lnet/lnet/router.c

   1 /*
   2  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
   3  *
   4  * Copyright (c) 2011, 2017, Intel Corporation.
   5  *
   6  *   This file is part of Lustre, https://wiki.whamcloud.com/
   7  *
   8  *   Portals is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Portals is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Portals; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_LNET
  24
  25 #include <linux/random.h>
  26 #include <lnet/lib-lnet.h>
  27
  28 #define LNET_NRB_TINY_MIN       512     /* min value for each CPT */
  29 #define LNET_NRB_TINY           (LNET_NRB_TINY_MIN * 4)
  30 #define LNET_NRB_SMALL_MIN      4096    /* min value for each CPT */
  31 #define LNET_NRB_SMALL          (LNET_NRB_SMALL_MIN * 4)
  32 #define LNET_NRB_SMALL_PAGES    1
  33 #define LNET_NRB_LARGE_MIN      256     /* min value for each CPT */
  34 #define LNET_NRB_LARGE          (LNET_NRB_LARGE_MIN * 4)
  35 #define LNET_NRB_LARGE_PAGES    ((LNET_MTU + PAGE_SIZE - 1) >> \
  36                                   PAGE_SHIFT)
  37
  38 static char *forwarding = "";
  39 module_param(forwarding, charp, 0444);
  40 MODULE_PARM_DESC(forwarding, "Explicitly enable/disable forwarding between networks");
  41
  42 static int tiny_router_buffers;
  43 module_param(tiny_router_buffers, int, 0444);
  44 MODULE_PARM_DESC(tiny_router_buffers, "# of 0 payload messages to buffer in the router");
  45 static int small_router_buffers;
  46 module_param(small_router_buffers, int, 0444);
  47 MODULE_PARM_DESC(small_router_buffers, "# of small (1 page) messages to buffer in the router");
  48 static int large_router_buffers;
  49 module_param(large_router_buffers, int, 0444);
  50 MODULE_PARM_DESC(large_router_buffers, "# of large messages to buffer in the router");
  51 static int peer_buffer_credits;
  52 module_param(peer_buffer_credits, int, 0444);
  53 MODULE_PARM_DESC(peer_buffer_credits, "# router buffer credits per peer");
  54
  55 static int auto_down = 1;
  56 module_param(auto_down, int, 0444);
  57 MODULE_PARM_DESC(auto_down, "Automatically mark peers down on comms error");
  58
  59 int
  60 lnet_peer_buffer_credits(struct lnet_net *net)
  61 {
  62         /* NI option overrides LNet default */
  63         if (net->net_tunables.lct_peer_rtr_credits > 0)
  64                 return net->net_tunables.lct_peer_rtr_credits;
  65         if (peer_buffer_credits > 0)
  66                 return peer_buffer_credits;
  67
  68         /* As an approximation, allow this peer the same number of router
  69          * buffers as it is allowed outstanding sends */
  70         return net->net_tunables.lct_peer_tx_credits;
  71 }
  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(struct lnet_peer_ni *lp, int notifylnd, int alive,
 101                    time64_t when)
 102 {
 103         if (lp->lpni_timestamp > when) { /* out of date information */
 104                 CDEBUG(D_NET, "Out of date\n");
 105                 return;
 106         }
 107
 108         /*
 109          * This function can be called with different cpt locks being
 110          * held. lpni_alive_count modification needs to be properly protected.
 111          * Significant reads to lpni_alive_count are also protected with
 112          * the same lock
 113          */
 114         spin_lock(&lp->lpni_lock);
 115
 116         lp->lpni_timestamp = when; /* update timestamp */
 117         lp->lpni_ping_deadline = 0;               /* disable ping timeout */
 118
 119         if (lp->lpni_alive_count != 0 &&          /* got old news */
 120             (!lp->lpni_alive) == (!alive)) {      /* new date for old news */
 121                 spin_unlock(&lp->lpni_lock);
 122                 CDEBUG(D_NET, "Old news\n");
 123                 return;
 124         }
 125
 126         /* Flag that notification is outstanding */
 127
 128         lp->lpni_alive_count++;
 129         lp->lpni_alive = (alive) ? 1 : 0;
 130         lp->lpni_notify = 1;
 131         lp->lpni_notifylnd = notifylnd;
 132         if (lp->lpni_alive)
 133                 lp->lpni_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 134
 135         spin_unlock(&lp->lpni_lock);
 136
 137         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive);
 138 }
 139
 140 /*
 141  * This function will always be called with lp->lpni_cpt lock held.
 142  */
 143 static void
 144 lnet_ni_notify_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
 145 {
 146         int alive;
 147         int notifylnd;
 148
 149         /* Notify only in 1 thread at any time to ensure ordered notification.
 150          * NB individual events can be missed; the only guarantee is that you
 151          * always get the most recent news */
 152
 153         spin_lock(&lp->lpni_lock);
 154
 155         if (lp->lpni_notifying || ni == NULL) {
 156                 spin_unlock(&lp->lpni_lock);
 157                 return;
 158         }
 159
 160         lp->lpni_notifying = 1;
 161
 162         /*
 163          * lp->lpni_notify needs to be protected because it can be set in
 164          * lnet_notify_locked().
 165          */
 166         while (lp->lpni_notify) {
 167                 alive     = lp->lpni_alive;
 168                 notifylnd = lp->lpni_notifylnd;
 169
 170                 lp->lpni_notifylnd = 0;
 171                 lp->lpni_notify    = 0;
 172
 173                 if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
 174                         spin_unlock(&lp->lpni_lock);
 175                         lnet_net_unlock(lp->lpni_cpt);
 176
 177                         /* A new notification could happen now; I'll handle it
 178                          * when control returns to me */
 179
 180                         (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid,
 181                                                           alive);
 182
 183                         lnet_net_lock(lp->lpni_cpt);
 184                         spin_lock(&lp->lpni_lock);
 185                 }
 186         }
 187
 188         lp->lpni_notifying = 0;
 189         spin_unlock(&lp->lpni_lock);
 190 }
 191
 192 static void
 193 lnet_rtr_addref_locked(struct lnet_peer_ni *lp)
 194 {
 195         LASSERT(atomic_read(&lp->lpni_refcount) > 0);
 196         LASSERT(lp->lpni_rtr_refcount >= 0);
 197
 198         /* lnet_net_lock must be exclusively locked */
 199         lp->lpni_rtr_refcount++;
 200         if (lp->lpni_rtr_refcount == 1) {
 201                 struct list_head *pos;
 202
 203                 /* a simple insertion sort */
 204                 list_for_each_prev(pos, &the_lnet.ln_routers) {
 205                         struct lnet_peer_ni *rtr;
 206
 207                         rtr = list_entry(pos, struct lnet_peer_ni,
 208                                          lpni_rtr_list);
 209                         if (rtr->lpni_nid < lp->lpni_nid)
 210                                 break;
 211                 }
 212
 213                 list_add(&lp->lpni_rtr_list, pos);
 214                 /* addref for the_lnet.ln_routers */
 215                 lnet_peer_ni_addref_locked(lp);
 216                 the_lnet.ln_routers_version++;
 217         }
 218 }
