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 #include "pim_addr.h"
  46
  47 /**
  48  * pim_sendmsg_zebra_rnh -- Format and send a nexthop register/Unregister
  49  *   command to Zebra.
  50  */
  51 void pim_sendmsg_zebra_rnh(struct pim_instance *pim, struct zclient *zclient,
  52                            struct pim_nexthop_cache *pnc, int command)
  53 {
  54         struct prefix *p;
  55         int ret;
  56
  57         p = &(pnc->rpf.rpf_addr);
  58         ret = zclient_send_rnh(zclient, command, p, SAFI_UNICAST, false, false,
  59                                pim->vrf->vrf_id);
  60         if (ret == ZCLIENT_SEND_FAILURE)
  61                 zlog_warn("sendmsg_nexthop: zclient_send_message() failed");
  62
  63         if (PIM_DEBUG_PIM_NHT)
  64                 zlog_debug(
  65                         "%s: NHT %sregistered addr %pFX(%s) with Zebra ret:%d ",
  66                         __func__,
  67                         (command == ZEBRA_NEXTHOP_REGISTER) ? " " : "de", p,
  68                         pim->vrf->name, ret);
  69
  70         return;
  71 }
  72
  73 struct pim_nexthop_cache *pim_nexthop_cache_find(struct pim_instance *pim,
  74                                                  struct pim_rpf *rpf)
  75 {
  76         struct pim_nexthop_cache *pnc = NULL;
  77         struct pim_nexthop_cache lookup;
  78
  79         lookup.rpf.rpf_addr = rpf->rpf_addr;
  80         pnc = hash_lookup(pim->rpf_hash, &lookup);
  81
  82         return pnc;
  83 }
  84
  85 static struct pim_nexthop_cache *pim_nexthop_cache_add(struct pim_instance *pim,
  86                                                        struct pim_rpf *rpf_addr)
  87 {
  88         struct pim_nexthop_cache *pnc;
  89         char hash_name[64];
  90
  91         pnc = XCALLOC(MTYPE_PIM_NEXTHOP_CACHE,
  92                       sizeof(struct pim_nexthop_cache));
  93         pnc->rpf.rpf_addr = rpf_addr->rpf_addr;
  94
  95         pnc = hash_get(pim->rpf_hash, pnc, hash_alloc_intern);
  96
  97         pnc->rp_list = list_new();
  98         pnc->rp_list->cmp = pim_rp_list_cmp;
  99
 100         snprintfrr(hash_name, sizeof(hash_name), "PNC %pFX(%s) Upstream Hash",
 101                    &pnc->rpf.rpf_addr, pim->vrf->name);
 102         pnc->upstream_hash = hash_create_size(8192, pim_upstream_hash_key,
 103                                               pim_upstream_equal, hash_name);
 104
 105         return pnc;
 106 }
 107
 108 static struct pim_nexthop_cache *pim_nht_get(struct pim_instance *pim,
 109                                              struct prefix *addr)
 110 {
 111         struct pim_nexthop_cache *pnc = NULL;
 112         struct pim_rpf rpf;
 113         struct zclient *zclient = NULL;
 114
 115         zclient = pim_zebra_zclient_get();
 116         memset(&rpf, 0, sizeof(struct pim_rpf));
 117         rpf.rpf_addr = *addr;
 118
 119         pnc = pim_nexthop_cache_find(pim, &rpf);
 120         if (!pnc) {
 121                 pnc = pim_nexthop_cache_add(pim, &rpf);
 122                 pim_sendmsg_zebra_rnh(pim, zclient, pnc,
 123                                       ZEBRA_NEXTHOP_REGISTER);
 124                 if (PIM_DEBUG_PIM_NHT_DETAIL)
 125                         zlog_debug(
 126                                 "%s: NHT cache and zebra notification added for %pFX(%s)",
 127                                 __func__, addr, pim->vrf->name);
 128         }
 129
 130         return pnc;
 131 }
 132
 133 /* TBD: this does several distinct things and should probably be split up.
