system-linux.c

   1 /*
   2  * netifd - network interface daemon
   3  * Copyright (C) 2012 Felix Fietkau <nbd@openwrt.org>
   4  * Copyright (C) 2013 Jo-Philipp Wich <jow@openwrt.org>
   5  * Copyright (C) 2013 Steven Barth <steven@midlink.org>
   6  * Copyright (C) 2014 Gioacchino Mazzurco <gio@eigenlab.org>
   7  * Copyright (C) 2017 Matthias Schiffer <mschiffer@universe-factory.net>
   8  * Copyright (C) 2018 Hans Dedecker <dedeckeh@gmail.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  */
  19 #define _GNU_SOURCE
  20
  21 #include <sys/socket.h>
  22 #include <sys/ioctl.h>
  23 #include <sys/stat.h>
  24 #include <sys/syscall.h>
  25
  26 #include <net/if.h>
  27 #include <net/if_arp.h>
  28
  29 #include <limits.h>
  30 #include <arpa/inet.h>
  31 #include <netinet/ether.h>
  32 #include <netinet/in.h>
  33
  34 #include <linux/rtnetlink.h>
  35 #include <linux/neighbour.h>
  36 #include <linux/sockios.h>
  37 #include <linux/ip.h>
  38 #include <linux/if_addr.h>
  39 #include <linux/if_link.h>
  40 #include <linux/if_vlan.h>
  41 #include <linux/if_bridge.h>
  42 #include <linux/if_tunnel.h>
  43 #include <linux/ip6_tunnel.h>
  44 #include <linux/ethtool.h>
  45 #include <linux/fib_rules.h>
  46 #include <linux/veth.h>
  47 #include <linux/version.h>
  48
  49 #include <sched.h>
  50
  51 #ifndef RTN_FAILED_POLICY
  52 #define RTN_FAILED_POLICY 12
  53 #endif
  54
  55 #ifndef IFA_F_NOPREFIXROUTE
  56 #define IFA_F_NOPREFIXROUTE 0x200
  57 #endif
  58
  59 #ifndef IFA_FLAGS
  60 #define IFA_FLAGS (IFA_MULTICAST + 1)
  61 #endif
  62
  63 #include <string.h>
  64 #include <fcntl.h>
  65 #include <glob.h>
  66 #include <time.h>
  67 #include <unistd.h>
  68
  69 #include <netlink/msg.h>
  70 #include <netlink/attr.h>
  71 #include <netlink/socket.h>
  72 #include <libubox/uloop.h>
  73
  74 #include "netifd.h"
  75 #include "device.h"
  76 #include "system.h"
  77 #include "utils.h"
  78
  79 struct event_socket {
  80         struct uloop_fd uloop;
  81         struct nl_sock *sock;
  82         int bufsize;
  83 };
  84
  85 static int sock_ioctl = -1;
  86 static struct nl_sock *sock_rtnl = NULL;
  87
  88 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  89 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  90 static int system_add_proto_tunnel(const char *name, const uint8_t proto,
  91                                         const unsigned int link, struct blob_attr **tb);
  92 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb);
  93
  94 static char dev_buf[256];
  95
  96 static void
  97 handler_nl_event(struct uloop_fd *u, unsigned int events)
  98 {
  99         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 100         int err;
 101         socklen_t errlen = sizeof(err);
 102
 103         if (!u->error) {
 104                 nl_recvmsgs_default(ev->sock);
 105                 return;
 106         }
 107
 108         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 109                 goto abort;
 110
 111         switch(err) {
 112         case ENOBUFS:
 113                 /* Increase rx buffer size on netlink socket */
 114                 ev->bufsize *= 2;
 115                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 116                         goto abort;
 117
 118                 /* Request full dump since some info got dropped */
 119                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 120                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 121                 break;
 122
 123         default:
 124                 goto abort;
 125         }
 126         u->error = false;
 127         return;
 128
 129 abort:
 130         uloop_fd_delete(&ev->uloop);
 131         return;
 132 }
 133
 134 static struct nl_sock *
 135 create_socket(int protocol, int groups)
 136 {
 137         struct nl_sock *sock;
 138
 139         sock = nl_socket_alloc();
 140         if (!sock)
 141                 return NULL;
 142
 143         if (groups)
 144                 nl_join_groups(sock, groups);
 145
 146         if (nl_connect(sock, protocol)) {
 147                 nl_socket_free(sock);
 148                 return NULL;
 149         }
 150
 151         return sock;
 152 }
 153
 154 static bool
 155 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 156                         uloop_fd_handler cb, int flags)
 157 {
 158         ev->sock = create_socket(protocol, groups);
 159         if (!ev->sock)
 160                 return false;
 161
 162         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 163         ev->uloop.cb = cb;
 164         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 165                 return false;
 166
 167         return true;
 168 }
 169
 170 static bool
 171 create_event_socket(struct event_socket *ev, int protocol,
 172                     int (*cb)(struct nl_msg *msg, void *arg))
 173 {
 174         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 175                 return false;
 176
 177         /* Install the valid custom callback handler */
 178         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 179
 180         /* Disable sequence number checking on event sockets */
 181         nl_socket_disable_seq_check(ev->sock);
 182
 183         /* Increase rx buffer size to 65K on event sockets */
 184         ev->bufsize = 65535;
 185         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 186                 return false;
 187
 188         return true;
 189 }
 190
 191 static bool
 192 create_hotplug_event_socket(struct event_socket *ev, int protocol,
 193                             void (*cb)(struct uloop_fd *u, unsigned int events))
 194 {
 195         if (!create_raw_event_socket(ev, protocol, 1, cb, ULOOP_ERROR_CB))
 196                 return false;
 197
 198         /* Increase rx buffer size to 65K on event sockets */
 199         ev->bufsize = 65535;
 200         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 201                 return false;
 202
 203         return true;
 204 }
 205
 206 static bool
 207 system_rtn_aton(const char *src, unsigned int *dst)
 208 {
 209         char *e;
 210         unsigned int n;
 211
 212         if (!strcmp(src, "local"))
 213                 n = RTN_LOCAL;
 214         else if (!strcmp(src, "nat"))
 215                 n = RTN_NAT;
 216         else if (!strcmp(src, "broadcast"))
 217                 n = RTN_BROADCAST;
 218         else if (!strcmp(src, "anycast"))
 219                 n = RTN_ANYCAST;
 220         else if (!strcmp(src, "multicast"))
 221                 n = RTN_MULTICAST;
 222         else if (!strcmp(src, "prohibit"))
 223                 n = RTN_PROHIBIT;
 224         else if (!strcmp(src, "unreachable"))
 225                 n = RTN_UNREACHABLE;
 226         else if (!strcmp(src, "blackhole"))
 227                 n = RTN_BLACKHOLE;
 228         else if (!strcmp(src, "xresolve"))
 229                 n = RTN_XRESOLVE;
 230         else if (!strcmp(src, "unicast"))
 231                 n = RTN_UNICAST;
 232         else if (!strcmp(src, "throw"))
 233                 n = RTN_THROW;
 234         else if (!strcmp(src, "failed_policy"))
 235                 n = RTN_FAILED_POLICY;
 236         else {
 237                 n = strtoul(src, &e, 0);
 238                 if (!e || *e || e == src || n > 255)
 239                         return false;
 240         }
 241
 242         *dst = n;
 243         return true;
 244 }
 245
 246 static bool
 247 system_tos_aton(const char *src, unsigned *dst)
 248 {
 249         char *e;
 250
 251         *dst = strtoul(src, &e, 16);
 252         if (e == src || *e || *dst > 255)
 253                 return false;
 254
 255         return true;
 256 }
 257
 258 int system_init(void)
 259 {
 260         static struct event_socket rtnl_event;
 261         static struct event_socket hotplug_event;
 262
 263         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 264         system_fd_set_cloexec(sock_ioctl);
 265
 266         /* Prepare socket for routing / address control */
 267         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 268         if (!sock_rtnl)
 269                 return -1;
 270
 271         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 272                 return -1;
 273
 274         if (!create_hotplug_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT,
 275                                          handle_hotplug_event))
 276                 return -1;
 277
 278         /* Receive network link events form kernel */
 279         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 280
 281         return 0;
 282 }
 283
 284 static void system_set_sysctl(const char *path, const char *val)
 285 {
 286         int fd;
 287
 288         fd = open(path, O_WRONLY);
 289         if (fd < 0)
 290                 return;
 291
 292         if (write(fd, val, strlen(val))) {}
 293         close(fd);
 294 }
 295
 296 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 297 {
 298         snprintf(dev_buf, sizeof(dev_buf), path, device);
 299         system_set_sysctl(dev_buf, val);
 300 }
 301
 302 static void system_set_disable_ipv6(struct device *dev, const char *val)
 303 {
 304         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 305 }
 306
 307 static void system_set_rpfilter(struct device *dev, const char *val)
 308 {
 309         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter", dev->ifname, val);
 310 }
 311
 312 static void system_set_acceptlocal(struct device *dev, const char *val)
 313 {
 314         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local", dev->ifname, val);
 315 }
 316
 317 static void system_set_igmpversion(struct device *dev, const char *val)
 318 {
 319         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version", dev->ifname, val);
 320 }
 321
 322 static void system_set_mldversion(struct device *dev, const char *val)
 323 {
 324         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version", dev->ifname, val);
 325 }
 326
 327 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 328 {
 329         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 330 }
 331
 332 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 333 {
 334         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 335 }
 336
 337 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 338 {
 339         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time", dev->ifname, val);
 340 }
 341
 342 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 343 {
 344         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time", dev->ifname, val);
 345 }
 346
 347 static void system_set_neigh4locktime(struct device *dev, const char *val)
 348 {
 349         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime", dev->ifname, val);
 350 }
 351
 352 static void system_set_dadtransmits(struct device *dev, const char *val)
 353 {
 354         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 355 }
 356
 357 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 358 {
 359         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 360 }
 361
 362 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 363 {
 364         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 365 }
 366
 367 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 368 {
 369         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 370 }
 371
 372 static void system_bridge_set_isolated(struct device *dev, const char *val)
 373 {
 374         system_set_dev_sysctl("/sys/class/net/%s/brport/isolated", dev->ifname, val);
 375 }
 376
 377 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 378 {
 379         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 380                                        "/sys/class/net/%s/brport/multicast_router",
 381                               dev->ifname, val);
 382 }
 383
 384 static void system_bridge_set_robustness(struct device *dev, const char *val)
 385 {
 386         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 387                               dev->ifname, val);
 388         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 389                               dev->ifname, val);
 390 }
 391
 392 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 393 {
 394         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 395                               dev->ifname, val);
 396 }
 397
 398 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 399 {
 400         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 401                               dev->ifname, val);
 402 }
 403
 404 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 405 {
 406         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 407                               dev->ifname, val);
 408 }
 409
 410 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 411 {
 412         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 413                               dev->ifname, val);
 414 }
 415
 416 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 417 {
 418         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 419                               dev->ifname, val);
 420 }
 421
 422 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 423 {
 424         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 425                               dev->ifname, val);
 426 }
 427
 428 static void system_bridge_set_stp_state(struct device *dev, const char *val)
 429 {
 430         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/stp_state", dev->ifname, val);
 431 }
 432
 433 static void system_bridge_set_forward_delay(struct device *dev, const char *val)
 434 {
 435         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/forward_delay", dev->ifname, val);
 436 }
 437
 438 static void system_bridge_set_priority(struct device *dev, const char *val)
 439 {
 440         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/priority", dev->ifname, val);
 441 }
 442
 443 static void system_bridge_set_ageing_time(struct device *dev, const char *val)
 444 {
 445         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/ageing_time", dev->ifname, val);
 446 }
 447
 448 static void system_bridge_set_hello_time(struct device *dev, const char *val)
 449 {
 450         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hello_time", dev->ifname, val);
 451 }
 452
 453 static void system_bridge_set_max_age(struct device *dev, const char *val)
 454 {
 455         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/max_age", dev->ifname, val);
 456 }
 457
 458 static void system_bridge_set_learning(struct device *dev, const char *val)
 459 {
 460         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 461 }
 462
 463 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 464 {
 465         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 466 }
 467
 468 static void system_set_sendredirects(struct device *dev, const char *val)
 469 {
 470         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 471 }
 472
 473 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 474 {
 475         int fd = -1, ret = -1;
 476
 477         fd = open(path, O_RDONLY);
 478         if (fd < 0)
 479                 goto out;
