system-linux.c

   1 /*
   2  * netifd - network interface daemon
   3  * Copyright (C) 2012 Felix Fietkau <nbd@openwrt.org>
   4  * Copyright (C) 2013 Jo-Philipp Wich <jow@openwrt.org>
   5  * Copyright (C) 2013 Steven Barth <steven@midlink.org>
   6  * Copyright (C) 2014 Gioacchino Mazzurco <gio@eigenlab.org>
   7  * Copyright (C) 2017 Matthias Schiffer <mschiffer@universe-factory.net>
   8  * Copyright (C) 2018 Hans Dedecker <dedeckeh@gmail.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  */
  19 #define _GNU_SOURCE
  20
  21 #include <sys/socket.h>
  22 #include <sys/ioctl.h>
  23 #include <sys/stat.h>
  24 #include <sys/syscall.h>
  25
  26 #include <net/if.h>
  27 #include <net/if_arp.h>
  28
  29 #include <arpa/inet.h>
  30 #include <netinet/ether.h>
  31 #include <netinet/in.h>
  32
  33 #include <linux/rtnetlink.h>
  34 #include <linux/neighbour.h>
  35 #include <linux/sockios.h>
  36 #include <linux/ip.h>
  37 #include <linux/if_addr.h>
  38 #include <linux/if_link.h>
  39 #include <linux/if_vlan.h>
  40 #include <linux/if_bridge.h>
  41 #include <linux/if_tunnel.h>
  42 #include <linux/ip6_tunnel.h>
  43 #include <linux/ethtool.h>
  44 #include <linux/fib_rules.h>
  45 #include <linux/veth.h>
  46 #include <linux/version.h>
  47
  48 #ifndef RTN_FAILED_POLICY
  49 #define RTN_FAILED_POLICY 12
  50 #endif
  51
  52 #ifndef IFA_F_NOPREFIXROUTE
  53 #define IFA_F_NOPREFIXROUTE 0x200
  54 #endif
  55
  56 #ifndef IFA_FLAGS
  57 #define IFA_FLAGS (IFA_MULTICAST + 1)
  58 #endif
  59
  60 #include <string.h>
  61 #include <fcntl.h>
  62 #include <glob.h>
  63 #include <time.h>
  64 #include <unistd.h>
  65
  66 #include <netlink/msg.h>
  67 #include <netlink/attr.h>
  68 #include <netlink/socket.h>
  69 #include <libubox/uloop.h>
  70
  71 #include "netifd.h"
  72 #include "device.h"
  73 #include "system.h"
  74
  75 struct event_socket {
  76         struct uloop_fd uloop;
  77         struct nl_sock *sock;
  78         int bufsize;
  79 };
  80
  81 static int sock_ioctl = -1;
  82 static struct nl_sock *sock_rtnl = NULL;
  83
  84 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  85 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  86 static int system_add_proto_tunnel(const char *name, const uint8_t proto,
  87                                         const unsigned int link, struct blob_attr **tb);
  88 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb);
  89
  90 static char dev_buf[256];
  91
  92 static void
  93 handler_nl_event(struct uloop_fd *u, unsigned int events)
  94 {
  95         struct event_socket *ev = container_of(u, struct event_socket, uloop);
  96         int err;
  97         socklen_t errlen = sizeof(err);
  98
  99         if (!u->error) {
 100                 nl_recvmsgs_default(ev->sock);
 101                 return;
 102         }
 103
 104         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 105                 goto abort;
 106
 107         switch(err) {
 108         case ENOBUFS:
 109                 /* Increase rx buffer size on netlink socket */
 110                 ev->bufsize *= 2;
 111                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 112                         goto abort;
 113
 114                 /* Request full dump since some info got dropped */
 115                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 116                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 117                 break;
 118
 119         default:
 120                 goto abort;
 121         }
 122         u->error = false;
 123         return;
 124
 125 abort:
 126         uloop_fd_delete(&ev->uloop);
 127         return;
 128 }
 129
 130 static struct nl_sock *
 131 create_socket(int protocol, int groups)
 132 {
 133         struct nl_sock *sock;
 134
 135         sock = nl_socket_alloc();
 136         if (!sock)
 137                 return NULL;
 138
 139         if (groups)
 140                 nl_join_groups(sock, groups);
 141
 142         if (nl_connect(sock, protocol)) {
 143                 nl_socket_free(sock);
 144                 return NULL;
 145         }
 146
 147         return sock;
 148 }
 149
 150 static bool
 151 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 152                         uloop_fd_handler cb, int flags)
 153 {
 154         ev->sock = create_socket(protocol, groups);
 155         if (!ev->sock)
 156                 return false;
 157
 158         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 159         ev->uloop.cb = cb;
 160         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 161                 return false;
 162
 163         return true;
 164 }
 165
 166 static bool
 167 create_event_socket(struct event_socket *ev, int protocol,
 168                     int (*cb)(struct nl_msg *msg, void *arg))
 169 {
 170         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 171                 return false;
 172
 173         /* Install the valid custom callback handler */
 174         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 175
 176         /* Disable sequence number checking on event sockets */
 177         nl_socket_disable_seq_check(ev->sock);
 178
 179         /* Increase rx buffer size to 65K on event sockets */
 180         ev->bufsize = 65535;
 181         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 182                 return false;
 183
 184         return true;
 185 }
 186
 187 static bool
 188 create_hotplug_event_socket(struct event_socket *ev, int protocol,
 189                             void (*cb)(struct uloop_fd *u, unsigned int events))
 190 {
 191         if (!create_raw_event_socket(ev, protocol, 1, cb, ULOOP_ERROR_CB))
 192                 return false;
 193
 194         /* Increase rx buffer size to 65K on event sockets */
 195         ev->bufsize = 65535;
 196         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 197                 return false;
 198
 199         return true;
 200 }
 201
 202 static bool
 203 system_rtn_aton(const char *src, unsigned int *dst)
 204 {
 205         char *e;
 206         unsigned int n;
 207
 208         if (!strcmp(src, "local"))
 209                 n = RTN_LOCAL;
 210         else if (!strcmp(src, "nat"))
 211                 n = RTN_NAT;
 212         else if (!strcmp(src, "broadcast"))
 213                 n = RTN_BROADCAST;
 214         else if (!strcmp(src, "anycast"))
 215                 n = RTN_ANYCAST;
 216         else if (!strcmp(src, "multicast"))
 217                 n = RTN_MULTICAST;
 218         else if (!strcmp(src, "prohibit"))
 219                 n = RTN_PROHIBIT;
 220         else if (!strcmp(src, "unreachable"))
 221                 n = RTN_UNREACHABLE;
 222         else if (!strcmp(src, "blackhole"))
 223                 n = RTN_BLACKHOLE;
 224         else if (!strcmp(src, "xresolve"))
 225                 n = RTN_XRESOLVE;
 226         else if (!strcmp(src, "unicast"))
 227                 n = RTN_UNICAST;
 228         else if (!strcmp(src, "throw"))
 229                 n = RTN_THROW;
 230         else if (!strcmp(src, "failed_policy"))
 231                 n = RTN_FAILED_POLICY;
 232         else {
 233                 n = strtoul(src, &e, 0);
 234                 if (!e || *e || e == src || n > 255)
 235                         return false;
 236         }
 237
 238         *dst = n;
 239         return true;
 240 }
 241
 242 static bool
 243 system_tos_aton(const char *src, unsigned *dst)
 244 {
 245         char *e;
 246
 247         *dst = strtoul(src, &e, 16);
 248         if (e == src || *e || *dst > 255)
 249                 return false;
 250
 251         return true;
 252 }
 253
 254 int system_init(void)
 255 {
 256         static struct event_socket rtnl_event;
 257         static struct event_socket hotplug_event;
 258
 259         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 260         system_fd_set_cloexec(sock_ioctl);
 261
 262         /* Prepare socket for routing / address control */
 263         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 264         if (!sock_rtnl)
 265                 return -1;
 266
 267         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 268                 return -1;
 269
 270         if (!create_hotplug_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT,
 271                                          handle_hotplug_event))
 272                 return -1;
 273
 274         /* Receive network link events form kernel */
 275         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 276
 277         return 0;
 278 }
 279
 280 static void system_set_sysctl(const char *path, const char *val)
 281 {
 282         int fd;
 283
 284         fd = open(path, O_WRONLY);
 285         if (fd < 0)
 286                 return;
 287
 288         if (write(fd, val, strlen(val))) {}
 289         close(fd);
 290 }
 291
 292 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 293 {
 294         snprintf(dev_buf, sizeof(dev_buf), path, device);
 295         system_set_sysctl(dev_buf, val);
 296 }
 297
 298 static void system_set_disable_ipv6(struct device *dev, const char *val)
 299 {
 300         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 301 }
 302
 303 static void system_set_rpfilter(struct device *dev, const char *val)
 304 {
 305         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter", dev->ifname, val);
 306 }
 307
 308 static void system_set_acceptlocal(struct device *dev, const char *val)
 309 {
 310         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local", dev->ifname, val);
 311 }
 312
 313 static void system_set_igmpversion(struct device *dev, const char *val)
 314 {
 315         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version", dev->ifname, val);
 316 }
 317
 318 static void system_set_mldversion(struct device *dev, const char *val)
 319 {
 320         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version", dev->ifname, val);
 321 }
 322
 323 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 324 {
 325         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 326 }
 327
 328 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 329 {
 330         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 331 }
 332
 333 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 334 {
 335         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time", dev->ifname, val);
 336 }
 337
 338 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 339 {
 340         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time", dev->ifname, val);
 341 }
 342
 343 static void system_set_neigh4locktime(struct device *dev, const char *val)
 344 {
 345         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime", dev->ifname, val);
 346 }
 347
 348 static void system_set_dadtransmits(struct device *dev, const char *val)
 349 {
 350         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 351 }
 352
 353 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 354 {
 355         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 356 }
 357
 358 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 359 {
 360         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 361 }
 362
 363 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 364 {
 365         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 366 }
 367
 368 static void system_bridge_set_isolated(struct device *dev, const char *val)
 369 {
 370         system_set_dev_sysctl("/sys/class/net/%s/brport/isolated", dev->ifname, val);
 371 }
 372
 373 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 374 {
 375         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 376                                        "/sys/class/net/%s/brport/multicast_router",
 377                               dev->ifname, val);
 378 }
 379
 380 static void system_bridge_set_robustness(struct device *dev, const char *val)
 381 {
 382         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 383                               dev->ifname, val);
 384         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 385                               dev->ifname, val);
 386 }
 387
 388 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 389 {
 390         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 391                               dev->ifname, val);
 392 }
 393
 394 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 395 {
 396         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 397                               dev->ifname, val);
 398 }
 399
 400 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 401 {
 402         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 403                               dev->ifname, val);
 404 }
 405
 406 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 407 {
 408         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 409                               dev->ifname, val);
 410 }
 411
 412 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 413 {
 414         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 415                               dev->ifname, val);
 416 }
 417
 418 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 419 {
 420         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 421                               dev->ifname, val);
 422 }
 423
 424 static void system_bridge_set_stp_state(struct device *dev, const char *val)
 425 {
 426         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/stp_state", dev->ifname, val);
 427 }
 428
 429 static void system_bridge_set_forward_delay(struct device *dev, const char *val)
 430 {
 431         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/forward_delay", dev->ifname, val);
 432 }
 433
 434 static void system_bridge_set_priority(struct device *dev, const char *val)
 435 {
 436         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/priority", dev->ifname, val);
 437 }
 438
 439 static void system_bridge_set_ageing_time(struct device *dev, const char *val)
 440 {
 441         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/ageing_time", dev->ifname, val);
 442 }
 443
 444 static void system_bridge_set_hello_time(struct device *dev, const char *val)
 445 {
 446         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hello_time", dev->ifname, val);
 447 }
 448
 449 static void system_bridge_set_max_age(struct device *dev, const char *val)
 450 {
 451         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/max_age", dev->ifname, val);
 452 }
 453
 454 static void system_bridge_set_learning(struct device *dev, const char *val)
 455 {
 456         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 457 }
 458
 459 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 460 {
 461         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 462 }
 463
 464 static void system_set_sendredirects(struct device *dev, const char *val)
 465 {
 466         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 467 }
