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 <limits.h>
  30 #include <arpa/inet.h>
  31 #include <netinet/ether.h>
  32 #include <netinet/in.h>
  33
  34 #include <linux/rtnetlink.h>
  35 #include <linux/neighbour.h>
  36 #include <linux/sockios.h>
  37 #include <linux/ip.h>
  38 #include <linux/if_addr.h>
  39 #include <linux/if_link.h>
  40 #include <linux/if_vlan.h>
  41 #include <linux/if_bridge.h>
  42 #include <linux/if_tunnel.h>
  43 #include <linux/ip6_tunnel.h>
  44 #include <linux/ethtool.h>
  45 #include <linux/fib_rules.h>
  46 #include <linux/veth.h>
  47 #include <linux/version.h>
  48
  49 #include <sched.h>
  50
  51 #ifndef RTN_FAILED_POLICY
  52 #define RTN_FAILED_POLICY 12
  53 #endif
  54
  55 #ifndef IFA_F_NOPREFIXROUTE
  56 #define IFA_F_NOPREFIXROUTE 0x200
  57 #endif
  58
  59 #ifndef IFA_FLAGS
  60 #define IFA_FLAGS (IFA_MULTICAST + 1)
  61 #endif
  62
  63 #include <string.h>
  64 #include <fcntl.h>
  65 #include <glob.h>
  66 #include <time.h>
  67 #include <unistd.h>
  68
  69 #include <netlink/msg.h>
  70 #include <netlink/attr.h>
  71 #include <netlink/socket.h>
  72 #include <libubox/uloop.h>
  73
  74 #include "netifd.h"
  75 #include "device.h"
  76 #include "system.h"
  77
  78 struct event_socket {
  79         struct uloop_fd uloop;
  80         struct nl_sock *sock;
  81         int bufsize;
  82 };
  83
  84 static int sock_ioctl = -1;
  85 static struct nl_sock *sock_rtnl = NULL;
  86
  87 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  88 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  89 static int system_add_proto_tunnel(const char *name, const uint8_t proto,
  90                                         const unsigned int link, struct blob_attr **tb);
  91 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb);
  92
  93 static char dev_buf[256];
  94
  95 static void
  96 handler_nl_event(struct uloop_fd *u, unsigned int events)
  97 {
  98         struct event_socket *ev = container_of(u, struct event_socket, uloop);
  99         int err;
 100         socklen_t errlen = sizeof(err);
 101
 102         if (!u->error) {
 103                 nl_recvmsgs_default(ev->sock);
 104                 return;
 105         }
 106
 107         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 108                 goto abort;
 109
 110         switch(err) {
 111         case ENOBUFS:
 112                 /* Increase rx buffer size on netlink socket */
 113                 ev->bufsize *= 2;
 114                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 115                         goto abort;
 116
 117                 /* Request full dump since some info got dropped */
 118                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 119                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 120                 break;
 121
 122         default:
 123                 goto abort;
 124         }
 125         u->error = false;
 126         return;
 127
 128 abort:
 129         uloop_fd_delete(&ev->uloop);
 130         return;
 131 }
 132
 133 static struct nl_sock *
 134 create_socket(int protocol, int groups)
 135 {
 136         struct nl_sock *sock;
 137
 138         sock = nl_socket_alloc();
 139         if (!sock)
 140                 return NULL;
 141
 142         if (groups)
 143                 nl_join_groups(sock, groups);
 144
 145         if (nl_connect(sock, protocol)) {
 146                 nl_socket_free(sock);
 147                 return NULL;
 148         }
 149
 150         return sock;
 151 }
 152
 153 static bool
 154 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 155                         uloop_fd_handler cb, int flags)
 156 {
 157         ev->sock = create_socket(protocol, groups);
 158         if (!ev->sock)
 159                 return false;
 160
 161         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 162         ev->uloop.cb = cb;
 163         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 164                 return false;
 165
 166         return true;
 167 }
 168
 169 static bool
 170 create_event_socket(struct event_socket *ev, int protocol,
 171                     int (*cb)(struct nl_msg *msg, void *arg))
 172 {
 173         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 174                 return false;
 175
 176         /* Install the valid custom callback handler */
 177         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 178
 179         /* Disable sequence number checking on event sockets */
 180         nl_socket_disable_seq_check(ev->sock);
 181
 182         /* Increase rx buffer size to 65K on event sockets */
 183         ev->bufsize = 65535;
 184         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 185                 return false;
 186
 187         return true;
 188 }
 189
 190 static bool
 191 create_hotplug_event_socket(struct event_socket *ev, int protocol,
 192                             void (*cb)(struct uloop_fd *u, unsigned int events))
 193 {
 194         if (!create_raw_event_socket(ev, protocol, 1, cb, ULOOP_ERROR_CB))
 195                 return false;
 196
 197         /* Increase rx buffer size to 65K on event sockets */
 198         ev->bufsize = 65535;
 199         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 200                 return false;
 201
 202         return true;
 203 }
 204
 205 static bool
 206 system_rtn_aton(const char *src, unsigned int *dst)
 207 {
 208         char *e;
 209         unsigned int n;
 210
 211         if (!strcmp(src, "local"))
 212                 n = RTN_LOCAL;
 213         else if (!strcmp(src, "nat"))
 214                 n = RTN_NAT;
 215         else if (!strcmp(src, "broadcast"))
 216                 n = RTN_BROADCAST;
 217         else if (!strcmp(src, "anycast"))
 218                 n = RTN_ANYCAST;
 219         else if (!strcmp(src, "multicast"))
 220                 n = RTN_MULTICAST;
 221         else if (!strcmp(src, "prohibit"))
 222                 n = RTN_PROHIBIT;
 223         else if (!strcmp(src, "unreachable"))
 224                 n = RTN_UNREACHABLE;
 225         else if (!strcmp(src, "blackhole"))
 226                 n = RTN_BLACKHOLE;
 227         else if (!strcmp(src, "xresolve"))
 228                 n = RTN_XRESOLVE;
 229         else if (!strcmp(src, "unicast"))
 230                 n = RTN_UNICAST;
 231         else if (!strcmp(src, "throw"))
 232                 n = RTN_THROW;
 233         else if (!strcmp(src, "failed_policy"))
 234                 n = RTN_FAILED_POLICY;
 235         else {
 236                 n = strtoul(src, &e, 0);
 237                 if (!e || *e || e == src || n > 255)
 238                         return false;
 239         }
 240
 241         *dst = n;
 242         return true;
 243 }
 244
 245 static bool
 246 system_tos_aton(const char *src, unsigned *dst)
 247 {
 248         char *e;
 249
 250         *dst = strtoul(src, &e, 16);
 251         if (e == src || *e || *dst > 255)
 252                 return false;
 253
 254         return true;
 255 }
 256
 257 int system_init(void)
 258 {
 259         static struct event_socket rtnl_event;
 260         static struct event_socket hotplug_event;
 261
 262         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 263         system_fd_set_cloexec(sock_ioctl);
 264
 265         /* Prepare socket for routing / address control */
 266         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 267         if (!sock_rtnl)
 268                 return -1;
 269
 270         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 271                 return -1;
 272
 273         if (!create_hotplug_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT,
 274                                          handle_hotplug_event))
 275                 return -1;
 276
 277         /* Receive network link events form kernel */
 278         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 279
 280         return 0;
 281 }
 282
 283 static void system_set_sysctl(const char *path, const char *val)
 284 {
 285         int fd;
 286
 287         fd = open(path, O_WRONLY);
 288         if (fd < 0)
 289                 return;
 290
 291         if (write(fd, val, strlen(val))) {}
 292         close(fd);
 293 }
 294
 295 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 296 {
 297         snprintf(dev_buf, sizeof(dev_buf), path, device);
 298         system_set_sysctl(dev_buf, val);
 299 }
 300
 301 static void system_set_disable_ipv6(struct device *dev, const char *val)
 302 {
 303         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 304 }
 305
 306 static void system_set_rpfilter(struct device *dev, const char *val)
 307 {
 308         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter", dev->ifname, val);
 309 }
 310
 311 static void system_set_acceptlocal(struct device *dev, const char *val)
 312 {
 313         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local", dev->ifname, val);
 314 }
 315
 316 static void system_set_igmpversion(struct device *dev, const char *val)
 317 {
 318         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version", dev->ifname, val);
 319 }
 320
 321 static void system_set_mldversion(struct device *dev, const char *val)
 322 {
 323         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version", dev->ifname, val);
 324 }
 325
 326 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 327 {
 328         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 329 }
 330
 331 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 332 {
 333         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 334 }
 335
 336 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 337 {
 338         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time", dev->ifname, val);
 339 }
 340
 341 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 342 {
 343         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time", dev->ifname, val);
 344 }
 345
 346 static void system_set_neigh4locktime(struct device *dev, const char *val)
 347 {
 348         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime", dev->ifname, val);
 349 }
 350
 351 static void system_set_dadtransmits(struct device *dev, const char *val)
 352 {
 353         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 354 }
 355
 356 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 357 {
 358         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 359 }
 360
 361 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 362 {
 363         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 364 }
 365
 366 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 367 {
 368         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 369 }
 370
 371 static void system_bridge_set_isolated(struct device *dev, const char *val)
 372 {
 373         system_set_dev_sysctl("/sys/class/net/%s/brport/isolated", dev->ifname, val);
 374 }
 375
 376 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 377 {
 378         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 379                                        "/sys/class/net/%s/brport/multicast_router",
 380                               dev->ifname, val);
 381 }
 382
 383 static void system_bridge_set_robustness(struct device *dev, const char *val)
 384 {
 385         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 386                               dev->ifname, val);
 387         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 388                               dev->ifname, val);
 389 }
 390
 391 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 392 {
 393         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 394                               dev->ifname, val);
 395 }
 396
 397 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 398 {
 399         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 400                               dev->ifname, val);
 401 }
 402
 403 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 404 {
 405         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 406                               dev->ifname, val);
 407 }
 408
 409 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 410 {
 411         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 412                               dev->ifname, val);
 413 }
 414
 415 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 416 {
 417         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 418                               dev->ifname, val);
 419 }
 420
 421 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 422 {
 423         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 424                               dev->ifname, val);
 425 }
 426
 427 static void system_bridge_set_stp_state(struct device *dev, const char *val)
 428 {
 429         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/stp_state", dev->ifname, val);
 430 }
 431
 432 static void system_bridge_set_forward_delay(struct device *dev, const char *val)
 433 {
 434         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/forward_delay", dev->ifname, val);
 435 }
 436
 437 static void system_bridge_set_priority(struct device *dev, const char *val)
 438 {
 439         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/priority", dev->ifname, val);
 440 }
 441
 442 static void system_bridge_set_ageing_time(struct device *dev, const char *val)
 443 {
 444         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/ageing_time", dev->ifname, val);
 445 }
 446
 447 static void system_bridge_set_hello_time(struct device *dev, const char *val)
 448 {
 449         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hello_time", dev->ifname, val);
 450 }
 451
 452 static void system_bridge_set_max_age(struct device *dev, const char *val)
 453 {
 454         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/max_age", dev->ifname, val);
 455 }
 456
 457 static void system_bridge_set_learning(struct device *dev, const char *val)
 458 {
 459         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 460 }
 461
 462 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 463 {
 464         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 465 }
 466
 467 static void system_set_sendredirects(struct device *dev, const char *val)
 468 {
 469         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 470 }
 471
 472 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 473 {
 474         int fd = -1, ret = -1;
 475
 476         fd = open(path, O_RDONLY);
 477         if (fd < 0)
 478                 goto out;
 479
 480         ssize_t len = read(fd, buf, buf_sz - 1);
 481         if (len < 0)
 482                 goto out;
 483
 484         ret = buf[len] = 0;
 485
 486 out:
 487         if (fd >= 0)
 488                 close(fd);
 489
 490         return ret;
 491 }
 492
 493 static int
 494 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 495 {
 496         snprintf(dev_buf, sizeof(dev_buf), path, device);
 497         return system_get_sysctl(dev_buf, buf, buf_sz);
