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_SIZE - 1) >> \
  31                                  PAGE_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, __u32 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 %d 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 != LNET_UNDEFINED_HOPS && (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 = lnet_is_route_alive(route);
 611                                         lnet_net_unlock(cpt);
 612                                         return 0;
 613                                 }
 614                         }
 615                 }
 616         }
 617
 618         lnet_net_unlock(cpt);
 619         return -ENOENT;
 620 }
 621
 622 void
 623 lnet_swap_pinginfo(lnet_ping_info_t *info)
 624 {
 625         int i;
 626         lnet_ni_status_t *stat;
 627
 628         __swab32s(&info->pi_magic);
 629         __swab32s(&info->pi_features);
 630         __swab32s(&info->pi_pid);
 631         __swab32s(&info->pi_nnis);
 632         for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 633                 stat = &info->pi_ni[i];
 634                 __swab64s(&stat->ns_nid);
 635                 __swab32s(&stat->ns_status);
 636         }
 637 }
 638
 639 /**
 640  * parse router-checker pinginfo, record number of down NIs for remote
 641  * networks on that router.
 642  */
 643 static void
 644 lnet_parse_rc_info(lnet_rc_data_t *rcd)
 645 {
 646         lnet_ping_info_t *info = rcd->rcd_pinginfo;
 647         struct lnet_peer *gw = rcd->rcd_gateway;
 648         lnet_route_t *rte;
 649
 650         if (!gw->lp_alive)
 651                 return;
 652
 653         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 654                 lnet_swap_pinginfo(info);
 655
 656         /* NB always racing with network! */
 657         if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
 658                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 659                        libcfs_nid2str(gw->lp_nid), info->pi_magic);
 660                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 661                 return;
 662         }
 663
 664         gw->lp_ping_feats = info->pi_features;
 665         if (!(gw->lp_ping_feats & LNET_PING_FEAT_MASK)) {
 666                 CDEBUG(D_NET, "%s: Unexpected features 0x%x\n",
 667                        libcfs_nid2str(gw->lp_nid), gw->lp_ping_feats);
 668                 return; /* nothing I can understand */
 669         }
 670
 671         if (!(gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS))
 672                 return; /* can't carry NI status info */
 673
 674         list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
 675                 int down = 0;
 676                 int up = 0;
 677                 int i;
 678
 679                 if (gw->lp_ping_feats & LNET_PING_FEAT_RTE_DISABLED) {
 680                         rte->lr_downis = 1;
 681                         continue;
 682                 }
 683
 684                 for (i = 0; i < info->pi_nnis && i < LNET_MAX_RTR_NIS; i++) {
 685                         lnet_ni_status_t *stat = &info->pi_ni[i];
 686                         lnet_nid_t nid = stat->ns_nid;
 687
 688                         if (nid == LNET_NID_ANY) {
 689                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 690                                        libcfs_nid2str(gw->lp_nid));
 691                                 gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 692                                 return;
 693                         }
 694
 695                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 696                                 continue;
 697
 698                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 699                                 down++;
 700                                 continue;
 701                         }
 702
 703                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 704                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 705                                         up = 1;
 706                                         break;
 707                                 }
 708                                 continue;
 709                         }
 710
 711                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 712                                libcfs_nid2str(gw->lp_nid), stat->ns_status);
 713                         gw->lp_ping_feats = LNET_PING_FEAT_INVAL;
 714                         return;
 715                 }
 716
 717                 if (up) { /* ignore downed NIs if NI for dest network is up */
 718                         rte->lr_downis = 0;
 719                         continue;
 720                 }
 721                 /**
 722                  * if @down is zero and this route is single-hop, it means
 723                  * we can't find NI for target network
 724                  */
 725                 if (!down && rte->lr_hops == 1)
 726                         down = 1;
 727
 728                 rte->lr_downis = down;
 729         }
 730 }
 731
 732 static void
 733 lnet_router_checker_event(lnet_event_t *event)
 734 {
 735         lnet_rc_data_t *rcd = event->md.user_ptr;
 736         struct lnet_peer *lp;
 737
 738         LASSERT(rcd);
 739
 740         if (event->unlinked) {
 741                 LNetInvalidateHandle(&rcd->rcd_mdh);
 742                 return;
 743         }
 744
 745         LASSERT(event->type == LNET_EVENT_SEND ||
 746                 event->type == LNET_EVENT_REPLY);
 747
 748         lp = rcd->rcd_gateway;
 749         LASSERT(lp);
 750
 751         /*
 752          * NB: it's called with holding lnet_res_lock, we have a few
 753          * places need to hold both locks at the same time, please take
 754          * care of lock ordering
 755          */
 756         lnet_net_lock(lp->lp_cpt);
 757         if (!lnet_isrouter(lp) || lp->lp_rcd != rcd) {
 758                 /* ignore if no longer a router or rcd is replaced */
 759                 goto out;
 760         }
 761
 762         if (event->type == LNET_EVENT_SEND) {
 763                 lp->lp_ping_notsent = 0;
 764                 if (!event->status)
 765                         goto out;
 766         }
 767
 768         /* LNET_EVENT_REPLY */
 769         /*
 770          * A successful REPLY means the router is up.  If _any_ comms
 771          * to the router fail I assume it's down (this will happen if
 772          * we ping alive routers to try to detect router death before
 773          * apps get burned).
