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  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2
  10  * as published by the Free Software Foundation
  11  *
  12  * This program is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  * GNU General Public License for more details.
  16  */
  17 #define _GNU_SOURCE
  18
  19 #include <sys/socket.h>
  20 #include <sys/ioctl.h>
  21 #include <sys/stat.h>
  22 #include <sys/syscall.h>
  23
  24 #include <net/if.h>
  25 #include <net/if_arp.h>
  26
  27 #include <arpa/inet.h>
  28 #include <netinet/in.h>
  29
  30 #include <linux/rtnetlink.h>
  31 #include <linux/sockios.h>
  32 #include <linux/ip.h>
  33 #include <linux/if_addr.h>
  34 #include <linux/if_link.h>
  35 #include <linux/if_vlan.h>
  36 #include <linux/if_bridge.h>
  37 #include <linux/if_tunnel.h>
  38 #include <linux/ip6_tunnel.h>
  39 #include <linux/ethtool.h>
  40 #include <linux/fib_rules.h>
  41 #include <linux/version.h>
  42
  43 #ifndef RTN_FAILED_POLICY
  44 #define RTN_FAILED_POLICY 12
  45 #endif
  46
  47 #ifndef IFA_F_NOPREFIXROUTE
  48 #define IFA_F_NOPREFIXROUTE 0x200
  49 #endif
  50
  51 #ifndef IFA_FLAGS
  52 #define IFA_FLAGS (IFA_MULTICAST + 1)
  53 #endif
  54
  55
  56 #include <string.h>
  57 #include <fcntl.h>
  58 #include <glob.h>
  59 #include <time.h>
  60 #include <unistd.h>
  61
  62 #include <netlink/msg.h>
  63 #include <netlink/attr.h>
  64 #include <netlink/socket.h>
  65 #include <libubox/uloop.h>
  66
  67 #include "netifd.h"
  68 #include "device.h"
  69 #include "system.h"
  70
  71 struct event_socket {
  72         struct uloop_fd uloop;
  73         struct nl_sock *sock;
  74         int bufsize;
  75 };
  76
  77 static int sock_ioctl = -1;
  78 static struct nl_sock *sock_rtnl = NULL;
  79
  80 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  81 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  82
  83 static char dev_buf[256];
  84
  85 static void
  86 handler_nl_event(struct uloop_fd *u, unsigned int events)
  87 {
  88         struct event_socket *ev = container_of(u, struct event_socket, uloop);
  89         int err;
  90         socklen_t errlen = sizeof(err);
  91
  92         if (!u->error) {
  93                 nl_recvmsgs_default(ev->sock);
  94                 return;
  95         }
  96
  97         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
  98                 goto abort;
  99
 100         switch(err) {
 101         case ENOBUFS:
 102                 // Increase rx buffer size on netlink socket
 103                 ev->bufsize *= 2;
 104                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 105                         goto abort;
 106
 107                 // Request full dump since some info got dropped
 108                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 109                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 110                 break;
 111
 112         default:
 113                 goto abort;
 114         }
 115         u->error = false;
 116         return;
 117
 118 abort:
 119         uloop_fd_delete(&ev->uloop);
 120         return;
 121 }
 122
 123 static struct nl_sock *
 124 create_socket(int protocol, int groups)
 125 {
 126         struct nl_sock *sock;
 127
 128         sock = nl_socket_alloc();
 129         if (!sock)
 130                 return NULL;
 131
 132         if (groups)
 133                 nl_join_groups(sock, groups);
 134
 135         if (nl_connect(sock, protocol))
 136                 return NULL;
 137
 138         return sock;
 139 }
 140
 141 static bool
 142 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 143                         uloop_fd_handler cb, int flags)
 144 {
 145         ev->sock = create_socket(protocol, groups);
 146         if (!ev->sock)
 147                 return false;
 148
 149         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 150         ev->uloop.cb = cb;
 151         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 152                 return false;
 153
 154         return true;
 155 }
 156
 157 static bool
 158 create_event_socket(struct event_socket *ev, int protocol,
 159                     int (*cb)(struct nl_msg *msg, void *arg))
 160 {
 161         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 162                 return false;
 163
 164         // Install the valid custom callback handler
 165         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 166
 167         // Disable sequence number checking on event sockets
 168         nl_socket_disable_seq_check(ev->sock);
 169
 170         // Increase rx buffer size to 65K on event sockets
 171         ev->bufsize = 65535;
 172         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 173                 return false;
 174
 175         return true;
 176 }
 177
 178 static bool
 179 system_rtn_aton(const char *src, unsigned int *dst)
 180 {
 181         char *e;
 182         unsigned int n;
 183
 184         if (!strcmp(src, "local"))
 185                 n = RTN_LOCAL;
 186         else if (!strcmp(src, "nat"))
 187                 n = RTN_NAT;
 188         else if (!strcmp(src, "broadcast"))
 189                 n = RTN_BROADCAST;
 190         else if (!strcmp(src, "anycast"))
 191                 n = RTN_ANYCAST;
 192         else if (!strcmp(src, "multicast"))
 193                 n = RTN_MULTICAST;
 194         else if (!strcmp(src, "prohibit"))
 195                 n = RTN_PROHIBIT;
 196         else if (!strcmp(src, "unreachable"))
 197                 n = RTN_UNREACHABLE;
 198         else if (!strcmp(src, "blackhole"))
 199                 n = RTN_BLACKHOLE;
 200         else if (!strcmp(src, "xresolve"))
 201                 n = RTN_XRESOLVE;
 202         else if (!strcmp(src, "unicast"))
 203                 n = RTN_UNICAST;
 204         else if (!strcmp(src, "throw"))
 205                 n = RTN_THROW;
 206         else if (!strcmp(src, "failed_policy"))
 207                 n = RTN_FAILED_POLICY;
 208         else {
 209                 n = strtoul(src, &e, 0);
 210                 if (!e || *e || e == src || n > 255)
 211                         return false;
 212         }
 213
 214         *dst = n;
 215         return true;
 216 }
 217
 218 static bool
 219 system_tos_aton(const char *src, unsigned *dst)
 220 {
 221         char *e;
 222
 223         *dst = strtoul(src, &e, 16);
 224         if (e == src || *e || *dst > 255)
 225                 return false;
 226
 227         return true;
 228 }
 229
 230 int system_init(void)
 231 {
 232         static struct event_socket rtnl_event;
 233         static struct event_socket hotplug_event;
 234
 235         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 236         system_fd_set_cloexec(sock_ioctl);
 237
 238         // Prepare socket for routing / address control
 239         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 240         if (!sock_rtnl)
 241                 return -1;
 242
 243         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 244                 return -1;
 245
 246         if (!create_raw_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT, 1,
 247                                      handle_hotplug_event, 0))
 248                 return -1;
 249
 250         // Receive network link events form kernel
 251         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 252
 253         return 0;
 254 }
 255
 256 static void system_set_sysctl(const char *path, const char *val)
 257 {
 258         int fd;
 259
 260         fd = open(path, O_WRONLY);
 261         if (fd < 0)
 262                 return;
 263
 264         if (write(fd, val, strlen(val))) {}
 265         close(fd);
 266 }
 267
 268 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 269 {
 270         snprintf(dev_buf, sizeof(dev_buf), path, device);
 271         system_set_sysctl(dev_buf, val);
 272 }
 273
 274 static void system_set_disable_ipv6(struct device *dev, const char *val)
 275 {
 276         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 277 }
 278
 279 static void system_set_rpfilter(struct device *dev, const char *val)
 280 {
 281         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter", dev->ifname, val);
 282 }
 283
 284 static void system_set_acceptlocal(struct device *dev, const char *val)
 285 {
 286         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local", dev->ifname, val);
 287 }
 288
 289 static void system_set_igmpversion(struct device *dev, const char *val)
 290 {
 291         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version", dev->ifname, val);
 292 }
 293
 294 static void system_set_mldversion(struct device *dev, const char *val)
 295 {
 296         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version", dev->ifname, val);
 297 }
 298
 299 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 300 {
 301         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 302 }
 303
 304 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 305 {
 306         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 307 }
 308
 309 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 310 {
 311         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time", dev->ifname, val);
 312 }
 313
 314 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 315 {
 316         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time", dev->ifname, val);
 317 }
 318
 319 static void system_set_dadtransmits(struct device *dev, const char *val)
 320 {
 321         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 322 }
 323
 324 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 325 {
 326         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 327 }
 328
 329 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 330 {
 331         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 332 }
 333
 334 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 335 {
 336         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 337 }
 338
 339 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 340 {
 341         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 342                                        "/sys/class/net/%s/brport/multicast_router",
 343                               dev->ifname, val);
 344 }
 345
 346 static void system_bridge_set_robustness(struct device *dev, const char *val)
 347 {
 348         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 349                               dev->ifname, val);
 350         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 351                               dev->ifname, val);
 352 }
 353
 354 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 355 {
 356         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 357                               dev->ifname, val);
 358 }
 359
 360 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 361 {
 362         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 363                               dev->ifname, val);
 364 }
 365
 366 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 367 {
 368         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 369                               dev->ifname, val);
 370 }
 371
 372 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 373 {
 374         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 375                               dev->ifname, val);
 376 }
 377
 378 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 379 {
 380         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 381                               dev->ifname, val);
 382 }
 383
 384 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 385 {
 386         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 387                               dev->ifname, val);
 388 }
 389
 390 static void system_bridge_set_learning(struct device *dev, const char *val)
 391 {
