secs;
 974 }
 975
 976 static void
 977 lnet_ping_router_locked(lnet_peer_t *rtr)
 978 {
 979         lnet_rc_data_t *rcd = NULL;
 980         unsigned long now = cfs_time_current();
 981         int secs;
 982
 983         lnet_peer_addref_locked(rtr);
 984
 985         if (rtr->lp_ping_deadline && /* ping timed out? */
 986             cfs_time_after(now, rtr->lp_ping_deadline))
 987                 lnet_notify_locked(rtr, 1, 0, now);
 988
 989         /* Run any outstanding notifications */
 990         lnet_ni_notify_locked(rtr->lp_ni, rtr);
 991
 992         if (!lnet_isrouter(rtr) ||
 993             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
 994                 /* router table changed or router checker is shutting down */
 995                 lnet_peer_decref_locked(rtr);
 996                 return;
 997         }
 998
 999         rcd = rtr->lp_rcd ?
1000               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
1001
1002         if (!rcd)
1003                 return;
1004
1005         secs = lnet_router_check_interval(rtr);
1006
1007         CDEBUG(D_NET,
1008                "rtr %s %d: deadline %lu ping_notsent %d alive %d alive_count %d lp_ping_timestamp %lu\n",
1009                libcfs_nid2str(rtr->lp_nid), secs,
1010                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
1011                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
1012
1013         if (secs && !rtr->lp_ping_notsent &&
1014             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
1015                                              cfs_time_seconds(secs)))) {
1016                 int rc;
1017                 lnet_process_id_t id;
1018                 lnet_handle_md_t mdh;
1019
1020                 id.nid = rtr->lp_nid;
1021                 id.pid = LNET_PID_LUSTRE;
1022                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1023
1024                 rtr->lp_ping_notsent   = 1;
1025                 rtr->lp_ping_timestamp = now;
1026
1027                 mdh = rcd->rcd_mdh;
1028
1029                 if (!rtr->lp_ping_deadline) {
1030                         rtr->lp_ping_deadline =
1031                                 cfs_time_shift(router_ping_timeout);
1032                 }
1033
1034                 lnet_net_unlock(rtr->lp_cpt);
1035
1036                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1037                              LNET_PROTO_PING_MATCHBITS, 0);
1038
1039                 lnet_net_lock(rtr->lp_cpt);
1040                 if (rc)
1041                         rtr->lp_ping_notsent = 0; /* no event pending */
1042         }
1043
1044         lnet_peer_decref_locked(rtr);
1045 }
1046
1047 int
1048 lnet_router_checker_start(void)
1049 {
1050         struct task_struct *task;
1051         int rc;
1052         int eqsz = 0;
1053
1054         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1055
1056         if (check_routers_before_use &&
1057             dead_router_check_interval <= 0) {
1058                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be set if 'check_routers_before_use' is set\n");
1059                 return -EINVAL;
1060         }
1061
1062         sema_init(&the_lnet.ln_rc_signal, 0);
1063
1064         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1065         if (rc) {
1066                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1067                 return -ENOMEM;
1068         }
1069
1070         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1071         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1072         if (IS_ERR(task)) {
1073                 rc = PTR_ERR(task);
1074                 CERROR("Can't start router checker thread: %d\n", rc);
1075                 /* block until event callback signals exit */
1076                 down(&the_lnet.ln_rc_signal);
1077                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1078                 LASSERT(!rc);
1079                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1080                 return -ENOMEM;
1081         }
1082
1083         if (check_routers_before_use) {
1084                 /*
1085                  * Note that a helpful side-effect of pinging all known routers
1086                  * at startup is that it makes them drop stale connections they
1087                  * may have to a previous instance of me.
