pimd/pim_nht.c

   1 /*
   2  * PIM for Quagga
   3  * Copyright (C) 2017 Cumulus Networks, Inc.
   4  * Chirag Shah
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License as published by
   8  * the Free Software Foundation; either version 2 of the License, or
   9  * (at your option) any later version.
  10  *
  11  * This program is distributed in the hope that it will be useful, but
  12  * WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along
  17  * with this program; see the file COPYING; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20 #include <zebra.h>
  21 #include "network.h"
  22 #include "zclient.h"
  23 #include "stream.h"
  24 #include "nexthop.h"
  25 #include "if.h"
  26 #include "hash.h"
  27 #include "jhash.h"
  28
  29 #include "lib/printfrr.h"
  30
  31 #include "pimd.h"
  32 #include "pimd/pim_nht.h"
  33 #include "log.h"
  34 #include "pim_time.h"
  35 #include "pim_oil.h"
  36 #include "pim_ifchannel.h"
  37 #include "pim_mroute.h"
  38 #include "pim_zebra.h"
  39 #include "pim_upstream.h"
  40 #include "pim_join.h"
  41 #include "pim_jp_agg.h"
  42 #include "pim_zebra.h"
  43 #include "pim_zlookup.h"
  44 #include "pim_rp.h"
  45
  46 /**
  47  * pim_sendmsg_zebra_rnh -- Format and send a nexthop register/Unregister
  48  *   command to Zebra.
  49  */
  50 void pim_sendmsg_zebra_rnh(struct pim_instance *pim, struct zclient *zclient,
  51                            struct pim_nexthop_cache *pnc, int command)
  52 {
  53         struct prefix *p;
  54         int ret;
  55
  56         p = &(pnc->rpf.rpf_addr);
  57         ret = zclient_send_rnh(zclient, command, p, false, false,
  58                                pim->vrf->vrf_id);
  59         if (ret == ZCLIENT_SEND_FAILURE)
  60                 zlog_warn("sendmsg_nexthop: zclient_send_message() failed");
  61
  62         if (PIM_DEBUG_PIM_NHT)
  63                 zlog_debug(
  64                         "%s: NHT %sregistered addr %pFX(%s) with Zebra ret:%d ",
  65                         __func__,
  66                         (command == ZEBRA_NEXTHOP_REGISTER) ? " " : "de", p,
  67                         pim->vrf->name, ret);
  68
  69         return;
  70 }
  71
  72 struct pim_nexthop_cache *pim_nexthop_cache_find(struct pim_instance *pim,
  73                                                  struct pim_rpf *rpf)
  74 {
  75         struct pim_nexthop_cache *pnc = NULL;
  76         struct pim_nexthop_cache lookup;
  77
  78         lookup.rpf.rpf_addr.family = rpf->rpf_addr.family;
  79         lookup.rpf.rpf_addr.prefixlen = rpf->rpf_addr.prefixlen;
  80         lookup.rpf.rpf_addr.u.prefix4.s_addr = rpf->rpf_addr.u.prefix4.s_addr;
  81
  82         pnc = hash_lookup(pim->rpf_hash, &lookup);
  83
  84         return pnc;
  85 }
  86
  87 static struct pim_nexthop_cache *pim_nexthop_cache_add(struct pim_instance *pim,
  88                                                        struct pim_rpf *rpf_addr)
  89 {
  90         struct pim_nexthop_cache *pnc;
  91         char hash_name[64];
  92
  93         pnc = XCALLOC(MTYPE_PIM_NEXTHOP_CACHE,
  94                       sizeof(struct pim_nexthop_cache));
  95         pnc->rpf.rpf_addr.family = rpf_addr->rpf_addr.family;
  96         pnc->rpf.rpf_addr.prefixlen = rpf_addr->rpf_addr.prefixlen;
  97         pnc->rpf.rpf_addr.u.prefix4.s_addr =
  98                 rpf_addr->rpf_addr.u.prefix4.s_addr;
  99
 100         pnc = hash_get(pim->rpf_hash, pnc, hash_alloc_intern);
 101
 102         pnc->rp_list = list_new();
 103         pnc->rp_list->cmp = pim_rp_list_cmp;
 104
 105         snprintfrr(hash_name, sizeof(hash_name), "PNC %pFX(%s) Upstream Hash",
 106                    &pnc->rpf.rpf_addr, pim->vrf->name);
 107         pnc->upstream_hash = hash_create_size(8192, pim_upstream_hash_key,
 108                                               pim_upstream_equal, hash_name);
 109
 110         return pnc;
 111 }
 112
 113 static struct pim_nexthop_cache *pim_nht_get(struct pim_instance *pim,
 114                                              struct prefix *addr)
 115 {
 116         struct pim_nexthop_cache *pnc = NULL;
 117         struct pim_rpf rpf;
 118         struct zclient *zclient = NULL;
 119
 120         zclient = pim_zebra_zclient_get();
 121         memset(&rpf, 0, sizeof(struct pim_rpf));
 122         rpf.rpf_addr.family = addr->family;
 123         rpf.rpf_addr.prefixlen = addr->prefixlen;
 124         rpf.rpf_addr.u.prefix4 = addr->u.prefix4;
 125
 126         pnc = pim_nexthop_cache_find(pim, &rpf);
 127         if (!pnc) {
 128                 pnc = pim_nexthop_cache_add(pim, &rpf);
 129                 pim_sendmsg_zebra_rnh(pim, zclient, pnc,
 130                                       ZEBRA_NEXTHOP_REGISTER);
 131                 if (PIM_DEBUG_PIM_NHT)
 132                         zlog_debug(
 133                                 "%s: NHT cache and zebra notification added for %pFX(%s)",
 134                                 __func__, addr, pim->vrf->name);
 135         }
 136
 137         return pnc;
 138 }
 139
 140 /* TBD: this does several distinct things and should probably be split up.