 392         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 393 }
 394
 395 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 396 {
 397         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 398 }
 399
 400 static void system_set_sendredirects(struct device *dev, const char *val)
 401 {
 402         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 403 }
 404
 405 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 406 {
 407         int fd = -1, ret = -1;
 408
 409         fd = open(path, O_RDONLY);
 410         if (fd < 0)
 411                 goto out;
 412
 413         ssize_t len = read(fd, buf, buf_sz - 1);
 414         if (len < 0)
 415                 goto out;
 416
 417         ret = buf[len] = 0;
 418
 419 out:
 420         if (fd >= 0)
 421                 close(fd);
 422
 423         return ret;
 424 }
 425
 426 static int
 427 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 428 {
 429         snprintf(dev_buf, sizeof(dev_buf), path, device);
 430         return system_get_sysctl(dev_buf, buf, buf_sz);
 431 }
 432
 433 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 434 {
 435         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 436                         dev->ifname, buf, buf_sz);
 437 }
 438
 439 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 440 {
 441         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 442                         dev->ifname, buf, buf_sz);
 443 }
 444
 445 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 446 {
 447         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 448                         dev->ifname, buf, buf_sz);
 449 }
 450
 451 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 452 {
 453         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 454                         dev->ifname, buf, buf_sz);
 455 }
 456
 457 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 458 {
 459         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 460                         dev->ifname, buf, buf_sz);
 461 }
 462
 463 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 464 {
 465         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 466                         dev->ifname, buf, buf_sz);
 467 }
 468
 469 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 470 {
 471         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 472                         dev->ifname, buf, buf_sz);
 473 }
 474
 475 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 476 {
 477         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 478                         dev->ifname, buf, buf_sz);
 479 }
 480
 481 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 482 {
 483         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 484                         dev->ifname, buf, buf_sz);
 485 }
 486
 487 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 488 {
 489         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 490                         dev->ifname, buf, buf_sz);
 491 }
 492
 493 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 494 {
 495         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 496                         dev->ifname, buf, buf_sz);
 497 }
 498
 499 // Evaluate netlink messages
 500 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 501 {
 502         struct nlmsghdr *nh = nlmsg_hdr(msg);
 503         struct nlattr *nla[__IFLA_MAX];
 504         int link_state = 0;
 505         char buf[10];
 506
 507         if (nh->nlmsg_type != RTM_NEWLINK)
 508                 goto out;
 509
 510         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 511         if (!nla[IFLA_IFNAME])
 512                 goto out;
 513
 514         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 515         if (!dev)
 516                 goto out;
 517
 518         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 519                 link_state = strtoul(buf, NULL, 0);
 520
 521         device_set_link(dev, link_state ? true : false);
 522
 523 out:
 524         return 0;
 525 }
 526
 527 static void
 528 handle_hotplug_msg(char *data, int size)
 529 {
 530         const char *subsystem = NULL, *interface = NULL;
 531         char *cur, *end, *sep;
 532         struct device *dev;
 533         int skip;
 534         bool add;
 535
 536         if (!strncmp(data, "add@", 4))
 537                 add = true;
 538         else if (!strncmp(data, "remove@", 7))
 539                 add = false;
 540         else
 541                 return;
 542
 543         skip = strlen(data) + 1;
 544         end = data + size;
 545
 546         for (cur = data + skip; cur < end; cur += skip) {
 547                 skip = strlen(cur) + 1;
 548
 549                 sep = strchr(cur, '=');
 550                 if (!sep)
 551                         continue;
 552
 553                 *sep = 0;
 554                 if (!strcmp(cur, "INTERFACE"))
 555                         interface = sep + 1;
 556                 else if (!strcmp(cur, "SUBSYSTEM")) {
 557                         subsystem = sep + 1;
 558                         if (strcmp(subsystem, "net") != 0)
 559                                 return;
 560                 }
 561                 if (subsystem && interface)
 562                         goto found;
 563         }
 564         return;
 565
 566 found:
 567         dev = device_find(interface);
 568         if (!dev)
 569                 return;
 570
 571         if (dev->type != &simple_device_type)
 572                 return;
 573
 574         if (add && system_if_force_external(dev->ifname))
 575                 return;
 576
 577         device_set_present(dev, add);
 578 }
 579
 580 static void
 581 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 582 {
 583         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 584         struct sockaddr_nl nla;
 585         unsigned char *buf = NULL;
 586         int size;
 587
 588         while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 589                 if (nla.nl_pid == 0)
 590                         handle_hotplug_msg((char *) buf, size);
 591
 592                 free(buf);
 593         }
 594 }
 595
 596 static int system_rtnl_call(struct nl_msg *msg)
 597 {
 598         int ret;
 599
 600         ret = nl_send_auto_complete(sock_rtnl, msg);
 601         nlmsg_free(msg);
 602
 603         if (ret < 0)
 604                 return ret;
 605
 606         return nl_wait_for_ack(sock_rtnl);
 607 }
 608
 609 int system_bridge_delbr(struct device *bridge)
 610 {
 611         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 612 }
 613
 614 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 615 {
 616         struct ifreq ifr;
 617
 618         memset(&ifr, 0, sizeof(ifr));
 619         if (dev)
 620                 ifr.ifr_ifindex = dev->ifindex;
 621         else
 622                 ifr.ifr_data = data;
 623         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name));
 624         return ioctl(sock_ioctl, cmd, &ifr);
 625 }
 626
 627 static bool system_is_bridge(const char *name, char *buf, int buflen)
 628 {
 629         struct stat st;
 630
 631         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 632         if (stat(buf, &st) < 0)
 633                 return false;
 634
 635         return true;
 636 }
 637
 638 static char *system_get_bridge(const char *name, char *buf, int buflen)
 639 {
 640         char *path;
 641         ssize_t len = -1;
 642         glob_t gl;
 643
 644         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 645         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 646                 return NULL;
 647
 648         if (gl.gl_pathc > 0)
 649                 len = readlink(gl.gl_pathv[0], buf, buflen);
 650
 651         globfree(&gl);
 652
 653         if (len < 0)
 654                 return NULL;
 655
 656         buf[len] = 0;
 657         path = strrchr(buf, '/');
 658         if (!path)
 659                 return NULL;
 660
 661         return path + 1;
 662 }
 663
 664 static void
 665 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 666 {
 667         bool mcast_to_ucast = dev->wireless_ap;
 668         bool hairpin = true;
 669
 670         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 671             !bridge->settings.multicast_to_unicast)
 672                 mcast_to_ucast = false;
 673
 674         if (!mcast_to_ucast || dev->wireless_isolate)
 675                 hairpin = false;
 676
 677         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 678         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 679 }
 680
 681 int system_bridge_addif(struct device *bridge, struct device *dev)
 682 {
 683         char buf[64];
 684         char *oldbr;
 685         int ret = 0;
 686
 687         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 688         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 689                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 690
 691         if (dev->wireless)
 692                 system_bridge_set_wireless(bridge, dev);
 693
 694         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 695                 snprintf(buf, sizeof(buf), "%i", dev->settings.multicast_router);
 696                 system_bridge_set_multicast_router(dev, buf, false);
 697         }
 698
 699         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 700             dev->settings.multicast_fast_leave)
 701                 system_bridge_set_multicast_fast_leave(dev, "1");
 702
 703         if (dev->settings.flags & DEV_OPT_LEARNING &&
 704             !dev->settings.learning)
 705                 system_bridge_set_learning(dev, "0");
 706
 707         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 708             !dev->settings.unicast_flood)
 709                 system_bridge_set_unicast_flood(dev, "0");
 710
 711         return ret;
 712 }
 713
 714 int system_bridge_delif(struct device *bridge, struct device *dev)
 715 {
 716         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 717 }
 718
 719 int system_if_resolve(struct device *dev)
 720 {
 721         struct ifreq ifr;
 722         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
 723         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 724                 return ifr.ifr_ifindex;
 725         else
 726                 return 0;
 727 }
 728
 729 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 730 {
 731         struct ifreq ifr;
 732
 733         memset(&ifr, 0, sizeof(ifr));
 734         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name));
 735         ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr);
 736         ifr.ifr_flags |= add;
 737         ifr.ifr_flags &= ~rem;
 738         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 739 }
 740
 741 struct clear_data {
 742         struct nl_msg *msg;
 743         struct device *dev;
 744         int type;
 745         int size;
 746         int af;
 747 };
 748
 749
 750 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 751 {
 752         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 753
 754         return ifa->ifa_index == ifindex;
 755 }
 756
 757 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 758 {
 759         struct rtmsg *r = NLMSG_DATA(hdr);
 760         struct nlattr *tb[__RTA_MAX];
 761
 762         if (r->rtm_protocol == RTPROT_KERNEL &&
 763             r->rtm_family == AF_INET6)
 764                 return false;
 765
 766         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 767         if (!tb[RTA_OIF])
 768                 return false;
 769
 770         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 771 }
 772
 773 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 774 {
 775         return true;
 776 }
 777
 778 static int cb_clear_event(struct nl_msg *msg, void *arg)
 779 {
 780         struct clear_data *clr = arg;
