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
 118         if (lp->lpni_alive_count != 0 &&          /* got old news */
 119             (!lp->lpni_alive) == (!alive)) {      /* new date for old news */
 120                 spin_unlock(&lp->lpni_lock);
 121                 CDEBUG(D_NET, "Old news\n");
 122                 return;
 123         }
 124
 125         /* Flag that notification is outstanding */
 126
 127         lp->lpni_alive_count++;
 128         lp->lpni_alive = (alive) ? 1 : 0;
 129         lp->lpni_notify = 1;
 130         lp->lpni_notifylnd = notifylnd;
 131         if (lp->lpni_alive)
 132                 lp->lpni_ping_feats = LNET_PING_FEAT_INVAL; /* reset */
 133
 134         spin_unlock(&lp->lpni_lock);
 135
 136         CDEBUG(D_NET, "set %s %d\n", libcfs_nid2str(lp->lpni_nid), alive);
 137 }
 138
 139 /*
 140  * This function will always be called with lp->lpni_cpt lock held.
 141  */
 142 static void
 143 lnet_ni_notify_locked(struct lnet_ni *ni, struct lnet_peer_ni *lp)
 144 {
 145         int alive;
 146         int notifylnd;
 147
 148         /* Notify only in 1 thread at any time to ensure ordered notification.
 149          * NB individual events can be missed; the only guarantee is that you
 150          * always get the most recent news */
 151
 152         spin_lock(&lp->lpni_lock);
 153
 154         if (lp->lpni_notifying || ni == NULL) {
 155                 spin_unlock(&lp->lpni_lock);
 156                 return;
 157         }
 158
 159         lp->lpni_notifying = 1;
 160
 161         /*
 162          * lp->lpni_notify needs to be protected because it can be set in
 163          * lnet_notify_locked().
 164          */
 165         while (lp->lpni_notify) {
 166                 alive     = lp->lpni_alive;
 167                 notifylnd = lp->lpni_notifylnd;
 168
 169                 lp->lpni_notifylnd = 0;
 170                 lp->lpni_notify    = 0;
 171
 172                 if (notifylnd && ni->ni_net->net_lnd->lnd_notify != NULL) {
 173                         spin_unlock(&lp->lpni_lock);
 174                         lnet_net_unlock(lp->lpni_cpt);
 175
 176                         /* A new notification could happen now; I'll handle it
 177                          * when control returns to me */
 178
 179                         (ni->ni_net->net_lnd->lnd_notify)(ni, lp->lpni_nid,
 180                                                           alive);
 181
 182                         lnet_net_lock(lp->lpni_cpt);
 183                         spin_lock(&lp->lpni_lock);
 184                 }
 185         }
 186
 187         lp->lpni_notifying = 0;
 188         spin_unlock(&lp->lpni_lock);
 189 }
 190
 191 struct lnet_remotenet *
 192 lnet_find_rnet_locked(__u32 net)
 193 {
 194         struct lnet_remotenet *rnet;
 195         struct list_head *tmp;
 196         struct list_head *rn_list;
 197
 198         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
 199
 200         rn_list = lnet_net2rnethash(net);
 201         list_for_each(tmp, rn_list) {
 202                 rnet = list_entry(tmp, struct lnet_remotenet, lrn_list);
 203
 204                 if (rnet->lrn_net == net)
 205                         return rnet;
 206         }
 207         return NULL;
 208 }
 209
 210 int
 211 lnet_add_route(__u32 net, __u32 hops, lnet_nid_t gateway,
 212                unsigned int priority)
 213 {
 214         net = net;
 215         hops = hops;
 216         gateway = gateway;
 217         priority = priority;
 218         return -EINVAL;
 219 }
 220
 221 /* TODO: reimplement lnet_check_routes() */
 222 int
 223 lnet_del_route(__u32 net, lnet_nid_t gw_nid)
 224 {
 225         net = net;
 226         gw_nid = gw_nid;
 227         return -EINVAL;
 228 }
 229
 230 void
 231 lnet_destroy_routes (void)
 232 {
 233         lnet_del_route(LNET_NIDNET(LNET_NID_ANY), LNET_NID_ANY);
 234 }
 235
 236 int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 237 {
 238         struct lnet_rtrbufpool *rbp;
 239         int i, rc = -ENOENT, j;
 240
 241         if (the_lnet.ln_rtrpools == NULL)
 242                 return rc;
 243
 244
 245         cfs_percpt_for_each(rbp, i, the_lnet.ln_rtrpools) {
 246                 if (i != cpt)
 247                         continue;
 248
 249                 lnet_net_lock(i);
 250                 for (j = 0; j < LNET_NRBPOOLS; j++) {
 251                         pool_cfg->pl_pools[j].pl_npages = rbp[j].rbp_npages;
 252                         pool_cfg->pl_pools[j].pl_nbuffers = rbp[j].rbp_nbuffers;
 253                         pool_cfg->pl_pools[j].pl_credits = rbp[j].rbp_credits;
