1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2021 Ng Peng Nam Sean 5 * Copyright (c) 2022 Alexander V. Chernikov <melifaro@FreeBSD.org> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 #include <sys/cdefs.h> 30 __FBSDID("$FreeBSD$"); 31 #include "opt_inet.h" 32 #include "opt_inet6.h" 33 #include "opt_route.h" 34 #include <sys/types.h> 35 #include <sys/malloc.h> 36 #include <sys/rmlock.h> 37 #include <sys/socket.h> 38 39 #include <net/if.h> 40 #include <net/route.h> 41 #include <net/route/nhop.h> 42 #include <net/route/route_ctl.h> 43 #include <net/route/route_var.h> 44 #include <netlink/netlink.h> 45 #include <netlink/netlink_ctl.h> 46 #include <netlink/netlink_route.h> 47 #include <netlink/route/route_var.h> 48 49 #define DEBUG_MOD_NAME nl_route 50 #define DEBUG_MAX_LEVEL LOG_DEBUG3 51 #include <netlink/netlink_debug.h> 52 _DECLARE_DEBUG(LOG_DEBUG); 53 54 static unsigned char 55 get_rtm_type(const struct nhop_object *nh) 56 { 57 int nh_flags = nh->nh_flags; 58 59 /* Use the fact that nhg runtime flags are only NHF_MULTIPATH */ 60 if (nh_flags & NHF_BLACKHOLE) 61 return (RTN_BLACKHOLE); 62 else if (nh_flags & NHF_REJECT) 63 return (RTN_PROHIBIT); 64 return (RTN_UNICAST); 65 } 66 67 static uint8_t 68 nl_get_rtm_protocol(const struct nhop_object *nh) 69 { 70 #ifdef ROUTE_MPATH 71 if (NH_IS_NHGRP(nh)) { 72 const struct nhgrp_object *nhg = (const struct nhgrp_object *)nh; 73 uint8_t origin = nhgrp_get_origin(nhg); 74 if (origin != RTPROT_UNSPEC) 75 return (origin); 76 nh = nhg->nhops[0]; 77 } 78 #endif 79 uint8_t origin = nhop_get_origin(nh); 80 if (origin != RTPROT_UNSPEC) 81 return (origin); 82 /* TODO: remove guesswork once all kernel users fill in origin */ 83 int rt_flags = nhop_get_rtflags(nh); 84 if (rt_flags & RTF_PROTO1) 85 return (RTPROT_ZEBRA); 86 if (rt_flags & RTF_STATIC) 87 return (RTPROT_STATIC); 88 return (RTPROT_KERNEL); 89 } 90 91 static int 92 get_rtmsg_type_from_rtsock(int cmd) 93 { 94 switch (cmd) { 95 case RTM_ADD: 96 case RTM_CHANGE: 97 case RTM_GET: 98 return NL_RTM_NEWROUTE; 99 case RTM_DELETE: 100 return NL_RTM_DELROUTE; 101 } 102 103 return (0); 104 } 105 106 /* 107 * fibnum heuristics 108 * 109 * if (dump && rtm_table == 0 && !rta_table) RT_ALL_FIBS 110 * msg rtm_table RTA_TABLE result 111 * RTM_GETROUTE/dump 0 - RT_ALL_FIBS 112 * RTM_GETROUTE/dump 1 - 1 113 * RTM_GETROUTE/get 0 - 0 114 * 115 */ 116 117 static struct nhop_object * 118 rc_get_nhop(const struct rib_cmd_info *rc) 119 { 120 return ((rc->rc_cmd == RTM_DELETE) ? rc->rc_nh_old : rc->rc_nh_new); 121 } 122 123 static void 124 dump_rc_nhop_gw(struct nl_writer *nw, const struct nhop_object *nh) 125 { 126 int upper_family; 127 128 switch (nhop_get_neigh_family(nh)) { 129 case AF_LINK: 130 /* onlink prefix, skip */ 131 break; 132 case AF_INET: 133 nlattr_add(nw, NL_RTA_GATEWAY, 4, &nh->gw4_sa.sin_addr); 134 break; 135 case AF_INET6: 136 upper_family = nhop_get_upper_family(nh); 137 if (upper_family == AF_INET6) { 138 nlattr_add(nw, NL_RTA_GATEWAY, 16, &nh->gw6_sa.sin6_addr); 139 } else if (upper_family == AF_INET) { 140 /* IPv4 over IPv6 */ 141 char buf[20]; 142 struct rtvia *via = (struct rtvia *)&buf[0]; 143 via->rtvia_family = AF_INET6; 144 memcpy(via->rtvia_addr, &nh->gw6_sa.sin6_addr, 16); 145 nlattr_add(nw, NL_RTA_VIA, 17, via); 146 } 147 break; 148 } 149 } 150 151 static void 152 dump_rc_nhop_mtu(struct nl_writer *nw, const struct