1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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 "opt_netlink.h" 30 31 #include <sys/cdefs.h> 32 __FBSDID("$FreeBSD$"); 33 #include "opt_inet.h" 34 #include "opt_inet6.h" 35 #include "opt_route.h" 36 #include <sys/types.h> 37 #include <sys/malloc.h> 38 #include <sys/rmlock.h> 39 #include <sys/socket.h> 40 41 #include <net/if.h> 42 #include <net/route.h> 43 #include <net/route/nhop.h> 44 #include <net/route/route_ctl.h> 45 #include <net/route/route_var.h> 46 #include <netinet6/scope6_var.h> 47 #include <netlink/netlink.h> 48 #include <netlink/netlink_ctl.h> 49 #include <netlink/netlink_route.h> 50 #include <netlink/route/route_var.h> 51 52 #define DEBUG_MOD_NAME nl_route 53 #define DEBUG_MAX_LEVEL LOG_DEBUG3 54 #include <netlink/netlink_debug.h> 55 _DECLARE_DEBUG(LOG_INFO); 56 57 static unsigned char 58 get_rtm_type(const struct nhop_object *nh) 59 { 60 int nh_flags = nh->nh_flags; 61 62 /* Use the fact that nhg runtime flags are only NHF_MULTIPATH */ 63 if (nh_flags & NHF_BLACKHOLE) 64 return (RTN_BLACKHOLE); 65 else if (nh_flags & NHF_REJECT) 66 return (RTN_PROHIBIT); 67 return (RTN_UNICAST); 68 } 69 70 static uint8_t 71 nl_get_rtm_protocol(const struct nhop_object *nh) 72 { 73 #ifdef ROUTE_MPATH 74 if (NH_IS_NHGRP(nh)) { 75 const struct nhgrp_object *nhg = (const struct nhgrp_object *)nh; 76 uint8_t origin = nhgrp_get_origin(nhg); 77 if (origin != RTPROT_UNSPEC) 78 return (origin); 79 nh = nhg->nhops[0]; 80 } 81 #endif 82 uint8_t origin = nhop_get_origin(nh); 83 if (origin != RTPROT_UNSPEC) 84 return (origin); 85 /* TODO: remove guesswork once all kernel users fill in origin */ 86 int rt_flags = nhop_get_rtflags(nh); 87 if (rt_flags & RTF_PROTO1) 88 return (RTPROT_ZEBRA); 89 if (rt_flags & RTF_STATIC) 90 return (RTPROT_STATIC); 91 return (RTPROT_KERNEL); 92 } 93 94 static int 95 get_rtmsg_type_from_rtsock(int cmd) 96 { 97 switch (cmd) { 98 case RTM_ADD: 99 case RTM_CHANGE: 100 case RTM_GET: 101 return NL_RTM_NEWROUTE; 102 case RTM_DELETE: 103 return NL_RTM_DELROUTE; 104 } 105 106 return (0); 107 } 108 109 /* 110 * fibnum heuristics 111 * 112 * if (dump && rtm_table == 0 && !rta_table) RT_ALL_FIBS 113 * msg rtm_table RTA_TABLE result 114 * RTM_GETROUTE/dump 0 - RT_ALL_FIBS 115 * RTM_GETROUTE/dump 1 - 1 116 * RTM_GETROUTE/get 0 - 0 117 * 118 */ 119 120 static struct nhop_object * 121 rc_get_nhop(const struct rib_cmd_info *rc) 122 { 123 return ((rc->rc_cmd == RTM_DELETE) ? rc->rc_nh_old : rc->rc_nh_new); 124 } 125 126 static void 127 dump_rc_nhop_gw(struct nl_writer *nw, const struct nhop_object *nh) 128 { 129 #ifdef INET6 130 int upper_family; 131 #endif 132 133 switch (nhop_get_neigh_family(nh)) { 134 case AF_LINK: 135 /* onlink prefix, skip */ 136 break; 137 case AF_INET: 138 nlattr_add(nw, NL_RTA_GATEWAY, 4, &nh->gw4_sa.sin_addr); 139 break; 140 #ifdef INET6 141 case AF_INET6: 142 upper_family = nhop_get_upper_family(nh); 143 if (upper_family == AF_INET6) { 144 struct in6_addr gw6 = nh->gw6_sa.sin6_addr; 145 in6_clearscope(&gw6); 146 147 nlattr_add(nw, NL_RTA_GATEWAY, 16, &gw6); 148 } else if (upper_family == AF_INET) { 149 /* IPv4 over IPv6 */ 150 struct in6_addr gw6 = nh->gw6_sa.sin6_addr; 151 in6_clearscope(&gw6); 152 153 char buf[20]; 154 struct rtvia *via = (struct rtvia *)&buf[0]; 155 via->rtvia_family = AF_INET6; 156 memcpy(via->rtvia_addr, &gw6, 16); 157 nlattr_add(nw, NL_RTA_VIA, 17, via); 158 } 159 break; 160 #endif 161 } 162 } 163 164 static void 165 dump_rc_nhop_mtu(struct nl_writer *nw, const struct nhop_object *nh) 166 { 167 int nla_len = sizeof(struct nlattr) * 2 + sizeof(uint32_t); 168 struct nlattr *nla = nlmsg_reserve_data(nw, nla_len, struct nlattr); 169 170 if (nla == NULL) 171 return; 172 nla->nla_type = NL_RTA_METRICS; 173 nla->nla_len = nla_len; 174 nla++; 175 