system-linux.c

   1 /*
   2  * netifd - network interface daemon
   3  * Copyright (C) 2012 Felix Fietkau <nbd@openwrt.org>
   4  * Copyright (C) 2013 Jo-Philipp Wich <jow@openwrt.org>
   5  * Copyright (C) 2013 Steven Barth <steven@midlink.org>
   6  * Copyright (C) 2014 Gioacchino Mazzurco <gio@eigenlab.org>
   7  * Copyright (C) 2017 Matthias Schiffer <mschiffer@universe-factory.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2
  11  * as published by the Free Software Foundation
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  */
  18 #define _GNU_SOURCE
  19
  20 #include <sys/socket.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/stat.h>
  23 #include <sys/syscall.h>
  24
  25 #include <net/if.h>
  26 #include <net/if_arp.h>
  27
  28 #include <arpa/inet.h>
  29 #include <netinet/ether.h>
  30 #include <netinet/in.h>
  31
  32 #include <linux/rtnetlink.h>
  33 #include <linux/sockios.h>
  34 #include <linux/ip.h>
  35 #include <linux/if_addr.h>
  36 #include <linux/if_link.h>
  37 #include <linux/if_vlan.h>
  38 #include <linux/if_bridge.h>
  39 #include <linux/if_tunnel.h>
  40 #include <linux/ip6_tunnel.h>
  41 #include <linux/ethtool.h>
  42 #include <linux/fib_rules.h>
  43 #include <linux/veth.h>
  44 #include <linux/version.h>
  45
  46 #ifndef RTN_FAILED_POLICY
  47 #define RTN_FAILED_POLICY 12
  48 #endif
  49
  50 #ifndef IFA_F_NOPREFIXROUTE
  51 #define IFA_F_NOPREFIXROUTE 0x200
  52 #endif
  53
  54 #ifndef IFA_FLAGS
  55 #define IFA_FLAGS (IFA_MULTICAST + 1)
  56 #endif
  57
  58 #include <string.h>
  59 #include <fcntl.h>
  60 #include <glob.h>
  61 #include <time.h>
  62 #include <unistd.h>
  63
  64 #include <netlink/msg.h>
  65 #include <netlink/attr.h>
  66 #include <netlink/socket.h>
  67 #include <libubox/uloop.h>
  68
  69 #include "netifd.h"
  70 #include "device.h"
  71 #include "system.h"
  72
  73 struct event_socket {
  74         struct uloop_fd uloop;
  75         struct nl_sock *sock;
  76         int bufsize;
  77 };
  78
  79 static int sock_ioctl = -1;
  80 static struct nl_sock *sock_rtnl = NULL;
  81
  82 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  83 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  84 static int system_add_proto_tunnel(const char *name, const uint8_t proto,
  85                                         const unsigned int link, struct blob_attr **tb);
  86 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb);
  87
  88 static char dev_buf[256];
  89
  90 static void
  91 handler_nl_event(struct uloop_fd *u, unsigned int events)
  92 {
  93         struct event_socket *ev = container_of(u, struct event_socket, uloop);
  94         int err;
  95         socklen_t errlen = sizeof(err);
  96
  97         if (!u->error) {
  98                 nl_recvmsgs_default(ev->sock);
  99                 return;
 100         }
 101
 102         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 103                 goto abort;
 104
 105         switch(err) {
 106         case ENOBUFS:
 107                 // Increase rx buffer size on netlink socket
 108                 ev->bufsize *= 2;
 109                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 110                         goto abort;
 111
 112                 // Request full dump since some info got dropped
 113                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 114                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 115                 break;
 116
 117         default:
 118                 goto abort;
 119         }
 120         u->error = false;
 121         return;
 122
 123 abort:
 124         uloop_fd_delete(&ev->uloop);
 125         return;
 126 }
 127
 128 static struct nl_sock *
 129 create_socket(int protocol, int groups)
 130 {
 131         struct nl_sock *sock;
 132
 133         sock = nl_socket_alloc();
 134         if (!sock)
 135                 return NULL;
 136
 137         if (groups)
 138                 nl_join_groups(sock, groups);
 139
 140         if (nl_connect(sock, protocol))
 141                 return NULL;
 142
 143         return sock;
 144 }
 145
 146 static bool
 147 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 148                         uloop_fd_handler cb, int flags)
 149 {
 150         ev->sock = create_socket(protocol, groups);
 151         if (!ev->sock)
 152                 return false;
 153
 154         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 155         ev->uloop.cb = cb;
 156         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 157                 return false;
 158
 159         return true;
 160 }
 161
 162 static bool
 163 create_event_socket(struct event_socket *ev, int protocol,
 164                     int (*cb)(struct nl_msg *msg, void *arg))
 165 {
 166         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 167                 return false;
 168
 169         // Install the valid custom callback handler
 170         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 171
 172         // Disable sequence number checking on event sockets
 173         nl_socket_disable_seq_check(ev->sock);
 174
 175         // Increase rx buffer size to 65K on event sockets
 176         ev->bufsize = 65535;
 177         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 178                 return false;
 179
 180         return true;
 181 }
 182
 183 static bool
 184 system_rtn_aton(const char *src, unsigned int *dst)
 185 {
 186         char *e;
 187         unsigned int n;
 188
 189         if (!strcmp(src, "local"))
 190                 n = RTN_LOCAL;
 191         else if (!strcmp(src, "nat"))
 192                 n = RTN_NAT;
 193         else if (!strcmp(src, "broadcast"))
 194                 n = RTN_BROADCAST;
 195         else if (!strcmp(src, "anycast"))
 196                 n = RTN_ANYCAST;
 197         else if (!strcmp(src, "multicast"))
 198                 n = RTN_MULTICAST;
 199         else if (!strcmp(src, "prohibit"))
 200                 n = RTN_PROHIBIT;
 201         else if (!strcmp(src, "unreachable"))
 202                 n = RTN_UNREACHABLE;
 203         else if (!strcmp(src, "blackhole"))
 204                 n = RTN_BLACKHOLE;
 205         else if (!strcmp(src, "xresolve"))
 206                 n = RTN_XRESOLVE;
 207         else if (!strcmp(src, "unicast"))
 208                 n = RTN_UNICAST;
 209         else if (!strcmp(src, "throw"))
 210                 n = RTN_THROW;
 211         else if (!strcmp(src, "failed_policy"))
 212                 n = RTN_FAILED_POLICY;
 213         else {
 214                 n = strtoul(src, &e, 0);
 215                 if (!e || *e || e == src || n > 255)
 216                         return false;
 217         }
 218
 219         *dst = n;
 220         return true;
 221 }
 222
 223 static bool
 224 system_tos_aton(const char *src, unsigned *dst)
 225 {
 226         char *e;
 227
 228         *dst = strtoul(src, &e, 16);
 229         if (e == src || *e || *dst > 255)
 230                 return false;
 231
 232         return true;
 233 }
 234
 235 int system_init(void)
 236 {
 237         static struct event_socket rtnl_event;
 238         static struct event_socket hotplug_event;
 239
 240         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 241         system_fd_set_cloexec(sock_ioctl);
 242
 243         // Prepare socket for routing / address control
 244         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 245         if (!sock_rtnl)
 246                 return -1;
 247
 248         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 249                 return -1;
 250
 251         if (!create_raw_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT, 1,
 252                                         handle_hotplug_event, 0))
 253                 return -1;
 254
 255         // Receive network link events form kernel
 256         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 257
 258         return 0;
 259 }
 260
 261 static void system_set_sysctl(const char *path, const char *val)
 262 {
 263         int fd;
 264
 265         fd = open(path, O_WRONLY);
 266         if (fd < 0)
 267                 return;
 268
 269         if (write(fd, val, strlen(val))) {}
 270         close(fd);
 271 }
 272
 273 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 274 {
 275         snprintf(dev_buf, sizeof(dev_buf), path, device);
 276         system_set_sysctl(dev_buf, val);
 277 }
 278
 279 static void system_set_disable_ipv6(struct device *dev, const char *val)
 280 {
 281         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 282 }
 283
 284 static void system_set_rpfilter(struct device *dev, const char *val)
 285 {
 286         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter", dev->ifname, val);
 287 }
 288
 289 static void system_set_acceptlocal(struct device *dev, const char *val)
 290 {
 291         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local", dev->ifname, val);
 292 }
 293
 294 static void system_set_igmpversion(struct device *dev, const char *val)
 295 {
 296         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version", dev->ifname, val);
 297 }
 298
 299 static void system_set_mldversion(struct device *dev, const char *val)
 300 {
 301         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version", dev->ifname, val);
 302 }
 303
 304 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 305 {
 306         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 307 }
 308
 309 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 310 {
 311         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 312 }
 313
 314 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 315 {
 316         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time", dev->ifname, val);
 317 }
 318
 319 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 320 {
 321         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time", dev->ifname, val);
 322 }
 323
 324 static void system_set_neigh4locktime(struct device *dev, const char *val)
 325 {
 326         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime", dev->ifname, val);
 327 }
 328
 329 static void system_set_dadtransmits(struct device *dev, const char *val)
 330 {
 331         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 332 }
 333
 334 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 335 {
 336         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 337 }
 338
 339 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 340 {
 341         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 342 }
 343
 344 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 345 {
 346         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 347 }
 348
 349 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 350 {
 351         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 352                                        "/sys/class/net/%s/brport/multicast_router",
 353                               dev->ifname, val);
 354 }
 355
 356 static void system_bridge_set_robustness(struct device *dev, const char *val)
 357 {
 358         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 359                               dev->ifname, val);
 360         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 361                               dev->ifname, val);
 362 }
 363
 364 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 365 {
 366         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 367                               dev->ifname, val);
 368 }
 369
 370 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 371 {
 372         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 373                               dev->ifname, val);
 374 }
 375
 376 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 377 {
 378         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 379                               dev->ifname, val);
 380 }
 381
 382 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 383 {
 384         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 385                               dev->ifname, val);
 386 }
 387
 388 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 389 {
 390         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 391                               dev->ifname, val);
 392 }
 393
 394 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 395 {
 396         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 397                               dev->ifname, val);
 398 }
 399
 400 static void system_bridge_set_learning(struct device *dev, const char *val)
 401 {
 402         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 403 }
 404
 405 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 406 {
 407         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 408 }
 409
 410 static void system_set_sendredirects(struct device *dev, const char *val)
 411 {
 412         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 413 }
 414
 415 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 416 {
 417         int fd = -1, ret = -1;
 418
 419         fd = open(path, O_RDONLY);
 420         if (fd < 0)
 421                 goto out;
 422
 423         ssize_t len = read(fd, buf, buf_sz - 1);
 424         if (len < 0)
 425                 goto out;
 426
 427         ret = buf[len] = 0;
 428
 429 out:
 430         if (fd >= 0)
 431                 close(fd);
 432
 433         return ret;
 434 }
 435