 774          */
 775         lnet_notify_locked(lp, 1, !event->status, cfs_time_current());
 776
 777         /*
 778          * The router checker will wake up very shortly and do the
 779          * actual notification.
 780          * XXX If 'lp' stops being a router before then, it will still
 781          * have the notification pending!!!
 782          */
 783         if (avoid_asym_router_failure && !event->status)
 784                 lnet_parse_rc_info(rcd);
 785
 786  out:
 787         lnet_net_unlock(lp->lp_cpt);
 788 }
 789
 790 static void
 791 lnet_wait_known_routerstate(void)
 792 {
 793         lnet_peer_t *rtr;
 794         struct list_head *entry;
 795         int all_known;
 796
 797         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
 798
 799         for (;;) {
 800                 int cpt = lnet_net_lock_current();
 801
 802                 all_known = 1;
 803                 list_for_each(entry, &the_lnet.ln_routers) {
 804                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
 805
 806                         if (!rtr->lp_alive_count) {
 807                                 all_known = 0;
 808                                 break;
 809                         }
 810                 }
 811
 812                 lnet_net_unlock(cpt);
 813
 814                 if (all_known)
 815                         return;
 816
 817                 set_current_state(TASK_UNINTERRUPTIBLE);
 818                 schedule_timeout(cfs_time_seconds(1));
 819         }
 820 }
 821
 822 void
 823 lnet_router_ni_update_locked(lnet_peer_t *gw, __u32 net)
 824 {
 825         lnet_route_t *rte;
 826
 827         if ((gw->lp_ping_feats & LNET_PING_FEAT_NI_STATUS)) {
 828                 list_for_each_entry(rte, &gw->lp_routes, lr_gwlist) {
 829                         if (rte->lr_net == net) {
 830                                 rte->lr_downis = 0;
 831                                 break;
 832                         }
 833                 }
 834         }
 835 }
 836
 837 static void
 838 lnet_update_ni_status_locked(void)
 839 {
 840         lnet_ni_t *ni;
 841         time64_t now;
 842         int timeout;
 843
 844         LASSERT(the_lnet.ln_routing);
 845
 846         timeout = router_ping_timeout +
 847                   max(live_router_check_interval, dead_router_check_interval);
 848
 849         now = ktime_get_real_seconds();
 850         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 851                 if (ni->ni_lnd->lnd_type == LOLND)
 852                         continue;
 853
 854                 if (now < ni->ni_last_alive + timeout)
 855                         continue;
 856
 857                 lnet_ni_lock(ni);
 858                 /* re-check with lock */
 859                 if (now < ni->ni_last_alive + timeout) {
 860                         lnet_ni_unlock(ni);
 861                         continue;
 862                 }
 863
 864                 LASSERT(ni->ni_status);
 865
 866                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 867                         CDEBUG(D_NET, "NI(%s:%d) status changed to down\n",
 868                                libcfs_nid2str(ni->ni_nid), timeout);
 869                         /*
 870                          * NB: so far, this is the only place to set
 871                          * NI status to "down"
 872                          */
 873                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 874                 }
 875                 lnet_ni_unlock(ni);
 876         }
 877 }
 878
 879 static void
 880 lnet_destroy_rc_data(lnet_rc_data_t *rcd)
 881 {
 882         LASSERT(list_empty(&rcd->rcd_list));
 883         /* detached from network */
 884         LASSERT(LNetHandleIsInvalid(rcd->rcd_mdh));
 885
 886         if (rcd->rcd_gateway) {
 887                 int cpt = rcd->rcd_gateway->lp_cpt;
 888
 889                 lnet_net_lock(cpt);
 890                 lnet_peer_decref_locked(rcd->rcd_gateway);
 891                 lnet_net_unlock(cpt);
 892         }
 893
 894         if (rcd->rcd_pinginfo)
 895                 LIBCFS_FREE(rcd->rcd_pinginfo, LNET_PINGINFO_SIZE);
 896
 897         LIBCFS_FREE(rcd, sizeof(*rcd));
 898 }
 899
 900 static lnet_rc_data_t *
 901 lnet_create_rc_data_locked(lnet_peer_t *gateway)
 902 {
 903         lnet_rc_data_t *rcd = NULL;
 904         lnet_ping_info_t *pi;
 905         int rc;
 906         int i;
 907
 908         lnet_net_unlock(gateway->lp_cpt);
 909
 910         LIBCFS_ALLOC(rcd, sizeof(*rcd));
 911         if (!rcd)
 912                 goto out;
 913
 914         LNetInvalidateHandle(&rcd->rcd_mdh);
 915         INIT_LIST_HEAD(&rcd->rcd_list);
 916
 917         LIBCFS_ALLOC(pi, LNET_PINGINFO_SIZE);
 918         if (!pi)
 919                 goto out;
 920
 921         for (i = 0; i < LNET_MAX_RTR_NIS; i++) {
 922                 pi->pi_ni[i].ns_nid = LNET_NID_ANY;
 923                 pi->pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 924         }
 925         rcd->rcd_pinginfo = pi;
 926
 927         LASSERT(!LNetHandleIsInvalid(the_lnet.ln_rc_eqh));