 134  * (checking state vs. returning pnc vs. adding upstream vs. adding rp)
 135  */
 136 int pim_find_or_track_nexthop(struct pim_instance *pim, struct prefix *addr,
 137                               struct pim_upstream *up, struct rp_info *rp,
 138                               struct pim_nexthop_cache *out_pnc)
 139 {
 140         struct pim_nexthop_cache *pnc;
 141         struct listnode *ch_node = NULL;
 142
 143         pnc = pim_nht_get(pim, addr);
 144
 145         assertf(up || rp, "addr=%pFX", addr);
 146
 147         if (rp != NULL) {
 148                 ch_node = listnode_lookup(pnc->rp_list, rp);
 149                 if (ch_node == NULL)
 150                         listnode_add_sort(pnc->rp_list, rp);
 151         }
 152
 153         if (up != NULL)
 154                 hash_get(pnc->upstream_hash, up, hash_alloc_intern);
 155
 156         if (CHECK_FLAG(pnc->flags, PIM_NEXTHOP_VALID)) {
 157                 if (out_pnc)
 158                         memcpy(out_pnc, pnc, sizeof(struct pim_nexthop_cache));
 159                 return 1;
 160         }
 161
 162         return 0;
 163 }
 164
 165 #if PIM_IPV == 4
 166 void pim_nht_bsr_add(struct pim_instance *pim, struct in_addr addr)
 167 {
 168         struct pim_nexthop_cache *pnc;
 169         struct prefix pfx;
 170
 171         pfx.family = AF_INET;
 172         pfx.prefixlen = IPV4_MAX_BITLEN;
 173         pfx.u.prefix4 = addr;
 174
 175         pnc = pim_nht_get(pim, &pfx);
 176
 177         pnc->bsr_count++;
 178 }
 179 #endif /* PIM_IPV == 4 */
 180
 181 static void pim_nht_drop_maybe(struct pim_instance *pim,
 182                                struct pim_nexthop_cache *pnc)
 183 {
 184         if (PIM_DEBUG_PIM_NHT)
 185                 zlog_debug(
 186                         "%s: NHT %pFX(%s) rp_list count:%d upstream count:%ld BSR count:%u",
 187                         __func__, &pnc->rpf.rpf_addr, pim->vrf->name,
 188                         pnc->rp_list->count, pnc->upstream_hash->count,
 189                         pnc->bsr_count);
 190
 191         if (pnc->rp_list->count == 0 && pnc->upstream_hash->count == 0
 192             && pnc->bsr_count == 0) {
 193                 struct zclient *zclient = pim_zebra_zclient_get();
 194
 195                 pim_sendmsg_zebra_rnh(pim, zclient, pnc,
 196                                       ZEBRA_NEXTHOP_UNREGISTER);
 197
 198                 list_delete(&pnc->rp_list);
 199                 hash_free(pnc->upstream_hash);
 200
 201                 hash_release(pim->rpf_hash, pnc);
 202                 if (pnc->nexthop)
 203                         nexthops_free(pnc->nexthop);
 204                 XFREE(MTYPE_PIM_NEXTHOP_CACHE, pnc);
 205         }
 206 }
 207
 208 void pim_delete_tracked_nexthop(struct pim_instance *pim, struct prefix *addr,
 209                                 struct pim_upstream *up, struct rp_info *rp)
 210 {
 211         struct pim_nexthop_cache *pnc = NULL;
 212         struct pim_nexthop_cache lookup;
 213         struct pim_upstream *upstream = NULL;
 214
 215         /* Remove from RPF hash if it is the last entry */
 216         lookup.rpf.rpf_addr = *addr;
 217         pnc = hash_lookup(pim->rpf_hash, &lookup);
 218         if (!pnc) {
 219                 zlog_warn("attempting to delete nonexistent NHT entry %pFX",
 220                           addr);
 221                 return;
 222         }
 223
 224         if (rp) {
 225                 /* Release the (*, G)upstream from pnc->upstream_hash,
 226                  * whose Group belongs to the RP getting deleted
 227                  */
 228                 frr_each (rb_pim_upstream, &pim->upstream_head, upstream) {
 229                         struct prefix grp;
 230                         struct rp_info *trp_info;
 231
 232                         if (!pim_addr_is_any(upstream->sg.src))
 233                                 continue;
 234
 235                         pim_addr_to_prefix(&grp, upstream->sg.grp);
 236                         trp_info = pim_rp_find_match_group(pim, &grp);
 237                         if (trp_info == rp)
 238                                 hash_release(pnc->upstream_hash, upstream);
 239                 }
 240                 listnode_delete(pnc->rp_list, rp);
 241         }
 242
 243         if (up)
 244                 hash_release(pnc->upstream_hash, up);
 245
 246         pim_nht_drop_maybe(pim, pnc);
 247 }
 248
 249 #if PIM_IPV == 4
 250 void pim_nht_bsr_del(struct pim_instance *pim, struct in_addr addr)
 251 {
 252         struct pim_nexthop_cache *pnc = NULL;
 253         struct pim_nexthop_cache lookup;
 254
 255         /*
 256          * Nothing to do here if the address to unregister
 257          * is 0.0.0.0 as that the BSR has not been registered
 258          * for tracking yet.
 259          */
 260         if (addr.s_addr == INADDR_ANY)
 261                 return;
 262
 263         lookup.rpf.rpf_addr.family = AF_INET;
 264         lookup.rpf.rpf_addr.prefixlen = IPV4_MAX_BITLEN;
 265         lookup.rpf.rpf_addr.u.prefix4 = addr;
 266
 267         pnc = hash_lookup(pim->rpf_hash, &lookup);
 268
 269         if (!pnc) {
 270                 zlog_warn("attempting to delete nonexistent NHT BSR entry %pI4",
 271                           &addr);
 272                 return;
 273         }
 274
 275         assertf(pnc->bsr_count > 0, "addr=%pI4", &addr);
 276         pnc->bsr_count--;
 277
 278         pim_nht_drop_maybe(pim, pnc);
 279 }
 280
 281 bool pim_nht_bsr_rpf_check(struct pim_instance *pim, struct in_addr bsr_addr,