 254                         pool_cfg->pl_pools[j].pl_mincredits = rbp[j].rbp_mincredits;
 255                 }
 256                 lnet_net_unlock(i);
 257                 rc = 0;
 258                 break;
 259         }
 260
 261         lnet_net_lock(LNET_LOCK_EX);
 262         pool_cfg->pl_routing = the_lnet.ln_routing;
 263         lnet_net_unlock(LNET_LOCK_EX);
 264
 265         return rc;
 266 }
 267
 268 int
 269 lnet_get_route(int idx, __u32 *net, __u32 *hops,
 270                lnet_nid_t *gateway, __u32 *alive, __u32 *priority)
 271 {
 272         struct list_head *e1;
 273         struct list_head *e2;
 274         struct lnet_remotenet *rnet;
 275         struct lnet_route        *route;
 276         int               cpt;
 277         int               i;
 278         struct list_head *rn_list;
 279
 280         cpt = lnet_net_lock_current();
 281
 282         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++) {
 283                 rn_list = &the_lnet.ln_remote_nets_hash[i];
 284                 list_for_each(e1, rn_list) {
 285                         rnet = list_entry(e1, struct lnet_remotenet, lrn_list);
 286
 287                         list_for_each(e2, &rnet->lrn_routes) {
 288                                 route = list_entry(e2, struct lnet_route,
 289                                                    lr_list);
 290
 291                                 if (idx-- == 0) {
 292                                         *net      = rnet->lrn_net;
 293                                         *hops     = route->lr_hops;
 294                                         *priority = route->lr_priority;
 295                                         *gateway  = route->lr_gateway->lp_primary_nid;
 296                                         *alive    = lnet_is_route_alive(route);
 297                                         lnet_net_unlock(cpt);
 298                                         return 0;
 299                                 }
 300                         }
 301                 }
 302         }
 303
 304         lnet_net_unlock(cpt);
 305         return -ENOENT;
 306 }
 307
 308 void
 309 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
 310 {
 311         struct lnet_ni_status *stat;
 312         int nnis;
 313         int i;
 314
 315         __swab32s(&pbuf->pb_info.pi_magic);
 316         __swab32s(&pbuf->pb_info.pi_features);
 317         __swab32s(&pbuf->pb_info.pi_pid);
 318         __swab32s(&pbuf->pb_info.pi_nnis);
 319         nnis = pbuf->pb_info.pi_nnis;
 320         if (nnis > pbuf->pb_nnis)
 321                 nnis = pbuf->pb_nnis;
 322         for (i = 0; i < nnis; i++) {
 323                 stat = &pbuf->pb_info.pi_ni[i];
 324                 __swab64s(&stat->ns_nid);
 325                 __swab32s(&stat->ns_status);
 326         }
 327         return;
 328 }
 329
 330 /**
 331  * TODO: re-implement
 332  */
 333 static void
 334 lnet_parse_rc_info(struct lnet_rc_data *rcd)
 335 {
 336         rcd = rcd;
 337 }
 338
 339 static void
 340 lnet_router_checker_event(struct lnet_event *event)
 341 {
 342         struct lnet_rc_data *rcd = event->md.user_ptr;
 343         struct lnet_peer_ni *lp;
 344
 345         LASSERT(rcd != NULL);
 346
 347         if (event->unlinked) {
 348                 LNetInvalidateMDHandle(&rcd->rcd_mdh);
 349                 return;
 350         }
 351
 352         LASSERT(event->type == LNET_EVENT_SEND ||
 353                 event->type == LNET_EVENT_REPLY);
 354
 355         lp = rcd->rcd_gateway;
 356         LASSERT(lp != NULL);
 357
 358          /* NB: it's called with holding lnet_res_lock, we have a few
 359           * places need to hold both locks at the same time, please take
 360           * care of lock ordering */
 361         lnet_net_lock(lp->lpni_cpt);
 362         if (!lnet_isrouter(lp) || lp->lpni_rcd != rcd) {
 363                 /* ignore if no longer a router or rcd is replaced */
 364                 goto out;
 365         }
 366
 367         if (event->type == LNET_EVENT_SEND) {
 368                 if (event->status == 0)
 369                         goto out;
 370         }
 371
 372         /* LNET_EVENT_REPLY */
 373         /* A successful REPLY means the router is up.  If _any_ comms
 374          * to the router fail I assume it's down (this will happen if
 375          * we ping alive routers to try to detect router death before
 376          * apps get burned). */
 377
 378         lnet_notify_locked(lp, 1, !event->status, ktime_get_seconds());
 379         /* The router checker will wake up very shortly and do the
 380          * actual notification.