1088                  */
1089                 lnet_wait_known_routerstate();
1090         }
1091
1092         return 0;
1093 }
1094
1095 void
1096 lnet_router_checker_stop(void)
1097 {
1098         int rc;
1099
1100         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1101                 return;
1102
1103         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1104         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1105         /* wakeup the RC thread if it's sleeping */
1106         wake_up(&the_lnet.ln_rc_waitq);
1107
1108         /* block until event callback signals exit */
1109         down(&the_lnet.ln_rc_signal);
1110         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1111
1112         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1113         LASSERT(!rc);
1114 }
1115
1116 static void
1117 lnet_prune_rc_data(int wait_unlink)
1118 {
1119         lnet_rc_data_t *rcd;
1120         lnet_rc_data_t *tmp;
1121         lnet_peer_t *lp;
1122         struct list_head head;
1123         int i = 2;
1124
1125         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1126                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1127                    list_empty(&the_lnet.ln_rcd_zombie)))
1128                 return;
1129
1130         INIT_LIST_HEAD(&head);
1131
1132         lnet_net_lock(LNET_LOCK_EX);
1133
1134         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1135                 /* router checker is stopping, prune all */
1136                 list_for_each_entry(lp, &the_lnet.ln_routers,
1137                                     lp_rtr_list) {
1138                         if (!lp->lp_rcd)
1139                                 continue;
1140
1141                         LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1142                         list_add(&lp->lp_rcd->rcd_list,
1143                                  &the_lnet.ln_rcd_deathrow);
1144                         lp->lp_rcd = NULL;
1145                 }
1146         }
1147
1148         /* unlink all RCDs on deathrow list */
1149         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1150
1151         if (!list_empty(&head)) {
1152                 lnet_net_unlock(LNET_LOCK_EX);
1153
1154                 list_for_each_entry(rcd, &head, rcd_list)
1155                         LNetMDUnlink(rcd->rcd_mdh);
1156
1157                 lnet_net_lock(LNET_LOCK_EX);
1158         }
1159
1160         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1161
1162         /* release all zombie RCDs */
1163         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1164                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1165                                          rcd_list) {
1166                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1167                                 list_move(&rcd->rcd_list, &head);
1168                 }
1169
1170                 wait_unlink = wait_unlink &&
1171                               !list_empty(&the_lnet.ln_rcd_zombie);
1172
1173                 lnet_net_unlock(LNET_LOCK_EX);
1174
1175                 while (!list_empty(&head)) {
1176                         rcd = list_entry(head.next,
1177                                          lnet_rc_data_t, rcd_list);
1178                         list_del_init(&rcd->rcd_list);
1179                         lnet_destroy_rc_data(rcd);
1180                 }
1181
1182                 if (!wait_unlink)
1183                         return;
1184
1185                 i++;
1186                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1187                        "Waiting for rc buffers to unlink\n");
1188                 set_current_state(TASK_UNINTERRUPTIBLE);
1189                 schedule_timeout(cfs_time_seconds(1) / 4);
1190
1191                 lnet_net_lock(LNET_LOCK_EX);
1192         }
1193
1194         lnet_net_unlock(LNET_LOCK_EX);
1195 }
1196
1197 /*
1198  * This function is called to check if the RC should block indefinitely.
1199  * It's called from lnet_router_checker() as well as being passed to
1200  * wait_event_interruptible() to avoid the lost wake_up problem.
1201  *
1202  * When it's called from wait_event_interruptible() it is necessary to
1203  * also not sleep if the rc state is not running to avoid a deadlock
1204  * when the system is shutting down
1205  */
1206 static inline bool
1207 lnet_router_checker_active(void)
1208 {
1209         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1210                 return true;
1211
1212         /*
1213          * Router Checker thread needs to run when routing is enabled in
1214          * order to call lnet_update_ni_status_locked()
1215          */
1216         if (the_lnet.ln_routing)
1217                 return true;
1218
1219         return !list_empty(&the_lnet.ln_routers) &&
1220                 (live_router_check_interval > 0 ||
1221                  dead_router_check_interval > 0);
1222 }
1223
1224 static int
1225 lnet_router_checker(void *arg)
1226 {
1227         lnet_peer_t *rtr;
1228         struct list_head *entry;
1229
1230         cfs_block_allsigs();
1231
1232         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1233
1234         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1235                 __u64 version;
1236                 int cpt;
1237                 int cpt2;
1238
1239                 cpt = lnet_net_lock_current();
1240 rescan:
1241                 version = the_lnet.ln_routers_version;
1242
1243                 list_for_each(entry, &the_lnet.ln_routers) {
1244                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1245
1246                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1247                         if (cpt != cpt2) {
1248                                 lnet_net_unlock(cpt);
1249                                 cpt = cpt2;
1250                                 lnet_net_lock(cpt);
1251                                 /* the routers list has changed */
1252                                 if (version != the_lnet.ln_routers_version)
1253                                         goto rescan;
1254                         }
1255
1256                         lnet_ping_router_locked(rtr);
1257
1258                         /* NB dropped lock */
1259                         if (version != the_lnet.ln_routers_version) {
1260                                 /* the routers list has changed */
1261                                 goto rescan;
1262                         }
1263                 }
1264
1265                 if (the_lnet.ln_routing)
1266                         lnet_update_ni_status_locked();
1267
1268                 lnet_net_unlock(cpt);
1269
1270                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1271
1272                 /*
1273                  * Call schedule_timeout() here always adds 1 to load average
1274                  * because kernel counts # active tasks as nr_running
1275                  * + nr_uninterruptible.