 498 }
 499
 500 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 501 {
 502         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 503                         dev->ifname, buf, buf_sz);
 504 }
 505
 506 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 507 {
 508         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 509                         dev->ifname, buf, buf_sz);
 510 }
 511
 512 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 513 {
 514         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 515                         dev->ifname, buf, buf_sz);
 516 }
 517
 518 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 519 {
 520         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 521                         dev->ifname, buf, buf_sz);
 522 }
 523
 524 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 525 {
 526         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 527                         dev->ifname, buf, buf_sz);
 528 }
 529
 530 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 531 {
 532         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 533                         dev->ifname, buf, buf_sz);
 534 }
 535
 536 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 537 {
 538         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 539                         dev->ifname, buf, buf_sz);
 540 }
 541
 542 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 543 {
 544         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 545                         dev->ifname, buf, buf_sz);
 546 }
 547
 548 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 549 {
 550         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 551                         dev->ifname, buf, buf_sz);
 552 }
 553
 554 static int system_get_neigh4locktime(struct device *dev, char *buf, const size_t buf_sz)
 555 {
 556         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime",
 557                         dev->ifname, buf, buf_sz);
 558 }
 559
 560 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 561 {
 562         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 563                         dev->ifname, buf, buf_sz);
 564 }
 565
 566 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 567 {
 568         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 569                         dev->ifname, buf, buf_sz);
 570 }
 571
 572 /* Evaluate netlink messages */
 573 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 574 {
 575         struct nlmsghdr *nh = nlmsg_hdr(msg);
 576         struct nlattr *nla[__IFLA_MAX];
 577         int link_state = 0;
 578         char buf[10];
 579
 580         if (nh->nlmsg_type != RTM_NEWLINK)
 581                 goto out;
 582
 583         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 584         if (!nla[IFLA_IFNAME])
 585                 goto out;
 586
 587         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 588         if (!dev)
 589                 goto out;
 590
 591         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 592                 link_state = strtoul(buf, NULL, 0);
 593
 594         device_set_link(dev, link_state ? true : false);
 595
 596 out:
 597         return 0;
 598 }
 599
 600 static void
 601 handle_hotplug_msg(char *data, int size)
 602 {
 603         const char *subsystem = NULL, *interface = NULL, *interface_old = NULL;
 604         char *cur, *end, *sep;
 605         struct device *dev;
 606         int skip;
 607         bool add, move = false;
 608
 609         if (!strncmp(data, "add@", 4))
 610                 add = true;
 611         else if (!strncmp(data, "remove@", 7))
 612                 add = false;
 613         else if (!strncmp(data, "move@", 5)) {
 614                 add = true;
 615                 move = true;
 616         }
 617         else
 618                 return;
 619
 620         skip = strlen(data) + 1;
 621         end = data + size;
 622
 623         for (cur = data + skip; cur < end; cur += skip) {
 624                 skip = strlen(cur) + 1;
 625
 626                 sep = strchr(cur, '=');
 627                 if (!sep)
 628                         continue;
 629
 630                 *sep = 0;
 631                 if (!strcmp(cur, "INTERFACE"))
 632                         interface = sep + 1;
 633                 else if (!strcmp(cur, "SUBSYSTEM")) {
 634                         subsystem = sep + 1;
 635                         if (strcmp(subsystem, "net") != 0)
 636                                 return;
 637                 } else if (!strcmp(cur, "DEVPATH_OLD")) {
 638                         interface_old = strrchr(sep + 1, '/');
 639                         if (interface_old)
 640                                 interface_old++;
 641                 }
 642         }
 643
 644         if (subsystem && interface) {
 645                 if (move && interface_old)
 646                         goto move;
 647                 else
 648                         goto found;
 649         }
 650
 651         return;
 652
 653 move:
 654         dev = device_find(interface_old);
 655         if (!dev)
 656                 goto found;
 657
 658         if (dev->type != &simple_device_type)
 659                 goto found;
 660
 661         device_set_present(dev, false);
 662
 663 found:
 664         dev = device_find(interface);
 665         if (!dev)
 666                 return;
 667
 668         if (dev->type != &simple_device_type)
 669                 return;
 670
 671         if (add && system_if_force_external(dev->ifname))
 672                 return;
 673
 674         device_set_present(dev, add);
 675 }
 676
 677 static void
 678 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 679 {
 680         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 681         struct sockaddr_nl nla;
 682         unsigned char *buf = NULL;
 683         int size;
 684         int err;
 685         socklen_t errlen = sizeof(err);
 686
 687         if (!u->error) {
 688                 while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 689                         if (nla.nl_pid == 0)
 690                                 handle_hotplug_msg((char *) buf, size);
 691
 692                         free(buf);
 693                 }
 694                 return;
 695         }
 696
 697         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 698                 goto abort;
 699
 700         switch(err) {
 701         case ENOBUFS:
 702                 /* Increase rx buffer size on netlink socket */
 703                 ev->bufsize *= 2;
 704                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 705                         goto abort;
 706                 break;
 707
 708         default:
 709                 goto abort;
 710         }
 711         u->error = false;
 712         return;
 713
 714 abort:
 715         uloop_fd_delete(&ev->uloop);
 716         return;
 717 }
 718
 719 static int system_rtnl_call(struct nl_msg *msg)
 720 {
 721         int ret;
 722
 723         ret = nl_send_auto_complete(sock_rtnl, msg);
 724         nlmsg_free(msg);
 725
 726         if (ret < 0)
 727                 return ret;
 728
 729         return nl_wait_for_ack(sock_rtnl);
 730 }
 731
 732 int system_bridge_delbr(struct device *bridge)
 733 {
 734         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 735 }
 736
 737 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 738 {
 739         struct ifreq ifr;
 740
 741         memset(&ifr, 0, sizeof(ifr));
 742         if (dev)
 743                 ifr.ifr_ifindex = dev->ifindex;
 744         else
 745                 ifr.ifr_data = data;
 746         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name) - 1);
 747         return ioctl(sock_ioctl, cmd, &ifr);
 748 }
 749
 750 static bool system_is_bridge(const char *name, char *buf, int buflen)
 751 {
 752         struct stat st;
 753
 754         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 755         if (stat(buf, &st) < 0)
 756                 return false;
 757
 758         return true;
 759 }
 760
 761 static char *system_get_bridge(const char *name, char *buf, int buflen)
 762 {
 763         char *path;
 764         ssize_t len = -1;
 765         glob_t gl;
 766
 767         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 768         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 769                 return NULL;
 770
 771         if (gl.gl_pathc > 0)
 772                 len = readlink(gl.gl_pathv[0], buf, buflen);
 773
 774         globfree(&gl);
 775
 776         if (len < 0)
 777                 return NULL;
 778
 779         buf[len] = 0;
 780         path = strrchr(buf, '/');
 781         if (!path)
 782                 return NULL;
 783
 784         return path + 1;
 785 }
 786
 787 static void
 788 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 789 {
 790         bool mcast_to_ucast = dev->wireless_ap;
 791         bool hairpin = true;
 792
 793         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 794             !bridge->settings.multicast_to_unicast)
 795                 mcast_to_ucast = false;
 796
 797         if (!mcast_to_ucast || dev->wireless_isolate)
 798                 hairpin = false;
 799
 800         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 801         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 802 }
 803
 804 int system_bridge_addif(struct device *bridge, struct device *dev)
 805 {
 806         char buf[64];
 807         char *oldbr;
 808         int ret = 0;
 809
 810         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 811         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 812                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 813
 814         if (dev->wireless)
 815                 system_bridge_set_wireless(bridge, dev);
 816
 817         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 818                 snprintf(buf, sizeof(buf), "%u", dev->settings.multicast_router);
 819                 system_bridge_set_multicast_router(dev, buf, false);
 820         }
 821
 822         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 823             dev->settings.multicast_fast_leave)
 824                 system_bridge_set_multicast_fast_leave(dev, "1");
 825
 826         if (dev->settings.flags & DEV_OPT_LEARNING &&
 827             !dev->settings.learning)
 828                 system_bridge_set_learning(dev, "0");
 829
 830         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 831             !dev->settings.unicast_flood)
 832                 system_bridge_set_unicast_flood(dev, "0");
 833
 834         if (dev->settings.flags & DEV_OPT_ISOLATE &&
 835             dev->settings.isolate)
 836                 system_bridge_set_isolated(dev, "1");
 837
 838         return ret;
 839 }
 840
 841 int system_bridge_delif(struct device *bridge, struct device *dev)
 842 {
 843         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 844 }
 845
 846 int system_if_resolve(struct device *dev)
 847 {
 848         struct ifreq ifr;
 849         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
 850         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 851                 return ifr.ifr_ifindex;
 852         else
 853                 return 0;
 854 }
 855
 856 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 857 {
 858         struct ifreq ifr;
 859
 860         memset(&ifr, 0, sizeof(ifr));
 861         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
 862         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) < 0)
 863                 return -1;
 864
 865         ifr.ifr_flags |= add;
 866         ifr.ifr_flags &= ~rem;
 867         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 868 }
 869
 870 struct clear_data {
 871         struct nl_msg *msg;
 872         struct device *dev;
 873         int type;
 874         int size;
 875         int af;
 876 };
 877
 878
 879 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 880 {
 881         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 882
 883         return ifa->ifa_index == ifindex;
 884 }
 885
 886 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 887 {
 888         struct rtmsg *r = NLMSG_DATA(hdr);
 889         struct nlattr *tb[__RTA_MAX];
 890
 891         if (r->rtm_protocol == RTPROT_KERNEL &&
 892             r->rtm_family == AF_INET6)
 893                 return false;
 894
 895         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 896         if (!tb[RTA_OIF])
 897                 return false;
 898
 899         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 900 }
 901
 902 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 903 {
 904         return true;
 905 }
 906
 907 static int cb_clear_event(struct nl_msg *msg, void *arg)
 908 {
 909         struct clear_data *clr = arg;
 910         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 911         bool (*cb)(struct nlmsghdr *, int ifindex);
 912         int type, ret;
 913
 914         switch(clr->type) {
 915         case RTM_GETADDR:
 916                 type = RTM_DELADDR;
 917                 if (hdr->nlmsg_type != RTM_NEWADDR)
 918                         return NL_SKIP;
 919
 920                 cb = check_ifaddr;
 921                 break;
 922         case RTM_GETROUTE:
 923                 type = RTM_DELROUTE;
 924                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 925                         return NL_SKIP;
 926
 927                 cb = check_route;
 928                 break;
 929         case RTM_GETRULE:
 930                 type = RTM_DELRULE;
 931                 if (hdr->nlmsg_type != RTM_NEWRULE)
 932                         return NL_SKIP;
 933
 934                 cb = check_rule;
 935                 break;
 936         default:
 937                 return NL_SKIP;
 938         }
 939
 940         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 941                 return NL_SKIP;
 942
 943         if (type == RTM_DELRULE)
 944                 D(SYSTEM, "Remove a rule\n");
 945         else
 946                 D(SYSTEM, "Remove %s from device %s\n",
 947                   type == RTM_DELADDR ? "an address" : "a route",
 948                   clr->dev->ifname);
 949
 950         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 951         hdr = nlmsg_hdr(clr->msg);
 952         hdr->nlmsg_type = type;
 953         hdr->nlmsg_flags = NLM_F_REQUEST;
 954
 955         nl_socket_disable_auto_ack(sock_rtnl);
 956         ret = nl_send_auto_complete(sock_rtnl, clr->msg);
 957         if (ret < 0) {
 958                 if (type == RTM_DELRULE)
 959                         D(SYSTEM, "Error deleting a rule: %d\n", ret);
 960                 else
 961                         D(SYSTEM, "Error deleting %s from device '%s': %d\n",
 962                                 type == RTM_DELADDR ? "an address" : "a route",