 219
 220 static void
 221 lnet_rtr_decref_locked(struct lnet_peer_ni *lp)
 222 {
 223         LASSERT(atomic_read(&lp->lpni_refcount) > 0);
 224         LASSERT(lp->lpni_rtr_refcount > 0);
 225
 226         /* lnet_net_lock must be exclusively locked */
 227         lp->lpni_rtr_refcount--;
 228         if (lp->lpni_rtr_refcount == 0) {
 229                 LASSERT(list_empty(&lp->lpni_routes));
 230
 231                 if (lp->lpni_rcd != NULL) {
 232                         list_add(&lp->lpni_rcd->rcd_list,
 233                                  &the_lnet.ln_rcd_deathrow);
 234                         lp->lpni_rcd = NULL;
 235                 }
 236
 237                 list_del(&lp->lpni_rtr_list);
 238                 /* decref for the_lnet.ln_routers */
 239                 lnet_peer_ni_decref_locked(lp);
 240                 the_lnet.ln_routers_version++;
 241         }
 242 }
 243
 244 struct lnet_remotenet *
 245 lnet_find_rnet_locked(__u32 net)
 246 {
 247         struct lnet_remotenet *rnet;
 248         struct list_head *tmp;
 249         struct list_head *rn_list;
 250
 251         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 252
 253         rn_list = lnet_net2rnethash(net);
 254         list_for_each(tmp, rn_list) {
 255                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 256
 257                 if (rnet->lrn_net == net)
 258                         return rnet;
 259         }
 260         return NULL;
 261 }
 262
 263 static void lnet_shuffle_seed(void)
 264 {
 265         static int seeded;
 266         struct lnet_ni *ni = NULL;
 267
 268         if (seeded)
 269                 return;
 270
 271         /* Nodes with small feet have little entropy
 272          * the NID for this node gives the most entropy in the low bits
 273          */
 274         while ((ni = lnet_get_next_ni_locked(NULL, ni)))
 275                 add_device_randomness(&ni->ni_nid, sizeof(ni->ni_nid));
 276
 277         seeded = 1;
 278         return;
 279 }
 280
 281 /* NB expects LNET_LOCK held */
 282 static void
 283 lnet_add_route_to_rnet(struct lnet_remotenet *rnet, struct lnet_route *route)
 284 {
 285         unsigned int      len = 0;
 286         unsigned int      offset = 0;
 287         struct list_head *e;
 288
 289         lnet_shuffle_seed();
 290
 291         list_for_each(e, &rnet->lrn_routes) {
 292                 len++;
 293         }
 294
 295         /* len+1 positions to add a new entry, also prevents division by 0 */
 296         offset = cfs_rand() % (len + 1);
 297         list_for_each(e, &rnet->lrn_routes) {
 298                 if (offset == 0)
 299                         break;
 300                 offset--;
 301         }
 302         list_add(&route->lr_list, e);
 303         list_add(&route->lr_gwlist, &route->lr_gateway->lpni_routes);
 304
 305         the_lnet.ln_remote_nets_version++;
 306         lnet_rtr_addref_locked(route->lr_gateway);
 307 }
 308
 309 int
 310 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 311                unsigned int priority)
 312 {
 313         struct list_head        *e;
 314         struct lnet_remotenet   *rnet;
 315         struct lnet_remotenet   *rnet2;
 316         struct lnet_route               *route;
 317         struct lnet_ni          *ni;
 318         struct lnet_peer_ni     *lpni;
 319         int                     add_route;
 320         int                     rc;
 321
 322         CDEBUG(D_NET, "Add route: net %s hops %d priority %u gw %s\n",
 323                libcfs_net2str(net), hops, priority, libcfs_nid2str(gateway));
 324
 325         if (gateway == LNET_NID_ANY ||
 326             LNET_NETTYP(LNET_NIDNET(gateway)) == LOLND ||
 327             net == LNET_NIDNET(LNET_NID_ANY) ||
 328             LNET_NETTYP(net) == LOLND ||
 329             LNET_NIDNET(gateway) == net ||
 330             (hops != LNET_UNDEFINED_HOPS && (hops < 1 || hops > 255)))
 331                 return -EINVAL;
 332
 333         if (lnet_islocalnet(net))       /* it's a local network */
 334                 return -EEXIST;
 335
 336         /* Assume net, route, all new */
 337         LIBCFS_ALLOC(route, sizeof(*route));
 338         LIBCFS_ALLOC(rnet, sizeof(*rnet));
 339         if (route == NULL || rnet == NULL) {
 340                 CERROR("Out of memory creating route %s %d %s\n",
 341                        libcfs_net2str(net), hops, libcfs_nid2str(gateway));
 342                 if (route != NULL)
 343                         LIBCFS_FREE(route, sizeof(*route));
 344                 if (rnet != NULL)
 345                         LIBCFS_FREE(rnet, sizeof(*rnet));
 346                 return -ENOMEM;
 347         }
 348
 349         INIT_LIST_HEAD(&rnet->lrn_routes);
 350         rnet->lrn_net = net;
 351         route->lr_hops = hops;
 352         route->lr_net = net;
 353         route->lr_priority = priority;
 354
 355         lnet_net_lock(LNET_LOCK_EX);
 356
 357         lpni = lnet_nid2peerni_ex(gateway, LNET_LOCK_EX);
 358         if (IS_ERR(lpni)) {
 359                 lnet_net_unlock(LNET_LOCK_EX);
 360
 361                 LIBCFS_FREE(route, sizeof(*route));
 362                 LIBCFS_FREE(rnet, sizeof(*rnet));
 363
 364                 rc = PTR_ERR(lpni);
 365                 if (rc == -EHOSTUNREACH) /* gateway is not on a local net. */
 366                         return rc;       /* ignore the route entry */
 367                 CERROR("Error %d creating route %s %d %s\n", rc,
 368                         libcfs_net2str(net), hops,
 369                         libcfs_nid2str(gateway));
 370                 return rc;
 371         }
 372         route->lr_gateway = lpni;
 373         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 374
 375         rnet2 = lnet_find_rnet_locked(net);
 376         if (rnet2 == NULL) {
 377                 /* new network */
 378                 list_add_tail(&rnet->lrn_list, lnet_net2rnethash(net));
 379                 rnet2 = rnet;
 380         }
 381
 382         /* Search for a duplicate route (it's a NOOP if it is) */
 383         add_route = 1;
 384         list_for_each(e, &rnet2->lrn_routes) {
 385                 struct lnet_route *route2;
 386
 387                 route2 = list_entry(e, struct lnet_route, lr_list);
 388                 if (route2->lr_gateway == route->lr_gateway) {
 389                         add_route = 0;
 390                         break;
 391                 }
 392
 393                 /* our lookups must be true */
 394                 LASSERT(route2->lr_gateway->lpni_nid != gateway);
 395         }
 396
 397         if (add_route) {
 398                 lnet_peer_ni_addref_locked(route->lr_gateway); /* +1 for notify */
 399                 lnet_add_route_to_rnet(rnet2, route);
 400
 401                 ni = lnet_get_next_ni_locked(route->lr_gateway->lpni_net, NULL);
 402                 lnet_net_unlock(LNET_LOCK_EX);
 403
 404                 /* XXX Assume alive */
 405                 if (ni->ni_net->net_lnd->lnd_notify != NULL)
 406                         (ni->ni_net->net_lnd->lnd_notify)(ni, gateway, 1);
 407
 408                 lnet_net_lock(LNET_LOCK_EX);
 409         }
 410
 411         /* -1 for notify or !add_route */