 436 static int
 437 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 438 {
 439         snprintf(dev_buf, sizeof(dev_buf), path, device);
 440         return system_get_sysctl(dev_buf, buf, buf_sz);
 441 }
 442
 443 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 444 {
 445         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 446                         dev->ifname, buf, buf_sz);
 447 }
 448
 449 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 450 {
 451         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 452                         dev->ifname, buf, buf_sz);
 453 }
 454
 455 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 456 {
 457         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 458                         dev->ifname, buf, buf_sz);
 459 }
 460
 461 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 462 {
 463         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 464                         dev->ifname, buf, buf_sz);
 465 }
 466
 467 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 468 {
 469         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 470                         dev->ifname, buf, buf_sz);
 471 }
 472
 473 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 474 {
 475         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 476                         dev->ifname, buf, buf_sz);
 477 }
 478
 479 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 480 {
 481         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 482                         dev->ifname, buf, buf_sz);
 483 }
 484
 485 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 486 {
 487         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 488                         dev->ifname, buf, buf_sz);
 489 }
 490
 491 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 492 {
 493         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 494                         dev->ifname, buf, buf_sz);
 495 }
 496
 497 static int system_get_neigh4locktime(struct device *dev, char *buf, const size_t buf_sz)
 498 {
 499         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime",
 500                         dev->ifname, buf, buf_sz);
 501 }
 502
 503 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 504 {
 505         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 506                         dev->ifname, buf, buf_sz);
 507 }
 508
 509 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 510 {
 511         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 512                         dev->ifname, buf, buf_sz);
 513 }
 514
 515 // Evaluate netlink messages
 516 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 517 {
 518         struct nlmsghdr *nh = nlmsg_hdr(msg);
 519         struct nlattr *nla[__IFLA_MAX];
 520         int link_state = 0;
 521         char buf[10];
 522
 523         if (nh->nlmsg_type != RTM_NEWLINK)
 524                 goto out;
 525
 526         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 527         if (!nla[IFLA_IFNAME])
 528                 goto out;
 529
 530         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 531         if (!dev)
 532                 goto out;
 533
 534         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 535                 link_state = strtoul(buf, NULL, 0);
 536
 537         device_set_link(dev, link_state ? true : false);
 538
 539 out:
 540         return 0;
 541 }
 542
 543 static void
 544 handle_hotplug_msg(char *data, int size)
 545 {
 546         const char *subsystem = NULL, *interface = NULL;
 547         char *cur, *end, *sep;
 548         struct device *dev;
 549         int skip;
 550         bool add;
 551
 552         if (!strncmp(data, "add@", 4))
 553                 add = true;
 554         else if (!strncmp(data, "remove@", 7))
 555                 add = false;
 556         else
 557                 return;
 558
 559         skip = strlen(data) + 1;
 560         end = data + size;
 561
 562         for (cur = data + skip; cur < end; cur += skip) {
 563                 skip = strlen(cur) + 1;
 564
 565                 sep = strchr(cur, '=');
 566                 if (!sep)
 567                         continue;
 568
 569                 *sep = 0;
 570                 if (!strcmp(cur, "INTERFACE"))
 571                         interface = sep + 1;
 572                 else if (!strcmp(cur, "SUBSYSTEM")) {
 573                         subsystem = sep + 1;
 574                         if (strcmp(subsystem, "net") != 0)
 575                                 return;
 576                 }
 577                 if (subsystem && interface)
 578                         goto found;
 579         }
 580         return;
 581
 582 found:
 583         dev = device_find(interface);
 584         if (!dev)
 585                 return;
 586
 587         if (dev->type != &simple_device_type)
 588                 return;
 589
 590         if (add && system_if_force_external(dev->ifname))
 591                 return;
 592
 593         device_set_present(dev, add);
 594 }
 595
 596 static void
 597 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 598 {
 599         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 600         struct sockaddr_nl nla;
 601         unsigned char *buf = NULL;
 602         int size;
 603
 604         while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 605                 if (nla.nl_pid == 0)
 606                         handle_hotplug_msg((char *) buf, size);
 607
 608                 free(buf);
 609         }
 610 }
 611
 612 static int system_rtnl_call(struct nl_msg *msg)
 613 {
 614         int ret;
 615
 616         ret = nl_send_auto_complete(sock_rtnl, msg);
 617         nlmsg_free(msg);
 618
 619         if (ret < 0)
 620                 return ret;
 621
 622         return nl_wait_for_ack(sock_rtnl);
 623 }
 624
 625 int system_bridge_delbr(struct device *bridge)
 626 {
 627         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 628 }
 629
 630 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 631 {
 632         struct ifreq ifr;
 633
 634         memset(&ifr, 0, sizeof(ifr));
 635         if (dev)
 636                 ifr.ifr_ifindex = dev->ifindex;
 637         else
 638                 ifr.ifr_data = data;
 639         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name));
 640         return ioctl(sock_ioctl, cmd, &ifr);
 641 }
 642
 643 static bool system_is_bridge(const char *name, char *buf, int buflen)
 644 {
 645         struct stat st;
 646
 647         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 648         if (stat(buf, &st) < 0)
 649                 return false;
 650
 651         return true;
 652 }
 653
 654 static char *system_get_bridge(const char *name, char *buf, int buflen)
 655 {
 656         char *path;
 657         ssize_t len = -1;
 658         glob_t gl;
 659
 660         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 661         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 662                 return NULL;
 663
 664         if (gl.gl_pathc > 0)
 665                 len = readlink(gl.gl_pathv[0], buf, buflen);
 666
 667         globfree(&gl);
 668
 669         if (len < 0)
 670                 return NULL;
 671
 672         buf[len] = 0;
 673         path = strrchr(buf, '/');
 674         if (!path)
 675                 return NULL;
 676
 677         return path + 1;
 678 }
 679
 680 static void
 681 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 682 {
 683         bool mcast_to_ucast = dev->wireless_ap;
 684         bool hairpin = true;
 685
 686         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 687             !bridge->settings.multicast_to_unicast)
 688                 mcast_to_ucast = false;
 689
 690         if (!mcast_to_ucast || dev->wireless_isolate)
 691                 hairpin = false;
 692
 693         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 694         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 695 }
 696
 697 int system_bridge_addif(struct device *bridge, struct device *dev)
 698 {
 699         char buf[64];
 700         char *oldbr;
 701         int ret = 0;
 702
 703         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 704         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 705                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 706
 707         if (dev->wireless)
 708                 system_bridge_set_wireless(bridge, dev);
 709
 710         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 711                 snprintf(buf, sizeof(buf), "%u", dev->settings.multicast_router);
 712                 system_bridge_set_multicast_router(dev, buf, false);
 713         }
 714
 715         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 716             dev->settings.multicast_fast_leave)
 717                 system_bridge_set_multicast_fast_leave(dev, "1");
 718
 719         if (dev->settings.flags & DEV_OPT_LEARNING &&
 720             !dev->settings.learning)
 721                 system_bridge_set_learning(dev, "0");
 722
 723         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 724             !dev->settings.unicast_flood)
 725                 system_bridge_set_unicast_flood(dev, "0");
 726
 727         return ret;
 728 }
 729
 730 int system_bridge_delif(struct device *bridge, struct device *dev)
 731 {
 732         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 733 }
 734
 735 int system_if_resolve(struct device *dev)
 736 {
 737         struct ifreq ifr;
 738         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
 739         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 740                 return ifr.ifr_ifindex;
 741         else
 742                 return 0;
 743 }
 744
 745 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 746 {
 747         struct ifreq ifr;
 748
 749         memset(&ifr, 0, sizeof(ifr));
 750         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name));
 751         ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr);
 752         ifr.ifr_flags |= add;
 753         ifr.ifr_flags &= ~rem;
 754         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 755 }
 756
 757 struct clear_data {
 758         struct nl_msg *msg;
 759         struct device *dev;
 760         int type;
 761         int size;
 762         int af;
 763 };
 764
 765
 766 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 767 {
 768         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 769
 770         return ifa->ifa_index == ifindex;
 771 }
 772
 773 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 774 {
 775         struct rtmsg *r = NLMSG_DATA(hdr);
 776         struct nlattr *tb[__RTA_MAX];
 777
 778         if (r->rtm_protocol == RTPROT_KERNEL &&
 779             r->rtm_family == AF_INET6)
 780                 return false;
 781
 782         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 783         if (!tb[RTA_OIF])
 784                 return false;
 785
 786         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 787 }
 788
 789 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 790 {
 791         return true;
 792 }
 793
 794 static int cb_clear_event(struct nl_msg *msg, void *arg)
 795 {
 796         struct clear_data *clr = arg;
 797         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 798         bool (*cb)(struct nlmsghdr *, int ifindex);
 799         int type;
 800
 801         switch(clr->type) {
 802         case RTM_GETADDR:
 803                 type = RTM_DELADDR;
 804                 if (hdr->nlmsg_type != RTM_NEWADDR)
 805                         return NL_SKIP;
 806
 807                 cb = check_ifaddr;
 808                 break;
 809         case RTM_GETROUTE:
 810                 type = RTM_DELROUTE;
 811                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 812                         return NL_SKIP;
 813
 814                 cb = check_route;
 815                 break;
 816         case RTM_GETRULE:
 817                 type = RTM_DELRULE;
 818                 if (hdr->nlmsg_type != RTM_NEWRULE)
 819                         return NL_SKIP;
 820
 821                 cb = check_rule;
 822                 break;
 823         default:
 824                 return NL_SKIP;
 825         }
 826
 827         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 828                 return NL_SKIP;
 829
 830         if (type == RTM_DELRULE)
 831                 D(SYSTEM, "Remove a rule\n");
 832         else
 833                 D(SYSTEM, "Remove %s from device %s\n",
 834                   type == RTM_DELADDR ? "an address" : "a route",
 835                   clr->dev->ifname);
 836         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 837         hdr = nlmsg_hdr(clr->msg);
 838         hdr->nlmsg_type = type;
 839         hdr->nlmsg_flags = NLM_F_REQUEST;
 840
 841         nl_socket_disable_auto_ack(sock_rtnl);
 842         nl_send_auto_complete(sock_rtnl, clr->msg);
 843         nl_socket_enable_auto_ack(sock_rtnl);
 844
 845         return NL_SKIP;
 846 }
 847
 848 static int
 849 cb_finish_event(struct nl_msg *msg, void *arg)
 850 {
 851         int *pending = arg;
 852         *pending = 0;
 853         return NL_STOP;
 854 }
 855
 856 static int
 857 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 858 {
 859         int *pending = arg;
 860         *pending = err->error;
 861         return NL_STOP;
