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         if (dev->type == &simple_device_type && !system_if_force_external(dev->ifname))
 596                 device_set_present(dev, true);
 597
 598         device_set_link(dev, link_state ? true : false);
 599
 600 out:
 601         return 0;
 602 }
 603
 604 static void
 605 handle_hotplug_msg(char *data, int size)
 606 {
 607         const char *subsystem = NULL, *interface = NULL, *interface_old = NULL;
 608         char *cur, *end, *sep;
 609         struct device *dev;
 610         int skip;
 611         bool add, move = false;
 612
 613         if (!strncmp(data, "add@", 4))
 614                 add = true;
 615         else if (!strncmp(data, "remove@", 7))
 616                 add = false;
 617         else if (!strncmp(data, "move@", 5)) {
 618                 add = true;
 619                 move = true;
 620         }
 621         else
 622                 return;
 623
 624         skip = strlen(data) + 1;
 625         end = data + size;
 626
 627         for (cur = data + skip; cur < end; cur += skip) {
 628                 skip = strlen(cur) + 1;
 629
 630                 sep = strchr(cur, '=');
 631                 if (!sep)
 632                         continue;
 633
 634                 *sep = 0;
 635                 if (!strcmp(cur, "INTERFACE"))
 636                         interface = sep + 1;
 637                 else if (!strcmp(cur, "SUBSYSTEM")) {
 638                         subsystem = sep + 1;
 639                         if (strcmp(subsystem, "net") != 0)
 640                                 return;
 641                 } else if (!strcmp(cur, "DEVPATH_OLD")) {
 642                         interface_old = strrchr(sep + 1, '/');
 643                         if (interface_old)
 644                                 interface_old++;
 645                 }
 646         }
 647
 648         if (subsystem && interface) {
 649                 if (move && interface_old)
 650                         goto move;
 651                 else
 652                         goto found;
 653         }
 654
 655         return;
 656
 657 move:
 658         dev = device_find(interface_old);
 659         if (!dev)
 660                 return;
 661
 662         if (dev->type != &simple_device_type)
 663                 goto found;
 664
 665         device_set_present(dev, false);
 666
 667         return;
 668
 669 found:
 670         dev = device_find(interface);
 671         if (!dev)
 672                 return;
 673
 674         if (dev->type != &simple_device_type)
 675                 return;
 676
 677         if (add && system_if_force_external(dev->ifname))
 678                 return;
 679
 680         device_set_present(dev, add);
 681 }
 682
 683 static void
 684 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 685 {
 686         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 687         struct sockaddr_nl nla;
 688         unsigned char *buf = NULL;
 689         int size;
 690         int err;
 691         socklen_t errlen = sizeof(err);
 692
 693         if (!u->error) {
 694                 while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 695                         if (nla.nl_pid == 0)
 696                                 handle_hotplug_msg((char *) buf, size);
 697
 698                         free(buf);
 699                 }
 700                 return;
 701         }
 702
 703         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 704                 goto abort;
 705
 706         switch(err) {
 707         case ENOBUFS:
 708                 /* Increase rx buffer size on netlink socket */
 709                 ev->bufsize *= 2;
 710                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 711                         goto abort;
 712                 break;
 713
 714         default:
 715                 goto abort;
 716         }
 717         u->error = false;
 718         return;
 719
 720 abort:
 721         uloop_fd_delete(&ev->uloop);
 722         return;
 723 }
 724
 725 static int system_rtnl_call(struct nl_msg *msg)
 726 {
 727         int ret;
 728
 729         ret = nl_send_auto_complete(sock_rtnl, msg);
 730         nlmsg_free(msg);
 731
 732         if (ret < 0)
 733                 return ret;
 734
 735         return nl_wait_for_ack(sock_rtnl);
 736 }
 737
 738 int system_bridge_delbr(struct device *bridge)
 739 {
 740         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 741 }
 742
 743 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 744 {
 745         struct ifreq ifr;
 746
 747         memset(&ifr, 0, sizeof(ifr));
 748         if (dev)
 749                 ifr.ifr_ifindex = dev->ifindex;
 750         else
 751                 ifr.ifr_data = data;
 752         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name) - 1);
 753         return ioctl(sock_ioctl, cmd, &ifr);
 754 }
 755
 756 static bool system_is_bridge(const char *name, char *buf, int buflen)
 757 {
 758         struct stat st;
 759
 760         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 761         if (stat(buf, &st) < 0)
 762                 return false;
 763
 764         return true;
 765 }
 766
 767 static char *system_get_bridge(const char *name, char *buf, int buflen)
 768 {
 769         char *path;
 770         ssize_t len = -1;
 771         glob_t gl;
 772
 773         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 774         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 775                 return NULL;