 141  * (checking state vs. returning pnc vs. adding upstream vs. adding rp)
 142  */
 143 int pim_find_or_track_nexthop(struct pim_instance *pim, struct prefix *addr,
 144                               struct pim_upstream *up, struct rp_info *rp,
 145                               struct pim_nexthop_cache *out_pnc)
 146 {
 147         struct pim_nexthop_cache *pnc;
 148         struct listnode *ch_node = NULL;
 149
 150         pnc = pim_nht_get(pim, addr);
 151
 152         assertf(up || rp, "addr=%pFX", addr);
 153
 154         if (rp != NULL) {
 155                 ch_node = listnode_lookup(pnc->rp_list, rp);
 156                 if (ch_node == NULL)
 157                         listnode_add_sort(pnc->rp_list, rp);
 158         }
 159
 160         if (up != NULL)
 161                 hash_get(pnc->upstream_hash, up, hash_alloc_intern);
 162
 163         if (CHECK_FLAG(pnc->flags, PIM_NEXTHOP_VALID)) {
 164                 if (out_pnc)
 165                         memcpy(out_pnc, pnc, sizeof(struct pim_nexthop_cache));
 166                 return 1;
 167         }
 168
 169         return 0;
 170 }
 171
 172 void pim_nht_bsr_add(struct pim_instance *pim, struct in_addr addr)
 173 {
 174         struct pim_nexthop_cache *pnc;
 175         struct prefix pfx;
 176
 177         pfx.family = AF_INET;
 178         pfx.prefixlen = IPV4_MAX_BITLEN;
 179         pfx.u.prefix4 = addr;
 180
 181         pnc = pim_nht_get(pim, &pfx);
 182
 183         pnc->bsr_count++;
 184 }
 185
 186 static void pim_nht_drop_maybe(struct pim_instance *pim,
 187                                struct pim_nexthop_cache *pnc)
 188 {
 189         if (PIM_DEBUG_PIM_NHT)
 190                 zlog_debug(
 191                         "%s: NHT %pFX(%s) rp_list count:%d upstream count:%ld BSR count:%u",
 192                         __func__, &pnc->rpf.rpf_addr, pim->vrf->name,
 193                         pnc->rp_list->count, pnc->upstream_hash->count,
 194                         pnc->bsr_count);
 195
 196         if (pnc->rp_list->count == 0 && pnc->upstream_hash->count == 0
 197             && pnc->bsr_count == 0) {
 198                 struct zclient *zclient = pim_zebra_zclient_get();
 199
 200                 pim_sendmsg_zebra_rnh(pim, zclient, pnc,
 201                                       ZEBRA_NEXTHOP_UNREGISTER);
 202
 203                 list_delete(&pnc->rp_list);
 204                 hash_free(pnc->upstream_hash);
 205
 206                 hash_release(pim->rpf_hash, pnc);
 207                 if (pnc->nexthop)
 208                         nexthops_free(pnc->nexthop);
 209                 XFREE(MTYPE_PIM_NEXTHOP_CACHE, pnc);
 210         }
 211 }
 212
 213 void pim_delete_tracked_nexthop(struct pim_instance *pim, struct prefix *addr,
 214                                 struct pim_upstream *up, struct rp_info *rp)
 215 {
 216         struct pim_nexthop_cache *pnc = NULL;
 217         struct pim_nexthop_cache lookup;
 218         struct pim_upstream *upstream = NULL;
 219
 220         /* Remove from RPF hash if it is the last entry */
 221         lookup.rpf.rpf_addr = *addr;
 222         pnc = hash_lookup(pim->rpf_hash, &lookup);
 223         if (!pnc) {
 224                 zlog_warn("attempting to delete nonexistent NHT entry %pFX",
 225                           addr);
 226                 return;
 227         }
 228
 229         if (rp) {
 230                 /* Release the (*, G)upstream from pnc->upstream_hash,
 231                  * whose Group belongs to the RP getting deleted
 232                  */
 233                 frr_each (rb_pim_upstream, &pim->upstream_head, upstream) {
 234                         struct prefix grp;
 235                         struct rp_info *trp_info;
 236
 237                         if (!pim_addr_is_any(upstream->sg.src))
 238                                 continue;
 239
 240                         grp.family = AF_INET;
 241                         grp.prefixlen = IPV4_MAX_BITLEN;
 242                         grp.u.prefix4 = upstream->sg.grp;
 243
 244                         trp_info = pim_rp_find_match_group(pim, &grp);
 245                         if (trp_info == rp)
 246                                 hash_release(pnc->upstream_hash, upstream);
 247                 }
 248                 listnode_delete(pnc->rp_list, rp);
 249         }
 250
 251         if (up)
 252                 hash_release(pnc->upstream_hash, up);
 253
 254         pim_nht_drop_maybe(pim, pnc);
 255 }
 256
 257 void pim_nht_bsr_del(struct pim_instance *pim, struct in_addr addr)
 258 {
 259         struct pim_nexthop_cache *pnc = NULL;
 260         struct pim_nexthop_cache lookup;
 261
 262         /*
 263          * Nothing to do here if the address to unregister
 264          * is 0.0.0.0 as that the BSR has not been registered
 265          * for tracking yet.