 928         rc = LNetMDBind((lnet_md_t){.start     = pi,
 929                                     .user_ptr  = rcd,
 930                                     .length    = LNET_PINGINFO_SIZE,
 931                                     .threshold = LNET_MD_THRESH_INF,
 932                                     .options   = LNET_MD_TRUNCATE,
 933                                     .eq_handle = the_lnet.ln_rc_eqh},
 934                         LNET_UNLINK,
 935                         &rcd->rcd_mdh);
 936         if (rc < 0) {
 937                 CERROR("Can't bind MD: %d\n", rc);
 938                 goto out;
 939         }
 940         LASSERT(!rc);
 941
 942         lnet_net_lock(gateway->lp_cpt);
 943         /* router table changed or someone has created rcd for this gateway */
 944         if (!lnet_isrouter(gateway) || gateway->lp_rcd) {
 945                 lnet_net_unlock(gateway->lp_cpt);
 946                 goto out;
 947         }
 948
 949         lnet_peer_addref_locked(gateway);
 950         rcd->rcd_gateway = gateway;
 951         gateway->lp_rcd = rcd;
 952         gateway->lp_ping_notsent = 0;
 953
 954         return rcd;
 955
 956  out:
 957         if (rcd) {
 958                 if (!LNetHandleIsInvalid(rcd->rcd_mdh)) {
 959                         rc = LNetMDUnlink(rcd->rcd_mdh);
 960                         LASSERT(!rc);
 961                 }
 962                 lnet_destroy_rc_data(rcd);
 963         }
 964
 965         lnet_net_lock(gateway->lp_cpt);
 966         return gateway->lp_rcd;
 967 }
 968
 969 static int
 970 lnet_router_check_interval(lnet_peer_t *rtr)
 971 {
 972         int secs;
 973
 974         secs = rtr->lp_alive ? live_router_check_interval :
 975                                dead_router_check_interval;
 976         if (secs < 0)
 977                 secs = 0;
 978
 979         return secs;
 980 }
 981
 982 static void
 983 lnet_ping_router_locked(lnet_peer_t *rtr)
 984 {
 985         lnet_rc_data_t *rcd = NULL;
 986         unsigned long now = cfs_time_current();
 987         int secs;
 988
 989         lnet_peer_addref_locked(rtr);
 990
 991         if (rtr->lp_ping_deadline && /* ping timed out? */
 992             cfs_time_after(now, rtr->lp_ping_deadline))
 993                 lnet_notify_locked(rtr, 1, 0, now);
 994
 995         /* Run any outstanding notifications */
 996         lnet_ni_notify_locked(rtr->lp_ni, rtr);
 997
 998         if (!lnet_isrouter(rtr) ||
 999             the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1000                 /* router table changed or router checker is shutting down */
1001                 lnet_peer_decref_locked(rtr);
1002                 return;
1003         }
1004
1005         rcd = rtr->lp_rcd ?
1006               rtr->lp_rcd : lnet_create_rc_data_locked(rtr);
1007
1008         if (!rcd)
1009                 return;
1010
1011         secs = lnet_router_check_interval(rtr);
1012
1013         CDEBUG(D_NET,
1014                "rtr %s %d: deadline %lu ping_notsent %d alive %d alive_count %d lp_ping_timestamp %lu\n",
1015                libcfs_nid2str(rtr->lp_nid), secs,
1016                rtr->lp_ping_deadline, rtr->lp_ping_notsent,
1017                rtr->lp_alive, rtr->lp_alive_count, rtr->lp_ping_timestamp);
1018
1019         if (secs && !rtr->lp_ping_notsent &&
1020             cfs_time_after(now, cfs_time_add(rtr->lp_ping_timestamp,
1021                                              cfs_time_seconds(secs)))) {
1022                 int rc;
1023                 lnet_process_id_t id;
1024                 lnet_handle_md_t mdh;
1025
1026                 id.nid = rtr->lp_nid;
1027                 id.pid = LNET_PID_LUSTRE;
1028                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1029
1030                 rtr->lp_ping_notsent   = 1;
1031                 rtr->lp_ping_timestamp = now;
1032
1033                 mdh = rcd->rcd_mdh;
1034
1035                 if (!rtr->lp_ping_deadline) {
1036                         rtr->lp_ping_deadline =
1037                                 cfs_time_shift(router_ping_timeout);
1038                 }
1039
1040                 lnet_net_unlock(rtr->lp_cpt);
1041
1042                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1043                              LNET_PROTO_PING_MATCHBITS, 0);
1044
1045                 lnet_net_lock(rtr->lp_cpt);
1046                 if (rc)
1047                         rtr->lp_ping_notsent = 0; /* no event pending */
1048         }
1049
1050         lnet_peer_decref_locked(rtr);
1051 }
1052
1053 int
1054 lnet_router_checker_start(void)
1055 {
1056         struct task_struct *task;
1057         int rc;
1058         int eqsz = 0;
1059
1060         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1061
1062         if (check_routers_before_use &&
1063             dead_router_check_interval <= 0) {
1064                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be set if 'check_routers_before_use' is set\n");
1065                 return -EINVAL;
1066         }
1067
1068         sema_init(&the_lnet.ln_rc_signal, 0);
1069
1070         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1071         if (rc) {