 412         lnet_peer_ni_decref_locked(route->lr_gateway);
 413         lnet_net_unlock(LNET_LOCK_EX);
 414
 415         rc = 0;
 416
 417         if (!add_route) {
 418                 rc = -EEXIST;
 419                 LIBCFS_FREE(route, sizeof(*route));
 420         }
 421
 422         if (rnet != rnet2)
 423                 LIBCFS_FREE(rnet, sizeof(*rnet));
 424
 425         /* kick start the monitor thread to handle the added route */
 426         wake_up(&the_lnet.ln_mt_waitq);
 427
 428         return rc;
 429 }
 430
 431 int
 432 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 433 {
 434         struct lnet_peer_ni     *gateway;
 435         struct lnet_remotenet   *rnet;
 436         struct lnet_route               *route;
 437         struct list_head        *e1;
 438         struct list_head        *e2;
 439         int                     rc = -ENOENT;
 440         struct list_head        *rn_list;
 441         int                     idx = 0;
 442
 443         CDEBUG(D_NET, "Del route: net %s : gw %s\n",
 444                libcfs_net2str(net), libcfs_nid2str(gw_nid));
 445
 446         /* NB Caller may specify either all routes via the given gateway
 447          * or a specific route entry actual NIDs) */
 448
 449         lnet_net_lock(LNET_LOCK_EX);
 450         if (net == LNET_NIDNET(LNET_NID_ANY))
 451                 rn_list = &the_lnet.ln_remote_nets_hash[0];
 452         else
 453                 rn_list = lnet_net2rnethash(net);
 454
 455 again:
 456         list_for_each(e1, rn_list) {
 457                 rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 458
 459                 if (!(net == LNET_NIDNET(LNET_NID_ANY) ||
 460                         net == rnet->lrn_net))
 461                         continue;
 462
 463                 list_for_each(e2, &rnet->lrn_routes) {
 464                         route = list_entry(e2, struct lnet_route, lr_list);
 465
 466                         gateway = route->lr_gateway;
 467                         if (!(gw_nid == LNET_NID_ANY ||
 468                               gw_nid == gateway->lpni_nid))
 469                                 continue;
 470
 471                         list_del(&route->lr_list);
 472                         list_del(&route->lr_gwlist);
 473                         the_lnet.ln_remote_nets_version++;
 474
 475                         if (list_empty(&rnet->lrn_routes))
 476                                 list_del(&rnet->lrn_list);
 477                         else
 478                                 rnet = NULL;
 479
 480                         lnet_rtr_decref_locked(gateway);
 481                         lnet_peer_ni_decref_locked(gateway);
 482
 483                         lnet_net_unlock(LNET_LOCK_EX);
 484
 485                         LIBCFS_FREE(route, sizeof(*route));
 486
 487                         if (rnet != NULL)
 488                                 LIBCFS_FREE(rnet, sizeof(*rnet));
 489
 490                         rc = 0;
 491                         lnet_net_lock(LNET_LOCK_EX);
 492                         goto again;
 493                 }
 494         }
 495
 496         if (net == LNET_NIDNET(LNET_NID_ANY) &&
 497             ++idx < LNET_REMOTE_NETS_HASH_SIZE) {
 498                 rn_list = &the_lnet.ln_remote_nets_hash[idx];
 499                 goto again;
 500         }
 501         lnet_net_unlock(LNET_LOCK_EX);
 502
 503         return rc;
 504 }
 505
 506 void
 507 lnet_destroy_routes (void)
 508 {
 509         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 510 }
 511
 512 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 513 {
 514         struct lnet_rtrbufpool *rbp;
 515         int i, rc = -ENOENT, j;
 516
 517         if (the_lnet.ln_rtrpools == NULL)
 518                 return rc;
 519
 520
 521         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 522                 if (i != cpt)
 523                         continue;
 524
 525                 lnet_net_lock(i);
 526                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 527                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 528                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 529                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 530                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 531                 }
 532                 lnet_net_unlock(i);
 533                 rc = 0;
 534                 break;
 535         }
 536
 537         lnet_net_lock(LNET_LOCK_EX);
 538         pool_cfg->pl_routing = the_lnet.ln_routing;
 539         lnet_net_unlock(LNET_LOCK_EX);
 540
 541         return rc;
 542 }
 543
 544 int
 545 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 546                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 547 {
 548         struct list_head *e1;
 549         struct list_head *e2;
 550         struct lnet_remotenet *rnet;
 551         struct lnet_route        *route;
 552         int               cpt;
 553         int               i;
 554         struct list_head *rn_list;
 555
 556         cpt = lnet_net_lock_current();
 557
 558         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 559                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 560                 list_for_each(e1, rn_list) {
 561                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 562
 563                         list_for_each(e2, &rnet->lrn_routes) {
 564                                 route = list_entry(e2, struct lnet_route,
 565                                                    lr_list);
 566
 567                                 if (idx-- == 0) {
 568                                         *net      = rnet->lrn_net;
 569                                         *hops     = route->lr_hops;
 570                                         *priority = route->lr_priority;
 571                                         *gateway  = route->lr_gateway->lpni_nid;
 572                                         *alive    = lnet_is_route_alive(route);
 573                                         lnet_net_unlock(cpt);
 574                                         return 0;
 575                                 }
 576                         }
 577                 }
 578         }
 579
 580         lnet_net_unlock(cpt);
 581         return -ENOENT;
 582 }
 583
 584 void
 585 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
 586 {
 587         struct lnet_ni_status *stat;
 588         int nnis;
 589         int i;
 590
 591         __swab32s(&pbuf->pb_info.pi_magic);
 592         __swab32s(&pbuf->pb_info.pi_features);
 593         __swab32s(&pbuf->pb_info.pi_pid);
 594         __swab32s(&pbuf->pb_info.pi_nnis);
 595         nnis = pbuf->pb_info.pi_nnis;
 596         if (nnis > pbuf->pb_nnis)
 597                 nnis = pbuf->pb_nnis;
 598         for (i = 0; i < nnis; i++) {
 599                 stat = &pbuf->pb_info.pi_ni[i];
 600                 __swab64s(&stat->ns_nid);
 601                 __swab32s(&stat->ns_status);
 602         }
 603         return;
 604 }
 605
 606 /**
 607  * parse router-checker pinginfo, record number of down NIs for remote
 608  * networks on that router.