 468
 469 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 470 {
 471         int fd = -1, ret = -1;
 472
 473         fd = open(path, O_RDONLY);
 474         if (fd < 0)
 475                 goto out;
 476
 477         ssize_t len = read(fd, buf, buf_sz - 1);
 478         if (len < 0)
 479                 goto out;
 480
 481         ret = buf[len] = 0;
 482
 483 out:
 484         if (fd >= 0)
 485                 close(fd);
 486
 487         return ret;
 488 }
 489
 490 static int
 491 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 492 {
 493         snprintf(dev_buf, sizeof(dev_buf), path, device);
 494         return system_get_sysctl(dev_buf, buf, buf_sz);
 495 }
 496
 497 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 498 {
 499         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 500                         dev->ifname, buf, buf_sz);
 501 }
 502
 503 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 504 {
 505         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 506                         dev->ifname, buf, buf_sz);
 507 }
 508
 509 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 510 {
 511         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 512                         dev->ifname, buf, buf_sz);
 513 }
 514
 515 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 516 {
 517         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 518                         dev->ifname, buf, buf_sz);
 519 }
 520
 521 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 522 {
 523         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 524                         dev->ifname, buf, buf_sz);
 525 }
 526
 527 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 528 {
 529         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 530                         dev->ifname, buf, buf_sz);
 531 }
 532
 533 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 534 {
 535         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 536                         dev->ifname, buf, buf_sz);
 537 }
 538
 539 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 540 {
 541         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 542                         dev->ifname, buf, buf_sz);
 543 }
 544
 545 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 546 {
 547         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 548                         dev->ifname, buf, buf_sz);
 549 }
 550
 551 static int system_get_neigh4locktime(struct device *dev, char *buf, const size_t buf_sz)
 552 {
 553         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime",
 554                         dev->ifname, buf, buf_sz);
 555 }
 556
 557 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 558 {
 559         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 560                         dev->ifname, buf, buf_sz);
 561 }
 562
 563 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 564 {
 565         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 566                         dev->ifname, buf, buf_sz);
 567 }
 568
 569 /* Evaluate netlink messages */
 570 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 571 {
 572         struct nlmsghdr *nh = nlmsg_hdr(msg);
 573         struct nlattr *nla[__IFLA_MAX];
 574         int link_state = 0;
 575         char buf[10];
 576
 577         if (nh->nlmsg_type != RTM_NEWLINK)
 578                 goto out;
 579
 580         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 581         if (!nla[IFLA_IFNAME])
 582                 goto out;
 583
 584         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 585         if (!dev)
 586                 goto out;
 587
 588         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 589                 link_state = strtoul(buf, NULL, 0);
 590
 591         device_set_link(dev, link_state ? true : false);
 592
 593 out:
 594         return 0;
 595 }
 596
 597 static void
 598 handle_hotplug_msg(char *data, int size)
 599 {
 600         const char *subsystem = NULL, *interface = NULL, *interface_old = NULL;
 601         char *cur, *end, *sep;
 602         struct device *dev;
 603         int skip;
 604         bool add, move = false;
 605
 606         if (!strncmp(data, "add@", 4))
 607                 add = true;
 608         else if (!strncmp(data, "remove@", 7))
 609                 add = false;
 610         else if (!strncmp(data, "move@", 5)) {
 611                 add = true;
 612                 move = true;
 613         }
 614         else
 615                 return;
 616
 617         skip = strlen(data) + 1;
 618         end = data + size;
 619
 620         for (cur = data + skip; cur < end; cur += skip) {
 621                 skip = strlen(cur) + 1;
 622
 623                 sep = strchr(cur, '=');
 624                 if (!sep)
 625                         continue;
 626
 627                 *sep = 0;
 628                 if (!strcmp(cur, "INTERFACE"))
 629                         interface = sep + 1;
 630                 else if (!strcmp(cur, "SUBSYSTEM")) {
 631                         subsystem = sep + 1;
 632                         if (strcmp(subsystem, "net") != 0)
 633                                 return;
 634                 } else if (!strcmp(cur, "DEVPATH_OLD")) {
 635                         interface_old = strrchr(sep + 1, '/');
 636                         if (interface_old)
 637                                 interface_old++;
 638                 }
 639         }
 640
 641         if (subsystem && interface) {
 642                 if (move && interface_old)
 643                         goto move;
 644                 else
 645                         goto found;
 646         }
 647
 648         return;
 649
 650 move:
 651         dev = device_find(interface_old);
 652         if (!dev)
 653                 goto found;
 654
 655         if (dev->type != &simple_device_type)
 656                 goto found;
 657
 658         device_set_present(dev, false);
 659
 660 found:
 661         dev = device_find(interface);
 662         if (!dev)
 663                 return;
 664
 665         if (dev->type != &simple_device_type)
 666                 return;
 667
 668         if (add && system_if_force_external(dev->ifname))
 669                 return;
 670
 671         device_set_present(dev, add);
 672 }
 673
 674 static void
 675 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 676 {
 677         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 678         struct sockaddr_nl nla;
 679         unsigned char *buf = NULL;
 680         int size;
 681         int err;
 682         socklen_t errlen = sizeof(err);
 683
 684         if (!u->error) {
 685                 while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 686                         if (nla.nl_pid == 0)
 687                                 handle_hotplug_msg((char *) buf, size);
 688
 689                         free(buf);
 690                 }
 691                 return;
 692         }
 693
 694         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 695                 goto abort;
 696
 697         switch(err) {
 698         case ENOBUFS:
 699                 /* Increase rx buffer size on netlink socket */
 700                 ev->bufsize *= 2;
 701                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 702                         goto abort;
 703                 break;
 704
 705         default:
 706                 goto abort;
 707         }
 708         u->error = false;
 709         return;
 710
 711 abort:
 712         uloop_fd_delete(&ev->uloop);
 713         return;
 714 }
 715
 716 static int system_rtnl_call(struct nl_msg *msg)
 717 {
 718         int ret;
 719
 720         ret = nl_send_auto_complete(sock_rtnl, msg);
 721         nlmsg_free(msg);
 722
 723         if (ret < 0)
 724                 return ret;
 725
 726         return nl_wait_for_ack(sock_rtnl);
 727 }
 728
 729 int system_bridge_delbr(struct device *bridge)
 730 {
 731         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 732 }
 733
 734 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 735 {
 736         struct ifreq ifr;
 737
 738         memset(&ifr, 0, sizeof(ifr));
 739         if (dev)
 740                 ifr.ifr_ifindex = dev->ifindex;
 741         else
 742                 ifr.ifr_data = data;
 743         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name) - 1);
 744         return ioctl(sock_ioctl, cmd, &ifr);
 745 }
 746
 747 static bool system_is_bridge(const char *name, char *buf, int buflen)
 748 {
 749         struct stat st;
 750
 751         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 752         if (stat(buf, &st) < 0)
 753                 return false;
 754
 755         return true;
 756 }
 757
 758 static char *system_get_bridge(const char *name, char *buf, int buflen)
 759 {
 760         char *path;
 761         ssize_t len = -1;
 762         glob_t gl;
 763
 764         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 765         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 766                 return NULL;
 767
 768         if (gl.gl_pathc > 0)
 769                 len = readlink(gl.gl_pathv[0], buf, buflen);
 770
 771         globfree(&gl);
 772
 773         if (len < 0)
 774                 return NULL;
 775
 776         buf[len] = 0;
 777         path = strrchr(buf, '/');
 778         if (!path)
 779                 return NULL;
 780
 781         return path + 1;
 782 }
 783
 784 static void
 785 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 786 {
 787         bool mcast_to_ucast = dev->wireless_ap;
 788         bool hairpin = true;
 789
 790         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 791             !bridge->settings.multicast_to_unicast)
 792                 mcast_to_ucast = false;
 793
 794         if (!mcast_to_ucast || dev->wireless_isolate)
 795                 hairpin = false;
 796
 797         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 798         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 799 }
 800
 801 int system_bridge_addif(struct device *bridge, struct device *dev)
 802 {
 803         char buf[64];
 804         char *oldbr;
 805         int ret = 0;
 806
 807         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 808         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 809                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 810
 811         if (dev->wireless)
 812                 system_bridge_set_wireless(bridge, dev);
 813
 814         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 815                 snprintf(buf, sizeof(buf), "%u", dev->settings.multicast_router);
 816                 system_bridge_set_multicast_router(dev, buf, false);
 817         }
 818
 819         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 820             dev->settings.multicast_fast_leave)
 821                 system_bridge_set_multicast_fast_leave(dev, "1");
 822
 823         if (dev->settings.flags & DEV_OPT_LEARNING &&
 824             !dev->settings.learning)
 825                 system_bridge_set_learning(dev, "0");
 826
 827         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 828             !dev->settings.unicast_flood)
 829                 system_bridge_set_unicast_flood(dev, "0");
 830
 831         if (dev->settings.flags & DEV_OPT_ISOLATE &&
 832             dev->settings.isolate)
 833                 system_bridge_set_isolated(dev, "1");
 834
 835         return ret;
 836 }
 837
 838 int system_bridge_delif(struct device *bridge, struct device *dev)
 839 {
 840         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 841 }
 842
 843 int system_if_resolve(struct device *dev)
 844 {
 845         struct ifreq ifr;
 846         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
 847         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 848                 return ifr.ifr_ifindex;
 849         else
 850                 return 0;
 851 }
 852
 853 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 854 {
 855         struct ifreq ifr;
 856
 857         memset(&ifr, 0, sizeof(ifr));
 858         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
 859         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) < 0)
 860                 return -1;
 861
 862         ifr.ifr_flags |= add;
 863         ifr.ifr_flags &= ~rem;
 864         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 865 }
 866
 867 struct clear_data {
 868         struct nl_msg *msg;
 869         struct device *dev;
 870         int type;
 871         int size;
 872         int af;
 873 };
 874
 875
 876 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 877 {
 878         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 879
 880         return ifa->ifa_index == ifindex;
 881 }
 882
 883 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 884 {
 885         struct rtmsg *r = NLMSG_DATA(hdr);
 886         struct nlattr *tb[__RTA_MAX];
 887
 888         if (r->rtm_protocol == RTPROT_KERNEL &&
 889             r->rtm_family == AF_INET6)
 890                 return false;
 891
 892         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 893         if (!tb[RTA_OIF])
 894                 return false;
 895
 896         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 897 }
 898
 899 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 900 {
 901         return true;
 902 }
 903
 904 static int cb_clear_event(struct nl_msg *msg, void *arg)
 905 {
 906         struct clear_data *clr = arg;
 907         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 908         bool (*cb)(struct nlmsghdr *, int ifindex);
 909         int type, ret;
 910
 911         switch(clr->type) {
 912         case RTM_GETADDR:
 913                 type = RTM_DELADDR;
 914                 if (hdr->nlmsg_type != RTM_NEWADDR)
 915                         return NL_SKIP;
 916
 917                 cb = check_ifaddr;
 918                 break;
 919         case RTM_GETROUTE:
 920                 type = RTM_DELROUTE;
 921                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 922                         return NL_SKIP;
 923
 924                 cb = check_route;
 925                 break;
 926         case RTM_GETRULE:
 927                 type = RTM_DELRULE;
 928                 if (hdr->nlmsg_type != RTM_NEWRULE)
 929                         return NL_SKIP;
 930
 931                 cb = check_rule;
 932                 break;
 933         default:
 934                 return NL_SKIP;
 935         }
 936
 937         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 938                 return NL_SKIP;
 939
 940         if (type == RTM_DELRULE)
 941                 D(SYSTEM, "Remove a rule\n");
 942         else
 943                 D(SYSTEM, "Remove %s from device %s\n",
 944                   type == RTM_DELADDR ? "an address" : "a route",
 945                   clr->dev->ifname);