1072                 CERROR("Can't allocate EQ(%d): %d\n", eqsz, rc);
1073                 return -ENOMEM;
1074         }
1075
1076         the_lnet.ln_rc_state = LNET_RC_STATE_RUNNING;
1077         task = kthread_run(lnet_router_checker, NULL, "router_checker");
1078         if (IS_ERR(task)) {
1079                 rc = PTR_ERR(task);
1080                 CERROR("Can't start router checker thread: %d\n", rc);
1081                 /* block until event callback signals exit */
1082                 down(&the_lnet.ln_rc_signal);
1083                 rc = LNetEQFree(the_lnet.ln_rc_eqh);
1084                 LASSERT(!rc);
1085                 the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1086                 return -ENOMEM;
1087         }
1088
1089         if (check_routers_before_use) {
1090                 /*
1091                  * Note that a helpful side-effect of pinging all known routers
1092                  * at startup is that it makes them drop stale connections they
1093                  * may have to a previous instance of me.
1094                  */
1095                 lnet_wait_known_routerstate();
1096         }
1097
1098         return 0;
1099 }
1100
1101 void
1102 lnet_router_checker_stop(void)
1103 {
1104         int rc;
1105
1106         if (the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN)
1107                 return;
1108
1109         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING);
1110         the_lnet.ln_rc_state = LNET_RC_STATE_STOPPING;
1111         /* wakeup the RC thread if it's sleeping */
1112         wake_up(&the_lnet.ln_rc_waitq);
1113
1114         /* block until event callback signals exit */
1115         down(&the_lnet.ln_rc_signal);
1116         LASSERT(the_lnet.ln_rc_state == LNET_RC_STATE_SHUTDOWN);
1117
1118         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1119         LASSERT(!rc);
1120 }
1121
1122 static void
1123 lnet_prune_rc_data(int wait_unlink)
1124 {
1125         lnet_rc_data_t *rcd;
1126         lnet_rc_data_t *tmp;
1127         lnet_peer_t *lp;
1128         struct list_head head;
1129         int i = 2;
1130
1131         if (likely(the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING &&
1132                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1133                    list_empty(&the_lnet.ln_rcd_zombie)))
1134                 return;
1135
1136         INIT_LIST_HEAD(&head);
1137
1138         lnet_net_lock(LNET_LOCK_EX);
1139
1140         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING) {
1141                 /* router checker is stopping, prune all */
1142                 list_for_each_entry(lp, &the_lnet.ln_routers,
1143                                     lp_rtr_list) {
1144                         if (!lp->lp_rcd)
1145                                 continue;
1146
1147                         LASSERT(list_empty(&lp->lp_rcd->rcd_list));
1148                         list_add(&lp->lp_rcd->rcd_list,
1149                                  &the_lnet.ln_rcd_deathrow);
1150                         lp->lp_rcd = NULL;
1151                 }
1152         }
1153
1154         /* unlink all RCDs on deathrow list */
1155         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1156
1157         if (!list_empty(&head)) {
1158                 lnet_net_unlock(LNET_LOCK_EX);
1159
1160                 list_for_each_entry(rcd, &head, rcd_list)
1161                         LNetMDUnlink(rcd->rcd_mdh);
1162
1163                 lnet_net_lock(LNET_LOCK_EX);
1164         }
1165
1166         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1167
1168         /* release all zombie RCDs */
1169         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1170                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1171                                          rcd_list) {
1172                         if (LNetHandleIsInvalid(rcd->rcd_mdh))
1173                                 list_move(&rcd->rcd_list, &head);
1174                 }
1175
1176                 wait_unlink = wait_unlink &&
1177                               !list_empty(&the_lnet.ln_rcd_zombie);
1178
1179                 lnet_net_unlock(LNET_LOCK_EX);
1180
1181                 while (!list_empty(&head)) {
1182                         rcd = list_entry(head.next,
1183                                          lnet_rc_data_t, rcd_list);
1184                         list_del_init(&rcd->rcd_list);
1185                         lnet_destroy_rc_data(rcd);
1186                 }
1187
1188                 if (!wait_unlink)
1189                         return;
1190
1191                 i++;
1192                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1193                        "Waiting for rc buffers to unlink\n");
1194                 set_current_state(TASK_UNINTERRUPTIBLE);
1195                 schedule_timeout(cfs_time_seconds(1) / 4);
1196
1197                 lnet_net_lock(LNET_LOCK_EX);
1198         }
1199
1200         lnet_net_unlock(LNET_LOCK_EX);
1201 }
1202
1203 /*
1204  * This function is called to check if the RC should block indefinitely.