 963                                 clr->dev->ifname, ret);
 964         }
 965
 966         nl_socket_enable_auto_ack(sock_rtnl);
 967
 968         return NL_SKIP;
 969 }
 970
 971 static int
 972 cb_finish_event(struct nl_msg *msg, void *arg)
 973 {
 974         int *pending = arg;
 975         *pending = 0;
 976         return NL_STOP;
 977 }
 978
 979 static int
 980 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 981 {
 982         int *pending = arg;
 983         *pending = err->error;
 984         return NL_STOP;
 985 }
 986
 987 static void
 988 system_if_clear_entries(struct device *dev, int type, int af)
 989 {
 990         struct clear_data clr;
 991         struct nl_cb *cb;
 992         struct rtmsg rtm = {
 993                 .rtm_family = af,
 994                 .rtm_flags = RTM_F_CLONED,
 995         };
 996         int flags = NLM_F_DUMP;
 997         int pending = 1;
 998
 999         clr.af = af;
1000         clr.dev = dev;
1001         clr.type = type;
1002         switch (type) {
1003         case RTM_GETADDR:
1004         case RTM_GETRULE:
1005                 clr.size = sizeof(struct rtgenmsg);
1006                 break;
1007         case RTM_GETROUTE:
1008                 clr.size = sizeof(struct rtmsg);
1009                 break;
1010         default:
1011                 return;
1012         }
1013
1014         cb = nl_cb_alloc(NL_CB_DEFAULT);
1015         if (!cb)
1016                 return;
1017
1018         clr.msg = nlmsg_alloc_simple(type, flags);
1019         if (!clr.msg)
1020                 goto out;
1021
1022         nlmsg_append(clr.msg, &rtm, clr.size, 0);
1023         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
1024         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
1025         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
1026
1027         if (nl_send_auto_complete(sock_rtnl, clr.msg) < 0)
1028                 goto free;
1029
1030         while (pending > 0)
1031                 nl_recvmsgs(sock_rtnl, cb);
1032
1033 free:
1034         nlmsg_free(clr.msg);
1035 out:
1036         nl_cb_put(cb);
1037 }
1038
1039 /*
1040  * Clear bridge (membership) state and bring down device
1041  */
1042 void system_if_clear_state(struct device *dev)
1043 {
1044         static char buf[256];
1045         char *bridge;
1046         device_set_ifindex(dev, system_if_resolve(dev));
1047
1048         if (dev->external || !dev->ifindex)
1049                 return;
1050
1051         system_if_flags(dev->ifname, 0, IFF_UP);
1052
1053         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
1054                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
1055                 system_bridge_delbr(dev);
1056                 return;
1057         }
1058
1059         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
1060         if (bridge) {
1061                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
1062                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
1063         }
1064
1065         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
1066         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
1067         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
1068         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
1069         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET);
1070         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET6);
1071         system_set_disable_ipv6(dev, "0");
1072 }
1073
1074 static inline unsigned long
1075 sec_to_jiffies(int val)
1076 {
1077         return (unsigned long) val * 100;
1078 }
1079
1080 static void system_bridge_conf_multicast_deps(struct device *bridge,
1081                                               struct bridge_config *cfg,
1082                                               char *buf,
1083                                               int buf_len)
1084 {
1085         int val;
1086
1087         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
1088             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
1089             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1090                 val = cfg->robustness * cfg->query_interval +
1091                         cfg->query_response_interval;
1092
1093                 snprintf(buf, buf_len, "%i", val);
1094                 system_bridge_set_membership_interval(bridge, buf);
1095
1096                 val = cfg->robustness * cfg->query_interval +
1097                         cfg->query_response_interval / 2;
1098
1099                 snprintf(buf, buf_len, "%i", val);
1100                 system_bridge_set_other_querier_timeout(bridge, buf);
1101         }
1102
1103         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1104                 val = cfg->query_interval / 4;
1105
1106                 snprintf(buf, buf_len, "%i", val);
1107                 system_bridge_set_startup_query_interval(bridge, buf);
1108         }
1109 }
1110
1111 static void system_bridge_conf_multicast(struct device *bridge,
1112                                          struct bridge_config *cfg,
1113                                          char *buf,
1114                                          int buf_len)
1115 {
1116         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
1117                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
1118
1119         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
1120                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
1121
1122         snprintf(buf, buf_len, "%i", cfg->hash_max);
1123         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
1124                 bridge->ifname, buf);
1125
1126         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
1127                 snprintf(buf, buf_len, "%u", bridge->settings.multicast_router);
1128                 system_bridge_set_multicast_router(bridge, buf, true);
1129         }
1130
1131         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
1132                 snprintf(buf, buf_len, "%i", cfg->robustness);
1133                 system_bridge_set_robustness(bridge, buf);
1134         }
1135
1136         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1137                 snprintf(buf, buf_len, "%i", cfg->query_interval);
1138                 system_bridge_set_query_interval(bridge, buf);
1139         }
1140
1141         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1142                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
1143                 system_bridge_set_query_response_interval(bridge, buf);
1144         }
1145
1146         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1147                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1148                 system_bridge_set_last_member_interval(bridge, buf);
1149         }
1150
1151         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1152 }
1153
1154 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1155 {
1156         char buf[64];
1157
1158         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1159                 return -1;
1160
1161         system_bridge_set_stp_state(bridge, cfg->stp ? "1" : "0");
1162
1163         snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->forward_delay));
1164         system_bridge_set_forward_delay(bridge, buf);
1165
1166         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1167
1168         snprintf(buf, sizeof(buf), "%d", cfg->priority);
1169         system_bridge_set_priority(bridge, buf);
1170
1171         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1172                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->ageing_time));
1173                 system_bridge_set_ageing_time(bridge, buf);
1174         }
1175
1176         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1177                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->hello_time));
1178                 system_bridge_set_hello_time(bridge, buf);
1179         }
1180
1181         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1182                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->max_age));
1183                 system_bridge_set_max_age(bridge, buf);
1184         }
1185
1186         return 0;
1187 }
1188
1189 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1190 {
1191         struct nl_msg *msg;
1192         struct nlattr *linkinfo, *data;
1193         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1194         int i, rv;
1195         static const struct {
1196                 const char *name;
1197                 enum macvlan_mode val;
1198         } modes[] = {
1199                 { "private", MACVLAN_MODE_PRIVATE },
1200                 { "vepa", MACVLAN_MODE_VEPA },
1201                 { "bridge", MACVLAN_MODE_BRIDGE },
1202                 { "passthru", MACVLAN_MODE_PASSTHRU },
1203         };
1204
1205         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1206
1207         if (!msg)
1208                 return -1;
1209
1210         nlmsg_append(msg, &iim, sizeof(iim), 0);
1211
1212         if (cfg->flags & MACVLAN_OPT_MACADDR)
1213                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1214         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1215         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1216
1217         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1218                 goto nla_put_failure;
1219
1220         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1221
1222         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1223                 goto nla_put_failure;
1224
1225         if (cfg->mode) {
1226                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1227                         if (strcmp(cfg->mode, modes[i].name) != 0)
1228                                 continue;
1229
1230                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1231                         break;
1232                 }
1233         }
1234
1235         nla_nest_end(msg, data);
1236         nla_nest_end(msg, linkinfo);
1237
1238         rv = system_rtnl_call(msg);
1239         if (rv)
1240                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1241
1242         return rv;
1243
1244 nla_put_failure:
1245         nlmsg_free(msg);
1246         return -ENOMEM;
1247 }
1248
1249 int system_link_netns_move(const char *ifname, int netns_fd)
1250 {
1251         struct nl_msg *msg;
1252         struct ifinfomsg iim = {
1253                 .ifi_family = AF_UNSPEC,
1254                 .ifi_index = 0,
1255         };
1256
1257         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST);
1258
1259         if (!msg)
1260                 return -1;
1261
1262         nlmsg_append(msg, &iim, sizeof(iim), 0);
1263         nla_put_string(msg, IFLA_IFNAME, ifname);
1264         nla_put_u32(msg, IFLA_NET_NS_FD, netns_fd);
1265         return system_rtnl_call(msg);
1266 }
1267
1268 static int system_link_del(const char *ifname)
1269 {
1270         struct nl_msg *msg;
1271         struct ifinfomsg iim = {
1272                 .ifi_family = AF_UNSPEC,
1273                 .ifi_index = 0,
1274         };
1275
1276         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1277
1278         if (!msg)
1279                 return -1;
1280
1281         nlmsg_append(msg, &iim, sizeof(iim), 0);
1282         nla_put_string(msg, IFLA_IFNAME, ifname);
1283         return system_rtnl_call(msg);
1284 }
1285
1286 int system_macvlan_del(struct device *macvlan)
1287 {
1288         return system_link_del(macvlan->ifname);
1289 }
1290
1291 int system_netns_open(const pid_t target_ns)
1292 {
1293         char pid_net_path[PATH_MAX];
1294
1295         snprintf(pid_net_path, sizeof(pid_net_path), "/proc/%u/ns/net", target_ns);
1296
1297         return open(pid_net_path, O_RDONLY);
1298 }
1299
1300 int system_netns_set(int netns_fd)
1301 {
1302         return setns(netns_fd, CLONE_NEWNET);
1303 }
1304
1305 int system_veth_add(struct device *veth, struct veth_config *cfg)
1306 {
1307         struct nl_msg *msg;
1308         struct ifinfomsg empty_iim = {};
1309         struct nlattr *linkinfo, *data, *veth_info;
1310         int rv;
1311
1312         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1313
1314         if (!msg)
1315                 return -1;
1316
1317         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1318
1319         if (cfg->flags & VETH_OPT_MACADDR)
1320                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1321         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1322
1323         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1324                 goto nla_put_failure;
1325
1326         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1327
1328         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1329                 goto nla_put_failure;
1330
1331         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1332                 goto nla_put_failure;
1333
1334         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1335
1336         if (cfg->flags & VETH_OPT_PEER_NAME)
1337                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1338         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1339                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1340
1341         nla_nest_end(msg, veth_info);
1342         nla_nest_end(msg, data);
1343         nla_nest_end(msg, linkinfo);
1344
1345         rv = system_rtnl_call(msg);
1346         if (rv) {
1347                 if (cfg->flags & VETH_OPT_PEER_NAME)
1348                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1349                 else
1350                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1351         }
1352
1353         return rv;
1354
1355 nla_put_failure:
1356         nlmsg_free(msg);
1357         return -ENOMEM;
1358 }
1359
1360 int system_veth_del(struct device *veth)
1361 {
1362         return system_link_del(veth->ifname);
1363 }
1364
1365 static int system_vlan(struct device *dev, int id)
1366 {
1367         struct vlan_ioctl_args ifr = {
1368                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1369                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1370         };
1371
1372         if (ioctl(sock_ioctl, SIOCSIFVLAN, &ifr) < 0)
1373                 return -1;
1374
1375         if (id < 0) {
1376                 ifr.cmd = DEL_VLAN_CMD;
1377                 ifr.u.VID = 0;
1378         } else {
1379                 ifr.cmd = ADD_VLAN_CMD;
1380                 ifr.u.VID = id;
1381         }
1382         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1383         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1384 }
1385
1386 int system_vlan_add(struct device *dev, int id)
1387 {
1388         return system_vlan(dev, id);
1389 }
1390
1391 int system_vlan_del(struct device *dev)
1392 {
1393         return system_vlan(dev, -1);
1394 }
1395
1396 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1397 {