 609  */
 610 static void
 611 lnet_parse_rc_info(struct lnet_rc_data *rcd)
 612 {
 613         struct lnet_ping_buffer *pbuf = rcd->rcd_pingbuffer;
 614         struct lnet_peer_ni     *gw   = rcd->rcd_gateway;
 615         struct lnet_route               *rte;
 616         int                     nnis;
 617
 618         if (!gw->lpni_alive || !pbuf)
 619                 return;
 620
 621         /*
 622          * Protect gw->lpni_ping_feats. This can be set from
 623          * lnet_notify_locked with different locks being held
 624          */
 625         spin_lock(&gw->lpni_lock);
 626
 627         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
 628                 lnet_swap_pinginfo(pbuf);
 629
 630         /* NB always racing with network! */
 631         if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
 632                 CDEBUG(D_NET, "%s: Unexpected magic %08x\n",
 633                        libcfs_nid2str(gw->lpni_nid), pbuf->pb_info.pi_magic);
 634                 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 635                 goto out;
 636         }
 637
 638         gw->lpni_ping_feats = pbuf->pb_info.pi_features;
 639
 640         /* Without NI status info there's nothing more to do. */
 641         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) == 0)
 642                 goto out;
 643
 644         /* Determine the number of NIs for which there is data. */
 645         nnis = pbuf->pb_info.pi_nnis;
 646         if (pbuf->pb_nnis < nnis) {
 647                 if (rcd->rcd_nnis < nnis)
 648                         rcd->rcd_nnis = nnis;
 649                 nnis = pbuf->pb_nnis;
 650         }
 651
 652         list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 653                 int     down = 0;
 654                 int     up = 0;
 655                 int     i;
 656
 657                 /* If routing disabled then the route is down. */
 658                 if ((gw->lpni_ping_feats & LNET_PING_FEAT_RTE_DISABLED) != 0) {
 659                         rte->lr_downis = 1;
 660                         continue;
 661                 }
 662
 663                 for (i = 0; i < nnis; i++) {
 664                         struct lnet_ni_status *stat = &pbuf->pb_info.pi_ni[i];
 665                         lnet_nid_t       nid = stat->ns_nid;
 666
 667                         if (nid == LNET_NID_ANY) {
 668                                 CDEBUG(D_NET, "%s: unexpected LNET_NID_ANY\n",
 669                                        libcfs_nid2str(gw->lpni_nid));
 670                                 gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 671                                 goto out;
 672                         }
 673
 674                         if (LNET_NETTYP(LNET_NIDNET(nid)) == LOLND)
 675                                 continue;
 676
 677                         if (stat->ns_status == LNET_NI_STATUS_DOWN) {
 678                                 down++;
 679                                 continue;
 680                         }
 681
 682                         if (stat->ns_status == LNET_NI_STATUS_UP) {
 683                                 if (LNET_NIDNET(nid) == rte->lr_net) {
 684                                         up = 1;
 685                                         break;
 686                                 }
 687                                 continue;
 688                         }
 689
 690                         CDEBUG(D_NET, "%s: Unexpected status 0x%x\n",
 691                                libcfs_nid2str(gw->lpni_nid), stat->ns_status);
 692                         gw->lpni_ping_feats = LNET_PING_FEAT_INVAL;
 693                         goto out;
 694                 }
 695
 696                 if (up) { /* ignore downed NIs if NI for dest network is up */
 697                         rte->lr_downis = 0;
 698                         continue;
 699                 }
 700                 /* if @down is zero and this route is single-hop, it means
 701                  * we can't find NI for target network */
 702                 if (down == 0 && rte->lr_hops == 1)
 703                         down = 1;
 704
 705                 rte->lr_downis = down;
 706         }
 707 out:
 708         spin_unlock(&gw->lpni_lock);
 709 }
 710
 711 static void
 712 lnet_router_checker_event(struct lnet_event *event)
 713 {
 714         struct lnet_rc_data *rcd = event->md.user_ptr;
 715         struct lnet_peer_ni *lp;
 716
 717         LASSERT(rcd != NULL);
 718
 719         if (event->unlinked) {
 720                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 721                 return;
 722         }
 723
 724         LASSERT(event->type == LNET_EVENT_SEND ||
 725                 event->type == LNET_EVENT_REPLY);
 726
 727         lp = rcd->rcd_gateway;
 728         LASSERT(lp != NULL);
 729
 730          /* NB: it's called with holding lnet_res_lock, we have a few
 731           * places need to hold both locks at the same time, please take
 732           * care of lock ordering */
 733         lnet_net_lock(lp->lpni_cpt);
 734         if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
 735                 /* ignore if no longer a router or rcd is replaced */
 736                 goto out;
 737         }
 738
 739         if (event->type == LNET_EVENT_SEND) {
 740                 lp->lpni_ping_notsent = 0;
 741                 if (event->status == 0)
 742                         goto out;
 743         }
 744
 745         /* LNET_EVENT_REPLY */
 746         /* A successful REPLY means the router is up.  If _any_ comms
 747          * to the router fail I assume it's down (this will happen if
 748          * we ping alive routers to try to detect router death before
 749          * apps get burned). */
 750
 751         lnet_notify_locked(lp, 1, !event->status, ktime_get_seconds());
 752         /* The router checker will wake up very shortly and do the
 753          * actual notification.
 754          * XXX If 'lp' stops being a router before then, it will still
 755          * have the notification pending!!! */
 756
 757         if (avoid_asym_router_failure && event->status == 0)
 758                 lnet_parse_rc_info(rcd);
 759
 760  out:
 761         lnet_net_unlock(lp->lpni_cpt);
 762 }
 763
 764 static void
 765 lnet_wait_known_routerstate(void)
 766 {
 767         struct lnet_peer_ni *rtr;
 768         struct list_head *entry;
 769         int all_known;
 770
 771         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
 772
 773         for (;;) {
 774                 int cpt = lnet_net_lock_current();
 775
 776                 all_known = 1;
 777                 list_for_each(entry, &the_lnet.ln_routers) {
 778                         rtr = list_entry(entry, struct lnet_peer_ni,
 779                                          lpni_rtr_list);
 780
 781                         spin_lock(&rtr->lpni_lock);
 782
 783                         if (rtr->lpni_alive_count == 0) {
 784                                 all_known = 0;
 785                                 spin_unlock(&rtr->lpni_lock);
 786                                 break;
 787                         }
 788                         spin_unlock(&rtr->lpni_lock);
 789                 }
 790
 791                 lnet_net_unlock(cpt);
 792
 793                 if (all_known)
 794                         return;
 795
 796                 set_current_state(TASK_UNINTERRUPTIBLE);
 797                 schedule_timeout(cfs_time_seconds(1));
 798         }
 799 }
 800
 801 void
 802 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
 803 {
 804         struct lnet_route *rte;
 805
 806         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0) {
 807                 list_for_each_entry(rte, &gw->lpni_routes, lr_gwlist) {
 808                         if (rte->lr_net == net) {
 809                                 rte->lr_downis = 0;
 810                                 break;
 811                         }
 812                 }
 813         }
 814 }
 815
 816 static void
 817 lnet_update_ni_status_locked(void)
 818 {
 819         struct lnet_ni *ni = NULL;
 820         time64_t now;
 821         time64_t timeout;
 822
 823         LASSERT(the_lnet.ln_routing);
 824
 825         timeout = router_ping_timeout +
 826                   MAX(live_router_check_interval, dead_router_check_interval);
 827
 828         now = ktime_get_real_seconds();
 829         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 830                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
 831                         continue;
 832
 833                 if (now < ni->ni_last_alive + timeout)
 834                         continue;
 835
 836                 lnet_ni_lock(ni);
 837                 /* re-check with lock */
 838                 if (now < ni->ni_last_alive + timeout) {
 839                         lnet_ni_unlock(ni);
 840                         continue;
 841                 }
 842
 843                 LASSERT(ni->ni_status != NULL);