 282                            struct interface *src_ifp, pim_addr src_ip)
 283 {
 284         struct pim_nexthop_cache *pnc = NULL;
 285         struct pim_nexthop_cache lookup;
 286         struct pim_neighbor *nbr = NULL;
 287         struct nexthop *nh;
 288         struct interface *ifp;
 289
 290         lookup.rpf.rpf_addr.family = AF_INET;
 291         lookup.rpf.rpf_addr.prefixlen = IPV4_MAX_BITLEN;
 292         lookup.rpf.rpf_addr.u.prefix4 = bsr_addr;
 293
 294         pnc = hash_lookup(pim->rpf_hash, &lookup);
 295         if (!pnc || !CHECK_FLAG(pnc->flags, PIM_NEXTHOP_ANSWER_RECEIVED)) {
 296                 /* BSM from a new freshly registered BSR - do a synchronous
 297                  * zebra query since otherwise we'd drop the first packet,
 298                  * leading to additional delay in picking up BSM data
 299                  */
 300
 301                 /* FIXME: this should really be moved into a generic NHT
 302                  * function that does "add and get immediate result" or maybe
 303                  * "check cache or get immediate result." But until that can
 304                  * be worked in, here's a copy of the code below :(
 305                  */
 306                 struct pim_zlookup_nexthop nexthop_tab[MULTIPATH_NUM];
 307                 ifindex_t i;
 308                 struct interface *ifp = NULL;
 309                 int num_ifindex;
 310
 311                 memset(nexthop_tab, 0, sizeof(nexthop_tab));
 312                 num_ifindex = zclient_lookup_nexthop(pim, nexthop_tab,
 313                                                      MULTIPATH_NUM, bsr_addr,
 314                                                      PIM_NEXTHOP_LOOKUP_MAX);
 315
 316                 if (num_ifindex <= 0)
 317                         return false;
 318
 319                 for (i = 0; i < num_ifindex; i++) {
 320                         struct pim_zlookup_nexthop *znh = &nexthop_tab[i];
 321
 322                         /* pim_zlookup_nexthop has no ->type */
 323
 324                         /* 1:1 match code below with znh instead of nh */
 325                         ifp = if_lookup_by_index(znh->ifindex,
 326                                                  pim->vrf->vrf_id);
 327
 328                         if (!ifp || !ifp->info)
 329                                 continue;
 330
 331                         if (if_is_loopback(ifp) && if_is_loopback(src_ifp))
 332                                 return true;
 333
 334                         nbr = pim_neighbor_find(ifp, znh->nexthop_addr);
 335                         if (!nbr)
 336                                 continue;
 337
 338                         return znh->ifindex == src_ifp->ifindex &&
 339                                (!pim_addr_cmp(znh->nexthop_addr, src_ip));
 340                 }
 341                 return false;
 342         }
 343
 344         if (!CHECK_FLAG(pnc->flags, PIM_NEXTHOP_VALID))
 345                 return false;
 346
 347         /* if we accept BSMs from more than one ECMP nexthop, this will cause
 348          * BSM message "multiplication" for each ECMP hop.  i.e. if you have
 349          * 4-way ECMP and 4 hops you end up with 256 copies of each BSM
 350          * message.
 351          *
 352          * so...  only accept the first (IPv4) valid nexthop as source.
 353          */
 354
 355         for (nh = pnc->nexthop; nh; nh = nh->next) {
 356                 pim_addr nhaddr;
 357
 358                 switch (nh->type) {
 359 #if PIM_IPV == 4
 360                 case NEXTHOP_TYPE_IPV4:
 361                         if (nh->ifindex == IFINDEX_INTERNAL)
 362                                 continue;
 363
 364                         /* fallthru */
 365                 case NEXTHOP_TYPE_IPV4_IFINDEX:
 366                         nhaddr = nh->gate.ipv4;
 367                         break;
 368 #else
 369                 case NEXTHOP_TYPE_IPV6:
 370                         if (nh->ifindex == IFINDEX_INTERNAL)
 371                                 continue;
 372
 373                         /* fallthru */
 374                 case NEXTHOP_TYPE_IPV6_IFINDEX:
 375                         nhaddr = nh->gate.ipv6;
 376                         break;
 377 #endif
 378                 case NEXTHOP_TYPE_IFINDEX:
 379                         nhaddr = bsr_addr;
 380                         break;
 381
 382                 default:
 383                         continue;
 384                 }
 385
 386                 ifp = if_lookup_by_index(nh->ifindex, pim->vrf->vrf_id);
 387                 if (!ifp || !ifp->info)
 388                         continue;
 389
 390                 if (if_is_loopback(ifp) && if_is_loopback(src_ifp))
 391                         return true;
 392
 393                 /* MRIB (IGP) may be pointing at a router where PIM is down */
 394                 nbr = pim_neighbor_find(ifp, nhaddr);
 395                 if (!nbr)
 396                         continue;
 397
 398                 return nh->ifindex == src_ifp->ifindex
 399                        && nhaddr.s_addr == src_ip.s_addr;
 400         }
 401         return false;
 402 }
 403 #endif /* PIM_IPV == 4 */
 404
 405 void pim_rp_nexthop_del(struct rp_info *rp_info)
 406 {
 407         rp_info->rp.source_nexthop.interface = NULL;
 408         rp_info->rp.source_nexthop.mrib_nexthop_addr = PIMADDR_ANY;
 409         rp_info->rp.source_nexthop.mrib_metric_preference =