1398         struct nl_msg *msg;
1399         struct nlattr *linkinfo, *data;
1400         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1401         int rv;
1402
1403         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1404
1405         if (!msg)
1406                 return -1;
1407
1408         nlmsg_append(msg, &iim, sizeof(iim), 0);
1409         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1410         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1411
1412         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1413                 goto nla_put_failure;
1414
1415         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1416
1417         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1418                 goto nla_put_failure;
1419
1420         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1421
1422 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1423         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1424 #else
1425         if(cfg->proto == VLAN_PROTO_8021AD)
1426                 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);
1427 #endif
1428
1429         nla_nest_end(msg, data);
1430         nla_nest_end(msg, linkinfo);
1431
1432         rv = system_rtnl_call(msg);
1433         if (rv)
1434                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1435
1436         return rv;
1437
1438 nla_put_failure:
1439         nlmsg_free(msg);
1440         return -ENOMEM;
1441 }
1442
1443 int system_vlandev_del(struct device *vlandev)
1444 {
1445         return system_link_del(vlandev->ifname);
1446 }
1447
1448 void
1449 system_if_get_settings(struct device *dev, struct device_settings *s)
1450 {
1451         struct ifreq ifr;
1452         char buf[10];
1453
1454         memset(&ifr, 0, sizeof(ifr));
1455         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1456
1457         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1458                 s->mtu = ifr.ifr_mtu;
1459                 s->flags |= DEV_OPT_MTU;
1460         }
1461
1462         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1463         if (s->mtu6 > 0)
1464                 s->flags |= DEV_OPT_MTU6;
1465
1466         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1467                 s->txqueuelen = ifr.ifr_qlen;
1468                 s->flags |= DEV_OPT_TXQUEUELEN;
1469         }
1470
1471         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1472                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1473                 s->flags |= DEV_OPT_MACADDR;
1474         }
1475
1476         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1477                 s->ipv6 = !strtoul(buf, NULL, 0);
1478                 s->flags |= DEV_OPT_IPV6;
1479         }
1480
1481         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1482                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1483                 s->flags |= DEV_OPT_PROMISC;
1484
1485                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1486                 s->flags |= DEV_OPT_MULTICAST;
1487         }
1488
1489         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1490                 s->rpfilter = strtoul(buf, NULL, 0);
1491                 s->flags |= DEV_OPT_RPFILTER;
1492         }
1493
1494         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1495                 s->acceptlocal = strtoul(buf, NULL, 0);
1496                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1497         }
1498
1499         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1500                 s->igmpversion = strtoul(buf, NULL, 0);
1501                 s->flags |= DEV_OPT_IGMPVERSION;
1502         }
1503
1504         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1505                 s->mldversion = strtoul(buf, NULL, 0);
1506                 s->flags |= DEV_OPT_MLDVERSION;
1507         }
1508
1509         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1510                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1511                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1512         }
1513
1514         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1515                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1516                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1517         }
1518
1519         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
1520                 s->neigh4locktime = strtol(buf, NULL, 0);
1521                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
1522         }
1523
1524         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1525                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1526                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1527         }
1528
1529         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1530                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1531                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1532         }
1533
1534         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1535                 s->dadtransmits = strtoul(buf, NULL, 0);
1536                 s->flags |= DEV_OPT_DADTRANSMITS;
1537         }
1538
1539         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1540                 s->sendredirects = strtoul(buf, NULL, 0);
1541                 s->flags |= DEV_OPT_SENDREDIRECTS;
1542         }
1543 }
1544
1545 void
1546 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1547 {
1548         struct ifreq ifr;
1549         char buf[12];
1550
1551         memset(&ifr, 0, sizeof(ifr));
1552         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1553         if (s->flags & DEV_OPT_MTU & apply_mask) {
1554                 ifr.ifr_mtu = s->mtu;
1555                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1556                         s->flags &= ~DEV_OPT_MTU;
1557         }
1558         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1559                 system_update_ipv6_mtu(dev, s->mtu6);
1560         }
1561         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1562                 ifr.ifr_qlen = s->txqueuelen;
1563                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1564                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1565         }
1566         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1567                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1568                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1569                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1570                         s->flags &= ~DEV_OPT_MACADDR;
1571         }
1572         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1573                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1574         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1575                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1576                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1577                         s->flags &= ~DEV_OPT_PROMISC;
1578         }
1579         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1580                 snprintf(buf, sizeof(buf), "%u", s->rpfilter);
1581                 system_set_rpfilter(dev, buf);
1582         }
1583         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1584                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1585         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1586                 snprintf(buf, sizeof(buf), "%u", s->igmpversion);
1587                 system_set_igmpversion(dev, buf);
1588         }
1589         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1590                 snprintf(buf, sizeof(buf), "%u", s->mldversion);
1591                 system_set_mldversion(dev, buf);
1592         }
1593         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1594                 snprintf(buf, sizeof(buf), "%u", s->neigh4reachabletime);
1595                 system_set_neigh4reachabletime(dev, buf);
1596                 snprintf(buf, sizeof(buf), "%u", s->neigh6reachabletime);
1597                 system_set_neigh6reachabletime(dev, buf);
1598         }
1599         if (s->flags & DEV_OPT_NEIGHLOCKTIME & apply_mask) {
1600                 snprintf(buf, sizeof(buf), "%d", s->neigh4locktime);
1601                 system_set_neigh4locktime(dev, buf);
1602         }
1603         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1604                 snprintf(buf, sizeof(buf), "%u", s->neigh4gcstaletime);
1605                 system_set_neigh4gcstaletime(dev, buf);
1606                 snprintf(buf, sizeof(buf), "%u", s->neigh6gcstaletime);
1607                 system_set_neigh6gcstaletime(dev, buf);
1608         }
1609         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1610                 snprintf(buf, sizeof(buf), "%u", s->dadtransmits);
1611                 system_set_dadtransmits(dev, buf);
1612         }
1613         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1614                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1615                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1616                         s->flags &= ~DEV_OPT_MULTICAST;
1617         }
1618         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1619                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1620 }
1621
1622 int system_if_up(struct device *dev)
1623 {
1624         system_if_get_settings(dev, &dev->orig_settings);
1625         /* Only keep orig settings based on what needs to be set */
1626         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1627         dev->orig_settings.flags &= dev->settings.flags;
1628         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1629         return system_if_flags(dev->ifname, IFF_UP, 0);
1630 }
1631
1632 int system_if_down(struct device *dev)
1633 {
1634         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1635         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1636         return ret;
1637 }
1638
1639 struct if_check_data {
1640         struct device *dev;
1641         int pending;
1642         int ret;
1643 };
1644
1645 #ifndef IFF_LOWER_UP
1646 #define IFF_LOWER_UP    0x10000
1647 #endif
1648
1649 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1650 {
1651         struct nlmsghdr *nh = nlmsg_hdr(msg);
1652         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1653         struct if_check_data *chk = (struct if_check_data *)arg;
1654
1655         if (nh->nlmsg_type != RTM_NEWLINK)
1656                 return NL_SKIP;
1657
1658         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1659         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1660
1661         return NL_OK;
1662 }
1663
1664 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1665 {
1666         struct if_check_data *chk = (struct if_check_data *)arg;
1667         chk->pending = 0;
1668         return NL_STOP;
1669 }
1670
1671 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1672 {
1673         struct if_check_data *chk = (struct if_check_data *)arg;
1674
1675         device_set_present(chk->dev, false);
1676         device_set_link(chk->dev, false);
1677         chk->pending = err->error;
1678
1679         return NL_STOP;
1680 }
1681
1682 int system_if_check(struct device *dev)
1683 {
1684         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1685         struct nl_msg *msg;
1686         struct ifinfomsg ifi = {
1687                 .ifi_family = AF_UNSPEC,
1688                 .ifi_index = 0,
1689         };
1690         struct if_check_data chk = {
1691                 .dev = dev,
1692                 .pending = 1,
1693         };
1694         int ret = 1;
1695
1696         if (!cb)
1697                 return ret;
1698
1699         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1700         if (!msg)
1701                 goto out;
1702
1703         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1704             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1705                 goto free;
1706
1707         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1708         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1709         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1710
1711         ret = nl_send_auto_complete(sock_rtnl, msg);
1712         if (ret < 0)
1713                 goto free;
1714
1715         while (chk.pending > 0)
1716                 nl_recvmsgs(sock_rtnl, cb);
1717
1718         ret = chk.pending;
1719
1720 free:
1721         nlmsg_free(msg);
1722 out:
1723         nl_cb_put(cb);
1724         return ret;
1725 }
1726
1727 struct device *
1728 system_if_get_parent(struct device *dev)
1729 {
1730         char buf[64], *devname;
1731         int ifindex, iflink, len;
1732         FILE *f;
1733
1734         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1735         f = fopen(buf, "r");
1736         if (!f)
1737                 return NULL;
1738
1739         len = fread(buf, 1, sizeof(buf) - 1, f);
1740         fclose(f);
1741
1742         if (len <= 0)
1743                 return NULL;
1744
1745         buf[len] = 0;
1746         iflink = strtoul(buf, NULL, 0);
1747         ifindex = system_if_resolve(dev);
1748         if (!iflink || iflink == ifindex)
1749                 return NULL;
1750
1751         devname = if_indextoname(iflink, buf);
1752         if (!devname)
1753                 return NULL;
1754
1755         return device_get(devname, true);
1756 }
1757
1758 static bool
1759 read_string_file(int dir_fd, const char *file, char *buf, int len)
1760 {
1761         bool ret = false;
1762         char *c;
1763         int fd;
1764
1765         fd = openat(dir_fd, file, O_RDONLY);
1766         if (fd < 0)
1767                 return false;
1768
1769 retry:
1770         len = read(fd, buf, len - 1);
1771         if (len < 0) {
1772                 if (errno == EINTR)
1773                         goto retry;
1774         } else if (len > 0) {
1775                         buf[len] = 0;
1776
1777                         c = strchr(buf, '\n');
1778                         if (c)
1779                                 *c = 0;
1780
1781                         ret = true;
1782         }
1783
1784         close(fd);
1785
1786         return ret;
1787 }
1788
1789 static bool
1790 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1791 {
1792         char buf[64];
1793         bool ret = false;
1794
1795         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1796         if (ret)
1797                 *val = strtoull(buf, NULL, 0);
1798
1799         return ret;
1800 }
1801
1802 /* Assume advertised flags == supported flags */
1803 static const struct {
1804         uint32_t mask;
1805         const char *name;
1806 } ethtool_link_modes[] = {
1807         { ADVERTISED_10baseT_Half, "10baseT-H" },
1808         { ADVERTISED_10baseT_Full, "10baseT-F" },