 862 }
 863
 864 static void
 865 system_if_clear_entries(struct device *dev, int type, int af)
 866 {
 867         struct clear_data clr;
 868         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
 869         struct rtmsg rtm = {
 870                 .rtm_family = af,
 871                 .rtm_flags = RTM_F_CLONED,
 872         };
 873         int flags = NLM_F_DUMP;
 874         int pending = 1;
 875
 876         clr.af = af;
 877         clr.dev = dev;
 878         clr.type = type;
 879         switch (type) {
 880         case RTM_GETADDR:
 881         case RTM_GETRULE:
 882                 clr.size = sizeof(struct rtgenmsg);
 883                 break;
 884         case RTM_GETROUTE:
 885                 clr.size = sizeof(struct rtmsg);
 886                 break;
 887         default:
 888                 return;
 889         }
 890
 891         if (!cb)
 892                 return;
 893
 894         clr.msg = nlmsg_alloc_simple(type, flags);
 895         if (!clr.msg)
 896                 goto out;
 897
 898         nlmsg_append(clr.msg, &rtm, clr.size, 0);
 899         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
 900         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
 901         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
 902
 903         nl_send_auto_complete(sock_rtnl, clr.msg);
 904         while (pending > 0)
 905                 nl_recvmsgs(sock_rtnl, cb);
 906
 907         nlmsg_free(clr.msg);
 908 out:
 909         nl_cb_put(cb);
 910 }
 911
 912 /*
 913  * Clear bridge (membership) state and bring down device
 914  */
 915 void system_if_clear_state(struct device *dev)
 916 {
 917         static char buf[256];
 918         char *bridge;
 919
 920         device_set_ifindex(dev, system_if_resolve(dev));
 921         if (dev->external || !dev->ifindex)
 922                 return;
 923
 924         system_if_flags(dev->ifname, 0, IFF_UP);
 925
 926         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
 927                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
 928                 system_bridge_delbr(dev);
 929                 return;
 930         }
 931
 932         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
 933         if (bridge) {
 934                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
 935                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
 936         }
 937
 938         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
 939         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
 940         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
 941         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
 942         system_set_disable_ipv6(dev, "0");
 943 }
 944
 945 static inline unsigned long
 946 sec_to_jiffies(int val)
 947 {
 948         return (unsigned long) val * 100;
 949 }
 950
 951 static void system_bridge_conf_multicast_deps(struct device *bridge,
 952                                               struct bridge_config *cfg,
 953                                               char *buf,
 954                                               int buf_len)
 955 {
 956         int val;
 957
 958         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
 959             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
 960             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
 961                 val = cfg->robustness * cfg->query_interval +
 962                         cfg->query_response_interval;
 963
 964                 snprintf(buf, buf_len, "%i", val);
 965                 system_bridge_set_membership_interval(bridge, buf);
 966
 967                 val = cfg->robustness * cfg->query_interval +
 968                         cfg->query_response_interval / 2;
 969
 970                 snprintf(buf, buf_len, "%i", val);
 971                 system_bridge_set_other_querier_timeout(bridge, buf);
 972         }
 973
 974         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
 975                 val = cfg->query_interval / 4;
 976
 977                 snprintf(buf, buf_len, "%i", val);
 978                 system_bridge_set_startup_query_interval(bridge, buf);
 979         }
 980 }
 981
 982 static void system_bridge_conf_multicast(struct device *bridge,
 983                                          struct bridge_config *cfg,
 984                                          char *buf,
 985                                          int buf_len)
 986 {
 987         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
 988                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
 989
 990         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
 991                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
 992
 993         snprintf(buf, buf_len, "%i", cfg->hash_max);
 994         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
 995                 bridge->ifname, buf);
 996
 997         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 998                 snprintf(buf, buf_len, "%u", bridge->settings.multicast_router);
 999                 system_bridge_set_multicast_router(bridge, buf, true);
1000         }
1001
1002         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
1003                 snprintf(buf, buf_len, "%i", cfg->robustness);
1004                 system_bridge_set_robustness(bridge, buf);
1005         }
1006
1007         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1008                 snprintf(buf, buf_len, "%i", cfg->query_interval);
1009                 system_bridge_set_query_interval(bridge, buf);
1010         }
1011
1012         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1013                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
1014                 system_bridge_set_query_response_interval(bridge, buf);
1015         }
1016
1017         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1018                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1019                 system_bridge_set_last_member_interval(bridge, buf);
1020         }
1021
1022         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1023 }
1024
1025 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1026 {
1027         char buf[64];
1028         unsigned long args[4] = {};
1029
1030         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1031                 return -1;
1032
1033         args[0] = BRCTL_SET_BRIDGE_STP_STATE;
1034         args[1] = !!cfg->stp;
1035         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1036
1037         args[0] = BRCTL_SET_BRIDGE_FORWARD_DELAY;
1038         args[1] = sec_to_jiffies(cfg->forward_delay);
1039         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1040
1041         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1042
1043         args[0] = BRCTL_SET_BRIDGE_PRIORITY;
1044         args[1] = cfg->priority;
1045         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1046
1047         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1048                 args[0] = BRCTL_SET_AGEING_TIME;
1049                 args[1] = sec_to_jiffies(cfg->ageing_time);
1050                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1051         }
1052
1053         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1054                 args[0] = BRCTL_SET_BRIDGE_HELLO_TIME;
1055                 args[1] = sec_to_jiffies(cfg->hello_time);
1056                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1057         }
1058
1059         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1060                 args[0] = BRCTL_SET_BRIDGE_MAX_AGE;
1061                 args[1] = sec_to_jiffies(cfg->max_age);
1062                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1063         }
1064
1065         return 0;
1066 }
1067
1068 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1069 {
1070         struct nl_msg *msg;
1071         struct nlattr *linkinfo, *data;
1072         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1073         int i, rv;
1074         static const struct {
1075                 const char *name;
1076                 enum macvlan_mode val;
1077         } modes[] = {
1078                 { "private", MACVLAN_MODE_PRIVATE },
1079                 { "vepa", MACVLAN_MODE_VEPA },
1080                 { "bridge", MACVLAN_MODE_BRIDGE },
1081                 { "passthru", MACVLAN_MODE_PASSTHRU },
1082         };
1083
1084         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1085
1086         if (!msg)
1087                 return -1;
1088
1089         nlmsg_append(msg, &iim, sizeof(iim), 0);
1090
1091         if (cfg->flags & MACVLAN_OPT_MACADDR)
1092                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1093         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1094         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1095
1096         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1097                 goto nla_put_failure;
1098
1099         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1100
1101         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1102                 goto nla_put_failure;
1103
1104         if (cfg->mode) {
1105                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1106                         if (strcmp(cfg->mode, modes[i].name) != 0)
1107                                 continue;
1108
1109                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1110                         break;
1111                 }
1112         }
1113
1114         nla_nest_end(msg, data);
1115         nla_nest_end(msg, linkinfo);
1116
1117         rv = system_rtnl_call(msg);
1118         if (rv)
1119                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1120
1121         return rv;
1122
1123 nla_put_failure:
1124         nlmsg_free(msg);
1125         return -ENOMEM;
1126 }
1127
1128 static int system_link_del(const char *ifname)
1129 {
1130         struct nl_msg *msg;
1131         struct ifinfomsg iim = {
1132                 .ifi_family = AF_UNSPEC,
1133                 .ifi_index = 0,
1134         };
1135
1136         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1137
1138         if (!msg)
1139                 return -1;
1140
1141         nlmsg_append(msg, &iim, sizeof(iim), 0);
1142         nla_put_string(msg, IFLA_IFNAME, ifname);
1143         return system_rtnl_call(msg);
1144 }
1145
1146 int system_macvlan_del(struct device *macvlan)
1147 {
1148         return system_link_del(macvlan->ifname);
1149 }
1150
1151 int system_veth_add(struct device *veth, struct veth_config *cfg)
1152 {
1153         struct nl_msg *msg;
1154         struct ifinfomsg empty_iim = {};
1155         struct nlattr *linkinfo, *data, *veth_info;
1156         int rv;
1157
1158         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1159
1160         if (!msg)
1161                 return -1;
1162
1163         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1164
1165         if (cfg->flags & VETH_OPT_MACADDR)
1166                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1167         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1168
1169         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1170                 goto nla_put_failure;
1171
1172         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1173
1174         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1175                 goto nla_put_failure;
1176
1177         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1178                 goto nla_put_failure;
1179
1180         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1181
1182         if (cfg->flags & VETH_OPT_PEER_NAME)
1183                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1184         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1185                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1186
1187         nla_nest_end(msg, veth_info);
1188         nla_nest_end(msg, data);
1189         nla_nest_end(msg, linkinfo);
1190
1191         rv = system_rtnl_call(msg);
1192         if (rv) {
1193                 if (cfg->flags & VETH_OPT_PEER_NAME)
1194                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1195                 else
1196                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1197         }
1198
1199         return rv;
1200
1201 nla_put_failure:
1202         nlmsg_free(msg);
1203         return -ENOMEM;
1204 }
1205
1206 int system_veth_del(struct device *veth)
1207 {
1208         return system_link_del(veth->ifname);
1209 }
1210
1211 static int system_vlan(struct device *dev, int id)
1212 {
1213         struct vlan_ioctl_args ifr = {
1214                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1215                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1216         };
1217
1218         ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1219
1220         if (id < 0) {
1221                 ifr.cmd = DEL_VLAN_CMD;
1222                 ifr.u.VID = 0;
1223         } else {
1224                 ifr.cmd = ADD_VLAN_CMD;
1225                 ifr.u.VID = id;
1226         }
1227         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1228         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1229 }
1230
1231 int system_vlan_add(struct device *dev, int id)
1232 {
1233         return system_vlan(dev, id);
1234 }
1235