 410                 router->infinite_assert_metric.metric_preference;
 411         rp_info->rp.source_nexthop.mrib_route_metric =
 412                 router->infinite_assert_metric.route_metric;
 413 }
 414
 415 /* Update RP nexthop info based on Nexthop update received from Zebra.*/
 416 static void pim_update_rp_nh(struct pim_instance *pim,
 417                              struct pim_nexthop_cache *pnc)
 418 {
 419         struct listnode *node = NULL;
 420         struct rp_info *rp_info = NULL;
 421
 422         /*Traverse RP list and update each RP Nexthop info */
 423         for (ALL_LIST_ELEMENTS_RO(pnc->rp_list, node, rp_info)) {
 424                 if (pim_rpf_addr_is_inaddr_any(&rp_info->rp))
 425                         continue;
 426
 427                 // Compute PIM RPF using cached nexthop
 428                 if (!pim_ecmp_nexthop_lookup(pim, &rp_info->rp.source_nexthop,
 429                                              &rp_info->rp.rpf_addr,
 430                                              &rp_info->group, 1))
 431                         pim_rp_nexthop_del(rp_info);
 432         }
 433 }
 434
 435 /* Update Upstream nexthop info based on Nexthop update received from Zebra.*/
 436 static int pim_update_upstream_nh_helper(struct hash_bucket *bucket, void *arg)
 437 {
 438         struct pim_instance *pim = (struct pim_instance *)arg;
 439         struct pim_upstream *up = (struct pim_upstream *)bucket->data;
 440
 441         enum pim_rpf_result rpf_result;
 442         struct pim_rpf old;
 443
 444         old.source_nexthop.interface = up->rpf.source_nexthop.interface;
 445         rpf_result = pim_rpf_update(pim, up, &old, __func__);
 446
 447         /* update kernel multicast forwarding cache (MFC); if the
 448          * RPF nbr is now unreachable the MFC has already been updated
 449          * by pim_rpf_clear
 450          */
 451         if (rpf_result != PIM_RPF_FAILURE)
 452                 pim_upstream_mroute_iif_update(up->channel_oil, __func__);
 453
 454         if (rpf_result == PIM_RPF_CHANGED ||
 455                 (rpf_result == PIM_RPF_FAILURE && old.source_nexthop.interface))
 456                 pim_zebra_upstream_rpf_changed(pim, up, &old);
 457
 458
 459         if (PIM_DEBUG_PIM_NHT) {
 460                 zlog_debug(
 461                         "%s: NHT upstream %s(%s) old ifp %s new ifp %s",
 462                         __func__, up->sg_str, pim->vrf->name,
 463                         old.source_nexthop.interface ? old.source_nexthop
 464                                                                .interface->name
 465                                                      : "Unknown",
 466                         up->rpf.source_nexthop.interface ? up->rpf.source_nexthop
 467                                                                    .interface->name
 468                                                          : "Unknown");
 469         }
 470
 471         return HASHWALK_CONTINUE;
 472 }
 473
 474 static int pim_update_upstream_nh(struct pim_instance *pim,
 475                                   struct pim_nexthop_cache *pnc)
 476 {
 477         hash_walk(pnc->upstream_hash, pim_update_upstream_nh_helper, pim);
 478
 479         pim_zebra_update_all_interfaces(pim);
 480
 481         return 0;
 482 }
 483
 484 uint32_t pim_compute_ecmp_hash(struct prefix *src, struct prefix *grp)
 485 {
 486         uint32_t hash_val;
 487
 488         if (!src)
 489                 return 0;
 490
 491         hash_val = prefix_hash_key(src);
 492         if (grp)
 493                 hash_val ^= prefix_hash_key(grp);
 494         return hash_val;
 495 }
 496
 497 static int pim_ecmp_nexthop_search(struct pim_instance *pim,
 498                                    struct pim_nexthop_cache *pnc,
 499                                    struct pim_nexthop *nexthop,
 500                                    struct prefix *src, struct prefix *grp,
 501                                    int neighbor_needed)
 502 {
 503         struct pim_neighbor *nbrs[MULTIPATH_NUM], *nbr = NULL;
 504         struct interface *ifps[MULTIPATH_NUM];
 505         struct nexthop *nh_node = NULL;
 506         ifindex_t first_ifindex;
 507         struct interface *ifp = NULL;
 508         uint32_t hash_val = 0, mod_val = 0;
 509         uint8_t nh_iter = 0, found = 0;
 510         uint32_t i, num_nbrs = 0;
 511         pim_addr nh_addr = nexthop->mrib_nexthop_addr;
 512         pim_addr src_addr = pim_addr_from_prefix(src);
 513         pim_addr grp_addr = pim_addr_from_prefix(grp);
 514
 515         if (!pnc || !pnc->nexthop_num || !nexthop)
 516                 return 0;
 517
 518         memset(&nbrs, 0, sizeof(nbrs));
 519         memset(&ifps, 0, sizeof(ifps));
 520
 521
 522         // Current Nexthop is VALID, check to stay on the current path.
 523         if (nexthop->interface && nexthop->interface->info &&
 524             (!pim_addr_is_any(nh_addr))) {
 525                 /* User configured knob to explicitly switch
 526                    to new path is disabled or current path
 527                    metric is less than nexthop update.
 528                  */
 529
 530                 if (pim->ecmp_rebalance_enable == 0) {
 531                         uint8_t curr_route_valid = 0;
 532                         // Check if current nexthop is present in new updated
 533                         // Nexthop list.
 534                         // If the current nexthop is not valid, candidate to
 535                         // choose new Nexthop.