nhop_object *nh) 153 { 154 int nla_len = sizeof(struct nlattr) * 2 + sizeof(uint32_t); 155 struct nlattr *nla = nlmsg_reserve_data(nw, nla_len, struct nlattr); 156 157 if (nla == NULL) 158 return; 159 nla->nla_type = NL_RTA_METRICS; 160 nla->nla_len = nla_len; 161 nla++; 162 nla->nla_type = NL_RTAX_MTU; 163 nla->nla_len = sizeof(struct nlattr) + sizeof(uint32_t); 164 *((uint32_t *)(nla + 1)) = nh->nh_mtu; 165 } 166 167 #ifdef ROUTE_MPATH 168 static void 169 dump_rc_nhg(struct nl_writer *nw, const struct nhgrp_object *nhg, struct rtmsg *rtm) 170 { 171 uint32_t uidx = nhgrp_get_uidx(nhg); 172 uint32_t num_nhops; 173 const struct weightened_nhop *wn = nhgrp_get_nhops(nhg, &num_nhops); 174 uint32_t base_rtflags = nhop_get_rtflags(wn[0].nh); 175 176 if (uidx != 0) 177 nlattr_add_u32(nw, NL_RTA_NH_ID, uidx); 178 179 nlattr_add_u32(nw, NL_RTA_RTFLAGS, base_rtflags); 180 int off = nlattr_add_nested(nw, NL_RTA_MULTIPATH); 181 if (off == 0) 182 return; 183 184 for (int i = 0; i < num_nhops; i++) { 185 int nh_off = nlattr_save_offset(nw); 186 struct rtnexthop *rtnh = nlmsg_reserve_object(nw, struct rtnexthop); 187 if (rtnh == NULL) 188 return; 189 rtnh->rtnh_flags = 0; 190 rtnh->rtnh_ifindex = wn[i].nh->nh_ifp->if_index; 191 rtnh->rtnh_hops = wn[i].weight; 192 dump_rc_nhop_gw(nw, wn[i].nh); 193 uint32_t rtflags = nhop_get_rtflags(wn[i].nh); 194 if (rtflags != base_rtflags) 195 nlattr_add_u32(nw, NL_RTA_RTFLAGS, rtflags); 196 if (rtflags & RTF_FIXEDMTU) 197 dump_rc_nhop_mtu(nw, wn[i].nh); 198 rtnh = nlattr_restore_offset(nw, nh_off, struct rtnexthop); 199 /* 200 * nlattr_add() allocates 4-byte aligned storage, no need to aligh 201 * length here 202 * */ 203 rtnh->rtnh_len = nlattr_save_offset(nw) - nh_off; 204 } 205 nlattr_set_len(nw, off); 206 } 207 #endif 208 209 static void 210 dump_rc_nhop(struct nl_writer *nw, const struct nhop_object *nh, struct rtmsg *rtm) 211 { 212 #ifdef ROUTE_MPATH 213 if (NH_IS_NHGRP(nh)) { 214 dump_rc_nhg(nw, (const struct nhgrp_object *)nh, rtm); 215 return; 216 } 217 #endif 218 uint32_t rtflags = nhop_get_rtflags(nh); 219 220 /* 221 * IPv4 over IPv6 222 * ('RTA_VIA', {'family': 10, 'addr': 'fe80::20c:29ff:fe67:2dd'}), ('RTA_OIF', 2), 223 * IPv4 w/ gw 224 * ('RTA_GATEWAY', '172.16.107.131'), ('RTA_OIF', 2)], 225 * Direct route: 226 * ('RTA_OIF', 2) 227 */ 228 if (nh->nh_flags & NHF_GATEWAY) 229 dump_rc_nhop_gw(nw, nh); 230 231 uint32_t uidx = nhop_get_uidx(nh); 232 if (uidx != 0) 233 nlattr_add_u32(nw, NL_RTA_NH_ID, uidx); 234 nlattr_add_u32(nw, NL_RTA_KNH_ID, nhop_get_idx(nh)); 235 nlattr_add_u32(nw, NL_RTA_RTFLAGS, rtflags); 236 237 if (rtflags & RTF_FIXEDMTU) 238 dump_rc_nhop_mtu(nw, nh); 239 uint32_t nh_expire = nhop_get_expire(nh); 240 if (nh_expire > 0) 241 nlattr_add_u32(nw, NL_RTA_EXPIRES, nh_expire - time_uptime); 242 243 /* In any case, fill outgoing interface */ 244 nlattr_add_u32(nw, NL_RTA_OIF, nh->nh_ifp->if_index); 245 } 246 247 /* 248 * Dumps output from a rib command into an rtmsg 249 */ 250 251 static int 252 dump_px(uint32_t fibnum, const struct nlmsghdr *hdr, 253 const struct rtentry *rt, struct route_nhop_data *rnd, 254 struct nl_writer *nw) 255 { 256 struct rtmsg *rtm; 257 int error = 0; 258 259 NET_EPOCH_ASSERT(); 260 261 if (!nlmsg_reply(nw, hdr, sizeof(struct rtmsg))) 262 goto enomem; 263 264 int family = rt_get_family(rt); 265 int rtm_off = nlattr_save_offset(nw); 266 rtm = nlmsg_reserve_object(nw, struct rtmsg); 267 rtm->rtm_family = family; 268 rtm->rtm_dst_len = 0; 269 rtm->rtm_src_len = 0; 270 rtm->rtm_tos = 0; 271 if (fibnum < 255) 272 rtm->rtm_table = (unsigned char)fibnum; 273 rtm->rtm_scope = RT_SCOPE_UNIVERSE; 274 rtm->rtm_protocol = nl_get_rtm_protocol(rnd->rnd_nhop); 275 rtm->rtm_type = get_rtm_type(rnd->rnd_nhop); 276 277 nlattr_add_u32(nw, NL_RTA_TABLE, fibnum); 278 279 int plen = 0; 280 #if defined(INET) || defined(INET6) 281 uint32_t scopeid; 282 #endif 283 switch (family) { 284 #ifdef INET 285 case AF_INET: 286 { 287 struct in_addr addr; 288 rt_get_inet_prefix_plen(rt, &addr, &plen, &scopeid); 289 nlattr_add(nw, NL_RTA_DST, 4, &addr); 290 break; 291 } 292 #endif 293 #ifdef INET6 294 case AF_INET6: 295 { 296 struct in6_addr addr; 297 rt_get_inet6_prefix_plen(rt, &addr, &plen, &scopeid); 298 nlattr_add(nw, NL_RTA_DST, 16, &addr); 299 break; 300 } 301 #endif 302 default: 303 FIB_LOG(LOG_NOTICE, fibnum, family, "unsupported rt family: %d", family); 304 error = EAFNOSUPPORT; 305 goto flush; 306 } 307 308 rtm = nlattr_restore_offset(nw, rtm_off, struct rtmsg); 309 if (plen > 0) 310 rtm->rtm_dst_len = plen; 311 dump_rc_nhop(nw, rnd->rnd_nhop, rtm); 312 313 if (nlmsg_end(nw)) 314 return (0); 315 enomem: 316 error = ENOMEM; 317 flush: 318 nlmsg_abort(nw); 319 return (error); 320 } 321 322 static int 323 family_to_group(int family) 324 { 325 switch (family) { 326 case AF_INET: 327 return (RTNLGRP_IPV4_ROUTE); 328 case AF_INET6: 329 return (RTNLGRP_IPV6_ROUTE); 330 } 331 return (0); 332 } 333 334 335 static void 336 report_operation(uint32_t fibnum, struct rib_cmd_info *rc, 337 struct nlpcb *nlp, struct nlmsghdr *hdr) 338 { 339 struct nl_writer nw; 340 341 uint32_t group_id = family_to_group(rt_get_family(rc->rc_rt)); 342 if (nlmsg_get_group_writer(&nw, NLMSG_SMALL, NETLINK_ROUTE, group_id)) { 343 struct route_nhop_data rnd = { 344 .rnd_nhop = rc_get_nhop(rc), 345 .rnd_weight = rc->rc_nh_weight, 346 }; 347 hdr->nlmsg_flags &= ~(NLM_F_REPLACE | NLM_F_CREATE); 348 hdr->nlmsg_flags &= ~(NLM_F_EXCL | NLM_F_APPEND); 349 switch (rc->rc_cmd) { 350 case RTM_ADD: 351 hdr->nlmsg_type = NL_RTM_NEWROUTE; 352 hdr->nlmsg_flags |= NLM_F_CREATE | NLM_F_EXCL; 353 break; 354 case RTM_CHANGE: 355 hdr->nlmsg_type = NL_RTM_NEWROUTE; 356 hdr->nlmsg_flags |= NLM_F_REPLACE; 357 break; 358 case RTM_DELETE: 359 hdr->nlmsg_type = NL_RTM_DELROUTE; 360 break; 361 } 362 dump_px(fibnum, hdr, rc->rc_rt, &rnd, &nw); 363 nlmsg_flush(&nw); 364 } 365 366 rtsock_callback_p->route_f(fibnum, rc); 367 } 368 369 struct rta_mpath_nh { 370 struct sockaddr *gw; 371 struct ifnet *ifp; 372 uint8_t rtnh_flags; 373 uint8_t rtnh_weight; 374 }; 375 376 #define _IN(_field) offsetof(struct rtnexthop, _field) 377 #define _OUT(_field) offsetof(struct rta_mpath_nh, _field) 378 const static struct nlattr_parser nla_p_rtnh[] = { 379 { .type = NL_RTA_GATEWAY, .off = _OUT(gw), .cb = nlattr_get_ip }, 380 { .type = NL_RTA_VIA, .off = _OUT(gw), .cb = nlattr_get_ipvia }, 381 }; 382 const static struct nlfield_parser nlf_p_rtnh[] = { 383 { .off_in = _IN(rtnh_flags), .off_out = _OUT(rtnh_flags), .cb = nlf_get_u8 }, 384 { .off_in = _IN(rtnh_hops), .off_out = _OUT(rtnh_weight), .cb = nlf_get_u8 }, 385 { .off_in = _IN(rtnh_ifindex), .off_out = _OUT(ifp), .cb = nlf_get_ifpz }, 386 }; 387 #undef _IN 388 #undef _OUT 389 NL_DECLARE_PARSER(mpath_parser, struct