1205  * It's called from lnet_router_checker() as well as being passed to
1206  * wait_event_interruptible() to avoid the lost wake_up problem.
1207  *
1208  * When it's called from wait_event_interruptible() it is necessary to
1209  * also not sleep if the rc state is not running to avoid a deadlock
1210  * when the system is shutting down
1211  */
1212 static inline bool
1213 lnet_router_checker_active(void)
1214 {
1215         if (the_lnet.ln_rc_state != LNET_RC_STATE_RUNNING)
1216                 return true;
1217
1218         /*
1219          * Router Checker thread needs to run when routing is enabled in
1220          * order to call lnet_update_ni_status_locked()
1221          */
1222         if (the_lnet.ln_routing)
1223                 return true;
1224
1225         return !list_empty(&the_lnet.ln_routers) &&
1226                 (live_router_check_interval > 0 ||
1227                  dead_router_check_interval > 0);
1228 }
1229
1230 static int
1231 lnet_router_checker(void *arg)
1232 {
1233         lnet_peer_t *rtr;
1234         struct list_head *entry;
1235
1236         cfs_block_allsigs();
1237
1238         while (the_lnet.ln_rc_state == LNET_RC_STATE_RUNNING) {
1239                 __u64 version;
1240                 int cpt;
1241                 int cpt2;
1242
1243                 cpt = lnet_net_lock_current();
1244 rescan:
1245                 version = the_lnet.ln_routers_version;
1246
1247                 list_for_each(entry, &the_lnet.ln_routers) {
1248                         rtr = list_entry(entry, lnet_peer_t, lp_rtr_list);
1249
1250                         cpt2 = lnet_cpt_of_nid_locked(rtr->lp_nid);
1251                         if (cpt != cpt2) {
1252                                 lnet_net_unlock(cpt);
1253                                 cpt = cpt2;
1254                                 lnet_net_lock(cpt);
1255                                 /* the routers list has changed */
1256                                 if (version != the_lnet.ln_routers_version)
1257                                         goto rescan;
1258                         }
1259
1260                         lnet_ping_router_locked(rtr);
1261
1262                         /* NB dropped lock */
1263                         if (version != the_lnet.ln_routers_version) {
1264                                 /* the routers list has changed */
1265                                 goto rescan;
1266                         }
1267                 }
1268
1269                 if (the_lnet.ln_routing)
1270                         lnet_update_ni_status_locked();
1271
1272                 lnet_net_unlock(cpt);
1273
1274                 lnet_prune_rc_data(0); /* don't wait for UNLINK */
1275
1276                 /*
1277                  * Call schedule_timeout() here always adds 1 to load average
1278                  * because kernel counts # active tasks as nr_running
1279                  * + nr_uninterruptible.
1280                  */
1281                 /*
1282                  * if there are any routes then wakeup every second.  If
1283                  * there are no routes then sleep indefinitely until woken
1284                  * up by a user adding a route
1285                  */
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         lnet_prune_rc_data(1); /* wait for UNLINK */
1296
1297         the_lnet.ln_rc_state = LNET_RC_STATE_SHUTDOWN;
1298         up(&the_lnet.ln_rc_signal);
1299         /* The unlink event callback will signal final completion */
1300         return 0;
1301 }
1302
1303 void
1304 lnet_destroy_rtrbuf(lnet_rtrbuf_t *rb, int npages)
1305 {
1306         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1307
1308         while (--npages >= 0)
1309                 __free_page(rb->rb_kiov[npages].kiov_page);
1310
1311         LIBCFS_FREE(rb, sz);
1312 }
1313
1314 static lnet_rtrbuf_t *
1315 lnet_new_rtrbuf(lnet_rtrbufpool_t *rbp, int cpt)
1316 {
1317         int npages = rbp->rbp_npages;
1318         int sz = offsetof(lnet_rtrbuf_t, rb_kiov[npages]);
1319         struct page *page;
1320         lnet_rtrbuf_t *rb;
1321         int i;
1322
1323         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1324         if (!rb)
1325                 return NULL;
1326
1327         rb->rb_pool = rbp;
1328
1329         for (i = 0; i < npages; i++) {
1330                 page = alloc_pages_node(
1331                                 cfs_cpt_spread_node(lnet_cpt_table(), cpt),
1332                                 GFP_KERNEL | __GFP_ZERO, 0);
1333                 if (!page) {
1334                         while (--i >= 0)
1335                                 __free_page(rb->rb_kiov[i].kiov_page);
1336
1337                         LIBCFS_FREE(rb, sz);
1338                         return NULL;
1339                 }
1340
1341                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1342                 rb->rb_kiov[i].kiov_offset = 0;
1343                 rb->rb_kiov[i].kiov_page = page;
1344         }
1345
1346         return rb;
1347 }
1348
1349 static void
1350 lnet_rtrpool_free_bufs(lnet_rtrbufpool_t *rbp, int cpt)
1351 {
1352         int npages = rbp->rbp_npages;
1353         struct list_head tmp;
1354         lnet_rtrbuf_t *rb;
1355         lnet_rtrbuf_t *temp;
1356
1357         if (!rbp->rbp_nbuffers) /* 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_req_nbuffers = 0;
1366         rbp->rbp_nbuffers = 0;
1367         rbp->rbp_credits = 0;
1368         rbp->rbp_mincredits = 0;
1369         lnet_net_unlock(cpt);
1370
1371         /* Free buffers on the free list. */
1372         list_for_each_entry_safe(rb, temp, &tmp, rb_list) {
1373                 list_del(&rb->rb_list);
1374                 lnet_destroy_rtrbuf(rb, npages);
1375         }
1376 }
1377
1378 static int
1379 lnet_rtrpool_adjust_bufs(lnet_rtrbufpool_t *rbp, int nbufs, int cpt)
1380 {
1381         struct list_head rb_list;
1382         lnet_rtrbuf_t *rb;
1383         int num_rb;
1384         int num_buffers = 0;
1385         int old_req_nbufs;
1386         int npages = rbp->rbp_npages;
1387
1388         lnet_net_lock(cpt);
1389         /*
1390          * If we are called for less buffers than already in the pool, we
1391          * just lower the req_nbuffers number and excess buffers will be
1392          * thrown away as they are returned to the free list.  Credits
1393          * then get adjusted as well.