 536                         for (nh_node = pnc->nexthop; nh_node;
 537                              nh_node = nh_node->next) {
 538                                 curr_route_valid = (nexthop->interface->ifindex
 539                                                     == nh_node->ifindex);
 540                                 if (curr_route_valid)
 541                                         break;
 542                         }
 543
 544                         if (curr_route_valid &&
 545                             !pim_if_connected_to_source(nexthop->interface,
 546                                                         src_addr)) {
 547                                 nbr = pim_neighbor_find(
 548                                         nexthop->interface,
 549                                         nexthop->mrib_nexthop_addr);
 550                                 if (!nbr
 551                                     && !if_is_loopback(nexthop->interface)) {
 552                                         if (PIM_DEBUG_PIM_NHT)
 553                                                 zlog_debug(
 554                                                         "%s: current nexthop does not have nbr ",
 555                                                         __func__);
 556                                 } else {
 557                                         /* update metric even if the upstream
 558                                          * neighbor stays unchanged
 559                                          */
 560                                         nexthop->mrib_metric_preference =
 561                                                 pnc->distance;
 562                                         nexthop->mrib_route_metric =
 563                                                 pnc->metric;
 564                                         if (PIM_DEBUG_PIM_NHT)
 565                                                 zlog_debug(
 566                                                         "%s: (%pPA,%pPA)(%s) current nexthop %s is valid, skipping new path selection",
 567                                                         __func__, &src_addr,
 568                                                         &grp_addr,
 569                                                         pim->vrf->name,
 570                                                         nexthop->interface->name);
 571                                         return 1;
 572                                 }
 573                         }
 574                 }
 575         }
 576
 577         /*
 578          * Look up all interfaces and neighbors,
 579          * store for later usage
 580          */
 581         for (nh_node = pnc->nexthop, i = 0; nh_node;
 582              nh_node = nh_node->next, i++) {
 583                 ifps[i] =
 584                         if_lookup_by_index(nh_node->ifindex, pim->vrf->vrf_id);
 585                 if (ifps[i]) {
 586 #if PIM_IPV == 4
 587                         pim_addr nhaddr = nh_node->gate.ipv4;
 588 #else
 589                         pim_addr nhaddr = nh_node->gate.ipv6;
 590 #endif
 591                         nbrs[i] = pim_neighbor_find(ifps[i], nhaddr);
 592                         if (nbrs[i] ||
 593                             pim_if_connected_to_source(ifps[i], src_addr))
 594                                 num_nbrs++;
 595                 }
 596         }
 597         if (pim->ecmp_enable) {
 598                 uint32_t consider = pnc->nexthop_num;
 599
 600                 if (neighbor_needed && num_nbrs < consider)
 601                         consider = num_nbrs;
 602
 603                 if (consider == 0)
 604                         return 0;
 605
 606                 // PIM ECMP flag is enable then choose ECMP path.
 607                 hash_val = pim_compute_ecmp_hash(src, grp);
 608                 mod_val = hash_val % consider;
 609         }
 610
 611         for (nh_node = pnc->nexthop; nh_node && (found == 0);
 612              nh_node = nh_node->next) {
 613                 first_ifindex = nh_node->ifindex;
 614                 ifp = ifps[nh_iter];
 615                 if (!ifp) {
 616                         if (PIM_DEBUG_PIM_NHT)
 617                                 zlog_debug(
 618                                         "%s %s: could not find interface for ifindex %d (address %pPA(%s))",
 619                                         __FILE__, __func__, first_ifindex,
 620                                         &src_addr, pim->vrf->name);
 621                         if (nh_iter == mod_val)
 622                                 mod_val++; // Select nexthpath
 623                         nh_iter++;
 624                         continue;
 625                 }
 626                 if (!ifp->info) {
 627                         if (PIM_DEBUG_PIM_NHT)
 628                                 zlog_debug(
 629                                         "%s: multicast not enabled on input interface %s(%s) (ifindex=%d, RPF for source %pPA)",
 630                                         __func__, ifp->name, pim->vrf->name,
 631                                         first_ifindex, &src_addr);
 632                         if (nh_iter == mod_val)
 633                                 mod_val++; // Select nexthpath
 634                         nh_iter++;
 635                         continue;
 636                 }
 637
 638                 if (neighbor_needed &&
 639                     !pim_if_connected_to_source(ifp, src_addr)) {
 640                         nbr = nbrs[nh_iter];
 641                         if (!nbr && !if_is_loopback(ifp)) {
 642                                 if (PIM_DEBUG_PIM_NHT)
 643                                         zlog_debug(
 644                                                 "%s: pim nbr not found on input interface %s(%s)",
 645                                                 __func__, ifp->name,
 646                                                 pim->vrf->name);
 647                                 if (nh_iter == mod_val)
 648                                         mod_val++; // Select nexthpath
 649                                 nh_iter++;
 650                                 continue;
 651                         }
 652                 }
 653
 654                 if (nh_iter == mod_val) {
 655                         nexthop->interface = ifp;
 656 #if PIM_IPV == 4
 657                         nexthop->mrib_nexthop_addr = nh_node->gate.ipv4;
 658 #else
 659                         nexthop->mrib_nexthop_addr = nh_node->gate.ipv6;
 660 #endif
 661                         nexthop->mrib_metric_preference = pnc->distance;
 662                         nexthop->mrib_route_metric = pnc->metric;
 663                         nexthop->last_lookup = src_addr;
 664                         nexthop->last_lookup_time = pim_time_monotonic_usec();
 665                         nexthop->nbr = nbr;
 666                         found = 1;
 667                         if (PIM_DEBUG_PIM_NHT)
 668                                 zlog_debug(
 669                                         "%s: (%pPA,%pPA)(%s) selected nhop interface %s addr %pPAs mod_val %u iter %d ecmp %d",
 670                                         __func__, &src_addr, &grp_addr,
 671                                         pim->vrf->name, ifp->name, &nh_addr,
 672                                         mod_val, nh_iter, pim->ecmp_enable);
 673                 }
 674                 nh_iter++;
 675         }
 676
 677         if (found)
 678                 return 1;
 679         else
 680                 return 0;
 681 }
 682
 683 /* This API is used to parse Registered address nexthop update coming from Zebra
 684  */
 685 int pim_parse_nexthop_update(ZAPI_CALLBACK_ARGS)
 686 {
 687         struct nexthop *nexthop;
 688         struct nexthop *nhlist_head = NULL;
 689         struct nexthop *nhlist_tail = NULL;
 690         int i;
 691         struct pim_rpf rpf;
 692         struct pim_nexthop_cache *pnc = NULL;
 693         struct interface *ifp = NULL;
 694         struct vrf *vrf = vrf_lookup_by_id(vrf_id);
 695         struct pim_instance *pim;
 696         struct zapi_route nhr;
 697         struct prefix match;
 698
 699         if (!vrf)
 700                 return 0;
 701         pim = vrf->info;
 702
 703         if (!zapi_nexthop_update_decode(zclient->ibuf, &match, &nhr)) {
 704                 zlog_err("%s: Decode of nexthop update from zebra failed",
 705                          __func__);
 706                 return 0;
 707         }
 708
 709         if (cmd == ZEBRA_NEXTHOP_UPDATE) {
 710                 prefix_copy(&rpf.rpf_addr, &match);
 711                 pnc = pim_nexthop_cache_find(pim, &rpf);
 712                 if (!pnc) {
 713                         if (PIM_DEBUG_PIM_NHT)
 714                                 zlog_debug(
 715                                         "%s: Skipping NHT update, addr %pFX is not in local cached DB.",
 716                                         __func__, &rpf.rpf_addr);
 717                         return 0;
 718                 }
 719         } else {
 720                 /*
 721                  * We do not currently handle ZEBRA_IMPORT_CHECK_UPDATE
 722                  */
 723                 return 0;
 724         }
 725
 726         pnc->last_update = pim_time_monotonic_usec();
 727
 728         if (nhr.nexthop_num) {
 729                 pnc->nexthop_num = 0; // Only increment for pim enabled rpf.