nla->nla_type = NL_RTAX_MTU; 176 nla->nla_len = sizeof(struct nlattr) + sizeof(uint32_t); 177 *((uint32_t *)(nla + 1)) = nh->nh_mtu; 178 } 179 180 #ifdef ROUTE_MPATH 181 static void 182 dump_rc_nhg(struct nl_writer *nw, const struct nhgrp_object *nhg, struct rtmsg *rtm) 183 { 184 uint32_t uidx = nhgrp_get_uidx(nhg); 185 uint32_t num_nhops; 186 const struct weightened_nhop *wn = nhgrp_get_nhops(nhg, &num_nhops); 187 uint32_t base_rtflags = nhop_get_rtflags(wn[0].nh); 188 189 if (uidx != 0) 190 nlattr_add_u32(nw, NL_RTA_NH_ID, uidx); 191 nlattr_add_u32(nw, NL_RTA_KNH_ID, nhgrp_get_idx(nhg)); 192 193 nlattr_add_u32(nw, NL_RTA_RTFLAGS, base_rtflags); 194 int off = nlattr_add_nested(nw, NL_RTA_MULTIPATH); 195 if (off == 0) 196 return; 197 198 for (int i = 0; i < num_nhops; i++) { 199 int nh_off = nlattr_save_offset(nw); 200 struct rtnexthop *rtnh = nlmsg_reserve_object(nw, struct rtnexthop); 201 if (rtnh == NULL) 202 return; 203 rtnh->rtnh_flags = 0; 204 rtnh->rtnh_ifindex = wn[i].nh->nh_ifp->if_index; 205 rtnh->rtnh_hops = wn[i].weight; 206 dump_rc_nhop_gw(nw, wn[i].nh); 207 uint32_t rtflags = nhop_get_rtflags(wn[i].nh); 208 if (rtflags != base_rtflags) 209 nlattr_add_u32(nw, NL_RTA_RTFLAGS, rtflags); 210 if (rtflags & RTF_FIXEDMTU) 211 dump_rc_nhop_mtu(nw, wn[i].nh); 212 rtnh = nlattr_restore_offset(nw, nh_off, struct rtnexthop); 213 /* 214 * nlattr_add() allocates 4-byte aligned storage, no need to aligh 215 * length here 216 * */ 217 rtnh->rtnh_len = nlattr_save_offset(nw) - nh_off; 218 } 219 nlattr_set_len(nw, off); 220 } 221 #endif 222 223 static void 224 dump_rc_nhop(struct nl_writer *nw, const struct route_nhop_data *rnd, struct rtmsg *rtm) 225 { 226 #ifdef ROUTE_MPATH 227 if (NH_IS_NHGRP(rnd->rnd_nhop)) { 228 dump_rc_nhg(nw, rnd->rnd_nhgrp, rtm); 229 return; 230 } 231 #endif 232 const struct nhop_object *nh = rnd->rnd_nhop; 233 uint32_t rtflags = nhop_get_rtflags(nh); 234 235 /* 236 * IPv4 over IPv6 237 * ('RTA_VIA', {'family': 10, 'addr': 'fe80::20c:29ff:fe67:2dd'}), ('RTA_OIF', 2), 238 * IPv4 w/ gw 239 * ('RTA_GATEWAY', '172.16.107.131'), ('RTA_OIF', 2)], 240 * Direct route: 241 * ('RTA_OIF', 2) 242 */ 243 if (nh->nh_flags & NHF_GATEWAY) 244 dump_rc_nhop_gw(nw, nh); 245 246 uint32_t uidx = nhop_get_uidx(nh); 247 if (uidx != 0) 248 nlattr_add_u32(nw, NL_RTA_NH_ID, uidx); 249 nlattr_add_u32(nw, NL_RTA_KNH_ID, nhop_get_idx(nh)); 250 nlattr_add_u32(nw, NL_RTA_RTFLAGS, rtflags); 251 252 if (rtflags & RTF_FIXEDMTU) 253 dump_rc_nhop_mtu(nw, nh); 254 uint32_t nh_expire = nhop_get_expire(nh); 255 if (nh_expire > 0) 256 nlattr_add_u32(nw, NL_RTA_EXPIRES, nh_expire - time_uptime); 257 258 /* In any case, fill outgoing interface */ 259 nlattr_add_u32(nw, NL_RTA_OIF, nh->nh_ifp->if_index); 260 261 if (rnd->rnd_weight != RT_DEFAULT_WEIGHT) 262 nlattr_add_u32(nw, NL_RTA_WEIGHT, rnd->rnd_weight); 263 } 264 265 /* 266 * Dumps output from a rib command into an rtmsg 267 */ 268 269 static int 270 dump_px(uint32_t fibnum, const struct nlmsghdr *hdr, 271 const struct rtentry *rt, struct route_nhop_data *rnd, 272 struct nl_writer *nw) 273 { 274 struct rtmsg *rtm; 275 int error = 0; 276 277 NET_EPOCH_ASSERT(); 278 279 if (!nlmsg_reply(nw, hdr, sizeof(struct rtmsg))) 280 goto enomem; 281 282 int family = rt_get_family(rt); 283 int rtm_off = nlattr_save_offset(nw); 284 rtm = nlmsg_reserve_object(nw, struct rtmsg); 285 rtm->rtm_family = family; 286 rtm->rtm_dst_len = 0; 287 rtm->rtm_src_len = 0; 288 rtm->rtm_tos = 0; 289 if (fibnum < 255) 290 rtm->rtm_table = (unsigned char)fibnum; 291 rtm->rtm_scope = RT_SCOPE_UNIVERSE; 292 rtm->rtm_protocol = nl_get_rtm_protocol(rnd->rnd_nhop); 293 rtm->rtm_type = get_rtm_type(rnd->rnd_nhop); 294 295 nlattr_add_u32(nw, NL_RTA_TABLE, fibnum); 296 297 int plen = 0; 298 #if defined(INET) || defined(INET6) 299 uint32_t scopeid; 300 #endif 301 