 381          * XXX If 'lp' stops being a router before then, it will still
 382          * have the notification pending!!! */
 383
 384         if (avoid_asym_router_failure && event->status == 0)
 385                 lnet_parse_rc_info(rcd);
 386
 387  out:
 388         lnet_net_unlock(lp->lpni_cpt);
 389 }
 390
 391 static void
 392 lnet_wait_known_routerstate(void)
 393 {
 394         struct lnet_peer *rtr;
 395         struct list_head *entry;
 396         int all_known;
 397
 398         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
 399
 400         for (;;) {
 401                 int cpt = lnet_net_lock_current();
 402
 403                 all_known = 1;
 404                 list_for_each(entry, &the_lnet.ln_routers) {
 405                         rtr = list_entry(entry, struct lnet_peer,
 406                                          lp_rtr_list);
 407
 408                         spin_lock(&rtr->lp_lock);
 409
 410                         if ((rtr->lp_state & LNET_PEER_DISCOVERED) == 0) {
 411                                 all_known = 0;
 412                                 spin_unlock(&rtr->lp_lock);
 413                                 break;
 414                         }
 415                         spin_unlock(&rtr->lp_lock);
 416                 }
 417
 418                 lnet_net_unlock(cpt);
 419
 420                 if (all_known)
 421                         return;
 422
 423                 set_current_state(TASK_UNINTERRUPTIBLE);
 424                 schedule_timeout(cfs_time_seconds(1));
 425         }
 426 }
 427
 428 /* TODO: reimplement */
 429 void
 430 lnet_router_ni_update_locked(struct lnet_peer_ni *gw, __u32 net)
 431 {
 432         struct lnet_route *rte;
 433         struct lnet_peer *lp;
 434
 435         if ((gw->lpni_ping_feats & LNET_PING_FEAT_NI_STATUS) != 0)
 436                 lp = gw->lpni_peer_net->lpn_peer;
 437         else
 438                 return;
 439
 440         list_for_each_entry(rte, &lp->lp_routes, lr_gwlist) {
 441                 if (rte->lr_net == net) {
 442                         rte->lr_downis = 0;
 443                         break;
 444                 }
 445         }
 446 }
 447
 448 static void
 449 lnet_update_ni_status_locked(void)
 450 {
 451         struct lnet_ni *ni = NULL;
 452         time64_t now;
 453         time64_t timeout;
 454
 455         LASSERT(the_lnet.ln_routing);
 456
 457         timeout = router_ping_timeout +
 458                   MAX(live_router_check_interval, dead_router_check_interval);
 459
 460         now = ktime_get_real_seconds();
 461         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
 462                 if (ni->ni_net->net_lnd->lnd_type == LOLND)
 463                         continue;
 464
 465                 if (now < ni->ni_last_alive + timeout)
 466                         continue;
 467
 468                 lnet_ni_lock(ni);
 469                 /* re-check with lock */
 470                 if (now < ni->ni_last_alive + timeout) {
 471                         lnet_ni_unlock(ni);
 472                         continue;
 473                 }
 474
 475                 LASSERT(ni->ni_status != NULL);
 476
 477                 if (ni->ni_status->ns_status != LNET_NI_STATUS_DOWN) {
 478                         CDEBUG(D_NET, "NI(%s:%lld) status changed to down\n",
 479                                libcfs_nid2str(ni->ni_nid), timeout);
 480                         /* NB: so far, this is the only place to set
 481                          * NI status to "down" */
 482                         ni->ni_status->ns_status = LNET_NI_STATUS_DOWN;
 483                 }
 484                 lnet_ni_unlock(ni);
 485         }
 486 }
 487
 488 int lnet_router_pre_mt_start(void)
 489 {
 490         int rc;
 491
 492         if (check_routers_before_use &&
 493             dead_router_check_interval <= 0) {
 494                 LCONSOLE_ERROR_MSG(0x10a, "'dead_router_check_interval' must be"
 495                                    " set if 'check_routers_before_use' is set"
 496                                    "\n");
 497                 return -EINVAL;
 498         }
 499
 500         rc = LNetEQAlloc(0, lnet_router_checker_event, &the_lnet.ln_rc_eqh);
 501         if (rc != 0) {
 502                 CERROR("Can't allocate EQ(0): %d\n", rc);
 503                 return -ENOMEM;
 504         }
 505
 506         return 0;
 507 }
 508
 509 void lnet_router_post_mt_start(void)
 510 {
 511         if (check_routers_before_use) {
 512                 /* Note that a helpful side-effect of pinging all known routers
 513                  * at startup is that it makes them drop stale connections they
 514                  * may have to a previous instance of me. */
 515                 lnet_wait_known_routerstate();
 516         }
 517 }
 518
 519 void
 520 lnet_router_cleanup(void)
 521 {
 522         int rc;
 523
 524         rc = LNetEQFree(the_lnet.ln_rc_eqh);
 525         LASSERT(rc == 0);
 526         return;
 527 }
 528
 529 void
 530 lnet_prune_rc_data(int wait_unlink)
 531 {
 532         wait_unlink = wait_unlink;
 533 }
 534
 535 /*
 536  * This function is called from the monitor thread to check if there are
 537  * any active routers that need to be checked.