 844
 845                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 846                         CDEBUG(D_NET, "NI(%s:%lld) status changed to down\n",
 847                                libcfs_nid2str(ni->ni_nid), timeout);
 848                         /* NB: so far, this is the only place to set
 849                          * NI status to "down" */
 850                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 851                 }
 852                 lnet_ni_unlock(ni);
 853         }
 854 }
 855
 856 static void
 857 lnet_destroy_rc_data(struct lnet_rc_data *rcd)
 858 {
 859         LASSERT(list_empty(&rcd->rcd_list));
 860         /* detached from network */
 861         LASSERT(LNetMDHandleIsInvalid(rcd->rcd_mdh));
 862
 863         if (rcd->rcd_gateway != NULL) {
 864                 int cpt = rcd->rcd_gateway->lpni_cpt;
 865
 866                 lnet_net_lock(cpt);
 867                 lnet_peer_ni_decref_locked(rcd->rcd_gateway);
 868                 lnet_net_unlock(cpt);
 869         }
 870
 871         if (rcd->rcd_pingbuffer != NULL)
 872                 lnet_ping_buffer_decref(rcd->rcd_pingbuffer);
 873
 874         LIBCFS_FREE(rcd, sizeof(*rcd));
 875 }
 876
 877 static struct lnet_rc_data *
 878 lnet_update_rc_data_locked(struct lnet_peer_ni *gateway)
 879 {
 880         struct lnet_handle_md mdh;
 881         struct lnet_rc_data *rcd;
 882         struct lnet_ping_buffer *pbuf = NULL;
 883         int nnis = LNET_INTERFACES_MIN;
 884         int rc;
 885         int i;
 886
 887         rcd = gateway->lpni_rcd;
 888         if (rcd) {
 889                 nnis = rcd->rcd_nnis;
 890                 mdh = rcd->rcd_mdh;
 891                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 892                 pbuf = rcd->rcd_pingbuffer;
 893                 rcd->rcd_pingbuffer = NULL;
 894         } else {
 895                 LNetInvalidateMDHandle(&mdh);
 896         }
 897
 898         lnet_net_unlock(gateway->lpni_cpt);
 899
 900         if (rcd) {
 901                 LNetMDUnlink(mdh);
 902                 lnet_ping_buffer_decref(pbuf);
 903         } else {
 904                 LIBCFS_ALLOC(rcd, sizeof(*rcd));
 905                 if (rcd == NULL)
 906                         goto out;
 907
 908                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 909                 INIT_LIST_HEAD(&rcd->rcd_list);
 910                 rcd->rcd_nnis = nnis;
 911         }
 912
 913         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
 914         if (pbuf == NULL)
 915                 goto out;
 916
 917         for (i = 0; i < nnis; i++) {
 918                 pbuf->pb_info.pi_ni[i].ns_nid = LNET_NID_ANY;
 919                 pbuf->pb_info.pi_ni[i].ns_status = LNET_NI_STATUS_INVALID;
 920         }
 921         rcd->rcd_pingbuffer = pbuf;
 922
 923         LASSERT(!LNetEQHandleIsInvalid(the_lnet.ln_rc_eqh));
 924         rc = LNetMDBind((struct lnet_md){.start     = &pbuf->pb_info,
 925                                     .user_ptr  = rcd,
 926                                     .length    = LNET_PING_INFO_SIZE(nnis),
 927                                     .threshold = LNET_MD_THRESH_INF,
 928                                     .options   = LNET_MD_TRUNCATE,
 929                                     .eq_handle = the_lnet.ln_rc_eqh},
 930                         LNET_UNLINK,
 931                         &rcd->rcd_mdh);
 932         if (rc < 0) {
 933                 CERROR("Can't bind MD: %d\n", rc);
 934                 goto out_ping_buffer_decref;
 935         }
 936         LASSERT(rc == 0);
 937
 938         lnet_net_lock(gateway->lpni_cpt);
 939         /* Check if this is still a router. */
 940         if (!lnet_isrouter(gateway))
 941                 goto out_unlock;
 942         /* Check if someone else installed router data. */
 943         if (gateway->lpni_rcd && gateway->lpni_rcd != rcd)
 944                 goto out_unlock;
 945
 946         /* Install and/or update the router data. */
 947         if (!gateway->lpni_rcd) {
 948                 lnet_peer_ni_addref_locked(gateway);
 949                 rcd->rcd_gateway = gateway;
 950                 gateway->lpni_rcd = rcd;
 951         }
 952         gateway->lpni_ping_notsent = 0;
 953
 954         return rcd;
 955
 956 out_unlock:
 957         lnet_net_unlock(gateway->lpni_cpt);
 958         rc = LNetMDUnlink(mdh);
 959         LASSERT(rc == 0);
 960 out_ping_buffer_decref:
 961         lnet_ping_buffer_decref(pbuf);
 962 out:
 963         if (rcd && rcd != gateway->lpni_rcd)
 964                 lnet_destroy_rc_data(rcd);
 965         lnet_net_lock(gateway->lpni_cpt);
 966         return gateway->lpni_rcd;
 967 }
 968
 969 static int
 970 lnet_router_check_interval(struct lnet_peer_ni *rtr)
 971 {
 972         int secs;
 973
 974         secs = rtr->lpni_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(struct lnet_peer_ni *rtr)
 984 {
 985         struct lnet_rc_data *rcd = NULL;
 986         time64_t now = ktime_get_seconds();
 987         time64_t secs;
 988         struct lnet_ni *ni;
 989
 990         lnet_peer_ni_addref_locked(rtr);
 991
 992         if (rtr->lpni_ping_deadline != 0 && /* ping timed out? */
 993             now >  rtr->lpni_ping_deadline)
 994                 lnet_notify_locked(rtr, 1, 0, now);
 995
 996         /* Run any outstanding notifications */
 997         ni = lnet_get_next_ni_locked(rtr->lpni_net, NULL);
 998         lnet_ni_notify_locked(ni, rtr);
 999
1000         if (!lnet_isrouter(rtr) ||
1001             the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
1002                 /* router table changed or router checker is shutting down */
1003                 lnet_peer_ni_decref_locked(rtr);
1004                 return;
1005         }
1006
1007         rcd = rtr->lpni_rcd;
1008
1009         /*
1010          * The response to the router checker ping could've timed out and
1011          * the mdh might've been invalidated, so we need to update it
1012          * again.
1013          */
1014         if (!rcd || rcd->rcd_nnis > rcd->rcd_pingbuffer->pb_nnis ||
1015             LNetMDHandleIsInvalid(rcd->rcd_mdh))
1016                 rcd = lnet_update_rc_data_locked(rtr);
1017         if (rcd == NULL)
1018                 return;
1019
1020         secs = lnet_router_check_interval(rtr);
1021
1022         CDEBUG(D_NET,
1023                "rtr %s %lld: deadline %lld ping_notsent %d alive %d "
1024                "alive_count %d lpni_ping_timestamp %lld\n",
1025                libcfs_nid2str(rtr->lpni_nid), secs,
1026                rtr->lpni_ping_deadline, rtr->lpni_ping_notsent,
1027                rtr->lpni_alive, rtr->lpni_alive_count, rtr->lpni_ping_timestamp);
1028
1029         if (secs != 0 && !rtr->lpni_ping_notsent &&
1030             now > rtr->lpni_ping_timestamp + secs) {
1031                 int               rc;
1032                 struct lnet_process_id id;
1033                 struct lnet_handle_md mdh;
1034
1035                 id.nid = rtr->lpni_nid;
1036                 id.pid = LNET_PID_LUSTRE;
1037                 CDEBUG(D_NET, "Check: %s\n", libcfs_id2str(id));
1038
1039                 rtr->lpni_ping_notsent   = 1;
1040                 rtr->lpni_ping_timestamp = now;
1041
1042                 mdh = rcd->rcd_mdh;
1043
1044                 if (rtr->lpni_ping_deadline == 0) {
1045                         rtr->lpni_ping_deadline = ktime_get_seconds() +
1046                                                   router_ping_timeout;
1047                 }
1048
1049                 lnet_net_unlock(rtr->lpni_cpt);
1050
1051                 rc = LNetGet(LNET_NID_ANY, mdh, id, LNET_RESERVED_PORTAL,
1052                              LNET_PROTO_PING_MATCHBITS, 0, false);
1053
1054                 lnet_net_lock(rtr->lpni_cpt);
1055                 if (rc != 0)
1056                         rtr->lpni_ping_notsent = 0; /* no event pending */
1057         }
1058
1059         lnet_peer_ni_decref_locked(rtr);
1060         return;
1061 }
1062
1063 int lnet_router_pre_mt_start(void)
1064 {
1065         int rc;
1066
1067         if (check_routers_before_use &&
1068             dead_router_check_interval <= 0) {
1069                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
1070                                    " set if 'check_routers_before_use' is set"
1071                                    "\n");
1072                 return -EINVAL;
1073         }
1074
1075         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
1076         if (rc != 0) {
1077                 CERROR("Can't allocate EQ(0): %d\n", rc);
1078                 return -ENOMEM;
1079         }
1080
1081         return 0;
1082 }
1083
1084 void lnet_router_post_mt_start(void)
1085 {
1086         if (check_routers_before_use) {
1087                 /* Note that a helpful side-effect of pinging all known routers
1088                  * at startup is that it makes them drop stale connections they
1089                  * may have to a previous instance of me. */