switch (family) { 302 #ifdef INET 303 case AF_INET: 304 { 305 struct in_addr addr; 306 rt_get_inet_prefix_plen(rt, &addr, &plen, &scopeid); 307 nlattr_add(nw, NL_RTA_DST, 4, &addr); 308 break; 309 } 310 #endif 311 #ifdef INET6 312 case AF_INET6: 313 { 314 struct in6_addr addr; 315 rt_get_inet6_prefix_plen(rt, &addr, &plen, &scopeid); 316 nlattr_add(nw, NL_RTA_DST, 16, &addr); 317 break; 318 } 319 #endif 320 default: 321 FIB_LOG(LOG_NOTICE, fibnum, family, "unsupported rt family: %d", family); 322 error = EAFNOSUPPORT; 323 goto flush; 324 } 325 326 rtm = nlattr_restore_offset(nw, rtm_off, struct rtmsg); 327 if (plen > 0) 328 rtm->rtm_dst_len = plen; 329 dump_rc_nhop(nw, rnd, rtm); 330 331 if (nlmsg_end(nw)) 332 return (0); 333 enomem: 334 error = ENOMEM; 335 flush: 336 nlmsg_abort(nw); 337 return (error); 338 } 339 340 static int 341 family_to_group(int family) 342 { 343 switch (family) { 344 case AF_INET: 345 return (RTNLGRP_IPV4_ROUTE); 346 case AF_INET6: 347 return (RTNLGRP_IPV6_ROUTE); 348 } 349 return (0); 350 } 351 352 static void 353 report_operation(uint32_t fibnum, struct rib_cmd_info *rc, 354 struct nlpcb *nlp, struct nlmsghdr *hdr) 355 { 356 struct nl_writer nw = {}; 357 uint32_t group_id = family_to_group(rt_get_family(rc->rc_rt)); 358 359 if (nlmsg_get_group_writer(&nw, NLMSG_SMALL, NETLINK_ROUTE, group_id)) { 360 struct route_nhop_data rnd = { 361 .rnd_nhop = rc_get_nhop(rc), 362 .rnd_weight = rc->rc_nh_weight, 363 }; 364 hdr->nlmsg_flags &= ~(NLM_F_REPLACE | NLM_F_CREATE); 365 hdr->nlmsg_flags &= ~(NLM_F_EXCL | NLM_F_APPEND); 366 switch (rc->rc_cmd) { 367 case RTM_ADD: 368 hdr->nlmsg_type = NL_RTM_NEWROUTE; 369 hdr->nlmsg_flags |= NLM_F_CREATE | NLM_F_EXCL; 370 break; 371 case RTM_CHANGE: 372 hdr->nlmsg_type = NL_RTM_NEWROUTE; 373 hdr->nlmsg_flags |= NLM_F_REPLACE; 374 break; 375 case RTM_DELETE: 376 hdr->nlmsg_type = NL_RTM_DELROUTE; 377 break; 378 } 379 dump_px(fibnum, hdr, rc->rc_rt, &rnd, &nw); 380 nlmsg_flush(&nw); 381 } 382 383 rtsock_callback_p->route_f(fibnum, rc); 384 } 385 386 static void 387 set_scope6(struct sockaddr *sa, struct ifnet *ifp) 388 { 389 #ifdef INET6 390 if (sa != NULL && sa->sa_family == AF_INET6 && ifp != NULL) { 391 struct sockaddr_in6 *sa6 = (struct sockaddr_in6 *)sa; 392 393 if (IN6_IS_ADDR_LINKLOCAL(&sa6->sin6_addr)) 394 in6_set_unicast_scopeid(&sa6->sin6_addr, if_getindex(ifp)); 395 } 396 #endif 397 } 398 399 struct rta_mpath_nh { 400 struct sockaddr *gw; 401 struct ifnet *ifp; 402 uint8_t rtnh_flags; 403 uint8_t rtnh_weight; 404 }; 405 406 #define _IN(_field) offsetof(struct rtnexthop, _field) 407 #define _OUT(_field) offsetof(struct rta_mpath_nh, _field) 408 const static struct nlattr_parser nla_p_rtnh[] = { 409 { .type = NL_RTA_GATEWAY, .off = _OUT(gw), .cb = nlattr_get_ip }, 410 { .type = NL_RTA_VIA, .off = _OUT(gw), .cb = nlattr_get_ipvia }, 411 }; 412 const static struct nlfield_parser nlf_p_rtnh[] = { 413 { .off_in = _IN(rtnh_flags), .off_out = _OUT(rtnh_flags), .cb = nlf_get_u8 }, 414 { .off_in = _IN(rtnh_hops), .off_out = _OUT(rtnh_weight), .cb = nlf_get_u8 }, 415 { .off_in = _IN(rtnh_ifindex), .off_out = _OUT(ifp), .cb = nlf_get_ifpz }, 416 }; 417 #undef _IN 418 #undef _OUT 419 420 static bool 421 post_p_rtnh(void *_attrs, struct nl_pstate *npt __unused) 422 { 423 struct rta_mpath_nh *attrs = (struct rta_mpath_nh *)_attrs; 424 425 set_scope6(attrs->gw, attrs->ifp); 426 return (true); 427 } 428 NL_DECLARE_PARSER_EXT(mpath_parser, struct rtnexthop, NULL, nlf_p_rtnh, nla_p_rtnh, post_p_rtnh); 429 430 struct rta_mpath { 431 int num_nhops; 432 struct rta_mpath_nh nhops[0]; 433 }; 434 435 static int 436 nlattr_get_multipath(struct nlattr *nla, struct nl_pstate *npt, const void *arg, void *target) 437 { 438 int data_len = nla->nla_len - sizeof(struct nlattr); 439 struct rtnexthop *rtnh; 440 441 int