1394          * If we already have enough buffers allocated to serve the
1395          * increase requested, then we can treat that the same way as we
1396          * do the decrease.
1397          */
1398         num_rb = nbufs - rbp->rbp_nbuffers;
1399         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1400                 rbp->rbp_req_nbuffers = nbufs;
1401                 lnet_net_unlock(cpt);
1402                 return 0;
1403         }
1404         /*
1405          * store the older value of rbp_req_nbuffers and then set it to
1406          * the new request to prevent lnet_return_rx_credits_locked() from
1407          * freeing buffers that we need to keep around
1408          */
1409         old_req_nbufs = rbp->rbp_req_nbuffers;
1410         rbp->rbp_req_nbuffers = nbufs;
1411         lnet_net_unlock(cpt);
1412
1413         INIT_LIST_HEAD(&rb_list);
1414
1415         /*
1416          * allocate the buffers on a local list first.  If all buffers are
1417          * allocated successfully then join this list to the rbp buffer
1418          * list. If not then free all allocated buffers.
1419          */
1420         while (num_rb-- > 0) {
1421                 rb = lnet_new_rtrbuf(rbp, cpt);
1422                 if (!rb) {
1423                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1424                                nbufs, npages);
1425
1426                         lnet_net_lock(cpt);
1427                         rbp->rbp_req_nbuffers = old_req_nbufs;
1428                         lnet_net_unlock(cpt);
1429
1430                         goto failed;
1431                 }
1432
1433                 list_add(&rb->rb_list, &rb_list);
1434                 num_buffers++;
1435         }
1436
1437         lnet_net_lock(cpt);
1438
1439         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1440         rbp->rbp_nbuffers += num_buffers;
1441         rbp->rbp_credits += num_buffers;
1442         rbp->rbp_mincredits = rbp->rbp_credits;
1443         /*
1444          * We need to schedule blocked msg using the newly
1445          * added buffers.
1446          */
1447         while (!list_empty(&rbp->rbp_bufs) &&
1448                !list_empty(&rbp->rbp_msgs))
1449                 lnet_schedule_blocked_locked(rbp);
1450
1451         lnet_net_unlock(cpt);
1452
1453         return 0;
1454
1455 failed:
1456         while (!list_empty(&rb_list)) {
1457                 rb = list_entry(rb_list.next, lnet_rtrbuf_t, rb_list);
1458                 list_del(&rb->rb_list);
1459                 lnet_destroy_rtrbuf(rb, npages);
1460         }
1461
1462         return -ENOMEM;
1463 }
1464
1465 static void
1466 lnet_rtrpool_init(lnet_rtrbufpool_t *rbp, int npages)
1467 {
1468         INIT_LIST_HEAD(&rbp->rbp_msgs);
1469         INIT_LIST_HEAD(&rbp->rbp_bufs);
1470
1471         rbp->rbp_npages = npages;
1472         rbp->rbp_credits = 0;
1473         rbp->rbp_mincredits = 0;
1474 }
1475
1476 void
1477 lnet_rtrpools_free(int keep_pools)
1478 {
1479         lnet_rtrbufpool_t *rtrp;
1480         int i;
1481
1482         if (!the_lnet.ln_rtrpools) /* uninitialized or freed */
1483                 return;
1484
1485         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1486                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1487                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1488                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1489         }
1490
1491         if (!keep_pools) {
1492                 cfs_percpt_free(the_lnet.ln_rtrpools);
1493                 the_lnet.ln_rtrpools = NULL;
1494         }
1495 }
1496
1497 static int
1498 lnet_nrb_tiny_calculate(void)
1499 {
1500         int nrbs = LNET_NRB_TINY;
1501
1502         if (tiny_router_buffers < 0) {
1503                 LCONSOLE_ERROR_MSG(0x10c,
1504                                    "tiny_router_buffers=%d invalid when routing enabled\n",
1505                                    tiny_router_buffers);
1506                 return -EINVAL;
1507         }
1508
1509         if (tiny_router_buffers > 0)
1510                 nrbs = tiny_router_buffers;
1511
1512         nrbs /= LNET_CPT_NUMBER;
1513         return max(nrbs, LNET_NRB_TINY_MIN);
1514 }
1515
1516 static int
1517 lnet_nrb_small_calculate(void)
1518 {
1519         int nrbs = LNET_NRB_SMALL;
1520
1521         if (small_router_buffers < 0) {