1809         { ADVERTISED_100baseT_Half, "100baseT-H" },
1810         { ADVERTISED_100baseT_Full, "100baseT-F" },
1811         { ADVERTISED_1000baseT_Half, "1000baseT-H" },
1812         { ADVERTISED_1000baseT_Full, "1000baseT-F" },
1813         { ADVERTISED_1000baseKX_Full, "1000baseKX-F" },
1814         { ADVERTISED_2500baseX_Full, "2500baseX-F" },
1815         { ADVERTISED_10000baseT_Full, "10000baseT-F" },
1816         { ADVERTISED_10000baseKX4_Full, "10000baseKX4-F" },
1817         { ADVERTISED_10000baseKR_Full, "10000baseKR-F" },
1818         { ADVERTISED_20000baseMLD2_Full, "20000baseMLD2-F" },
1819         { ADVERTISED_20000baseKR2_Full, "20000baseKR2-F" },
1820         { ADVERTISED_40000baseKR4_Full, "40000baseKR4-F" },
1821         { ADVERTISED_40000baseCR4_Full, "40000baseCR4-F" },
1822         { ADVERTISED_40000baseSR4_Full, "40000baseSR4-F" },
1823         { ADVERTISED_40000baseLR4_Full, "40000baseLR4-F" },
1824 #ifdef ADVERTISED_56000baseKR4_Full
1825         { ADVERTISED_56000baseKR4_Full, "56000baseKR4-F" },
1826         { ADVERTISED_56000baseCR4_Full, "56000baseCR4-F" },
1827         { ADVERTISED_56000baseSR4_Full, "56000baseSR4-F" },
1828         { ADVERTISED_56000baseLR4_Full, "56000baseLR4-F" },
1829 #endif
1830 };
1831
1832 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1833 {
1834         int i;
1835         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1836                 if (mask & ethtool_link_modes[i].mask)
1837                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1838         }
1839 }
1840
1841 bool
1842 system_if_force_external(const char *ifname)
1843 {
1844         char buf[64];
1845         struct stat s;
1846
1847         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1848         return stat(buf, &s) == 0;
1849 }
1850
1851 int
1852 system_if_dump_info(struct device *dev, struct blob_buf *b)
1853 {
1854         struct ethtool_cmd ecmd;
1855         struct ifreq ifr;
1856         char *s;
1857         void *c;
1858
1859         memset(&ecmd, 0, sizeof(ecmd));
1860         memset(&ifr, 0, sizeof(ifr));
1861         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1862         ifr.ifr_data = (caddr_t) &ecmd;
1863         ecmd.cmd = ETHTOOL_GSET;
1864
1865         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1866                 c = blobmsg_open_array(b, "link-advertising");
1867                 system_add_link_modes(b, ecmd.advertising);
1868                 blobmsg_close_array(b, c);
1869
1870                 c = blobmsg_open_array(b, "link-partner-advertising");
1871                 system_add_link_modes(b, ecmd.lp_advertising);
1872                 blobmsg_close_array(b, c);
1873
1874                 c = blobmsg_open_array(b, "link-supported");
1875                 system_add_link_modes(b, ecmd.supported);
1876                 blobmsg_close_array(b, c);
1877
1878                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1879                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1880                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1881                 blobmsg_add_string_buffer(b);
1882
1883                 blobmsg_add_u8(b, "autoneg", !!ecmd.autoneg);
1884         }
1885
1886         return 0;
1887 }
1888
1889 int
1890 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1891 {
1892         const char *const counters[] = {
1893                 "collisions",     "rx_frame_errors",   "tx_compressed",
1894                 "multicast",      "rx_length_errors",  "tx_dropped",
1895                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1896                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1897                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1898                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1899                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1900                 "rx_fifo_errors", "tx_carrier_errors",
1901         };
1902         char buf[64];
1903         int stats_dir;
1904         int i;
1905         uint64_t val = 0;
1906
1907         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1908         stats_dir = open(buf, O_DIRECTORY);
1909         if (stats_dir < 0)
1910                 return -1;
1911
1912         for (i = 0; i < ARRAY_SIZE(counters); i++)
1913                 if (read_uint64_file(stats_dir, counters[i], &val))
1914                         blobmsg_add_u64(b, counters[i], val);
1915
1916         close(stats_dir);
1917         return 0;
1918 }
1919
1920 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1921 {
1922         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1923         int alen = v4 ? 4 : 16;
1924         unsigned int flags = 0;
1925         struct ifaddrmsg ifa = {
1926                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1927                 .ifa_prefixlen = addr->mask,
1928                 .ifa_index = dev->ifindex,
1929         };
1930
1931         struct nl_msg *msg;
1932         if (cmd == RTM_NEWADDR)
1933                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1934
1935         msg = nlmsg_alloc_simple(cmd, flags);
1936         if (!msg)
1937                 return -1;
1938
1939         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1940         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1941         if (v4) {
1942                 if (addr->broadcast)
1943                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1944                 if (addr->point_to_point)
1945                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1946         } else {
1947                 time_t now = system_get_rtime();
1948                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1949
1950                 if (addr->preferred_until) {
1951                         int64_t preferred = addr->preferred_until - now;
1952                         if (preferred < 0)
1953                                 preferred = 0;
1954                         else if (preferred > UINT32_MAX)
1955                                 preferred = UINT32_MAX;
1956
1957                         cinfo.ifa_prefered = preferred;
1958                 }
1959
1960                 if (addr->valid_until) {
1961                         int64_t valid = addr->valid_until - now;
1962                         if (valid <= 0) {
1963                                 nlmsg_free(msg);
1964                                 return -1;
1965                         }
1966                         else if (valid > UINT32_MAX)
1967                                 valid = UINT32_MAX;
1968
1969                         cinfo.ifa_valid = valid;
1970                 }
1971
1972                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
1973
1974                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
1975                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
1976         }
1977
1978         return system_rtnl_call(msg);
1979 }
1980
1981 int system_add_address(struct device *dev, struct device_addr *addr)
1982 {
1983         return system_addr(dev, addr, RTM_NEWADDR);
1984 }
1985
1986 int system_del_address(struct device *dev, struct device_addr *addr)
1987 {
1988         return system_addr(dev, addr, RTM_DELADDR);
1989 }
1990
1991 static int system_neigh(struct device *dev, struct device_neighbor *neighbor, int cmd)
1992 {
1993         int alen = ((neighbor->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1994         unsigned int flags = 0;
1995         struct ndmsg ndm = {
1996                 .ndm_family = (alen == 4) ? AF_INET : AF_INET6,
1997                 .ndm_ifindex = dev->ifindex,
1998                 .ndm_state = NUD_PERMANENT,
1999                 .ndm_flags = (neighbor->proxy ? NTF_PROXY : 0) | (neighbor->router ? NTF_ROUTER : 0),
2000         };
2001         struct nl_msg *msg;
2002
2003         if (cmd == RTM_NEWNEIGH)
2004                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
2005
2006         msg = nlmsg_alloc_simple(cmd, flags);
2007
2008         if (!msg)
2009                 return -1;
2010
2011         nlmsg_append(msg, &ndm, sizeof(ndm), 0);
2012
2013         nla_put(msg, NDA_DST, alen, &neighbor->addr);
2014         if (neighbor->flags & DEVNEIGH_MAC)
2015                 nla_put(msg, NDA_LLADDR, sizeof(neighbor->macaddr), &neighbor->macaddr);
2016
2017
2018         return system_rtnl_call(msg);
2019 }
2020
2021 int system_add_neighbor(struct device *dev, struct device_neighbor *neighbor)
2022 {
2023         return system_neigh(dev, neighbor, RTM_NEWNEIGH);
2024 }
2025
2026 int system_del_neighbor(struct device *dev, struct device_neighbor *neighbor)
2027 {
2028         return system_neigh(dev, neighbor, RTM_DELNEIGH);
2029 }
2030
2031 static int system_rt(struct device *dev, struct device_route *route, int cmd)
2032 {
2033         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
2034         bool have_gw;
2035         unsigned int flags = 0;
2036
2037         if (alen == 4)
2038                 have_gw = !!route->nexthop.in.s_addr;
2039         else
2040                 have_gw = route->nexthop.in6.s6_addr32[0] ||
2041                         route->nexthop.in6.s6_addr32[1] ||
2042                         route->nexthop.in6.s6_addr32[2] ||
2043                         route->nexthop.in6.s6_addr32[3];
2044
2045         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
2046                         ? route->table : RT_TABLE_MAIN;
2047
2048         struct rtmsg rtm = {
2049                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2050                 .rtm_dst_len = route->mask,
2051                 .rtm_src_len = route->sourcemask,
2052                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
2053                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
2054                 .rtm_scope = RT_SCOPE_NOWHERE,
2055                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
2056                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
2057         };
2058         struct nl_msg *msg;
2059
2060         if (cmd == RTM_NEWROUTE) {
2061                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
2062
2063                 if (!dev) { /* Add null-route */
2064                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2065                         rtm.rtm_type = RTN_UNREACHABLE;
2066                 }
2067                 else
2068                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
2069         }
2070
2071         if (route->flags & DEVROUTE_TYPE) {
2072                 rtm.rtm_type = route->type;
2073                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
2074                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
2075                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
2076                                 rtm.rtm_table = RT_TABLE_LOCAL;
2077                 }
2078
2079                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
2080                         rtm.rtm_scope = RT_SCOPE_HOST;
2081                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
2082                                 rtm.rtm_type == RTN_ANYCAST) {
2083                         rtm.rtm_scope = RT_SCOPE_LINK;
2084                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
2085                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY ||
2086                                 rtm.rtm_type == RTN_THROW) {
2087                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2088                         dev = NULL;
2089                 }
2090         }
2091
2092         msg = nlmsg_alloc_simple(cmd, flags);
2093         if (!msg)
2094                 return -1;
2095
2096         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2097
2098         if (route->mask)
2099                 nla_put(msg, RTA_DST, alen, &route->addr);
2100
2101         if (route->sourcemask) {
2102                 if (rtm.rtm_family == AF_INET)
2103                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
2104                 else
2105                         nla_put(msg, RTA_SRC, alen, &route->source);
2106         }
2107
2108         if (route->metric > 0)
2109                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
2110
2111         if (have_gw)
2112                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
2113
2114         if (dev)
2115                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
2116
2117         if (table >= 256)
2118                 nla_put_u32(msg, RTA_TABLE, table);
2119
2120         if (route->flags & DEVROUTE_MTU) {
2121                 struct nlattr *metrics;
2122
2123                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
2124                         goto nla_put_failure;
2125
2126                 nla_put_u32(msg, RTAX_MTU, route->mtu);
2127
2128                 nla_nest_end(msg, metrics);
2129         }
2130
2131         return system_rtnl_call(msg);
2132
2133 nla_put_failure:
2134         nlmsg_free(msg);
2135         return -ENOMEM;
2136 }
2137
2138 int system_add_route(struct device *dev, struct device_route *route)
2139 {
2140         return system_rt(dev, route, RTM_NEWROUTE);
2141 }
2142
2143 int system_del_route(struct device *dev, struct device_route *route)
2144 {
2145         return system_rt(dev, route, RTM_DELROUTE);
2146 }
2147
2148 int system_flush_routes(void)
2149 {
2150         const char *names[] = {
2151                 "/proc/sys/net/ipv4/route/flush",
2152                 "/proc/sys/net/ipv6/route/flush"
2153         };
2154         int fd, i;
2155
2156         for (i = 0; i < ARRAY_SIZE(names); i++) {
2157                 fd = open(names[i], O_WRONLY);
2158                 if (fd < 0)
2159                         continue;
2160
2161                 if (write(fd, "-1", 2)) {}
2162                 close(fd);
2163         }
2164         return 0;
2165 }
2166
2167 bool system_resolve_rt_type(const char *type, unsigned int *id)
2168 {
2169         return system_rtn_aton(type, id);
2170 }
2171
2172 bool system_resolve_rt_proto(const char *type, unsigned int *id)
2173 {
2174         FILE *f;
2175         char *e, buf[128];
2176         unsigned int n, proto = 256;
2177         n = strtoul(type, &e, 0);
2178         if (!*e && e != type)
2179                 proto = n;
2180         else if (!strcmp(type, "unspec"))
2181                 proto = RTPROT_UNSPEC;
2182         else if (!strcmp(type, "kernel"))
2183                 proto = RTPROT_KERNEL;
2184         else if (!strcmp(type, "boot"))
2185                 proto = RTPROT_BOOT;
2186         else if (!strcmp(type, "static"))
2187                 proto = RTPROT_STATIC;
2188         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
2189                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
2190                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2191                                 continue;
2192