 538  */
 539 inline bool
 540 lnet_router_checker_active(void)
 541 {
 542         if (the_lnet.ln_mt_state != LNET_MT_STATE_RUNNING)
 543                 return true;
 544
 545         /* Router Checker thread needs to run when routing is enabled in
 546          * order to call lnet_update_ni_status_locked() */
 547         if (the_lnet.ln_routing)
 548                 return true;
 549
 550         /* if there are routers that need to be cleaned up then do so */
 551         if (!list_empty(&the_lnet.ln_rcd_deathrow) ||
 552             !list_empty(&the_lnet.ln_rcd_zombie))
 553                 return true;
 554
 555         return !list_empty(&the_lnet.ln_routers) &&
 556                 (live_router_check_interval > 0 ||
 557                  dead_router_check_interval > 0);
 558 }
 559
 560 void
 561 lnet_check_routers(void)
 562 {
 563         struct lnet_peer *rtr;
 564         struct list_head *entry;
 565         __u64   version;
 566         int     cpt;
 567
 568         cpt = lnet_net_lock_current();
 569 rescan:
 570         version = the_lnet.ln_routers_version;
 571
 572         list_for_each(entry, &the_lnet.ln_routers) {
 573                 rtr = list_entry(entry, struct lnet_peer,
 574                                  lp_rtr_list);
 575
 576                 /* TODO use discovery to determine if router is alive */
 577
 578                 /* NB dropped lock */
 579                 if (version != the_lnet.ln_routers_version) {
 580                         /* the routers list has changed */
 581                         goto rescan;
 582                 }
 583         }
 584
 585         if (the_lnet.ln_routing)
 586                 lnet_update_ni_status_locked();
 587
 588         lnet_net_unlock(cpt);
 589
 590         lnet_prune_rc_data(0); /* don't wait for UNLINK */
 591 }
 592
 593 void
 594 lnet_destroy_rtrbuf(struct lnet_rtrbuf *rb, int npages)
 595 {
 596         int sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
 597
 598         while (--npages >= 0)
 599                 __free_page(rb->rb_kiov[npages].kiov_page);
 600
 601         LIBCFS_FREE(rb, sz);
 602 }
 603
 604 static struct lnet_rtrbuf *
 605 lnet_new_rtrbuf(struct lnet_rtrbufpool *rbp, int cpt)
 606 {
 607         int            npages = rbp->rbp_npages;
 608         int            sz = offsetof(struct lnet_rtrbuf, rb_kiov[npages]);
 609         struct page   *page;
 610         struct lnet_rtrbuf *rb;
 611         int            i;
 612
 613         LIBCFS_CPT_ALLOC(rb, lnet_cpt_table(), cpt, sz);
 614         if (rb == NULL)
 615                 return NULL;
 616
 617         rb->rb_pool = rbp;
 618
 619         for (i = 0; i < npages; i++) {
 620                 page = cfs_page_cpt_alloc(lnet_cpt_table(), cpt,
 621                                           GFP_KERNEL | __GFP_ZERO);
 622                 if (page == NULL) {
 623                         while (--i >= 0)
 624                                 __free_page(rb->rb_kiov[i].kiov_page);
 625
 626                         LIBCFS_FREE(rb, sz);
 627                         return NULL;
 628                 }
 629
 630                 rb->rb_kiov[i].kiov_len = PAGE_SIZE;
 631                 rb->rb_kiov[i].kiov_offset = 0;
 632                 rb->rb_kiov[i].kiov_page = page;
 633         }
 634
 635         return rb;
 636 }
 637
 638 static void
 639 lnet_rtrpool_free_bufs(struct lnet_rtrbufpool *rbp, int cpt)
 640 {
 641         int npages = rbp->rbp_npages;
 642         struct lnet_rtrbuf *rb;
 643         struct list_head tmp;
 644
 645         if (rbp->rbp_nbuffers == 0) /* not initialized or already freed */
 646                 return;
 647
 648         INIT_LIST_HEAD(&tmp);
 649
 650         lnet_net_lock(cpt);
 651         list_splice_init(&rbp->rbp_msgs, &tmp);
 652         lnet_drop_routed_msgs_locked(&tmp, cpt);
 653         list_splice_init(&rbp->rbp_bufs, &tmp);
 654         rbp->rbp_req_nbuffers = 0;
 655         rbp->rbp_nbuffers = rbp->rbp_credits = 0;
 656         rbp->rbp_mincredits = 0;
 657         lnet_net_unlock(cpt);
 658
 659         /* Free buffers on the free list. */