 266          */
 267         if (addr.s_addr == INADDR_ANY)
 268                 return;
 269
 270         lookup.rpf.rpf_addr.family = AF_INET;
 271         lookup.rpf.rpf_addr.prefixlen = IPV4_MAX_BITLEN;
 272         lookup.rpf.rpf_addr.u.prefix4 = addr;
 273
 274         pnc = hash_lookup(pim->rpf_hash, &lookup);
 275
 276         if (!pnc) {
 277                 zlog_warn("attempting to delete nonexistent NHT BSR entry %pI4",
 278                           &addr);
 279                 return;
 280         }
 281
 282         assertf(pnc->bsr_count > 0, "addr=%pI4", &addr);
 283         pnc->bsr_count--;
 284
 285         pim_nht_drop_maybe(pim, pnc);
 286 }
 287
 288 bool pim_nht_bsr_rpf_check(struct pim_instance *pim, struct in_addr bsr_addr,
 289                            struct interface *src_ifp, struct in_addr src_ip)
 290 {
 291         struct pim_nexthop_cache *pnc = NULL;
 292         struct pim_nexthop_cache lookup;
 293         struct pim_neighbor *nbr = NULL;
 294         struct nexthop *nh;
 295         struct interface *ifp;
 296
 297         lookup.rpf.rpf_addr.family = AF_INET;
 298         lookup.rpf.rpf_addr.prefixlen = IPV4_MAX_BITLEN;
 299         lookup.rpf.rpf_addr.u.prefix4 = bsr_addr;
 300
 301         pnc = hash_lookup(pim->rpf_hash, &lookup);
 302         if (!pnc || !CHECK_FLAG(pnc->flags, PIM_NEXTHOP_ANSWER_RECEIVED)) {
 303                 /* BSM from a new freshly registered BSR - do a synchronous
 304                  * zebra query since otherwise we'd drop the first packet,
 305                  * leading to additional delay in picking up BSM data
 306                  */
 307
 308                 /* FIXME: this should really be moved into a generic NHT
 309                  * function that does "add and get immediate result" or maybe
 310                  * "check cache or get immediate result." But until that can
 311                  * be worked in, here's a copy of the code below :(
 312                  */
 313                 struct pim_zlookup_nexthop nexthop_tab[MULTIPATH_NUM];
 314                 ifindex_t i;
 315                 struct interface *ifp = NULL;
 316                 int num_ifindex;
 317
 318                 memset(nexthop_tab, 0, sizeof(nexthop_tab));
 319                 num_ifindex = zclient_lookup_nexthop(pim, nexthop_tab,
 320                                                      MULTIPATH_NUM, bsr_addr,
 321                                                      PIM_NEXTHOP_LOOKUP_MAX);
 322
 323                 if (num_ifindex <= 0)
 324                         return false;
 325
 326                 for (i = 0; i < num_ifindex; i++) {
 327                         struct pim_zlookup_nexthop *znh = &nexthop_tab[i];
 328
 329                         /* pim_zlookup_nexthop has no ->type */
 330
 331                         /* 1:1 match code below with znh instead of nh */
 332                         ifp = if_lookup_by_index(znh->ifindex,
 333                                                  pim->vrf->vrf_id);
 334
 335                         if (!ifp || !ifp->info)
 336                                 continue;
 337
 338                         if (if_is_loopback(ifp) && if_is_loopback(src_ifp))
 339                                 return true;
 340
 341                         nbr = pim_neighbor_find(ifp,
 342                                                 znh->nexthop_addr.u.prefix4);
 343                         if (!nbr)
 344                                 continue;
 345
 346                         return znh->ifindex == src_ifp->ifindex
 347                                && znh->nexthop_addr.u.prefix4.s_addr
 348                                           == src_ip.s_addr;
 349                 }
 350                 return false;
 351         }
 352
 353         if (!CHECK_FLAG(pnc->flags, PIM_NEXTHOP_VALID))
 354                 return false;
 355
 356         /* if we accept BSMs from more than one ECMP nexthop, this will cause
 357          * BSM message "multiplication" for each ECMP hop.  i.e. if you have
 358          * 4-way ECMP and 4 hops you end up with 256 copies of each BSM
 359          * message.
 360          *
 361          * so...  only accept the first (IPv4) valid nexthop as source.
 362          */
 363
 364         for (nh = pnc->nexthop; nh; nh = nh->next) {
 365                 struct in_addr nhaddr;
 366
 367                 switch (nh->type) {
 368                 case NEXTHOP_TYPE_IPV4:
 369                         if (nh->ifindex == IFINDEX_INTERNAL)
 370                                 continue;
 371
 372                         /* fallthru */
 373                 case NEXTHOP_TYPE_IPV4_IFINDEX:
 374                         nhaddr = nh->gate.ipv4;
 375                         break;
 376
 377                 case NEXTHOP_TYPE_IFINDEX:
 378                         nhaddr = bsr_addr;
 379                         break;
 380
 381                 default:
 382                         continue;
 383                 }
 384
 385                 ifp = if_lookup_by_index(nh->ifindex, pim->vrf->vrf_id);
 386                 if (!ifp || !ifp->info)
 387                         continue;
 388
 389                 if (if_is_loopback(ifp) && if_is_loopback(src_ifp))
 390                         return true;
 391
 392                 /* MRIB (IGP) may be pointing at a router where PIM is down */
 393                 nbr = pim_neighbor_find(ifp, nhaddr);
 394                 if (!nbr)
 395                         continue;
 396
 397                 return nh->ifindex == src_ifp->ifindex
 398                        && nhaddr.s_addr == src_ip.s_addr;
 399         }
 400         return false;
 401 }
 402
 403 void pim_rp_nexthop_del(struct rp_info *rp_info)
 404 {
 405         rp_info->rp.source_nexthop.interface = NULL;
 406         rp_info->rp.source_nexthop.mrib_nexthop_addr.u.prefix4.s_addr =