1276                  */
1277                 /*
1278                  * if there are any routes then wakeup every second.  If
1279                  * there are no routes then sleep indefinitely until woken
1280                  * up by a user adding a route
1281                  */
1282                 if (!lnet_router_checker_active())
1283                         wait_event_interruptible(the_lnet.ln_rc_waitq,
1284                                                  lnet_router_checker_active());
1285                 else
1286                         wait_event_interruptible_timeout(the_lnet.ln_rc_waitq,
1287                                                          false,
1288                                                          cfs_time_seconds(1));
1289         }
1290
1291         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_STOPPING);
1292
1293         lnet_prune_rc_data(1); /* wait for UNLINK */
1294
1295         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1296         up(&the_lnet.ln_rc_signal);
1297         /* The unlink event callback will signal final completion */
1298         return 0;
1299 }
1300
1301 void
1302 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1303 {
1304         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1305
1306         while (--npages >= 0)
1307                 __free_page(rb->rb_kiov[npages].kiov_page);
1308
1309         LIBCFS_FREE(rb, sz);
1310 }
1311
1312 static lnet_rtrbuf_t *
1313 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1314 {
1315         int npages = rbp->rbp_npages;
1316         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1317         struct page *page;
1318         lnet_rtrbuf_t *rb;
1319         int i;
1320
1321         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1322         if (!rb)
1323                 return NULL;
1324
1325         rb->rb_pool = rbp;
1326
1327         for (i = 0; i < npages; i++) {
1328                 page = alloc_pages_node(
1329                                 cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1330                                 GFP_KERNEL | __GFP_ZERO, 0);
1331                 if (!page) {
1332                         while (--i >= 0)
1333                                 __free_page(rb->rb_kiov[i].kiov_page);
1334
1335                         LIBCFS_FREE(rb, sz);
1336                         return NULL;
1337                 }
1338
1339                 rb->rb_kiov[i].kiov_len = PAGE_CACHE_SIZE;
1340                 rb->rb_kiov[i].kiov_offset = 0;
1341                 rb->rb_kiov[i].kiov_page = page;
1342         }
1343
1344         return rb;
1345 }
1346
1347 static void
1348 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1349 {
1350         int npages = rbp->rbp_npages;
1351         struct list_head tmp;
1352         lnet_rtrbuf_t *rb;
1353
1354         if (!rbp->rbp_nbuffers) /* not initialized or already freed */
1355                 return;
1356
1357         INIT_LIST_HEAD(&tmp);
1358
1359         lnet_net_lock(cpt);
1360         lnet_drop_routed_msgs_locked(&rbp->rbp_msgs, cpt);
1361         list_splice_init(&rbp->rbp_bufs, &tmp);
1362         rbp->rbp_nbuffers = 0;
1363         rbp->rbp_credits = 0;
1364         rbp->rbp_mincredits = 0;
1365         lnet_net_unlock(cpt);
1366
1367         /* Free buffers on the free list. */
1368         while (!list_empty(&tmp)) {
1369                 rb = list_entry(tmp.next, lnet_rtrbuf_t, rb_list);
1370                 list_del(&rb->rb_list);
1371                 lnet_destroy_rtrbuf(rb, npages);
1372         }
1373 }
1374
1375 static int
1376 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1377 {
1378         struct list_head rb_list;
1379         lnet_rtrbuf_t *rb;
1380         int num_rb;
1381         int num_buffers = 0;
1382         int npages = rbp->rbp_npages;
1383
1384         /*
1385          * If we are called for less buffers than already in the pool, we
1386          * just lower the nbuffers number and excess buffers will be
1387          * thrown away as they are returned to the free list.  Credits
1388          * then get adjusted as well.