1090                 lnet_wait_known_routerstate();
1091         }
1092 }
1093
1094 void
1095 lnet_router_cleanup(void)
1096 {
1097         int rc;
1098
1099         rc = LNetEQFree(the_lnet.ln_rc_eqh);
1100         LASSERT(rc == 0);
1101         return;
1102 }
1103
1104 void
1105 lnet_prune_rc_data(int wait_unlink)
1106 {
1107         struct lnet_rc_data *rcd;
1108         struct lnet_rc_data *tmp;
1109         struct lnet_peer_ni *lp;
1110         struct list_head head;
1111         int i = 2;
1112
1113         if (likely(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING &&
1114                    list_empty(&the_lnet.ln_rcd_deathrow) &&
1115                    list_empty(&the_lnet.ln_rcd_zombie)))
1116                 return;
1117
1118         INIT_LIST_HEAD(&head);
1119
1120         lnet_net_lock(LNET_LOCK_EX);
1121
1122         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING) {
1123                 /* router checker is stopping, prune all */
1124                 list_for_each_entry(lp, &the_lnet.ln_routers,
1125                                     lpni_rtr_list) {
1126                         if (lp->lpni_rcd == NULL)
1127                                 continue;
1128
1129                         LASSERT(list_empty(&lp->lpni_rcd->rcd_list));
1130                         list_add(&lp->lpni_rcd->rcd_list,
1131                                  &the_lnet.ln_rcd_deathrow);
1132                         lp->lpni_rcd = NULL;
1133                 }
1134         }
1135
1136         /* unlink all RCDs on deathrow list */
1137         list_splice_init(&the_lnet.ln_rcd_deathrow, &head);
1138
1139         if (!list_empty(&head)) {
1140                 lnet_net_unlock(LNET_LOCK_EX);
1141
1142                 list_for_each_entry(rcd, &head, rcd_list)
1143                         LNetMDUnlink(rcd->rcd_mdh);
1144
1145                 lnet_net_lock(LNET_LOCK_EX);
1146         }
1147
1148         list_splice_init(&head, &the_lnet.ln_rcd_zombie);
1149
1150         /* release all zombie RCDs */
1151         while (!list_empty(&the_lnet.ln_rcd_zombie)) {
1152                 list_for_each_entry_safe(rcd, tmp, &the_lnet.ln_rcd_zombie,
1153                                          rcd_list) {
1154                         if (LNetMDHandleIsInvalid(rcd->rcd_mdh))
1155                                 list_move(&rcd->rcd_list, &head);
1156                 }
1157
1158                 wait_unlink = wait_unlink &&
1159                               !list_empty(&the_lnet.ln_rcd_zombie);
1160
1161                 lnet_net_unlock(LNET_LOCK_EX);
1162
1163                 while (!list_empty(&head)) {
1164                         rcd = list_entry(head.next,
1165                                          struct lnet_rc_data, rcd_list);
1166                         list_del_init(&rcd->rcd_list);
1167                         lnet_destroy_rc_data(rcd);
1168                 }
1169
1170                 if (!wait_unlink)
1171                         return;
1172
1173                 i++;
1174                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
1175                        "Waiting for rc buffers to unlink\n");
1176                 set_current_state(TASK_UNINTERRUPTIBLE);
1177                 schedule_timeout(cfs_time_seconds(1) / 4);
1178
1179                 lnet_net_lock(LNET_LOCK_EX);
1180         }
1181
1182         lnet_net_unlock(LNET_LOCK_EX);
1183 }
1184
1185 /*
1186  * This function is called from the monitor thread to check if there are
1187  * any active routers that need to be checked.
1188  */
1189 inline bool
1190 lnet_router_checker_active(void)
1191 {
1192         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
1193                 return true;
1194
1195         /* Router Checker thread needs to run when routing is enabled in
1196          * order to call lnet_update_ni_status_locked() */
1197         if (the_lnet.ln_routing)
1198                 return true;
1199
1200         /* if there are routers that need to be cleaned up then do so */
1201         if (!list_empty(&the_lnet.ln_rcd_deathrow) ||
1202             !list_empty(&the_lnet.ln_rcd_zombie))
1203                 return true;
1204
1205         return !list_empty(&the_lnet.ln_routers) &&
1206                 (live_router_check_interval > 0 ||
1207                  dead_router_check_interval > 0);
1208 }
1209
1210 void
1211 lnet_check_routers(void)
1212 {
1213         struct lnet_peer_ni *rtr;
1214         struct list_head *entry;
1215         __u64   version;
1216         int     cpt;
1217         int     cpt2;
1218
1219         cpt = lnet_net_lock_current();
1220 rescan:
1221         version = the_lnet.ln_routers_version;
1222
1223         list_for_each(entry, &the_lnet.ln_routers) {
1224                 rtr = list_entry(entry, struct lnet_peer_ni,
1225                                         lpni_rtr_list);
1226
1227                 cpt2 = rtr->lpni_cpt;
1228                 if (cpt != cpt2) {
1229                         lnet_net_unlock(cpt);
1230                         cpt = cpt2;
1231                         lnet_net_lock(cpt);
1232                         /* the routers list has changed */
1233                         if (version != the_lnet.ln_routers_version)
1234                                 goto rescan;
1235                 }
1236
1237                 lnet_ping_router_locked(rtr);
1238
1239                 /* NB dropped lock */
1240                 if (version != the_lnet.ln_routers_version) {
1241                         /* the routers list has changed */
1242                         goto rescan;
1243                 }
1244         }
1245
1246         if (the_lnet.ln_routing)
1247                 lnet_update_ni_status_locked();
1248
1249         lnet_net_unlock(cpt);
1250
1251         lnet_prune_rc_data(0); /* don't wait for UNLINK */
1252 }
1253
1254 void
1255 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
1256 {
1257         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1258
1259         while (--npages >= 0)
1260                 __free_page(rb->rb_kiov[npages].kiov_page);
1261
1262         LIBCFS_FREE(rb, sz);
1263 }
1264
1265 static struct lnet_rtrbuf *
1266 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
1267 {
1268         int            npages = rbp->rbp_npages;
1269         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
1270         struct page   *page;
1271         struct lnet_rtrbuf *rb;
1272         int            i;
1273
1274         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
1275         if (rb == NULL)
1276                 return NULL;
1277
1278         rb->rb_pool = rbp;
1279
1280         for (i = 0; i < npages; i++) {
1281                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
1282                                           GFP_KERNEL | __GFP_ZERO);
1283                 if (page == NULL) {
1284                         while (--i >= 0)
1285                                 __free_page(rb->rb_kiov[i].kiov_page);
1286
1287                         LIBCFS_FREE(rb, sz);
1288                         return NULL;
1289                 }
1290
1291                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
1292                 rb->rb_kiov[i].kiov_offset = 0;
1293                 rb->rb_kiov[i].kiov_page = page;
1294         }
1295
1296         return rb;
1297 }
1298
1299 static void
1300 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
1301 {
1302         int npages = rbp->rbp_npages;
1303         struct lnet_rtrbuf *rb;
1304         struct list_head tmp;
1305
1306         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
1307                 return;
1308
1309         INIT_LIST_HEAD(&tmp);
1310
1311         lnet_net_lock(cpt);
1312         list_splice_init(&rbp->rbp_msgs, &tmp);
1313         lnet_drop_routed_msgs_locked(&tmp, cpt);
1314         list_splice_init(&rbp->rbp_bufs, &tmp);
1315         rbp->rbp_req_nbuffers = 0;
1316         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
1317         rbp->rbp_mincredits = 0;
1318         lnet_net_unlock(cpt);
1319
1320         /* Free buffers on the free list. */
1321         while (!list_empty(&tmp)) {
1322                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
1323                 list_del(&rb->rb_list);
1324                 lnet_destroy_rtrbuf(rb, npages);
1325         }
1326 }
1327
1328 static int
1329 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
1330 {
1331         struct list_head rb_list;
1332         struct lnet_rtrbuf *rb;
1333         int             num_rb;
1334         int             num_buffers = 0;
1335         int             old_req_nbufs;
1336         int             npages = rbp->rbp_npages;
1337
1338         lnet_net_lock(cpt);
1339         /* If we are called for less buffers than already in the pool, we
1340          * just lower the req_nbuffers number and excess buffers will be
1341          * thrown away as they are returned to the free list.  Credits
1342          * then get adjusted as well.