1236 int system_vlan_del(struct device *dev)
1237 {
1238         return system_vlan(dev, -1);
1239 }
1240
1241 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1242 {
1243         struct nl_msg *msg;
1244         struct nlattr *linkinfo, *data;
1245         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1246         int rv;
1247
1248         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1249
1250         if (!msg)
1251                 return -1;
1252
1253         nlmsg_append(msg, &iim, sizeof(iim), 0);
1254         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1255         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1256
1257         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1258                 goto nla_put_failure;
1259
1260         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1261
1262         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1263                 goto nla_put_failure;
1264
1265         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1266
1267 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1268         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1269 #else
1270         if(cfg->proto == VLAN_PROTO_8021AD)
1271                 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);
1272 #endif
1273
1274         nla_nest_end(msg, data);
1275         nla_nest_end(msg, linkinfo);
1276
1277         rv = system_rtnl_call(msg);
1278         if (rv)
1279                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1280
1281         return rv;
1282
1283 nla_put_failure:
1284         nlmsg_free(msg);
1285         return -ENOMEM;
1286 }
1287
1288 int system_vlandev_del(struct device *vlandev)
1289 {
1290         return system_link_del(vlandev->ifname);
1291 }
1292
1293 void
1294 system_if_get_settings(struct device *dev, struct device_settings *s)
1295 {
1296         struct ifreq ifr;
1297         char buf[10];
1298
1299         memset(&ifr, 0, sizeof(ifr));
1300         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1301
1302         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1303                 s->mtu = ifr.ifr_mtu;
1304                 s->flags |= DEV_OPT_MTU;
1305         }
1306
1307         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1308         if (s->mtu6 > 0)
1309                 s->flags |= DEV_OPT_MTU6;
1310
1311         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1312                 s->txqueuelen = ifr.ifr_qlen;
1313                 s->flags |= DEV_OPT_TXQUEUELEN;
1314         }
1315
1316         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1317                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1318                 s->flags |= DEV_OPT_MACADDR;
1319         }
1320
1321         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1322                 s->ipv6 = !strtoul(buf, NULL, 0);
1323                 s->flags |= DEV_OPT_IPV6;
1324         }
1325
1326         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1327                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1328                 s->flags |= DEV_OPT_PROMISC;
1329
1330                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1331                 s->flags |= DEV_OPT_MULTICAST;
1332         }
1333
1334         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1335                 s->rpfilter = strtoul(buf, NULL, 0);
1336                 s->flags |= DEV_OPT_RPFILTER;
1337         }
1338
1339         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1340                 s->acceptlocal = strtoul(buf, NULL, 0);
1341                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1342         }
1343
1344         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1345                 s->igmpversion = strtoul(buf, NULL, 0);
1346                 s->flags |= DEV_OPT_IGMPVERSION;
1347         }
1348
1349         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1350                 s->mldversion = strtoul(buf, NULL, 0);
1351                 s->flags |= DEV_OPT_MLDVERSION;
1352         }
1353
1354         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1355                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1356                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1357         }
1358
1359         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1360                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1361                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1362         }
1363
1364         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
1365                 s->neigh4locktime = strtoul(buf, NULL, 0);
1366                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
1367         }
1368
1369         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1370                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1371                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1372         }
1373
1374         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1375                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1376                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1377         }
1378
1379         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1380                 s->dadtransmits = strtoul(buf, NULL, 0);
1381                 s->flags |= DEV_OPT_DADTRANSMITS;
1382         }
1383
1384         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1385                 s->sendredirects = strtoul(buf, NULL, 0);
1386                 s->flags |= DEV_OPT_SENDREDIRECTS;
1387         }
1388 }
1389
1390 static void
1391 system_if_set_rps_xps_val(const char *path, int val)
1392 {
1393         char val_buf[8];
1394         glob_t gl;
1395         int i;
1396
1397         if (glob(path, 0, NULL, &gl))
1398                 return;
1399
1400         snprintf(val_buf, sizeof(val_buf), "%x", val);
1401         for (i = 0; i < gl.gl_pathc; i++)
1402                 system_set_sysctl(gl.gl_pathv[i], val_buf);
1403
1404         globfree(&gl);
1405 }
1406
1407 static void
1408 system_if_apply_rps_xps(struct device *dev, struct device_settings *s)
1409 {
1410         long n_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1411         int val;
1412
1413         if (n_cpus < 2)
1414                 return;
1415
1416         val = (1 << n_cpus) - 1;
1417         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/rps_cpus", dev->ifname);
1418         system_if_set_rps_xps_val(dev_buf, s->rps ? val : 0);
1419
1420         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/xps_cpus", dev->ifname);
1421         system_if_set_rps_xps_val(dev_buf, s->xps ? val : 0);
1422 }
1423
1424 void
1425 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1426 {
1427         struct ifreq ifr;
1428         char buf[12];
1429
1430         memset(&ifr, 0, sizeof(ifr));
1431         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1432         if (s->flags & DEV_OPT_MTU & apply_mask) {
1433                 ifr.ifr_mtu = s->mtu;
1434                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1435                         s->flags &= ~DEV_OPT_MTU;
1436         }
1437         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1438                 system_update_ipv6_mtu(dev, s->mtu6);
1439         }
1440         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1441                 ifr.ifr_qlen = s->txqueuelen;
1442                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1443                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1444         }
1445         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1446                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1447                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1448                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1449                         s->flags &= ~DEV_OPT_MACADDR;
1450         }
1451         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1452                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1453         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1454                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1455                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1456                         s->flags &= ~DEV_OPT_PROMISC;
1457         }
1458         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1459                 snprintf(buf, sizeof(buf), "%u", s->rpfilter);
1460                 system_set_rpfilter(dev, buf);
1461         }
1462         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1463                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1464         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1465                 snprintf(buf, sizeof(buf), "%u", s->igmpversion);
1466                 system_set_igmpversion(dev, buf);
1467         }
1468         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1469                 snprintf(buf, sizeof(buf), "%u", s->mldversion);
1470                 system_set_mldversion(dev, buf);
1471         }
1472         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1473                 snprintf(buf, sizeof(buf), "%u", s->neigh4reachabletime);
1474                 system_set_neigh4reachabletime(dev, buf);
1475                 snprintf(buf, sizeof(buf), "%u", s->neigh6reachabletime);
1476                 system_set_neigh6reachabletime(dev, buf);
1477         }
1478         if (s->flags & DEV_OPT_NEIGHLOCKTIME & apply_mask) {
1479                 snprintf(buf, sizeof(buf), "%u", s->neigh4locktime);
1480                 system_set_neigh4locktime(dev, buf);
1481         }
1482         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1483                 snprintf(buf, sizeof(buf), "%u", s->neigh4gcstaletime);
1484                 system_set_neigh4gcstaletime(dev, buf);
1485                 snprintf(buf, sizeof(buf), "%u", s->neigh6gcstaletime);
1486                 system_set_neigh6gcstaletime(dev, buf);
1487         }
1488         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1489                 snprintf(buf, sizeof(buf), "%u", s->dadtransmits);
1490                 system_set_dadtransmits(dev, buf);
1491         }
1492         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1493                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1494                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1495                         s->flags &= ~DEV_OPT_MULTICAST;
1496         }
1497         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1498                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1499
1500         system_if_apply_rps_xps(dev, s);
1501 }
1502
1503 int system_if_up(struct device *dev)
1504 {
1505         system_if_get_settings(dev, &dev->orig_settings);
1506         /* Only keep orig settings based on what needs to be set */
1507         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1508         dev->orig_settings.flags &= dev->settings.flags;
1509         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1510         return system_if_flags(dev->ifname, IFF_UP, 0);
1511 }
1512
1513 int system_if_down(struct device *dev)
1514 {
1515         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1516         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1517         return ret;
1518 }
1519
1520 struct if_check_data {
1521         struct device *dev;
1522         int pending;
1523         int ret;
1524 };
1525
1526 #ifndef IFF_LOWER_UP
1527 #define IFF_LOWER_UP    0x10000
1528 #endif
1529
1530 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1531 {
1532         struct nlmsghdr *nh = nlmsg_hdr(msg);
1533         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1534         struct if_check_data *chk = (struct if_check_data *)arg;
1535
1536         if (nh->nlmsg_type != RTM_NEWLINK)
1537                 return NL_SKIP;
1538
1539         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1540         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1541
1542         return NL_OK;
1543 }
1544
1545 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1546 {
1547         struct if_check_data *chk = (struct if_check_data *)arg;
1548         chk->pending = 0;
1549         return NL_STOP;
1550 }
1551
1552 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1553 {
1554         struct if_check_data *chk = (struct if_check_data *)arg;
1555
1556         device_set_present(chk->dev, false);
1557         device_set_link(chk->dev, false);
1558         chk->pending = err->error;
1559
1560         return NL_STOP;
1561 }
1562
1563 int system_if_check(struct device *dev)
1564 {
1565         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1566         struct nl_msg *msg;
1567         struct ifinfomsg ifi = {
1568                 .ifi_family = AF_UNSPEC,
1569                 .ifi_index = 0,
1570         };
1571         struct if_check_data chk = {
1572                 .dev = dev,
1573                 .pending = 1,
1574         };
1575         int ret = 1;
1576
1577         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1578         if (!msg)
1579                 goto out;
1580
1581         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1582             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1583                 goto free;
1584