 480
 481         ssize_t len = read(fd, buf, buf_sz - 1);
 482         if (len < 0)
 483                 goto out;
 484
 485         ret = buf[len] = 0;
 486
 487 out:
 488         if (fd >= 0)
 489                 close(fd);
 490
 491         return ret;
 492 }
 493
 494 static int
 495 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 496 {
 497         snprintf(dev_buf, sizeof(dev_buf), path, device);
 498         return system_get_sysctl(dev_buf, buf, buf_sz);
 499 }
 500
 501 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 502 {
 503         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 504                         dev->ifname, buf, buf_sz);
 505 }
 506
 507 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 508 {
 509         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 510                         dev->ifname, buf, buf_sz);
 511 }
 512
 513 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 514 {
 515         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 516                         dev->ifname, buf, buf_sz);
 517 }
 518
 519 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 520 {
 521         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 522                         dev->ifname, buf, buf_sz);
 523 }
 524
 525 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 526 {
 527         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 528                         dev->ifname, buf, buf_sz);
 529 }
 530
 531 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 532 {
 533         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 534                         dev->ifname, buf, buf_sz);
 535 }
 536
 537 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 538 {
 539         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 540                         dev->ifname, buf, buf_sz);
 541 }
 542
 543 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 544 {
 545         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 546                         dev->ifname, buf, buf_sz);
 547 }
 548
 549 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 550 {
 551         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 552                         dev->ifname, buf, buf_sz);
 553 }
 554
 555 static int system_get_neigh4locktime(struct device *dev, char *buf, const size_t buf_sz)
 556 {
 557         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime",
 558                         dev->ifname, buf, buf_sz);
 559 }
 560
 561 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 562 {
 563         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 564                         dev->ifname, buf, buf_sz);
 565 }
 566
 567 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 568 {
 569         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 570                         dev->ifname, buf, buf_sz);
 571 }
 572
 573 /* Evaluate netlink messages */
 574 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 575 {
 576         struct nlmsghdr *nh = nlmsg_hdr(msg);
 577         struct nlattr *nla[__IFLA_MAX];
 578         int link_state = 0;
 579         char buf[10];
 580
 581         if (nh->nlmsg_type != RTM_NEWLINK)
 582                 goto out;
 583
 584         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 585         if (!nla[IFLA_IFNAME])
 586                 goto out;
 587
 588         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 589         if (!dev)
 590                 goto out;
 591
 592         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 593                 link_state = strtoul(buf, NULL, 0);
 594
 595         device_set_link(dev, link_state ? true : false);
 596
 597 out:
 598         return 0;
 599 }
 600
 601 static void
 602 handle_hotplug_msg(char *data, int size)
 603 {
 604         const char *subsystem = NULL, *interface = NULL, *interface_old = NULL;
 605         char *cur, *end, *sep;
 606         struct device *dev;
 607         int skip;
 608         bool add, move = false;
 609
 610         if (!strncmp(data, "add@", 4))
 611                 add = true;
 612         else if (!strncmp(data, "remove@", 7))
 613                 add = false;
 614         else if (!strncmp(data, "move@", 5)) {
 615                 add = true;
 616                 move = true;
 617         }
 618         else
 619                 return;
 620
 621         skip = strlen(data) + 1;
 622         end = data + size;
 623
 624         for (cur = data + skip; cur < end; cur += skip) {
 625                 skip = strlen(cur) + 1;
 626
 627                 sep = strchr(cur, '=');
 628                 if (!sep)
 629                         continue;
 630
 631                 *sep = 0;
 632                 if (!strcmp(cur, "INTERFACE"))
 633                         interface = sep + 1;
 634                 else if (!strcmp(cur, "SUBSYSTEM")) {
 635                         subsystem = sep + 1;
 636                         if (strcmp(subsystem, "net") != 0)
 637                                 return;
 638                 } else if (!strcmp(cur, "DEVPATH_OLD")) {
 639                         interface_old = strrchr(sep + 1, '/');
 640                         if (interface_old)
 641                                 interface_old++;
 642                 }
 643         }
 644
 645         if (subsystem && interface) {
 646                 if (move && interface_old)
 647                         goto move;
 648                 else
 649                         goto found;
 650         }
 651
 652         return;
 653
 654 move:
 655         dev = device_find(interface_old);
 656         if (!dev)
 657                 goto found;
 658
 659         if (dev->type != &simple_device_type)
 660                 goto found;
 661
 662         device_set_present(dev, false);
 663
 664 found:
 665         dev = device_find(interface);
 666         if (!dev)
 667                 return;
 668
 669         if (dev->type != &simple_device_type)
 670                 return;
 671
 672         if (add && system_if_force_external(dev->ifname))
 673                 return;
 674
 675         device_set_present(dev, add);
 676 }
 677
 678 static void
 679 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 680 {
 681         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 682         struct sockaddr_nl nla;
 683         unsigned char *buf = NULL;
 684         int size;
 685         int err;
 686         socklen_t errlen = sizeof(err);
 687
 688         if (!u->error) {
 689                 while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 690                         if (nla.nl_pid == 0)
 691                                 handle_hotplug_msg((char *) buf, size);
 692
 693                         free(buf);
 694                 }
 695                 return;
 696         }
 697
 698         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 699                 goto abort;
 700
 701         switch(err) {
 702         case ENOBUFS:
 703                 /* Increase rx buffer size on netlink socket */
 704                 ev->bufsize *= 2;
 705                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 706                         goto abort;
 707                 break;
 708
 709         default:
 710                 goto abort;
 711         }
 712         u->error = false;
 713         return;
 714
 715 abort:
 716         uloop_fd_delete(&ev->uloop);
 717         return;
 718 }
 719
 720 static int system_rtnl_call(struct nl_msg *msg)
 721 {
 722         int ret;
 723
 724         ret = nl_send_auto_complete(sock_rtnl, msg);
 725         nlmsg_free(msg);
 726
 727         if (ret < 0)
 728                 return ret;
 729
 730         return nl_wait_for_ack(sock_rtnl);
 731 }
 732
 733 int system_bridge_delbr(struct device *bridge)
 734 {
 735         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 736 }
 737
 738 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 739 {
 740         struct ifreq ifr;
 741
 742         memset(&ifr, 0, sizeof(ifr));
 743         if (dev)
 744                 ifr.ifr_ifindex = dev->ifindex;
 745         else
 746                 ifr.ifr_data = data;
 747         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name) - 1);
 748         return ioctl(sock_ioctl, cmd, &ifr);
 749 }
 750
 751 static bool system_is_bridge(const char *name, char *buf, int buflen)
 752 {
 753         struct stat st;
 754
 755         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 756         if (stat(buf, &st) < 0)
 757                 return false;
 758
 759         return true;
 760 }
 761
 762 static char *system_get_bridge(const char *name, char *buf, int buflen)
 763 {
 764         char *path;
 765         ssize_t len = -1;
 766         glob_t gl;
 767
 768         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 769         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 770                 return NULL;
 771
 772         if (gl.gl_pathc > 0)
 773                 len = readlink(gl.gl_pathv[0], buf, buflen);
 774
 775         globfree(&gl);
 776
 777         if (len < 0)
 778                 return NULL;
 779
 780         buf[len] = 0;
 781         path = strrchr(buf, '/');
 782         if (!path)
 783                 return NULL;
 784
 785         return path + 1;
 786 }
 787
 788 static void
 789 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 790 {
 791         bool mcast_to_ucast = dev->wireless_ap;
 792         bool hairpin = true;
 793
 794         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 795             !bridge->settings.multicast_to_unicast)
 796                 mcast_to_ucast = false;
 797
 798         if (!mcast_to_ucast || dev->wireless_isolate)
 799                 hairpin = false;
 800
 801         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 802         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 803 }
 804
 805 int system_bridge_addif(struct device *bridge, struct device *dev)
 806 {
 807         char buf[64];
 808         char *oldbr;
 809         int ret = 0;
 810
 811         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 812         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 813                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 814
 815         if (dev->wireless)
 816                 system_bridge_set_wireless(bridge, dev);
 817
 818         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 819                 snprintf(buf, sizeof(buf), "%u", dev->settings.multicast_router);
 820                 system_bridge_set_multicast_router(dev, buf, false);
 821         }
 822
 823         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 824             dev->settings.multicast_fast_leave)
 825                 system_bridge_set_multicast_fast_leave(dev, "1");
 826
 827         if (dev->settings.flags & DEV_OPT_LEARNING &&
 828             !dev->settings.learning)
 829                 system_bridge_set_learning(dev, "0");
 830
 831         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 832             !dev->settings.unicast_flood)
 833                 system_bridge_set_unicast_flood(dev, "0");
 834
 835         if (dev->settings.flags & DEV_OPT_ISOLATE &&
 836             dev->settings.isolate)
 837                 system_bridge_set_isolated(dev, "1");
 838
 839         return ret;
 840 }
 841
 842 int system_bridge_delif(struct device *bridge, struct device *dev)
 843 {
 844         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 845 }
 846
 847 int system_if_resolve(struct device *dev)
 848 {
 849         struct ifreq ifr;
 850         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
 851         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 852                 return ifr.ifr_ifindex;
 853         else
 854                 return 0;
 855 }
 856
 857 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 858 {
 859         struct ifreq ifr;
 860
 861         memset(&ifr, 0, sizeof(ifr));
 862         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
 863         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) < 0)
 864                 return -1;
 865
 866         ifr.ifr_flags |= add;
 867         ifr.ifr_flags &= ~rem;
 868         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 869 }
 870
 871 struct clear_data {
 872         struct nl_msg *msg;
 873         struct device *dev;
 874         int type;
 875         int size;
 876         int af;
 877 };
 878
 879
 880 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 881 {
 882         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 883
 884         return ifa->ifa_index == ifindex;
 885 }
 886
 887 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 888 {
 889         struct rtmsg *r = NLMSG_DATA(hdr);
 890         struct nlattr *tb[__RTA_MAX];
 891
 892         if (r->rtm_protocol == RTPROT_KERNEL &&
 893             r->rtm_family == AF_INET6)
 894                 return false;
 895
 896         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 897         if (!tb[RTA_OIF])
 898                 return false;
 899
 900         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 901 }
 902
 903 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 904 {
 905         return true;
 906 }
 907
 908 static int cb_clear_event(struct nl_msg *msg, void *arg)
 909 {
 910         struct clear_data *clr = arg;
 911         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 912         bool (*cb)(struct nlmsghdr *, int ifindex);
 913         int type, ret;
 914
 915         switch(clr->type) {
 916         case RTM_GETADDR:
 917                 type = RTM_DELADDR;
 918                 if (hdr->nlmsg_type != RTM_NEWADDR)
 919                         return NL_SKIP;
 920
 921                 cb = check_ifaddr;
 922                 break;
 923         case RTM_GETROUTE:
 924                 type = RTM_DELROUTE;
 925                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 926                         return NL_SKIP;
 927
 928                 cb = check_route;
 929                 break;
 930         case RTM_GETRULE:
 931                 type = RTM_DELRULE;
 932                 if (hdr->nlmsg_type != RTM_NEWRULE)
 933                         return NL_SKIP;
 934
 935                 cb = check_rule;
 936                 break;
 937         default:
 938                 return NL_SKIP;
 939         }
 940
 941         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 942                 return NL_SKIP;
 943
 944         if (type == RTM_DELRULE)
 945                 D(SYSTEM, "Remove a rule\n");
 946         else
 947                 D(SYSTEM, "Remove %s from device %s\n",
 948                   type == RTM_DELADDR ? "an address" : "a route",
 949                   clr->dev->ifname);
 950
 951         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 952         hdr = nlmsg_hdr(clr->msg);
 953         hdr->nlmsg_type = type;
 954         hdr->nlmsg_flags = NLM_F_REQUEST;
 955
 956         nl_socket_disable_auto_ack(sock_rtnl);
 957         ret = nl_send_auto_complete(sock_rtnl, clr->msg);
 958         if (ret < 0) {
 959                 if (type == RTM_DELRULE)
 960                         D(SYSTEM, "Error deleting a rule: %d\n", ret);