 407                 PIM_NET_INADDR_ANY;
 408         rp_info->rp.source_nexthop.mrib_metric_preference =
 409                 router->infinite_assert_metric.metric_preference;
 410         rp_info->rp.source_nexthop.mrib_route_metric =
 411                 router->infinite_assert_metric.route_metric;
 412 }
 413
 414 /* Update RP nexthop info based on Nexthop update received from Zebra.*/
 415 static void pim_update_rp_nh(struct pim_instance *pim,
 416                              struct pim_nexthop_cache *pnc)
 417 {
 418         struct listnode *node = NULL;
 419         struct rp_info *rp_info = NULL;
 420
 421         /*Traverse RP list and update each RP Nexthop info */
 422         for (ALL_LIST_ELEMENTS_RO(pnc->rp_list, node, rp_info)) {
 423                 if (rp_info->rp.rpf_addr.u.prefix4.s_addr == INADDR_NONE)
 424                         continue;
 425
 426                 // Compute PIM RPF using cached nexthop
 427                 if (!pim_ecmp_nexthop_lookup(pim, &rp_info->rp.source_nexthop,
 428                                              &rp_info->rp.rpf_addr,
 429                                              &rp_info->group, 1))
 430                         pim_rp_nexthop_del(rp_info);
 431         }
 432 }
 433
 434 /* Update Upstream nexthop info based on Nexthop update received from Zebra.*/
 435 static int pim_update_upstream_nh_helper(struct hash_bucket *bucket, void *arg)
 436 {
 437         struct pim_instance *pim = (struct pim_instance *)arg;
 438         struct pim_upstream *up = (struct pim_upstream *)bucket->data;
 439
 440         enum pim_rpf_result rpf_result;
 441         struct pim_rpf old;
 442
 443         old.source_nexthop.interface = up->rpf.source_nexthop.interface;
 444         rpf_result = pim_rpf_update(pim, up, &old, __func__);
 445
 446         /* update kernel multicast forwarding cache (MFC); if the
 447          * RPF nbr is now unreachable the MFC has already been updated
 448          * by pim_rpf_clear
 449          */
 450         if (rpf_result != PIM_RPF_FAILURE)
 451                 pim_upstream_mroute_iif_update(up->channel_oil, __func__);
 452
 453         if (rpf_result == PIM_RPF_CHANGED ||
 454                 (rpf_result == PIM_RPF_FAILURE && old.source_nexthop.interface))
 455                 pim_zebra_upstream_rpf_changed(pim, up, &old);
 456
 457
 458         if (PIM_DEBUG_PIM_NHT) {
 459                 zlog_debug(
 460                         "%s: NHT upstream %s(%s) old ifp %s new ifp %s",
 461                         __func__, up->sg_str, pim->vrf->name,
 462                         old.source_nexthop.interface ? old.source_nexthop
 463                                                                .interface->name
 464                                                      : "Unknown",
 465                         up->rpf.source_nexthop.interface ? up->rpf.source_nexthop
 466                                                                    .interface->name
 467                                                          : "Unknown");
 468         }
 469
 470         return HASHWALK_CONTINUE;
 471 }
 472
 473 static int pim_update_upstream_nh(struct pim_instance *pim,
 474                                   struct pim_nexthop_cache *pnc)
 475 {
 476         hash_walk(pnc->upstream_hash, pim_update_upstream_nh_helper, pim);
 477
 478         pim_zebra_update_all_interfaces(pim);
 479
 480         return 0;
 481 }
 482
 483 uint32_t pim_compute_ecmp_hash(struct prefix *src, struct prefix *grp)
 484 {
 485         uint32_t hash_val;
 486         uint32_t s = 0, g = 0;
 487
 488         if ((!src))
 489                 return 0;
 490
 491         switch (src->family) {
 492         case AF_INET: {
 493                 s = src->u.prefix4.s_addr;
 494                 s = s == 0 ? 1 : s;
 495                 if (grp)
 496                         g = grp->u.prefix4.s_addr;
 497         } break;
 498         default:
 499                 break;
 500         }
 501
 502         hash_val = jhash_2words(g, s, 101);
 503         return hash_val;
 504 }
 505
 506 static int pim_ecmp_nexthop_search(struct pim_instance *pim,
 507                                    struct pim_nexthop_cache *pnc,
 508                                    struct pim_nexthop *nexthop,
 509                                    struct prefix *src, struct prefix *grp,
 510                                    int neighbor_needed)
 511 {
 512         struct pim_neighbor *nbrs[MULTIPATH_NUM], *nbr = NULL;
 513         struct interface *ifps[MULTIPATH_NUM];
 514         struct nexthop *nh_node = NULL;
 515         ifindex_t first_ifindex;
 516         struct interface *ifp = NULL;
 517         uint32_t hash_val = 0, mod_val = 0;
 518         uint8_t nh_iter = 0, found = 0;
 519         uint32_t i, num_nbrs = 0;
 520
 521         if (!pnc || !pnc->nexthop_num || !nexthop)
 522                 return 0;
 523
 524         memset(&nbrs, 0, sizeof(nbrs));
 525         memset(&ifps, 0, sizeof(ifps));
 526
 527         // Current Nexthop is VALID, check to stay on the current path.
 528         if (nexthop->interface && nexthop->interface->info
 529             && nexthop->mrib_nexthop_addr.u.prefix4.s_addr
 530                        != PIM_NET_INADDR_ANY) {
 531                 /* User configured knob to explicitly switch
 532                    to new path is disabled or current path
 533                    metric is less than nexthop update.
 534                  */
 535
 536                 if (pim->ecmp_rebalance_enable == 0) {
 537                         uint8_t curr_route_valid = 0;
 538                         // Check if current nexthop is present in new updated
 539                         // Nexthop list.
 540                         // If the current nexthop is not valid, candidate to
 541                         // choose new Nexthop.