1389          */
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         /*
1400          * allocate the buffers on a local list first.  If all buffers are
1401          * allocated successfully then join this list to the rbp buffer
1402          * list. If not then free all allocated buffers.
1403          */
1404         num_rb = rbp->rbp_nbuffers;
1405
1406         while (num_rb < nbufs) {
1407                 rb = lnet_new_rtrbuf(rbp, cpt);
1408                 if (!rb) {
1409                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1410                                nbufs, npages);
1411                         goto failed;
1412                 }
1413
1414                 list_add(&rb->rb_list, &rb_list);
1415                 num_buffers++;
1416                 num_rb++;
1417         }
1418
1419         lnet_net_lock(cpt);
1420
1421         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1422         rbp->rbp_nbuffers += num_buffers;
1423         rbp->rbp_credits += num_buffers;
1424         rbp->rbp_mincredits = rbp->rbp_credits;
1425         /*
1426          * We need to schedule blocked msg using the newly
1427          * added buffers.
1428          */
1429         while (!list_empty(&rbp->rbp_bufs) &&
1430                !list_empty(&rbp->rbp_msgs))
1431                 lnet_schedule_blocked_locked(rbp);
1432
1433         lnet_net_unlock(cpt);
1434
1435         return 0;
1436
1437 failed:
1438         while (!list_empty(&rb_list)) {
1439                 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1440                 list_del(&rb->rb_list);
1441                 lnet_destroy_rtrbuf(rb, npages);
1442         }
1443
1444         return -ENOMEM;
1445 }
1446
1447 static void
1448 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1449 {
1450         INIT_LIST_HEAD(&rbp->rbp_msgs);
1451         INIT_LIST_HEAD(&rbp->rbp_bufs);
1452
1453         rbp->rbp_npages = npages;
1454         rbp->rbp_credits = 0;
1455         rbp->rbp_mincredits = 0;
1456 }
1457
1458 void
1459 lnet_rtrpools_free(int keep_pools)
1460 {
1461         lnet_rtrbufpool_t *rtrp;
1462         int i;
1463
1464         if (!the_lnet.ln_rtrpools) /* uninitialized or freed */
1465                 return;
1466
1467         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1468                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1469                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1470                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1471         }
1472
1473         if (!keep_pools) {
1474                 cfs_percpt_free(the_lnet.ln_rtrpools);
1475                 the_lnet.ln_rtrpools = NULL;
1476         }
1477 }
1478
1479 static int
1480 lnet_nrb_tiny_calculate(void)
1481 {
1482         int nrbs = LNET_NRB_TINY;
1483
1484         if (tiny_router_buffers < 0) {
1485                 LCONSOLE_ERROR_MSG(0x10c,
1486                                    "tiny_router_buffers=%d invalid when routing enabled\n",
1487                                    tiny_router_buffers);
1488                 return -1;
1489         }
1490
1491         if (tiny_router_buffers > 0)
1492                 nrbs = tiny_router_buffers;
1493
1494         nrbs /= LNET_CPT_NUMBER;
1495         return max(nrbs, LNET_NRB_TINY_MIN);
1496 }
1497
1498 static int
1499 lnet_nrb_small_calculate(void)
1500 {
1501         int nrbs = LNET_NRB_SMALL;
1502
1503         if (small_router_buffers < 0) {
1504                 LCONSOLE_ERROR_MSG(0x10c,
1505                                    "small_router_buffers=%d invalid when routing enabled\n",
1506                                    small_router_buffers);
1507                 return -1;
1508         }
1509
1510         if (small_router_buffers > 0)
1511                 nrbs = small_router_buffers;
1512
1513         nrbs /= LNET_CPT_NUMBER;
1514         return max(nrbs, LNET_NRB_SMALL_MIN);
1515 }
1516
1517 static int
1518 lnet_nrb_large_calculate(void)
1519 {
1520         int nrbs = LNET_NRB_LARGE;
1521
1522         if (large_router_buffers < 0) {
1523                 LCONSOLE_ERROR_MSG(0x10c,
1524                                    "large_router_buffers=%d invalid when routing enabled\n",
1525                                    large_router_buffers);
1526                 return -1;
1527         }
1528
1529         if (large_router_buffers > 0)
1530                 nrbs = large_router_buffers;
1531
1532         nrbs /= LNET_CPT_NUMBER;
1533         return max(nrbs, LNET_NRB_LARGE_MIN);
1534 }
1535
1536 int
1537 lnet_rtrpools_alloc(int im_a_router)
1538 {
1539         lnet_rtrbufpool_t *rtrp;
1540         int nrb_tiny;
1541         int nrb_small;
1542         int nrb_large;
1543         int rc;
1544         int i;
1545
1546         if (!strcmp(forwarding, "")) {