max_nhops = data_len / sizeof(struct rtnexthop); 442 443 struct rta_mpath *mp = npt_alloc(npt, (max_nhops + 2) * sizeof(struct rta_mpath_nh)); 444 mp->num_nhops = 0; 445 446 for (rtnh = (struct rtnexthop *)(nla + 1); data_len > 0; ) { 447 struct rta_mpath_nh *mpnh = &mp->nhops[mp->num_nhops++]; 448 449 int error = nl_parse_header(rtnh, rtnh->rtnh_len, &mpath_parser, 450 npt, mpnh); 451 if (error != 0) { 452 NLMSG_REPORT_ERR_MSG(npt, "RTA_MULTIPATH: nexhop %d: parse failed", 453 mp->num_nhops - 1); 454 return (error); 455 } 456 457 int len = NL_ITEM_ALIGN(rtnh->rtnh_len); 458 data_len -= len; 459 rtnh = (struct rtnexthop *)((char *)rtnh + len); 460 } 461 if (data_len != 0 || mp->num_nhops == 0) { 462 NLMSG_REPORT_ERR_MSG(npt, "invalid RTA_MULTIPATH attr"); 463 return (EINVAL); 464 } 465 466 *((struct rta_mpath **)target) = mp; 467 return (0); 468 } 469 470 471 struct nl_parsed_route { 472 struct sockaddr *rta_dst; 473 struct sockaddr *rta_gw; 474 struct ifnet *rta_oif; 475 struct rta_mpath *rta_multipath; 476 uint32_t rta_table; 477 uint32_t rta_rtflags; 478 uint32_t rta_nh_id; 479 uint32_t rta_weight; 480 uint32_t rtax_mtu; 481 uint8_t rtm_family; 482 uint8_t rtm_dst_len; 483 uint8_t rtm_protocol; 484 uint8_t rtm_type; 485 uint32_t rtm_flags; 486 }; 487 488 #define _IN(_field) offsetof(struct rtmsg, _field) 489 #define _OUT(_field) offsetof(struct nl_parsed_route, _field) 490 static struct nlattr_parser nla_p_rtmetrics[] = { 491 { .type = NL_RTAX_MTU, .off = _OUT(rtax_mtu), .cb = nlattr_get_uint32 }, 492 }; 493 NL_DECLARE_ATTR_PARSER(metrics_parser, nla_p_rtmetrics); 494 495 static const struct nlattr_parser nla_p_rtmsg[] = { 496 { .type = NL_RTA_DST, .off = _OUT(rta_dst), .cb = nlattr_get_ip }, 497 { .type = NL_RTA_OIF, .off = _OUT(rta_oif), .cb = nlattr_get_ifp }, 498 { .type = NL_RTA_GATEWAY, .off = _OUT(rta_gw), .cb = nlattr_get_ip }, 499 { .type = NL_RTA_METRICS, .arg = &metrics_parser, .cb = nlattr_get_nested }, 500 { .type = NL_RTA_MULTIPATH, .off = _OUT(rta_multipath), .cb = nlattr_get_multipath }, 501 { .type = NL_RTA_WEIGHT, .off = _OUT(rta_weight), .cb = nlattr_get_uint32 }, 502 { .type = NL_RTA_RTFLAGS, .off = _OUT(rta_rtflags), .cb = nlattr_get_uint32 }, 503 { .type = NL_RTA_TABLE, .off = _OUT(rta_table), .cb = nlattr_get_uint32 }, 504 { .type = NL_RTA_VIA, .off = _OUT(rta_gw), .cb = nlattr_get_ipvia }, 505 { .type = NL_RTA_NH_ID, .off = _OUT(rta_nh_id), .cb = nlattr_get_uint32 }, 506 }; 507 508 static const struct nlfield_parser nlf_p_rtmsg[] = { 509 { .off_in = _IN(rtm_family), .off_out = _OUT(rtm_family), .cb = nlf_get_u8 }, 510 { .off_in = _IN(rtm_dst_len), .off_out = _OUT(rtm_dst_len), .cb = nlf_get_u8 }, 511 { .off_in = _IN(rtm_protocol), .off_out = _OUT(rtm_protocol), .cb = nlf_get_u8 }, 512 { .off_in = _IN(rtm_type), .off_out = _OUT(rtm_type), .cb = nlf_get_u8 }, 513 { .off_in = _IN(rtm_flags), .off_out = _OUT(rtm_flags), .cb = nlf_get_u32 }, 514 }; 515 #undef _IN 516 #undef _OUT 517 518 static bool 519 post_p_rtmsg(void *_attrs, struct nl_pstate *npt __unused) 520 { 521 struct nl_parsed_route *attrs = (struct nl_parsed_route *)_attrs; 522 523 set_scope6(attrs->rta_dst, attrs->rta_oif); 524 set_scope6(attrs->rta_gw, attrs->rta_oif); 525 return (true); 526 } 527 NL_DECLARE_PARSER_EXT(rtm_parser, struct rtmsg, NULL, nlf_p_rtmsg, nla_p_rtmsg, post_p_rtmsg); 528 529 struct netlink_walkargs { 530 struct nl_writer *nw; 531 struct route_nhop_data rnd; 532 struct nlmsghdr hdr; 533 struct nlpcb *nlp; 534 uint32_t fibnum; 535 int family; 536 int error; 537 int count; 538 int dumped; 539 int dumped_tables; 540 }; 541 542 static int 543 dump_rtentry(struct rtentry *rt, void *_arg) 544 { 545 struct netlink_walkargs *wa = (struct netlink_walkargs *)_arg; 546 int error; 547 548 wa->count++; 549 if (wa->error != 0) 550 return (0); 551 if (!rt_is_exportable(rt, nlp_get_cred(wa->nlp))) 552 return (0); 553 wa->dumped++; 554 555 rt_get_rnd(rt, &wa->rnd); 556 557 error = dump_px(wa->fibnum, &wa->hdr, rt, &wa->rnd, wa->nw); 558 559 IF_DEBUG_LEVEL(LOG_DEBUG3) { 560 char rtbuf[INET6_ADDRSTRLEN + 5]; 561 FIB_LOG(LOG_DEBUG3, wa->fibnum, wa->family, 562 "Dump %s, offset %u, error %d", 563 rt_print_buf(rt, rtbuf, sizeof(rtbuf)), 564 wa->nw->offset, error); 565 } 566 wa->error = error; 567 568 return (0); 569 } 570 571 static void 572 dump_rtable_one(struct netlink_walkargs *wa, uint32_t fibnum, int family) 573 { 574 FIB_LOG(LOG_DEBUG2, fibnum, family, "Start dump"); 575 wa->count = 0; 576 wa->dumped = 0; 577 578 rib_walk(fibnum, family, false, dump_rtentry, wa); 579 580 wa->dumped_tables++; 581 582 FIB_LOG(LOG_DEBUG2, fibnum, family, "End dump, iterated %d dumped %d", 583 wa->count, wa->dumped); 584 NL_LOG(LOG_DEBUG2, "Current offset: %d", wa->nw->offset); 585 } 586 587 static int 588 dump_rtable_fib(struct netlink_walkargs *wa, uint32_t fibnum, int family) 589 { 590 wa->fibnum = fibnum; 591 592 if (family == AF_UNSPEC) { 593 for (int i = 0; i < AF_MAX; i++) { 594 if (rt_tables_get_rnh(fibnum, i) != 0) { 595 wa->family = i; 596 dump_rtable_one(wa, fibnum, i); 597 if (wa->error != 0) 598 break; 599 } 600 } 601 } else { 602 if (rt_tables_get_rnh(fibnum, family) != 0) { 603 wa->family = family; 604 dump_rtable_one(wa, fibnum, family); 605 } 606 } 607 608 return (wa->error); 609 } 610 611 static int 612 handle_rtm_getroute(struct nlpcb *nlp, struct nl_parsed_route *attrs, 613 struct nlmsghdr *hdr, struct nl_pstate *npt) 614 { 615 RIB_RLOCK_TRACKER; 616 struct rib_head *rnh; 617 const struct rtentry *rt; 618 struct route_nhop_data rnd; 619 uint32_t fibnum = attrs->rta_table; 620 sa_family_t family = attrs->rtm_family; 621 622 if (attrs->rta_dst == NULL) { 623 NLMSG_REPORT_ERR_MSG(npt, "No RTA_DST supplied"); 624 return (EINVAL); 625 } 626 627 rnh = rt_tables_get_rnh(fibnum, family); 628 if (rnh == NULL) 629 return (EAFNOSUPPORT); 630 631 RIB_RLOCK(rnh); 632 633 struct sockaddr *dst = attrs->rta_dst; 634 635 if (attrs->rtm_flags & RTM_F_PREFIX) 636 rt = rib_lookup_prefix_plen(rnh, dst, attrs->rtm_dst_len, &rnd); 637 else 638 rt = (const struct rtentry *)rnh->rnh_matchaddr(dst, &rnh->head); 639 if (rt == NULL) { 640 RIB_RUNLOCK(rnh); 641 return (ESRCH); 642 } 643 644 rt_get_rnd(rt, &rnd); 645 rnd.rnd_nhop = nhop_select_func(rnd.rnd_nhop, 0); 646 647 RIB_RUNLOCK(rnh); 648 649 if (!rt_is_exportable(rt, nlp_get_cred(nlp))) 650 return (ESRCH); 651 652 IF_DEBUG_LEVEL(LOG_DEBUG2) { 653 char rtbuf[NHOP_PRINT_BUFSIZE] __unused, nhbuf[NHOP_PRINT_BUFSIZE] __unused; 654 FIB_LOG(LOG_DEBUG2, fibnum, family, "getroute completed: got %s for %s", 655 nhop_print_buf_any(rnd.rnd_nhop, nhbuf, sizeof(nhbuf)), 656 rt_print_buf(rt, rtbuf, sizeof(rtbuf))); 657 } 658 659 hdr->nlmsg_type = NL_RTM_NEWROUTE; 660 dump_px(fibnum, hdr, rt, &rnd, npt->nw); 661 662 return (0); 663 } 664 665 static int 666 handle_rtm_dump(struct nlpcb *nlp, uint32_t fibnum, int family, 667 struct nlmsghdr *hdr, struct nl_writer *nw) 668 { 669 struct netlink_walkargs wa = { 670 .nlp = nlp, 671 .nw = nw, 672 .hdr.nlmsg_pid = hdr->nlmsg_pid, 673 .hdr.nlmsg_seq = hdr->nlmsg_seq, 674 .hdr.nlmsg_type = NL_RTM_NEWROUTE, 675 .hdr.nlmsg_flags = hdr->nlmsg_flags | NLM_F_MULTI, 676 }; 677 678 if (fibnum == RT_TABLE_UNSPEC) { 679 for (int i = 0; i < V_rt_numfibs; i++) { 680 dump_rtable_fib(&wa, fibnum, family); 681 if (wa.error != 0) 682 break; 683 } 684 } else 685 dump_rtable_fib(&wa, fibnum, family); 686 687 if (wa.error == 0 && wa.dumped_tables == 0) { 688 FIB_LOG(LOG_DEBUG, fibnum, family, "incorrect fibnum/family"); 689 wa.error = ESRCH; 690 // How do we propagate it? 691 } 692 693 if (!nlmsg_end_dump(wa.nw, wa.error, &wa.hdr)) { 694 NL_LOG(LOG_DEBUG, "Unable to