2193                         n = strtoul(e, NULL, 10);
2194                         e = strtok(NULL, " \t\n");
2195
2196                         if (e && !strcmp(e, type)) {
2197                                 proto = n;
2198                                 break;
2199                         }
2200                 }
2201                 fclose(f);
2202         }
2203
2204         if (proto > 255)
2205                 return false;
2206
2207         *id = proto;
2208         return true;
2209 }
2210
2211 bool system_resolve_rt_table(const char *name, unsigned int *id)
2212 {
2213         FILE *f;
2214         char *e, buf[128];
2215         unsigned int n, table = RT_TABLE_UNSPEC;
2216
2217         /* first try to parse table as number */
2218         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2219                 table = n;
2220
2221         /* handle well known aliases */
2222         else if (!strcmp(name, "default"))
2223                 table = RT_TABLE_DEFAULT;
2224         else if (!strcmp(name, "main"))
2225                 table = RT_TABLE_MAIN;
2226         else if (!strcmp(name, "local"))
2227                 table = RT_TABLE_LOCAL;
2228
2229         /* try to look up name in /etc/iproute2/rt_tables */
2230         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2231         {
2232                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2233                 {
2234                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2235                                 continue;
2236
2237                         n = strtoul(e, NULL, 10);
2238                         e = strtok(NULL, " \t\n");
2239
2240                         if (e && !strcmp(e, name))
2241                         {
2242                                 table = n;
2243                                 break;
2244                         }
2245                 }
2246
2247                 fclose(f);
2248         }
2249
2250         if (table == RT_TABLE_UNSPEC)
2251                 return false;
2252
2253         *id = table;
2254         return true;
2255 }
2256
2257 bool system_is_default_rt_table(unsigned int id)
2258 {
2259         return (id == RT_TABLE_MAIN);
2260 }
2261
2262 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2263 {
2264         char *e;
2265         unsigned int n;
2266
2267         if (!strcmp(filter, "strict"))
2268                 n = 1;
2269         else if (!strcmp(filter, "loose"))
2270                 n = 2;
2271         else {
2272                 n = strtoul(filter, &e, 0);
2273                 if (*e || e == filter || n > 2)
2274                         return false;
2275         }
2276
2277         *id = n;
2278         return true;
2279 }
2280
2281 static int system_iprule(struct iprule *rule, int cmd)
2282 {
2283         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2284
2285         struct nl_msg *msg;
2286         struct rtmsg rtm = {
2287                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2288                 .rtm_protocol = RTPROT_STATIC,
2289                 .rtm_scope = RT_SCOPE_UNIVERSE,
2290                 .rtm_table = RT_TABLE_UNSPEC,
2291                 .rtm_type = RTN_UNSPEC,
2292                 .rtm_flags = 0,
2293         };
2294
2295         if (cmd == RTM_NEWRULE)
2296                 rtm.rtm_type = RTN_UNICAST;
2297
2298         if (rule->invert)
2299                 rtm.rtm_flags |= FIB_RULE_INVERT;
2300
2301         if (rule->flags & IPRULE_SRC)
2302                 rtm.rtm_src_len = rule->src_mask;
2303
2304         if (rule->flags & IPRULE_DEST)
2305                 rtm.rtm_dst_len = rule->dest_mask;
2306
2307         if (rule->flags & IPRULE_TOS)
2308                 rtm.rtm_tos = rule->tos;
2309
2310         if (rule->flags & IPRULE_LOOKUP) {
2311                 if (rule->lookup < 256)
2312                         rtm.rtm_table = rule->lookup;
2313         }
2314
2315         if (rule->flags & IPRULE_ACTION)
2316                 rtm.rtm_type = rule->action;
2317         else if (rule->flags & IPRULE_GOTO)
2318                 rtm.rtm_type = FR_ACT_GOTO;
2319         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2320                 rtm.rtm_type = FR_ACT_NOP;
2321
2322         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2323
2324         if (!msg)
2325                 return -1;
2326
2327         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2328
2329         if (rule->flags & IPRULE_IN)
2330                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2331
2332         if (rule->flags & IPRULE_OUT)
2333                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2334
2335         if (rule->flags & IPRULE_SRC)
2336                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2337
2338         if (rule->flags & IPRULE_DEST)
2339                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2340
2341         if (rule->flags & IPRULE_PRIORITY)
2342                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2343         else if (cmd == RTM_NEWRULE)
2344                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2345
2346         if (rule->flags & IPRULE_FWMARK)
2347                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2348
2349         if (rule->flags & IPRULE_FWMASK)
2350                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2351
2352         if (rule->flags & IPRULE_LOOKUP) {
2353                 if (rule->lookup >= 256)
2354                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2355         }
2356
2357         if (rule->flags & IPRULE_SUP_PREFIXLEN)
2358                 nla_put_u32(msg, FRA_SUPPRESS_PREFIXLEN, rule->sup_prefixlen);
2359
2360         if (rule->flags & IPRULE_GOTO)
2361                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2362
2363         return system_rtnl_call(msg);
2364 }
2365
2366 int system_add_iprule(struct iprule *rule)
2367 {
2368         return system_iprule(rule, RTM_NEWRULE);
2369 }
2370
2371 int system_del_iprule(struct iprule *rule)
2372 {
2373         return system_iprule(rule, RTM_DELRULE);
2374 }
2375
2376 int system_flush_iprules(void)
2377 {
2378         int rv = 0;
2379         struct iprule rule;
2380
2381         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2382         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2383
2384         memset(&rule, 0, sizeof(rule));
2385
2386
2387         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2388
2389         rule.priority = 0;
2390         rule.lookup = RT_TABLE_LOCAL;
2391         rv |= system_iprule(&rule, RTM_NEWRULE);
2392
2393         rule.priority = 32766;
2394         rule.lookup = RT_TABLE_MAIN;
2395         rv |= system_iprule(&rule, RTM_NEWRULE);
2396
2397         rule.priority = 32767;
2398         rule.lookup = RT_TABLE_DEFAULT;
2399         rv |= system_iprule(&rule, RTM_NEWRULE);
2400
2401
2402         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2403
2404         rule.priority = 0;
2405         rule.lookup = RT_TABLE_LOCAL;
2406         rv |= system_iprule(&rule, RTM_NEWRULE);
2407
2408         rule.priority = 32766;
2409         rule.lookup = RT_TABLE_MAIN;
2410         rv |= system_iprule(&rule, RTM_NEWRULE);
2411
2412         return rv;
2413 }
2414
2415 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2416 {
2417         return system_rtn_aton(action, id);
2418 }
2419
2420 time_t system_get_rtime(void)
2421 {
2422         struct timespec ts;
2423         struct timeval tv;
2424
2425         if (clock_gettime(CLOCK_MONOTONIC, &ts) == 0)
2426                 return ts.tv_sec;
2427
2428         if (gettimeofday(&tv, NULL) == 0)
2429                 return tv.tv_sec;
2430
2431         return 0;
2432 }
2433
2434 #ifndef IP_DF
2435 #define IP_DF       0x4000
2436 #endif
2437
2438 static int tunnel_ioctl(const char *name, int cmd, void *p)
2439 {
2440         struct ifreq ifr;
2441
2442         memset(&ifr, 0, sizeof(ifr));
2443         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name) - 1);
2444         ifr.ifr_ifru.ifru_data = p;
2445         return ioctl(sock_ioctl, cmd, &ifr);
2446 }
2447
2448 #ifdef IFLA_IPTUN_MAX
2449 static int system_add_ip6_tunnel(const char *name, const unsigned int link,
2450                                  struct blob_attr **tb)
2451 {
2452         struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2453                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2454         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2455         struct blob_attr *cur;
2456         int ret = 0, ttl = 0;
2457
2458         if (!nlm)
2459                 return -1;
2460
2461         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2462         nla_put_string(nlm, IFLA_IFNAME, name);
2463
2464         if (link)
2465                 nla_put_u32(nlm, IFLA_LINK, link);
2466
2467         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2468         if (!linkinfo) {
2469                 ret = -ENOMEM;
2470                 goto failure;
2471         }
2472
2473         nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2474         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2475         if (!infodata) {
2476                 ret = -ENOMEM;
2477                 goto failure;
2478         }
2479
2480         if (link)
2481                 nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2482
2483         if ((cur = tb[TUNNEL_ATTR_TTL]))
2484                 ttl = blobmsg_get_u32(cur);
2485
2486         nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2487         nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2488
2489         struct in6_addr in6buf;
2490         if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2491                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2492                         ret = -EINVAL;
2493                         goto failure;
2494                 }
2495                 nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2496         }
2497
2498         if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2499                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2500                         ret = -EINVAL;
2501                         goto failure;
2502                 }
2503                 nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2504         }
2505
2506         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2507                 struct blob_attr *tb_data[__IPIP6_DATA_ATTR_MAX];
2508                 uint32_t tun_flags = IP6_TNL_F_IGN_ENCAP_LIMIT;
2509
2510                 blobmsg_parse(ipip6_data_attr_list.params, __IPIP6_DATA_ATTR_MAX, tb_data,
2511                         blobmsg_data(cur), blobmsg_len(cur));
2512
2513                 if ((cur = tb_data[IPIP6_DATA_ENCAPLIMIT])) {
2514                         char *str = blobmsg_get_string(cur);
2515
2516                         if (strcmp(str, "ignore")) {
2517                                 char *e;
2518                                 unsigned encap_limit = strtoul(str, &e, 0);
2519
2520                                 if (e == str || *e || encap_limit > 255) {
2521                                         ret = -EINVAL;
2522                                         goto failure;
2523                                 }
2524
2525                                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, encap_limit);
2526                                 tun_flags &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2527                         }
2528                 }
2529
2530 #ifdef IFLA_IPTUN_FMR_MAX
2531                 if ((cur = tb_data[IPIP6_DATA_FMRS])) {
2532                         struct blob_attr *rcur;
2533                         unsigned rrem, fmrcnt = 0;
2534                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2535
2536                         if (!fmrs) {
2537                                 ret = -ENOMEM;
2538                                 goto failure;
2539                         }
2540
2541                         blobmsg_for_each_attr(rcur, cur, rrem) {
2542                                 struct blob_attr *tb_fmr[__FMR_DATA_ATTR_MAX], *tb_cur;
2543                                 struct in6_addr ip6prefix;
2544                                 struct in_addr ip4prefix;
2545                                 unsigned ip4len, ip6len, ealen, offset;
2546
2547                                 blobmsg_parse(fmr_data_attr_list.params, __FMR_DATA_ATTR_MAX, tb_fmr,
2548                                                 blobmsg_data(rcur), blobmsg_len(rcur));
2549
2550                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX6]) ||
2551                                                 !parse_ip_and_netmask(AF_INET6,
2552                                                         blobmsg_data(tb_cur), &ip6prefix,
2553                                                         &ip6len)) {
2554                                         ret = -EINVAL;
2555                                         goto failure;
2556                                 }
2557
2558                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX4]) ||
2559                                                 !parse_ip_and_netmask(AF_INET,
2560                                                         blobmsg_data(tb_cur), &ip4prefix,
2561                                                         &ip4len)) {
2562                                         ret = -EINVAL;
2563                                         goto failure;
2564                                 }
2565
2566                                 if (!(tb_cur = tb_fmr[FMR_DATA_EALEN])) {
2567                                         ret = -EINVAL;
2568                                         goto failure;
2569                                 }
2570                                 ealen = blobmsg_get_u32(tb_cur);
2571
2572                                 if (!(tb_cur = tb_fmr[FMR_DATA_OFFSET])) {
2573                                         ret = -EINVAL;
2574                                         goto failure;
2575                                 }
2576                                 offset = blobmsg_get_u32(tb_cur);
2577
2578                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2579                                 if (!rule) {
2580                                         ret = -ENOMEM;
2581                                         goto failure;
2582                                 }
2583
2584                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2585                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2586                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2587                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2588                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2589                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2590
2591                                 nla_nest_end(nlm, rule);
2592                         }