1522                 LCONSOLE_ERROR_MSG(0x10c,
1523                                    "small_router_buffers=%d invalid when routing enabled\n",
1524                                    small_router_buffers);
1525                 return -EINVAL;
1526         }
1527
1528         if (small_router_buffers > 0)
1529                 nrbs = small_router_buffers;
1530
1531         nrbs /= LNET_CPT_NUMBER;
1532         return max(nrbs, LNET_NRB_SMALL_MIN);
1533 }
1534
1535 static int
1536 lnet_nrb_large_calculate(void)
1537 {
1538         int nrbs = LNET_NRB_LARGE;
1539
1540         if (large_router_buffers < 0) {
1541                 LCONSOLE_ERROR_MSG(0x10c,
1542                                    "large_router_buffers=%d invalid when routing enabled\n",
1543                                    large_router_buffers);
1544                 return -EINVAL;
1545         }
1546
1547         if (large_router_buffers > 0)
1548                 nrbs = large_router_buffers;
1549
1550         nrbs /= LNET_CPT_NUMBER;
1551         return max(nrbs, LNET_NRB_LARGE_MIN);
1552 }
1553
1554 int
1555 lnet_rtrpools_alloc(int im_a_router)
1556 {
1557         lnet_rtrbufpool_t *rtrp;
1558         int nrb_tiny;
1559         int nrb_small;
1560         int nrb_large;
1561         int rc;
1562         int i;
1563
1564         if (!strcmp(forwarding, "")) {
1565                 /* not set either way */
1566                 if (!im_a_router)
1567                         return 0;
1568         } else if (!strcmp(forwarding, "disabled")) {
1569                 /* explicitly disabled */
1570                 return 0;
1571         } else if (!strcmp(forwarding, "enabled")) {
1572                 /* explicitly enabled */
1573         } else {
1574                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either 'enabled' or 'disabled'\n");
1575                 return -EINVAL;
1576         }
1577
1578         nrb_tiny = lnet_nrb_tiny_calculate();
1579         if (nrb_tiny < 0)
1580                 return -EINVAL;
1581
1582         nrb_small = lnet_nrb_small_calculate();
1583         if (nrb_small < 0)
1584                 return -EINVAL;
1585
1586         nrb_large = lnet_nrb_large_calculate();
1587         if (nrb_large < 0)
1588                 return -EINVAL;
1589
1590         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1591                                                 LNET_NRBPOOLS *
1592                                                 sizeof(lnet_rtrbufpool_t));
1593         if (!the_lnet.ln_rtrpools) {
1594                 LCONSOLE_ERROR_MSG(0x10c,
1595                                    "Failed to initialize router buffe pool\n");
1596                 return -ENOMEM;
1597         }
1598
1599         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1600                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1601                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1602                                               nrb_tiny, i);
1603                 if (rc)
1604                         goto failed;
1605
1606                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1607                                   LNET_NRB_SMALL_PAGES);
1608                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1609                                               nrb_small, i);
1610                 if (rc)
1611                         goto failed;
1612
1613                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1614                                   LNET_NRB_LARGE_PAGES);
1615                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1616                                               nrb_large, i);
1617                 if (rc)
1618                         goto failed;
1619         }
1620
1621         lnet_net_lock(LNET_LOCK_EX);
1622         the_lnet.ln_routing = 1;
1623         lnet_net_unlock(LNET_LOCK_EX);
1624
1625         return 0;
1626
1627  failed:
1628         lnet_rtrpools_free(0);
1629         return rc;
1630 }
1631
1632 static int
1633 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1634 {
1635         int nrb = 0;
1636         int rc = 0;
1637         int i;
1638         lnet_rtrbufpool_t *rtrp;
1639
1640         /*
1641          * If the provided values for each buffer pool are different than the
1642          * configured values, we need to take action.