1547                 /* not set either way */
1548                 if (!im_a_router)
1549                         return 0;
1550         } else if (!strcmp(forwarding, "disabled")) {
1551                 /* explicitly disabled */
1552                 return 0;
1553         } else if (!strcmp(forwarding, "enabled")) {
1554                 /* explicitly enabled */
1555         } else {
1556                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either 'enabled' or 'disabled'\n");
1557                 return -EINVAL;
1558         }
1559
1560         nrb_tiny = lnet_nrb_tiny_calculate();
1561         if (nrb_tiny < 0)
1562                 return -EINVAL;
1563
1564         nrb_small = lnet_nrb_small_calculate();
1565         if (nrb_small < 0)
1566                 return -EINVAL;
1567
1568         nrb_large = lnet_nrb_large_calculate();
1569         if (nrb_large < 0)
1570                 return -EINVAL;
1571
1572         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1573                                                 LNET_NRBPOOLS *
1574                                                 sizeof(lnet_rtrbufpool_t));
1575         if (!the_lnet.ln_rtrpools) {
1576                 LCONSOLE_ERROR_MSG(0x10c,
1577                                    "Failed to initialize router buffe pool\n");
1578                 return -ENOMEM;
1579         }
1580
1581         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1582                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1583                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1584                                               nrb_tiny, i);
1585                 if (rc)
1586                         goto failed;
1587
1588                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1589                                   LNET_NRB_SMALL_PAGES);
1590                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1591                                               nrb_small, i);
1592                 if (rc)
1593                         goto failed;
1594
1595                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1596                                   LNET_NRB_LARGE_PAGES);
1597                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1598                                               nrb_large, i);
1599                 if (rc)
1600                         goto failed;
1601         }
1602
1603         lnet_net_lock(LNET_LOCK_EX);
1604         the_lnet.ln_routing = 1;
1605         lnet_net_unlock(LNET_LOCK_EX);
1606
1607         return 0;
1608
1609  failed:
1610         lnet_rtrpools_free(0);
1611         return rc;
1612 }
1613
1614 static int
1615 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1616 {
1617         int nrb = 0;
1618         int rc = 0;
1619         int i;
1620         lnet_rtrbufpool_t *rtrp;
1621
1622         /*
1623          * If the provided values for each buffer pool are different than the
1624          * configured values, we need to take action.
1625          */
1626         if (tiny >= 0) {
1627                 tiny_router_buffers = tiny;
1628                 nrb = lnet_nrb_tiny_calculate();
1629                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1630                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1631                                                       nrb, i);
1632                         if (rc)
1633                                 return rc;
1634                 }
1635         }
1636         if (small >= 0) {
1637                 small_router_buffers = small;
1638                 nrb = lnet_nrb_small_calculate();
1639                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1640                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1641                                                       nrb, i);
1642                         if (rc)
1643                                 return rc;
1644                 }
1645         }
1646         if (large >= 0) {
1647                 large_router_buffers = large;
1648                 nrb = lnet_nrb_large_calculate();
1649                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1650                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1651                                                       nrb, i);
1652                         if (rc)
1653                                 return rc;
1654                 }
1655         }
1656
1657         return 0;
1658 }
1659
1660 int
1661 lnet_rtrpools_adjust(int tiny, int small, int large)
1662 {
1663         /*
1664          * this function doesn't revert the changes if adding new buffers
1665          * failed.  It's up to the user space caller to revert the
1666          * changes.