 781         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 782         bool (*cb)(struct nlmsghdr *, int ifindex);
 783         int type;
 784
 785         switch(clr->type) {
 786         case RTM_GETADDR:
 787                 type = RTM_DELADDR;
 788                 if (hdr->nlmsg_type != RTM_NEWADDR)
 789                         return NL_SKIP;
 790
 791                 cb = check_ifaddr;
 792                 break;
 793         case RTM_GETROUTE:
 794                 type = RTM_DELROUTE;
 795                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 796                         return NL_SKIP;
 797
 798                 cb = check_route;
 799                 break;
 800         case RTM_GETRULE:
 801                 type = RTM_DELRULE;
 802                 if (hdr->nlmsg_type != RTM_NEWRULE)
 803                         return NL_SKIP;
 804
 805                 cb = check_rule;
 806                 break;
 807         default:
 808                 return NL_SKIP;
 809         }
 810
 811         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 812                 return NL_SKIP;
 813
 814         if (type == RTM_DELRULE)
 815                 D(SYSTEM, "Remove a rule\n");
 816         else
 817                 D(SYSTEM, "Remove %s from device %s\n",
 818                   type == RTM_DELADDR ? "an address" : "a route",
 819                   clr->dev->ifname);
 820         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 821         hdr = nlmsg_hdr(clr->msg);
 822         hdr->nlmsg_type = type;
 823         hdr->nlmsg_flags = NLM_F_REQUEST;
 824
 825         nl_socket_disable_auto_ack(sock_rtnl);
 826         nl_send_auto_complete(sock_rtnl, clr->msg);
 827         nl_socket_enable_auto_ack(sock_rtnl);
 828
 829         return NL_SKIP;
 830 }
 831
 832 static int
 833 cb_finish_event(struct nl_msg *msg, void *arg)
 834 {
 835         int *pending = arg;
 836         *pending = 0;
 837         return NL_STOP;
 838 }
 839
 840 static int
 841 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 842 {
 843         int *pending = arg;
 844         *pending = err->error;
 845         return NL_STOP;
 846 }
 847
 848 static void
 849 system_if_clear_entries(struct device *dev, int type, int af)
 850 {
 851         struct clear_data clr;
 852         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
 853         struct rtmsg rtm = {
 854                 .rtm_family = af,
 855                 .rtm_flags = RTM_F_CLONED,
 856         };
 857         int flags = NLM_F_DUMP;
 858         int pending = 1;
 859
 860         clr.af = af;
 861         clr.dev = dev;
 862         clr.type = type;
 863         switch (type) {
 864         case RTM_GETADDR:
 865         case RTM_GETRULE:
 866                 clr.size = sizeof(struct rtgenmsg);
 867                 break;
 868         case RTM_GETROUTE:
 869                 clr.size = sizeof(struct rtmsg);
 870                 break;
 871         default:
 872                 return;
 873         }
 874
 875         if (!cb)
 876                 return;
 877
 878         clr.msg = nlmsg_alloc_simple(type, flags);
 879         if (!clr.msg)
 880                 goto out;
 881
 882         nlmsg_append(clr.msg, &rtm, clr.size, 0);
 883         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
 884         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
 885         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
 886
 887         nl_send_auto_complete(sock_rtnl, clr.msg);
 888         while (pending > 0)
 889                 nl_recvmsgs(sock_rtnl, cb);
 890
 891         nlmsg_free(clr.msg);
 892 out:
 893         nl_cb_put(cb);
 894 }
 895
 896 /*
 897  * Clear bridge (membership) state and bring down device
 898  */
 899 void system_if_clear_state(struct device *dev)
 900 {
 901         static char buf[256];
 902         char *bridge;
 903
 904         device_set_ifindex(dev, system_if_resolve(dev));
 905         if (dev->external || !dev->ifindex)
 906                 return;
 907
 908         system_if_flags(dev->ifname, 0, IFF_UP);
 909
 910         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
 911                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
 912                 system_bridge_delbr(dev);
 913                 return;
 914         }
 915
 916         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
 917         if (bridge) {
 918                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
 919                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
 920         }
 921
 922         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
 923         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
 924         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
 925         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
 926         system_set_disable_ipv6(dev, "0");
 927 }
 928
 929 static inline unsigned long
 930 sec_to_jiffies(int val)
 931 {
 932         return (unsigned long) val * 100;
 933 }
 934
 935 static void system_bridge_conf_multicast_deps(struct device *bridge,
 936                                               struct bridge_config *cfg,
 937                                               char *buf,
 938                                               int buf_len)
 939 {
 940         int val;
 941
 942         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
 943             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
 944             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
 945                 val = cfg->robustness * cfg->query_interval +
 946                         cfg->query_response_interval;
 947
 948                 snprintf(buf, buf_len, "%i", val);
 949                 system_bridge_set_membership_interval(bridge, buf);
 950
 951                 val = cfg->robustness * cfg->query_interval +
 952                         cfg->query_response_interval / 2;
 953
 954                 snprintf(buf, buf_len, "%i", val);
 955                 system_bridge_set_other_querier_timeout(bridge, buf);
 956         }
 957
 958         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
 959                 val = cfg->query_interval / 4;
 960
 961                 snprintf(buf, buf_len, "%i", val);
 962                 system_bridge_set_startup_query_interval(bridge, buf);
 963         }
 964 }
 965
 966 static void system_bridge_conf_multicast(struct device *bridge,
 967                                          struct bridge_config *cfg,
 968                                          char *buf,
 969                                          int buf_len)
 970 {
 971         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
 972                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
 973
 974         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
 975                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
 976
 977         snprintf(buf, buf_len, "%i", cfg->hash_max);
 978         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
 979                 bridge->ifname, buf);
 980
 981         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 982                 snprintf(buf, buf_len, "%i", bridge->settings.multicast_router);
 983                 system_bridge_set_multicast_router(bridge, buf, true);
 984         }
 985
 986         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
 987                 snprintf(buf, buf_len, "%i", cfg->robustness);
 988                 system_bridge_set_robustness(bridge, buf);
 989         }
 990
 991         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
 992                 snprintf(buf, buf_len, "%i", cfg->query_interval);
 993                 system_bridge_set_query_interval(bridge, buf);
 994         }
 995
 996         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
 997                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
 998                 system_bridge_set_query_response_interval(bridge, buf);
 999         }
1000
1001         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1002                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1003                 system_bridge_set_last_member_interval(bridge, buf);
1004         }
1005
1006         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1007 }
1008
1009 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1010 {
1011         char buf[64];
1012         unsigned long args[4] = {};
1013
1014         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1015                 return -1;
1016
1017         args[0] = BRCTL_SET_BRIDGE_STP_STATE;
1018         args[1] = !!cfg->stp;
1019         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1020
1021         args[0] = BRCTL_SET_BRIDGE_FORWARD_DELAY;
1022         args[1] = sec_to_jiffies(cfg->forward_delay);
1023         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1024
1025         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1026
1027         args[0] = BRCTL_SET_BRIDGE_PRIORITY;
1028         args[1] = cfg->priority;
1029         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1030
1031         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1032                 args[0] = BRCTL_SET_AGEING_TIME;
1033                 args[1] = sec_to_jiffies(cfg->ageing_time);
1034                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1035         }
1036
1037         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1038                 args[0] = BRCTL_SET_BRIDGE_HELLO_TIME;
1039                 args[1] = sec_to_jiffies(cfg->hello_time);
1040                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1041         }
1042
1043         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1044                 args[0] = BRCTL_SET_BRIDGE_MAX_AGE;
1045                 args[1] = sec_to_jiffies(cfg->max_age);
1046                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1047         }
1048
1049         return 0;
1050 }
1051
1052 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1053 {
1054         struct nl_msg *msg;
1055         struct nlattr *linkinfo, *data;
1056         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1057         int i, rv;
1058         static const struct {
1059                 const char *name;
1060                 enum macvlan_mode val;
1061         } modes[] = {
1062                 { "private", MACVLAN_MODE_PRIVATE },
1063                 { "vepa", MACVLAN_MODE_VEPA },
1064                 { "bridge", MACVLAN_MODE_BRIDGE },
1065                 { "passthru", MACVLAN_MODE_PASSTHRU },
1066         };
1067
1068         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1069
1070         if (!msg)
1071                 return -1;
1072
1073         nlmsg_append(msg, &iim, sizeof(iim), 0);
1074
1075         if (cfg->flags & MACVLAN_OPT_MACADDR)
1076                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1077         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1078         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1079
1080         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1081                 goto nla_put_failure;
1082
1083         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1084
1085         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1086                 goto nla_put_failure;
1087
1088         if (cfg->mode) {
1089                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1090                         if (strcmp(cfg->mode, modes[i].name) != 0)
1091                                 continue;
1092
1093                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1094                         break;
1095                 }
1096         }
1097
1098         nla_nest_end(msg, data);
1099         nla_nest_end(msg, linkinfo);
1100
1101         rv = system_rtnl_call(msg);
1102         if (rv)