 660         while (!list_empty(&tmp)) {
 661                 rb = list_entry(tmp.next, struct lnet_rtrbuf, rb_list);
 662                 list_del(&rb->rb_list);
 663                 lnet_destroy_rtrbuf(rb, npages);
 664         }
 665 }
 666
 667 static int
 668 lnet_rtrpool_adjust_bufs(struct lnet_rtrbufpool *rbp, int nbufs, int cpt)
 669 {
 670         struct list_head rb_list;
 671         struct lnet_rtrbuf *rb;
 672         int             num_rb;
 673         int             num_buffers = 0;
 674         int             old_req_nbufs;
 675         int             npages = rbp->rbp_npages;
 676
 677         lnet_net_lock(cpt);
 678         /* If we are called for less buffers than already in the pool, we
 679          * just lower the req_nbuffers number and excess buffers will be
 680          * thrown away as they are returned to the free list.  Credits
 681          * then get adjusted as well.
 682          * If we already have enough buffers allocated to serve the
 683          * increase requested, then we can treat that the same way as we
 684          * do the decrease. */
 685         num_rb = nbufs - rbp->rbp_nbuffers;
 686         if (nbufs <= rbp->rbp_req_nbuffers || num_rb <= 0) {
 687                 rbp->rbp_req_nbuffers = nbufs;
 688                 lnet_net_unlock(cpt);
 689                 return 0;
 690         }
 691         /* store the older value of rbp_req_nbuffers and then set it to
 692          * the new request to prevent lnet_return_rx_credits_locked() from
 693          * freeing buffers that we need to keep around */
 694         old_req_nbufs = rbp->rbp_req_nbuffers;
 695         rbp->rbp_req_nbuffers = nbufs;
 696         lnet_net_unlock(cpt);
 697
 698         INIT_LIST_HEAD(&rb_list);
 699
 700         /* allocate the buffers on a local list first.  If all buffers are
 701          * allocated successfully then join this list to the rbp buffer
 702          * list.  If not then free all allocated buffers. */
 703         while (num_rb-- > 0) {
 704                 rb = lnet_new_rtrbuf(rbp, cpt);
 705                 if (rb == NULL) {
 706                         CERROR("Failed to allocate %d route bufs of %d pages\n",
 707                                nbufs, npages);
 708
 709                         lnet_net_lock(cpt);
 710                         rbp->rbp_req_nbuffers = old_req_nbufs;
 711                         lnet_net_unlock(cpt);
 712
 713                         goto failed;
 714                 }
 715
 716                 list_add(&rb->rb_list, &rb_list);
 717                 num_buffers++;
 718         }
 719
 720         lnet_net_lock(cpt);
 721
 722         list_splice_tail(&rb_list, &rbp->rbp_bufs);
 723         rbp->rbp_nbuffers += num_buffers;
 724         rbp->rbp_credits += num_buffers;
 725         rbp->rbp_mincredits = rbp->rbp_credits;
 726         /* We need to schedule blocked msg using the newly
 727          * added buffers. */
 728         while (!list_empty(&rbp->rbp_bufs) &&
 729                !list_empty(&rbp->rbp_msgs))
 730                 lnet_schedule_blocked_locked(rbp);
 731
 732         lnet_net_unlock(cpt);
 733
 734         return 0;
 735
 736 failed:
 737         while (!list_empty(&rb_list)) {
 738                 rb = list_entry(rb_list.next, struct lnet_rtrbuf, rb_list);
 739                 list_del(&rb->rb_list);
 740                 lnet_destroy_rtrbuf(rb, npages);
 741         }
 742
 743         return -ENOMEM;
 744 }
 745
 746 static void
 747 lnet_rtrpool_init(struct lnet_rtrbufpool *rbp, int npages)
 748 {
 749         INIT_LIST_HEAD(&rbp->rbp_msgs);
 750         INIT_LIST_HEAD(&rbp->rbp_bufs);
 751
 752         rbp->rbp_npages = npages;
 753         rbp->rbp_credits = 0;
 754         rbp->rbp_mincredits = 0;
 755 }
 756
 757 void
 758 lnet_rtrpools_free(int keep_pools)
 759 {
 760         struct lnet_rtrbufpool *rtrp;
 761         int               i;
 762
 763         if (the_lnet.ln_rtrpools == NULL) /* uninitialized or freed */
 764                 return;
 765
 766         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