 776
 777         if (gl.gl_pathc > 0)
 778                 len = readlink(gl.gl_pathv[0], buf, buflen);
 779
 780         globfree(&gl);
 781
 782         if (len < 0)
 783                 return NULL;
 784
 785         buf[len] = 0;
 786         path = strrchr(buf, '/');
 787         if (!path)
 788                 return NULL;
 789
 790         return path + 1;
 791 }
 792
 793 static void
 794 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 795 {
 796         bool mcast_to_ucast = dev->wireless_ap;
 797         bool hairpin = true;
 798
 799         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 800             !bridge->settings.multicast_to_unicast)
 801                 mcast_to_ucast = false;
 802
 803         if (!mcast_to_ucast || dev->wireless_isolate)
 804                 hairpin = false;
 805
 806         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 807         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 808 }
 809
 810 int system_bridge_addif(struct device *bridge, struct device *dev)
 811 {
 812         char buf[64];
 813         char *oldbr;
 814         int ret = 0;
 815
 816         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 817         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 818                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 819
 820         if (dev->wireless)
 821                 system_bridge_set_wireless(bridge, dev);
 822
 823         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 824                 snprintf(buf, sizeof(buf), "%u", dev->settings.multicast_router);
 825                 system_bridge_set_multicast_router(dev, buf, false);
 826         }
 827
 828         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 829             dev->settings.multicast_fast_leave)
 830                 system_bridge_set_multicast_fast_leave(dev, "1");
 831
 832         if (dev->settings.flags & DEV_OPT_LEARNING &&
 833             !dev->settings.learning)
 834                 system_bridge_set_learning(dev, "0");
 835
 836         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 837             !dev->settings.unicast_flood)
 838                 system_bridge_set_unicast_flood(dev, "0");
 839
 840         if (dev->settings.flags & DEV_OPT_ISOLATE &&
 841             dev->settings.isolate)
 842                 system_bridge_set_isolated(dev, "1");
 843
 844         return ret;
 845 }
 846
 847 int system_bridge_delif(struct device *bridge, struct device *dev)
 848 {
 849         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 850 }
 851
 852 int system_if_resolve(struct device *dev)
 853 {
 854         struct ifreq ifr;
 855         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
 856         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 857                 return ifr.ifr_ifindex;
 858         else
 859                 return 0;
 860 }
 861
 862 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 863 {
 864         struct ifreq ifr;
 865
 866         memset(&ifr, 0, sizeof(ifr));
 867         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
 868         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) < 0)
 869                 return -1;
 870
 871         ifr.ifr_flags |= add;
 872         ifr.ifr_flags &= ~rem;
 873         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 874 }
 875
 876 struct clear_data {
 877         struct nl_msg *msg;
 878         struct device *dev;
 879         int type;
 880         int size;
 881         int af;
 882 };
 883
 884
 885 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 886 {
 887         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 888
 889         return ifa->ifa_index == ifindex;
 890 }
 891
 892 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 893 {
 894         struct rtmsg *r = NLMSG_DATA(hdr);
 895         struct nlattr *tb[__RTA_MAX];
 896
 897         if (r->rtm_protocol == RTPROT_KERNEL &&
 898             r->rtm_family == AF_INET6)
 899                 return false;
 900
 901         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 902         if (!tb[RTA_OIF])
 903                 return false;
 904
 905         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 906 }
 907
 908 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 909 {
 910         return true;
 911 }
 912
 913 static int cb_clear_event(struct nl_msg *msg, void *arg)
 914 {
 915         struct clear_data *clr = arg;
 916         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 917         bool (*cb)(struct nlmsghdr *, int ifindex);
 918         int type, ret;
 919
 920         switch(clr->type) {
 921         case RTM_GETADDR:
 922                 type = RTM_DELADDR;
 923                 if (hdr->nlmsg_type != RTM_NEWADDR)
 924                         return NL_SKIP;
 925
 926                 cb = check_ifaddr;
 927                 break;
 928         case RTM_GETROUTE:
 929                 type = RTM_DELROUTE;
 930                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 931                         return NL_SKIP;
 932
 933                 cb = check_route;
 934                 break;
 935         case RTM_GETRULE:
 936                 type = RTM_DELRULE;
 937                 if (hdr->nlmsg_type != RTM_NEWRULE)
 938                         return NL_SKIP;
 939
 940                 cb = check_rule;
 941                 break;