1103                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1104
1105         return rv;
1106
1107 nla_put_failure:
1108         nlmsg_free(msg);
1109         return -ENOMEM;
1110 }
1111
1112 static int system_link_del(const char *ifname)
1113 {
1114         struct nl_msg *msg;
1115         struct ifinfomsg iim = {
1116                 .ifi_family = AF_UNSPEC,
1117                 .ifi_index = 0,
1118         };
1119
1120         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1121
1122         if (!msg)
1123                 return -1;
1124
1125         nlmsg_append(msg, &iim, sizeof(iim), 0);
1126         nla_put_string(msg, IFLA_IFNAME, ifname);
1127         return system_rtnl_call(msg);
1128 }
1129
1130 int system_macvlan_del(struct device *macvlan)
1131 {
1132         return system_link_del(macvlan->ifname);
1133 }
1134
1135 static int system_vlan(struct device *dev, int id)
1136 {
1137         struct vlan_ioctl_args ifr = {
1138                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1139                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1140         };
1141
1142         ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1143
1144         if (id < 0) {
1145                 ifr.cmd = DEL_VLAN_CMD;
1146                 ifr.u.VID = 0;
1147         } else {
1148                 ifr.cmd = ADD_VLAN_CMD;
1149                 ifr.u.VID = id;
1150         }
1151         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1152         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1153 }
1154
1155 int system_vlan_add(struct device *dev, int id)
1156 {
1157         return system_vlan(dev, id);
1158 }
1159
1160 int system_vlan_del(struct device *dev)
1161 {
1162         return system_vlan(dev, -1);
1163 }
1164
1165 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1166 {
1167         struct nl_msg *msg;
1168         struct nlattr *linkinfo, *data;
1169         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1170         int rv;
1171
1172         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1173
1174         if (!msg)
1175                 return -1;
1176
1177         nlmsg_append(msg, &iim, sizeof(iim), 0);
1178         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1179         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1180
1181         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1182                 goto nla_put_failure;
1183
1184         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1185
1186         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1187                 goto nla_put_failure;
1188
1189         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1190
1191 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1192         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1193 #else
1194         if(cfg->proto == VLAN_PROTO_8021AD)
1195                 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);
1196 #endif
1197
1198         nla_nest_end(msg, data);
1199         nla_nest_end(msg, linkinfo);
1200
1201         rv = system_rtnl_call(msg);
1202         if (rv)
1203                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1204
1205         return rv;
1206
1207 nla_put_failure:
1208         nlmsg_free(msg);
1209         return -ENOMEM;
1210 }
1211
1212 int system_vlandev_del(struct device *vlandev)
1213 {
1214         return system_link_del(vlandev->ifname);
1215 }
1216
1217 void
1218 system_if_get_settings(struct device *dev, struct device_settings *s)
1219 {
1220         struct ifreq ifr;
1221         char buf[10];
1222
1223         memset(&ifr, 0, sizeof(ifr));
1224         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1225
1226         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1227                 s->mtu = ifr.ifr_mtu;
1228                 s->flags |= DEV_OPT_MTU;
1229         }
1230
1231         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1232         if (s->mtu6 > 0)
1233                 s->flags |= DEV_OPT_MTU6;
1234
1235         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1236                 s->txqueuelen = ifr.ifr_qlen;
1237                 s->flags |= DEV_OPT_TXQUEUELEN;
1238         }
1239
1240         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1241                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1242                 s->flags |= DEV_OPT_MACADDR;
1243         }
1244
1245         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1246                 s->ipv6 = !strtoul(buf, NULL, 0);
1247                 s->flags |= DEV_OPT_IPV6;
1248         }
1249
1250         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1251                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1252                 s->flags |= DEV_OPT_PROMISC;
1253
1254                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1255                 s->flags |= DEV_OPT_MULTICAST;
1256         }
1257
1258         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1259                 s->rpfilter = strtoul(buf, NULL, 0);
1260                 s->flags |= DEV_OPT_RPFILTER;
1261         }
1262
1263         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1264                 s->acceptlocal = strtoul(buf, NULL, 0);
1265                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1266         }
1267
1268         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1269                 s->igmpversion = strtoul(buf, NULL, 0);
1270                 s->flags |= DEV_OPT_IGMPVERSION;
1271         }
1272
1273         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1274                 s->mldversion = strtoul(buf, NULL, 0);
1275                 s->flags |= DEV_OPT_MLDVERSION;
1276         }
1277
1278         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1279                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1280                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1281         }
1282
1283         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1284                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1285                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1286         }
1287
1288         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1289                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1290                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1291         }
1292
1293         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1294                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1295                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1296         }
1297
1298         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1299                 s->dadtransmits = strtoul(buf, NULL, 0);
1300                 s->flags |= DEV_OPT_DADTRANSMITS;
1301         }
1302
1303         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1304                 s->sendredirects = strtoul(buf, NULL, 0);
1305                 s->flags |= DEV_OPT_SENDREDIRECTS;
1306         }
1307 }
1308
1309 static void
1310 system_if_set_rps_xps_val(const char *path, int val)
1311 {
1312         char val_buf[8];
1313         glob_t gl;
1314         int i;
1315
1316         if (glob(path, 0, NULL, &gl))
1317                 return;
1318
1319         snprintf(val_buf, sizeof(val_buf), "%x", val);
1320         for (i = 0; i < gl.gl_pathc; i++)
1321                 system_set_sysctl(gl.gl_pathv[i], val_buf);
1322
1323         globfree(&gl);
1324 }
1325
1326 static void
1327 system_if_apply_rps_xps(struct device *dev, struct device_settings *s)
1328 {
1329         long n_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1330         int val;
1331
1332         if (n_cpus < 2)
1333                 return;
1334
1335         val = (1 << n_cpus) - 1;
1336         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/rps_cpus", dev->ifname);
1337         system_if_set_rps_xps_val(dev_buf, s->rps ? val : 0);
1338
1339         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/xps_cpus", dev->ifname);
1340         system_if_set_rps_xps_val(dev_buf, s->xps ? val : 0);
1341 }
1342
1343 void
1344 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1345 {
1346         struct ifreq ifr;
1347         char buf[12];
1348
1349         memset(&ifr, 0, sizeof(ifr));
1350         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1351         if (s->flags & DEV_OPT_MTU & apply_mask) {
1352                 ifr.ifr_mtu = s->mtu;
1353                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1354                         s->flags &= ~DEV_OPT_MTU;
1355         }
1356         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1357                 system_update_ipv6_mtu(dev, s->mtu6);
1358         }
1359         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1360                 ifr.ifr_qlen = s->txqueuelen;
1361                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1362                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1363         }
1364         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1365                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1366                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1367                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1368                         s->flags &= ~DEV_OPT_MACADDR;
1369         }
1370         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1371                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1372         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1373                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1374                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1375                         s->flags &= ~DEV_OPT_PROMISC;
1376         }
1377         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1378                 snprintf(buf, sizeof(buf), "%d", s->rpfilter);
1379                 system_set_rpfilter(dev, buf);
1380         }
1381         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1382                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1383         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1384                 snprintf(buf, sizeof(buf), "%d", s->igmpversion);
1385                 system_set_igmpversion(dev, buf);
1386         }
1387         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1388                 snprintf(buf, sizeof(buf), "%d", s->mldversion);
1389                 system_set_mldversion(dev, buf);
1390         }
1391         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1392                 snprintf(buf, sizeof(buf), "%d", s->neigh4reachabletime);
1393                 system_set_neigh4reachabletime(dev, buf);
1394                 snprintf(buf, sizeof(buf), "%d", s->neigh6reachabletime);
1395                 system_set_neigh6reachabletime(dev, buf);
1396         }
1397         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1398                 snprintf(buf, sizeof(buf), "%d", s->neigh4gcstaletime);
1399                 system_set_neigh4gcstaletime(dev, buf);
1400                 snprintf(buf, sizeof(buf), "%d", s->neigh6gcstaletime);
1401                 system_set_neigh6gcstaletime(dev, buf);
1402         }
1403         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1404                 snprintf(buf, sizeof(buf), "%d", s->dadtransmits);
1405                 system_set_dadtransmits(dev, buf);
1406         }
1407         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1408                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1409                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1410                         s->flags &= ~DEV_OPT_MULTICAST;
1411         }
1412         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1413                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1414