 767                 lnet_rtrpool_free_bufs(&rtrp[LNET_TINY_BUF_IDX], i);
 768                 lnet_rtrpool_free_bufs(&rtrp[LNET_SMALL_BUF_IDX], i);
 769                 lnet_rtrpool_free_bufs(&rtrp[LNET_LARGE_BUF_IDX], i);
 770         }
 771
 772         if (!keep_pools) {
 773                 cfs_percpt_free(the_lnet.ln_rtrpools);
 774                 the_lnet.ln_rtrpools = NULL;
 775         }
 776 }
 777
 778 static int
 779 lnet_nrb_tiny_calculate(void)
 780 {
 781         int     nrbs = LNET_NRB_TINY;
 782
 783         if (tiny_router_buffers < 0) {
 784                 LCONSOLE_ERROR_MSG(0x10c,
 785                                    "tiny_router_buffers=%d invalid when "
 786                                    "routing enabled\n", tiny_router_buffers);
 787                 return -EINVAL;
 788         }
 789
 790         if (tiny_router_buffers > 0)
 791                 nrbs = tiny_router_buffers;
 792
 793         nrbs /= LNET_CPT_NUMBER;
 794         return max(nrbs, LNET_NRB_TINY_MIN);
 795 }
 796
 797 static int
 798 lnet_nrb_small_calculate(void)
 799 {
 800         int     nrbs = LNET_NRB_SMALL;
 801
 802         if (small_router_buffers < 0) {
 803                 LCONSOLE_ERROR_MSG(0x10c,
 804                                    "small_router_buffers=%d invalid when "
 805                                    "routing enabled\n", small_router_buffers);
 806                 return -EINVAL;
 807         }
 808
 809         if (small_router_buffers > 0)
 810                 nrbs = small_router_buffers;
 811
 812         nrbs /= LNET_CPT_NUMBER;
 813         return max(nrbs, LNET_NRB_SMALL_MIN);
 814 }
 815
 816 static int
 817 lnet_nrb_large_calculate(void)
 818 {
 819         int     nrbs = LNET_NRB_LARGE;
 820
 821         if (large_router_buffers < 0) {
 822                 LCONSOLE_ERROR_MSG(0x10c,
 823                                    "large_router_buffers=%d invalid when "
 824                                    "routing enabled\n", large_router_buffers);
 825                 return -EINVAL;
 826         }
 827
 828         if (large_router_buffers > 0)
 829                 nrbs = large_router_buffers;
 830
 831         nrbs /= LNET_CPT_NUMBER;
 832         return max(nrbs, LNET_NRB_LARGE_MIN);
 833 }
 834
 835 int
 836 lnet_rtrpools_alloc(int im_a_router)
 837 {
 838         struct lnet_rtrbufpool *rtrp;
 839         int     nrb_tiny;
 840         int     nrb_small;
 841         int     nrb_large;
 842         int     rc;
 843         int     i;
 844
 845         if (!strcmp(forwarding, "")) {
 846                 /* not set either way */
 847                 if (!im_a_router)
 848                         return 0;
 849         } else if (!strcmp(forwarding, "disabled")) {
 850                 /* explicitly disabled */
 851                 return 0;
 852         } else if (!strcmp(forwarding, "enabled")) {
 853                 /* explicitly enabled */
 854         } else {
 855                 LCONSOLE_ERROR_MSG(0x10b, "'forwarding' not set to either "
 856                                    "'enabled' or 'disabled'\n");
 857                 return -EINVAL;
 858         }
 859
 860         nrb_tiny = lnet_nrb_tiny_calculate();
 861         if (nrb_tiny < 0)
 862                 return -EINVAL;
 863
 864         nrb_small = lnet_nrb_small_calculate();
 865         if (nrb_small < 0)
 866                 return -EINVAL;
 867
 868         nrb_large = lnet_nrb_large_calculate();
 869         if (nrb_large < 0)
 870                 return -EINVAL;
 871
 872         the_lnet.ln_rtrpools = cfs_percpt_alloc(lnet_cpt_table(),
 873                                                 LNET_NRBPOOLS *
 874                                                 sizeof(struct lnet_rtrbufpool));
 875         if (the_lnet.ln_rtrpools == NULL) {
 876                 LCONSOLE_ERROR_MSG(0x10c,
 877                                    "Failed to initialize router buffe pool\n");
 878                 return -ENOMEM;
 879         }
 880
 881         cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
 882                 lnet_rtrpool_init(&rtrp[LNET_TINY_BUF_IDX], 0);