1585         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1586         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1587         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1588
1589         nl_send_auto_complete(sock_rtnl, msg);
1590         while (chk.pending > 0)
1591                 nl_recvmsgs(sock_rtnl, cb);
1592
1593         ret = chk.pending;
1594
1595 free:
1596         nlmsg_free(msg);
1597 out:
1598         nl_cb_put(cb);
1599         return ret;
1600 }
1601
1602 struct device *
1603 system_if_get_parent(struct device *dev)
1604 {
1605         char buf[64], *devname;
1606         int ifindex, iflink, len;
1607         FILE *f;
1608
1609         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1610         f = fopen(buf, "r");
1611         if (!f)
1612                 return NULL;
1613
1614         len = fread(buf, 1, sizeof(buf) - 1, f);
1615         fclose(f);
1616
1617         if (len <= 0)
1618                 return NULL;
1619
1620         buf[len] = 0;
1621         iflink = strtoul(buf, NULL, 0);
1622         ifindex = system_if_resolve(dev);
1623         if (!iflink || iflink == ifindex)
1624                 return NULL;
1625
1626         devname = if_indextoname(iflink, buf);
1627         if (!devname)
1628                 return NULL;
1629
1630         return device_get(devname, true);
1631 }
1632
1633 static bool
1634 read_string_file(int dir_fd, const char *file, char *buf, int len)
1635 {
1636         bool ret = false;
1637         char *c;
1638         int fd;
1639
1640         fd = openat(dir_fd, file, O_RDONLY);
1641         if (fd < 0)
1642                 return false;
1643
1644 retry:
1645         len = read(fd, buf, len - 1);
1646         if (len < 0) {
1647                 if (errno == EINTR)
1648                         goto retry;
1649         } else if (len > 0) {
1650                         buf[len] = 0;
1651
1652                         c = strchr(buf, '\n');
1653                         if (c)
1654                                 *c = 0;
1655
1656                         ret = true;
1657         }
1658
1659         close(fd);
1660
1661         return ret;
1662 }
1663
1664 static bool
1665 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1666 {
1667         char buf[64];
1668         bool ret = false;
1669
1670         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1671         if (ret)
1672                 *val = strtoull(buf, NULL, 0);
1673
1674         return ret;
1675 }
1676
1677 /* Assume advertised flags == supported flags */
1678 static const struct {
1679         uint32_t mask;
1680         const char *name;
1681 } ethtool_link_modes[] = {
1682         { ADVERTISED_10baseT_Half, "10H" },
1683         { ADVERTISED_10baseT_Full, "10F" },
1684         { ADVERTISED_100baseT_Half, "100H" },
1685         { ADVERTISED_100baseT_Full, "100F" },
1686         { ADVERTISED_1000baseT_Half, "1000H" },
1687         { ADVERTISED_1000baseT_Full, "1000F" },
1688 };
1689
1690 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1691 {
1692         int i;
1693         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1694                 if (mask & ethtool_link_modes[i].mask)
1695                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1696         }
1697 }
1698
1699 bool
1700 system_if_force_external(const char *ifname)
1701 {
1702         char buf[64];
1703         struct stat s;
1704
1705         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1706         return stat(buf, &s) == 0;
1707 }
1708
1709 int
1710 system_if_dump_info(struct device *dev, struct blob_buf *b)
1711 {
1712         struct ethtool_cmd ecmd;
1713         struct ifreq ifr;
1714         char buf[64], *s;
1715         void *c;
1716         int dir_fd;
1717
1718         snprintf(buf, sizeof(buf), "/sys/class/net/%s", dev->ifname);
1719         dir_fd = open(buf, O_DIRECTORY);
1720
1721         memset(&ecmd, 0, sizeof(ecmd));
1722         memset(&ifr, 0, sizeof(ifr));
1723         strcpy(ifr.ifr_name, dev->ifname);
1724         ifr.ifr_data = (caddr_t) &ecmd;
1725         ecmd.cmd = ETHTOOL_GSET;
1726
1727         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1728                 c = blobmsg_open_array(b, "link-advertising");
1729                 system_add_link_modes(b, ecmd.advertising);
1730                 blobmsg_close_array(b, c);
1731
1732                 c = blobmsg_open_array(b, "link-supported");
1733                 system_add_link_modes(b, ecmd.supported);
1734                 blobmsg_close_array(b, c);
1735
1736                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1737                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1738                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1739                 blobmsg_add_string_buffer(b);
1740         }
1741
1742         close(dir_fd);
1743         return 0;
1744 }
1745
1746 int
1747 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1748 {
1749         const char *const counters[] = {
1750                 "collisions",     "rx_frame_errors",   "tx_compressed",
1751                 "multicast",      "rx_length_errors",  "tx_dropped",
1752                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1753                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1754                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1755                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1756                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1757                 "rx_fifo_errors", "tx_carrier_errors",
1758         };
1759         char buf[64];
1760         int stats_dir;
1761         int i;
1762         uint64_t val = 0;
1763
1764         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1765         stats_dir = open(buf, O_DIRECTORY);
1766         if (stats_dir < 0)
1767                 return -1;
1768
1769         for (i = 0; i < ARRAY_SIZE(counters); i++)
1770                 if (read_uint64_file(stats_dir, counters[i], &val))
1771                         blobmsg_add_u64(b, counters[i], val);
1772
1773         close(stats_dir);
1774         return 0;
1775 }
1776
1777 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1778 {
1779         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1780         int alen = v4 ? 4 : 16;
1781         unsigned int flags = 0;
1782         struct ifaddrmsg ifa = {
1783                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1784                 .ifa_prefixlen = addr->mask,
1785                 .ifa_index = dev->ifindex,
1786         };
1787
1788         struct nl_msg *msg;
1789         if (cmd == RTM_NEWADDR)
1790                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1791
1792         msg = nlmsg_alloc_simple(cmd, flags);
1793         if (!msg)
1794                 return -1;
1795
1796         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1797         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1798         if (v4) {
1799                 if (addr->broadcast)
1800                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1801                 if (addr->point_to_point)
1802                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1803         } else {
1804                 time_t now = system_get_rtime();
1805                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1806
1807                 if (addr->preferred_until) {
1808                         int64_t preferred = addr->preferred_until - now;
1809                         if (preferred < 0)
1810                                 preferred = 0;
1811                         else if (preferred > UINT32_MAX)
1812                                 preferred = UINT32_MAX;
1813
1814                         cinfo.ifa_prefered = preferred;
1815                 }
1816
1817                 if (addr->valid_until) {
1818                         int64_t valid = addr->valid_until - now;
1819                         if (valid <= 0) {
1820                                 nlmsg_free(msg);
1821                                 return -1;
1822                         }
1823                         else if (valid > UINT32_MAX)
1824                                 valid = UINT32_MAX;
1825
1826                         cinfo.ifa_valid = valid;
1827                 }
1828
1829                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
1830
1831                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
1832                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
1833         }
1834
1835         return system_rtnl_call(msg);
1836 }
1837
1838 int system_add_address(struct device *dev, struct device_addr *addr)
1839 {
1840         return system_addr(dev, addr, RTM_NEWADDR);
1841 }
1842
1843 int system_del_address(struct device *dev, struct device_addr *addr)
1844 {
1845         return system_addr(dev, addr, RTM_DELADDR);
1846 }
1847
1848 static int system_rt(struct device *dev, struct device_route *route, int cmd)
1849 {
1850         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1851         bool have_gw;
1852         unsigned int flags = 0;
1853
1854         if (alen == 4)
1855                 have_gw = !!route->nexthop.in.s_addr;
1856         else
1857                 have_gw = route->nexthop.in6.s6_addr32[0] ||
1858                         route->nexthop.in6.s6_addr32[1] ||
1859                         route->nexthop.in6.s6_addr32[2] ||
1860                         route->nexthop.in6.s6_addr32[3];
1861
1862         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
1863                         ? route->table : RT_TABLE_MAIN;
1864
1865         struct rtmsg rtm = {
1866                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
1867                 .rtm_dst_len = route->mask,
1868                 .rtm_src_len = route->sourcemask,
1869                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
1870                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
1871                 .rtm_scope = RT_SCOPE_NOWHERE,
1872                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
1873                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
1874         };
1875         struct nl_msg *msg;
1876
1877         if (cmd == RTM_NEWROUTE) {
1878                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1879
1880                 if (!dev) { // Add null-route
1881                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1882                         rtm.rtm_type = RTN_UNREACHABLE;
1883                 }
1884                 else
1885                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
1886         }
1887
1888         if (route->flags & DEVROUTE_TYPE) {
1889                 rtm.rtm_type = route->type;
1890                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
1891                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
1892                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
1893                                 rtm.rtm_table = RT_TABLE_LOCAL;
1894                 }
1895
1896                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
1897                         rtm.rtm_scope = RT_SCOPE_HOST;
1898                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
1899                                 rtm.rtm_type == RTN_ANYCAST) {
1900                         rtm.rtm_scope = RT_SCOPE_LINK;
1901                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
1902                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY ||
1903                                 rtm.rtm_type == RTN_THROW) {
1904                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1905                         dev = NULL;
1906                 }
1907         }
1908
1909         msg = nlmsg_alloc_simple(cmd, flags);
1910         if (!msg)
1911                 return -1;
1912
1913         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
1914
1915         if (route->mask)
1916                 nla_put(msg, RTA_DST, alen, &route->addr);
1917
1918         if (route->sourcemask) {
1919                 if (rtm.rtm_family == AF_INET)
1920                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
1921                 else
1922                         nla_put(msg, RTA_SRC, alen, &route->source);
1923         }
1924
1925         if (route->metric > 0)
1926                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
1927
1928         if (have_gw)
1929                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
1930
1931         if (dev)
1932                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
1933
1934         if (table >= 256)
1935                 nla_put_u32(msg, RTA_TABLE, table);
1936
1937         if (route->flags & DEVROUTE_MTU) {
1938                 struct nlattr *metrics;
1939
1940                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
1941                         goto nla_put_failure;
1942
1943                 nla_put_u32(msg, RTAX_MTU, route->mtu);
1944
1945                 nla_nest_end(msg, metrics);
1946         }
1947
1948         return system_rtnl_call(msg);
1949
1950 nla_put_failure:
1951         nlmsg_free(msg);
1952         return -ENOMEM;
1953 }
1954
1955 int system_add_route(struct device *dev, struct device_route *route)
1956 {
1957         return system_rt(dev, route, RTM_NEWROUTE);
1958 }
1959
1960 int system_del_route(struct device *dev, struct device_route *route)