 542                         for (nh_node = pnc->nexthop; nh_node;
 543                              nh_node = nh_node->next) {
 544                                 curr_route_valid = (nexthop->interface->ifindex
 545                                                     == nh_node->ifindex);
 546                                 if (curr_route_valid)
 547                                         break;
 548                         }
 549
 550                         if (curr_route_valid
 551                             && !pim_if_connected_to_source(nexthop->interface,
 552                                                            src->u.prefix4)) {
 553                                 nbr = pim_neighbor_find(
 554                                         nexthop->interface,
 555                                         nexthop->mrib_nexthop_addr.u.prefix4);
 556                                 if (!nbr
 557                                     && !if_is_loopback(nexthop->interface)) {
 558                                         if (PIM_DEBUG_PIM_NHT)
 559                                                 zlog_debug(
 560                                                         "%s: current nexthop does not have nbr ",
 561                                                         __func__);
 562                                 } else {
 563                                         /* update metric even if the upstream
 564                                          * neighbor stays unchanged
 565                                          */
 566                                         nexthop->mrib_metric_preference =
 567                                                 pnc->distance;
 568                                         nexthop->mrib_route_metric =
 569                                                 pnc->metric;
 570                                         if (PIM_DEBUG_PIM_NHT) {
 571                                                 char src_str[INET_ADDRSTRLEN];
 572                                                 pim_inet4_dump("<addr?>",
 573                                                                src->u.prefix4,
 574                                                                src_str,
 575                                                                sizeof(src_str));
 576                                                 char grp_str[INET_ADDRSTRLEN];
 577                                                 pim_inet4_dump("<addr?>",
 578                                                                grp->u.prefix4,
 579                                                                grp_str,
 580                                                                sizeof(grp_str));
 581                                                 zlog_debug(
 582                                                         "%s: (%s,%s)(%s) current nexthop %s is valid, skipping new path selection",
 583                                                         __func__, src_str,
 584                                                         grp_str, pim->vrf->name,
 585                                                         nexthop->interface->name);
 586                                         }
 587                                         return 1;
 588                                 }
 589                         }
 590                 }
 591         }
 592
 593         /*
 594          * Look up all interfaces and neighbors,
 595          * store for later usage
 596          */
 597         for (nh_node = pnc->nexthop, i = 0; nh_node;
 598              nh_node = nh_node->next, i++) {
 599                 ifps[i] =
 600                         if_lookup_by_index(nh_node->ifindex, pim->vrf->vrf_id);
 601                 if (ifps[i]) {
 602                         nbrs[i] = pim_neighbor_find(ifps[i],
 603                                                     nh_node->gate.ipv4);
 604                         if (nbrs[i] || pim_if_connected_to_source(ifps[i],
 605
 606                                                                   src->u.prefix4))
 607                                 num_nbrs++;
 608                 }
 609         }
 610         if (pim->ecmp_enable) {
 611                 uint32_t consider = pnc->nexthop_num;
 612
 613                 if (neighbor_needed && num_nbrs < consider)
 614                         consider = num_nbrs;
 615
 616                 if (consider == 0)
 617                         return 0;
 618
 619                 // PIM ECMP flag is enable then choose ECMP path.
 620                 hash_val = pim_compute_ecmp_hash(src, grp);
 621                 mod_val = hash_val % consider;
 622         }
 623
 624         for (nh_node = pnc->nexthop; nh_node && (found == 0);
 625              nh_node = nh_node->next) {
 626                 first_ifindex = nh_node->ifindex;
 627                 ifp = ifps[nh_iter];
 628                 if (!ifp) {
 629                         if (PIM_DEBUG_PIM_NHT) {
 630                                 char addr_str[INET_ADDRSTRLEN];
 631                                 pim_inet4_dump("<addr?>", src->u.prefix4,
 632                                                addr_str, sizeof(addr_str));
 633                                 zlog_debug(
 634                                         "%s %s: could not find interface for ifindex %d (address %s(%s))",
 635                                         __FILE__, __func__, first_ifindex,
 636                                         addr_str, pim->vrf->name);
 637                         }
 638                         if (nh_iter == mod_val)
 639                                 mod_val++; // Select nexthpath
 640                         nh_iter++;
 641                         continue;
 642                 }
 643                 if (!ifp->info) {
 644                         if (PIM_DEBUG_PIM_NHT) {
 645                                 char addr_str[INET_ADDRSTRLEN];
 646                                 pim_inet4_dump("<addr?>", src->u.prefix4,
 647                                                addr_str, sizeof(addr_str));
 648                                 zlog_debug(
 649                                         "%s: multicast not enabled on input interface %s(%s) (ifindex=%d, RPF for source %s)",
 650                                         __func__, ifp->name, pim->vrf->name,
 651                                         first_ifindex, addr_str);
 652                         }
 653                         if (nh_iter == mod_val)
 654                                 mod_val++; // Select nexthpath
 655                         nh_iter++;
 656                         continue;
 657                 }
 658
 659                 if (neighbor_needed
 660                     && !pim_if_connected_to_source(ifp, src->u.prefix4)) {
 661                         nbr = nbrs[nh_iter];
 662                         if (!nbr && !if_is_loopback(ifp)) {
 663                                 if (PIM_DEBUG_PIM_NHT)
 664                                         zlog_debug(
 665                                                 "%s: pim nbr not found on input interface %s(%s)",
 666                                                 __func__, ifp->name,
 667                                                 pim->vrf->name);
 668                                 if (nh_iter == mod_val)
 669                                         mod_val++; // Select nexthpath
 670                                 nh_iter++;
 671                                 continue;
 672                         }
 673                 }
 674
 675                 if (nh_iter == mod_val) {
 676                         nexthop->interface = ifp;
 677                         nexthop->mrib_nexthop_addr.family = AF_INET;
 678                         nexthop->mrib_nexthop_addr.prefixlen = IPV4_MAX_BITLEN;
 679                         nexthop->mrib_nexthop_addr.u.prefix4 =
 680                                 nh_node->gate.ipv4;
 681                         nexthop->mrib_metric_preference = pnc->distance;
 682                         nexthop->mrib_route_metric = pnc->metric;
 683                         nexthop->last_lookup = src->u.prefix4;
 684                         nexthop->last_lookup_time = pim_time_monotonic_usec();
 685                         nexthop->nbr = nbr;
 686                         found = 1;
 687                         if (PIM_DEBUG_PIM_NHT) {
 688                                 char buf[INET_ADDRSTRLEN];
 689                                 char buf2[INET_ADDRSTRLEN];
 690                                 char buf3[INET_ADDRSTRLEN];