 883                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
 884                                               nrb_tiny, i);
 885                 if (rc != 0)
 886                         goto failed;
 887
 888                 lnet_rtrpool_init(&rtrp[LNET_SMALL_BUF_IDX],
 889                                   LNET_NRB_SMALL_PAGES);
 890                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
 891                                               nrb_small, i);
 892                 if (rc != 0)
 893                         goto failed;
 894
 895                 lnet_rtrpool_init(&rtrp[LNET_LARGE_BUF_IDX],
 896                                   LNET_NRB_LARGE_PAGES);
 897                 rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
 898                                               nrb_large, i);
 899                 if (rc != 0)
 900                         goto failed;
 901         }
 902
 903         lnet_net_lock(LNET_LOCK_EX);
 904         the_lnet.ln_routing = 1;
 905         lnet_net_unlock(LNET_LOCK_EX);
 906         return 0;
 907
 908  failed:
 909         lnet_rtrpools_free(0);
 910         return rc;
 911 }
 912
 913 static int
 914 lnet_rtrpools_adjust_helper(int tiny, int small, int large)
 915 {
 916         int nrb = 0;
 917         int rc = 0;
 918         int i;
 919         struct lnet_rtrbufpool *rtrp;
 920
 921         /* If the provided values for each buffer pool are different than the
 922          * configured values, we need to take action. */
 923         if (tiny >= 0) {
 924                 tiny_router_buffers = tiny;
 925                 nrb = lnet_nrb_tiny_calculate();
 926                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
 927                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_TINY_BUF_IDX],
 928                                                       nrb, i);
 929                         if (rc != 0)
 930                                 return rc;
 931                 }
 932         }
 933         if (small >= 0) {
 934                 small_router_buffers = small;
 935                 nrb = lnet_nrb_small_calculate();
 936                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
 937                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_SMALL_BUF_IDX],
 938                                                       nrb, i);
 939                         if (rc != 0)
 940                                 return rc;
 941                 }
 942         }
 943         if (large >= 0) {
 944                 large_router_buffers = large;
 945                 nrb = lnet_nrb_large_calculate();
 946                 cfs_percpt_for_each(rtrp, i, the_lnet.ln_rtrpools) {
 947                         rc = lnet_rtrpool_adjust_bufs(&rtrp[LNET_LARGE_BUF_IDX],
 948                                                       nrb, i);
 949                         if (rc != 0)
 950                                 return rc;
 951                 }
 952         }
 953
 954         return 0;
 955 }
 956
 957 int
 958 lnet_rtrpools_adjust(int tiny, int small, int large)
 959 {
 960         /* this function doesn't revert the changes if adding new buffers
 961          * failed.  It's up to the user space caller to revert the
 962          * changes. */
 963
 964         if (!the_lnet.ln_routing)
 965                 return 0;
 966
 967         return lnet_rtrpools_adjust_helper(tiny, small, large);
 968 }
 969
 970 int
 971 lnet_rtrpools_enable(void)
 972 {
 973         int rc = 0;
 974
 975         if (the_lnet.ln_routing)
 976                 return 0;
 977
 978         if (the_lnet.ln_rtrpools == NULL)
 979                 /* If routing is turned off, and we have never
 980                  * initialized the pools before, just call the
 981                  * standard buffer pool allocation routine as
 982                  * if we are just configuring this for the first
 983                  * time. */
 984                 rc = lnet_rtrpools_alloc(1);
 985         else
 986                 rc = lnet_rtrpools_adjust_helper(0, 0, 0);
 987         if (rc != 0)