 730
 731                 for (i = 0; i < nhr.nexthop_num; i++) {
 732                         nexthop = nexthop_from_zapi_nexthop(&nhr.nexthops[i]);
 733                         switch (nexthop->type) {
 734                         case NEXTHOP_TYPE_IFINDEX:
 735                                 /*
 736                                  * Connected route (i.e. no nexthop), use
 737                                  * RPF address from nexthop cache (i.e.
 738                                  * destination) as PIM nexthop.
 739                                  */
 740 #if PIM_IPV == 4
 741                                 nexthop->type = NEXTHOP_TYPE_IPV4_IFINDEX;
 742                                 nexthop->gate.ipv4 =
 743                                         pnc->rpf.rpf_addr.u.prefix4;
 744 #else
 745                                 nexthop->type = NEXTHOP_TYPE_IPV6_IFINDEX;
 746                                 nexthop->gate.ipv6 =
 747                                         pnc->rpf.rpf_addr.u.prefix6;
 748 #endif
 749                                 break;
 750 #if PIM_IPV == 4
 751                         /* RFC5549 IPv4-over-IPv6 nexthop handling:
 752                          * if we get an IPv6 nexthop in IPv4 PIM, hunt down a
 753                          * PIM neighbor and use that instead.
 754                          */
 755                         case NEXTHOP_TYPE_IPV6_IFINDEX: {
 756                                 struct interface *ifp1 = NULL;
 757                                 struct pim_neighbor *nbr = NULL;
 758
 759                                 ifp1 = if_lookup_by_index(nexthop->ifindex,
 760                                                           pim->vrf->vrf_id);
 761
 762                                 if (!ifp1)
 763                                         nbr = NULL;
 764                                 else
 765                                         /* FIXME: should really use nbr's
 766                                          * secondary address list here
 767                                          */
 768                                         nbr = pim_neighbor_find_if(ifp1);
 769
 770                                 /* Overwrite with Nbr address as NH addr */
 771                                 if (nbr)
 772                                         nexthop->gate.ipv4 = nbr->source_addr;
 773                                 else
 774                                         // Mark nexthop address to 0 until PIM
 775                                         // Nbr is resolved.
 776                                         nexthop->gate.ipv4 = PIMADDR_ANY;
 777
 778                                 break;
 779                         }
 780 #else
 781                         case NEXTHOP_TYPE_IPV6_IFINDEX:
 782 #endif
 783                         case NEXTHOP_TYPE_IPV6:
 784                         case NEXTHOP_TYPE_IPV4:
 785                         case NEXTHOP_TYPE_IPV4_IFINDEX:
 786                         case NEXTHOP_TYPE_BLACKHOLE:
 787                                 /* nothing to do for the other nexthop types */
 788                                 break;
 789                         }
 790
 791                         ifp = if_lookup_by_index(nexthop->ifindex,
 792                                                  pim->vrf->vrf_id);
 793                         if (!ifp) {
 794                                 if (PIM_DEBUG_PIM_NHT) {
 795                                         char buf[NEXTHOP_STRLEN];
 796                                         zlog_debug(
 797                                                 "%s: could not find interface for ifindex %d(%s) (addr %s)",
 798                                                 __func__, nexthop->ifindex,
 799                                                 pim->vrf->name,
 800                                                 nexthop2str(nexthop, buf,
 801                                                             sizeof(buf)));
 802                                 }
 803                                 nexthop_free(nexthop);
 804                                 continue;
 805                         }
 806
 807                         if (PIM_DEBUG_PIM_NHT) {
 808 #if PIM_IPV == 4
 809                                 pim_addr nhaddr = nexthop->gate.ipv4;
 810 #else
 811                                 pim_addr nhaddr = nexthop->gate.ipv6;
 812 #endif
 813                                 zlog_debug(
 814                                         "%s: NHT addr %pFX(%s) %d-nhop via %pPA(%s) type %d distance:%u metric:%u ",
 815                                         __func__, &match, pim->vrf->name, i + 1,
 816                                         &nhaddr, ifp->name, nexthop->type,
 817                                         nhr.distance, nhr.metric);
 818                         }
 819
 820                         if (!ifp->info) {
 821                                 /*
 822                                  * Though Multicast is not enabled on this
 823                                  * Interface store it in database otheriwse we
 824                                  * may miss this update and this will not cause
 825                                  * any issue, because while choosing the path we
 826                                  * are ommitting the Interfaces which are not
 827                                  * multicast enabled
 828                                  */
 829                                 if (PIM_DEBUG_PIM_NHT) {
 830                                         char buf[NEXTHOP_STRLEN];
 831
 832                                         zlog_debug(
 833                                                 "%s: multicast not enabled on input interface %s(%s) (ifindex=%d, addr %s)",
 834                                                 __func__, ifp->name,
 835                                                 pim->vrf->name,
 836                                                 nexthop->ifindex,
 837                                                 nexthop2str(nexthop, buf,
 838                                                             sizeof(buf)));
 839                                 }
 840                         }
 841
 842                         if (nhlist_tail) {
 843                                 nhlist_tail->next = nexthop;
 844                                 nhlist_tail = nexthop;
 845                         } else {
 846                                 nhlist_tail = nexthop;
 847                                 nhlist_head = nexthop;
 848                         }
 849                         // Only keep track of nexthops which are PIM enabled.