1667          */
1668         if (!the_lnet.ln_routing)
1669                 return 0;
1670
1671         return lnet_rtrpools_adjust_helper(tiny, small, large);
1672 }
1673
1674 int
1675 lnet_rtrpools_enable(void)
1676 {
1677         int rc;
1678
1679         if (the_lnet.ln_routing)
1680                 return 0;
1681
1682         if (!the_lnet.ln_rtrpools)
1683                 /*
1684                  * If routing is turned off, and we have never
1685                  * initialized the pools before, just call the
1686                  * standard buffer pool allocation routine as
1687                  * if we are just configuring this for the first
1688                  * time.
1689                  */
1690                 return lnet_rtrpools_alloc(1);
1691
1692         rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1693         if (rc)
1694                 return rc;
1695
1696         lnet_net_lock(LNET_LOCK_EX);
1697         the_lnet.ln_routing = 1;
1698
1699         the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1700         lnet_net_unlock(LNET_LOCK_EX);
1701
1702         return 0;
1703 }
1704
1705 void
1706 lnet_rtrpools_disable(void)
1707 {
1708         if (!the_lnet.ln_routing)
1709                 return;
1710
1711         lnet_net_lock(LNET_LOCK_EX);
1712         the_lnet.ln_routing = 0;
1713         the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1714
1715         tiny_router_buffers = 0;
1716         small_router_buffers = 0;
1717         large_router_buffers = 0;
1718         lnet_net_unlock(LNET_LOCK_EX);
1719         lnet_rtrpools_free(1);
1720 }
1721
1722 int
1723 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1724 {
1725         struct lnet_peer *lp = NULL;
1726         unsigned long now = cfs_time_current();
1727         int cpt = lnet_cpt_of_nid(nid);
1728
1729         LASSERT(!in_interrupt());
1730
1731         CDEBUG(D_NET, "%s notifying %s: %s\n",
1732                !ni ? "userspace" : libcfs_nid2str(ni->ni_nid),
1733                libcfs_nid2str(nid),
1734                alive ? "up" : "down");
1735
1736         if (ni &&
1737             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1738                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1739                       libcfs_nid2str(nid), alive ? "birth" : "death",
1740                       libcfs_nid2str(ni->ni_nid));
1741                 return -EINVAL;
1742         }
1743
1744         /* can't do predictions... */
1745         if (cfs_time_after(when, now)) {
1746                 CWARN("Ignoring prediction from %s of %s %s %ld seconds in the future\n",
1747                       !ni ? "userspace" : libcfs_nid2str(ni->ni_nid),
1748                       libcfs_nid2str(nid), alive ? "up" : "down",
1749                       cfs_duration_sec(cfs_time_sub(when, now)));
1750                 return -EINVAL;
1751         }
1752
1753         if (ni && !alive &&          /* LND telling me she's down */
1754             !auto_down) {                      /* auto-down disabled */
1755                 CDEBUG(D_NET, "Auto-down disabled\n");
1756                 return 0;
1757         }
1758
1759         lnet_net_lock(cpt);
1760
1761         if (the_lnet.ln_shutdown) {
1762                 lnet_net_unlock(cpt);
1763                 return -ESHUTDOWN;
1764         }
1765
1766         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1767         if (!lp) {
1768                 /* nid not found */
1769                 lnet_net_unlock(cpt);
1770                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1771                 return 0;
1772         }
1773
1774         /*
1775          * We can't fully trust LND on reporting exact peer last_alive
1776          * if he notifies us about dead peer. For example ksocklnd can
1777          * call us with when == _time_when_the_node_was_booted_ if
1778          * no connections were successfully established
1779          */
1780         if (ni && !alive && when < lp->lp_last_alive)
1781                 when = lp->lp_last_alive;
1782
1783         lnet_notify_locked(lp, !ni, alive, when);
1784
1785         lnet_ni_notify_locked(ni, lp);
1786
1787         lnet_peer_decref_locked(lp);
1788
1789         lnet_net_unlock(cpt);
1790         return 0;
1791 }
1792 EXPORT_SYMBOL(lnet_notify);