 961                 else
 962                         D(SYSTEM, "Error deleting %s from device '%s': %d\n",
 963                                 type == RTM_DELADDR ? "an address" : "a route",
 964                                 clr->dev->ifname, ret);
 965         }
 966
 967         nl_socket_enable_auto_ack(sock_rtnl);
 968
 969         return NL_SKIP;
 970 }
 971
 972 static int
 973 cb_finish_event(struct nl_msg *msg, void *arg)
 974 {
 975         int *pending = arg;
 976         *pending = 0;
 977         return NL_STOP;
 978 }
 979
 980 static int
 981 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 982 {
 983         int *pending = arg;
 984         *pending = err->error;
 985         return NL_STOP;
 986 }
 987
 988 static void
 989 system_if_clear_entries(struct device *dev, int type, int af)
 990 {
 991         struct clear_data clr;
 992         struct nl_cb *cb;
 993         struct rtmsg rtm = {
 994                 .rtm_family = af,
 995                 .rtm_flags = RTM_F_CLONED,
 996         };
 997         int flags = NLM_F_DUMP;
 998         int pending = 1;
 999
1000         clr.af = af;
1001         clr.dev = dev;
1002         clr.type = type;
1003         switch (type) {
1004         case RTM_GETADDR:
1005         case RTM_GETRULE:
1006                 clr.size = sizeof(struct rtgenmsg);
1007                 break;
1008         case RTM_GETROUTE:
1009                 clr.size = sizeof(struct rtmsg);
1010                 break;
1011         default:
1012                 return;
1013         }
1014
1015         cb = nl_cb_alloc(NL_CB_DEFAULT);
1016         if (!cb)
1017                 return;
1018
1019         clr.msg = nlmsg_alloc_simple(type, flags);
1020         if (!clr.msg)
1021                 goto out;
1022
1023         nlmsg_append(clr.msg, &rtm, clr.size, 0);
1024         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
1025         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
1026         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
1027
1028         if (nl_send_auto_complete(sock_rtnl, clr.msg) < 0)
1029                 goto free;
1030
1031         while (pending > 0)
1032                 nl_recvmsgs(sock_rtnl, cb);
1033
1034 free:
1035         nlmsg_free(clr.msg);
1036 out:
1037         nl_cb_put(cb);
1038 }
1039
1040 /*
1041  * Clear bridge (membership) state and bring down device
1042  */
1043 void system_if_clear_state(struct device *dev)
1044 {
1045         static char buf[256];
1046         char *bridge;
1047         device_set_ifindex(dev, system_if_resolve(dev));
1048
1049         if (dev->external || !dev->ifindex)
1050                 return;
1051
1052         system_if_flags(dev->ifname, 0, IFF_UP);
1053
1054         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
1055                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
1056                 system_bridge_delbr(dev);
1057                 return;
1058         }
1059
1060         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
1061         if (bridge) {
1062                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
1063                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
1064         }
1065
1066         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
1067         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
1068         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
1069         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
1070         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET);
1071         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET6);
1072         system_set_disable_ipv6(dev, "0");
1073 }
1074
1075 static inline unsigned long
1076 sec_to_jiffies(int val)
1077 {
1078         return (unsigned long) val * 100;
1079 }
1080
1081 static void system_bridge_conf_multicast_deps(struct device *bridge,
1082                                               struct bridge_config *cfg,
1083                                               char *buf,
1084                                               int buf_len)
1085 {
1086         int val;
1087
1088         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
1089             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
1090             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1091                 val = cfg->robustness * cfg->query_interval +
1092                         cfg->query_response_interval;
1093
1094                 snprintf(buf, buf_len, "%i", val);
1095                 system_bridge_set_membership_interval(bridge, buf);
1096
1097                 val = cfg->robustness * cfg->query_interval +
1098                         cfg->query_response_interval / 2;
1099
1100                 snprintf(buf, buf_len, "%i", val);
1101                 system_bridge_set_other_querier_timeout(bridge, buf);
1102         }
1103
1104         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1105                 val = cfg->query_interval / 4;
1106
1107                 snprintf(buf, buf_len, "%i", val);
1108                 system_bridge_set_startup_query_interval(bridge, buf);
1109         }
1110 }
1111
1112 static void system_bridge_conf_multicast(struct device *bridge,
1113                                          struct bridge_config *cfg,
1114                                          char *buf,
1115                                          int buf_len)
1116 {
1117         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
1118                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
1119
1120         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
1121                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
1122
1123         snprintf(buf, buf_len, "%i", cfg->hash_max);
1124         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
1125                 bridge->ifname, buf);
1126
1127         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
1128                 snprintf(buf, buf_len, "%u", bridge->settings.multicast_router);
1129                 system_bridge_set_multicast_router(bridge, buf, true);
1130         }
1131
1132         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
1133                 snprintf(buf, buf_len, "%i", cfg->robustness);
1134                 system_bridge_set_robustness(bridge, buf);
1135         }
1136
1137         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1138                 snprintf(buf, buf_len, "%i", cfg->query_interval);
1139                 system_bridge_set_query_interval(bridge, buf);
1140         }
1141
1142         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1143                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
1144                 system_bridge_set_query_response_interval(bridge, buf);
1145         }
1146
1147         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1148                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1149                 system_bridge_set_last_member_interval(bridge, buf);
1150         }
1151
1152         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1153 }
1154
1155 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1156 {
1157         char buf[64];
1158
1159         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1160                 return -1;
1161
1162         system_bridge_set_stp_state(bridge, cfg->stp ? "1" : "0");
1163
1164         snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->forward_delay));
1165         system_bridge_set_forward_delay(bridge, buf);
1166
1167         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1168
1169         snprintf(buf, sizeof(buf), "%d", cfg->priority);
1170         system_bridge_set_priority(bridge, buf);
1171
1172         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1173                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->ageing_time));
1174                 system_bridge_set_ageing_time(bridge, buf);
1175         }
1176
1177         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1178                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->hello_time));
1179                 system_bridge_set_hello_time(bridge, buf);
1180         }
1181
1182         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1183                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->max_age));
1184                 system_bridge_set_max_age(bridge, buf);
1185         }
1186
1187         return 0;
1188 }
1189
1190 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1191 {
1192         struct nl_msg *msg;
1193         struct nlattr *linkinfo, *data;
1194         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1195         int i, rv;
1196         static const struct {
1197                 const char *name;
1198                 enum macvlan_mode val;
1199         } modes[] = {
1200                 { "private", MACVLAN_MODE_PRIVATE },
1201                 { "vepa", MACVLAN_MODE_VEPA },
1202                 { "bridge", MACVLAN_MODE_BRIDGE },
1203                 { "passthru", MACVLAN_MODE_PASSTHRU },
1204         };
1205
1206         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1207
1208         if (!msg)
1209                 return -1;
1210
1211         nlmsg_append(msg, &iim, sizeof(iim), 0);
1212
1213         if (cfg->flags & MACVLAN_OPT_MACADDR)
1214                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1215         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1216         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1217
1218         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1219                 goto nla_put_failure;
1220
1221         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1222
1223         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1224                 goto nla_put_failure;
1225
1226         if (cfg->mode) {
1227                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1228                         if (strcmp(cfg->mode, modes[i].name) != 0)
1229                                 continue;
1230
1231                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1232                         break;
1233                 }
1234         }
1235
1236         nla_nest_end(msg, data);
1237         nla_nest_end(msg, linkinfo);
1238
1239         rv = system_rtnl_call(msg);
1240         if (rv)
1241                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1242
1243         return rv;
1244
1245 nla_put_failure:
1246         nlmsg_free(msg);
1247         return -ENOMEM;
1248 }
1249
1250 int system_link_netns_move(struct device *dev, int netns_fd, const char *target_ifname)
1251 {
1252         struct nl_msg *msg;
1253         struct ifinfomsg iim = {
1254                 .ifi_family = AF_UNSPEC,
1255         };
1256
1257         if (!dev)
1258                 return -1;
1259
1260         iim.ifi_index = system_if_resolve(dev);
1261         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST);
1262
1263         if (!msg)
1264                 return -1;
1265
1266         nlmsg_append(msg, &iim, sizeof(iim), 0);
1267         if (target_ifname)
1268                 nla_put_string(msg, IFLA_IFNAME, target_ifname);
1269
1270         nla_put_u32(msg, IFLA_NET_NS_FD, netns_fd);
1271         return system_rtnl_call(msg);
1272 }
1273
1274 static int system_link_del(const char *ifname)
1275 {
1276         struct nl_msg *msg;
1277         struct ifinfomsg iim = {
1278                 .ifi_family = AF_UNSPEC,
1279                 .ifi_index = 0,
1280         };
1281
1282         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1283
1284         if (!msg)
1285                 return -1;
1286
1287         nlmsg_append(msg, &iim, sizeof(iim), 0);
1288         nla_put_string(msg, IFLA_IFNAME, ifname);
1289         return system_rtnl_call(msg);
1290 }
1291
1292 int system_macvlan_del(struct device *macvlan)
1293 {
1294         return system_link_del(macvlan->ifname);
1295 }
1296
1297 int system_netns_open(const pid_t target_ns)
1298 {
1299         char pid_net_path[PATH_MAX];
1300
1301         snprintf(pid_net_path, sizeof(pid_net_path), "/proc/%u/ns/net", target_ns);
1302
1303         return open(pid_net_path, O_RDONLY);
1304 }
1305
1306 int system_netns_set(int netns_fd)
1307 {
1308         return setns(netns_fd, CLONE_NEWNET);
1309 }
1310
1311 int system_veth_add(struct device *veth, struct veth_config *cfg)
1312 {
1313         struct nl_msg *msg;
1314         struct ifinfomsg empty_iim = {};
1315         struct nlattr *linkinfo, *data, *veth_info;
1316         int rv;
1317
1318         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1319
1320         if (!msg)
1321                 return -1;
1322
1323         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1324
1325         if (cfg->flags & VETH_OPT_MACADDR)
1326                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1327         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1328
1329         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1330                 goto nla_put_failure;
1331
1332         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1333
1334         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1335                 goto nla_put_failure;
1336
1337         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1338                 goto nla_put_failure;
1339
1340         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1341
1342         if (cfg->flags & VETH_OPT_PEER_NAME)
1343                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1344         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1345                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1346
1347         nla_nest_end(msg, veth_info);
1348         nla_nest_end(msg, data);
1349         nla_nest_end(msg, linkinfo);
1350
1351         rv = system_rtnl_call(msg);
1352         if (rv) {
1353                 if (cfg->flags & VETH_OPT_PEER_NAME)
1354                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1355                 else
1356                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1357         }
1358
1359         return rv;
1360
1361 nla_put_failure:
1362         nlmsg_free(msg);
1363         return -ENOMEM;
1364 }
1365
1366 int system_veth_del(struct device *veth)
1367 {
1368         return system_link_del(veth->ifname);
1369 }
1370
1371 static int system_vlan(struct device *dev, int id)
1372 {
1373         struct vlan_ioctl_args ifr = {
1374                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1375                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1376         };
1377
1378         if (ioctl(sock_ioctl, SIOCSIFVLAN, &ifr) < 0)
1379                 return -1;
1380
1381         if (id < 0) {
1382                 ifr.cmd = DEL_VLAN_CMD;
1383                 ifr.u.VID = 0;
1384         } else {
1385                 ifr.cmd = ADD_VLAN_CMD;
1386                 ifr.u.VID = id;
1387         }
1388         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1389         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1390 }
1391
1392 int system_vlan_add(struct device *dev, int id)
1393 {
1394         return system_vlan(dev, id);
1395 }
1396
1397 int system_vlan_del(struct device *dev)