 946
 947         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 948         hdr = nlmsg_hdr(clr->msg);
 949         hdr->nlmsg_type = type;
 950         hdr->nlmsg_flags = NLM_F_REQUEST;
 951
 952         nl_socket_disable_auto_ack(sock_rtnl);
 953         ret = nl_send_auto_complete(sock_rtnl, clr->msg);
 954         if (ret < 0) {
 955                 if (type == RTM_DELRULE)
 956                         D(SYSTEM, "Error deleting a rule: %d\n", ret);
 957                 else
 958                         D(SYSTEM, "Error deleting %s from device '%s': %d\n",
 959                                 type == RTM_DELADDR ? "an address" : "a route",
 960                                 clr->dev->ifname, ret);
 961         }
 962
 963         nl_socket_enable_auto_ack(sock_rtnl);
 964
 965         return NL_SKIP;
 966 }
 967
 968 static int
 969 cb_finish_event(struct nl_msg *msg, void *arg)
 970 {
 971         int *pending = arg;
 972         *pending = 0;
 973         return NL_STOP;
 974 }
 975
 976 static int
 977 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 978 {
 979         int *pending = arg;
 980         *pending = err->error;
 981         return NL_STOP;
 982 }
 983
 984 static void
 985 system_if_clear_entries(struct device *dev, int type, int af)
 986 {
 987         struct clear_data clr;
 988         struct nl_cb *cb;
 989         struct rtmsg rtm = {
 990                 .rtm_family = af,
 991                 .rtm_flags = RTM_F_CLONED,
 992         };
 993         int flags = NLM_F_DUMP;
 994         int pending = 1;
 995
 996         clr.af = af;
 997         clr.dev = dev;
 998         clr.type = type;
 999         switch (type) {
1000         case RTM_GETADDR:
1001         case RTM_GETRULE:
1002                 clr.size = sizeof(struct rtgenmsg);
1003                 break;
1004         case RTM_GETROUTE:
1005                 clr.size = sizeof(struct rtmsg);
1006                 break;
1007         default:
1008                 return;
1009         }
1010
1011         cb = nl_cb_alloc(NL_CB_DEFAULT);
1012         if (!cb)
1013                 return;
1014
1015         clr.msg = nlmsg_alloc_simple(type, flags);
1016         if (!clr.msg)
1017                 goto out;
1018
1019         nlmsg_append(clr.msg, &rtm, clr.size, 0);
1020         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
1021         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
1022         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
1023
1024         if (nl_send_auto_complete(sock_rtnl, clr.msg) < 0)
1025                 goto free;
1026
1027         while (pending > 0)
1028                 nl_recvmsgs(sock_rtnl, cb);
1029
1030 free:
1031         nlmsg_free(clr.msg);
1032 out:
1033         nl_cb_put(cb);
1034 }
1035
1036 /*
1037  * Clear bridge (membership) state and bring down device
1038  */
1039 void system_if_clear_state(struct device *dev)
1040 {
1041         static char buf[256];
1042         char *bridge;
1043         device_set_ifindex(dev, system_if_resolve(dev));
1044
1045         if (dev->external || !dev->ifindex)
1046                 return;
1047
1048         system_if_flags(dev->ifname, 0, IFF_UP);
1049
1050         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
1051                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
1052                 system_bridge_delbr(dev);
1053                 return;
1054         }
1055
1056         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
1057         if (bridge) {
1058                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
1059                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
1060         }
1061
1062         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
1063         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
1064         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
1065         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
1066         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET);
1067         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET6);
1068         system_set_disable_ipv6(dev, "0");
1069 }
1070
1071 static inline unsigned long
1072 sec_to_jiffies(int val)
1073 {
1074         return (unsigned long) val * 100;
1075 }
1076
1077 static void system_bridge_conf_multicast_deps(struct device *bridge,
1078                                               struct bridge_config *cfg,
1079                                               char *buf,
1080                                               int buf_len)
1081 {
1082         int val;
1083
1084         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
1085             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
1086             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1087                 val = cfg->robustness * cfg->query_interval +
1088                         cfg->query_response_interval;
1089
1090                 snprintf(buf, buf_len, "%i", val);
1091                 system_bridge_set_membership_interval(bridge, buf);
1092
1093                 val = cfg->robustness * cfg->query_interval +
1094                         cfg->query_response_interval / 2;
1095
1096                 snprintf(buf, buf_len, "%i", val);
1097                 system_bridge_set_other_querier_timeout(bridge, buf);
1098         }
1099
1100         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1101                 val = cfg->query_interval / 4;
1102
1103                 snprintf(buf, buf_len, "%i", val);
1104                 system_bridge_set_startup_query_interval(bridge, buf);
1105         }
1106 }
1107
1108 static void system_bridge_conf_multicast(struct device *bridge,
1109                                          struct bridge_config *cfg,
1110                                          char *buf,
1111                                          int buf_len)
1112 {
1113         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
1114                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
1115
1116         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
1117                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
1118
1119         snprintf(buf, buf_len, "%i", cfg->hash_max);
1120         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
1121                 bridge->ifname, buf);
1122
1123         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
1124                 snprintf(buf, buf_len, "%u", bridge->settings.multicast_router);
1125                 system_bridge_set_multicast_router(bridge, buf, true);
1126         }
1127
1128         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
1129                 snprintf(buf, buf_len, "%i", cfg->robustness);
1130                 system_bridge_set_robustness(bridge, buf);
1131         }
1132
1133         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1134                 snprintf(buf, buf_len, "%i", cfg->query_interval);
1135                 system_bridge_set_query_interval(bridge, buf);
1136         }
1137
1138         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1139                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
1140                 system_bridge_set_query_response_interval(bridge, buf);
1141         }
1142
1143         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1144                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1145                 system_bridge_set_last_member_interval(bridge, buf);
1146         }
1147
1148         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1149 }
1150
1151 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1152 {
1153         char buf[64];
1154
1155         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1156                 return -1;
1157
1158         system_bridge_set_stp_state(bridge, cfg->stp ? "1" : "0");
1159
1160         snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->forward_delay));
1161         system_bridge_set_forward_delay(bridge, buf);
1162
1163         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1164
1165         snprintf(buf, sizeof(buf), "%d", cfg->priority);
1166         system_bridge_set_priority(bridge, buf);
1167
1168         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1169                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->ageing_time));
1170                 system_bridge_set_ageing_time(bridge, buf);
1171         }
1172
1173         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1174                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->hello_time));
1175                 system_bridge_set_hello_time(bridge, buf);
1176         }
1177
1178         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1179                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->max_age));
1180                 system_bridge_set_max_age(bridge, buf);
1181         }
1182
1183         return 0;
1184 }
1185
1186 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1187 {
1188         struct nl_msg *msg;
1189         struct nlattr *linkinfo, *data;
1190         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1191         int i, rv;
1192         static const struct {
1193                 const char *name;
1194                 enum macvlan_mode val;
1195         } modes[] = {
1196                 { "private", MACVLAN_MODE_PRIVATE },
1197                 { "vepa", MACVLAN_MODE_VEPA },
1198                 { "bridge", MACVLAN_MODE_BRIDGE },
1199                 { "passthru", MACVLAN_MODE_PASSTHRU },
1200         };
1201
1202         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1203
1204         if (!msg)
1205                 return -1;
1206
1207         nlmsg_append(msg, &iim, sizeof(iim), 0);
1208
1209         if (cfg->flags & MACVLAN_OPT_MACADDR)
1210                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1211         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1212         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1213
1214         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1215                 goto nla_put_failure;
1216
1217         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1218
1219         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1220                 goto nla_put_failure;
1221
1222         if (cfg->mode) {
1223                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1224                         if (strcmp(cfg->mode, modes[i].name) != 0)
1225                                 continue;
1226
1227                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1228                         break;
1229                 }
1230         }
1231
1232         nla_nest_end(msg, data);
1233         nla_nest_end(msg, linkinfo);
1234
1235         rv = system_rtnl_call(msg);
1236         if (rv)
1237                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1238
1239         return rv;
1240
1241 nla_put_failure:
1242         nlmsg_free(msg);
1243         return -ENOMEM;
1244 }
1245
1246 static int system_link_del(const char *ifname)
1247 {
1248         struct nl_msg *msg;
1249         struct ifinfomsg iim = {
1250                 .ifi_family = AF_UNSPEC,
1251                 .ifi_index = 0,
1252         };
1253
1254         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1255
1256         if (!msg)
1257                 return -1;
1258
1259         nlmsg_append(msg, &iim, sizeof(iim), 0);
1260         nla_put_string(msg, IFLA_IFNAME, ifname);
1261         return system_rtnl_call(msg);
1262 }
1263
1264 int system_macvlan_del(struct device *macvlan)
1265 {
1266         return system_link_del(macvlan->ifname);
1267 }
1268
1269 int system_veth_add(struct device *veth, struct veth_config *cfg)
1270 {
1271         struct nl_msg *msg;
1272         struct ifinfomsg empty_iim = {};
1273         struct nlattr *linkinfo, *data, *veth_info;
1274         int rv;
1275
1276         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1277
1278         if (!msg)
1279                 return -1;
1280
1281         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1282
1283         if (cfg->flags & VETH_OPT_MACADDR)
1284                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1285         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1286
1287         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1288                 goto nla_put_failure;
1289
1290         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1291
1292         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1293                 goto nla_put_failure;
1294
1295         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1296                 goto nla_put_failure;
1297
1298         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1299
1300         if (cfg->flags & VETH_OPT_PEER_NAME)
1301                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1302         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1303                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1304
1305         nla_nest_end(msg, veth_info);
1306         nla_nest_end(msg, data);
1307         nla_nest_end(msg, linkinfo);
1308
1309         rv = system_rtnl_call(msg);
1310         if (rv) {
1311                 if (cfg->flags & VETH_OPT_PEER_NAME)
1312                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1313                 else
1314                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1315         }
1316
1317         return rv;
1318
1319 nla_put_failure:
1320         nlmsg_free(msg);
1321         return -ENOMEM;
1322 }
1323
1324 int system_veth_del(struct device *veth)
1325 {
1326         return system_link_del(veth->ifname);
1327 }
1328
1329 static int system_vlan(struct device *dev, int id)
1330 {
1331         struct vlan_ioctl_args ifr = {
1332                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1333                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1334         };
1335
1336         if (ioctl(sock_ioctl, SIOCSIFVLAN, &ifr) < 0)
1337                 return -1;
1338
1339         if (id < 0) {
1340                 ifr.cmd = DEL_VLAN_CMD;
1341                 ifr.u.VID = 0;
1342         } else {
1343                 ifr.cmd = ADD_VLAN_CMD;
1344                 ifr.u.VID = id;
1345         }
1346         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1347         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1348 }
1349
1350 int system_vlan_add(struct device *dev, int id)
1351 {
1352         return system_vlan(dev, id);
1353 }
1354
1355 int system_vlan_del(struct device *dev)
1356 {
1357         return system_vlan(dev, -1);
1358 }
1359
1360 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1361 {
1362         struct nl_msg *msg;
1363         struct nlattr *linkinfo, *data;
1364         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1365         int rv;
1366
1367         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1368
1369         if (!msg)