 942         default:
 943                 return NL_SKIP;
 944         }
 945
 946         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 947                 return NL_SKIP;
 948
 949         if (type == RTM_DELRULE)
 950                 D(SYSTEM, "Remove a rule\n");
 951         else
 952                 D(SYSTEM, "Remove %s from device %s\n",
 953                   type == RTM_DELADDR ? "an address" : "a route",
 954                   clr->dev->ifname);
 955
 956         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 957         hdr = nlmsg_hdr(clr->msg);
 958         hdr->nlmsg_type = type;
 959         hdr->nlmsg_flags = NLM_F_REQUEST;
 960
 961         nl_socket_disable_auto_ack(sock_rtnl);
 962         ret = nl_send_auto_complete(sock_rtnl, clr->msg);
 963         if (ret < 0) {
 964                 if (type == RTM_DELRULE)
 965                         D(SYSTEM, "Error deleting a rule: %d\n", ret);
 966                 else
 967                         D(SYSTEM, "Error deleting %s from device '%s': %d\n",
 968                                 type == RTM_DELADDR ? "an address" : "a route",
 969                                 clr->dev->ifname, ret);
 970         }
 971
 972         nl_socket_enable_auto_ack(sock_rtnl);
 973
 974         return NL_SKIP;
 975 }
 976
 977 static int
 978 cb_finish_event(struct nl_msg *msg, void *arg)
 979 {
 980         int *pending = arg;
 981         *pending = 0;
 982         return NL_STOP;
 983 }
 984
 985 static int
 986 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 987 {
 988         int *pending = arg;
 989         *pending = err->error;
 990         return NL_STOP;
 991 }
 992
 993 static void
 994 system_if_clear_entries(struct device *dev, int type, int af)
 995 {
 996         struct clear_data clr;
 997         struct nl_cb *cb;
 998         struct rtmsg rtm = {
 999                 .rtm_family = af,
1000                 .rtm_flags = RTM_F_CLONED,
1001         };
1002         int flags = NLM_F_DUMP;
1003         int pending = 1;
1004
1005         clr.af = af;
1006         clr.dev = dev;
1007         clr.type = type;
1008         switch (type) {
1009         case RTM_GETADDR:
1010         case RTM_GETRULE:
1011                 clr.size = sizeof(struct rtgenmsg);
1012                 break;
1013         case RTM_GETROUTE:
1014                 clr.size = sizeof(struct rtmsg);
1015                 break;
1016         default:
1017                 return;
1018         }
1019
1020         cb = nl_cb_alloc(NL_CB_DEFAULT);
1021         if (!cb)
1022                 return;
1023
1024         clr.msg = nlmsg_alloc_simple(type, flags);
1025         if (!clr.msg)
1026                 goto out;
1027
1028         nlmsg_append(clr.msg, &rtm, clr.size, 0);
1029         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
1030         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
1031         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
1032
1033         if (nl_send_auto_complete(sock_rtnl, clr.msg) < 0)
1034                 goto free;
1035
1036         while (pending > 0)
1037                 nl_recvmsgs(sock_rtnl, cb);
1038
1039 free:
1040         nlmsg_free(clr.msg);
1041 out:
1042         nl_cb_put(cb);
1043 }
1044
1045 /*
1046  * Clear bridge (membership) state and bring down device
1047  */
1048 void system_if_clear_state(struct device *dev)
1049 {
1050         static char buf[256];
1051         char *bridge;
1052         device_set_ifindex(dev, system_if_resolve(dev));
1053
1054         if (dev->external || !dev->ifindex)
1055                 return;
1056
1057         system_if_flags(dev->ifname, 0, IFF_UP);
1058
1059         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
1060                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
1061                 system_bridge_delbr(dev);
1062                 return;
1063         }
1064
1065         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
1066         if (bridge) {
1067                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
1068                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
1069         }
1070
1071         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
1072         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
1073         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
1074         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
1075         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET);
1076         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET6);
1077         system_set_disable_ipv6(dev, "0");
1078 }
1079
1080 static inline unsigned long
1081 sec_to_jiffies(int val)
1082 {
1083         return (unsigned long) val * 100;
1084 }
1085
1086 static void system_bridge_conf_multicast_deps(struct device *bridge,
1087                                               struct bridge_config *cfg,
1088                                               char *buf,
1089                                               int buf_len)
1090 {
1091         int val;
1092
1093         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
1094             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
1095             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1096                 val = cfg->robustness * cfg->query_interval +
1097                         cfg->query_response_interval;
1098
1099                 snprintf(buf, buf_len, "%i", val);