1398 {
1399         return system_vlan(dev, -1);
1400 }
1401
1402 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1403 {
1404         struct nl_msg *msg;
1405         struct nlattr *linkinfo, *data, *qos;
1406         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1407         struct vlan_qos_mapping *elem;
1408         struct ifla_vlan_qos_mapping nl_qos_map;
1409         int rv;
1410
1411         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1412
1413         if (!msg)
1414                 return -1;
1415
1416         nlmsg_append(msg, &iim, sizeof(iim), 0);
1417         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1418         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1419
1420         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1421                 goto nla_put_failure;
1422
1423         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1424
1425         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1426                 goto nla_put_failure;
1427
1428         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1429
1430 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1431         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1432 #else
1433         if(cfg->proto == VLAN_PROTO_8021AD)
1434                 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);
1435 #endif
1436
1437         if (!(qos = nla_nest_start(msg, IFLA_VLAN_INGRESS_QOS)))
1438                 goto nla_put_failure;
1439
1440         vlist_simple_for_each_element(&cfg->ingress_qos_mapping_list, elem, node) {
1441                 nl_qos_map.from = elem->from;
1442                 nl_qos_map.to = elem->to;
1443                 nla_put(msg, IFLA_VLAN_QOS_MAPPING, sizeof(nl_qos_map), &nl_qos_map);
1444         }
1445         nla_nest_end(msg, qos);
1446
1447         if (!(qos = nla_nest_start(msg, IFLA_VLAN_EGRESS_QOS)))
1448                 goto nla_put_failure;
1449
1450         vlist_simple_for_each_element(&cfg->egress_qos_mapping_list, elem, node) {
1451                 nl_qos_map.from = elem->from;
1452                 nl_qos_map.to = elem->to;
1453                 nla_put(msg, IFLA_VLAN_QOS_MAPPING, sizeof(nl_qos_map), &nl_qos_map);
1454         }
1455         nla_nest_end(msg, qos);
1456
1457         nla_nest_end(msg, data);
1458         nla_nest_end(msg, linkinfo);
1459
1460         rv = system_rtnl_call(msg);
1461         if (rv)
1462                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1463
1464         return rv;
1465
1466 nla_put_failure:
1467         nlmsg_free(msg);
1468         return -ENOMEM;
1469 }
1470
1471 int system_vlandev_del(struct device *vlandev)
1472 {
1473         return system_link_del(vlandev->ifname);
1474 }
1475
1476 void
1477 system_if_get_settings(struct device *dev, struct device_settings *s)
1478 {
1479         struct ifreq ifr;
1480         char buf[10];
1481
1482         memset(&ifr, 0, sizeof(ifr));
1483         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1484
1485         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1486                 s->mtu = ifr.ifr_mtu;
1487                 s->flags |= DEV_OPT_MTU;
1488         }
1489
1490         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1491         if (s->mtu6 > 0)
1492                 s->flags |= DEV_OPT_MTU6;
1493
1494         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1495                 s->txqueuelen = ifr.ifr_qlen;
1496                 s->flags |= DEV_OPT_TXQUEUELEN;
1497         }
1498
1499         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1500                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1501                 s->flags |= DEV_OPT_MACADDR;
1502         }
1503
1504         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1505                 s->ipv6 = !strtoul(buf, NULL, 0);
1506                 s->flags |= DEV_OPT_IPV6;
1507         }
1508
1509         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1510                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1511                 s->flags |= DEV_OPT_PROMISC;
1512
1513                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1514                 s->flags |= DEV_OPT_MULTICAST;
1515         }
1516
1517         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1518                 s->rpfilter = strtoul(buf, NULL, 0);
1519                 s->flags |= DEV_OPT_RPFILTER;
1520         }
1521
1522         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1523                 s->acceptlocal = strtoul(buf, NULL, 0);
1524                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1525         }
1526
1527         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1528                 s->igmpversion = strtoul(buf, NULL, 0);
1529                 s->flags |= DEV_OPT_IGMPVERSION;
1530         }
1531
1532         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1533                 s->mldversion = strtoul(buf, NULL, 0);
1534                 s->flags |= DEV_OPT_MLDVERSION;
1535         }
1536
1537         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1538                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1539                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1540         }
1541
1542         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1543                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1544                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1545         }
1546
1547         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
1548                 s->neigh4locktime = strtol(buf, NULL, 0);
1549                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
1550         }
1551
1552         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1553                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1554                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1555         }
1556
1557         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1558                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1559                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1560         }
1561
1562         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1563                 s->dadtransmits = strtoul(buf, NULL, 0);
1564                 s->flags |= DEV_OPT_DADTRANSMITS;
1565         }
1566
1567         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1568                 s->sendredirects = strtoul(buf, NULL, 0);
1569                 s->flags |= DEV_OPT_SENDREDIRECTS;
1570         }
1571 }
1572
1573 void
1574 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1575 {
1576         struct ifreq ifr;
1577         char buf[12];
1578
1579         memset(&ifr, 0, sizeof(ifr));
1580         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1581         if (s->flags & DEV_OPT_MTU & apply_mask) {
1582                 ifr.ifr_mtu = s->mtu;
1583                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1584                         s->flags &= ~DEV_OPT_MTU;
1585         }
1586         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1587                 system_update_ipv6_mtu(dev, s->mtu6);
1588         }
1589         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1590                 ifr.ifr_qlen = s->txqueuelen;
1591                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1592                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1593         }
1594         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1595                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1596                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1597                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1598                         s->flags &= ~DEV_OPT_MACADDR;
1599         }
1600         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1601                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1602         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1603                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1604                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1605                         s->flags &= ~DEV_OPT_PROMISC;
1606         }
1607         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1608                 snprintf(buf, sizeof(buf), "%u", s->rpfilter);
1609                 system_set_rpfilter(dev, buf);
1610         }
1611         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1612                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1613         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1614                 snprintf(buf, sizeof(buf), "%u", s->igmpversion);
1615                 system_set_igmpversion(dev, buf);
1616         }
1617         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1618                 snprintf(buf, sizeof(buf), "%u", s->mldversion);
1619                 system_set_mldversion(dev, buf);
1620         }
1621         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1622                 snprintf(buf, sizeof(buf), "%u", s->neigh4reachabletime);
1623                 system_set_neigh4reachabletime(dev, buf);
1624                 snprintf(buf, sizeof(buf), "%u", s->neigh6reachabletime);
1625                 system_set_neigh6reachabletime(dev, buf);
1626         }
1627         if (s->flags & DEV_OPT_NEIGHLOCKTIME & apply_mask) {
1628                 snprintf(buf, sizeof(buf), "%d", s->neigh4locktime);
1629                 system_set_neigh4locktime(dev, buf);
1630         }
1631         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1632                 snprintf(buf, sizeof(buf), "%u", s->neigh4gcstaletime);
1633                 system_set_neigh4gcstaletime(dev, buf);
1634                 snprintf(buf, sizeof(buf), "%u", s->neigh6gcstaletime);
1635                 system_set_neigh6gcstaletime(dev, buf);
1636         }
1637         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1638                 snprintf(buf, sizeof(buf), "%u", s->dadtransmits);
1639                 system_set_dadtransmits(dev, buf);
1640         }
1641         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1642                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1643                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1644                         s->flags &= ~DEV_OPT_MULTICAST;
1645         }
1646         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1647                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1648 }
1649
1650 int system_if_up(struct device *dev)
1651 {
1652         system_if_get_settings(dev, &dev->orig_settings);
1653         /* Only keep orig settings based on what needs to be set */
1654         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1655         dev->orig_settings.flags &= dev->settings.flags;
1656         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1657         return system_if_flags(dev->ifname, IFF_UP, 0);
1658 }
1659
1660 int system_if_down(struct device *dev)
1661 {
1662         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1663         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1664         return ret;
1665 }
1666
1667 struct if_check_data {
1668         struct device *dev;
1669         int pending;
1670         int ret;
1671 };
1672
1673 #ifndef IFF_LOWER_UP
1674 #define IFF_LOWER_UP    0x10000
1675 #endif
1676
1677 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1678 {
1679         struct nlmsghdr *nh = nlmsg_hdr(msg);
1680         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1681         struct if_check_data *chk = (struct if_check_data *)arg;
1682
1683         if (nh->nlmsg_type != RTM_NEWLINK)
1684                 return NL_SKIP;
1685
1686         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1687         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1688
1689         return NL_OK;
1690 }
1691
1692 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1693 {
1694         struct if_check_data *chk = (struct if_check_data *)arg;
1695         chk->pending = 0;
1696         return NL_STOP;
1697 }
1698
1699 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1700 {
1701         struct if_check_data *chk = (struct if_check_data *)arg;
1702
1703         device_set_present(chk->dev, false);
1704         device_set_link(chk->dev, false);
1705         chk->pending = err->error;
1706
1707         return NL_STOP;
1708 }
1709
1710 int system_if_check(struct device *dev)
1711 {
1712         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1713         struct nl_msg *msg;
1714         struct ifinfomsg ifi = {
1715                 .ifi_family = AF_UNSPEC,
1716                 .ifi_index = 0,
1717         };
1718         struct if_check_data chk = {
1719                 .dev = dev,
1720                 .pending = 1,
1721         };
1722         int ret = 1;
1723
1724         if (!cb)
1725                 return ret;
1726
1727         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1728         if (!msg)
1729                 goto out;
1730
1731         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1732             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1733                 goto free;
1734
1735         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1736         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1737         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1738
1739         ret = nl_send_auto_complete(sock_rtnl, msg);
1740         if (ret < 0)
1741                 goto free;
1742
1743         while (chk.pending > 0)
1744                 nl_recvmsgs(sock_rtnl, cb);
1745
1746         ret = chk.pending;
1747
1748 free:
1749         nlmsg_free(msg);
1750 out:
1751         nl_cb_put(cb);
1752         return ret;
1753 }
1754
1755 struct device *
1756 system_if_get_parent(struct device *dev)
1757 {
1758         char buf[64], *devname;
1759         int ifindex, iflink, len;
1760         FILE *f;
1761
1762         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1763         f = fopen(buf, "r");
1764         if (!f)
1765                 return NULL;
1766
1767         len = fread(buf, 1, sizeof(buf) - 1, f);
1768         fclose(f);
1769
1770         if (len <= 0)
1771                 return NULL;
1772
1773         buf[len] = 0;
1774         iflink = strtoul(buf, NULL, 0);
1775         ifindex = system_if_resolve(dev);
1776         if (!iflink || iflink == ifindex)
1777                 return NULL;
1778
1779         devname = if_indextoname(iflink, buf);
1780         if (!devname)
1781                 return NULL;
1782
1783         return device_get(devname, true);
1784 }
1785
1786 static bool
1787 read_string_file(int dir_fd, const char *file, char *buf, int len)
1788 {
1789         bool ret = false;
1790         char *c;
1791         int fd;
1792
1793         fd = openat(dir_fd, file, O_RDONLY);
1794         if (fd < 0)
1795                 return false;
1796
1797 retry:
1798         len = read(fd, buf, len - 1);
1799         if (len < 0) {
1800                 if (errno == EINTR)
1801                         goto retry;
1802         } else if (len > 0) {