1370                 return -1;
1371
1372         nlmsg_append(msg, &iim, sizeof(iim), 0);
1373         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1374         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1375
1376         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1377                 goto nla_put_failure;
1378
1379         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1380
1381         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1382                 goto nla_put_failure;
1383
1384         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1385
1386 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1387         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1388 #else
1389         if(cfg->proto == VLAN_PROTO_8021AD)
1390                 netifd_log_message(L_WARNING, "%s Your kernel is older than linux 3.10.0, 802.1ad is not supported defaulting to 802.1q", vlandev->type->name);
1391 #endif
1392
1393         nla_nest_end(msg, data);
1394         nla_nest_end(msg, linkinfo);
1395
1396         rv = system_rtnl_call(msg);
1397         if (rv)
1398                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1399
1400         return rv;
1401
1402 nla_put_failure:
1403         nlmsg_free(msg);
1404         return -ENOMEM;
1405 }
1406
1407 int system_vlandev_del(struct device *vlandev)
1408 {
1409         return system_link_del(vlandev->ifname);
1410 }
1411
1412 void
1413 system_if_get_settings(struct device *dev, struct device_settings *s)
1414 {
1415         struct ifreq ifr;
1416         char buf[10];
1417
1418         memset(&ifr, 0, sizeof(ifr));
1419         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1420
1421         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1422                 s->mtu = ifr.ifr_mtu;
1423                 s->flags |= DEV_OPT_MTU;
1424         }
1425
1426         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1427         if (s->mtu6 > 0)
1428                 s->flags |= DEV_OPT_MTU6;
1429
1430         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1431                 s->txqueuelen = ifr.ifr_qlen;
1432                 s->flags |= DEV_OPT_TXQUEUELEN;
1433         }
1434
1435         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1436                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1437                 s->flags |= DEV_OPT_MACADDR;
1438         }
1439
1440         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1441                 s->ipv6 = !strtoul(buf, NULL, 0);
1442                 s->flags |= DEV_OPT_IPV6;
1443         }
1444
1445         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1446                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1447                 s->flags |= DEV_OPT_PROMISC;
1448
1449                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1450                 s->flags |= DEV_OPT_MULTICAST;
1451         }
1452
1453         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1454                 s->rpfilter = strtoul(buf, NULL, 0);
1455                 s->flags |= DEV_OPT_RPFILTER;
1456         }
1457
1458         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1459                 s->acceptlocal = strtoul(buf, NULL, 0);
1460                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1461         }
1462
1463         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1464                 s->igmpversion = strtoul(buf, NULL, 0);
1465                 s->flags |= DEV_OPT_IGMPVERSION;
1466         }
1467
1468         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1469                 s->mldversion = strtoul(buf, NULL, 0);
1470                 s->flags |= DEV_OPT_MLDVERSION;
1471         }
1472
1473         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1474                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1475                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1476         }
1477
1478         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1479                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1480                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1481         }
1482
1483         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
1484                 s->neigh4locktime = strtol(buf, NULL, 0);
1485                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
1486         }
1487
1488         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1489                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1490                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1491         }
1492
1493         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1494                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1495                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1496         }
1497
1498         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1499                 s->dadtransmits = strtoul(buf, NULL, 0);
1500                 s->flags |= DEV_OPT_DADTRANSMITS;
1501         }
1502
1503         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1504                 s->sendredirects = strtoul(buf, NULL, 0);
1505                 s->flags |= DEV_OPT_SENDREDIRECTS;
1506         }
1507 }
1508
1509 void
1510 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1511 {
1512         struct ifreq ifr;
1513         char buf[12];
1514
1515         memset(&ifr, 0, sizeof(ifr));
1516         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1517         if (s->flags & DEV_OPT_MTU & apply_mask) {
1518                 ifr.ifr_mtu = s->mtu;
1519                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1520                         s->flags &= ~DEV_OPT_MTU;
1521         }
1522         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1523                 system_update_ipv6_mtu(dev, s->mtu6);
1524         }
1525         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1526                 ifr.ifr_qlen = s->txqueuelen;
1527                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1528                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1529         }
1530         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1531                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1532                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1533                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1534                         s->flags &= ~DEV_OPT_MACADDR;
1535         }
1536         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1537                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1538         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1539                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1540                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1541                         s->flags &= ~DEV_OPT_PROMISC;
1542         }
1543         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1544                 snprintf(buf, sizeof(buf), "%u", s->rpfilter);
1545                 system_set_rpfilter(dev, buf);
1546         }
1547         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1548                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1549         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1550                 snprintf(buf, sizeof(buf), "%u", s->igmpversion);
1551                 system_set_igmpversion(dev, buf);
1552         }
1553         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1554                 snprintf(buf, sizeof(buf), "%u", s->mldversion);
1555                 system_set_mldversion(dev, buf);
1556         }
1557         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1558                 snprintf(buf, sizeof(buf), "%u", s->neigh4reachabletime);
1559                 system_set_neigh4reachabletime(dev, buf);
1560                 snprintf(buf, sizeof(buf), "%u", s->neigh6reachabletime);
1561                 system_set_neigh6reachabletime(dev, buf);
1562         }
1563         if (s->flags & DEV_OPT_NEIGHLOCKTIME & apply_mask) {
1564                 snprintf(buf, sizeof(buf), "%d", s->neigh4locktime);
1565                 system_set_neigh4locktime(dev, buf);
1566         }
1567         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1568                 snprintf(buf, sizeof(buf), "%u", s->neigh4gcstaletime);
1569                 system_set_neigh4gcstaletime(dev, buf);
1570                 snprintf(buf, sizeof(buf), "%u", s->neigh6gcstaletime);
1571                 system_set_neigh6gcstaletime(dev, buf);
1572         }
1573         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1574                 snprintf(buf, sizeof(buf), "%u", s->dadtransmits);
1575                 system_set_dadtransmits(dev, buf);
1576         }
1577         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1578                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1579                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1580                         s->flags &= ~DEV_OPT_MULTICAST;
1581         }
1582         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1583                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1584 }
1585
1586 int system_if_up(struct device *dev)
1587 {
1588         system_if_get_settings(dev, &dev->orig_settings);
1589         /* Only keep orig settings based on what needs to be set */
1590         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1591         dev->orig_settings.flags &= dev->settings.flags;
1592         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1593         return system_if_flags(dev->ifname, IFF_UP, 0);
1594 }
1595
1596 int system_if_down(struct device *dev)
1597 {
1598         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1599         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1600         return ret;
1601 }
1602
1603 struct if_check_data {
1604         struct device *dev;
1605         int pending;
1606         int ret;
1607 };
1608
1609 #ifndef IFF_LOWER_UP
1610 #define IFF_LOWER_UP    0x10000
1611 #endif
1612
1613 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1614 {
1615         struct nlmsghdr *nh = nlmsg_hdr(msg);
1616         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1617         struct if_check_data *chk = (struct if_check_data *)arg;
1618
1619         if (nh->nlmsg_type != RTM_NEWLINK)
1620                 return NL_SKIP;
1621
1622         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1623         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1624
1625         return NL_OK;
1626 }
1627
1628 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1629 {
1630         struct if_check_data *chk = (struct if_check_data *)arg;
1631         chk->pending = 0;
1632         return NL_STOP;
1633 }
1634
1635 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1636 {
1637         struct if_check_data *chk = (struct if_check_data *)arg;
1638
1639         device_set_present(chk->dev, false);
1640         device_set_link(chk->dev, false);
1641         chk->pending = err->error;
1642
1643         return NL_STOP;
1644 }
1645
1646 int system_if_check(struct device *dev)
1647 {
1648         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1649         struct nl_msg *msg;
1650         struct ifinfomsg ifi = {
1651                 .ifi_family = AF_UNSPEC,
1652                 .ifi_index = 0,
1653         };
1654         struct if_check_data chk = {
1655                 .dev = dev,
1656                 .pending = 1,
1657         };
1658         int ret = 1;
1659
1660         if (!cb)
1661                 return ret;
1662
1663         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1664         if (!msg)
1665                 goto out;
1666
1667         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1668             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1669                 goto free;
1670
1671         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1672         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1673         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1674
1675         ret = nl_send_auto_complete(sock_rtnl, msg);
1676         if (ret < 0)
1677                 goto free;
1678
1679         while (chk.pending > 0)
1680                 nl_recvmsgs(sock_rtnl, cb);
1681
1682         ret = chk.pending;
1683
1684 free:
1685         nlmsg_free(msg);
1686 out:
1687         nl_cb_put(cb);
1688         return ret;
1689 }
1690
1691 struct device *
1692 system_if_get_parent(struct device *dev)
1693 {
1694         char buf[64], *devname;
1695         int ifindex, iflink, len;
1696         FILE *f;
1697
1698         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1699         f = fopen(buf, "r");
1700         if (!f)
1701                 return NULL;
1702
1703         len = fread(buf, 1, sizeof(buf) - 1, f);
1704         fclose(f);
1705
1706         if (len <= 0)
1707                 return NULL;
1708
1709         buf[len] = 0;
1710         iflink = strtoul(buf, NULL, 0);
1711         ifindex = system_if_resolve(dev);
1712         if (!iflink || iflink == ifindex)
1713                 return NULL;
1714
1715         devname = if_indextoname(iflink, buf);
1716         if (!devname)
1717                 return NULL;
1718
1719         return device_get(devname, true);
1720 }
1721
1722 static bool
1723 read_string_file(int dir_fd, const char *file, char *buf, int len)
1724 {
1725         bool ret = false;
1726         char *c;
1727         int fd;
1728
1729         fd = openat(dir_fd, file, O_RDONLY);
1730         if (fd < 0)
1731                 return false;
1732
1733 retry:
1734         len = read(fd, buf, len - 1);
1735         if (len < 0) {
1736                 if (errno == EINTR)
1737                         goto retry;
1738         } else if (len > 0) {
1739                         buf[len] = 0;
1740
1741                         c = strchr(buf, '\n');
1742                         if (c)
1743                                 *c = 0;
1744
1745                         ret = true;
1746         }
1747
1748         close(fd);
1749
1750         return ret;
1751 }
1752
1753 static bool
1754 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1755 {
1756         char buf[64];
1757         bool ret = false;
1758
1759         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1760         if (ret)
1761                 *val = strtoull(buf, NULL, 0);
1762
1763         return ret;
1764 }
1765
1766 /* Assume advertised flags == supported flags */
1767 static const struct {
1768         uint32_t mask;
1769         const char *name;
1770 } ethtool_link_modes[] = {
1771         { ADVERTISED_10baseT_Half, "10baseT-H" },
1772         { ADVERTISED_10baseT_Full, "10baseT-F" },
1773         { ADVERTISED_100baseT_Half, "100baseT-H" },
1774         { ADVERTISED_100baseT_Full, "100baseT-F" },