1961 {
1962         return system_rt(dev, route, RTM_DELROUTE);
1963 }
1964
1965 int system_flush_routes(void)
1966 {
1967         const char *names[] = {
1968                 "/proc/sys/net/ipv4/route/flush",
1969                 "/proc/sys/net/ipv6/route/flush"
1970         };
1971         int fd, i;
1972
1973         for (i = 0; i < ARRAY_SIZE(names); i++) {
1974                 fd = open(names[i], O_WRONLY);
1975                 if (fd < 0)
1976                         continue;
1977
1978                 if (write(fd, "-1", 2)) {}
1979                 close(fd);
1980         }
1981         return 0;
1982 }
1983
1984 bool system_resolve_rt_type(const char *type, unsigned int *id)
1985 {
1986         return system_rtn_aton(type, id);
1987 }
1988
1989 bool system_resolve_rt_proto(const char *type, unsigned int *id)
1990 {
1991         FILE *f;
1992         char *e, buf[128];
1993         unsigned int n, proto = 256;
1994         n = strtoul(type, &e, 0);
1995         if (!*e && e != type)
1996                 proto = n;
1997         else if (!strcmp(type, "unspec"))
1998                 proto = RTPROT_UNSPEC;
1999         else if (!strcmp(type, "kernel"))
2000                 proto = RTPROT_KERNEL;
2001         else if (!strcmp(type, "boot"))
2002                 proto = RTPROT_BOOT;
2003         else if (!strcmp(type, "static"))
2004                 proto = RTPROT_STATIC;
2005         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
2006                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
2007                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2008                                 continue;
2009
2010                         n = strtoul(e, NULL, 10);
2011                         e = strtok(NULL, " \t\n");
2012
2013                         if (e && !strcmp(e, type)) {
2014                                 proto = n;
2015                                 break;
2016                         }
2017                 }
2018                 fclose(f);
2019         }
2020
2021         if (proto > 255)
2022                 return false;
2023
2024         *id = proto;
2025         return true;
2026 }
2027
2028 bool system_resolve_rt_table(const char *name, unsigned int *id)
2029 {
2030         FILE *f;
2031         char *e, buf[128];
2032         unsigned int n, table = RT_TABLE_UNSPEC;
2033
2034         /* first try to parse table as number */
2035         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2036                 table = n;
2037
2038         /* handle well known aliases */
2039         else if (!strcmp(name, "default"))
2040                 table = RT_TABLE_DEFAULT;
2041         else if (!strcmp(name, "main"))
2042                 table = RT_TABLE_MAIN;
2043         else if (!strcmp(name, "local"))
2044                 table = RT_TABLE_LOCAL;
2045
2046         /* try to look up name in /etc/iproute2/rt_tables */
2047         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2048         {
2049                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2050                 {
2051                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2052                                 continue;
2053
2054                         n = strtoul(e, NULL, 10);
2055                         e = strtok(NULL, " \t\n");
2056
2057                         if (e && !strcmp(e, name))
2058                         {
2059                                 table = n;
2060                                 break;
2061                         }
2062                 }
2063
2064                 fclose(f);
2065         }
2066
2067         if (table == RT_TABLE_UNSPEC)
2068                 return false;
2069
2070         *id = table;
2071         return true;
2072 }
2073
2074 bool system_is_default_rt_table(unsigned int id)
2075 {
2076         return (id == RT_TABLE_MAIN);
2077 }
2078
2079 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2080 {
2081         char *e;
2082         unsigned int n;
2083
2084         if (!strcmp(filter, "strict"))
2085                 n = 1;
2086         else if (!strcmp(filter, "loose"))
2087                 n = 2;
2088         else {
2089                 n = strtoul(filter, &e, 0);
2090                 if (*e || e == filter || n > 2)
2091                         return false;
2092         }
2093
2094         *id = n;
2095         return true;
2096 }
2097
2098 static int system_iprule(struct iprule *rule, int cmd)
2099 {
2100         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2101
2102         struct nl_msg *msg;
2103         struct rtmsg rtm = {
2104                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2105                 .rtm_protocol = RTPROT_STATIC,
2106                 .rtm_scope = RT_SCOPE_UNIVERSE,
2107                 .rtm_table = RT_TABLE_UNSPEC,
2108                 .rtm_type = RTN_UNSPEC,
2109                 .rtm_flags = 0,
2110         };
2111
2112         if (cmd == RTM_NEWRULE)
2113                 rtm.rtm_type = RTN_UNICAST;
2114
2115         if (rule->invert)
2116                 rtm.rtm_flags |= FIB_RULE_INVERT;
2117
2118         if (rule->flags & IPRULE_SRC)
2119                 rtm.rtm_src_len = rule->src_mask;
2120
2121         if (rule->flags & IPRULE_DEST)
2122                 rtm.rtm_dst_len = rule->dest_mask;
2123
2124         if (rule->flags & IPRULE_TOS)
2125                 rtm.rtm_tos = rule->tos;
2126
2127         if (rule->flags & IPRULE_LOOKUP) {
2128                 if (rule->lookup < 256)
2129                         rtm.rtm_table = rule->lookup;
2130         }
2131
2132         if (rule->flags & IPRULE_ACTION)
2133                 rtm.rtm_type = rule->action;
2134         else if (rule->flags & IPRULE_GOTO)
2135                 rtm.rtm_type = FR_ACT_GOTO;
2136         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2137                 rtm.rtm_type = FR_ACT_NOP;
2138
2139         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2140
2141         if (!msg)
2142                 return -1;
2143
2144         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2145
2146         if (rule->flags & IPRULE_IN)
2147                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2148
2149         if (rule->flags & IPRULE_OUT)
2150                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2151
2152         if (rule->flags & IPRULE_SRC)
2153                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2154
2155         if (rule->flags & IPRULE_DEST)
2156                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2157
2158         if (rule->flags & IPRULE_PRIORITY)
2159                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2160         else if (cmd == RTM_NEWRULE)
2161                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2162
2163         if (rule->flags & IPRULE_FWMARK)
2164                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2165
2166         if (rule->flags & IPRULE_FWMASK)
2167                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2168
2169         if (rule->flags & IPRULE_LOOKUP) {
2170                 if (rule->lookup >= 256)
2171                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2172         }
2173
2174         if (rule->flags & IPRULE_SUP_PREFIXLEN)
2175                 nla_put_u32(msg, FRA_SUPPRESS_PREFIXLEN, rule->sup_prefixlen);
2176
2177         if (rule->flags & IPRULE_GOTO)
2178                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2179
2180         return system_rtnl_call(msg);
2181 }
2182
2183 int system_add_iprule(struct iprule *rule)
2184 {
2185         return system_iprule(rule, RTM_NEWRULE);
2186 }
2187
2188 int system_del_iprule(struct iprule *rule)
2189 {
2190         return system_iprule(rule, RTM_DELRULE);
2191 }
2192
2193 int system_flush_iprules(void)
2194 {
2195         int rv = 0;
2196         struct iprule rule;
2197
2198         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2199         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2200
2201         memset(&rule, 0, sizeof(rule));
2202
2203
2204         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2205
2206         rule.priority = 0;
2207         rule.lookup = RT_TABLE_LOCAL;
2208         rv |= system_iprule(&rule, RTM_NEWRULE);
2209
2210         rule.priority = 32766;
2211         rule.lookup = RT_TABLE_MAIN;
2212         rv |= system_iprule(&rule, RTM_NEWRULE);
2213
2214         rule.priority = 32767;
2215         rule.lookup = RT_TABLE_DEFAULT;
2216         rv |= system_iprule(&rule, RTM_NEWRULE);
2217
2218
2219         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2220
2221         rule.priority = 0;
2222         rule.lookup = RT_TABLE_LOCAL;
2223         rv |= system_iprule(&rule, RTM_NEWRULE);
2224
2225         rule.priority = 32766;
2226         rule.lookup = RT_TABLE_MAIN;
2227         rv |= system_iprule(&rule, RTM_NEWRULE);
2228
2229         return rv;
2230 }
2231
2232 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2233 {
2234         return system_rtn_aton(action, id);
2235 }
2236
2237 time_t system_get_rtime(void)
2238 {
2239         struct timespec ts;
2240         struct timeval tv;
2241
2242         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
2243                 return ts.tv_sec;
2244
2245         if (gettimeofday(&tv, NULL) == 0)
2246                 return tv.tv_sec;
2247
2248         return 0;
2249 }
2250
2251 #ifndef IP_DF
2252 #define IP_DF       0x4000
2253 #endif
2254
2255 static int tunnel_ioctl(const char *name, int cmd, void *p)
2256 {
2257         struct ifreq ifr;
2258
2259         memset(&ifr, 0, sizeof(ifr));
2260         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name));
2261         ifr.ifr_ifru.ifru_data = p;
2262         return ioctl(sock_ioctl, cmd, &ifr);
2263 }
2264
2265 #ifdef IFLA_IPTUN_MAX
2266 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2267 static int system_add_gre_tunnel(const char *name, const char *kind,
2268                                  const unsigned int link, struct blob_attr **tb, bool v6)
2269 {
2270         struct nl_msg *nlm;
2271         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2272         struct blob_attr *cur;
2273         uint32_t ikey = 0, okey = 0, flags = 0, flowinfo = 0;
2274         uint16_t iflags = 0, oflags = 0;
2275         uint8_t tos = 0;
2276         int ret = 0, ttl = 0;
2277
2278         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2279         if (!nlm)
2280                 return -1;
2281
2282         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2283         nla_put_string(nlm, IFLA_IFNAME, name);
2284
2285         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2286         if (!linkinfo) {
2287                 ret = -ENOMEM;
2288                 goto failure;
2289         }
2290
2291         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2292         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2293         if (!infodata) {
2294                 ret = -ENOMEM;
2295                 goto failure;
2296         }
2297
2298         if (link)
2299                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2300
2301         if ((cur = tb[TUNNEL_ATTR_TTL]))
2302                 ttl = blobmsg_get_u32(cur);
2303
2304         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2305                 char *str = blobmsg_get_string(cur);
2306                 if (strcmp(str, "inherit")) {
2307                         unsigned uval;
2308
2309                         if (!system_tos_aton(str, &uval)) {
2310                                 ret = -EINVAL;
2311                                 goto failure;
2312                         }
2313
2314                         if (v6)
2315                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2316                         else
2317                                 tos = uval;
2318                 } else {
2319                         if (v6)
2320                                 flags |= IP6_TNL_F_USE_ORIG_TCLASS;
2321                         else
2322                                 tos = 1;
2323                 }
2324         }
2325
2326         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2327                 struct blob_attr *tb_data[__GRE_DATA_ATTR_MAX];
2328
2329                 blobmsg_parse(gre_data_attr_list.params, __GRE_DATA_ATTR_MAX, tb_data,
2330                         blobmsg_data(cur), blobmsg_len(cur));
2331
2332                 if ((cur = tb_data[GRE_DATA_IKEY])) {
2333                         if ((ikey = blobmsg_get_u32(cur)))
2334                                 iflags |= GRE_KEY;
2335                 }
2336
2337                 if ((cur = tb_data[GRE_DATA_OKEY])) {
2338                         if ((okey = blobmsg_get_u32(cur)))
2339                                 oflags |= GRE_KEY;
2340                 }
2341
2342                 if ((cur = tb_data[GRE_DATA_ICSUM])) {
2343                         if (blobmsg_get_bool(cur))
2344                                 iflags |= GRE_CSUM;
2345                 }
2346
2347                 if ((cur = tb_data[GRE_DATA_OCSUM])) {
2348                         if (blobmsg_get_bool(cur))
2349                                 oflags |= GRE_CSUM;
2350                 }
2351
2352                 if ((cur = tb_data[GRE_DATA_ISEQNO])) {
2353                         if (blobmsg_get_bool(cur))
2354                                 iflags |= GRE_SEQ;
2355                 }
2356