2593
2594                         nla_nest_end(nlm, fmrs);
2595                 }
2596 #endif
2597                 if (tun_flags)
2598                         nla_put_u32(nlm, IFLA_IPTUN_FLAGS, tun_flags);
2599         }
2600
2601         nla_nest_end(nlm, infodata);
2602         nla_nest_end(nlm, linkinfo);
2603
2604         return system_rtnl_call(nlm);
2605
2606 failure:
2607         nlmsg_free(nlm);
2608         return ret;
2609 }
2610 #endif
2611
2612 #ifdef IFLA_IPTUN_MAX
2613 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2614 static int system_add_gre_tunnel(const char *name, const char *kind,
2615                                  const unsigned int link, struct blob_attr **tb, bool v6)
2616 {
2617         struct nl_msg *nlm;
2618         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2619         struct blob_attr *cur;
2620         uint32_t ikey = 0, okey = 0, flowinfo = 0, flags6 = IP6_TNL_F_IGN_ENCAP_LIMIT;
2621         uint16_t iflags = 0, oflags = 0;
2622         uint8_t tos = 0;
2623         int ret = 0, ttl = 0;
2624         unsigned encap_limit = 0;
2625
2626         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2627         if (!nlm)
2628                 return -1;
2629
2630         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2631         nla_put_string(nlm, IFLA_IFNAME, name);
2632
2633         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2634         if (!linkinfo) {
2635                 ret = -ENOMEM;
2636                 goto failure;
2637         }
2638
2639         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2640         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2641         if (!infodata) {
2642                 ret = -ENOMEM;
2643                 goto failure;
2644         }
2645
2646         if (link)
2647                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2648
2649         if ((cur = tb[TUNNEL_ATTR_TTL]))
2650                 ttl = blobmsg_get_u32(cur);
2651
2652         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2653                 char *str = blobmsg_get_string(cur);
2654                 if (strcmp(str, "inherit")) {
2655                         unsigned uval;
2656
2657                         if (!system_tos_aton(str, &uval)) {
2658                                 ret = -EINVAL;
2659                                 goto failure;
2660                         }
2661
2662                         if (v6)
2663                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2664                         else
2665                                 tos = uval;
2666                 } else {
2667                         if (v6)
2668                                 flags6 |= IP6_TNL_F_USE_ORIG_TCLASS;
2669                         else
2670                                 tos = 1;
2671                 }
2672         }
2673
2674         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2675                 struct blob_attr *tb_data[__GRE_DATA_ATTR_MAX];
2676
2677                 blobmsg_parse(gre_data_attr_list.params, __GRE_DATA_ATTR_MAX, tb_data,
2678                         blobmsg_data(cur), blobmsg_len(cur));
2679
2680                 if ((cur = tb_data[GRE_DATA_IKEY])) {
2681                         if ((ikey = blobmsg_get_u32(cur)))
2682                                 iflags |= GRE_KEY;
2683                 }
2684
2685                 if ((cur = tb_data[GRE_DATA_OKEY])) {
2686                         if ((okey = blobmsg_get_u32(cur)))
2687                                 oflags |= GRE_KEY;
2688                 }
2689
2690                 if ((cur = tb_data[GRE_DATA_ICSUM])) {
2691                         if (blobmsg_get_bool(cur))
2692                                 iflags |= GRE_CSUM;
2693                 }
2694
2695                 if ((cur = tb_data[GRE_DATA_OCSUM])) {
2696                         if (blobmsg_get_bool(cur))
2697                                 oflags |= GRE_CSUM;
2698                 }
2699
2700                 if ((cur = tb_data[GRE_DATA_ISEQNO])) {
2701                         if (blobmsg_get_bool(cur))
2702                                 iflags |= GRE_SEQ;
2703                 }
2704
2705                 if ((cur = tb_data[GRE_DATA_OSEQNO])) {
2706                         if (blobmsg_get_bool(cur))
2707                                 oflags |= GRE_SEQ;
2708                 }
2709
2710                 if ((cur = tb_data[GRE_DATA_ENCAPLIMIT])) {
2711                         char *str = blobmsg_get_string(cur);
2712
2713                         if (strcmp(str, "ignore")) {
2714                                 char *e;
2715
2716                                 encap_limit = strtoul(str, &e, 0);
2717
2718                                 if (e == str || *e || encap_limit > 255) {
2719                                         ret = -EINVAL;
2720                                         goto failure;
2721                                 }
2722
2723                                 flags6 &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2724                         }
2725                 }
2726         }
2727
2728         if (v6) {
2729                 struct in6_addr in6buf;
2730                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2731                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2732                                 ret = -EINVAL;
2733                                 goto failure;
2734                         }
2735                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2736                 }
2737
2738                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2739                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2740                                 ret = -EINVAL;
2741                                 goto failure;
2742                         }
2743                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2744                 }
2745
2746                 if (!(flags6 & IP6_TNL_F_IGN_ENCAP_LIMIT))
2747                         nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, encap_limit);
2748
2749                 if (flowinfo)
2750                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2751
2752                 if (flags6)
2753                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags6);
2754
2755                 if (!ttl)
2756                         ttl = 64;
2757         } else {
2758                 struct in_addr inbuf;
2759                 bool set_df = true;
2760
2761                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2762                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2763                                 ret = -EINVAL;
2764                                 goto failure;
2765                         }
2766                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2767                 }
2768
2769                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2770                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2771                                 ret = -EINVAL;
2772                                 goto failure;
2773                         }
2774                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2775
2776                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2777                                 if (!okey) {
2778                                         okey = inbuf.s_addr;
2779                                         oflags |= GRE_KEY;
2780                                 }
2781
2782                                 if (!ikey) {
2783                                         ikey = inbuf.s_addr;
2784                                         iflags |= GRE_KEY;
2785                                 }
2786                         }
2787                 }
2788
2789                 if ((cur = tb[TUNNEL_ATTR_DF]))
2790                         set_df = blobmsg_get_bool(cur);
2791
2792                 if (!set_df) {
2793                         /* ttl != 0 and nopmtudisc are incompatible */
2794                         if (ttl) {
2795                                 ret = -EINVAL;
2796                                 goto failure;
2797                         }
2798                 } else if (!ttl)
2799                         ttl = 64;
2800
2801                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2802
2803                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2804         }
2805
2806         if (ttl)
2807                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2808
2809         if (oflags)
2810                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2811
2812         if (iflags)
2813                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2814
2815         if (okey)
2816                 nla_put_u32(nlm, IFLA_GRE_OKEY, htonl(okey));
2817
2818         if (ikey)
2819                 nla_put_u32(nlm, IFLA_GRE_IKEY, htonl(ikey));
2820
2821         nla_nest_end(nlm, infodata);
2822         nla_nest_end(nlm, linkinfo);
2823
2824         return system_rtnl_call(nlm);
2825
2826 failure:
2827         nlmsg_free(nlm);
2828         return ret;
2829 }
2830 #endif
2831
2832 #ifdef IFLA_VTI_MAX
2833 static int system_add_vti_tunnel(const char *name, const char *kind,
2834                                  const unsigned int link, struct blob_attr **tb, bool v6)
2835 {
2836         struct nl_msg *nlm;
2837         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2838         struct blob_attr *cur;
2839         int ret = 0;
2840
2841         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2842         if (!nlm)
2843                 return -1;
2844
2845         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2846         nla_put_string(nlm, IFLA_IFNAME, name);
2847
2848         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2849         if (!linkinfo) {
2850                 ret = -ENOMEM;
2851                 goto failure;
2852         }
2853
2854         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2855         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2856         if (!infodata) {
2857                 ret = -ENOMEM;
2858                 goto failure;
2859         }
2860
2861         if (link)
2862                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2863
2864         if (v6) {
2865                 struct in6_addr in6buf;
2866                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2867                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2868                                 ret = -EINVAL;
2869                                 goto failure;
2870                         }
2871                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2872                 }
2873
2874                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2875                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2876                                 ret = -EINVAL;
2877                                 goto failure;
2878                         }
2879                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2880                 }
2881
2882         } else {
2883                 struct in_addr inbuf;
2884
2885                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2886                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2887                                 ret = -EINVAL;
2888                                 goto failure;
2889                         }
2890                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2891                 }
2892
2893                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2894                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2895                                 ret = -EINVAL;
2896                                 goto failure;
2897                         }
2898                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2899                 }
2900
2901         }
2902
2903         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2904                 struct blob_attr *tb_data[__VTI_DATA_ATTR_MAX];
2905                 uint32_t ikey = 0, okey = 0;
2906
2907                 blobmsg_parse(vti_data_attr_list.params, __VTI_DATA_ATTR_MAX, tb_data,
2908                         blobmsg_data(cur), blobmsg_len(cur));
2909
2910                 if ((cur = tb_data[VTI_DATA_IKEY])) {
2911                         if ((ikey = blobmsg_get_u32(cur)))
2912                                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2913                 }
2914
2915                 if ((cur = tb_data[VTI_DATA_OKEY])) {
2916                         if ((okey = blobmsg_get_u32(cur)))
2917                                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2918                 }
2919         }
2920
2921         nla_nest_end(nlm, infodata);
2922         nla_nest_end(nlm, linkinfo);
2923
2924         return system_rtnl_call(nlm);
2925
2926 failure:
2927         nlmsg_free(nlm);
2928         return ret;
2929 }
2930 #endif
2931
2932 #ifdef IFLA_XFRM_MAX
2933 static int system_add_xfrm_tunnel(const char *name, const char *kind,
2934                                  const unsigned int link, struct blob_attr **tb)
2935 {
2936         struct nl_msg *nlm;
2937         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2938         struct blob_attr *cur;
2939         int ret = 0;
2940
2941         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2942         if (!nlm)
2943                 return -1;
2944
2945         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2946         nla_put_string(nlm, IFLA_IFNAME, name);
2947
2948         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2949         if (!linkinfo) {
2950                 ret = -ENOMEM;
2951                 goto failure;
2952         }
2953
2954         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2955         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2956         if (!infodata) {
2957                 ret = -ENOMEM;
2958                 goto failure;
2959         }
2960
2961         if (link)
2962                 nla_put_u32(nlm, IFLA_XFRM_LINK, link);
2963
2964         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2965                 struct blob_attr *tb_data[__XFRM_DATA_ATTR_MAX];
2966                 uint32_t if_id = 0;
2967
2968                 blobmsg_parse(xfrm_data_attr_list.params, __XFRM_DATA_ATTR_MAX, tb_data,
2969                         blobmsg_data(cur), blobmsg_len(cur));
2970
2971                 if ((cur = tb_data[XFRM_DATA_IF_ID])) {
2972                         if ((if_id = blobmsg_get_u32(cur)))
2973                                 nla_put_u32(nlm, IFLA_XFRM_IF_ID, if_id);
2974                 }
2975
2976         }
2977
2978         nla_nest_end(nlm, infodata);
2979         nla_nest_end(nlm, linkinfo);
2980
2981         return system_rtnl_call(nlm);
2982
2983 failure:
2984         nlmsg_free(nlm);
2985         return ret;
2986 }