1775         { ADVERTISED_1000baseT_Half, "1000baseT-H" },
1776         { ADVERTISED_1000baseT_Full, "1000baseT-F" },
1777         { ADVERTISED_1000baseKX_Full, "1000baseKX-F" },
1778         { ADVERTISED_2500baseX_Full, "2500baseX-F" },
1779         { ADVERTISED_10000baseT_Full, "10000baseT-F" },
1780         { ADVERTISED_10000baseKX4_Full, "10000baseKX4-F" },
1781         { ADVERTISED_10000baseKR_Full, "10000baseKR-F" },
1782         { ADVERTISED_20000baseMLD2_Full, "20000baseMLD2-F" },
1783         { ADVERTISED_20000baseKR2_Full, "20000baseKR2-F" },
1784         { ADVERTISED_40000baseKR4_Full, "40000baseKR4-F" },
1785         { ADVERTISED_40000baseCR4_Full, "40000baseCR4-F" },
1786         { ADVERTISED_40000baseSR4_Full, "40000baseSR4-F" },
1787         { ADVERTISED_40000baseLR4_Full, "40000baseLR4-F" },
1788 #ifdef ADVERTISED_56000baseKR4_Full
1789         { ADVERTISED_56000baseKR4_Full, "56000baseKR4-F" },
1790         { ADVERTISED_56000baseCR4_Full, "56000baseCR4-F" },
1791         { ADVERTISED_56000baseSR4_Full, "56000baseSR4-F" },
1792         { ADVERTISED_56000baseLR4_Full, "56000baseLR4-F" },
1793 #endif
1794 };
1795
1796 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1797 {
1798         int i;
1799         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1800                 if (mask & ethtool_link_modes[i].mask)
1801                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1802         }
1803 }
1804
1805 bool
1806 system_if_force_external(const char *ifname)
1807 {
1808         char buf[64];
1809         struct stat s;
1810
1811         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1812         return stat(buf, &s) == 0;
1813 }
1814
1815 int
1816 system_if_dump_info(struct device *dev, struct blob_buf *b)
1817 {
1818         struct ethtool_cmd ecmd;
1819         struct ifreq ifr;
1820         char *s;
1821         void *c;
1822
1823         memset(&ecmd, 0, sizeof(ecmd));
1824         memset(&ifr, 0, sizeof(ifr));
1825         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1826         ifr.ifr_data = (caddr_t) &ecmd;
1827         ecmd.cmd = ETHTOOL_GSET;
1828
1829         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1830                 c = blobmsg_open_array(b, "link-advertising");
1831                 system_add_link_modes(b, ecmd.advertising);
1832                 blobmsg_close_array(b, c);
1833
1834                 c = blobmsg_open_array(b, "link-partner-advertising");
1835                 system_add_link_modes(b, ecmd.lp_advertising);
1836                 blobmsg_close_array(b, c);
1837
1838                 c = blobmsg_open_array(b, "link-supported");
1839                 system_add_link_modes(b, ecmd.supported);
1840                 blobmsg_close_array(b, c);
1841
1842                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1843                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1844                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1845                 blobmsg_add_string_buffer(b);
1846
1847                 blobmsg_add_u8(b, "autoneg", !!ecmd.autoneg);
1848         }
1849
1850         return 0;
1851 }
1852
1853 int
1854 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1855 {
1856         const char *const counters[] = {
1857                 "collisions",     "rx_frame_errors",   "tx_compressed",
1858                 "multicast",      "rx_length_errors",  "tx_dropped",
1859                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1860                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1861                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1862                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1863                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1864                 "rx_fifo_errors", "tx_carrier_errors",
1865         };
1866         char buf[64];
1867         int stats_dir;
1868         int i;
1869         uint64_t val = 0;
1870
1871         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1872         stats_dir = open(buf, O_DIRECTORY);
1873         if (stats_dir < 0)
1874                 return -1;
1875
1876         for (i = 0; i < ARRAY_SIZE(counters); i++)
1877                 if (read_uint64_file(stats_dir, counters[i], &val))
1878                         blobmsg_add_u64(b, counters[i], val);
1879
1880         close(stats_dir);
1881         return 0;
1882 }
1883
1884 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1885 {
1886         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1887         int alen = v4 ? 4 : 16;
1888         unsigned int flags = 0;
1889         struct ifaddrmsg ifa = {
1890                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1891                 .ifa_prefixlen = addr->mask,
1892                 .ifa_index = dev->ifindex,
1893         };
1894
1895         struct nl_msg *msg;
1896         if (cmd == RTM_NEWADDR)
1897                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1898
1899         msg = nlmsg_alloc_simple(cmd, flags);
1900         if (!msg)
1901                 return -1;
1902
1903         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1904         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1905         if (v4) {
1906                 if (addr->broadcast)
1907                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1908                 if (addr->point_to_point)
1909                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1910         } else {
1911                 time_t now = system_get_rtime();
1912                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1913
1914                 if (addr->preferred_until) {
1915                         int64_t preferred = addr->preferred_until - now;
1916                         if (preferred < 0)
1917                                 preferred = 0;
1918                         else if (preferred > UINT32_MAX)
1919                                 preferred = UINT32_MAX;
1920
1921                         cinfo.ifa_prefered = preferred;
1922                 }
1923
1924                 if (addr->valid_until) {
1925                         int64_t valid = addr->valid_until - now;
1926                         if (valid <= 0) {
1927                                 nlmsg_free(msg);
1928                                 return -1;
1929                         }
1930                         else if (valid > UINT32_MAX)
1931                                 valid = UINT32_MAX;
1932
1933                         cinfo.ifa_valid = valid;
1934                 }
1935
1936                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
1937
1938                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
1939                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
1940         }
1941
1942         return system_rtnl_call(msg);
1943 }
1944
1945 int system_add_address(struct device *dev, struct device_addr *addr)
1946 {
1947         return system_addr(dev, addr, RTM_NEWADDR);
1948 }
1949
1950 int system_del_address(struct device *dev, struct device_addr *addr)
1951 {
1952         return system_addr(dev, addr, RTM_DELADDR);
1953 }
1954
1955 static int system_neigh(struct device *dev, struct device_neighbor *neighbor, int cmd)
1956 {
1957         int alen = ((neighbor->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1958         unsigned int flags = 0;
1959         struct ndmsg ndm = {
1960                 .ndm_family = (alen == 4) ? AF_INET : AF_INET6,
1961                 .ndm_ifindex = dev->ifindex,
1962                 .ndm_state = NUD_PERMANENT,
1963                 .ndm_flags = (neighbor->proxy ? NTF_PROXY : 0) | (neighbor->router ? NTF_ROUTER : 0),
1964         };
1965         struct nl_msg *msg;
1966
1967         if (cmd == RTM_NEWNEIGH)
1968                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1969
1970         msg = nlmsg_alloc_simple(cmd, flags);
1971
1972         if (!msg)
1973                 return -1;
1974
1975         nlmsg_append(msg, &ndm, sizeof(ndm), 0);
1976
1977         nla_put(msg, NDA_DST, alen, &neighbor->addr);
1978         if (neighbor->flags & DEVNEIGH_MAC)
1979                 nla_put(msg, NDA_LLADDR, sizeof(neighbor->macaddr), &neighbor->macaddr);
1980
1981
1982         return system_rtnl_call(msg);
1983 }
1984
1985 int system_add_neighbor(struct device *dev, struct device_neighbor *neighbor)
1986 {
1987         return system_neigh(dev, neighbor, RTM_NEWNEIGH);
1988 }
1989
1990 int system_del_neighbor(struct device *dev, struct device_neighbor *neighbor)
1991 {
1992         return system_neigh(dev, neighbor, RTM_DELNEIGH);
1993 }
1994
1995 static int system_rt(struct device *dev, struct device_route *route, int cmd)
1996 {
1997         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1998         bool have_gw;
1999         unsigned int flags = 0;
2000
2001         if (alen == 4)
2002                 have_gw = !!route->nexthop.in.s_addr;
2003         else
2004                 have_gw = route->nexthop.in6.s6_addr32[0] ||
2005                         route->nexthop.in6.s6_addr32[1] ||
2006                         route->nexthop.in6.s6_addr32[2] ||
2007                         route->nexthop.in6.s6_addr32[3];
2008
2009         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
2010                         ? route->table : RT_TABLE_MAIN;
2011
2012         struct rtmsg rtm = {
2013                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2014                 .rtm_dst_len = route->mask,
2015                 .rtm_src_len = route->sourcemask,
2016                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
2017                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
2018                 .rtm_scope = RT_SCOPE_NOWHERE,
2019                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
2020                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
2021         };
2022         struct nl_msg *msg;
2023
2024         if (cmd == RTM_NEWROUTE) {
2025                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
2026
2027                 if (!dev) { /* Add null-route */
2028                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2029                         rtm.rtm_type = RTN_UNREACHABLE;
2030                 }
2031                 else
2032                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
2033         }
2034
2035         if (route->flags & DEVROUTE_TYPE) {
2036                 rtm.rtm_type = route->type;
2037                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
2038                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
2039                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
2040                                 rtm.rtm_table = RT_TABLE_LOCAL;
2041                 }
2042
2043                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
2044                         rtm.rtm_scope = RT_SCOPE_HOST;
2045                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
2046                                 rtm.rtm_type == RTN_ANYCAST) {
2047                         rtm.rtm_scope = RT_SCOPE_LINK;
2048                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
2049                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY ||
2050                                 rtm.rtm_type == RTN_THROW) {
2051                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2052                         dev = NULL;
2053                 }
2054         }
2055
2056         msg = nlmsg_alloc_simple(cmd, flags);
2057         if (!msg)
2058                 return -1;
2059
2060         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2061
2062         if (route->mask)
2063                 nla_put(msg, RTA_DST, alen, &route->addr);
2064
2065         if (route->sourcemask) {
2066                 if (rtm.rtm_family == AF_INET)
2067                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
2068                 else
2069                         nla_put(msg, RTA_SRC, alen, &route->source);
2070         }
2071
2072         if (route->metric > 0)
2073                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
2074
2075         if (have_gw)
2076                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
2077
2078         if (dev)
2079                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
2080
2081         if (table >= 256)
2082                 nla_put_u32(msg, RTA_TABLE, table);
2083
2084         if (route->flags & DEVROUTE_MTU) {
2085                 struct nlattr *metrics;
2086
2087                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
2088                         goto nla_put_failure;
2089
2090                 nla_put_u32(msg, RTAX_MTU, route->mtu);
2091
2092                 nla_nest_end(msg, metrics);
2093         }
2094
2095         return system_rtnl_call(msg);
2096
2097 nla_put_failure:
2098         nlmsg_free(msg);
2099         return -ENOMEM;
2100 }
2101
2102 int system_add_route(struct device *dev, struct device_route *route)
2103 {
2104         return system_rt(dev, route, RTM_NEWROUTE);
2105 }
2106
2107 int system_del_route(struct device *dev, struct device_route *route)
2108 {
2109         return system_rt(dev, route, RTM_DELROUTE);
2110 }
2111
2112 int system_flush_routes(void)
2113 {
2114         const char *names[] = {
2115                 "/proc/sys/net/ipv4/route/flush",
2116                 "/proc/sys/net/ipv6/route/flush"
2117         };
2118         int fd, i;
2119
2120         for (i = 0; i < ARRAY_SIZE(names); i++) {
2121                 fd = open(names[i], O_WRONLY);
2122                 if (fd < 0)
2123                         continue;
2124
2125                 if (write(fd, "-1", 2)) {}
2126                 close(fd);
2127         }
2128         return 0;
2129 }
2130
2131 bool system_resolve_rt_type(const char *type, unsigned int *id)
2132 {
2133         return system_rtn_aton(type, id);
2134 }
2135
2136 bool system_resolve_rt_proto(const char *type, unsigned int *id)
2137 {
2138         FILE *f;
2139         char *e, buf[128];
2140         unsigned int n, proto = 256;
2141         n = strtoul(type, &e, 0);
2142         if (!*e && e != type)
2143                 proto = n;
2144         else if (!strcmp(type, "unspec"))
2145                 proto = RTPROT_UNSPEC;
2146         else if (!strcmp(type, "kernel"))
2147                 proto = RTPROT_KERNEL;
2148         else if (!strcmp(type, "boot"))
2149                 proto = RTPROT_BOOT;
2150         else if (!strcmp(type, "static"))
2151                 proto = RTPROT_STATIC;
2152         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
2153                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
2154                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2155                                 continue;
2156
2157                         n = strtoul(e, NULL, 10);
2158                         e = strtok(NULL, " \t\n");