1415         system_if_apply_rps_xps(dev, s);
1416 }
1417
1418 int system_if_up(struct device *dev)
1419 {
1420         system_if_get_settings(dev, &dev->orig_settings);
1421         /* Only keep orig settings based on what needs to be set */
1422         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1423         dev->orig_settings.flags &= dev->settings.flags;
1424         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1425         return system_if_flags(dev->ifname, IFF_UP, 0);
1426 }
1427
1428 int system_if_down(struct device *dev)
1429 {
1430         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1431         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1432         return ret;
1433 }
1434
1435 struct if_check_data {
1436         struct device *dev;
1437         int pending;
1438         int ret;
1439 };
1440
1441 #ifndef IFF_LOWER_UP
1442 #define IFF_LOWER_UP    0x10000
1443 #endif
1444
1445 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1446 {
1447         struct nlmsghdr *nh = nlmsg_hdr(msg);
1448         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1449         struct if_check_data *chk = (struct if_check_data *)arg;
1450
1451         if (nh->nlmsg_type != RTM_NEWLINK)
1452                 return NL_SKIP;
1453
1454         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1455         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1456
1457         return NL_OK;
1458 }
1459
1460 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1461 {
1462         struct if_check_data *chk = (struct if_check_data *)arg;
1463         chk->pending = 0;
1464         return NL_STOP;
1465 }
1466
1467 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1468 {
1469         struct if_check_data *chk = (struct if_check_data *)arg;
1470
1471         device_set_present(chk->dev, false);
1472         device_set_link(chk->dev, false);
1473         chk->pending = err->error;
1474
1475         return NL_STOP;
1476 }
1477
1478 int system_if_check(struct device *dev)
1479 {
1480         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1481         struct nl_msg *msg;
1482         struct ifinfomsg ifi = {
1483                 .ifi_family = AF_UNSPEC,
1484                 .ifi_index = 0,
1485         };
1486         struct if_check_data chk = {
1487                 .dev = dev,
1488                 .pending = 1,
1489         };
1490         int ret = 1;
1491
1492         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1493         if (!msg)
1494                 goto out;
1495
1496         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1497             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1498                 goto free;
1499
1500         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1501         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1502         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1503
1504         nl_send_auto_complete(sock_rtnl, msg);
1505         while (chk.pending > 0)
1506                 nl_recvmsgs(sock_rtnl, cb);
1507
1508         ret = chk.pending;
1509
1510 free:
1511         nlmsg_free(msg);
1512 out:
1513         nl_cb_put(cb);
1514         return ret;
1515 }
1516
1517 struct device *
1518 system_if_get_parent(struct device *dev)
1519 {
1520         char buf[64], *devname;
1521         int ifindex, iflink, len;
1522         FILE *f;
1523
1524         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1525         f = fopen(buf, "r");
1526         if (!f)
1527                 return NULL;
1528
1529         len = fread(buf, 1, sizeof(buf) - 1, f);
1530         fclose(f);
1531
1532         if (len <= 0)
1533                 return NULL;
1534
1535         buf[len] = 0;
1536         iflink = strtoul(buf, NULL, 0);
1537         ifindex = system_if_resolve(dev);
1538         if (!iflink || iflink == ifindex)
1539                 return NULL;
1540
1541         devname = if_indextoname(iflink, buf);
1542         if (!devname)
1543                 return NULL;
1544
1545         return device_get(devname, true);
1546 }
1547
1548 static bool
1549 read_string_file(int dir_fd, const char *file, char *buf, int len)
1550 {
1551         bool ret = false;
1552         char *c;
1553         int fd;
1554
1555         fd = openat(dir_fd, file, O_RDONLY);
1556         if (fd < 0)
1557                 return false;
1558
1559 retry:
1560         len = read(fd, buf, len - 1);
1561         if (len < 0) {
1562                 if (errno == EINTR)
1563                         goto retry;
1564         } else if (len > 0) {
1565                         buf[len] = 0;
1566
1567                         c = strchr(buf, '\n');
1568                         if (c)
1569                                 *c = 0;
1570
1571                         ret = true;
1572         }
1573
1574         close(fd);
1575
1576         return ret;
1577 }
1578
1579 static bool
1580 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1581 {
1582         char buf[64];
1583         bool ret = false;
1584
1585         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1586         if (ret)
1587                 *val = strtoull(buf, NULL, 0);
1588
1589         return ret;
1590 }
1591
1592 /* Assume advertised flags == supported flags */
1593 static const struct {
1594         uint32_t mask;
1595         const char *name;
1596 } ethtool_link_modes[] = {
1597         { ADVERTISED_10baseT_Half, "10H" },
1598         { ADVERTISED_10baseT_Full, "10F" },
1599         { ADVERTISED_100baseT_Half, "100H" },
1600         { ADVERTISED_100baseT_Full, "100F" },
1601         { ADVERTISED_1000baseT_Half, "1000H" },
1602         { ADVERTISED_1000baseT_Full, "1000F" },
1603 };
1604
1605 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1606 {
1607         int i;
1608         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1609                 if (mask & ethtool_link_modes[i].mask)
1610                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1611         }
1612 }
1613
1614 bool
1615 system_if_force_external(const char *ifname)
1616 {
1617         char buf[64];
1618         struct stat s;
1619
1620         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1621         return stat(buf, &s) == 0;
1622 }
1623
1624 int
1625 system_if_dump_info(struct device *dev, struct blob_buf *b)
1626 {
1627         struct ethtool_cmd ecmd;
1628         struct ifreq ifr;
1629         char buf[64], *s;
1630         void *c;
1631         int dir_fd;
1632
1633         snprintf(buf, sizeof(buf), "/sys/class/net/%s", dev->ifname);
1634         dir_fd = open(buf, O_DIRECTORY);
1635
1636         memset(&ecmd, 0, sizeof(ecmd));
1637         memset(&ifr, 0, sizeof(ifr));
1638         strcpy(ifr.ifr_name, dev->ifname);
1639         ifr.ifr_data = (caddr_t) &ecmd;
1640         ecmd.cmd = ETHTOOL_GSET;
1641
1642         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1643                 c = blobmsg_open_array(b, "link-advertising");
1644                 system_add_link_modes(b, ecmd.advertising);
1645                 blobmsg_close_array(b, c);
1646
1647                 c = blobmsg_open_array(b, "link-supported");
1648                 system_add_link_modes(b, ecmd.supported);
1649                 blobmsg_close_array(b, c);
1650
1651                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1652                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1653                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1654                 blobmsg_add_string_buffer(b);
1655         }
1656
1657         close(dir_fd);
1658         return 0;
1659 }
1660
1661 int
1662 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1663 {
1664         const char *const counters[] = {
1665                 "collisions",     "rx_frame_errors",   "tx_compressed",
1666                 "multicast",      "rx_length_errors",  "tx_dropped",
1667                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1668                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1669                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1670                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1671                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1672                 "rx_fifo_errors", "tx_carrier_errors",
1673         };
1674         char buf[64];
1675         int stats_dir;
1676         int i;
1677         uint64_t val = 0;
1678
1679         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1680         stats_dir = open(buf, O_DIRECTORY);
1681         if (stats_dir < 0)
1682                 return -1;
1683
1684         for (i = 0; i < ARRAY_SIZE(counters); i++)
1685                 if (read_uint64_file(stats_dir, counters[i], &val))
1686                         blobmsg_add_u64(b, counters[i], val);
1687
1688         close(stats_dir);
1689         return 0;
1690 }
1691
1692 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1693 {
1694         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1695         int alen = v4 ? 4 : 16;
1696         unsigned int flags = 0;
1697         struct ifaddrmsg ifa = {
1698                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1699                 .ifa_prefixlen = addr->mask,
1700                 .ifa_index = dev->ifindex,
1701         };
1702
1703         struct nl_msg *msg;
1704         if (cmd == RTM_NEWADDR)
1705                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1706
1707         msg = nlmsg_alloc_simple(cmd, flags);
1708         if (!msg)
1709                 return -1;
1710
1711         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1712         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1713         if (v4) {
1714                 if (addr->broadcast)
1715                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1716                 if (addr->point_to_point)
1717                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1718         } else {
1719                 time_t now = system_get_rtime();
1720                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1721
1722                 if (addr->preferred_until) {
1723                         int64_t preferred = addr->preferred_until - now;
1724                         if (preferred < 0)
1725                                 preferred = 0;
1726                         else if (preferred > UINT32_MAX)
1727                                 preferred = UINT32_MAX;
1728
1729                         cinfo.ifa_prefered = preferred;
1730                 }
1731
1732                 if (addr->valid_until) {
1733                         int64_t valid = addr->valid_until - now;
1734                         if (valid <= 0) {
1735                                 nlmsg_free(msg);
1736                                 return -1;
1737                         }
1738                         else if (valid > UINT32_MAX)
1739                                 valid = UINT32_MAX;
1740
1741                         cinfo.ifa_valid = valid;
1742                 }
1743
1744                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
1745
1746                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
1747                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
1748         }
1749
1750         return system_rtnl_call(msg);
1751 }
1752
1753 int system_add_address(struct device *dev, struct device_addr *addr)
1754 {
1755         return system_addr(dev, addr, RTM_NEWADDR);
1756 }
1757
1758 int system_del_address(struct device *dev, struct device_addr *addr)
1759 {
1760         return system_addr(dev, addr, RTM_DELADDR);
1761 }
1762
1763 static int system_rt(struct device *dev, struct device_route *route, int cmd)
1764 {
1765         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1766         bool have_gw;
1767         unsigned int flags = 0;
1768
1769         if (alen == 4)