2987 #endif
2988
2989 #ifdef IFLA_VXLAN_MAX
2990 static int system_add_vxlan(const char *name, const unsigned int link, struct blob_attr **tb, bool v6)
2991 {
2992         struct blob_attr *tb_data[__VXLAN_DATA_ATTR_MAX];
2993         struct nl_msg *msg;
2994         struct nlattr *linkinfo, *data;
2995         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
2996         struct blob_attr *cur;
2997         int ret = 0;
2998
2999         if ((cur = tb[TUNNEL_ATTR_DATA]))
3000                 blobmsg_parse(vxlan_data_attr_list.params, __VXLAN_DATA_ATTR_MAX, tb_data,
3001                         blobmsg_data(cur), blobmsg_len(cur));
3002         else
3003                 return -EINVAL;
3004
3005         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
3006
3007         if (!msg)
3008                 return -1;
3009
3010         nlmsg_append(msg, &iim, sizeof(iim), 0);
3011
3012         nla_put_string(msg, IFLA_IFNAME, name);
3013
3014         if ((cur = tb_data[VXLAN_DATA_ATTR_MACADDR])) {
3015                 struct ether_addr *ea = ether_aton(blobmsg_get_string(cur));
3016                 if (!ea) {
3017                         ret = -EINVAL;
3018                         goto failure;
3019                 }
3020
3021                 nla_put(msg, IFLA_ADDRESS, ETH_ALEN, ea);
3022         }
3023
3024         if ((cur = tb[TUNNEL_ATTR_MTU])) {
3025                 uint32_t mtu = blobmsg_get_u32(cur);
3026                 nla_put_u32(msg, IFLA_MTU, mtu);
3027         }
3028
3029         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO))) {
3030                 ret = -ENOMEM;
3031                 goto failure;
3032         }
3033
3034         nla_put_string(msg, IFLA_INFO_KIND, "vxlan");
3035
3036         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA))) {
3037                 ret = -ENOMEM;
3038                 goto failure;
3039         }
3040
3041         if (link)
3042                 nla_put_u32(msg, IFLA_VXLAN_LINK, link);
3043
3044         if ((cur = tb_data[VXLAN_DATA_ATTR_ID])) {
3045                 uint32_t id = blobmsg_get_u32(cur);
3046                 if (id >= (1u << 24) - 1) {
3047                         ret = -EINVAL;
3048                         goto failure;
3049                 }
3050
3051                 nla_put_u32(msg, IFLA_VXLAN_ID, id);
3052         }
3053
3054         if (v6) {
3055                 struct in6_addr in6buf;
3056                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3057                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3058                                 ret = -EINVAL;
3059                                 goto failure;
3060                         }
3061                         nla_put(msg, IFLA_VXLAN_LOCAL6, sizeof(in6buf), &in6buf);
3062                 }
3063
3064                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3065                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3066                                 ret = -EINVAL;
3067                                 goto failure;
3068                         }
3069                         nla_put(msg, IFLA_VXLAN_GROUP6, sizeof(in6buf), &in6buf);
3070                 }
3071         } else {
3072                 struct in_addr inbuf;
3073
3074                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3075                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3076                                 ret = -EINVAL;
3077                                 goto failure;
3078                         }
3079                         nla_put(msg, IFLA_VXLAN_LOCAL, sizeof(inbuf), &inbuf);
3080                 }
3081
3082                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3083                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3084                                 ret = -EINVAL;
3085                                 goto failure;
3086                         }
3087                         nla_put(msg, IFLA_VXLAN_GROUP, sizeof(inbuf), &inbuf);
3088                 }
3089         }
3090
3091         uint32_t port = 4789;
3092         if ((cur = tb_data[VXLAN_DATA_ATTR_PORT])) {
3093                 port = blobmsg_get_u32(cur);
3094                 if (port < 1 || port > 65535) {
3095                         ret = -EINVAL;
3096                         goto failure;
3097                 }
3098         }
3099         nla_put_u16(msg, IFLA_VXLAN_PORT, htons(port));
3100
3101         if ((cur = tb_data[VXLAN_DATA_ATTR_RXCSUM])) {
3102                 bool rxcsum = blobmsg_get_bool(cur);
3103                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_RX, !rxcsum);
3104         }
3105
3106         if ((cur = tb_data[VXLAN_DATA_ATTR_TXCSUM])) {
3107                 bool txcsum = blobmsg_get_bool(cur);
3108                 nla_put_u8(msg, IFLA_VXLAN_UDP_CSUM, txcsum);
3109                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_TX, !txcsum);
3110         }
3111
3112         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3113                 char *str = blobmsg_get_string(cur);
3114                 unsigned tos = 1;
3115
3116                 if (strcmp(str, "inherit")) {
3117                         if (!system_tos_aton(str, &tos)) {
3118                                 ret = -EINVAL;
3119                                 goto failure;
3120                         }
3121                 }
3122
3123                 nla_put_u8(msg, IFLA_VXLAN_TOS, tos);
3124         }
3125
3126         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3127                 uint32_t ttl = blobmsg_get_u32(cur);
3128                 if (ttl < 1 || ttl > 255) {
3129                         ret = -EINVAL;
3130                         goto failure;
3131                 }
3132
3133                 nla_put_u8(msg, IFLA_VXLAN_TTL, ttl);
3134         }
3135
3136         nla_nest_end(msg, data);
3137         nla_nest_end(msg, linkinfo);
3138
3139         ret = system_rtnl_call(msg);
3140         if (ret)
3141                 D(SYSTEM, "Error adding vxlan '%s': %d\n", name, ret);
3142
3143         return ret;
3144
3145 failure:
3146         nlmsg_free(msg);
3147         return ret;
3148 }
3149 #endif
3150
3151 static int system_add_sit_tunnel(const char *name, const unsigned int link, struct blob_attr **tb)
3152 {
3153         struct blob_attr *cur;
3154         int ret = 0;
3155
3156         if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
3157                 return -1;
3158
3159 #ifdef SIOCADD6RD
3160         if ((cur = tb[TUNNEL_ATTR_DATA])) {
3161                 struct blob_attr *tb_data[__SIXRD_DATA_ATTR_MAX];
3162                 unsigned int mask;
3163                 struct ip_tunnel_6rd p6;
3164
3165                 blobmsg_parse(sixrd_data_attr_list.params, __SIXRD_DATA_ATTR_MAX, tb_data,
3166                         blobmsg_data(cur), blobmsg_len(cur));
3167
3168                 memset(&p6, 0, sizeof(p6));
3169
3170                 if ((cur = tb_data[SIXRD_DATA_PREFIX])) {
3171                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
3172                                                 &p6.prefix, &mask) || mask > 128) {
3173                                 ret = -EINVAL;
3174                                 goto failure;
3175                         }
3176
3177                         p6.prefixlen = mask;
3178                 }
3179
3180                 if ((cur = tb_data[SIXRD_DATA_RELAY_PREFIX])) {
3181                         if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
3182                                                 &p6.relay_prefix, &mask) || mask > 32) {
3183                                 ret = -EINVAL;
3184                                 goto failure;
3185                         }
3186
3187                         p6.relay_prefixlen = mask;
3188                 }
3189
3190                 if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
3191                         ret = -1;
3192                         goto failure;
3193                 }
3194         }
3195 #endif
3196
3197         return ret;
3198
3199 failure:
3200         __system_del_ip_tunnel(name, tb);
3201         return ret;
3202 }
3203
3204 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
3205 {
3206         struct blob_attr *cur;
3207         bool set_df = true;
3208         struct ip_tunnel_parm p  = {
3209                 .link = link,
3210                 .iph = {
3211                         .version = 4,
3212                         .ihl = 5,
3213                         .protocol = proto,
3214                 }
3215         };
3216
3217         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
3218                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
3219                 return -EINVAL;
3220
3221         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
3222                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
3223                 return -EINVAL;
3224
3225         if ((cur = tb[TUNNEL_ATTR_DF]))
3226                 set_df = blobmsg_get_bool(cur);
3227
3228         if ((cur = tb[TUNNEL_ATTR_TTL]))
3229                 p.iph.ttl = blobmsg_get_u32(cur);
3230
3231         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3232                 char *str = blobmsg_get_string(cur);
3233                 if (strcmp(str, "inherit")) {
3234                         unsigned uval;
3235
3236                         if (!system_tos_aton(str, &uval))
3237                                 return -EINVAL;
3238
3239                         p.iph.tos = uval;
3240                 } else
3241                         p.iph.tos = 1;
3242         }
3243
3244         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
3245         /* ttl !=0 and nopmtudisc are incompatible */
3246         if (p.iph.ttl && p.iph.frag_off == 0)
3247                 return -EINVAL;
3248
3249         strncpy(p.name, name, sizeof(p.name) - 1);
3250
3251         switch (p.iph.protocol) {
3252         case IPPROTO_IPIP:
3253                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
3254         case IPPROTO_IPV6:
3255                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
3256         default:
3257                 break;
3258         }
3259         return -1;
3260 }
3261
3262 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
3263 {
3264         struct blob_attr *cur;
3265         const char *str;
3266
3267         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3268                 return -EINVAL;
3269         str = blobmsg_data(cur);
3270
3271         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
3272             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
3273             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6") ||
3274             !strcmp(str, "vxlan") || !strcmp(str, "vxlan6") ||
3275             !strcmp(str, "xfrm"))
3276                 return system_link_del(name);
3277         else
3278                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
3279 }
3280
3281 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
3282 {
3283         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3284
3285         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3286                 blob_data(attr), blob_len(attr));
3287
3288         return __system_del_ip_tunnel(name, tb);
3289 }
3290
3291 int system_update_ipv6_mtu(struct device *dev, int mtu)
3292 {
3293         int ret = -1;
3294         char buf[64];
3295         int fd;
3296
3297         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
3298                         dev->ifname);
3299
3300         fd = open(buf, O_RDWR);
3301         if (fd < 0)
3302                 return ret;
3303
3304         if (!mtu) {
3305                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
3306                 if (len < 0)
3307                         goto out;
3308
3309                 buf[len] = 0;
3310                 ret = atoi(buf);
3311         } else {
3312                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
3313                         ret = mtu;
3314         }
3315
3316 out:
3317         close(fd);
3318         return ret;
3319 }
3320
3321 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
3322 {
3323         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3324         struct blob_attr *cur;
3325         const char *str;
3326
3327         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3328                 blob_data(attr), blob_len(attr));
3329
3330         __system_del_ip_tunnel(name, tb);
3331
3332         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3333                 return -EINVAL;
3334         str = blobmsg_data(cur);
3335
3336         unsigned int ttl = 0;
3337         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3338                 ttl = blobmsg_get_u32(cur);
3339                 if (ttl > 255)
3340                         return -EINVAL;
3341         }
3342
3343         unsigned int link = 0;
3344         if ((cur = tb[TUNNEL_ATTR_LINK])) {
3345                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
3346                 if (!iface)
3347                         return -EINVAL;
3348
3349                 if (iface->l3_dev.dev)
3350                         link = iface->l3_dev.dev->ifindex;
3351         }
3352
3353         if (!strcmp(str, "sit"))
3354                 return system_add_sit_tunnel(name, link, tb);
3355 #ifdef IFLA_IPTUN_MAX
3356         else if (!strcmp(str, "ipip6")) {
3357                 return system_add_ip6_tunnel(name, link, tb);
3358         } else if (!strcmp(str, "greip")) {
3359                 return system_add_gre_tunnel(name, "gre", link, tb, false);
3360         } else if (!strcmp(str, "gretapip"))  {
3361                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
3362         } else if (!strcmp(str, "greip6")) {
3363                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
3364         } else if (!strcmp(str, "gretapip6")) {
3365                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
3366 #ifdef IFLA_VTI_MAX
3367         } else if (!strcmp(str, "vtiip")) {
3368                 return system_add_vti_tunnel(name, "vti", link, tb, false);
3369         } else if (!strcmp(str, "vtiip6")) {
3370                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
3371 #endif
3372 #ifdef IFLA_XFRM_MAX
3373         } else if (!strcmp(str, "xfrm")) {
3374                 return system_add_xfrm_tunnel(name, "xfrm", link, tb);
3375 #endif
3376 #ifdef IFLA_VXLAN_MAX
3377         } else if(!strcmp(str, "vxlan")) {
3378                 return system_add_vxlan(name, link, tb, false);
3379         } else if(!strcmp(str, "vxlan6")) {
3380                 return system_add_vxlan(name, link, tb, true);
3381 #endif
3382 #endif
3383         } else if (!strcmp(str, "ipip")) {
3384                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
3385         }
3386         else
3387                 return -EINVAL;
3388
3389         return 0;
3390 }