2159
2160                         if (e && !strcmp(e, type)) {
2161                                 proto = n;
2162                                 break;
2163                         }
2164                 }
2165                 fclose(f);
2166         }
2167
2168         if (proto > 255)
2169                 return false;
2170
2171         *id = proto;
2172         return true;
2173 }
2174
2175 bool system_resolve_rt_table(const char *name, unsigned int *id)
2176 {
2177         FILE *f;
2178         char *e, buf[128];
2179         unsigned int n, table = RT_TABLE_UNSPEC;
2180
2181         /* first try to parse table as number */
2182         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2183                 table = n;
2184
2185         /* handle well known aliases */
2186         else if (!strcmp(name, "default"))
2187                 table = RT_TABLE_DEFAULT;
2188         else if (!strcmp(name, "main"))
2189                 table = RT_TABLE_MAIN;
2190         else if (!strcmp(name, "local"))
2191                 table = RT_TABLE_LOCAL;
2192
2193         /* try to look up name in /etc/iproute2/rt_tables */
2194         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2195         {
2196                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2197                 {
2198                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2199                                 continue;
2200
2201                         n = strtoul(e, NULL, 10);
2202                         e = strtok(NULL, " \t\n");
2203
2204                         if (e && !strcmp(e, name))
2205                         {
2206                                 table = n;
2207                                 break;
2208                         }
2209                 }
2210
2211                 fclose(f);
2212         }
2213
2214         if (table == RT_TABLE_UNSPEC)
2215                 return false;
2216
2217         *id = table;
2218         return true;
2219 }
2220
2221 bool system_is_default_rt_table(unsigned int id)
2222 {
2223         return (id == RT_TABLE_MAIN);
2224 }
2225
2226 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2227 {
2228         char *e;
2229         unsigned int n;
2230
2231         if (!strcmp(filter, "strict"))
2232                 n = 1;
2233         else if (!strcmp(filter, "loose"))
2234                 n = 2;
2235         else {
2236                 n = strtoul(filter, &e, 0);
2237                 if (*e || e == filter || n > 2)
2238                         return false;
2239         }
2240
2241         *id = n;
2242         return true;
2243 }
2244
2245 static int system_iprule(struct iprule *rule, int cmd)
2246 {
2247         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2248
2249         struct nl_msg *msg;
2250         struct rtmsg rtm = {
2251                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2252                 .rtm_protocol = RTPROT_STATIC,
2253                 .rtm_scope = RT_SCOPE_UNIVERSE,
2254                 .rtm_table = RT_TABLE_UNSPEC,
2255                 .rtm_type = RTN_UNSPEC,
2256                 .rtm_flags = 0,
2257         };
2258
2259         if (cmd == RTM_NEWRULE)
2260                 rtm.rtm_type = RTN_UNICAST;
2261
2262         if (rule->invert)
2263                 rtm.rtm_flags |= FIB_RULE_INVERT;
2264
2265         if (rule->flags & IPRULE_SRC)
2266                 rtm.rtm_src_len = rule->src_mask;
2267
2268         if (rule->flags & IPRULE_DEST)
2269                 rtm.rtm_dst_len = rule->dest_mask;
2270
2271         if (rule->flags & IPRULE_TOS)
2272                 rtm.rtm_tos = rule->tos;
2273
2274         if (rule->flags & IPRULE_LOOKUP) {
2275                 if (rule->lookup < 256)
2276                         rtm.rtm_table = rule->lookup;
2277         }
2278
2279         if (rule->flags & IPRULE_ACTION)
2280                 rtm.rtm_type = rule->action;
2281         else if (rule->flags & IPRULE_GOTO)
2282                 rtm.rtm_type = FR_ACT_GOTO;
2283         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2284                 rtm.rtm_type = FR_ACT_NOP;
2285
2286         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2287
2288         if (!msg)
2289                 return -1;
2290
2291         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2292
2293         if (rule->flags & IPRULE_IN)
2294                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2295
2296         if (rule->flags & IPRULE_OUT)
2297                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2298
2299         if (rule->flags & IPRULE_SRC)
2300                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2301
2302         if (rule->flags & IPRULE_DEST)
2303                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2304
2305         if (rule->flags & IPRULE_PRIORITY)
2306                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2307         else if (cmd == RTM_NEWRULE)
2308                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2309
2310         if (rule->flags & IPRULE_FWMARK)
2311                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2312
2313         if (rule->flags & IPRULE_FWMASK)
2314                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2315
2316         if (rule->flags & IPRULE_LOOKUP) {
2317                 if (rule->lookup >= 256)
2318                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2319         }
2320
2321         if (rule->flags & IPRULE_SUP_PREFIXLEN)
2322                 nla_put_u32(msg, FRA_SUPPRESS_PREFIXLEN, rule->sup_prefixlen);
2323
2324         if (rule->flags & IPRULE_GOTO)
2325                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2326
2327         return system_rtnl_call(msg);
2328 }
2329
2330 int system_add_iprule(struct iprule *rule)
2331 {
2332         return system_iprule(rule, RTM_NEWRULE);
2333 }
2334
2335 int system_del_iprule(struct iprule *rule)
2336 {
2337         return system_iprule(rule, RTM_DELRULE);
2338 }
2339
2340 int system_flush_iprules(void)
2341 {
2342         int rv = 0;
2343         struct iprule rule;
2344
2345         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2346         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2347
2348         memset(&rule, 0, sizeof(rule));
2349
2350
2351         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2352
2353         rule.priority = 0;
2354         rule.lookup = RT_TABLE_LOCAL;
2355         rv |= system_iprule(&rule, RTM_NEWRULE);
2356
2357         rule.priority = 32766;
2358         rule.lookup = RT_TABLE_MAIN;
2359         rv |= system_iprule(&rule, RTM_NEWRULE);
2360
2361         rule.priority = 32767;
2362         rule.lookup = RT_TABLE_DEFAULT;
2363         rv |= system_iprule(&rule, RTM_NEWRULE);
2364
2365
2366         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2367
2368         rule.priority = 0;
2369         rule.lookup = RT_TABLE_LOCAL;
2370         rv |= system_iprule(&rule, RTM_NEWRULE);
2371
2372         rule.priority = 32766;
2373         rule.lookup = RT_TABLE_MAIN;
2374         rv |= system_iprule(&rule, RTM_NEWRULE);
2375
2376         return rv;
2377 }
2378
2379 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2380 {
2381         return system_rtn_aton(action, id);
2382 }
2383
2384 time_t system_get_rtime(void)
2385 {
2386         struct timespec ts;
2387         struct timeval tv;
2388
2389         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
2390                 return ts.tv_sec;
2391
2392         if (gettimeofday(&tv, NULL) == 0)
2393                 return tv.tv_sec;
2394
2395         return 0;
2396 }
2397
2398 #ifndef IP_DF
2399 #define IP_DF       0x4000
2400 #endif
2401
2402 static int tunnel_ioctl(const char *name, int cmd, void *p)
2403 {
2404         struct ifreq ifr;
2405
2406         memset(&ifr, 0, sizeof(ifr));
2407         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name) - 1);
2408         ifr.ifr_ifru.ifru_data = p;
2409         return ioctl(sock_ioctl, cmd, &ifr);
2410 }
2411
2412 #ifdef IFLA_IPTUN_MAX
2413 static int system_add_ip6_tunnel(const char *name, const unsigned int link,
2414                                  struct blob_attr **tb)
2415 {
2416         struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2417                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2418         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2419         struct blob_attr *cur;
2420         int ret = 0, ttl = 0;
2421
2422         if (!nlm)
2423                 return -1;
2424
2425         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2426         nla_put_string(nlm, IFLA_IFNAME, name);
2427
2428         if (link)
2429                 nla_put_u32(nlm, IFLA_LINK, link);
2430
2431         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2432         if (!linkinfo) {
2433                 ret = -ENOMEM;
2434                 goto failure;
2435         }
2436
2437         nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2438         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2439         if (!infodata) {
2440                 ret = -ENOMEM;
2441                 goto failure;
2442         }
2443
2444         if (link)
2445                 nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2446
2447         if ((cur = tb[TUNNEL_ATTR_TTL]))
2448                 ttl = blobmsg_get_u32(cur);
2449
2450         nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2451         nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2452
2453         struct in6_addr in6buf;
2454         if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2455                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2456                         ret = -EINVAL;
2457                         goto failure;
2458                 }
2459                 nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2460         }
2461
2462         if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2463                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2464                         ret = -EINVAL;
2465                         goto failure;
2466                 }
2467                 nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2468         }
2469
2470         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2471                 struct blob_attr *tb_data[__IPIP6_DATA_ATTR_MAX];
2472                 uint32_t tun_flags = IP6_TNL_F_IGN_ENCAP_LIMIT;
2473
2474                 blobmsg_parse(ipip6_data_attr_list.params, __IPIP6_DATA_ATTR_MAX, tb_data,
2475                         blobmsg_data(cur), blobmsg_len(cur));
2476
2477                 if ((cur = tb_data[IPIP6_DATA_ENCAPLIMIT])) {
2478                         char *str = blobmsg_get_string(cur);
2479
2480                         if (strcmp(str, "ignore")) {
2481                                 char *e;
2482                                 unsigned encap_limit = strtoul(str, &e, 0);
2483
2484                                 if (e == str || *e || encap_limit > 255) {
2485                                         ret = -EINVAL;
2486                                         goto failure;
2487                                 }
2488
2489                                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, encap_limit);
2490                                 tun_flags &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2491                         }
2492                 }
2493
2494 #ifdef IFLA_IPTUN_FMR_MAX
2495                 if ((cur = tb_data[IPIP6_DATA_FMRS])) {
2496                         struct blob_attr *rcur;
2497                         unsigned rrem, fmrcnt = 0;
2498                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2499
2500                         if (!fmrs) {
2501                                 ret = -ENOMEM;
2502                                 goto failure;
2503                         }
2504
2505                         blobmsg_for_each_attr(rcur, cur, rrem) {
2506                                 struct blob_attr *tb_fmr[__FMR_DATA_ATTR_MAX], *tb_cur;
2507                                 struct in6_addr ip6prefix;
2508                                 struct in_addr ip4prefix;
2509                                 unsigned ip4len, ip6len, ealen, offset;
2510
2511                                 blobmsg_parse(fmr_data_attr_list.params, __FMR_DATA_ATTR_MAX, tb_fmr,
2512                                                 blobmsg_data(rcur), blobmsg_len(rcur));
2513
2514                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX6]) ||
2515                                                 !parse_ip_and_netmask(AF_INET6,
2516                                                         blobmsg_data(tb_cur), &ip6prefix,
2517                                                         &ip6len)) {
2518                                         ret = -EINVAL;
2519                                         goto failure;
2520                                 }
2521
2522                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX4]) ||
2523                                                 !parse_ip_and_netmask(AF_INET,
2524                                                         blobmsg_data(tb_cur), &ip4prefix,
2525                                                         &ip4len)) {
2526                                         ret = -EINVAL;
2527                                         goto failure;
2528                                 }
2529
2530                                 if (!(tb_cur = tb_fmr[FMR_DATA_EALEN])) {
2531                                         ret = -EINVAL;
2532                                         goto failure;
2533                                 }
2534                                 ealen = blobmsg_get_u32(tb_cur);
2535
2536                                 if (!(tb_cur = tb_fmr[FMR_DATA_OFFSET])) {
2537                                         ret = -EINVAL;
2538                                         goto failure;
2539                                 }
2540                                 offset = blobmsg_get_u32(tb_cur);
2541
2542                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2543                                 if (!rule) {
2544                                         ret = -ENOMEM;
2545                                         goto failure;
2546                                 }
2547
2548                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2549                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2550                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2551                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2552                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2553                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2554
2555                                 nla_nest_end(nlm, rule);
2556                         }
2557
2558                         nla_nest_end(nlm, fmrs);