 850                         pnc->nexthop_num++;
 851                 }
 852                 /* Reset existing pnc->nexthop before assigning new list */
 853                 nexthops_free(pnc->nexthop);
 854                 pnc->nexthop = nhlist_head;
 855                 if (pnc->nexthop_num) {
 856                         pnc->flags |= PIM_NEXTHOP_VALID;
 857                         pnc->distance = nhr.distance;
 858                         pnc->metric = nhr.metric;
 859                 }
 860         } else {
 861                 pnc->flags &= ~PIM_NEXTHOP_VALID;
 862                 pnc->nexthop_num = nhr.nexthop_num;
 863                 nexthops_free(pnc->nexthop);
 864                 pnc->nexthop = NULL;
 865         }
 866         SET_FLAG(pnc->flags, PIM_NEXTHOP_ANSWER_RECEIVED);
 867
 868         if (PIM_DEBUG_PIM_NHT)
 869                 zlog_debug(
 870                         "%s: NHT Update for %pFX(%s) num_nh %d num_pim_nh %d vrf:%u up %ld rp %d",
 871                         __func__, &match, pim->vrf->name, nhr.nexthop_num,
 872                         pnc->nexthop_num, vrf_id, pnc->upstream_hash->count,
 873                         listcount(pnc->rp_list));
 874
 875         pim_rpf_set_refresh_time(pim);
 876
 877         if (listcount(pnc->rp_list))
 878                 pim_update_rp_nh(pim, pnc);
 879         if (pnc->upstream_hash->count)
 880                 pim_update_upstream_nh(pim, pnc);
 881
 882         return 0;
 883 }
 884
 885 int pim_ecmp_nexthop_lookup(struct pim_instance *pim,
 886                             struct pim_nexthop *nexthop, struct prefix *src,
 887                             struct prefix *grp, int neighbor_needed)
 888 {
 889         struct pim_nexthop_cache *pnc;
 890         struct pim_zlookup_nexthop nexthop_tab[MULTIPATH_NUM];
 891         struct pim_neighbor *nbrs[MULTIPATH_NUM], *nbr = NULL;
 892         struct pim_rpf rpf;
 893         int num_ifindex;
 894         struct interface *ifps[MULTIPATH_NUM], *ifp;
 895         int first_ifindex;
 896         int found = 0;
 897         uint8_t i = 0;
 898         uint32_t hash_val = 0, mod_val = 0;
 899         uint32_t num_nbrs = 0;
 900         pim_addr src_addr = pim_addr_from_prefix(src);
 901
 902         if (PIM_DEBUG_PIM_NHT_DETAIL)
 903                 zlog_debug("%s: Looking up: %pPA(%s), last lookup time: %lld",
 904                            __func__, &src_addr, pim->vrf->name,
 905                            nexthop->last_lookup_time);
 906
 907         rpf.rpf_addr = *src;
 908
 909         pnc = pim_nexthop_cache_find(pim, &rpf);
 910         if (pnc) {
 911                 if (CHECK_FLAG(pnc->flags, PIM_NEXTHOP_ANSWER_RECEIVED))
 912                     return pim_ecmp_nexthop_search(pim, pnc, nexthop, src, grp,
 913                                                    neighbor_needed);
 914         }
 915
 916         memset(nexthop_tab, 0,
 917                sizeof(struct pim_zlookup_nexthop) * MULTIPATH_NUM);
 918         num_ifindex = zclient_lookup_nexthop(pim, nexthop_tab, MULTIPATH_NUM,
 919                                              src_addr, PIM_NEXTHOP_LOOKUP_MAX);
 920         if (num_ifindex < 1) {
 921                 if (PIM_DEBUG_PIM_NHT)
 922                         zlog_warn(
 923                                 "%s: could not find nexthop ifindex for address %pPA(%s)",
 924                                 __func__, &src_addr, pim->vrf->name);
 925                 return 0;
 926         }
 927
 928         memset(&nbrs, 0, sizeof(nbrs));
 929         memset(&ifps, 0, sizeof(ifps));
 930
 931         /*
 932          * Look up all interfaces and neighbors,
 933          * store for later usage
 934          */
 935         for (i = 0; i < num_ifindex; i++) {
 936                 ifps[i] = if_lookup_by_index(nexthop_tab[i].ifindex,
 937                                              pim->vrf->vrf_id);
 938                 if (ifps[i]) {
 939                         nbrs[i] = pim_neighbor_find(
 940                                 ifps[i], nexthop_tab[i].nexthop_addr);
 941                         if (nbrs[i] ||
 942                             pim_if_connected_to_source(ifps[i], src_addr))
 943                                 num_nbrs++;
 944                 }
 945         }
 946
 947         // If PIM ECMP enable then choose ECMP path.