2357                 if ((cur = tb[GRE_DATA_OSEQNO])) {
2358                         if (blobmsg_get_bool(cur))
2359                                 oflags |= GRE_SEQ;
2360                 }
2361         }
2362
2363         if (v6) {
2364                 struct in6_addr in6buf;
2365                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2366                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2367                                 ret = -EINVAL;
2368                                 goto failure;
2369                         }
2370                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2371                 }
2372
2373                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2374                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2375                                 ret = -EINVAL;
2376                                 goto failure;
2377                         }
2378                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2379                 }
2380                 nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, 4);
2381
2382                 if (flowinfo)
2383                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2384
2385                 if (flags)
2386                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags);
2387
2388                 if (!ttl)
2389                         ttl = 64;
2390         } else {
2391                 struct in_addr inbuf;
2392                 bool set_df = true;
2393
2394                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2395                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2396                                 ret = -EINVAL;
2397                                 goto failure;
2398                         }
2399                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2400                 }
2401
2402                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2403                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2404                                 ret = -EINVAL;
2405                                 goto failure;
2406                         }
2407                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2408
2409                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2410                                 if (!okey) {
2411                                         okey = inbuf.s_addr;
2412                                         oflags |= GRE_KEY;
2413                                 }
2414
2415                                 if (!ikey) {
2416                                         ikey = inbuf.s_addr;
2417                                         iflags |= GRE_KEY;
2418                                 }
2419                         }
2420                 }
2421
2422                 if ((cur = tb[TUNNEL_ATTR_DF]))
2423                         set_df = blobmsg_get_bool(cur);
2424
2425                 if (!set_df) {
2426                         /* ttl != 0 and nopmtudisc are incompatible */
2427                         if (ttl) {
2428                                 ret = -EINVAL;
2429                                 goto failure;
2430                         }
2431                 } else if (!ttl)
2432                         ttl = 64;
2433
2434                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2435
2436                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2437         }
2438
2439         if (ttl)
2440                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2441
2442         if (oflags)
2443                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2444
2445         if (iflags)
2446                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2447
2448         if (okey)
2449                 nla_put_u32(nlm, IFLA_GRE_OKEY, okey);
2450
2451         if (ikey)
2452                 nla_put_u32(nlm, IFLA_GRE_IKEY, ikey);
2453
2454         nla_nest_end(nlm, infodata);
2455         nla_nest_end(nlm, linkinfo);
2456
2457         return system_rtnl_call(nlm);
2458
2459 failure:
2460         nlmsg_free(nlm);
2461         return ret;
2462 }
2463 #endif
2464
2465 #ifdef IFLA_VTI_MAX
2466 static int system_add_vti_tunnel(const char *name, const char *kind,
2467                                  const unsigned int link, struct blob_attr **tb, bool v6)
2468 {
2469         struct nl_msg *nlm;
2470         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2471         struct blob_attr *cur;
2472         int ret = 0;
2473
2474         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2475         if (!nlm)
2476                 return -1;
2477
2478         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2479         nla_put_string(nlm, IFLA_IFNAME, name);
2480
2481         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2482         if (!linkinfo) {
2483                 ret = -ENOMEM;
2484                 goto failure;
2485         }
2486
2487         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2488         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2489         if (!infodata) {
2490                 ret = -ENOMEM;
2491                 goto failure;
2492         }
2493
2494         if (link)
2495                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2496
2497         if (v6) {
2498                 struct in6_addr in6buf;
2499                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2500                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2501                                 ret = -EINVAL;
2502                                 goto failure;
2503                         }
2504                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2505                 }
2506
2507                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2508                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2509                                 ret = -EINVAL;
2510                                 goto failure;
2511                         }
2512                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2513                 }
2514
2515         } else {
2516                 struct in_addr inbuf;
2517
2518                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2519                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2520                                 ret = -EINVAL;
2521                                 goto failure;
2522                         }
2523                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2524                 }
2525
2526                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2527                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2528                                 ret = -EINVAL;
2529                                 goto failure;
2530                         }
2531                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2532                 }
2533
2534         }
2535
2536         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2537                 struct blob_attr *tb_data[__VTI_DATA_ATTR_MAX];
2538                 uint32_t ikey = 0, okey = 0;
2539
2540                 blobmsg_parse(vti_data_attr_list.params, __VTI_DATA_ATTR_MAX, tb_data,
2541                         blobmsg_data(cur), blobmsg_len(cur));
2542
2543                 if ((cur = tb_data[VTI_DATA_IKEY])) {
2544                         if ((ikey = blobmsg_get_u32(cur)))
2545                                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2546                 }
2547
2548                 if ((cur = tb_data[VTI_DATA_OKEY])) {
2549                         if ((okey = blobmsg_get_u32(cur)))
2550                                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2551                 }
2552         }
2553
2554         nla_nest_end(nlm, infodata);
2555         nla_nest_end(nlm, linkinfo);
2556
2557         return system_rtnl_call(nlm);
2558
2559 failure:
2560         nlmsg_free(nlm);
2561         return ret;
2562 }
2563 #endif
2564
2565 #ifdef IFLA_VXLAN_MAX
2566 static int system_add_vxlan(const char *name, const unsigned int link, struct blob_attr **tb, bool v6)
2567 {
2568         struct blob_attr *tb_data[__VXLAN_DATA_ATTR_MAX];
2569         struct nl_msg *msg;
2570         struct nlattr *linkinfo, *data;
2571         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
2572         struct blob_attr *cur;
2573         int ret = 0;
2574
2575         if ((cur = tb[TUNNEL_ATTR_DATA]))
2576                 blobmsg_parse(vxlan_data_attr_list.params, __VXLAN_DATA_ATTR_MAX, tb_data,
2577                         blobmsg_data(cur), blobmsg_len(cur));
2578         else
2579                 return -EINVAL;
2580
2581         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
2582
2583         if (!msg)
2584                 return -1;
2585
2586         nlmsg_append(msg, &iim, sizeof(iim), 0);
2587
2588         nla_put_string(msg, IFLA_IFNAME, name);
2589
2590         if ((cur = tb_data[VXLAN_DATA_ATTR_MACADDR])) {
2591                 struct ether_addr *ea = ether_aton(blobmsg_get_string(cur));
2592                 if (!ea) {
2593                         ret = -EINVAL;
2594                         goto failure;
2595                 }
2596
2597                 nla_put(msg, IFLA_ADDRESS, ETH_ALEN, ea);
2598         }
2599
2600         if ((cur = tb[TUNNEL_ATTR_MTU])) {
2601                 uint32_t mtu = blobmsg_get_u32(cur);
2602                 nla_put_u32(msg, IFLA_MTU, mtu);
2603         }
2604
2605         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO))) {
2606                 ret = -ENOMEM;
2607                 goto failure;
2608         }
2609
2610         nla_put_string(msg, IFLA_INFO_KIND, "vxlan");
2611
2612         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA))) {
2613                 ret = -ENOMEM;
2614                 goto failure;
2615         }
2616
2617         if (link)
2618                 nla_put_u32(msg, IFLA_VXLAN_LINK, link);
2619
2620         if ((cur = tb_data[VXLAN_DATA_ATTR_ID])) {
2621                 uint32_t id = blobmsg_get_u32(cur);
2622                 if (id >= (1u << 24) - 1) {
2623                         ret = -EINVAL;
2624                         goto failure;
2625                 }
2626
2627                 nla_put_u32(msg, IFLA_VXLAN_ID, id);
2628         }
2629
2630         if (v6) {
2631                 struct in6_addr in6buf;
2632                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2633                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2634                                 ret = -EINVAL;
2635                                 goto failure;
2636                         }
2637                         nla_put(msg, IFLA_VXLAN_LOCAL6, sizeof(in6buf), &in6buf);
2638                 }
2639
2640                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2641                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2642                                 ret = -EINVAL;
2643                                 goto failure;
2644                         }
2645                         nla_put(msg, IFLA_VXLAN_GROUP6, sizeof(in6buf), &in6buf);
2646                 }
2647         } else {
2648                 struct in_addr inbuf;
2649
2650                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2651                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2652                                 ret = -EINVAL;
2653                                 goto failure;
2654                         }
2655                         nla_put(msg, IFLA_VXLAN_LOCAL, sizeof(inbuf), &inbuf);
2656                 }
2657
2658                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2659                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2660                                 ret = -EINVAL;
2661                                 goto failure;
2662                         }
2663                         nla_put(msg, IFLA_VXLAN_GROUP, sizeof(inbuf), &inbuf);
2664                 }
2665         }
2666
2667         uint32_t port = 4789;
2668         if ((cur = tb_data[VXLAN_DATA_ATTR_PORT])) {
2669                 port = blobmsg_get_u32(cur);
2670                 if (port < 1 || port > 65535) {
2671                         ret = -EINVAL;
2672                         goto failure;
2673                 }
2674         }
2675         nla_put_u16(msg, IFLA_VXLAN_PORT, htons(port));
2676
2677         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2678                 char *str = blobmsg_get_string(cur);
2679                 unsigned tos = 1;
2680
2681                 if (strcmp(str, "inherit")) {
2682                         if (!system_tos_aton(str, &tos))
2683                                 return -EINVAL;
2684                 }
2685
2686                 nla_put_u8(msg, IFLA_VXLAN_TOS, tos);
2687         }
2688
2689         if ((cur = tb[TUNNEL_ATTR_TTL])) {
2690                 uint32_t ttl = blobmsg_get_u32(cur);
2691                 if (ttl < 1 || ttl > 255) {
2692                         ret = -EINVAL;
2693                         goto failure;
2694                 }
2695
2696                 nla_put_u8(msg, IFLA_VXLAN_TTL, ttl);
2697         }
2698
2699         nla_nest_end(msg, data);
2700         nla_nest_end(msg, linkinfo);
2701
2702         ret = system_rtnl_call(msg);
2703         if (ret)
2704                 D(SYSTEM, "Error adding vxlan '%s': %d\n", name, ret);