 691                                 pim_inet4_dump("<src?>", src->u.prefix4, buf2,
 692                                                sizeof(buf2));
 693                                 pim_inet4_dump("<grp?>", grp->u.prefix4, buf3,
 694                                                sizeof(buf3));
 695                                 pim_inet4_dump(
 696                                         "<rpf?>",
 697                                         nexthop->mrib_nexthop_addr.u.prefix4,
 698                                         buf, sizeof(buf));
 699                                 zlog_debug(
 700                                         "%s: (%s,%s)(%s) selected nhop interface %s addr %s mod_val %u iter %d ecmp %d",
 701                                         __func__, buf2, buf3, pim->vrf->name,
 702                                         ifp->name, buf, mod_val, nh_iter,
 703                                         pim->ecmp_enable);
 704                         }
 705                 }
 706                 nh_iter++;
 707         }
 708
 709         if (found)
 710                 return 1;
 711         else
 712                 return 0;
 713 }
 714
 715 /* This API is used to parse Registered address nexthop update coming from Zebra
 716  */
 717 int pim_parse_nexthop_update(ZAPI_CALLBACK_ARGS)
 718 {
 719         struct nexthop *nexthop;
 720         struct nexthop *nhlist_head = NULL;
 721         struct nexthop *nhlist_tail = NULL;
 722         int i;
 723         struct pim_rpf rpf;
 724         struct pim_nexthop_cache *pnc = NULL;
 725         struct pim_neighbor *nbr = NULL;
 726         struct interface *ifp = NULL;
 727         struct interface *ifp1 = NULL;
 728         struct vrf *vrf = vrf_lookup_by_id(vrf_id);
 729         struct pim_instance *pim;
 730         struct zapi_route nhr;
 731
 732         if (!vrf)
 733                 return 0;
 734         pim = vrf->info;
 735
 736         if (!zapi_nexthop_update_decode(zclient->ibuf, &nhr)) {
 737                 zlog_err("%s: Decode of nexthop update from zebra failed",
 738                          __func__);
 739                 return 0;
 740         }
 741
 742         if (cmd == ZEBRA_NEXTHOP_UPDATE) {
 743                 prefix_copy(&rpf.rpf_addr, &nhr.prefix);
 744                 pnc = pim_nexthop_cache_find(pim, &rpf);
 745                 if (!pnc) {
 746                         if (PIM_DEBUG_PIM_NHT)
 747                                 zlog_debug(
 748                                         "%s: Skipping NHT update, addr %pFX is not in local cached DB.",
 749                                         __func__, &rpf.rpf_addr);
 750                         return 0;
 751                 }
 752         } else {
 753                 /*
 754                  * We do not currently handle ZEBRA_IMPORT_CHECK_UPDATE
 755                  */
 756                 return 0;
 757         }
 758
 759         pnc->last_update = pim_time_monotonic_usec();
 760
 761         if (nhr.nexthop_num) {
 762                 pnc->nexthop_num = 0; // Only increment for pim enabled rpf.
 763
 764                 for (i = 0; i < nhr.nexthop_num; i++) {
 765                         nexthop = nexthop_from_zapi_nexthop(&nhr.nexthops[i]);
 766                         switch (nexthop->type) {
 767                         case NEXTHOP_TYPE_IPV4:
 768                         case NEXTHOP_TYPE_IPV4_IFINDEX:
 769                         case NEXTHOP_TYPE_IPV6:
 770                         case NEXTHOP_TYPE_BLACKHOLE:
 771                                 break;
 772                         case NEXTHOP_TYPE_IFINDEX:
 773                                 /*
 774                                  * Connected route (i.e. no nexthop), use
 775                                  * RPF address from nexthop cache (i.e.
 776                                  * destination) as PIM nexthop.
 777                                  */
 778                                 nexthop->type = NEXTHOP_TYPE_IPV4_IFINDEX;
 779                                 nexthop->gate.ipv4 =
 780                                         pnc->rpf.rpf_addr.u.prefix4;
 781                                 break;
 782                         case NEXTHOP_TYPE_IPV6_IFINDEX:
 783                                 ifp1 = if_lookup_by_index(nexthop->ifindex,
 784                                                           pim->vrf->vrf_id);
 785
 786                                 if (!ifp1)
 787                                         nbr = NULL;
 788                                 else
 789                                         nbr = pim_neighbor_find_if(ifp1);
 790                                 /* Overwrite with Nbr address as NH addr */
 791                                 if (nbr)
 792                                         nexthop->gate.ipv4 = nbr->source_addr;
 793                                 else {
 794                                         // Mark nexthop address to 0 until PIM
 795                                         // Nbr is resolved.
 796                                         nexthop->gate.ipv4.s_addr =
 797                                                 PIM_NET_INADDR_ANY;
 798                                 }
 799
 800                                 break;
 801                         }
 802
 803                         ifp = if_lookup_by_index(nexthop->ifindex,
 804                                                  pim->vrf->vrf_id);
 805                         if (!ifp) {
 806                                 if (PIM_DEBUG_PIM_NHT) {
 807                                         char buf[NEXTHOP_STRLEN];
 808                                         zlog_debug(
 809                                                 "%s: could not find interface for ifindex %d(%s) (addr %s)",
 810                                                 __func__, nexthop->ifindex,
 811                                                 pim->vrf->name,
 812                                                 nexthop2str(nexthop, buf,
 813                                                             sizeof(buf)));
 814                                 }
 815                                 nexthop_free(nexthop);
 816                                 continue;
 817                         }
 818
 819                         if (PIM_DEBUG_PIM_NHT)
 820                                 zlog_debug(
 821                                         "%s: NHT addr %pFX(%s) %d-nhop via %pI4(%s) type %d distance:%u metric:%u ",
 822                                         __func__, &nhr.prefix, pim->vrf->name,
 823                                         i + 1, &nexthop->gate.ipv4,
 824                                         ifp->name, nexthop->type, nhr.distance,
 825                                         nhr.metric);
 826
 827                         if (!ifp->info) {
 828                                 /*
 829                                  * Though Multicast is not enabled on this
 830                                  * Interface store it in database otheriwse we
 831                                  * may miss this update and this will not cause
 832                                  * any issue, because while choosing the path we
 833                                  * are ommitting the Interfaces which are not
 834                                  * multicast enabled
 835                                  */
 836                                 if (PIM_DEBUG_PIM_NHT) {
 837                                         char buf[NEXTHOP_STRLEN];
 838
 839                                         zlog_debug(
 840                                                 "%s: multicast not enabled on input interface %s(%s) (ifindex=%d, addr %s)",
 841                                                 __func__, ifp->name,
 842                                                 pim->vrf->name,
 843                                                 nexthop->ifindex,
 844                                                 nexthop2str(nexthop, buf,
 845                                                             sizeof(buf)));
 846                                 }
 847                         }
 848
 849                         if (nhlist_tail) {
 850                                 nhlist_tail->next = nexthop;
 851                                 nhlist_tail = nexthop;
 852                         } else {
 853                                 nhlist_tail = nexthop;
 854                                 nhlist_head = nexthop;
 855                         }
 856                         // Only keep track of nexthops which are PIM enabled.