1770                 have_gw = !!route->nexthop.in.s_addr;
1771         else
1772                 have_gw = route->nexthop.in6.s6_addr32[0] ||
1773                         route->nexthop.in6.s6_addr32[1] ||
1774                         route->nexthop.in6.s6_addr32[2] ||
1775                         route->nexthop.in6.s6_addr32[3];
1776
1777         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
1778                         ? route->table : RT_TABLE_MAIN;
1779
1780         struct rtmsg rtm = {
1781                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
1782                 .rtm_dst_len = route->mask,
1783                 .rtm_src_len = route->sourcemask,
1784                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
1785                 .rtm_protocol = (route->flags & DEVADDR_KERNEL) ? RTPROT_KERNEL : RTPROT_STATIC,
1786                 .rtm_scope = RT_SCOPE_NOWHERE,
1787                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
1788                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
1789         };
1790         struct nl_msg *msg;
1791
1792         if (cmd == RTM_NEWROUTE) {
1793                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1794
1795                 if (!dev) { // Add null-route
1796                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1797                         rtm.rtm_type = RTN_UNREACHABLE;
1798                 }
1799                 else
1800                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
1801         }
1802
1803         if (route->flags & DEVROUTE_TYPE) {
1804                 rtm.rtm_type = route->type;
1805                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
1806                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
1807                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
1808                                 rtm.rtm_table = RT_TABLE_LOCAL;
1809                 }
1810
1811                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
1812                         rtm.rtm_scope = RT_SCOPE_HOST;
1813                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
1814                                 rtm.rtm_type == RTN_ANYCAST) {
1815                         rtm.rtm_scope = RT_SCOPE_LINK;
1816                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
1817                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY) {
1818                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1819                         dev = NULL;
1820                 }
1821         }
1822
1823         msg = nlmsg_alloc_simple(cmd, flags);
1824         if (!msg)
1825                 return -1;
1826
1827         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
1828
1829         if (route->mask)
1830                 nla_put(msg, RTA_DST, alen, &route->addr);
1831
1832         if (route->sourcemask) {
1833                 if (rtm.rtm_family == AF_INET)
1834                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
1835                 else
1836                         nla_put(msg, RTA_SRC, alen, &route->source);
1837         }
1838
1839         if (route->metric > 0)
1840                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
1841
1842         if (have_gw)
1843                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
1844
1845         if (dev)
1846                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
1847
1848         if (table >= 256)
1849                 nla_put_u32(msg, RTA_TABLE, table);
1850
1851         if (route->flags & DEVROUTE_MTU) {
1852                 struct nlattr *metrics;
1853
1854                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
1855                         goto nla_put_failure;
1856
1857                 nla_put_u32(msg, RTAX_MTU, route->mtu);
1858
1859                 nla_nest_end(msg, metrics);
1860         }
1861
1862         return system_rtnl_call(msg);
1863
1864 nla_put_failure:
1865         nlmsg_free(msg);
1866         return -ENOMEM;
1867 }
1868
1869 int system_add_route(struct device *dev, struct device_route *route)
1870 {
1871         return system_rt(dev, route, RTM_NEWROUTE);
1872 }
1873
1874 int system_del_route(struct device *dev, struct device_route *route)
1875 {
1876         return system_rt(dev, route, RTM_DELROUTE);
1877 }
1878
1879 int system_flush_routes(void)
1880 {
1881         const char *names[] = {
1882                 "/proc/sys/net/ipv4/route/flush",
1883                 "/proc/sys/net/ipv6/route/flush"
1884         };
1885         int fd, i;
1886
1887         for (i = 0; i < ARRAY_SIZE(names); i++) {
1888                 fd = open(names[i], O_WRONLY);
1889                 if (fd < 0)
1890                         continue;
1891
1892                 if (write(fd, "-1", 2)) {}
1893                 close(fd);
1894         }
1895         return 0;
1896 }
1897
1898 bool system_resolve_rt_type(const char *type, unsigned int *id)
1899 {
1900         return system_rtn_aton(type, id);
1901 }
1902
1903 bool system_resolve_rt_table(const char *name, unsigned int *id)
1904 {
1905         FILE *f;
1906         char *e, buf[128];
1907         unsigned int n, table = RT_TABLE_UNSPEC;
1908
1909         /* first try to parse table as number */
1910         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
1911                 table = n;
1912
1913         /* handle well known aliases */
1914         else if (!strcmp(name, "default"))
1915                 table = RT_TABLE_DEFAULT;
1916         else if (!strcmp(name, "main"))
1917                 table = RT_TABLE_MAIN;
1918         else if (!strcmp(name, "local"))
1919                 table = RT_TABLE_LOCAL;
1920
1921         /* try to look up name in /etc/iproute2/rt_tables */
1922         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
1923         {
1924                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
1925                 {
1926                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
1927                                 continue;
1928
1929                         n = strtoul(e, NULL, 10);
1930                         e = strtok(NULL, " \t\n");
1931
1932                         if (e && !strcmp(e, name))
1933                         {
1934                                 table = n;
1935                                 break;
1936                         }
1937                 }
1938
1939                 fclose(f);
1940         }
1941
1942         if (table == RT_TABLE_UNSPEC)
1943                 return false;
1944
1945         *id = table;
1946         return true;
1947 }
1948
1949 bool system_is_default_rt_table(unsigned int id)
1950 {
1951         return (id == RT_TABLE_MAIN);
1952 }
1953
1954 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
1955 {
1956         char *e;
1957         unsigned int n;
1958
1959         if (!strcmp(filter, "strict"))
1960                 n = 1;
1961         else if (!strcmp(filter, "loose"))
1962                 n = 2;
1963         else {
1964                 n = strtoul(filter, &e, 0);
1965                 if (*e || e == filter || n > 2)
1966                         return false;
1967         }
1968
1969         *id = n;
1970         return true;
1971 }
1972
1973 static int system_iprule(struct iprule *rule, int cmd)
1974 {
1975         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
1976
1977         struct nl_msg *msg;
1978         struct rtmsg rtm = {
1979                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
1980                 .rtm_protocol = RTPROT_STATIC,
1981                 .rtm_scope = RT_SCOPE_UNIVERSE,
1982                 .rtm_table = RT_TABLE_UNSPEC,
1983                 .rtm_type = RTN_UNSPEC,
1984                 .rtm_flags = 0,
1985         };
1986
1987         if (cmd == RTM_NEWRULE)
1988                 rtm.rtm_type = RTN_UNICAST;
1989
1990         if (rule->invert)
1991                 rtm.rtm_flags |= FIB_RULE_INVERT;
1992
1993         if (rule->flags & IPRULE_SRC)
1994                 rtm.rtm_src_len = rule->src_mask;
1995
1996         if (rule->flags & IPRULE_DEST)
1997                 rtm.rtm_dst_len = rule->dest_mask;
1998
1999         if (rule->flags & IPRULE_TOS)
2000                 rtm.rtm_tos = rule->tos;
2001
2002         if (rule->flags & IPRULE_LOOKUP) {
2003                 if (rule->lookup < 256)
2004                         rtm.rtm_table = rule->lookup;
2005         }
2006
2007         if (rule->flags & IPRULE_ACTION)
2008                 rtm.rtm_type = rule->action;
2009         else if (rule->flags & IPRULE_GOTO)
2010                 rtm.rtm_type = FR_ACT_GOTO;
2011         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2012                 rtm.rtm_type = FR_ACT_NOP;
2013
2014         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2015
2016         if (!msg)
2017                 return -1;
2018
2019         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2020
2021         if (rule->flags & IPRULE_IN)
2022                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2023
2024         if (rule->flags & IPRULE_OUT)
2025                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2026
2027         if (rule->flags & IPRULE_SRC)
2028                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2029
2030         if (rule->flags & IPRULE_DEST)
2031                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2032
2033         if (rule->flags & IPRULE_PRIORITY)
2034                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2035         else if (cmd == RTM_NEWRULE)
2036                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2037
2038         if (rule->flags & IPRULE_FWMARK)
2039                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2040
2041         if (rule->flags & IPRULE_FWMASK)
2042                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2043
2044         if (rule->flags & IPRULE_LOOKUP) {
2045                 if (rule->lookup >= 256)
2046                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2047         }
2048
2049         if (rule->flags & IPRULE_GOTO)
2050                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2051
2052         return system_rtnl_call(msg);
2053 }
2054
2055 int system_add_iprule(struct iprule *rule)
2056 {
2057         return system_iprule(rule, RTM_NEWRULE);
2058 }
2059
2060 int system_del_iprule(struct iprule *rule)
2061 {
2062         return system_iprule(rule, RTM_DELRULE);
2063 }
2064
2065 int system_flush_iprules(void)
2066 {
2067         int rv = 0;
2068         struct iprule rule;
2069
2070         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2071         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2072
2073         memset(&rule, 0, sizeof(rule));
2074
2075
2076         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2077
2078         rule.priority = 0;
2079         rule.lookup = RT_TABLE_LOCAL;
2080         rv |= system_iprule(&rule, RTM_NEWRULE);
2081
2082         rule.priority = 32766;
2083         rule.lookup = RT_TABLE_MAIN;
2084         rv |= system_iprule(&rule, RTM_NEWRULE);
2085
2086         rule.priority = 32767;
2087         rule.lookup = RT_TABLE_DEFAULT;
2088         rv |= system_iprule(&rule, RTM_NEWRULE);
2089
2090
2091         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2092
2093         rule.priority = 0;
2094         rule.lookup = RT_TABLE_LOCAL;
2095         rv |= system_iprule(&rule, RTM_NEWRULE);
2096
2097         rule.priority = 32766;
2098         rule.lookup = RT_TABLE_MAIN;
2099         rv |= system_iprule(&rule, RTM_NEWRULE);
2100
2101         return rv;
2102 }
2103
2104 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2105 {
2106         return system_rtn_aton(action, id);
2107 }
2108
2109 time_t system_get_rtime(void)
2110 {
2111         struct timespec ts;