rtnexthop, nlf_p_rtnh, nla_p_rtnh); 390 391 struct rta_mpath { 392 int num_nhops; 393 struct rta_mpath_nh nhops[0]; 394 }; 395 396 static int 397 nlattr_get_multipath(struct nlattr *nla, struct nl_pstate *npt, const void *arg, void *target) 398 { 399 int data_len = nla->nla_len - sizeof(struct nlattr); 400 struct rtnexthop *rtnh; 401 402 int max_nhops = data_len / sizeof(struct rtnexthop); 403 404 struct rta_mpath *mp = npt_alloc(npt, (max_nhops + 2) * sizeof(struct rta_mpath_nh)); 405 mp->num_nhops = 0; 406 407 for (rtnh = (struct rtnexthop *)(nla + 1); data_len > 0; ) { 408 struct rta_mpath_nh *mpnh = &mp->nhops[mp->num_nhops++]; 409 410 int error = nl_parse_header(rtnh, rtnh->rtnh_len, &mpath_parser, 411 npt, mpnh); 412 if (error != 0) { 413 NLMSG_REPORT_ERR_MSG(npt, "RTA_MULTIPATH: nexhop %d: parse failed", 414 mp->num_nhops - 1); 415 return (error); 416 } 417 418 int len = NL_ITEM_ALIGN(rtnh->rtnh_len); 419 data_len -= len; 420 rtnh = (struct rtnexthop *)((char *)rtnh + len); 421 } 422 if (data_len != 0 || mp->num_nhops == 0) { 423 NLMSG_REPORT_ERR_MSG(npt, "invalid RTA_MULTIPATH attr"); 424 return (EINVAL); 425 } 426 427 *((struct rta_mpath **)target) = mp; 428 return (0); 429 } 430 431 432 struct nl_parsed_route { 433 struct sockaddr *rta_dst; 434 struct sockaddr *rta_gw; 435 struct ifnet *rta_oif; 436 struct rta_mpath *rta_multipath; 437 uint32_t rta_table; 438 uint32_t rta_rtflags; 439 uint32_t rta_nh_id; 440 uint32_t rtax_mtu; 441 uint8_t rtm_family; 442 uint8_t rtm_dst_len; 443 uint8_t rtm_protocol; 444 }; 445 446 #define _IN(_field) offsetof(struct rtmsg, _field) 447 #define _OUT(_field) offsetof(struct nl_parsed_route, _field) 448 static struct nlattr_parser nla_p_rtmetrics[] = { 449 { .type = NL_RTAX_MTU, .off = _OUT(rtax_mtu), .cb = nlattr_get_uint32 }, 450 }; 451 NL_DECLARE_ATTR_PARSER(metrics_parser, nla_p_rtmetrics); 452 453 static const struct nlattr_parser nla_p_rtmsg[] = { 454 { .type = NL_RTA_DST, .off = _OUT(rta_dst), .cb = nlattr_get_ip }, 455 { .type = NL_RTA_OIF, .off = _OUT(rta_oif), .cb = nlattr_get_ifp }, 456 { .type = NL_RTA_GATEWAY, .off = _OUT(rta_gw), .cb = nlattr_get_ip }, 457 { .type = NL_RTA_METRICS, .arg = &metrics_parser, .cb = nlattr_get_nested }, 458 { .type = NL_RTA_MULTIPATH, .off = _OUT(rta_multipath), .cb = nlattr_get_multipath }, 459 { .type = NL_RTA_RTFLAGS, .off = _OUT(rta_rtflags), .cb = nlattr_get_uint32 }, 460 { .type = NL_RTA_TABLE, .off = _OUT(rta_table), .cb = nlattr_get_uint32 }, 461 { .type = NL_RTA_VIA, .off = _OUT(rta_gw), .cb = nlattr_get_ipvia }, 462 { .type = NL_RTA_NH_ID, .off = _OUT(rta_nh_id), .cb = nlattr_get_uint32 }, 463 }; 464 465 static const struct nlfield_parser nlf_p_rtmsg[] = { 466 {.off_in = _IN(rtm_family), .off_out = _OUT(rtm_family), .cb = nlf_get_u8 }, 467 {.off_in = _IN(rtm_dst_len), .off_out = _OUT(rtm_dst_len), .cb = nlf_get_u8 }, 468 {.off_in = _IN(rtm_protocol), .off_out = _OUT(rtm_protocol), .cb = nlf_get_u8 }, 469 }; 470 #undef _IN 471 #undef _OUT 472 NL_DECLARE_PARSER(rtm_parser, struct rtmsg, nlf_p_rtmsg, nla_p_rtmsg); 473 474 struct netlink_walkargs { 475 struct nl_writer *nw; 476 struct route_nhop_data rnd; 477 struct nlmsghdr hdr; 478 struct nlpcb *nlp; 479 uint32_t fibnum; 480 int family; 481 int error; 482 int count; 483 int dumped; 484 int dumped_tables; 485 }; 486 487 static int 488 dump_rtentry(struct rtentry *rt, void *_arg) 489 { 490 struct