 857                         pnc->nexthop_num++;
 858                 }
 859                 /* Reset existing pnc->nexthop before assigning new list */
 860                 nexthops_free(pnc->nexthop);
 861                 pnc->nexthop = nhlist_head;
 862                 if (pnc->nexthop_num) {
 863                         pnc->flags |= PIM_NEXTHOP_VALID;
 864                         pnc->distance = nhr.distance;
 865                         pnc->metric = nhr.metric;
 866                 }
 867         } else {
 868                 pnc->flags &= ~PIM_NEXTHOP_VALID;
 869                 pnc->nexthop_num = nhr.nexthop_num;
 870                 nexthops_free(pnc->nexthop);
 871                 pnc->nexthop = NULL;
 872         }
 873         SET_FLAG(pnc->flags, PIM_NEXTHOP_ANSWER_RECEIVED);
 874
 875         if (PIM_DEBUG_PIM_NHT)
 876                 zlog_debug(
 877                         "%s: NHT Update for %pFX(%s) num_nh %d num_pim_nh %d vrf:%u up %ld rp %d",
 878                         __func__, &nhr.prefix, pim->vrf->name, nhr.nexthop_num,
 879                         pnc->nexthop_num, vrf_id, pnc->upstream_hash->count,
 880                         listcount(pnc->rp_list));
 881
 882         pim_rpf_set_refresh_time(pim);
 883
 884         if (listcount(pnc->rp_list))
 885                 pim_update_rp_nh(pim, pnc);
 886         if (pnc->upstream_hash->count)
 887                 pim_update_upstream_nh(pim, pnc);
 888
 889         return 0;
 890 }
 891
 892 int pim_ecmp_nexthop_lookup(struct pim_instance *pim,
 893                             struct pim_nexthop *nexthop, struct prefix *src,
 894                             struct prefix *grp, int neighbor_needed)
 895 {
 896         struct pim_nexthop_cache *pnc;
 897         struct pim_zlookup_nexthop nexthop_tab[MULTIPATH_NUM];
 898         struct pim_neighbor *nbrs[MULTIPATH_NUM], *nbr = NULL;
 899         struct pim_rpf rpf;
 900         int num_ifindex;
 901         struct interface *ifps[MULTIPATH_NUM], *ifp;
 902         int first_ifindex;
 903         int found = 0;
 904         uint8_t i = 0;
 905         uint32_t hash_val = 0, mod_val = 0;
 906         uint32_t num_nbrs = 0;
 907         char addr_str[PREFIX_STRLEN];
 908
 909         if (PIM_DEBUG_PIM_NHT) {
 910                 pim_inet4_dump("<addr?>", src->u.prefix4, addr_str,
 911                                sizeof(addr_str));
 912                 zlog_debug("%s: Looking up: %s(%s), last lookup time: %lld",
 913                            __func__, addr_str, pim->vrf->name,
 914                            nexthop->last_lookup_time);
 915         }
 916
 917         memset(&rpf, 0, sizeof(struct pim_rpf));
 918         rpf.rpf_addr.family = AF_INET;
 919         rpf.rpf_addr.prefixlen = IPV4_MAX_BITLEN;
 920         rpf.rpf_addr.u.prefix4 = src->u.prefix4;
 921
 922         pnc = pim_nexthop_cache_find(pim, &rpf);
 923         if (pnc) {
 924                 if (CHECK_FLAG(pnc->flags, PIM_NEXTHOP_ANSWER_RECEIVED))
 925                     return pim_ecmp_nexthop_search(pim, pnc, nexthop, src, grp,
 926                                                    neighbor_needed);
 927         }
 928
 929         memset(nexthop_tab, 0,
 930                sizeof(struct pim_zlookup_nexthop) * MULTIPATH_NUM);
 931         num_ifindex =
 932                 zclient_lookup_nexthop(pim, nexthop_tab, MULTIPATH_NUM,
 933                                        src->u.prefix4, PIM_NEXTHOP_LOOKUP_MAX);
 934         if (num_ifindex < 1) {
 935                 if (PIM_DEBUG_PIM_NHT)
 936                         zlog_warn(
 937                                 "%s: could not find nexthop ifindex for address %s(%s)",
 938                                 __func__, addr_str, pim->vrf->name);
 939                 return 0;
 940         }
 941
 942         memset(&nbrs, 0, sizeof(nbrs));
 943         memset(&ifps, 0, sizeof(ifps));
 944
 945         /*
 946          * Look up all interfaces and neighbors,
 947          * store for later usage
 948          */
 949         for (i = 0; i < num_ifindex; i++) {
 950                 ifps[i] = if_lookup_by_index(nexthop_tab[i].ifindex,
 951                                              pim->vrf->vrf_id);
 952                 if (ifps[i]) {
 953                         nbrs[i] = pim_neighbor_find(
 954                                 ifps[i], nexthop_tab[i].nexthop_addr.u.prefix4);
 955                         if (nbrs[i]
 956                             || pim_if_connected_to_source(ifps[i],
 957                                                           src->u.prefix4))
 958                                 num_nbrs++;
 959                 }
 960         }
 961
 962         // If PIM ECMP enable then choose ECMP path.