1343          * If we already have enough buffers allocated to serve the
1344          * increase requested, then we can treat that the same way as we
1345          * do the decrease. */
1346         num_rb = nbufs - rbp->rbp_nbuffers;
1347         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
1348                 rbp->rbp_req_nbuffers = nbufs;
1349                 lnet_net_unlock(cpt);
1350                 return 0;
1351         }
1352         /* store the older value of rbp_req_nbuffers and then set it to
1353          * the new request to prevent lnet_return_rx_credits_locked() from
1354          * freeing buffers that we need to keep around */
1355         old_req_nbufs = rbp->rbp_req_nbuffers;
1356         rbp->rbp_req_nbuffers = nbufs;
1357         lnet_net_unlock(cpt);
1358
1359         INIT_LIST_HEAD(&rb_list);
1360
1361         /* allocate the buffers on a local list first.  If all buffers are
1362          * allocated successfully then join this list to the rbp buffer
1363          * list.  If not then free all allocated buffers. */
1364         while (num_rb-- > 0) {
1365                 rb = lnet_new_rtrbuf(rbp, cpt);
1366                 if (rb == NULL) {
1367                         CERROR("Failed to allocate %d route bufs of %d pages\n",
1368                                nbufs, npages);
1369
1370                         lnet_net_lock(cpt);
1371                         rbp->rbp_req_nbuffers = old_req_nbufs;
1372                         lnet_net_unlock(cpt);
1373
1374                         goto failed;
1375                 }
1376
1377                 list_add(&rb->rb_list, &rb_list);
1378                 num_buffers++;
1379         }
1380
1381         lnet_net_lock(cpt);
1382
1383         list_splice_tail(&rb_list, &rbp->rbp_bufs);
1384         rbp->rbp_nbuffers += num_buffers;
1385         rbp->rbp_credits += num_buffers;
1386         rbp->rbp_mincredits = rbp->rbp_credits;
1387         /* We need to schedule blocked msg using the newly
1388          * added buffers. */
1389         while (!list_empty(&rbp->rbp_bufs) &&
1390                !list_empty(&rbp->rbp_msgs))
1391                 lnet_schedule_blocked_locked(rbp);
1392
1393         lnet_net_unlock(cpt);
1394
1395         return 0;
1396
1397 failed:
1398         while (!list_empty(&rb_list)) {
1399                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
1400                 list_del(&rb->rb_list);
1401                 lnet_destroy_rtrbuf(rb, npages);
1402         }
1403
1404         return -ENOMEM;
1405 }
1406
1407 static void
1408 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
1409 {
1410         INIT_LIST_HEAD(&rbp->rbp_msgs);
1411         INIT_LIST_HEAD(&rbp->rbp_bufs);
1412
1413         rbp->rbp_npages = npages;
1414         rbp->rbp_credits = 0;
1415         rbp->rbp_mincredits = 0;
1416 }
1417
1418 void
1419 lnet_rtrpools_free(int keep_pools)
1420 {
1421         struct lnet_rtrbufpool *rtrp;
1422         int               i;
1423
1424         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
1425                 return;
1426
1427         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1428                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
1429                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
1430                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
1431         }
1432
1433         if (!keep_pools) {
1434                 cfs_percpt_free(the_lnet.ln_rtrpools);
1435                 the_lnet.ln_rtrpools = NULL;
1436         }
1437 }
1438
1439 static int
1440 lnet_nrb_tiny_calculate(void)
1441 {
1442         int     nrbs = LNET_NRB_TINY;
1443
1444         if (tiny_router_buffers < 0) {
1445                 LCONSOLE_ERROR_MSG(0x10c,
1446                                    "tiny_router_buffers=%d invalid when "
1447                                    "routing enabled\n", tiny_router_buffers);
1448                 return -EINVAL;
1449         }
1450
1451         if (tiny_router_buffers > 0)
1452                 nrbs = tiny_router_buffers;
1453
1454         nrbs /= LNET_CPT_NUMBER;
1455         return max(nrbs, LNET_NRB_TINY_MIN);
1456 }
1457
1458 static int
1459 lnet_nrb_small_calculate(void)
1460 {
1461         int     nrbs = LNET_NRB_SMALL;
1462
1463         if (small_router_buffers < 0) {
1464                 LCONSOLE_ERROR_MSG(0x10c,
1465                                    "small_router_buffers=%d invalid when "
1466                                    "routing enabled\n", small_router_buffers);
1467                 return -EINVAL;
1468         }
1469
1470         if (small_router_buffers > 0)
1471                 nrbs = small_router_buffers;
1472
1473         nrbs /= LNET_CPT_NUMBER;
1474         return max(nrbs, LNET_NRB_SMALL_MIN);
1475 }
1476
1477 static int
1478 lnet_nrb_large_calculate(void)
1479 {
1480         int     nrbs = LNET_NRB_LARGE;
1481
1482         if (large_router_buffers < 0) {
1483                 LCONSOLE_ERROR_MSG(0x10c,
1484                                    "large_router_buffers=%d invalid when "
1485                                    "routing enabled\n", large_router_buffers);
1486                 return -EINVAL;
1487         }
1488
1489         if (large_router_buffers > 0)
1490                 nrbs = large_router_buffers;
1491
1492         nrbs /= LNET_CPT_NUMBER;
1493         return max(nrbs, LNET_NRB_LARGE_MIN);
1494 }
1495
1496 int
1497 lnet_rtrpools_alloc(int im_a_router)
1498 {
1499         struct lnet_rtrbufpool *rtrp;
1500         int     nrb_tiny;
1501         int     nrb_small;
1502         int     nrb_large;
1503         int     rc;
1504         int     i;
1505
1506         if (!strcmp(forwarding, "")) {
1507                 /* not set either way */
1508                 if (!im_a_router)
1509                         return 0;
1510         } else if (!strcmp(forwarding, "disabled")) {
1511                 /* explicitly disabled */
1512                 return 0;
1513         } else if (!strcmp(forwarding, "enabled")) {
1514                 /* explicitly enabled */
1515         } else {
1516                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
1517                                    "'enabled' or 'disabled'\n");
1518                 return -EINVAL;
1519         }
1520
1521         nrb_tiny = lnet_nrb_tiny_calculate();
1522         if (nrb_tiny < 0)
1523                 return -EINVAL;
1524
1525         nrb_small = lnet_nrb_small_calculate();
1526         if (nrb_small < 0)
1527                 return -EINVAL;
1528
1529         nrb_large = lnet_nrb_large_calculate();
1530         if (nrb_large < 0)
1531                 return -EINVAL;
1532
1533         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
1534                                                 LNET_NRBPOOLS *
1535                                                 sizeof(struct lnet_rtrbufpool));
1536         if (the_lnet.ln_rtrpools == NULL) {
1537                 LCONSOLE_ERROR_MSG(0x10c,
1538                                    "Failed to initialize router buffe pool\n");
1539                 return -ENOMEM;
1540         }
1541
1542         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1543                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
1544                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1545                                               nrb_tiny, i);
1546                 if (rc != 0)
1547                         goto failed;
1548
1549                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
1550                                   LNET_NRB_SMALL_PAGES);
1551                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1552                                               nrb_small, i);