finalize the dump"); 695 return (ENOMEM); 696 } 697 698 return (wa.error); 699 } 700 701 static struct nhop_object * 702 finalize_nhop(struct nhop_object *nh, const struct sockaddr *dst, int *perror) 703 { 704 /* 705 * The following MUST be filled: 706 * nh_ifp, nh_ifa, nh_gw 707 */ 708 if (nh->gw_sa.sa_family == 0) { 709 /* 710 * Empty gateway. Can be direct route with RTA_OIF set. 711 */ 712 if (nh->nh_ifp != NULL) 713 nhop_set_direct_gw(nh, nh->nh_ifp); 714 else { 715 NL_LOG(LOG_DEBUG, "empty gateway and interface, skipping"); 716 *perror = EINVAL; 717 return (NULL); 718 } 719 /* Both nh_ifp and gateway are set */ 720 } else { 721 /* Gateway is set up, we can derive ifp if not set */ 722 if (nh->nh_ifp == NULL) { 723 uint32_t fibnum = nhop_get_fibnum(nh); 724 uint32_t flags = 0; 725 726 if (nh->nh_flags & NHF_GATEWAY) 727 flags = RTF_GATEWAY; 728 else if (nh->nh_flags & NHF_HOST) 729 flags = RTF_HOST; 730 731 struct ifaddr *ifa = ifa_ifwithroute(flags, dst, &nh->gw_sa, fibnum); 732 if (ifa == NULL) { 733 NL_LOG(LOG_DEBUG, "Unable to determine ifp, skipping"); 734 *perror = EINVAL; 735 return (NULL); 736 } 737 nhop_set_transmit_ifp(nh, ifa->ifa_ifp); 738 } 739 } 740 /* Both nh_ifp and gateway are set */ 741 if (nh->nh_ifa == NULL) { 742 const struct sockaddr *gw_sa = &nh->gw_sa; 743 744 if (gw_sa->sa_family != dst->sa_family) { 745 /* 746 * Use dst as the target for determining the default 747 * preferred ifa IF 748 * 1) the gateway is link-level (e.g. direct route) 749 * 2) the gateway family is different (e.g. IPv4 over IPv6). 750 */ 751 gw_sa = dst; 752 } 753 754 struct ifaddr *ifa = ifaof_ifpforaddr(gw_sa, nh->nh_ifp); 755 if (ifa == NULL) { 756 NL_LOG(LOG_DEBUG, "Unable to determine ifa, skipping"); 757 *perror = EINVAL; 758 return (NULL); 759 } 760 nhop_set_src(nh, ifa); 761 } 762 763 return (nhop_get_nhop(nh, perror)); 764 } 765 766 static int 767 get_pxflag(const struct nl_parsed_route *attrs) 768 { 769 int pxflag = 0; 770 switch (attrs->rtm_family) { 771 case AF_INET: 772 if (attrs->rtm_dst_len == 32) 773 pxflag = NHF_HOST; 774 else if (attrs->rtm_dst_len == 0) 775 pxflag = NHF_DEFAULT; 776 break; 777 case AF_INET6: 778 if (attrs->rtm_dst_len == 128) 779 pxflag = NHF_HOST; 780 else if (attrs->rtm_dst_len == 0) 781 pxflag = NHF_DEFAULT; 782 break; 783 } 784 785 return (pxflag); 786 } 787 788 static int 789 get_op_flags(int nlm_flags) 790 { 791 int op_flags = 0; 792 793 op_flags |= (nlm_flags & NLM_F_REPLACE) ? RTM_F_REPLACE : 0; 794 op_flags |= (nlm_flags & NLM_F_EXCL) ? RTM_F_EXCL : 0; 795 op_flags |= (nlm_flags & NLM_F_CREATE) ? RTM_F_CREATE : 0; 796 op_flags |= (nlm_flags & NLM_F_APPEND) ? RTM_F_APPEND : 0; 797 798 return (op_flags); 799 } 800 801 #ifdef ROUTE_MPATH 802 static int 803 create_nexthop_one(struct nl_parsed_route *attrs, struct rta_mpath_nh *mpnh, 804 struct nl_pstate *npt, struct nhop_object **pnh) 805 { 806 int error; 807 808 if (mpnh->gw == NULL) 809 return (EINVAL); 810 811 struct nhop_object *nh = nhop_alloc(attrs->rta_table, attrs->rtm_family); 812 if (nh == NULL) 813 return (ENOMEM); 814 815 error = nl_set_nexthop_gw(nh, mpnh->gw, mpnh->ifp, npt); 816 if (error != 0) { 817 nhop_free(nh); 818 return (error); 819 } 820 if (mpnh->ifp != NULL) 821 nhop_set_transmit_ifp(nh, mpnh->ifp); 822 nhop_set_pxtype_flag(nh, get_pxflag(attrs)); 823 nhop_set_rtflags(nh, attrs->rta_rtflags); 824 if (attrs->rtm_protocol > RTPROT_STATIC) 825 nhop_set_origin(nh, attrs->rtm_protocol); 826 827 *pnh = finalize_nhop(nh, attrs->rta_dst, &error); 828 829 return (error); 830 } 831 #endif 832 833 static struct nhop_object * 834 create_nexthop_from_attrs(struct nl_parsed_route *attrs, 835 struct nl_pstate *npt, int *perror) 836 { 837 struct nhop_object *nh = NULL; 838 int error = 0; 839 840 if (attrs->rta_multipath != NULL) { 841 #ifdef ROUTE_MPATH 