 963         if (pim->ecmp_enable) {
 964                 uint32_t consider = num_ifindex;
 965
 966                 if (neighbor_needed && num_nbrs < consider)
 967                         consider = num_nbrs;
 968
 969                 if (consider == 0)
 970                         return 0;
 971
 972                 hash_val = pim_compute_ecmp_hash(src, grp);
 973                 mod_val = hash_val % consider;
 974                 if (PIM_DEBUG_PIM_NHT_DETAIL)
 975                         zlog_debug("%s: hash_val %u mod_val %u", __func__,
 976                                    hash_val, mod_val);
 977         }
 978
 979         i = 0;
 980         while (!found && (i < num_ifindex)) {
 981                 first_ifindex = nexthop_tab[i].ifindex;
 982
 983                 ifp = ifps[i];
 984                 if (!ifp) {
 985                         if (PIM_DEBUG_PIM_NHT)
 986                                 zlog_debug(
 987                                         "%s %s: could not find interface for ifindex %d (address %s(%s))",
 988                                         __FILE__, __func__, first_ifindex,
 989                                         addr_str, pim->vrf->name);
 990                         if (i == mod_val)
 991                                 mod_val++;
 992                         i++;
 993                         continue;
 994                 }
 995
 996                 if (!ifp->info) {
 997                         if (PIM_DEBUG_PIM_NHT)
 998                                 zlog_debug(
 999                                         "%s: multicast not enabled on input interface %s(%s) (ifindex=%d, RPF for source %s)",
1000                                         __func__, ifp->name, pim->vrf->name,
1001                                         first_ifindex, addr_str);
1002                         if (i == mod_val)
1003                                 mod_val++;
1004                         i++;
1005                         continue;
1006                 }
1007                 if (neighbor_needed
1008                     && !pim_if_connected_to_source(ifp, src->u.prefix4)) {
1009                         nbr = nbrs[i];
1010                         if (PIM_DEBUG_PIM_NHT_DETAIL)
1011                                 zlog_debug("ifp name: %s(%s), pim nbr: %p",
1012                                            ifp->name, pim->vrf->name, nbr);
1013                         if (!nbr && !if_is_loopback(ifp)) {
1014                                 if (i == mod_val)
1015                                         mod_val++;
1016                                 i++;
1017                                 if (PIM_DEBUG_PIM_NHT)
1018                                         zlog_debug(
1019                                                 "%s: NBR not found on input interface %s(%s) (RPF for source %s)",
1020                                                 __func__, ifp->name,
1021                                                 pim->vrf->name, addr_str);
1022                                 continue;
1023                         }
1024                 }
1025
1026                 if (i == mod_val) {
1027                         if (PIM_DEBUG_PIM_NHT) {
1028                                 char nexthop_str[PREFIX_STRLEN];
1029
1030                                 pim_addr_dump("<nexthop?>",
1031                                               &nexthop_tab[i].nexthop_addr,
1032                                               nexthop_str, sizeof(nexthop_str));
1033                                 zlog_debug(
1034                                         "%s: found nhop %s for addr %s interface %s(%s) metric %d dist %d",
1035                                         __func__, nexthop_str, addr_str,
1036                                         ifp->name, pim->vrf->name,
1037                                         nexthop_tab[i].route_metric,
1038                                         nexthop_tab[i].protocol_distance);
1039                         }
1040                         /* update nexthop data */
1041                         nexthop->interface = ifp;
1042                         nexthop->mrib_nexthop_addr =
1043                                 nexthop_tab[i].nexthop_addr;
1044                         nexthop->mrib_metric_preference =
1045                                 nexthop_tab[i].protocol_distance;
1046                         nexthop->mrib_route_metric =
1047                                 nexthop_tab[i].route_metric;
1048                         nexthop->last_lookup = src->u.prefix4;
1049                         nexthop->last_lookup_time = pim_time_monotonic_usec();
1050                         nexthop->nbr = nbr;
1051                         found = 1;
1052                 }
1053                 i++;
1054         }
1055
1056         if (found)
1057                 return 1;
1058         else
1059                 return 0;
1060 }
1061
1062 int pim_ecmp_fib_lookup_if_vif_index(struct pim_instance *pim,
1063                                      struct prefix *src, struct prefix *grp)
1064 {
1065         struct pim_nexthop nhop;
1066         int vif_index;
1067         ifindex_t ifindex;
1068         char addr_str[PREFIX_STRLEN];
1069
1070         if (PIM_DEBUG_PIM_NHT)
1071                 pim_inet4_dump("<addr?>", src->u.prefix4, addr_str,
1072                                sizeof(addr_str));
1073
1074         memset(&nhop, 0, sizeof(nhop));
1075         if (!pim_ecmp_nexthop_lookup(pim, &nhop, src, grp, 1)) {
1076                 if (PIM_DEBUG_PIM_NHT)
1077                         zlog_debug(
1078                                 "%s: could not find nexthop ifindex for address %s(%s)",
1079                                 __func__, addr_str, pim->vrf->name);
1080                 return -1;
1081         }
1082
1083         ifindex = nhop.interface->ifindex;
1084         if (PIM_DEBUG_PIM_NHT)
1085                 zlog_debug(
1086                         "%s: found nexthop ifindex=%d (interface %s(%s)) for address %s",
1087                         __func__, ifindex,
1088                         ifindex2ifname(ifindex, pim->vrf->vrf_id),
1089                         pim->vrf->name, addr_str);
1090
1091         vif_index = pim_if_find_vifindex_by_ifindex(pim, ifindex);
1092
1093         if (vif_index < 0) {
1094                 if (PIM_DEBUG_PIM_NHT) {
1095                         zlog_debug(
1096                                 "%s: low vif_index=%d(%s) < 1 nexthop for address %s",
1097                                 __func__, vif_index, pim->vrf->name, addr_str);
1098                 }
1099                 return -2;
1100         }
1101
1102         return vif_index;
1103 }