1803                         buf[len] = 0;
1804
1805                         c = strchr(buf, '\n');
1806                         if (c)
1807                                 *c = 0;
1808
1809                         ret = true;
1810         }
1811
1812         close(fd);
1813
1814         return ret;
1815 }
1816
1817 static bool
1818 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1819 {
1820         char buf[64];
1821         bool ret = false;
1822
1823         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1824         if (ret)
1825                 *val = strtoull(buf, NULL, 0);
1826
1827         return ret;
1828 }
1829
1830 /* Assume advertised flags == supported flags */
1831 static const struct {
1832         uint32_t mask;
1833         const char *name;
1834 } ethtool_link_modes[] = {
1835         { ADVERTISED_10baseT_Half, "10baseT-H" },
1836         { ADVERTISED_10baseT_Full, "10baseT-F" },
1837         { ADVERTISED_100baseT_Half, "100baseT-H" },
1838         { ADVERTISED_100baseT_Full, "100baseT-F" },
1839         { ADVERTISED_1000baseT_Half, "1000baseT-H" },
1840         { ADVERTISED_1000baseT_Full, "1000baseT-F" },
1841         { ADVERTISED_1000baseKX_Full, "1000baseKX-F" },
1842         { ADVERTISED_2500baseX_Full, "2500baseX-F" },
1843         { ADVERTISED_10000baseT_Full, "10000baseT-F" },
1844         { ADVERTISED_10000baseKX4_Full, "10000baseKX4-F" },
1845         { ADVERTISED_10000baseKR_Full, "10000baseKR-F" },
1846         { ADVERTISED_20000baseMLD2_Full, "20000baseMLD2-F" },
1847         { ADVERTISED_20000baseKR2_Full, "20000baseKR2-F" },
1848         { ADVERTISED_40000baseKR4_Full, "40000baseKR4-F" },
1849         { ADVERTISED_40000baseCR4_Full, "40000baseCR4-F" },
1850         { ADVERTISED_40000baseSR4_Full, "40000baseSR4-F" },
1851         { ADVERTISED_40000baseLR4_Full, "40000baseLR4-F" },
1852 #ifdef ADVERTISED_56000baseKR4_Full
1853         { ADVERTISED_56000baseKR4_Full, "56000baseKR4-F" },
1854         { ADVERTISED_56000baseCR4_Full, "56000baseCR4-F" },
1855         { ADVERTISED_56000baseSR4_Full, "56000baseSR4-F" },
1856         { ADVERTISED_56000baseLR4_Full, "56000baseLR4-F" },
1857 #endif
1858 };
1859
1860 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1861 {
1862         int i;
1863         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1864                 if (mask & ethtool_link_modes[i].mask)
1865                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1866         }
1867 }
1868
1869 bool
1870 system_if_force_external(const char *ifname)
1871 {
1872         char buf[64];
1873         struct stat s;
1874
1875         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1876         return stat(buf, &s) == 0;
1877 }
1878
1879 int
1880 system_if_dump_info(struct device *dev, struct blob_buf *b)
1881 {
1882         struct ethtool_cmd ecmd;
1883         struct ifreq ifr;
1884         char *s;
1885         void *c;
1886
1887         memset(&ecmd, 0, sizeof(ecmd));
1888         memset(&ifr, 0, sizeof(ifr));
1889         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1890         ifr.ifr_data = (caddr_t) &ecmd;
1891         ecmd.cmd = ETHTOOL_GSET;
1892
1893         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1894                 c = blobmsg_open_array(b, "link-advertising");
1895                 system_add_link_modes(b, ecmd.advertising);
1896                 blobmsg_close_array(b, c);
1897
1898                 c = blobmsg_open_array(b, "link-partner-advertising");
1899                 system_add_link_modes(b, ecmd.lp_advertising);
1900                 blobmsg_close_array(b, c);
1901
1902                 c = blobmsg_open_array(b, "link-supported");
1903                 system_add_link_modes(b, ecmd.supported);
1904                 blobmsg_close_array(b, c);
1905
1906                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1907                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1908                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1909                 blobmsg_add_string_buffer(b);
1910
1911                 blobmsg_add_u8(b, "autoneg", !!ecmd.autoneg);
1912         }
1913
1914         return 0;
1915 }
1916
1917 int
1918 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1919 {
1920         const char *const counters[] = {
1921                 "collisions",     "rx_frame_errors",   "tx_compressed",
1922                 "multicast",      "rx_length_errors",  "tx_dropped",
1923                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1924                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1925                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1926                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1927                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1928                 "rx_fifo_errors", "tx_carrier_errors",
1929         };
1930         char buf[64];
1931         int stats_dir;
1932         int i;
1933         uint64_t val = 0;
1934
1935         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1936         stats_dir = open(buf, O_DIRECTORY);
1937         if (stats_dir < 0)
1938                 return -1;
1939
1940         for (i = 0; i < ARRAY_SIZE(counters); i++)
1941                 if (read_uint64_file(stats_dir, counters[i], &val))
1942                         blobmsg_add_u64(b, counters[i], val);
1943
1944         close(stats_dir);
1945         return 0;
1946 }
1947
1948 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1949 {
1950         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1951         int alen = v4 ? 4 : 16;
1952         unsigned int flags = 0;
1953         struct ifaddrmsg ifa = {
1954                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1955                 .ifa_prefixlen = addr->mask,
1956                 .ifa_index = dev->ifindex,
1957         };
1958
1959         struct nl_msg *msg;
1960         if (cmd == RTM_NEWADDR)
1961                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1962
1963         msg = nlmsg_alloc_simple(cmd, flags);
1964         if (!msg)
1965                 return -1;
1966
1967         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1968         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1969         if (v4) {
1970                 if (addr->broadcast)
1971                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1972                 if (addr->point_to_point)
1973                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1974         } else {
1975                 time_t now = system_get_rtime();
1976                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1977
1978                 if (addr->preferred_until) {
1979                         int64_t preferred = addr->preferred_until - now;
1980                         if (preferred < 0)
1981                                 preferred = 0;
1982                         else if (preferred > UINT32_MAX)
1983                                 preferred = UINT32_MAX;
1984
1985                         cinfo.ifa_prefered = preferred;
1986                 }
1987
1988                 if (addr->valid_until) {
1989                         int64_t valid = addr->valid_until - now;
1990                         if (valid <= 0) {
1991                                 nlmsg_free(msg);
1992                                 return -1;
1993                         }
1994                         else if (valid > UINT32_MAX)
1995                                 valid = UINT32_MAX;
1996
1997                         cinfo.ifa_valid = valid;
1998                 }
1999
2000                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
2001
2002                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
2003                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
2004         }
2005
2006         return system_rtnl_call(msg);
2007 }
2008
2009 int system_add_address(struct device *dev, struct device_addr *addr)
2010 {
2011         return system_addr(dev, addr, RTM_NEWADDR);
2012 }
2013
2014 int system_del_address(struct device *dev, struct device_addr *addr)
2015 {
2016         return system_addr(dev, addr, RTM_DELADDR);
2017 }
2018
2019 static int system_neigh(struct device *dev, struct device_neighbor *neighbor, int cmd)
2020 {
2021         int alen = ((neighbor->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
2022         unsigned int flags = 0;
2023         struct ndmsg ndm = {
2024                 .ndm_family = (alen == 4) ? AF_INET : AF_INET6,
2025                 .ndm_ifindex = dev->ifindex,
2026                 .ndm_state = NUD_PERMANENT,
2027                 .ndm_flags = (neighbor->proxy ? NTF_PROXY : 0) | (neighbor->router ? NTF_ROUTER : 0),
2028         };
2029         struct nl_msg *msg;
2030
2031         if (cmd == RTM_NEWNEIGH)
2032                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
2033
2034         msg = nlmsg_alloc_simple(cmd, flags);
2035
2036         if (!msg)
2037                 return -1;
2038
2039         nlmsg_append(msg, &ndm, sizeof(ndm), 0);
2040
2041         nla_put(msg, NDA_DST, alen, &neighbor->addr);
2042         if (neighbor->flags & DEVNEIGH_MAC)
2043                 nla_put(msg, NDA_LLADDR, sizeof(neighbor->macaddr), &neighbor->macaddr);
2044
2045
2046         return system_rtnl_call(msg);
2047 }
2048
2049 int system_add_neighbor(struct device *dev, struct device_neighbor *neighbor)
2050 {
2051         return system_neigh(dev, neighbor, RTM_NEWNEIGH);
2052 }
2053
2054 int system_del_neighbor(struct device *dev, struct device_neighbor *neighbor)
2055 {
2056         return system_neigh(dev, neighbor, RTM_DELNEIGH);
2057 }
2058
2059 static int system_rt(struct device *dev, struct device_route *route, int cmd)
2060 {
2061         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
2062         bool have_gw;
2063         unsigned int flags = 0;
2064
2065         if (alen == 4)
2066                 have_gw = !!route->nexthop.in.s_addr;
2067         else
2068                 have_gw = route->nexthop.in6.s6_addr32[0] ||
2069                         route->nexthop.in6.s6_addr32[1] ||
2070                         route->nexthop.in6.s6_addr32[2] ||
2071                         route->nexthop.in6.s6_addr32[3];
2072
2073         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
2074                         ? route->table : RT_TABLE_MAIN;
2075
2076         struct rtmsg rtm = {
2077                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2078                 .rtm_dst_len = route->mask,
2079                 .rtm_src_len = route->sourcemask,
2080                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
2081                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
2082                 .rtm_scope = RT_SCOPE_NOWHERE,
2083                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
2084                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
2085         };
2086         struct nl_msg *msg;
2087
2088         if (cmd == RTM_NEWROUTE) {
2089                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
2090
2091                 if (!dev) { /* Add null-route */
2092                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2093                         rtm.rtm_type = RTN_UNREACHABLE;
2094                 }
2095                 else
2096                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
2097         }
2098
2099         if (route->flags & DEVROUTE_TYPE) {
2100                 rtm.rtm_type = route->type;
2101                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
2102                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
2103                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
2104                                 rtm.rtm_table = RT_TABLE_LOCAL;
2105                 }
2106
2107                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
2108                         rtm.rtm_scope = RT_SCOPE_HOST;
2109                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
2110                                 rtm.rtm_type == RTN_ANYCAST) {
2111                         rtm.rtm_scope = RT_SCOPE_LINK;
2112                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
2113                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY ||
2114                                 rtm.rtm_type == RTN_THROW) {
2115                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2116                         dev = NULL;
2117                 }
2118         }
2119
2120         msg = nlmsg_alloc_simple(cmd, flags);
2121         if (!msg)
2122                 return -1;
2123
2124         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2125
2126         if (route->mask)
2127                 nla_put(msg, RTA_DST, alen, &route->addr);
2128
2129         if (route->sourcemask) {
2130                 if (rtm.rtm_family == AF_INET)
2131                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
2132                 else
2133                         nla_put(msg, RTA_SRC, alen, &route->source);
2134         }
2135
2136         if (route->metric > 0)
2137                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
2138
2139         if (have_gw)
2140                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
2141
2142         if (dev)
2143                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
2144
2145         if (table >= 256)
2146                 nla_put_u32(msg, RTA_TABLE, table);
2147
2148         if (route->flags & DEVROUTE_MTU) {
2149                 struct nlattr *metrics;
2150
2151                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
2152                         goto nla_put_failure;
2153
2154                 nla_put_u32(msg, RTAX_MTU, route->mtu);
2155
2156                 nla_nest_end(msg, metrics);
2157         }
2158
2159         return system_rtnl_call(msg);
2160
2161 nla_put_failure:
2162         nlmsg_free(msg);
2163         return -ENOMEM;
2164 }
2165
2166 int system_add_route(struct device *dev, struct device_route *route)
2167 {
2168         return system_rt(dev, route, RTM_NEWROUTE);
2169 }
2170
2171 int system_del_route(struct device *dev, struct device_route *route)
2172 {
2173         return system_rt(dev, route, RTM_DELROUTE);
2174 }
2175
2176 int system_flush_routes(void)
2177 {
2178         const char *names[] = {
2179                 "/proc/sys/net/ipv4/route/flush",
2180                 "/proc/sys/net/ipv6/route/flush"
2181         };
2182         int fd, i;
2183
2184         for (i = 0; i < ARRAY_SIZE(names); i++) {
2185                 fd = open(names[i], O_WRONLY);
2186                 if (fd < 0)
2187                         continue;
2188
2189                 if (write(fd, "-1", 2)) {}
2190                 close(fd);
2191         }
2192         return 0;
2193 }
2194
2195 bool system_resolve_rt_type(const char *type, unsigned int *id)
2196 {
2197         return system_rtn_aton(type, id);
2198 }
2199
2200 bool system_resolve_rt_proto(const char *type, unsigned int *id)
2201 {
2202         FILE *f;
2203         char *e, buf[128];
2204         unsigned int n, proto = 256;
2205         n = strtoul(type, &e, 0);