842 /* Multipath w/o explicit nexthops */ 843 int num_nhops = attrs->rta_multipath->num_nhops; 844 struct weightened_nhop *wn = npt_alloc(npt, sizeof(*wn) * num_nhops); 845 846 for (int i = 0; i < num_nhops; i++) { 847 struct rta_mpath_nh *mpnh = &attrs->rta_multipath->nhops[i]; 848 849 error = create_nexthop_one(attrs, mpnh, npt, &wn[i].nh); 850 if (error != 0) { 851 for (int j = 0; j < i; j++) 852 nhop_free(wn[j].nh); 853 break; 854 } 855 wn[i].weight = mpnh->rtnh_weight > 0 ? mpnh->rtnh_weight : 1; 856 } 857 if (error == 0) { 858 struct rib_head *rh = nhop_get_rh(wn[0].nh); 859 struct nhgrp_object *nhg; 860 861 nhg = nhgrp_alloc(rh->rib_fibnum, rh->rib_family, 862 wn, num_nhops, perror); 863 if (nhg != NULL) { 864 if (attrs->rtm_protocol > RTPROT_STATIC) 865 nhgrp_set_origin(nhg, attrs->rtm_protocol); 866 nhg = nhgrp_get_nhgrp(nhg, perror); 867 } 868 for (int i = 0; i < num_nhops; i++) 869 nhop_free(wn[i].nh); 870 if (nhg != NULL) 871 return ((struct nhop_object *)nhg); 872 error = *perror; 873 } 874 #else 875 error = ENOTSUP; 876 #endif 877 *perror = error; 878 } else { 879 nh = nhop_alloc(attrs->rta_table, attrs->rtm_family); 880 if (nh == NULL) { 881 *perror = ENOMEM; 882 return (NULL); 883 } 884 if (attrs->rta_gw != NULL) { 885 *perror = nl_set_nexthop_gw(nh, attrs->rta_gw, attrs->rta_oif, npt); 886 if (*perror != 0) { 887 nhop_free(nh); 888 return (NULL); 889 } 890 } 891 if (attrs->rta_oif != NULL) 892 nhop_set_transmit_ifp(nh, attrs->rta_oif); 893 if (attrs->rtax_mtu != 0) 894 nhop_set_mtu(nh, attrs->rtax_mtu, true); 895 if (attrs->rta_rtflags & RTF_BROADCAST) 896 nhop_set_broadcast(nh, true); 897 if (attrs->rtm_protocol > RTPROT_STATIC) 898 nhop_set_origin(nh, attrs->rtm_protocol); 899 nhop_set_pxtype_flag(nh, get_pxflag(attrs)); 900 nhop_set_rtflags(nh, attrs->rta_rtflags); 901 902 switch (attrs->rtm_type) { 903 case RTN_UNICAST: 904 break; 905 case RTN_BLACKHOLE: 906 nhop_set_blackhole(nh, RTF_BLACKHOLE); 907 break; 908 case RTN_PROHIBIT: 909 case RTN_UNREACHABLE: 910 nhop_set_blackhole(nh, RTF_REJECT); 911 break; 912 /* TODO: return ENOTSUP for other types if strict option is set */ 913 } 914 915 nh = finalize_nhop(nh, attrs->rta_dst, perror); 916 } 917 918 return (nh); 919 } 920 921 static int 922 rtnl_handle_newroute(struct nlmsghdr *hdr, struct nlpcb *nlp, 923 struct nl_pstate *npt) 924 { 925 struct rib_cmd_info rc = {}; 926 struct nhop_object *nh = NULL; 927 int error; 928 929 struct nl_parsed_route attrs = {}; 930 error = nl_parse_nlmsg(hdr, &rtm_parser, npt, &attrs); 931 if (error != 0) 932 return (error); 933 934 /* Check if we have enough data */ 935 if (attrs.rta_dst == NULL) { 936 NL_LOG(LOG_DEBUG, "missing RTA_DST"); 937 return (EINVAL); 938 } 939 940 if (attrs.rta_table >= V_rt_numfibs) { 941 NLMSG_REPORT_ERR_MSG(npt, "invalid fib"); 942 return (EINVAL); 943 } 944 945 if (attrs.rta_nh_id != 0) { 946 /* Referenced uindex */ 947 int pxflag = get_pxflag(&attrs); 948 nh = nl_find_nhop(attrs.rta_table, attrs.rtm_family, attrs.rta_nh_id, 949 pxflag, &error); 950 if (error != 0) 951 return (error); 952 } else { 953 nh = create_nexthop_from_attrs(&attrs, npt, &error); 954 if (error != 0) { 955 NL_LOG(LOG_DEBUG, "Error creating nexthop"); 956 return (error); 957 } 958 } 959 960 if (!NH_IS_NHGRP(nh) && attrs.rta_weight == 0) 961 attrs.rta_weight = RT_DEFAULT_WEIGHT; 962 struct route_nhop_data rnd = { .rnd_nhop = nh, .rnd_weight = attrs.rta_weight }; 963 int op_flags = get_op_flags(hdr->nlmsg_flags); 964 965 error = rib_add_route_px(attrs.rta_table, attrs.rta_dst, attrs.rtm_dst_len, 966 &rnd, op_flags, &rc); 967 if (error == 0) 968 report_operation(attrs.rta_table, &rc, nlp, hdr); 969 return (error); 970 } 971 972 static int 973 path_match_func(const struct rtentry *rt, const struct nhop_object *nh, void *_data) 974 { 975 struct