2559                 }
2560 #endif
2561                 if (tun_flags)
2562                         nla_put_u32(nlm, IFLA_IPTUN_FLAGS, tun_flags);
2563         }
2564
2565         nla_nest_end(nlm, infodata);
2566         nla_nest_end(nlm, linkinfo);
2567
2568         return system_rtnl_call(nlm);
2569
2570 failure:
2571         nlmsg_free(nlm);
2572         return ret;
2573 }
2574 #endif
2575
2576 #ifdef IFLA_IPTUN_MAX
2577 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2578 static int system_add_gre_tunnel(const char *name, const char *kind,
2579                                  const unsigned int link, struct blob_attr **tb, bool v6)
2580 {
2581         struct nl_msg *nlm;
2582         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2583         struct blob_attr *cur;
2584         uint32_t ikey = 0, okey = 0, flowinfo = 0, flags6 = IP6_TNL_F_IGN_ENCAP_LIMIT;
2585         uint16_t iflags = 0, oflags = 0;
2586         uint8_t tos = 0;
2587         int ret = 0, ttl = 0;
2588         unsigned encap_limit = 0;
2589
2590         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2591         if (!nlm)
2592                 return -1;
2593
2594         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2595         nla_put_string(nlm, IFLA_IFNAME, name);
2596
2597         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2598         if (!linkinfo) {
2599                 ret = -ENOMEM;
2600                 goto failure;
2601         }
2602
2603         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2604         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2605         if (!infodata) {
2606                 ret = -ENOMEM;
2607                 goto failure;
2608         }
2609
2610         if (link)
2611                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2612
2613         if ((cur = tb[TUNNEL_ATTR_TTL]))
2614                 ttl = blobmsg_get_u32(cur);
2615
2616         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2617                 char *str = blobmsg_get_string(cur);
2618                 if (strcmp(str, "inherit")) {
2619                         unsigned uval;
2620
2621                         if (!system_tos_aton(str, &uval)) {
2622                                 ret = -EINVAL;
2623                                 goto failure;
2624                         }
2625
2626                         if (v6)
2627                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2628                         else
2629                                 tos = uval;
2630                 } else {
2631                         if (v6)
2632                                 flags6 |= IP6_TNL_F_USE_ORIG_TCLASS;
2633                         else
2634                                 tos = 1;
2635                 }
2636         }
2637
2638         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2639                 struct blob_attr *tb_data[__GRE_DATA_ATTR_MAX];
2640
2641                 blobmsg_parse(gre_data_attr_list.params, __GRE_DATA_ATTR_MAX, tb_data,
2642                         blobmsg_data(cur), blobmsg_len(cur));
2643
2644                 if ((cur = tb_data[GRE_DATA_IKEY])) {
2645                         if ((ikey = blobmsg_get_u32(cur)))
2646                                 iflags |= GRE_KEY;
2647                 }
2648
2649                 if ((cur = tb_data[GRE_DATA_OKEY])) {
2650                         if ((okey = blobmsg_get_u32(cur)))
2651                                 oflags |= GRE_KEY;
2652                 }
2653
2654                 if ((cur = tb_data[GRE_DATA_ICSUM])) {
2655                         if (blobmsg_get_bool(cur))
2656                                 iflags |= GRE_CSUM;
2657                 }
2658
2659                 if ((cur = tb_data[GRE_DATA_OCSUM])) {
2660                         if (blobmsg_get_bool(cur))
2661                                 oflags |= GRE_CSUM;
2662                 }
2663
2664                 if ((cur = tb_data[GRE_DATA_ISEQNO])) {
2665                         if (blobmsg_get_bool(cur))
2666                                 iflags |= GRE_SEQ;
2667                 }
2668
2669                 if ((cur = tb_data[GRE_DATA_OSEQNO])) {
2670                         if (blobmsg_get_bool(cur))
2671                                 oflags |= GRE_SEQ;
2672                 }
2673
2674                 if ((cur = tb_data[GRE_DATA_ENCAPLIMIT])) {
2675                         char *str = blobmsg_get_string(cur);
2676
2677                         if (strcmp(str, "ignore")) {
2678                                 char *e;
2679
2680                                 encap_limit = strtoul(str, &e, 0);
2681
2682                                 if (e == str || *e || encap_limit > 255) {
2683                                         ret = -EINVAL;
2684                                         goto failure;
2685                                 }
2686
2687                                 flags6 &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2688                         }
2689                 }
2690         }
2691
2692         if (v6) {
2693                 struct in6_addr in6buf;
2694                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2695                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2696                                 ret = -EINVAL;
2697                                 goto failure;
2698                         }
2699                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2700                 }
2701
2702                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2703                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2704                                 ret = -EINVAL;
2705                                 goto failure;
2706                         }
2707                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2708                 }
2709
2710                 if (!(flags6 & IP6_TNL_F_IGN_ENCAP_LIMIT))
2711                         nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, encap_limit);
2712
2713                 if (flowinfo)
2714                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2715
2716                 if (flags6)
2717                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags6);
2718
2719                 if (!ttl)
2720                         ttl = 64;
2721         } else {
2722                 struct in_addr inbuf;
2723                 bool set_df = true;
2724
2725                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2726                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2727                                 ret = -EINVAL;
2728                                 goto failure;
2729                         }
2730                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2731                 }
2732
2733                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2734                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2735                                 ret = -EINVAL;
2736                                 goto failure;
2737                         }
2738                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2739
2740                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2741                                 if (!okey) {
2742                                         okey = inbuf.s_addr;
2743                                         oflags |= GRE_KEY;
2744                                 }
2745
2746                                 if (!ikey) {
2747                                         ikey = inbuf.s_addr;
2748                                         iflags |= GRE_KEY;
2749                                 }
2750                         }
2751                 }
2752
2753                 if ((cur = tb[TUNNEL_ATTR_DF]))
2754                         set_df = blobmsg_get_bool(cur);
2755
2756                 if (!set_df) {
2757                         /* ttl != 0 and nopmtudisc are incompatible */
2758                         if (ttl) {
2759                                 ret = -EINVAL;
2760                                 goto failure;
2761                         }
2762                 } else if (!ttl)
2763                         ttl = 64;
2764
2765                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2766
2767                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2768         }
2769
2770         if (ttl)
2771                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2772
2773         if (oflags)
2774                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2775
2776         if (iflags)
2777                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2778
2779         if (okey)
2780                 nla_put_u32(nlm, IFLA_GRE_OKEY, htonl(okey));
2781
2782         if (ikey)
2783                 nla_put_u32(nlm, IFLA_GRE_IKEY, htonl(ikey));
2784
2785         nla_nest_end(nlm, infodata);
2786         nla_nest_end(nlm, linkinfo);
2787
2788         return system_rtnl_call(nlm);
2789
2790 failure:
2791         nlmsg_free(nlm);
2792         return ret;
2793 }
2794 #endif
2795
2796 #ifdef IFLA_VTI_MAX
2797 static int system_add_vti_tunnel(const char *name, const char *kind,
2798                                  const unsigned int link, struct blob_attr **tb, bool v6)
2799 {
2800         struct nl_msg *nlm;
2801         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2802         struct blob_attr *cur;
2803         int ret = 0;
2804
2805         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2806         if (!nlm)
2807                 return -1;
2808
2809         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2810         nla_put_string(nlm, IFLA_IFNAME, name);
2811
2812         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2813         if (!linkinfo) {
2814                 ret = -ENOMEM;
2815                 goto failure;
2816         }
2817
2818         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2819         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2820         if (!infodata) {
2821                 ret = -ENOMEM;
2822                 goto failure;
2823         }
2824
2825         if (link)
2826                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2827
2828         if (v6) {
2829                 struct in6_addr in6buf;
2830                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2831                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2832                                 ret = -EINVAL;
2833                                 goto failure;
2834                         }
2835                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2836                 }
2837
2838                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2839                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2840                                 ret = -EINVAL;
2841                                 goto failure;
2842                         }
2843                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2844                 }
2845
2846         } else {
2847                 struct in_addr inbuf;
2848
2849                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2850                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2851                                 ret = -EINVAL;
2852                                 goto failure;
2853                         }
2854                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2855                 }
2856
2857                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2858                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2859                                 ret = -EINVAL;
2860                                 goto failure;
2861                         }
2862                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2863                 }
2864
2865         }
2866
2867         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2868                 struct blob_attr *tb_data[__VTI_DATA_ATTR_MAX];
2869                 uint32_t ikey = 0, okey = 0;
2870
2871                 blobmsg_parse(vti_data_attr_list.params, __VTI_DATA_ATTR_MAX, tb_data,
2872                         blobmsg_data(cur), blobmsg_len(cur));
2873
2874                 if ((cur = tb_data[VTI_DATA_IKEY])) {
2875                         if ((ikey = blobmsg_get_u32(cur)))
2876                                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2877                 }
2878
2879                 if ((cur = tb_data[VTI_DATA_OKEY])) {
2880                         if ((okey = blobmsg_get_u32(cur)))
2881                                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2882                 }
2883         }
2884
2885         nla_nest_end(nlm, infodata);
2886         nla_nest_end(nlm, linkinfo);
2887
2888         return system_rtnl_call(nlm);
2889
2890 failure:
2891         nlmsg_free(nlm);
2892         return ret;
2893 }
2894 #endif
2895
2896 #ifdef IFLA_XFRM_MAX
2897 static int system_add_xfrm_tunnel(const char *name, const char *kind,
2898                                  const unsigned int link, struct blob_attr **tb)
2899 {
2900         struct nl_msg *nlm;
2901         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2902         struct blob_attr *cur;
2903         int ret = 0;
2904
2905         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2906         if (!nlm)
2907                 return -1;
2908
2909         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2910         nla_put_string(nlm, IFLA_IFNAME, name);
2911
2912         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2913         if (!linkinfo) {
2914                 ret = -ENOMEM;
2915                 goto failure;
2916         }
2917
2918         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2919         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2920         if (!infodata) {
2921                 ret = -ENOMEM;
2922                 goto failure;
2923         }
2924
2925         if (link)
2926                 nla_put_u32(nlm, IFLA_XFRM_LINK, link);
2927
2928         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2929                 struct blob_attr *tb_data[__XFRM_DATA_ATTR_MAX];
2930                 uint32_t if_id = 0;
2931
2932                 blobmsg_parse(xfrm_data_attr_list.params, __XFRM_DATA_ATTR_MAX, tb_data,
2933                         blobmsg_data(cur), blobmsg_len(cur));
2934
2935                 if ((cur = tb_data[XFRM_DATA_IF_ID])) {
2936                         if ((if_id = blobmsg_get_u32(cur)))
2937                                 nla_put_u32(nlm, IFLA_XFRM_IF_ID, if_id);
2938                 }
2939
2940         }
2941
2942         nla_nest_end(nlm, infodata);
2943         nla_nest_end(nlm, linkinfo);
2944
2945         return system_rtnl_call(nlm);
2946
2947 failure:
2948         nlmsg_free(nlm);
2949         return ret;
2950 }
2951 #endif
2952
2953 #ifdef IFLA_VXLAN_MAX