1100                 system_bridge_set_membership_interval(bridge, buf);
1101
1102                 val = cfg->robustness * cfg->query_interval +
1103                         cfg->query_response_interval / 2;
1104
1105                 snprintf(buf, buf_len, "%i", val);
1106                 system_bridge_set_other_querier_timeout(bridge, buf);
1107         }
1108
1109         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1110                 val = cfg->query_interval / 4;
1111
1112                 snprintf(buf, buf_len, "%i", val);
1113                 system_bridge_set_startup_query_interval(bridge, buf);
1114         }
1115 }
1116
1117 static void system_bridge_conf_multicast(struct device *bridge,
1118                                          struct bridge_config *cfg,
1119                                          char *buf,
1120                                          int buf_len)
1121 {
1122         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
1123                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
1124
1125         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
1126                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
1127
1128         snprintf(buf, buf_len, "%i", cfg->hash_max);
1129         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
1130                 bridge->ifname, buf);
1131
1132         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
1133                 snprintf(buf, buf_len, "%u", bridge->settings.multicast_router);
1134                 system_bridge_set_multicast_router(bridge, buf, true);
1135         }
1136
1137         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
1138                 snprintf(buf, buf_len, "%i", cfg->robustness);
1139                 system_bridge_set_robustness(bridge, buf);
1140         }
1141
1142         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1143                 snprintf(buf, buf_len, "%i", cfg->query_interval);
1144                 system_bridge_set_query_interval(bridge, buf);
1145         }
1146
1147         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1148                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
1149                 system_bridge_set_query_response_interval(bridge, buf);
1150         }
1151
1152         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1153                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1154                 system_bridge_set_last_member_interval(bridge, buf);
1155         }
1156
1157         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1158 }
1159
1160 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1161 {
1162         char buf[64];
1163
1164         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1165                 return -1;
1166
1167         system_bridge_set_stp_state(bridge, cfg->stp ? "1" : "0");
1168
1169         snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->forward_delay));
1170         system_bridge_set_forward_delay(bridge, buf);
1171
1172         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1173
1174         snprintf(buf, sizeof(buf), "%d", cfg->priority);
1175         system_bridge_set_priority(bridge, buf);
1176
1177         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1178                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->ageing_time));
1179                 system_bridge_set_ageing_time(bridge, buf);
1180         }
1181
1182         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1183                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->hello_time));
1184                 system_bridge_set_hello_time(bridge, buf);
1185         }
1186
1187         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1188                 snprintf(buf, sizeof(buf), "%lu", sec_to_jiffies(cfg->max_age));
1189                 system_bridge_set_max_age(bridge, buf);
1190         }
1191
1192         return 0;
1193 }
1194
1195 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1196 {
1197         struct nl_msg *msg;
1198         struct nlattr *linkinfo, *data;
1199         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1200         int i, rv;
1201         static const struct {
1202                 const char *name;
1203                 enum macvlan_mode val;
1204         } modes[] = {
1205                 { "private", MACVLAN_MODE_PRIVATE },
1206                 { "vepa", MACVLAN_MODE_VEPA },
1207                 { "bridge", MACVLAN_MODE_BRIDGE },
1208                 { "passthru", MACVLAN_MODE_PASSTHRU },
1209         };
1210
1211         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1212
1213         if (!msg)
1214                 return -1;
1215
1216         nlmsg_append(msg, &iim, sizeof(iim), 0);
1217
1218         if (cfg->flags & MACVLAN_OPT_MACADDR)
1219                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1220         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1221         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1222
1223         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1224                 goto nla_put_failure;
1225
1226         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1227
1228         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1229                 goto nla_put_failure;
1230
1231         if (cfg->mode) {
1232                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1233                         if (strcmp(cfg->mode, modes[i].name) != 0)