 988                 return rc;
 989
 990         lnet_net_lock(LNET_LOCK_EX);
 991         the_lnet.ln_routing = 1;
 992
 993         the_lnet.ln_ping_target->pb_info.pi_features &=
 994                 ~LNET_PING_FEAT_RTE_DISABLED;
 995         lnet_net_unlock(LNET_LOCK_EX);
 996
 997         return rc;
 998 }
 999
1000 void
1001 lnet_rtrpools_disable(void)
1002 {
1003         if (!the_lnet.ln_routing)
1004                 return;
1005
1006         lnet_net_lock(LNET_LOCK_EX);
1007         the_lnet.ln_routing = 0;
1008         the_lnet.ln_ping_target->pb_info.pi_features |=
1009                 LNET_PING_FEAT_RTE_DISABLED;
1010
1011         tiny_router_buffers = 0;
1012         small_router_buffers = 0;
1013         large_router_buffers = 0;
1014         lnet_net_unlock(LNET_LOCK_EX);
1015         lnet_rtrpools_free(1);
1016 }
1017
1018 int
1019 lnet_notify(struct lnet_ni *ni, lnet_nid_t nid, int alive, time64_t when)
1020 {
1021         struct lnet_peer_ni *lp = NULL;
1022         time64_t now = ktime_get_seconds();
1023         int cpt = lnet_cpt_of_nid(nid, ni);
1024
1025         LASSERT (!in_interrupt ());
1026
1027         CDEBUG (D_NET, "%s notifying %s: %s\n",
1028                 (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1029                 libcfs_nid2str(nid),
1030                 alive ? "up" : "down");
1031
1032         if (ni != NULL &&
1033             LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid)) {
1034                 CWARN("Ignoring notification of %s %s by %s (different net)\n",
1035                       libcfs_nid2str(nid), alive ? "birth" : "death",
1036                       libcfs_nid2str(ni->ni_nid));
1037                 return -EINVAL;
1038         }
1039
1040         /* can't do predictions... */
1041         if (when > now) {
1042                 CWARN("Ignoring prediction from %s of %s %s "
1043                       "%lld seconds in the future\n",
1044                       (ni == NULL) ? "userspace" : libcfs_nid2str(ni->ni_nid),
1045                       libcfs_nid2str(nid), alive ? "up" : "down", when - now);
1046                 return -EINVAL;
1047         }
1048
1049         if (ni != NULL && !alive &&             /* LND telling me she's down */
1050             !auto_down) {                       /* auto-down disabled */
1051                 CDEBUG(D_NET, "Auto-down disabled\n");
1052                 return 0;
1053         }
1054
1055         lnet_net_lock(cpt);
1056
1057         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
1058                 lnet_net_unlock(cpt);
1059                 return -ESHUTDOWN;
1060         }
1061
1062         lp = lnet_find_peer_ni_locked(nid);
1063         if (lp == NULL) {
1064                 /* nid not found */
1065                 lnet_net_unlock(cpt);
1066                 CDEBUG(D_NET, "%s not found\n", libcfs_nid2str(nid));
1067                 return 0;
1068         }
1069
1070         /*
1071          * It is possible for this function to be called for the same peer
1072          * but with different NIs. We want to synchronize the notification
1073          * between the different calls. So we will use the lpni_cpt to
1074          * grab the net lock.
1075          */
1076         if (lp->lpni_cpt != cpt) {
1077                 lnet_net_unlock(cpt);
1078                 cpt = lp->lpni_cpt;
1079                 lnet_net_lock(cpt);
1080         }
1081
1082         /* We can't fully trust LND on reporting exact peer last_alive
1083          * if he notifies us about dead peer. For example ksocklnd can
1084          * call us with when == _time_when_the_node_was_booted_ if
1085          * no connections were successfully established */
1086         if (ni != NULL && !alive && when < lp->lpni_last_alive)
1087                 when = lp->lpni_last_alive;
1088
1089         lnet_notify_locked(lp, ni == NULL, alive, when);
1090
1091         if (ni != NULL)
1092                 lnet_ni_notify_locked(ni, lp);
1093
1094         lnet_peer_ni_decref_locked(lp);
1095
1096         lnet_net_unlock(cpt);
1097         return 0;
1098 }
1099 EXPORT_SYMBOL(lnet_notify);