 948         if (pim->ecmp_enable) {
 949                 uint32_t consider = num_ifindex;
 950
 951                 if (neighbor_needed && num_nbrs < consider)
 952                         consider = num_nbrs;
 953
 954                 if (consider == 0)
 955                         return 0;
 956
 957                 hash_val = pim_compute_ecmp_hash(src, grp);
 958                 mod_val = hash_val % consider;
 959                 if (PIM_DEBUG_PIM_NHT_DETAIL)
 960                         zlog_debug("%s: hash_val %u mod_val %u", __func__,
 961                                    hash_val, mod_val);
 962         }
 963
 964         i = 0;
 965         while (!found && (i < num_ifindex)) {
 966                 first_ifindex = nexthop_tab[i].ifindex;
 967
 968                 ifp = ifps[i];
 969                 if (!ifp) {
 970                         if (PIM_DEBUG_PIM_NHT)
 971                                 zlog_debug(
 972                                         "%s %s: could not find interface for ifindex %d (address %pPA(%s))",
 973                                         __FILE__, __func__, first_ifindex,
 974                                         &src_addr, pim->vrf->name);
 975                         if (i == mod_val)
 976                                 mod_val++;
 977                         i++;
 978                         continue;
 979                 }
 980
 981                 if (!ifp->info) {
 982                         if (PIM_DEBUG_PIM_NHT)
 983                                 zlog_debug(
 984                                         "%s: multicast not enabled on input interface %s(%s) (ifindex=%d, RPF for source %pPA)",
 985                                         __func__, ifp->name, pim->vrf->name,
 986                                         first_ifindex, &src_addr);
 987                         if (i == mod_val)
 988                                 mod_val++;
 989                         i++;
 990                         continue;
 991                 }
 992                 if (neighbor_needed &&
 993                     !pim_if_connected_to_source(ifp, src_addr)) {
 994                         nbr = nbrs[i];
 995                         if (PIM_DEBUG_PIM_NHT_DETAIL)
 996                                 zlog_debug("ifp name: %s(%s), pim nbr: %p",
 997                                            ifp->name, pim->vrf->name, nbr);
 998                         if (!nbr && !if_is_loopback(ifp)) {
 999                                 if (i == mod_val)
1000                                         mod_val++;
1001                                 if (PIM_DEBUG_PIM_NHT)
1002                                         zlog_debug(
1003                                                 "%s: NBR (%pPA) not found on input interface %s(%s) (RPF for source %pPA)",
1004                                                 __func__,
1005                                                 &nexthop_tab[i].nexthop_addr,
1006                                                 ifp->name, pim->vrf->name,
1007                                                 &src_addr);
1008                                 i++;
1009                                 continue;
1010                         }
1011                 }
1012
1013                 if (i == mod_val) {
1014                         if (PIM_DEBUG_PIM_NHT)
1015                                 zlog_debug(
1016                                         "%s: found nhop %pPA for addr %pPA interface %s(%s) metric %d dist %d",
1017                                         __func__, &nexthop_tab[i].nexthop_addr,
1018                                         &src_addr, ifp->name, pim->vrf->name,
1019                                         nexthop_tab[i].route_metric,
1020                                         nexthop_tab[i].protocol_distance);
1021                         /* update nexthop data */
1022                         nexthop->interface = ifp;
1023                         nexthop->mrib_nexthop_addr =
1024                                 nexthop_tab[i].nexthop_addr;
1025                         nexthop->mrib_metric_preference =
1026                                 nexthop_tab[i].protocol_distance;
1027                         nexthop->mrib_route_metric =
1028                                 nexthop_tab[i].route_metric;
1029                         nexthop->last_lookup = src_addr;
1030                         nexthop->last_lookup_time = pim_time_monotonic_usec();
1031                         nexthop->nbr = nbr;
1032                         found = 1;
1033                 }
1034                 i++;
1035         }
1036
1037         if (found)
1038                 return 1;
1039         else
1040                 return 0;
1041 }
1042
1043 int pim_ecmp_fib_lookup_if_vif_index(struct pim_instance *pim,
1044                                      struct prefix *src, struct prefix *grp)
1045 {
1046         struct pim_nexthop nhop;
1047         int vif_index;
1048         ifindex_t ifindex;
1049         pim_addr src_addr;
1050
1051         if (PIM_DEBUG_PIM_NHT_DETAIL) {
1052                 src_addr = pim_addr_from_prefix(src);
1053         }
1054
1055         memset(&nhop, 0, sizeof(nhop));
1056         if (!pim_ecmp_nexthop_lookup(pim, &nhop, src, grp, 1)) {
1057                 if (PIM_DEBUG_PIM_NHT)
1058                         zlog_debug(
1059                                 "%s: could not find nexthop ifindex for address %pPA(%s)",
1060                                 __func__, &src_addr, pim->vrf->name);
1061                 return -1;
1062         }
1063
1064         ifindex = nhop.interface->ifindex;
1065         if (PIM_DEBUG_PIM_NHT)
1066                 zlog_debug(
1067                         "%s: found nexthop ifindex=%d (interface %s(%s)) for address %pPA",
1068                         __func__, ifindex,
1069                         ifindex2ifname(ifindex, pim->vrf->vrf_id),
1070                         pim->vrf->name, &src_addr);
1071
1072         vif_index = pim_if_find_vifindex_by_ifindex(pim, ifindex);
1073
1074         if (vif_index < 0) {
1075                 if (PIM_DEBUG_PIM_NHT) {
1076                         zlog_debug(
1077                                 "%s: low vif_index=%d(%s) < 1 nexthop for address %pPA",
1078                                 __func__, vif_index, pim->vrf->name, &src_addr);
1079                 }
1080                 return -2;
1081         }
1082
1083         return vif_index;
1084 }