1553                 if (rc != 0)
1554                         goto failed;
1555
1556                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
1557                                   LNET_NRB_LARGE_PAGES);
1558                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1559                                               nrb_large, i);
1560                 if (rc != 0)
1561                         goto failed;
1562         }
1563
1564         lnet_net_lock(LNET_LOCK_EX);
1565         the_lnet.ln_routing = 1;
1566         lnet_net_unlock(LNET_LOCK_EX);
1567         return 0;
1568
1569  failed:
1570         lnet_rtrpools_free(0);
1571         return rc;
1572 }
1573
1574 static int
1575 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
1576 {
1577         int nrb = 0;
1578         int rc = 0;
1579         int i;
1580         struct lnet_rtrbufpool *rtrp;
1581
1582         /* If the provided values for each buffer pool are different than the
1583          * configured values, we need to take action. */
1584         if (tiny >= 0) {
1585                 tiny_router_buffers = tiny;
1586                 nrb = lnet_nrb_tiny_calculate();
1587                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1588                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
1589                                                       nrb, i);
1590                         if (rc != 0)
1591                                 return rc;
1592                 }
1593         }
1594         if (small >= 0) {
1595                 small_router_buffers = small;
1596                 nrb = lnet_nrb_small_calculate();
1597                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1598                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
1599                                                       nrb, i);
1600                         if (rc != 0)
1601                                 return rc;
1602                 }
1603         }
1604         if (large >= 0) {
1605                 large_router_buffers = large;
1606                 nrb = lnet_nrb_large_calculate();
1607                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
1608                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
1609                                                       nrb, i);
1610                         if (rc != 0)
1611                                 return rc;
1612                 }
1613         }
1614
1615         return 0;
1616 }
1617
1618 int
1619 lnet_rtrpools_adjust(int tiny, int small, int large)
1620 {
1621         /* this function doesn't revert the changes if adding new buffers
1622          * failed.  It's up to the user space caller to revert the
1623          * changes. */
1624
1625         if (!the_lnet.ln_routing)
1626                 return 0;
1627
1628         return lnet_rtrpools_adjust_helper(tiny, small, large);
1629 }
1630
1631 int
1632 lnet_rtrpools_enable(void)
1633 {
1634         int rc = 0;
1635
1636         if (the_lnet.ln_routing)
1637                 return 0;
1638
1639         if (the_lnet.ln_rtrpools == NULL)
1640                 /* If routing is turned off, and we have never
1641                  * initialized the pools before, just call the
1642                  * standard buffer pool allocation routine as
1643                  * if we are just configuring this for the first
1644                  * time. */
1645                 rc = lnet_rtrpools_alloc(1);
1646         else
1647                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
1648         if (rc != 0)
1649                 return rc;
1650
1651         lnet_net_lock(LNET_LOCK_EX);
1652         the_lnet.ln_routing = 1;
1653
1654         the_lnet.ln_ping_target->pb_info.pi_features &=
1655                 ~LNET_PING_FEAT_RTE_DISABLED;
1656         lnet_net_unlock(LNET_LOCK_EX);
1657
1658         return rc;
1659 }
1660
1661 void
1662 lnet_rtrpools_disable(void)
1663 {
1664         if (!the_lnet.ln_routing)
1665                 return;
1666
1667         lnet_net_lock(LNET_LOCK_EX);
1668         the_lnet.ln_routing = 0;
1669         the_lnet.ln_ping_target->pb_info.pi_features |=
1670                 LNET_PING_FEAT_RTE_DISABLED;
1671
1672         tiny_router_buffers = 0;
1673         small_router_buffers = 0;
1674         large_router_buffers = 0;
1675         lnet_net_unlock(LNET_LOCK_EX);
1676         lnet_rtrpools_free(1);
1677 }
1678
1679 int
1680 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, time64_t when)
1681 {
1682         struct lnet_peer_ni *lp = NULL;
1683         time64_t now = ktime_get_seconds();
1684         int cpt = lnet_cpt_of_nid(nid, ni);
1685
1686         LASSERT (!in_interrupt ());
1687
1688         CDEBUG (D_NET, "%s notifying %s: %s\n",
1689                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1690                 libcfs_nid2str(nid),
1691                 alive ? "up" : "down");
1692
1693         if (ni != NULL &&
1694             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1695                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1696                       libcfs_nid2str(nid), alive ? "birth" : "death",
1697                       libcfs_nid2str(ni->ni_nid));
1698                 return -EINVAL;
1699         }
1700
1701         /* can't do predictions... */
1702         if (when > now) {
1703                 CWARN("Ignoring prediction from %s of %s %s "
1704                       "%lld seconds in the future\n",
1705                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1706                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1707                 return -EINVAL;
1708         }
1709
1710         if (ni != NULL && !alive &&             /* LND telling me she's down */
1711             !auto_down) {                       /* auto-down disabled */
1712                 CDEBUG(D_NET, "Auto-down disabled\n");
1713                 return 0;
1714         }
1715
1716         lnet_net_lock(cpt);
1717
1718         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1719                 lnet_net_unlock(cpt);
1720                 return -ESHUTDOWN;
1721         }
1722
1723         lp = lnet_find_peer_ni_locked(nid);
1724         if (lp == NULL) {
1725                 /* nid not found */
1726                 lnet_net_unlock(cpt);
1727                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1728                 return 0;
1729         }
1730
1731         /*
1732          * It is possible for this function to be called for the same peer
1733          * but with different NIs. We want to synchronize the notification
1734          * between the different calls. So we will use the lpni_cpt to
1735          * grab the net lock.
1736          */
1737         if (lp->lpni_cpt != cpt) {
1738                 lnet_net_unlock(cpt);
1739                 cpt = lp->lpni_cpt;
1740                 lnet_net_lock(cpt);
1741         }
1742
1743         /* We can't fully trust LND on reporting exact peer last_alive
1744          * if he notifies us about dead peer. For example ksocklnd can
1745          * call us with when == _time_when_the_node_was_booted_ if
1746          * no connections were successfully established */
1747         if (ni != NULL && !alive && when < lp->lpni_last_alive)
1748                 when = lp->lpni_last_alive;
1749
1750         lnet_notify_locked(lp, ni == NULL, alive, when);
1751
1752         if (ni != NULL)
1753                 lnet_ni_notify_locked(ni, lp);
1754
1755         lnet_peer_ni_decref_locked(lp);
1756
1757         lnet_net_unlock(cpt);
1758         return 0;
1759 }
1760 EXPORT_SYMBOL(lnet_notify);