2112         struct timeval tv;
2113
2114         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
2115                 return ts.tv_sec;
2116
2117         if (gettimeofday(&tv, NULL) == 0)
2118                 return tv.tv_sec;
2119
2120         return 0;
2121 }
2122
2123 #ifndef IP_DF
2124 #define IP_DF       0x4000
2125 #endif
2126
2127 static int tunnel_ioctl(const char *name, int cmd, void *p)
2128 {
2129         struct ifreq ifr;
2130
2131         memset(&ifr, 0, sizeof(ifr));
2132         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name));
2133         ifr.ifr_ifru.ifru_data = p;
2134         return ioctl(sock_ioctl, cmd, &ifr);
2135 }
2136
2137 #ifdef IFLA_IPTUN_MAX
2138 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2139 static int system_add_gre_tunnel(const char *name, const char *kind,
2140                                  const unsigned int link, struct blob_attr **tb, bool v6)
2141 {
2142         struct nl_msg *nlm;
2143         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2144         struct blob_attr *cur;
2145         uint32_t ikey = 0, okey = 0, flags = 0, flowinfo = 0;
2146         uint16_t iflags = 0, oflags = 0;
2147         uint8_t tos = 0;
2148         int ret = 0, ttl = 0;
2149
2150         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2151         if (!nlm)
2152                 return -1;
2153
2154         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2155         nla_put_string(nlm, IFLA_IFNAME, name);
2156
2157         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2158         if (!linkinfo) {
2159                 ret = -ENOMEM;
2160                 goto failure;
2161         }
2162
2163         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2164         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2165         if (!infodata) {
2166                 ret = -ENOMEM;
2167                 goto failure;
2168         }
2169
2170         if (link)
2171                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2172
2173         if ((cur = tb[TUNNEL_ATTR_TTL]))
2174                 ttl = blobmsg_get_u32(cur);
2175
2176         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2177                 char *str = blobmsg_get_string(cur);
2178                 if (strcmp(str, "inherit")) {
2179                         unsigned uval;
2180
2181                         if (!system_tos_aton(str, &uval)) {
2182                                 ret = -EINVAL;
2183                                 goto failure;
2184                         }
2185
2186                         if (v6)
2187                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2188                         else
2189                                 tos = uval;
2190                 } else {
2191                         if (v6)
2192                                 flags |= IP6_TNL_F_USE_ORIG_TCLASS;
2193                         else
2194                                 tos = 1;
2195                 }
2196         }
2197
2198         if ((cur = tb[TUNNEL_ATTR_INFO]) && (blobmsg_type(cur) == BLOBMSG_TYPE_STRING)) {
2199                 uint8_t icsum, ocsum, iseqno, oseqno;
2200                 if (sscanf(blobmsg_get_string(cur), "%u,%u,%hhu,%hhu,%hhu,%hhu",
2201                         &ikey, &okey, &icsum, &ocsum, &iseqno, &oseqno) < 6) {
2202                         ret = -EINVAL;
2203                         goto failure;
2204                 }
2205
2206                 if (ikey)
2207                         iflags |= GRE_KEY;
2208
2209                 if (okey)
2210                         oflags |= GRE_KEY;
2211
2212                 if (icsum)
2213                         iflags |= GRE_CSUM;
2214
2215                 if (ocsum)
2216                         oflags |= GRE_CSUM;
2217
2218                 if (iseqno)
2219                         iflags |= GRE_SEQ;
2220
2221                 if (oseqno)
2222                         oflags |= GRE_SEQ;
2223         }
2224
2225         if (v6) {
2226                 struct in6_addr in6buf;
2227                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2228                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2229                                 ret = -EINVAL;
2230                                 goto failure;
2231                         }
2232                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2233                 }
2234
2235                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2236                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2237                                 ret = -EINVAL;
2238                                 goto failure;
2239                         }
2240                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2241                 }
2242                 nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, 4);
2243
2244                 if (flowinfo)
2245                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2246
2247                 if (flags)
2248                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags);
2249
2250                 if (!ttl)
2251                         ttl = 64;
2252         } else {
2253                 struct in_addr inbuf;
2254                 bool set_df = true;
2255
2256                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2257                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2258                                 ret = -EINVAL;
2259                                 goto failure;
2260                         }
2261                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2262                 }
2263
2264                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2265                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2266                                 ret = -EINVAL;
2267                                 goto failure;
2268                         }
2269                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2270
2271                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2272                                 if (!okey) {
2273                                         okey = inbuf.s_addr;
2274                                         oflags |= GRE_KEY;
2275                                 }
2276
2277                                 if (!ikey) {
2278                                         ikey = inbuf.s_addr;
2279                                         iflags |= GRE_KEY;
2280                                 }
2281                         }
2282                 }
2283
2284                 if ((cur = tb[TUNNEL_ATTR_DF]))
2285                         set_df = blobmsg_get_bool(cur);
2286
2287                 if (!set_df) {
2288                         /* ttl != 0 and nopmtudisc are incompatible */
2289                         if (ttl) {
2290                                 ret = -EINVAL;
2291                                 goto failure;
2292                         }
2293                 } else if (!ttl)
2294                         ttl = 64;
2295
2296                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2297
2298                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2299         }
2300
2301         if (ttl)
2302                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2303
2304         if (oflags)
2305                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2306
2307         if (iflags)
2308                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2309
2310         if (okey)
2311                 nla_put_u32(nlm, IFLA_GRE_OKEY, okey);
2312
2313         if (ikey)
2314                 nla_put_u32(nlm, IFLA_GRE_IKEY, ikey);
2315
2316         nla_nest_end(nlm, infodata);
2317         nla_nest_end(nlm, linkinfo);
2318
2319         return system_rtnl_call(nlm);
2320
2321 failure:
2322         nlmsg_free(nlm);
2323         return ret;
2324 }
2325 #endif
2326
2327 #ifdef IFLA_VTI_MAX
2328 static int system_add_vti_tunnel(const char *name, const char *kind,
2329                                  const unsigned int link, struct blob_attr **tb, bool v6)
2330 {
2331         struct nl_msg *nlm;
2332         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2333         struct blob_attr *cur;
2334         uint32_t ikey = 0, okey = 0;
2335         int ret = 0;
2336
2337         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2338         if (!nlm)
2339                 return -1;
2340
2341         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2342         nla_put_string(nlm, IFLA_IFNAME, name);
2343
2344         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2345         if (!linkinfo) {
2346                 ret = -ENOMEM;
2347                 goto failure;
2348         }
2349
2350         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2351         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2352         if (!infodata) {
2353                 ret = -ENOMEM;
2354                 goto failure;
2355         }
2356
2357         if (link)
2358                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2359
2360         if ((cur = tb[TUNNEL_ATTR_INFO]) && (blobmsg_type(cur) == BLOBMSG_TYPE_STRING)) {
2361                 if (sscanf(blobmsg_get_string(cur), "%u,%u",
2362                         &ikey, &okey) < 2) {
2363                         ret = -EINVAL;
2364                         goto failure;
2365                 }
2366         }
2367
2368         if (v6) {
2369                 struct in6_addr in6buf;
2370                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2371                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2372                                 ret = -EINVAL;
2373                                 goto failure;
2374                         }
2375                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2376                 }
2377
2378                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2379                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2380                                 ret = -EINVAL;
2381                                 goto failure;
2382                         }
2383                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2384                 }
2385
2386         } else {
2387                 struct in_addr inbuf;
2388
2389                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2390                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2391                                 ret = -EINVAL;
2392                                 goto failure;
2393                         }
2394                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2395                 }
2396
2397                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2398                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2399                                 ret = -EINVAL;
2400                                 goto failure;
2401                         }
2402                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2403                 }
2404
2405         }
2406
2407         if (okey)
2408                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2409
2410         if (ikey)
2411                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2412
2413         nla_nest_end(nlm, infodata);
2414         nla_nest_end(nlm, linkinfo);
2415
2416         return system_rtnl_call(nlm);
2417
2418 failure:
2419         nlmsg_free(nlm);
2420         return ret;
2421 }
2422 #endif
2423
2424 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
2425 {
2426         struct blob_attr *cur;
2427         bool set_df = true;
2428         struct ip_tunnel_parm p  = {
2429                 .link = link,
2430                 .iph = {
2431                         .version = 4,
2432                         .ihl = 5,
2433                         .protocol = proto,