nl_parsed_route *attrs = (struct nl_parsed_route *)_data; 976 977 if ((attrs->rta_gw != NULL) && !rib_match_gw(rt, nh, attrs->rta_gw)) 978 return (0); 979 980 if ((attrs->rta_oif != NULL) && (attrs->rta_oif != nh->nh_ifp)) 981 return (0); 982 983 return (1); 984 } 985 986 static int 987 rtnl_handle_delroute(struct nlmsghdr *hdr, struct nlpcb *nlp, 988 struct nl_pstate *npt) 989 { 990 struct rib_cmd_info rc; 991 int error; 992 993 struct nl_parsed_route attrs = {}; 994 error = nl_parse_nlmsg(hdr, &rtm_parser, npt, &attrs); 995 if (error != 0) 996 return (error); 997 998 if (attrs.rta_dst == NULL) { 999 NLMSG_REPORT_ERR_MSG(npt, "RTA_DST is not set"); 1000 return (ESRCH); 1001 } 1002 1003 if (attrs.rta_table >= V_rt_numfibs) { 1004 NLMSG_REPORT_ERR_MSG(npt, "invalid fib"); 1005 return (EINVAL); 1006 } 1007 1008 error = rib_del_route_px(attrs.rta_table, attrs.rta_dst, 1009 attrs.rtm_dst_len, path_match_func, &attrs, 0, &rc); 1010 if (error == 0) 1011 report_operation(attrs.rta_table, &rc, nlp, hdr); 1012 return (error); 1013 } 1014 1015 static int 1016 rtnl_handle_getroute(struct nlmsghdr *hdr, struct nlpcb *nlp, struct nl_pstate *npt) 1017 { 1018 int error; 1019 1020 struct nl_parsed_route attrs = {}; 1021 error = nl_parse_nlmsg(hdr, &rtm_parser, npt, &attrs); 1022 if (error != 0) 1023 return (error); 1024 1025 if (attrs.rta_table >= V_rt_numfibs) { 1026 NLMSG_REPORT_ERR_MSG(npt, "invalid fib"); 1027 return (EINVAL); 1028 } 1029 1030 if (hdr->nlmsg_flags & NLM_F_DUMP) 1031 error = handle_rtm_dump(nlp, attrs.rta_table, attrs.rtm_family, hdr, npt->nw); 1032 else 1033 error = handle_rtm_getroute(nlp, &attrs, hdr, npt); 1034 1035 return (error); 1036 } 1037 1038 void 1039 rtnl_handle_route_event(uint32_t fibnum, const struct rib_cmd_info *rc) 1040 { 1041 struct nl_writer nw = {}; 1042 int family, nlm_flags = 0; 1043 1044 family = rt_get_family(rc->rc_rt); 1045 1046 /* XXX: check if there are active listeners first */ 1047 1048 /* TODO: consider passing PID/type/seq */ 1049 switch (rc->rc_cmd) { 1050 case RTM_ADD: 1051 nlm_flags = NLM_F_EXCL | NLM_F_CREATE; 1052 break; 1053 case RTM_CHANGE: 1054 nlm_flags = NLM_F_REPLACE; 1055 break; 1056 case RTM_DELETE: 1057 nlm_flags = 0; 1058 break; 1059 } 1060 IF_DEBUG_LEVEL(LOG_DEBUG2) { 1061 char rtbuf[NHOP_PRINT_BUFSIZE] __unused; 1062 FIB_LOG(LOG_DEBUG2, fibnum, family, 1063 "received event %s for %s / nlm_flags=%X", 1064 rib_print_cmd(rc->rc_cmd), 1065 rt_print_buf(rc->rc_rt, rtbuf, sizeof(rtbuf)), 1066 nlm_flags); 1067 } 1068 1069 struct nlmsghdr hdr = { 1070 .nlmsg_flags = nlm_flags, 1071 .nlmsg_type = get_rtmsg_type_from_rtsock(rc->rc_cmd), 1072 }; 1073 1074 struct route_nhop_data rnd = { 1075 .rnd_nhop = rc_get_nhop(rc), 1076 .rnd_weight = rc->rc_nh_weight, 1077 }; 1078 1079 uint32_t group_id = family_to_group(family); 1080 if (!nlmsg_get_group_writer(&nw, NLMSG_SMALL, NETLINK_ROUTE, group_id)) { 1081 NL_LOG(LOG_DEBUG, "error allocating event buffer"); 1082 return; 1083 } 1084 1085 dump_px(fibnum, &hdr, rc->rc_rt, &rnd, &nw); 1086 nlmsg_flush(&nw); 1087 } 1088 1089 static const struct rtnl_cmd_handler cmd_handlers[] = { 1090 { 1091 .cmd = NL_RTM_GETROUTE, 1092 .name = "RTM_GETROUTE", 1093 .cb = &rtnl_handle_getroute, 1094 .flags = RTNL_F_ALLOW_NONVNET_JAIL, 1095 }, 1096 { 1097 .cmd = NL_RTM_DELROUTE, 1098 .name = "RTM_DELROUTE", 1099 .cb = &rtnl_handle_delroute, 1100 .priv = PRIV_NET_ROUTE, 1101 }, 1102 { 1103 .cmd = NL_RTM_NEWROUTE, 1104 .name = "RTM_NEWROUTE", 1105 .cb = &rtnl_handle_newroute, 1106 .priv = PRIV_NET_ROUTE, 1107 } 1108 }; 1109 1110 static const struct nlhdr_parser *all_parsers[] = {&mpath_parser, &metrics_parser, &rtm_parser}; 1111 1112 void 1113 rtnl_routes_init(void) 1114 { 1115 NL_VERIFY_PARSERS(all_parsers); 1116 rtnl_register_messages(cmd_handlers, NL_ARRAY_LEN(cmd_handlers)); 1117 } 1118