2954 static int system_add_vxlan(const char *name, const unsigned int link, struct blob_attr **tb, bool v6)
2955 {
2956         struct blob_attr *tb_data[__VXLAN_DATA_ATTR_MAX];
2957         struct nl_msg *msg;
2958         struct nlattr *linkinfo, *data;
2959         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
2960         struct blob_attr *cur;
2961         int ret = 0;
2962
2963         if ((cur = tb[TUNNEL_ATTR_DATA]))
2964                 blobmsg_parse(vxlan_data_attr_list.params, __VXLAN_DATA_ATTR_MAX, tb_data,
2965                         blobmsg_data(cur), blobmsg_len(cur));
2966         else
2967                 return -EINVAL;
2968
2969         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
2970
2971         if (!msg)
2972                 return -1;
2973
2974         nlmsg_append(msg, &iim, sizeof(iim), 0);
2975
2976         nla_put_string(msg, IFLA_IFNAME, name);
2977
2978         if ((cur = tb_data[VXLAN_DATA_ATTR_MACADDR])) {
2979                 struct ether_addr *ea = ether_aton(blobmsg_get_string(cur));
2980                 if (!ea) {
2981                         ret = -EINVAL;
2982                         goto failure;
2983                 }
2984
2985                 nla_put(msg, IFLA_ADDRESS, ETH_ALEN, ea);
2986         }
2987
2988         if ((cur = tb[TUNNEL_ATTR_MTU])) {
2989                 uint32_t mtu = blobmsg_get_u32(cur);
2990                 nla_put_u32(msg, IFLA_MTU, mtu);
2991         }
2992
2993         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO))) {
2994                 ret = -ENOMEM;
2995                 goto failure;
2996         }
2997
2998         nla_put_string(msg, IFLA_INFO_KIND, "vxlan");
2999
3000         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA))) {
3001                 ret = -ENOMEM;
3002                 goto failure;
3003         }
3004
3005         if (link)
3006                 nla_put_u32(msg, IFLA_VXLAN_LINK, link);
3007
3008         if ((cur = tb_data[VXLAN_DATA_ATTR_ID])) {
3009                 uint32_t id = blobmsg_get_u32(cur);
3010                 if (id >= (1u << 24) - 1) {
3011                         ret = -EINVAL;
3012                         goto failure;
3013                 }
3014
3015                 nla_put_u32(msg, IFLA_VXLAN_ID, id);
3016         }
3017
3018         if (v6) {
3019                 struct in6_addr in6buf;
3020                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3021                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3022                                 ret = -EINVAL;
3023                                 goto failure;
3024                         }
3025                         nla_put(msg, IFLA_VXLAN_LOCAL6, sizeof(in6buf), &in6buf);
3026                 }
3027
3028                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3029                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3030                                 ret = -EINVAL;
3031                                 goto failure;
3032                         }
3033                         nla_put(msg, IFLA_VXLAN_GROUP6, sizeof(in6buf), &in6buf);
3034                 }
3035         } else {
3036                 struct in_addr inbuf;
3037
3038                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3039                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3040                                 ret = -EINVAL;
3041                                 goto failure;
3042                         }
3043                         nla_put(msg, IFLA_VXLAN_LOCAL, sizeof(inbuf), &inbuf);
3044                 }
3045
3046                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3047                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3048                                 ret = -EINVAL;
3049                                 goto failure;
3050                         }
3051                         nla_put(msg, IFLA_VXLAN_GROUP, sizeof(inbuf), &inbuf);
3052                 }
3053         }
3054
3055         uint32_t port = 4789;
3056         if ((cur = tb_data[VXLAN_DATA_ATTR_PORT])) {
3057                 port = blobmsg_get_u32(cur);
3058                 if (port < 1 || port > 65535) {
3059                         ret = -EINVAL;
3060                         goto failure;
3061                 }
3062         }
3063         nla_put_u16(msg, IFLA_VXLAN_PORT, htons(port));
3064
3065         if ((cur = tb_data[VXLAN_DATA_ATTR_RXCSUM])) {
3066                 bool rxcsum = blobmsg_get_bool(cur);
3067                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_RX, !rxcsum);
3068         }
3069
3070         if ((cur = tb_data[VXLAN_DATA_ATTR_TXCSUM])) {
3071                 bool txcsum = blobmsg_get_bool(cur);
3072                 nla_put_u8(msg, IFLA_VXLAN_UDP_CSUM, txcsum);
3073                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_TX, !txcsum);
3074         }
3075
3076         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3077                 char *str = blobmsg_get_string(cur);
3078                 unsigned tos = 1;
3079
3080                 if (strcmp(str, "inherit")) {
3081                         if (!system_tos_aton(str, &tos)) {
3082                                 ret = -EINVAL;
3083                                 goto failure;
3084                         }
3085                 }
3086
3087                 nla_put_u8(msg, IFLA_VXLAN_TOS, tos);
3088         }
3089
3090         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3091                 uint32_t ttl = blobmsg_get_u32(cur);
3092                 if (ttl < 1 || ttl > 255) {
3093                         ret = -EINVAL;
3094                         goto failure;
3095                 }
3096
3097                 nla_put_u8(msg, IFLA_VXLAN_TTL, ttl);
3098         }
3099
3100         nla_nest_end(msg, data);
3101         nla_nest_end(msg, linkinfo);
3102
3103         ret = system_rtnl_call(msg);
3104         if (ret)
3105                 D(SYSTEM, "Error adding vxlan '%s': %d\n", name, ret);
3106
3107         return ret;
3108
3109 failure:
3110         nlmsg_free(msg);
3111         return ret;
3112 }
3113 #endif
3114
3115 static int system_add_sit_tunnel(const char *name, const unsigned int link, struct blob_attr **tb)
3116 {
3117         struct blob_attr *cur;
3118         int ret = 0;
3119
3120         if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
3121                 return -1;
3122
3123 #ifdef SIOCADD6RD
3124         if ((cur = tb[TUNNEL_ATTR_DATA])) {
3125                 struct blob_attr *tb_data[__SIXRD_DATA_ATTR_MAX];
3126                 unsigned int mask;
3127                 struct ip_tunnel_6rd p6;
3128
3129                 blobmsg_parse(sixrd_data_attr_list.params, __SIXRD_DATA_ATTR_MAX, tb_data,
3130                         blobmsg_data(cur), blobmsg_len(cur));
3131
3132                 memset(&p6, 0, sizeof(p6));
3133
3134                 if ((cur = tb_data[SIXRD_DATA_PREFIX])) {
3135                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
3136                                                 &p6.prefix, &mask) || mask > 128) {
3137                                 ret = -EINVAL;
3138                                 goto failure;
3139                         }
3140
3141                         p6.prefixlen = mask;
3142                 }
3143
3144                 if ((cur = tb_data[SIXRD_DATA_RELAY_PREFIX])) {
3145                         if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
3146                                                 &p6.relay_prefix, &mask) || mask > 32) {
3147                                 ret = -EINVAL;
3148                                 goto failure;
3149                         }
3150
3151                         p6.relay_prefixlen = mask;
3152                 }
3153
3154                 if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
3155                         ret = -1;
3156                         goto failure;
3157                 }
3158         }
3159 #endif
3160
3161         return ret;
3162
3163 failure:
3164         __system_del_ip_tunnel(name, tb);
3165         return ret;
3166 }
3167
3168 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
3169 {
3170         struct blob_attr *cur;
3171         bool set_df = true;
3172         struct ip_tunnel_parm p  = {
3173                 .link = link,
3174                 .iph = {
3175                         .version = 4,
3176                         .ihl = 5,
3177                         .protocol = proto,
3178                 }
3179         };
3180
3181         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
3182                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
3183                 return -EINVAL;
3184
3185         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
3186                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
3187                 return -EINVAL;
3188
3189         if ((cur = tb[TUNNEL_ATTR_DF]))
3190                 set_df = blobmsg_get_bool(cur);
3191
3192         if ((cur = tb[TUNNEL_ATTR_TTL]))
3193                 p.iph.ttl = blobmsg_get_u32(cur);
3194
3195         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3196                 char *str = blobmsg_get_string(cur);
3197                 if (strcmp(str, "inherit")) {
3198                         unsigned uval;
3199
3200                         if (!system_tos_aton(str, &uval))
3201                                 return -EINVAL;
3202
3203                         p.iph.tos = uval;
3204                 } else
3205                         p.iph.tos = 1;
3206         }
3207
3208         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
3209         /* ttl !=0 and nopmtudisc are incompatible */
3210         if (p.iph.ttl && p.iph.frag_off == 0)
3211                 return -EINVAL;
3212
3213         strncpy(p.name, name, sizeof(p.name) - 1);
3214
3215         switch (p.iph.protocol) {
3216         case IPPROTO_IPIP:
3217                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
3218         case IPPROTO_IPV6:
3219                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
3220         default:
3221                 break;
3222         }
3223         return -1;
3224 }
3225
3226 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
3227 {
3228         struct blob_attr *cur;
3229         const char *str;
3230
3231         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3232                 return -EINVAL;
3233         str = blobmsg_data(cur);
3234
3235         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
3236             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
3237             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6") ||
3238             !strcmp(str, "vxlan") || !strcmp(str, "vxlan6") ||
3239             !strcmp(str, "xfrm"))
3240                 return system_link_del(name);
3241         else
3242                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
3243 }
3244
3245 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
3246 {
3247         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3248
3249         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3250                 blob_data(attr), blob_len(attr));
3251
3252         return __system_del_ip_tunnel(name, tb);
3253 }
3254
3255 int system_update_ipv6_mtu(struct device *dev, int mtu)
3256 {
3257         int ret = -1;
3258         char buf[64];
3259         int fd;
3260
3261         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
3262                         dev->ifname);
3263
3264         fd = open(buf, O_RDWR);
3265         if (fd < 0)
3266                 return ret;
3267
3268         if (!mtu) {
3269                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
3270                 if (len < 0)
3271                         goto out;
3272
3273                 buf[len] = 0;
3274                 ret = atoi(buf);
3275         } else {
3276                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
3277                         ret = mtu;
3278         }
3279
3280 out:
3281         close(fd);
3282         return ret;
3283 }
3284
3285 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
3286 {
3287         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3288         struct blob_attr *cur;
3289         const char *str;
3290
3291         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3292                 blob_data(attr), blob_len(attr));
3293
3294         __system_del_ip_tunnel(name, tb);
3295
3296         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3297                 return -EINVAL;
3298         str = blobmsg_data(cur);
3299
3300         unsigned int ttl = 0;
3301         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3302                 ttl = blobmsg_get_u32(cur);
3303                 if (ttl > 255)
3304                         return -EINVAL;
3305         }
3306
3307         unsigned int link = 0;
3308         if ((cur = tb[TUNNEL_ATTR_LINK])) {
3309                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
3310                 if (!iface)
3311                         return -EINVAL;
3312
3313                 if (iface->l3_dev.dev)
3314                         link = iface->l3_dev.dev->ifindex;
3315         }
3316
3317         if (!strcmp(str, "sit"))
3318                 return system_add_sit_tunnel(name, link, tb);
3319 #ifdef IFLA_IPTUN_MAX
3320         else if (!strcmp(str, "ipip6")) {
3321                 return system_add_ip6_tunnel(name, link, tb);
3322         } else if (!strcmp(str, "greip")) {
3323                 return system_add_gre_tunnel(name, "gre", link, tb, false);
3324         } else if (!strcmp(str, "gretapip"))  {
3325                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
3326         } else if (!strcmp(str, "greip6")) {
3327                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
3328         } else if (!strcmp(str, "gretapip6")) {
3329                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
3330 #ifdef IFLA_VTI_MAX
3331         } else if (!strcmp(str, "vtiip")) {
3332                 return system_add_vti_tunnel(name, "vti", link, tb, false);
3333         } else if (!strcmp(str, "vtiip6")) {
3334                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
3335 #endif
3336 #ifdef IFLA_XFRM_MAX
3337         } else if (!strcmp(str, "xfrm")) {
3338                 return system_add_xfrm_tunnel(name, "xfrm", link, tb);
3339 #endif
3340 #ifdef IFLA_VXLAN_MAX
3341         } else if(!strcmp(str, "vxlan")) {
3342                 return system_add_vxlan(name, link, tb, false);
3343         } else if(!strcmp(str, "vxlan6")) {
3344                 return system_add_vxlan(name, link, tb, true);
3345 #endif
3346 #endif
3347         } else if (!strcmp(str, "ipip")) {
3348                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
3349         }
3350         else
3351                 return -EINVAL;
3352
3353         return 0;
3354 }