2206         if (!*e && e != type)
2207                 proto = n;
2208         else if (!strcmp(type, "unspec"))
2209                 proto = RTPROT_UNSPEC;
2210         else if (!strcmp(type, "kernel"))
2211                 proto = RTPROT_KERNEL;
2212         else if (!strcmp(type, "boot"))
2213                 proto = RTPROT_BOOT;
2214         else if (!strcmp(type, "static"))
2215                 proto = RTPROT_STATIC;
2216         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
2217                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
2218                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2219                                 continue;
2220
2221                         n = strtoul(e, NULL, 10);
2222                         e = strtok(NULL, " \t\n");
2223
2224                         if (e && !strcmp(e, type)) {
2225                                 proto = n;
2226                                 break;
2227                         }
2228                 }
2229                 fclose(f);
2230         }
2231
2232         if (proto > 255)
2233                 return false;
2234
2235         *id = proto;
2236         return true;
2237 }
2238
2239 bool system_resolve_rt_table(const char *name, unsigned int *id)
2240 {
2241         FILE *f;
2242         char *e, buf[128];
2243         unsigned int n, table = RT_TABLE_UNSPEC;
2244
2245         /* first try to parse table as number */
2246         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2247                 table = n;
2248
2249         /* handle well known aliases */
2250         else if (!strcmp(name, "default"))
2251                 table = RT_TABLE_DEFAULT;
2252         else if (!strcmp(name, "main"))
2253                 table = RT_TABLE_MAIN;
2254         else if (!strcmp(name, "local"))
2255                 table = RT_TABLE_LOCAL;
2256
2257         /* try to look up name in /etc/iproute2/rt_tables */
2258         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2259         {
2260                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2261                 {
2262                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2263                                 continue;
2264
2265                         n = strtoul(e, NULL, 10);
2266                         e = strtok(NULL, " \t\n");
2267
2268                         if (e && !strcmp(e, name))
2269                         {
2270                                 table = n;
2271                                 break;
2272                         }
2273                 }
2274
2275                 fclose(f);
2276         }
2277
2278         if (table == RT_TABLE_UNSPEC)
2279                 return false;
2280
2281         *id = table;
2282         return true;
2283 }
2284
2285 bool system_is_default_rt_table(unsigned int id)
2286 {
2287         return (id == RT_TABLE_MAIN);
2288 }
2289
2290 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2291 {
2292         char *e;
2293         unsigned int n;
2294
2295         if (!strcmp(filter, "strict"))
2296                 n = 1;
2297         else if (!strcmp(filter, "loose"))
2298                 n = 2;
2299         else {
2300                 n = strtoul(filter, &e, 0);
2301                 if (*e || e == filter || n > 2)
2302                         return false;
2303         }
2304
2305         *id = n;
2306         return true;
2307 }
2308
2309 static int system_iprule(struct iprule *rule, int cmd)
2310 {
2311         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2312
2313         struct nl_msg *msg;
2314         struct rtmsg rtm = {
2315                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2316                 .rtm_protocol = RTPROT_STATIC,
2317                 .rtm_scope = RT_SCOPE_UNIVERSE,
2318                 .rtm_table = RT_TABLE_UNSPEC,
2319                 .rtm_type = RTN_UNSPEC,
2320                 .rtm_flags = 0,
2321         };
2322
2323         if (cmd == RTM_NEWRULE)
2324                 rtm.rtm_type = RTN_UNICAST;
2325
2326         if (rule->invert)
2327                 rtm.rtm_flags |= FIB_RULE_INVERT;
2328
2329         if (rule->flags & IPRULE_SRC)
2330                 rtm.rtm_src_len = rule->src_mask;
2331
2332         if (rule->flags & IPRULE_DEST)
2333                 rtm.rtm_dst_len = rule->dest_mask;
2334
2335         if (rule->flags & IPRULE_TOS)
2336                 rtm.rtm_tos = rule->tos;
2337
2338         if (rule->flags & IPRULE_LOOKUP) {
2339                 if (rule->lookup < 256)
2340                         rtm.rtm_table = rule->lookup;
2341         }
2342
2343         if (rule->flags & IPRULE_ACTION)
2344                 rtm.rtm_type = rule->action;
2345         else if (rule->flags & IPRULE_GOTO)
2346                 rtm.rtm_type = FR_ACT_GOTO;
2347         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2348                 rtm.rtm_type = FR_ACT_NOP;
2349
2350         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2351
2352         if (!msg)
2353                 return -1;
2354
2355         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2356
2357         if (rule->flags & IPRULE_IN)
2358                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2359
2360         if (rule->flags & IPRULE_OUT)
2361                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2362
2363         if (rule->flags & IPRULE_SRC)
2364                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2365
2366         if (rule->flags & IPRULE_DEST)
2367                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2368
2369         if (rule->flags & IPRULE_PRIORITY)
2370                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2371         else if (cmd == RTM_NEWRULE)
2372                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2373
2374         if (rule->flags & IPRULE_FWMARK)
2375                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2376
2377         if (rule->flags & IPRULE_FWMASK)
2378                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2379
2380         if (rule->flags & IPRULE_LOOKUP) {
2381                 if (rule->lookup >= 256)
2382                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2383         }
2384
2385         if (rule->flags & IPRULE_SUP_PREFIXLEN)
2386                 nla_put_u32(msg, FRA_SUPPRESS_PREFIXLEN, rule->sup_prefixlen);
2387
2388         if (rule->flags & IPRULE_GOTO)
2389                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2390
2391         return system_rtnl_call(msg);
2392 }
2393
2394 int system_add_iprule(struct iprule *rule)
2395 {
2396         return system_iprule(rule, RTM_NEWRULE);
2397 }
2398
2399 int system_del_iprule(struct iprule *rule)
2400 {
2401         return system_iprule(rule, RTM_DELRULE);
2402 }
2403
2404 int system_flush_iprules(void)
2405 {
2406         int rv = 0;
2407         struct iprule rule;
2408
2409         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2410         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2411
2412         memset(&rule, 0, sizeof(rule));
2413
2414
2415         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2416
2417         rule.priority = 0;
2418         rule.lookup = RT_TABLE_LOCAL;
2419         rv |= system_iprule(&rule, RTM_NEWRULE);
2420
2421         rule.priority = 32766;
2422         rule.lookup = RT_TABLE_MAIN;
2423         rv |= system_iprule(&rule, RTM_NEWRULE);
2424
2425         rule.priority = 32767;
2426         rule.lookup = RT_TABLE_DEFAULT;
2427         rv |= system_iprule(&rule, RTM_NEWRULE);
2428
2429
2430         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2431
2432         rule.priority = 0;
2433         rule.lookup = RT_TABLE_LOCAL;
2434         rv |= system_iprule(&rule, RTM_NEWRULE);
2435
2436         rule.priority = 32766;
2437         rule.lookup = RT_TABLE_MAIN;
2438         rv |= system_iprule(&rule, RTM_NEWRULE);
2439
2440         return rv;
2441 }
2442
2443 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2444 {
2445         return system_rtn_aton(action, id);
2446 }
2447
2448 time_t system_get_rtime(void)
2449 {
2450         struct timespec ts;
2451         struct timeval tv;
2452
2453         if (clock_gettime(CLOCK_MONOTONIC, &ts) == 0)
2454                 return ts.tv_sec;
2455
2456         if (gettimeofday(&tv, NULL) == 0)
2457                 return tv.tv_sec;
2458
2459         return 0;
2460 }
2461
2462 #ifndef IP_DF
2463 #define IP_DF       0x4000
2464 #endif
2465
2466 static int tunnel_ioctl(const char *name, int cmd, void *p)
2467 {
2468         struct ifreq ifr;
2469
2470         memset(&ifr, 0, sizeof(ifr));
2471         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name) - 1);
2472         ifr.ifr_ifru.ifru_data = p;
2473         return ioctl(sock_ioctl, cmd, &ifr);
2474 }
2475
2476 #ifdef IFLA_IPTUN_MAX
2477 static int system_add_ip6_tunnel(const char *name, const unsigned int link,
2478                                  struct blob_attr **tb)
2479 {
2480         struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2481                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2482         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2483         struct blob_attr *cur;
2484         int ret = 0, ttl = 0;
2485
2486         if (!nlm)
2487                 return -1;
2488
2489         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2490         nla_put_string(nlm, IFLA_IFNAME, name);
2491
2492         if (link)
2493                 nla_put_u32(nlm, IFLA_LINK, link);
2494
2495         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2496         if (!linkinfo) {
2497                 ret = -ENOMEM;
2498                 goto failure;
2499         }
2500
2501         nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2502         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2503         if (!infodata) {
2504                 ret = -ENOMEM;
2505                 goto failure;
2506         }
2507
2508         if (link)
2509                 nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2510
2511         if ((cur = tb[TUNNEL_ATTR_TTL]))
2512                 ttl = blobmsg_get_u32(cur);
2513
2514         nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2515         nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2516
2517         struct in6_addr in6buf;
2518         if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2519                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2520                         ret = -EINVAL;
2521                         goto failure;
2522                 }
2523                 nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2524         }
2525
2526         if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2527                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2528                         ret = -EINVAL;
2529                         goto failure;
2530                 }
2531                 nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2532         }
2533
2534         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2535                 struct blob_attr *tb_data[__IPIP6_DATA_ATTR_MAX];
2536                 uint32_t tun_flags = IP6_TNL_F_IGN_ENCAP_LIMIT;
2537
2538                 blobmsg_parse(ipip6_data_attr_list.params, __IPIP6_DATA_ATTR_MAX, tb_data,
2539                         blobmsg_data(cur), blobmsg_len(cur));
2540
2541                 if ((cur = tb_data[IPIP6_DATA_ENCAPLIMIT])) {
2542                         char *str = blobmsg_get_string(cur);
2543
2544                         if (strcmp(str, "ignore")) {
2545                                 char *e;
2546                                 unsigned encap_limit = strtoul(str, &e, 0);
2547
2548                                 if (e == str || *e || encap_limit > 255) {
2549                                         ret = -EINVAL;
2550                                         goto failure;
2551                                 }
2552
2553                                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, encap_limit);
2554                                 tun_flags &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2555                         }
2556                 }
2557
2558 #ifdef IFLA_IPTUN_FMR_MAX
2559                 if ((cur = tb_data[IPIP6_DATA_FMRS])) {
2560                         struct blob_attr *rcur;
2561                         unsigned rrem, fmrcnt = 0;
2562                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2563
2564                         if (!fmrs) {
2565                                 ret = -ENOMEM;
2566                                 goto failure;
2567                         }
2568
2569                         blobmsg_for_each_attr(rcur, cur, rrem) {
2570                                 struct blob_attr *tb_fmr[__FMR_DATA_ATTR_MAX], *tb_cur;
2571                                 struct in6_addr ip6prefix;
2572                                 struct in_addr ip4prefix;
2573                                 unsigned ip4len, ip6len, ealen, offset;
2574
2575                                 blobmsg_parse(fmr_data_attr_list.params, __FMR_DATA_ATTR_MAX, tb_fmr,
2576                                                 blobmsg_data(rcur), blobmsg_len(rcur));
2577
2578                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX6]) ||
2579                                                 !parse_ip_and_netmask(AF_INET6,
2580                                                         blobmsg_data(tb_cur), &ip6prefix,
2581                                                         &ip6len)) {
2582                                         ret = -EINVAL;
2583                                         goto failure;
2584                                 }
2585
2586                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX4]) ||
2587                                                 !parse_ip_and_netmask(AF_INET,
2588                                                         blobmsg_data(tb_cur), &ip4prefix,
2589                                                         &ip4len)) {
2590                                         ret = -EINVAL;
2591                                         goto failure;
2592                                 }
2593
2594                                 if (!(tb_cur = tb_fmr[FMR_DATA_EALEN])) {
2595                                         ret = -EINVAL;
2596                                         goto failure;
2597                                 }
2598                                 ealen = blobmsg_get_u32(tb_cur);
2599
2600                                 if (!(tb_cur = tb_fmr[FMR_DATA_OFFSET])) {
2601                                         ret = -EINVAL;
2602                                         goto failure;
2603                                 }
2604                                 offset = blobmsg_get_u32(tb_cur);
2605
2606                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2607                                 if (!rule) {
2608                                         ret = -ENOMEM;
2609                                         goto failure;
2610                                 }
2611