netlink_walkargs *wa = (struct netlink_walkargs *)_arg; 491 int error; 492 493 wa->count++; 494 if (wa->error != 0) 495 return (0); 496 wa->dumped++; 497 498 rt_get_rnd(rt, &wa->rnd); 499 500 error = dump_px(wa->fibnum, &wa->hdr, rt, &wa->rnd, wa->nw); 501 502 IF_DEBUG_LEVEL(LOG_DEBUG3) { 503 char rtbuf[INET6_ADDRSTRLEN + 5]; 504 FIB_LOG(LOG_DEBUG3, wa->fibnum, wa->family, 505 "Dump %s, offset %u, error %d", 506 rt_print_buf(rt, rtbuf, sizeof(rtbuf)), 507 wa->nw->offset, error); 508 } 509 wa->error = error; 510 511 return (0); 512 } 513 514 static void 515 dump_rtable_one(struct netlink_walkargs *wa, uint32_t fibnum, int family) 516 { 517 FIB_LOG(LOG_DEBUG2, fibnum, family, "Start dump"); 518 wa->count = 0; 519 wa->dumped = 0; 520 521 rib_walk(fibnum, family, false, dump_rtentry, wa); 522 523 wa->dumped_tables++; 524 525 FIB_LOG(LOG_DEBUG2, fibnum, family, "End dump, iterated %d dumped %d", 526 wa->count, wa->dumped); 527 NL_LOG(LOG_DEBUG2, "Current offset: %d", wa->nw->offset); 528 } 529 530 static int 531 dump_rtable_fib(struct netlink_walkargs *wa, uint32_t fibnum, int family) 532 { 533 wa->fibnum = fibnum; 534 535 if (family == AF_UNSPEC) { 536 for (int i = 0; i < AF_MAX; i++) { 537 if (rt_tables_get_rnh(fibnum, i) != 0) { 538 wa->family = i; 539 dump_rtable_one(wa, fibnum, i); 540 if (wa->error != 0) 541 break; 542 } 543 } 544 } else { 545 if (rt_tables_get_rnh(fibnum, family) != 0) { 546 wa->family = family; 547 dump_rtable_one(wa, fibnum, family); 548 } 549 } 550 551 return (wa->error); 552 } 553 554 static int 555 handle_rtm_getroute(struct nlpcb *nlp, struct nl_parsed_route *attrs, 556 struct nlmsghdr *hdr, struct nl_pstate *npt) 557 { 558 RIB_RLOCK_TRACKER; 559 struct rib_head *rnh; 560 struct rtentry *rt; 561 uint32_t fibnum = attrs->rta_table; 562 sa_family_t family = attrs->rtm_family; 563 564 if (attrs->rta_dst == NULL) { 565 NLMSG_REPORT_ERR_MSG(npt, "No RTA_DST supplied"); 566 return (EINVAL); 567 } 568 569 FIB_LOG(LOG_DEBUG, fibnum, family, "getroute called"); 570 571 rnh = rt_tables_get_rnh(fibnum, family); 572 if (rnh == NULL) 573 return (EAFNOSUPPORT); 574 575 RIB_RLOCK(rnh); 576 577 rt = (struct rtentry *)rnh->rnh_matchaddr(attrs->rta_dst, &rnh->head); 578 if (rt == NULL) { 579 RIB_RUNLOCK(rnh); 580 return (ESRCH); 581 } 582 583 struct route_nhop_data rnd; 584 rt_get_rnd(rt, &rnd); 585 rnd.rnd_nhop = nhop_select_func(rnd.rnd_nhop, 0); 586 587 RIB_RUNLOCK(rnh); 588 589 IF_DEBUG_LEVEL(LOG_DEBUG2) { 590 char rtbuf[NHOP_PRINT_BUFSIZE] __unused, nhbuf[NHOP_PRINT_BUFSIZE] __unused; 591 FIB_LOG(LOG_DEBUG2, fibnum, family, "getroute completed: got %s for %s", 592 nhop_print_buf_any(rnd.rnd_nhop, nhbuf, sizeof(nhbuf)), 593 rt_print_buf(rt, rtbuf, sizeof(rtbuf))); 594 } 595 596 hdr->nlmsg_type = NL_RTM_NEWROUTE; 597 dump_px(fibnum, hdr, rt, &rnd, npt->nw); 598 599 return (0); 600 } 601 602 static int 603 handle_rtm_dump(struct nlpcb *nlp, uint32_t fibnum, int family, 604 struct nlmsghdr *hdr, struct nl_writer *nw) 605 { 606 struct netlink_walkargs wa = { 607 .nlp = nlp, 608 .nw = nw, 609 .hdr.nlmsg_pid = hdr->nlmsg_pid, 610 .hdr.nlmsg_seq = hdr->nlmsg_seq, 611 .hdr.nlmsg_type = NL_RTM_NEWROUTE, 612 .hdr.nlmsg_flags = hdr->nlmsg_flags | NLM_F_MULTI, 613 }; 614 615 if (fibnum == RT_TABLE_UNSPEC) { 616 for (int i = 0; i < V_rt_numfibs; i++) { 617 dump_rtable_fib(&wa, fibnum, family); 618 if (wa.error != 0) 619 break; 620 } 621 } else 622 dump_rtable_fib(&wa, fibnum, family); 623 624 if (wa.error == 0 && wa.dumped_tables == 0) { 625 FIB_LOG(LOG_DEBUG, fibnum, family, "incorrect fibnum/family"); 626 wa.error = ESRCH; 627 // How do we propagate it? 628 } 629 630 if (!nlmsg_end_dump(wa.nw, wa.error, &wa.hdr)) { 631 NL_LOG(LOG_DEBUG, "Unable to finalize the dump"); 632 return (ENOMEM); 633 } 634 635 return (wa.error); 636 } 637 638 static struct nhop_object * 639 finalize_nhop(struct nhop_object *nh, int *perror) 640 { 641 /* 642 * The following MUST be filled: 643 * nh_ifp, nh_ifa, nh_gw 644 */ 645 if (nh->gw_sa.sa_family == 0) { 646 /* 647 * Empty gateway. Can be direct route with RTA_OIF set. 648 */ 649 if (nh->nh_ifp != NULL) 650 nhop_set_direct_gw(nh, nh->nh_ifp); 651 else { 652 NL_LOG(LOG_DEBUG, "empty gateway and interface, skipping"); 653 *perror = EINVAL; 654 return (NULL); 655 } 656 /* Both nh_ifp and gateway are set */ 657 } else { 658 /* Gateway is set up, we can derive ifp if not set */ 659 if (nh->nh_ifp == NULL) { 660 struct ifaddr *ifa = ifa_ifwithnet(&nh->gw_sa, 1, nhop_get_fibnum(nh)); 661 if (ifa == NULL) { 662 NL_LOG(LOG_DEBUG, "Unable to determine ifp, skipping"); 663 *perror = EINVAL; 664 return (NULL); 665 } 666 nhop_set_transmit_ifp(nh, ifa->ifa_ifp); 667 } 668 } 669 /* Both nh_ifp and gateway are set */ 670 if (nh->nh_ifa == NULL) { 671 struct ifaddr *ifa = ifaof_ifpforaddr(&nh->gw_sa, nh->nh_ifp); 672 if (ifa == NULL) { 673 NL_LOG(LOG_DEBUG, "Unable to determine ifa, skipping"); 674 *perror = EINVAL; 675 return (NULL); 676 } 677 nhop_set_src(nh, ifa); 678 } 679 680 return (nhop_get_nhop(nh, perror)); 681 } 682 683 static int 684 get_pxflag(const struct nl_parsed_route *attrs) 685 { 686 int pxflag = 0; 687 switch (attrs->rtm_family) { 688 case AF_INET: 689 if (attrs->rtm_dst_len == 32) 690 pxflag = NHF_HOST; 691 else if (attrs->rtm_dst_len == 0) 692 pxflag = NHF_DEFAULT; 693 break; 694 case AF_INET6: 695 if (attrs->rtm_dst_len == 32) 696 pxflag = NHF_HOST; 697 else if (attrs->rtm_dst_len == 0) 698 pxflag = NHF_DEFAULT; 699 break; 700 } 701 702 return (pxflag); 703 } 704 705 static int 706 get_op_flags(int nlm_flags) 707 { 708 int op_flags = 0; 709 710 op_flags |= (nlm_flags & NLM_F_REPLACE) ? RTM_F_REPLACE : 0; 711 op_flags |= (nlm_flags & NLM_F_EXCL) ? RTM_F_EXCL : 0; 712 op_flags |= (nlm_flags & NLM_F_CREATE) ? RTM_F_CREATE : 0; 713 op_flags |= (nlm_flags & NLM_F_APPEND) ? RTM_F_APPEND : 0; 714 715 return (op_flags); 716 } 717 718 #ifdef ROUTE_MPATH 719 static int 720 create_nexthop_one(struct nl_parsed_route *attrs, struct rta_mpath_nh *mpnh, 721 struct nl_pstate *npt, struct nhop_object **pnh) 722 { 723 int error; 724 725 if (mpnh->gw == NULL) 726 return (EINVAL); 727 728 struct nhop_object *nh = nhop_alloc(attrs->rta_table, attrs->rtm_family); 729 if (nh == NULL) 730 return (ENOMEM); 731 732 nhop_set_gw(nh, mpnh->gw, true); 733 if (mpnh->ifp != NULL) 734 nhop_set_transmit_ifp(nh, mpnh->ifp); 735 nhop_set_rtflags(nh, attrs->rta_rtflags); 736 if (attrs->rtm_protocol > RTPROT_STATIC) 737 nhop_set_origin(nh, attrs->rtm_protocol); 738 739 *pnh = finalize_nhop(nh, &error); 740 741 return (error); 742 } 743 #endif 744 745 static struct nhop_object * 746 create_nexthop_from_attrs(struct nl_parsed_route *attrs, 747 struct nl_pstate *npt, int *perror) 748 { 749 struct nhop_object *nh = NULL; 750 int error = 0; 751 752 if (attrs->rta_multipath != NULL) { 753 #ifdef ROUTE_MPATH 754 /* Multipath w/o explicit nexthops */ 755 int num_nhops = attrs->rta_multipath->num_nhops; 756 struct weightened_nhop *wn = npt_alloc(npt, sizeof(*wn) * num_nhops); 757 758 for (int i = 0; i < num_nhops; i++) { 759 struct rta_mpath_nh *mpnh = &attrs->rta_multipath->nhops[i]; 760 761 error = create_nexthop_one(attrs, mpnh, npt, &wn[i].nh); 762 if (error != 0) { 763 for (int j = 0; j < i; j++) 764 nhop_free(wn[j].nh); 765 break; 766 } 767 wn[i].weight = mpnh->rtnh_weight > 0 ? mpnh->rtnh_weight : 1; 768 } 769 if (error == 0) { 770 struct rib_head *rh = nhop_get_rh(wn[0].nh); 771 struct nhgrp_object *nhg; 772 773 nhg = nhgrp_alloc(rh->rib_fibnum, rh->rib_family, 774 wn, num_nhops, perror); 775 if (nhg != NULL) { 776 if (attrs->rtm_protocol > RTPROT_STATIC) 777 nhgrp_set_origin(nhg, attrs->rtm_protocol); 778 nhg = nhgrp_get_nhgrp(nhg, perror); 779 } 780 for (int i = 0; i < num_nhops; i++) 781 nhop_free(wn[i].nh); 782 if (nhg != NULL) 783 return ((struct nhop_object *)nhg); 784 error = *perror; 785 } 786 #else 787 error = ENOTSUP; 788 #endif 789 *perror = error; 790 } else { 791 nh = nhop_alloc(attrs->rta_table, attrs->rtm_family); 792 if (nh == NULL) { 793 *perror = ENOMEM; 794 return (NULL); 795 } 796 if (attrs->rta_gw != NULL) 797 nhop_set_gw(nh, attrs->rta_gw, true); 798 if (attrs->rta_oif != NULL) 799 nhop_set_transmit_ifp(nh, attrs->rta_oif); 800 if (attrs->rtax_mtu != 0) 801 nhop_set_mtu(nh, attrs->rtax_mtu, true); 802 if (attrs->rta_rtflags & RTF_BROADCAST) 803 nhop_set_broadcast(nh, true); 804 if (attrs->rta_rtflags & RTF_BLACKHOLE) 805 nhop_set_blackhole(nh, NHF_BLACKHOLE); 806 if (attrs->rta_rtflags & RTF_REJECT) 807 nhop_set_blackhole(nh, NHF_REJECT); 808 nhop_set_rtflags(nh, attrs->rta_rtflags); 809 if (attrs->rtm_protocol > RTPROT_STATIC) 810 nhop_set_origin(nh, attrs->rtm_protocol); 811 nh = finalize_nhop(nh, perror); 812 } 813 814 return (nh); 815 } 816 817 static int 818 rtnl_handle_newroute(struct nlmsghdr *hdr, struct nlpcb *nlp, 819 struct nl_pstate *npt) 820 { 821 struct rib_cmd_info rc = {}; 822 struct nhop_object *nh = NULL; 823 int error; 824 825 struct nl_parsed_route attrs = {}; 826 error = nl_parse_nlmsg(hdr, &rtm_parser, npt, &attrs); 827 if (error != 0) 828 return (error); 829 830 /* Check if we have enough data */ 831 if (attrs.rta_dst == NULL) { 832 NL_LOG(LOG_DEBUG, "missing RTA_DST"); 833 return (EINVAL); 834 } 835 836 if (attrs.rta_nh_id != 0) { 837 /* Referenced uindex */ 838 int pxflag = get_pxflag(&attrs); 839 nh = nl_find_nhop(attrs.rta_table, attrs.rtm_family, attrs.rta_nh_id, 840 pxflag, &error); 841 if (error != 0) 842 return (error); 843 } else { 844 nh = create_nexthop_from_attrs(&attrs, npt, &error); 845 if (error != 0) { 846 NL_LOG(LOG_DEBUG, "Error creating nexthop"); 847 return (error); 848 } 849 } 850 851 int weight = NH_IS_NHGRP(nh) ? 0 : RT_DEFAULT_WEIGHT; 852 struct route_nhop_data rnd = { .rnd_nhop = nh, .rnd_weight = weight }; 853 int op_flags = get_op_flags(hdr->nlmsg_flags); 854 855 error = rib_add_route_px(attrs.rta_table, attrs.rta_dst, attrs.rtm_dst_len, 856 &rnd, op_flags, &rc); 857 if (error == 0) 858 report_operation(attrs.rta_table, &rc, nlp, hdr); 859 return (error); 860 } 861 862 static int 863 path_match_func(const struct rtentry *rt, const struct nhop_object *nh, void *_data) 864 { 865 struct nl_parsed_route *attrs = (struct