2705
2706         return ret;
2707
2708 failure:
2709         nlmsg_free(msg);
2710         return ret;
2711 }
2712 #endif
2713
2714 static int system_add_sit_tunnel(const char *name, const unsigned int link, struct blob_attr **tb)
2715 {
2716         struct blob_attr *cur;
2717         int ret = 0;
2718
2719         if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
2720                 return -1;
2721
2722 #ifdef SIOCADD6RD
2723         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2724                 struct blob_attr *tb_data[__SIXRD_DATA_ATTR_MAX];
2725                 unsigned int mask;
2726                 struct ip_tunnel_6rd p6;
2727
2728                 blobmsg_parse(sixrd_data_attr_list.params, __SIXRD_DATA_ATTR_MAX, tb_data,
2729                         blobmsg_data(cur), blobmsg_len(cur));
2730
2731                 memset(&p6, 0, sizeof(p6));
2732
2733                 if ((cur = tb_data[SIXRD_DATA_PREFIX])) {
2734                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
2735                                                 &p6.prefix, &mask) || mask > 128) {
2736                                 ret = -EINVAL;
2737                                 goto failure;
2738                         }
2739
2740                         p6.prefixlen = mask;
2741                 }
2742
2743                 if ((cur = tb_data[SIXRD_DATA_RELAY_PREFIX])) {
2744                         if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
2745                                                 &p6.relay_prefix, &mask) || mask > 32) {
2746                                 ret = -EINVAL;
2747                                 goto failure;
2748                         }
2749
2750                         p6.relay_prefixlen = mask;
2751                 }
2752
2753                 if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
2754                         ret = -1;
2755                         goto failure;
2756                 }
2757         }
2758 #endif
2759
2760         return ret;
2761
2762 failure:
2763         __system_del_ip_tunnel(name, tb);
2764         return ret;
2765 }
2766
2767 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
2768 {
2769         struct blob_attr *cur;
2770         bool set_df = true;
2771         struct ip_tunnel_parm p  = {
2772                 .link = link,
2773                 .iph = {
2774                         .version = 4,
2775                         .ihl = 5,
2776                         .protocol = proto,
2777                 }
2778         };
2779
2780         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
2781                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
2782                 return -EINVAL;
2783
2784         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
2785                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
2786                 return -EINVAL;
2787
2788         if ((cur = tb[TUNNEL_ATTR_DF]))
2789                 set_df = blobmsg_get_bool(cur);
2790
2791         if ((cur = tb[TUNNEL_ATTR_TTL]))
2792                 p.iph.ttl = blobmsg_get_u32(cur);
2793
2794         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2795                 char *str = blobmsg_get_string(cur);
2796                 if (strcmp(str, "inherit")) {
2797                         unsigned uval;
2798
2799                         if (!system_tos_aton(str, &uval))
2800                                 return -EINVAL;
2801
2802                         p.iph.tos = uval;
2803                 } else
2804                         p.iph.tos = 1;
2805         }
2806
2807         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
2808         /* ttl !=0 and nopmtudisc are incompatible */
2809         if (p.iph.ttl && p.iph.frag_off == 0)
2810                 return -EINVAL;
2811
2812         strncpy(p.name, name, sizeof(p.name));
2813
2814         switch (p.iph.protocol) {
2815         case IPPROTO_IPIP:
2816                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
2817         case IPPROTO_IPV6:
2818                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
2819         default:
2820                 break;
2821         }
2822         return -1;
2823 }
2824
2825 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
2826 {
2827         struct blob_attr *cur;
2828         const char *str;
2829
2830         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2831                 return -EINVAL;
2832         str = blobmsg_data(cur);
2833
2834         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
2835             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
2836             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6") ||
2837             !strcmp(str, "vxlan") || !strcmp(str, "vxlan6"))
2838                 return system_link_del(name);
2839         else
2840                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
2841 }
2842
2843 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
2844 {
2845         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2846
2847         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2848                 blob_data(attr), blob_len(attr));
2849
2850         return __system_del_ip_tunnel(name, tb);
2851 }
2852
2853 int system_update_ipv6_mtu(struct device *dev, int mtu)
2854 {
2855         int ret = -1;
2856         char buf[64];
2857         int fd;
2858
2859         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
2860                         dev->ifname);
2861
2862         fd = open(buf, O_RDWR);
2863         if (fd < 0)
2864                 return ret;
2865
2866         if (!mtu) {
2867                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
2868                 if (len < 0)
2869                         goto out;
2870
2871                 buf[len] = 0;
2872                 ret = atoi(buf);
2873         } else {
2874                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
2875                         ret = mtu;
2876         }
2877
2878 out:
2879         close(fd);
2880         return ret;
2881 }
2882
2883 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
2884 {
2885         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2886         struct blob_attr *cur;
2887         const char *str;
2888
2889         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2890                 blob_data(attr), blob_len(attr));
2891
2892         __system_del_ip_tunnel(name, tb);
2893
2894         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2895                 return -EINVAL;
2896         str = blobmsg_data(cur);
2897
2898         unsigned int ttl = 0;
2899         if ((cur = tb[TUNNEL_ATTR_TTL])) {
2900                 ttl = blobmsg_get_u32(cur);
2901                 if (ttl > 255)
2902                         return -EINVAL;
2903         }
2904
2905         unsigned int link = 0;
2906         if ((cur = tb[TUNNEL_ATTR_LINK])) {
2907                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
2908                 if (!iface)
2909                         return -EINVAL;
2910
2911                 if (iface->l3_dev.dev)
2912                         link = iface->l3_dev.dev->ifindex;
2913         }
2914
2915         if (!strcmp(str, "sit"))
2916                 return system_add_sit_tunnel(name, link, tb);
2917 #ifdef IFLA_IPTUN_MAX
2918         else if (!strcmp(str, "ipip6")) {
2919                 struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2920                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2921                 struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2922                 int ret = 0;
2923
2924                 if (!nlm)
2925                         return -1;
2926
2927                 nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2928                 nla_put_string(nlm, IFLA_IFNAME, name);
2929
2930                 if (link)
2931                         nla_put_u32(nlm, IFLA_LINK, link);
2932
2933                 struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2934                 if (!linkinfo) {
2935                         ret = -ENOMEM;
2936                         goto failure;
2937                 }
2938                 nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2939                 struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2940                 if (!infodata) {
2941                         ret = -ENOMEM;
2942                         goto failure;
2943                 }
2944
2945                 if (link)
2946                         nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2947
2948                 nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2949                 nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2950                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, 4);
2951
2952                 struct in6_addr in6buf;
2953                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2954                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2955                                 ret = -EINVAL;
2956                                 goto failure;
2957                         }
2958                         nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2959                 }
2960
2961                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2962                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2963                                 ret = -EINVAL;
2964                                 goto failure;
2965                         }
2966                         nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2967                 }
2968
2969 #ifdef IFLA_IPTUN_FMR_MAX
2970                 if ((cur = tb[TUNNEL_ATTR_FMRS])) {
2971                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2972
2973                         struct blob_attr *fmr;
2974                         unsigned rem, fmrcnt = 0;
2975                         blobmsg_for_each_attr(fmr, cur, rem) {
2976                                 if (blobmsg_type(fmr) != BLOBMSG_TYPE_STRING)
2977                                         continue;
2978
2979                                 unsigned ip4len, ip6len, ealen, offset = 6;
2980                                 char ip6buf[48];
2981                                 char ip4buf[16];
2982
2983                                 if (sscanf(blobmsg_get_string(fmr), "%47[^/]/%u,%15[^/]/%u,%u,%u",
2984                                                 ip6buf, &ip6len, ip4buf, &ip4len, &ealen, &offset) < 5) {
2985                                         ret = -EINVAL;
2986                                         goto failure;
2987                                 }
2988
2989                                 struct in6_addr ip6prefix;
2990                                 struct in_addr ip4prefix;
2991                                 if (inet_pton(AF_INET6, ip6buf, &ip6prefix) != 1 ||
2992                                                 inet_pton(AF_INET, ip4buf, &ip4prefix) != 1) {
2993                                         ret = -EINVAL;
2994                                         goto failure;
2995                                 }
2996
2997                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2998
2999                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
3000                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
3001                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
3002                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
3003                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
3004                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
3005
3006                                 nla_nest_end(nlm, rule);
3007                         }
3008
3009                         nla_nest_end(nlm, fmrs);
3010                 }
3011 #endif
3012
3013                 nla_nest_end(nlm, infodata);
3014                 nla_nest_end(nlm, linkinfo);
3015
3016                 return system_rtnl_call(nlm);
3017 failure:
3018                 nlmsg_free(nlm);
3019                 return ret;
3020         } else if (!strcmp(str, "greip")) {
3021                 return system_add_gre_tunnel(name, "gre", link, tb, false);
3022         } else if (!strcmp(str, "gretapip"))  {
3023                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
3024         } else if (!strcmp(str, "greip6")) {
3025                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
3026         } else if (!strcmp(str, "gretapip6")) {
3027                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
3028 #ifdef IFLA_VTI_MAX
3029         } else if (!strcmp(str, "vtiip")) {
3030                 return system_add_vti_tunnel(name, "vti", link, tb, false);
3031         } else if (!strcmp(str, "vtiip6")) {
3032                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
3033 #endif
3034 #ifdef IFLA_VXLAN_MAX
3035         } else if(!strcmp(str, "vxlan")) {
3036                 return system_add_vxlan(name, link, tb, false);
3037         } else if(!strcmp(str, "vxlan6")) {
3038                 return system_add_vxlan(name, link, tb, true);
3039 #endif
3040 #endif
3041         } else if (!strcmp(str, "ipip")) {
3042                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
3043         }
3044         else
3045                 return -EINVAL;
3046
3047         return 0;
3048 }