2612                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2613                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2614                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2615                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2616                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2617                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2618
2619                                 nla_nest_end(nlm, rule);
2620                         }
2621
2622                         nla_nest_end(nlm, fmrs);
2623                 }
2624 #endif
2625                 if (tun_flags)
2626                         nla_put_u32(nlm, IFLA_IPTUN_FLAGS, tun_flags);
2627         }
2628
2629         nla_nest_end(nlm, infodata);
2630         nla_nest_end(nlm, linkinfo);
2631
2632         return system_rtnl_call(nlm);
2633
2634 failure:
2635         nlmsg_free(nlm);
2636         return ret;
2637 }
2638 #endif
2639
2640 #ifdef IFLA_IPTUN_MAX
2641 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2642 static int system_add_gre_tunnel(const char *name, const char *kind,
2643                                  const unsigned int link, struct blob_attr **tb, bool v6)
2644 {
2645         struct nl_msg *nlm;
2646         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2647         struct blob_attr *cur;
2648         uint32_t ikey = 0, okey = 0, flowinfo = 0, flags6 = IP6_TNL_F_IGN_ENCAP_LIMIT;
2649         uint16_t iflags = 0, oflags = 0;
2650         uint8_t tos = 0;
2651         int ret = 0, ttl = 0;
2652         unsigned encap_limit = 0;
2653
2654         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2655         if (!nlm)
2656                 return -1;
2657
2658         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2659         nla_put_string(nlm, IFLA_IFNAME, name);
2660
2661         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2662         if (!linkinfo) {
2663                 ret = -ENOMEM;
2664                 goto failure;
2665         }
2666
2667         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2668         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2669         if (!infodata) {
2670                 ret = -ENOMEM;
2671                 goto failure;
2672         }
2673
2674         if (link)
2675                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2676
2677         if ((cur = tb[TUNNEL_ATTR_TTL]))
2678                 ttl = blobmsg_get_u32(cur);
2679
2680         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2681                 char *str = blobmsg_get_string(cur);
2682                 if (strcmp(str, "inherit")) {
2683                         unsigned uval;
2684
2685                         if (!system_tos_aton(str, &uval)) {
2686                                 ret = -EINVAL;
2687                                 goto failure;
2688                         }
2689
2690                         if (v6)
2691                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2692                         else
2693                                 tos = uval;
2694                 } else {
2695                         if (v6)
2696                                 flags6 |= IP6_TNL_F_USE_ORIG_TCLASS;
2697                         else
2698                                 tos = 1;
2699                 }
2700         }
2701
2702         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2703                 struct blob_attr *tb_data[__GRE_DATA_ATTR_MAX];
2704
2705                 blobmsg_parse(gre_data_attr_list.params, __GRE_DATA_ATTR_MAX, tb_data,
2706                         blobmsg_data(cur), blobmsg_len(cur));
2707
2708                 if ((cur = tb_data[GRE_DATA_IKEY])) {
2709                         if ((ikey = blobmsg_get_u32(cur)))
2710                                 iflags |= GRE_KEY;
2711                 }
2712
2713                 if ((cur = tb_data[GRE_DATA_OKEY])) {
2714                         if ((okey = blobmsg_get_u32(cur)))
2715                                 oflags |= GRE_KEY;
2716                 }
2717
2718                 if ((cur = tb_data[GRE_DATA_ICSUM])) {
2719                         if (blobmsg_get_bool(cur))
2720                                 iflags |= GRE_CSUM;
2721                 }
2722
2723                 if ((cur = tb_data[GRE_DATA_OCSUM])) {
2724                         if (blobmsg_get_bool(cur))
2725                                 oflags |= GRE_CSUM;
2726                 }
2727
2728                 if ((cur = tb_data[GRE_DATA_ISEQNO])) {
2729                         if (blobmsg_get_bool(cur))
2730                                 iflags |= GRE_SEQ;
2731                 }
2732
2733                 if ((cur = tb_data[GRE_DATA_OSEQNO])) {
2734                         if (blobmsg_get_bool(cur))
2735                                 oflags |= GRE_SEQ;
2736                 }
2737
2738                 if ((cur = tb_data[GRE_DATA_ENCAPLIMIT])) {
2739                         char *str = blobmsg_get_string(cur);
2740
2741                         if (strcmp(str, "ignore")) {
2742                                 char *e;
2743
2744                                 encap_limit = strtoul(str, &e, 0);
2745
2746                                 if (e == str || *e || encap_limit > 255) {
2747                                         ret = -EINVAL;
2748                                         goto failure;
2749                                 }
2750
2751                                 flags6 &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2752                         }
2753                 }
2754         }
2755
2756         if (v6) {
2757                 struct in6_addr in6buf;
2758                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2759                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2760                                 ret = -EINVAL;
2761                                 goto failure;
2762                         }
2763                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2764                 }
2765
2766                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2767                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2768                                 ret = -EINVAL;
2769                                 goto failure;
2770                         }
2771                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2772                 }
2773
2774                 if (!(flags6 & IP6_TNL_F_IGN_ENCAP_LIMIT))
2775                         nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, encap_limit);
2776
2777                 if (flowinfo)
2778                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2779
2780                 if (flags6)
2781                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags6);
2782
2783                 if (!ttl)
2784                         ttl = 64;
2785         } else {
2786                 struct in_addr inbuf;
2787                 bool set_df = true;
2788
2789                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2790                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2791                                 ret = -EINVAL;
2792                                 goto failure;
2793                         }
2794                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2795                 }
2796
2797                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2798                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2799                                 ret = -EINVAL;
2800                                 goto failure;
2801                         }
2802                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2803
2804                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2805                                 if (!okey) {
2806                                         okey = inbuf.s_addr;
2807                                         oflags |= GRE_KEY;
2808                                 }
2809
2810                                 if (!ikey) {
2811                                         ikey = inbuf.s_addr;
2812                                         iflags |= GRE_KEY;
2813                                 }
2814                         }
2815                 }
2816
2817                 if ((cur = tb[TUNNEL_ATTR_DF]))
2818                         set_df = blobmsg_get_bool(cur);
2819
2820                 if (!set_df) {
2821                         /* ttl != 0 and nopmtudisc are incompatible */
2822                         if (ttl) {
2823                                 ret = -EINVAL;
2824                                 goto failure;
2825                         }
2826                 } else if (!ttl)
2827                         ttl = 64;
2828
2829                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2830
2831                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2832         }
2833
2834         if (ttl)
2835                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2836
2837         if (oflags)
2838                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2839
2840         if (iflags)
2841                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2842
2843         if (okey)
2844                 nla_put_u32(nlm, IFLA_GRE_OKEY, htonl(okey));
2845
2846         if (ikey)
2847                 nla_put_u32(nlm, IFLA_GRE_IKEY, htonl(ikey));
2848
2849         nla_nest_end(nlm, infodata);
2850         nla_nest_end(nlm, linkinfo);
2851
2852         return system_rtnl_call(nlm);
2853
2854 failure:
2855         nlmsg_free(nlm);
2856         return ret;
2857 }
2858 #endif
2859
2860 #ifdef IFLA_VTI_MAX
2861 static int system_add_vti_tunnel(const char *name, const char *kind,
2862                                  const unsigned int link, struct blob_attr **tb, bool v6)
2863 {
2864         struct nl_msg *nlm;
2865         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2866         struct blob_attr *cur;
2867         int ret = 0;
2868
2869         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2870         if (!nlm)
2871                 return -1;
2872
2873         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2874         nla_put_string(nlm, IFLA_IFNAME, name);
2875
2876         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2877         if (!linkinfo) {
2878                 ret = -ENOMEM;
2879                 goto failure;
2880         }
2881
2882         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2883         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2884         if (!infodata) {
2885                 ret = -ENOMEM;
2886                 goto failure;
2887         }
2888
2889         if (link)
2890                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2891
2892         if (v6) {
2893                 struct in6_addr in6buf;
2894                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2895                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2896                                 ret = -EINVAL;
2897                                 goto failure;
2898                         }
2899                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2900                 }
2901
2902                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2903                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2904                                 ret = -EINVAL;
2905                                 goto failure;
2906                         }
2907                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2908                 }
2909
2910         } else {
2911                 struct in_addr inbuf;
2912
2913                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2914                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2915                                 ret = -EINVAL;
2916                                 goto failure;
2917                         }
2918                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2919                 }
2920
2921                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2922                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2923                                 ret = -EINVAL;
2924                                 goto failure;
2925                         }
2926                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2927                 }
2928
2929         }
2930
2931         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2932                 struct blob_attr *tb_data[__VTI_DATA_ATTR_MAX];
2933                 uint32_t ikey = 0, okey = 0;
2934
2935                 blobmsg_parse(vti_data_attr_list.params, __VTI_DATA_ATTR_MAX, tb_data,
2936                         blobmsg_data(cur), blobmsg_len(cur));
2937
2938                 if ((cur = tb_data[VTI_DATA_IKEY])) {
2939                         if ((ikey = blobmsg_get_u32(cur)))
2940                                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2941                 }
2942
2943                 if ((cur = tb_data[VTI_DATA_OKEY])) {
2944                         if ((okey = blobmsg_get_u32(cur)))
2945                                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2946                 }
2947         }
2948
2949         nla_nest_end(nlm, infodata);
2950         nla_nest_end(nlm, linkinfo);
2951
2952         return system_rtnl_call(nlm);
2953
2954 failure:
2955         nlmsg_free(nlm);
2956         return ret;
2957 }
2958 #endif
2959
2960 #ifdef IFLA_XFRM_MAX
2961 static int system_add_xfrm_tunnel(const char *name, const char *kind,
2962                                  const unsigned int link, struct blob_attr **tb)
2963 {
2964         struct nl_msg *nlm;
2965         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2966         struct blob_attr *cur;
2967         int ret = 0;
2968
2969         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2970         if (!nlm)
2971                 return -1;
2972
2973         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2974         nla_put_string(nlm, IFLA_IFNAME, name);
2975
2976         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2977         if (!linkinfo) {
2978                 ret = -ENOMEM;
2979                 goto failure;
2980         }
2981
2982         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2983         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2984         if (!infodata) {
2985                 ret = -ENOMEM;
2986                 goto failure;
2987         }
2988
2989         if (link)
2990                 nla_put_u32(nlm, IFLA_XFRM_LINK, link);
2991
2992         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2993                 struct blob_attr *tb_data[__XFRM_DATA_ATTR_MAX];
2994                 uint32_t if_id = 0;
2995
2996                 blobmsg_parse(xfrm_data_attr_list.params, __XFRM_DATA_ATTR_MAX, tb_data,
2997                         blobmsg_data(cur), blobmsg_len(cur));
2998
2999                 if ((cur = tb_data[XFRM_DATA_IF_ID])) {
3000                         if ((if_id = blobmsg_get_u32(cur)))
3001                                 nla_put_u32(nlm, IFLA_XFRM_IF_ID, if_id);
3002                 }
3003