2434                 }
2435         };
2436
2437         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
2438                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
2439                 return -EINVAL;
2440
2441         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
2442                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
2443                 return -EINVAL;
2444
2445         if ((cur = tb[TUNNEL_ATTR_DF]))
2446                 set_df = blobmsg_get_bool(cur);
2447
2448         if ((cur = tb[TUNNEL_ATTR_TTL]))
2449                 p.iph.ttl = blobmsg_get_u32(cur);
2450
2451         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2452                 char *str = blobmsg_get_string(cur);
2453                 if (strcmp(str, "inherit")) {
2454                         unsigned uval;
2455
2456                         if (!system_tos_aton(str, &uval))
2457                                 return -EINVAL;
2458
2459                         p.iph.tos = uval;
2460                 } else
2461                         p.iph.tos = 1;
2462         }
2463
2464         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
2465         /* ttl !=0 and nopmtudisc are incompatible */
2466         if (p.iph.ttl && p.iph.frag_off == 0)
2467                 return -EINVAL;
2468
2469         strncpy(p.name, name, sizeof(p.name));
2470
2471         switch (p.iph.protocol) {
2472         case IPPROTO_IPIP:
2473                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
2474         case IPPROTO_IPV6:
2475                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
2476         default:
2477                 break;
2478         }
2479         return -1;
2480 }
2481
2482 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
2483 {
2484         struct blob_attr *cur;
2485         const char *str;
2486
2487         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2488                 return -EINVAL;
2489         str = blobmsg_data(cur);
2490
2491         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
2492             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
2493             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6"))
2494                 return system_link_del(name);
2495         else
2496                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
2497 }
2498
2499 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
2500 {
2501         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2502
2503         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2504                 blob_data(attr), blob_len(attr));
2505
2506         return __system_del_ip_tunnel(name, tb);
2507 }
2508
2509 int system_update_ipv6_mtu(struct device *dev, int mtu)
2510 {
2511         int ret = -1;
2512         char buf[64];
2513         int fd;
2514
2515         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
2516                         dev->ifname);
2517
2518         fd = open(buf, O_RDWR);
2519         if (fd < 0)
2520                 return ret;
2521
2522         if (!mtu) {
2523                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
2524                 if (len < 0)
2525                         goto out;
2526
2527                 buf[len] = 0;
2528                 ret = atoi(buf);
2529         } else {
2530                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
2531                         ret = mtu;
2532         }
2533
2534 out:
2535         close(fd);
2536         return ret;
2537 }
2538
2539 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
2540 {
2541         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2542         struct blob_attr *cur;
2543         const char *str;
2544
2545         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2546                 blob_data(attr), blob_len(attr));
2547
2548         __system_del_ip_tunnel(name, tb);
2549
2550         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2551                 return -EINVAL;
2552         str = blobmsg_data(cur);
2553
2554         unsigned int ttl = 0;
2555         if ((cur = tb[TUNNEL_ATTR_TTL])) {
2556                 ttl = blobmsg_get_u32(cur);
2557                 if (ttl > 255)
2558                         return -EINVAL;
2559         }
2560
2561         unsigned int link = 0;
2562         if ((cur = tb[TUNNEL_ATTR_LINK])) {
2563                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
2564                 if (!iface)
2565                         return -EINVAL;
2566
2567                 if (iface->l3_dev.dev)
2568                         link = iface->l3_dev.dev->ifindex;
2569         }
2570
2571         if (!strcmp(str, "sit")) {
2572                 if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
2573                         return -1;
2574
2575 #ifdef SIOCADD6RD
2576                 if ((cur = tb[TUNNEL_ATTR_6RD_PREFIX])) {
2577                         unsigned int mask;
2578                         struct ip_tunnel_6rd p6;
2579
2580                         memset(&p6, 0, sizeof(p6));
2581
2582                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
2583                                                 &p6.prefix, &mask) || mask > 128)
2584                                 return -EINVAL;
2585                         p6.prefixlen = mask;
2586
2587                         if ((cur = tb[TUNNEL_ATTR_6RD_RELAY_PREFIX])) {
2588                                 if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
2589                                                         &p6.relay_prefix, &mask) || mask > 32)
2590                                         return -EINVAL;
2591                                 p6.relay_prefixlen = mask;
2592                         }
2593
2594                         if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
2595                                 __system_del_ip_tunnel(name, tb);
2596                                 return -1;
2597                         }
2598                 }
2599 #endif
2600 #ifdef IFLA_IPTUN_MAX
2601         } else if (!strcmp(str, "ipip6")) {
2602                 struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2603                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2604                 struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2605                 int ret = 0;
2606
2607                 if (!nlm)
2608                         return -1;
2609
2610                 nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2611                 nla_put_string(nlm, IFLA_IFNAME, name);
2612
2613                 if (link)
2614                         nla_put_u32(nlm, IFLA_LINK, link);
2615
2616                 struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2617                 if (!linkinfo) {
2618                         ret = -ENOMEM;
2619                         goto failure;
2620                 }
2621                 nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2622                 struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2623                 if (!infodata) {
2624                         ret = -ENOMEM;
2625                         goto failure;
2626                 }
2627
2628                 if (link)
2629                         nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2630
2631                 nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2632                 nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2633                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, 4);
2634
2635                 struct in6_addr in6buf;
2636                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2637                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2638                                 ret = -EINVAL;
2639                                 goto failure;
2640                         }
2641                         nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2642                 }
2643
2644                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2645                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2646                                 ret = -EINVAL;
2647                                 goto failure;
2648                         }
2649                         nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2650                 }
2651
2652 #ifdef IFLA_IPTUN_FMR_MAX
2653                 if ((cur = tb[TUNNEL_ATTR_FMRS])) {
2654                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2655
2656                         struct blob_attr *fmr;
2657                         unsigned rem, fmrcnt = 0;
2658                         blobmsg_for_each_attr(fmr, cur, rem) {
2659                                 if (blobmsg_type(fmr) != BLOBMSG_TYPE_STRING)
2660                                         continue;
2661
2662                                 unsigned ip4len, ip6len, ealen, offset = 6;
2663                                 char ip6buf[48];
2664                                 char ip4buf[16];
2665
2666                                 if (sscanf(blobmsg_get_string(fmr), "%47[^/]/%u,%15[^/]/%u,%u,%u",
2667                                                 ip6buf, &ip6len, ip4buf, &ip4len, &ealen, &offset) < 5) {
2668                                         ret = -EINVAL;
2669                                         goto failure;
2670                                 }
2671
2672                                 struct in6_addr ip6prefix;
2673                                 struct in_addr ip4prefix;
2674                                 if (inet_pton(AF_INET6, ip6buf, &ip6prefix) != 1 ||
2675                                                 inet_pton(AF_INET, ip4buf, &ip4prefix) != 1) {
2676                                         ret = -EINVAL;
2677                                         goto failure;
2678                                 }
2679
2680                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2681
2682                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2683                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2684                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2685                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2686                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2687                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2688
2689                                 nla_nest_end(nlm, rule);
2690                         }
2691
2692                         nla_nest_end(nlm, fmrs);
2693                 }
2694 #endif
2695
2696                 nla_nest_end(nlm, infodata);
2697                 nla_nest_end(nlm, linkinfo);
2698
2699                 return system_rtnl_call(nlm);
2700 failure:
2701                 nlmsg_free(nlm);
2702                 return ret;
2703         } else if (!strcmp(str, "greip")) {
2704                 return system_add_gre_tunnel(name, "gre", link, tb, false);
2705         } else if (!strcmp(str, "gretapip"))  {
2706                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
2707         } else if (!strcmp(str, "greip6")) {
2708                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
2709         } else if (!strcmp(str, "gretapip6")) {
2710                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
2711 #ifdef IFLA_VTI_MAX
2712         } else if (!strcmp(str, "vtiip")) {
2713                 return system_add_vti_tunnel(name, "vti", link, tb, false);
2714         } else if (!strcmp(str, "vtiip6")) {
2715                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
2716 #endif
2717 #endif
2718         } else if (!strcmp(str, "ipip")) {
2719                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
2720         }
2721         else
2722                 return -EINVAL;
2723
2724         return 0;
2725 }