nl_parsed_route *)_data; 866 867 if ((attrs->rta_gw != NULL) && !rib_match_gw(rt, nh, attrs->rta_gw)) 868 return (0); 869 870 if ((attrs->rta_oif != NULL) && (attrs->rta_oif != nh->nh_ifp)) 871 return (0); 872 873 return (1); 874 } 875 876 static int 877 rtnl_handle_delroute(struct nlmsghdr *hdr, struct nlpcb *nlp, 878 struct nl_pstate *npt) 879 { 880 struct rib_cmd_info rc; 881 int error; 882 883 struct nl_parsed_route attrs = {}; 884 error = nl_parse_nlmsg(hdr, &rtm_parser, npt, &attrs); 885 if (error != 0) 886 return (error); 887 888 if (attrs.rta_dst == NULL) { 889 NLMSG_REPORT_ERR_MSG(npt, "RTA_DST is not set"); 890 return (ESRCH); 891 } 892 893 error = rib_del_route_px(attrs.rta_table, attrs.rta_dst, 894 attrs.rtm_dst_len, path_match_func, &attrs, 0, &rc); 895 if (error == 0) 896 report_operation(attrs.rta_table, &rc, nlp, hdr); 897 return (error); 898 } 899 900 static int 901 rtnl_handle_getroute(struct nlmsghdr *hdr, struct nlpcb *nlp, struct nl_pstate *npt) 902 { 903 int error; 904 905 struct nl_parsed_route attrs = {}; 906 error = nl_parse_nlmsg(hdr, &rtm_parser, npt, &attrs); 907 if (error != 0) 908 return (error); 909 910 if (hdr->nlmsg_flags & NLM_F_DUMP) 911 error = handle_rtm_dump(nlp, attrs.rta_table, attrs.rtm_family, hdr, npt->nw); 912 else 913 error = handle_rtm_getroute(nlp, &attrs, hdr, npt); 914 915 return (error); 916 } 917 918 void 919 rtnl_handle_route_event(uint32_t fibnum, const struct rib_cmd_info *rc) 920 { 921 int family, nlm_flags = 0; 922 923 struct nl_writer nw; 924 925 family = rt_get_family(rc->rc_rt); 926 927 /* XXX: check if there are active listeners first */ 928 929 /* TODO: consider passing PID/type/seq */ 930 switch (rc->rc_cmd) { 931 case RTM_ADD: 932 nlm_flags = NLM_F_EXCL | NLM_F_CREATE; 933 break; 934 case RTM_CHANGE: 935 nlm_flags = NLM_F_REPLACE; 936 break; 937 case RTM_DELETE: 938 nlm_flags = 0; 939 break; 940 } 941 IF_DEBUG_LEVEL(LOG_DEBUG2) { 942 char rtbuf[NHOP_PRINT_BUFSIZE] __unused; 943 FIB_LOG(LOG_DEBUG2, fibnum, family, 944 "received event %s for %s / nlm_flags=%X", 945 rib_print_cmd(rc->rc_cmd), 946 rt_print_buf(rc->rc_rt, rtbuf, sizeof(rtbuf)), 947 nlm_flags); 948 } 949 950 struct nlmsghdr hdr = { 951 .nlmsg_flags = nlm_flags, 952 .nlmsg_type = get_rtmsg_type_from_rtsock(rc->rc_cmd), 953 }; 954 955 struct route_nhop_data rnd = { 956 .rnd_nhop = rc_get_nhop(rc), 957 .rnd_weight = rc->rc_nh_weight, 958 }; 959 960 uint32_t group_id = family_to_group(family); 961 if (!nlmsg_get_group_writer(&nw, NLMSG_SMALL, NETLINK_ROUTE, group_id)) { 962 NL_LOG(LOG_DEBUG, "error allocating event buffer"); 963 return; 964 } 965 966 dump_px(fibnum, &hdr, rc->rc_rt, &rnd, &nw); 967 nlmsg_flush(&nw); 968 } 969 970 static const struct rtnl_cmd_handler cmd_handlers[] = { 971 { 972 .cmd = NL_RTM_GETROUTE, 973 .name = "RTM_GETROUTE", 974 .cb = &rtnl_handle_getroute, 975 }, 976 { 977 .cmd = NL_RTM_DELROUTE, 978 .name = "RTM_DELROUTE", 979 .cb = &rtnl_handle_delroute, 980 .priv = PRIV_NET_ROUTE, 981 }, 982 { 983 .cmd = NL_RTM_NEWROUTE, 984 .name = "RTM_NEWROUTE", 985 .cb = &rtnl_handle_newroute, 986 .priv = PRIV_NET_ROUTE, 987 } 988 }; 989 990 static const struct nlhdr_parser *all_parsers[] = {&mpath_parser, &metrics_parser, &rtm_parser}; 991 992 void 993 rtnl_routes_init(void) 994 { 995 NL_VERIFY_PARSERS(all_parsers); 996 rtnl_register_messages(cmd_handlers, NL_ARRAY_LEN(cmd_handlers)); 997 } 998