3004         }
3005
3006         nla_nest_end(nlm, infodata);
3007         nla_nest_end(nlm, linkinfo);
3008
3009         return system_rtnl_call(nlm);
3010
3011 failure:
3012         nlmsg_free(nlm);
3013         return ret;
3014 }
3015 #endif
3016
3017 #ifdef IFLA_VXLAN_MAX
3018 static int system_add_vxlan(const char *name, const unsigned int link, struct blob_attr **tb, bool v6)
3019 {
3020         struct blob_attr *tb_data[__VXLAN_DATA_ATTR_MAX];
3021         struct nl_msg *msg;
3022         struct nlattr *linkinfo, *data;
3023         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
3024         struct blob_attr *cur;
3025         int ret = 0;
3026
3027         if ((cur = tb[TUNNEL_ATTR_DATA]))
3028                 blobmsg_parse(vxlan_data_attr_list.params, __VXLAN_DATA_ATTR_MAX, tb_data,
3029                         blobmsg_data(cur), blobmsg_len(cur));
3030         else
3031                 return -EINVAL;
3032
3033         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
3034
3035         if (!msg)
3036                 return -1;
3037
3038         nlmsg_append(msg, &iim, sizeof(iim), 0);
3039
3040         nla_put_string(msg, IFLA_IFNAME, name);
3041
3042         if ((cur = tb_data[VXLAN_DATA_ATTR_MACADDR])) {
3043                 struct ether_addr *ea = ether_aton(blobmsg_get_string(cur));
3044                 if (!ea) {
3045                         ret = -EINVAL;
3046                         goto failure;
3047                 }
3048
3049                 nla_put(msg, IFLA_ADDRESS, ETH_ALEN, ea);
3050         }
3051
3052         if ((cur = tb[TUNNEL_ATTR_MTU])) {
3053                 uint32_t mtu = blobmsg_get_u32(cur);
3054                 nla_put_u32(msg, IFLA_MTU, mtu);
3055         }
3056
3057         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO))) {
3058                 ret = -ENOMEM;
3059                 goto failure;
3060         }
3061
3062         nla_put_string(msg, IFLA_INFO_KIND, "vxlan");
3063
3064         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA))) {
3065                 ret = -ENOMEM;
3066                 goto failure;
3067         }
3068
3069         if (link)
3070                 nla_put_u32(msg, IFLA_VXLAN_LINK, link);
3071
3072         if ((cur = tb_data[VXLAN_DATA_ATTR_ID])) {
3073                 uint32_t id = blobmsg_get_u32(cur);
3074                 if (id >= (1u << 24) - 1) {
3075                         ret = -EINVAL;
3076                         goto failure;
3077                 }
3078
3079                 nla_put_u32(msg, IFLA_VXLAN_ID, id);
3080         }
3081
3082         if (v6) {
3083                 struct in6_addr in6buf;
3084                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3085                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3086                                 ret = -EINVAL;
3087                                 goto failure;
3088                         }
3089                         nla_put(msg, IFLA_VXLAN_LOCAL6, sizeof(in6buf), &in6buf);
3090                 }
3091
3092                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3093                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3094                                 ret = -EINVAL;
3095                                 goto failure;
3096                         }
3097                         nla_put(msg, IFLA_VXLAN_GROUP6, sizeof(in6buf), &in6buf);
3098                 }
3099         } else {
3100                 struct in_addr inbuf;
3101
3102                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3103                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3104                                 ret = -EINVAL;
3105                                 goto failure;
3106                         }
3107                         nla_put(msg, IFLA_VXLAN_LOCAL, sizeof(inbuf), &inbuf);
3108                 }
3109
3110                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3111                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3112                                 ret = -EINVAL;
3113                                 goto failure;
3114                         }
3115                         nla_put(msg, IFLA_VXLAN_GROUP, sizeof(inbuf), &inbuf);
3116                 }
3117         }
3118
3119         uint32_t port = 4789;
3120         if ((cur = tb_data[VXLAN_DATA_ATTR_PORT])) {
3121                 port = blobmsg_get_u32(cur);
3122                 if (port < 1 || port > 65535) {
3123                         ret = -EINVAL;
3124                         goto failure;
3125                 }
3126         }
3127         nla_put_u16(msg, IFLA_VXLAN_PORT, htons(port));
3128
3129         if ((cur = tb_data[VXLAN_DATA_ATTR_RXCSUM])) {
3130                 bool rxcsum = blobmsg_get_bool(cur);
3131                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_RX, !rxcsum);
3132         }
3133
3134         if ((cur = tb_data[VXLAN_DATA_ATTR_TXCSUM])) {
3135                 bool txcsum = blobmsg_get_bool(cur);
3136                 nla_put_u8(msg, IFLA_VXLAN_UDP_CSUM, txcsum);
3137                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_TX, !txcsum);
3138         }
3139
3140         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3141                 char *str = blobmsg_get_string(cur);
3142                 unsigned tos = 1;
3143
3144                 if (strcmp(str, "inherit")) {
3145                         if (!system_tos_aton(str, &tos)) {
3146                                 ret = -EINVAL;
3147                                 goto failure;
3148                         }
3149                 }
3150
3151                 nla_put_u8(msg, IFLA_VXLAN_TOS, tos);
3152         }
3153
3154         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3155                 uint32_t ttl = blobmsg_get_u32(cur);
3156                 if (ttl < 1 || ttl > 255) {
3157                         ret = -EINVAL;
3158                         goto failure;
3159                 }
3160
3161                 nla_put_u8(msg, IFLA_VXLAN_TTL, ttl);
3162         }
3163
3164         nla_nest_end(msg, data);
3165         nla_nest_end(msg, linkinfo);
3166
3167         ret = system_rtnl_call(msg);
3168         if (ret)
3169                 D(SYSTEM, "Error adding vxlan '%s': %d\n", name, ret);
3170
3171         return ret;
3172
3173 failure:
3174         nlmsg_free(msg);
3175         return ret;
3176 }
3177 #endif
3178
3179 static int system_add_sit_tunnel(const char *name, const unsigned int link, struct blob_attr **tb)
3180 {
3181         struct blob_attr *cur;
3182         int ret = 0;
3183
3184         if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
3185                 return -1;
3186
3187 #ifdef SIOCADD6RD
3188         if ((cur = tb[TUNNEL_ATTR_DATA])) {
3189                 struct blob_attr *tb_data[__SIXRD_DATA_ATTR_MAX];
3190                 unsigned int mask;
3191                 struct ip_tunnel_6rd p6;
3192
3193                 blobmsg_parse(sixrd_data_attr_list.params, __SIXRD_DATA_ATTR_MAX, tb_data,
3194                         blobmsg_data(cur), blobmsg_len(cur));
3195
3196                 memset(&p6, 0, sizeof(p6));
3197
3198                 if ((cur = tb_data[SIXRD_DATA_PREFIX])) {
3199                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
3200                                                 &p6.prefix, &mask) || mask > 128) {
3201                                 ret = -EINVAL;
3202                                 goto failure;
3203                         }
3204
3205                         p6.prefixlen = mask;
3206                 }
3207
3208                 if ((cur = tb_data[SIXRD_DATA_RELAY_PREFIX])) {
3209                         if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
3210                                                 &p6.relay_prefix, &mask) || mask > 32) {
3211                                 ret = -EINVAL;
3212                                 goto failure;
3213                         }
3214
3215                         p6.relay_prefixlen = mask;
3216                 }
3217
3218                 if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
3219                         ret = -1;
3220                         goto failure;
3221                 }
3222         }
3223 #endif
3224
3225         return ret;
3226
3227 failure:
3228         __system_del_ip_tunnel(name, tb);
3229         return ret;
3230 }
3231
3232 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
3233 {
3234         struct blob_attr *cur;
3235         bool set_df = true;
3236         struct ip_tunnel_parm p  = {
3237                 .link = link,
3238                 .iph = {
3239                         .version = 4,
3240                         .ihl = 5,
3241                         .protocol = proto,
3242                 }
3243         };
3244
3245         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
3246                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
3247                 return -EINVAL;
3248
3249         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
3250                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
3251                 return -EINVAL;
3252
3253         if ((cur = tb[TUNNEL_ATTR_DF]))
3254                 set_df = blobmsg_get_bool(cur);
3255
3256         if ((cur = tb[TUNNEL_ATTR_TTL]))
3257                 p.iph.ttl = blobmsg_get_u32(cur);
3258
3259         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3260                 char *str = blobmsg_get_string(cur);
3261                 if (strcmp(str, "inherit")) {
3262                         unsigned uval;
3263
3264                         if (!system_tos_aton(str, &uval))
3265                                 return -EINVAL;
3266
3267                         p.iph.tos = uval;
3268                 } else
3269                         p.iph.tos = 1;
3270         }
3271
3272         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
3273         /* ttl !=0 and nopmtudisc are incompatible */
3274         if (p.iph.ttl && p.iph.frag_off == 0)
3275                 return -EINVAL;
3276
3277         strncpy(p.name, name, sizeof(p.name) - 1);
3278
3279         switch (p.iph.protocol) {
3280         case IPPROTO_IPIP:
3281                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
3282         case IPPROTO_IPV6:
3283                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
3284         default:
3285                 break;
3286         }
3287         return -1;
3288 }
3289
3290 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
3291 {
3292         struct blob_attr *cur;
3293         const char *str;
3294
3295         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3296                 return -EINVAL;
3297         str = blobmsg_data(cur);
3298
3299         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
3300             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
3301             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6") ||
3302             !strcmp(str, "vxlan") || !strcmp(str, "vxlan6") ||
3303             !strcmp(str, "xfrm"))
3304                 return system_link_del(name);
3305         else
3306                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
3307 }
3308
3309 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
3310 {
3311         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3312
3313         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3314                 blob_data(attr), blob_len(attr));
3315
3316         return __system_del_ip_tunnel(name, tb);
3317 }
3318
3319 int system_update_ipv6_mtu(struct device *dev, int mtu)
3320 {
3321         int ret = -1;
3322         char buf[64];
3323         int fd;
3324
3325         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
3326                         dev->ifname);
3327
3328         fd = open(buf, O_RDWR);
3329         if (fd < 0)
3330                 return ret;
3331
3332         if (!mtu) {
3333                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
3334                 if (len < 0)
3335                         goto out;
3336
3337                 buf[len] = 0;
3338                 ret = atoi(buf);
3339         } else {
3340                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
3341                         ret = mtu;
3342         }
3343
3344 out:
3345         close(fd);
3346         return ret;
3347 }
3348
3349 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
3350 {
3351         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3352         struct blob_attr *cur;
3353         const char *str;
3354
3355         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3356                 blob_data(attr), blob_len(attr));
3357
3358         __system_del_ip_tunnel(name, tb);
3359
3360         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3361                 return -EINVAL;
3362         str = blobmsg_data(cur);
3363
3364         unsigned int ttl = 0;
3365         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3366                 ttl = blobmsg_get_u32(cur);
3367                 if (ttl > 255)
3368                         return -EINVAL;
3369         }
3370
3371         unsigned int link = 0;
3372         if ((cur = tb[TUNNEL_ATTR_LINK])) {
3373                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
3374                 if (!iface)
3375                         return -EINVAL;
3376
3377                 if (iface->l3_dev.dev)
3378                         link = iface->l3_dev.dev->ifindex;
3379         }
3380
3381         if (!strcmp(str, "sit"))
3382                 return system_add_sit_tunnel(name, link, tb);
3383 #ifdef IFLA_IPTUN_MAX
3384         else if (!strcmp(str, "ipip6")) {
3385                 return system_add_ip6_tunnel(name, link, tb);
3386         } else if (!strcmp(str, "greip")) {
3387                 return system_add_gre_tunnel(name, "gre", link, tb, false);
3388         } else if (!strcmp(str, "gretapip"))  {
3389                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
3390         } else if (!strcmp(str, "greip6")) {
3391                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
3392         } else if (!strcmp(str, "gretapip6")) {
3393                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
3394 #ifdef IFLA_VTI_MAX
3395         } else if (!strcmp(str, "vtiip")) {
3396                 return system_add_vti_tunnel(name, "vti", link, tb, false);
3397         } else if (!strcmp(str, "vtiip6")) {
3398                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
3399 #endif
3400 #ifdef IFLA_XFRM_MAX
3401         } else if (!strcmp(str, "xfrm")) {
3402                 return system_add_xfrm_tunnel(name, "xfrm", link, tb);
3403 #endif
3404 #ifdef IFLA_VXLAN_MAX
3405         } else if(!strcmp(str, "vxlan")) {
3406                 return system_add_vxlan(name, link, tb, false);
3407         } else if(!strcmp(str, "vxlan6")) {
3408                 return system_add_vxlan(name, link, tb, true);
3409 #endif
3410 #endif
3411         } else if (!strcmp(str, "ipip")) {
3412                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
3413         }
3414         else
3415                 return -EINVAL;
3416
3417         return 0;
3418 }