1643          */
1644         if (tiny >= 0) {
1645                 tiny_router_buffers = tiny;
1646                 nrb = lnet_nrb_tiny_calculate();
1647                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1648                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1649                                                       nrb, i);
1650                         if (rc)
1651                                 return rc;
1652                 }
1653         }
1654         if (small >= 0) {
1655                 small_router_buffers = small;
1656                 nrb = lnet_nrb_small_calculate();
1657                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1658                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1659                                                       nrb, i);
1660                         if (rc)
1661                                 return rc;
1662                 }
1663         }
1664         if (large >= 0) {
1665                 large_router_buffers = large;
1666                 nrb = lnet_nrb_large_calculate();
1667                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1668                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1669                                                       nrb, i);
1670                         if (rc)
1671                                 return rc;
1672                 }
1673         }
1674
1675         return 0;
1676 }
1677
1678 int
1679 lnet_rtrpools_adjust(int tiny, int small, int large)
1680 {
1681         /*
1682          * this function doesn't revert the changes if adding new buffers
1683          * failed.  It's up to the user space caller to revert the
1684          * changes.
1685          */
1686         if (!the_lnet.ln_routing)
1687                 return 0;
1688
1689         return lnet_rtrpools_adjust_helper(tiny, small, large);
1690 }
1691
1692 int
1693 lnet_rtrpools_enable(void)
1694 {
1695         int rc;
1696
1697         if (the_lnet.ln_routing)
1698                 return 0;
1699
1700         if (!the_lnet.ln_rtrpools)
1701                 /*
1702                  * If routing is turned off, and we have never
1703                  * initialized the pools before, just call the
1704                  * standard buffer pool allocation routine as
1705                  * if we are just configuring this for the first
1706                  * time.
1707                  */
1708                 return lnet_rtrpools_alloc(1);
1709
1710         rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1711         if (rc)
1712                 return rc;
1713
1714         lnet_net_lock(LNET_LOCK_EX);
1715         the_lnet.ln_routing = 1;
1716
1717         the_lnet.ln_ping_info->pi_features &= ~LNET_PING_FEAT_RTE_DISABLED;
1718         lnet_net_unlock(LNET_LOCK_EX);
1719
1720         return 0;
1721 }
1722
1723 void
1724 lnet_rtrpools_disable(void)
1725 {
1726         if (!the_lnet.ln_routing)
1727                 return;
1728
1729         lnet_net_lock(LNET_LOCK_EX);
1730         the_lnet.ln_routing = 0;
1731         the_lnet.ln_ping_info->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1732
1733         tiny_router_buffers = 0;
1734         small_router_buffers = 0;
1735         large_router_buffers = 0;
1736         lnet_net_unlock(LNET_LOCK_EX);
1737         lnet_rtrpools_free(1);
1738 }
1739
1740 int
1741 lnet_notify(lnet_ni_t *ni, lnet_nid_t nid, int alive, unsigned long when)
1742 {
1743         struct lnet_peer *lp = NULL;
1744         unsigned long now = cfs_time_current();
1745         int cpt = lnet_cpt_of_nid(nid);
1746
1747         LASSERT(!in_interrupt());
1748
1749         CDEBUG(D_NET, "%s notifying %s: %s\n",
1750                !ni ? "userspace" : libcfs_nid2str(ni->ni_nid),
1751                libcfs_nid2str(nid),
1752                alive ? "up" : "down");
1753
1754         if (ni &&
1755             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1756                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1757                       libcfs_nid2str(nid), alive ? "birth" : "death",
1758                       libcfs_nid2str(ni->ni_nid));
1759                 return -EINVAL;
1760         }
1761
1762         /* can't do predictions... */
1763         if (cfs_time_after(when, now)) {
1764                 CWARN("Ignoring prediction from %s of %s %s %ld seconds in the future\n",
1765                       !ni ? "userspace" : libcfs_nid2str(ni->ni_nid),
1766                       libcfs_nid2str(nid), alive ? "up" : "down",
1767                       cfs_duration_sec(cfs_time_sub(when, now)));
1768                 return -EINVAL;
1769         }
1770
1771         if (ni && !alive &&          /* LND telling me she's down */
1772             !auto_down) {                      /* auto-down disabled */
1773                 CDEBUG(D_NET, "Auto-down disabled\n");
1774                 return 0;
1775         }
1776
1777         lnet_net_lock(cpt);
1778
1779         if (the_lnet.ln_shutdown) {
1780                 lnet_net_unlock(cpt);
1781                 return -ESHUTDOWN;
1782         }
1783
1784         lp = lnet_find_peer_locked(the_lnet.ln_peer_tables[cpt], nid);
1785         if (!lp) {
1786                 /* nid not found */
1787                 lnet_net_unlock(cpt);
1788                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1789                 return 0;
1790         }
1791
1792         /*
1793          * We can't fully trust LND on reporting exact peer last_alive
1794          * if he notifies us about dead peer. For example ksocklnd can
1795          * call us with when == _time_when_the_node_was_booted_ if
1796          * no connections were successfully established
1797          */
1798         if (ni && !alive && when < lp->lp_last_alive)
1799                 when = lp->lp_last_alive;
1800
1801         lnet_notify_locked(lp, !ni, alive, when);
1802
1803         if (ni)
1804                 lnet_ni_notify_locked(ni, lp);
1805
1806         lnet_peer_decref_locked(lp);
1807
1808         lnet_net_unlock(cpt);
1809         return 0;
1810 }
1811 EXPORT_SYMBOL(lnet_notify);