1234                                 continue;
1235
1236                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1237                         break;
1238                 }
1239         }
1240
1241         nla_nest_end(msg, data);
1242         nla_nest_end(msg, linkinfo);
1243
1244         rv = system_rtnl_call(msg);
1245         if (rv)
1246                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1247
1248         return rv;
1249
1250 nla_put_failure:
1251         nlmsg_free(msg);
1252         return -ENOMEM;
1253 }
1254
1255 int system_link_netns_move(struct device *dev, int netns_fd, const char *target_ifname)
1256 {
1257         struct nl_msg *msg;
1258         struct ifinfomsg iim = {
1259                 .ifi_family = AF_UNSPEC,
1260         };
1261
1262         if (!dev)
1263                 return -1;
1264
1265         iim.ifi_index = system_if_resolve(dev);
1266         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST);
1267
1268         if (!msg)
1269                 return -1;
1270
1271         nlmsg_append(msg, &iim, sizeof(iim), 0);
1272         if (target_ifname)
1273                 nla_put_string(msg, IFLA_IFNAME, target_ifname);
1274
1275         nla_put_u32(msg, IFLA_NET_NS_FD, netns_fd);
1276         return system_rtnl_call(msg);
1277 }
1278
1279 static int system_link_del(const char *ifname)
1280 {
1281         struct nl_msg *msg;
1282         struct ifinfomsg iim = {
1283                 .ifi_family = AF_UNSPEC,
1284                 .ifi_index = 0,
1285         };
1286
1287         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1288
1289         if (!msg)
1290                 return -1;
1291
1292         nlmsg_append(msg, &iim, sizeof(iim), 0);
1293         nla_put_string(msg, IFLA_IFNAME, ifname);
1294         return system_rtnl_call(msg);
1295 }
1296
1297 int system_macvlan_del(struct device *macvlan)
1298 {
1299         return system_link_del(macvlan->ifname);
1300 }
1301
1302 int system_netns_open(const pid_t target_ns)
1303 {
1304         char pid_net_path[PATH_MAX];
1305
1306         snprintf(pid_net_path, sizeof(pid_net_path), "/proc/%u/ns/net", target_ns);
1307
1308         return open(pid_net_path, O_RDONLY);
1309 }
1310
1311 int system_netns_set(int netns_fd)
1312 {
1313         return setns(netns_fd, CLONE_NEWNET);
1314 }
1315
1316 int system_veth_add(struct device *veth, struct veth_config *cfg)
1317 {
1318         struct nl_msg *msg;
1319         struct ifinfomsg empty_iim = {};
1320         struct nlattr *linkinfo, *data, *veth_info;
1321         int rv;
1322
1323         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1324
1325         if (!msg)
1326                 return -1;
1327
1328         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1329
1330         if (cfg->flags & VETH_OPT_MACADDR)
1331                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1332         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1333
1334         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1335                 goto nla_put_failure;
1336
1337         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1338
1339         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1340                 goto nla_put_failure;
1341
1342         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1343                 goto nla_put_failure;
1344
1345         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1346
1347         if (cfg->flags & VETH_OPT_PEER_NAME)
1348                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1349         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1350                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1351
1352         nla_nest_end(msg, veth_info);
1353         nla_nest_end(msg, data);
1354         nla_nest_end(msg, linkinfo);
1355
1356         rv = system_rtnl_call(msg);
1357         if (rv) {
1358                 if (cfg->flags & VETH_OPT_PEER_NAME)
1359                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1360                 else
1361                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1362         }
1363
1364         return rv;
1365
1366 nla_put_failure:
1367         nlmsg_free(msg);
1368         return -ENOMEM;
1369 }
1370
1371 int system_veth_del(struct device *veth)
1372 {
1373         return system_link_del(veth->ifname);
1374 }
1375
1376 static int system_vlan(struct device *dev, int id)
1377 {
1378         struct vlan_ioctl_args ifr = {
1379                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1380                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1381         };
1382
1383         if (ioctl(sock_ioctl, SIOCSIFVLAN, &ifr) < 0)
1384                 return -1;
1385
1386         if (id < 0) {
1387                 ifr.cmd = DEL_VLAN_CMD;
1388                 ifr.u.VID = 0;
1389         } else {
1390                 ifr.cmd = ADD_VLAN_CMD;
1391                 ifr.u.VID = id;
1392         }
1393         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1394         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1395 }
1396
1397 int system_vlan_add(struct device *dev, int id)
1398 {
1399         return system_vlan(dev, id);
1400 }
1401
1402 int system_vlan_del(struct device *dev)
1403 {
1404         return system_vlan(dev, -1);
1405 }
1406
1407 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1408 {
1409         struct nl_msg *msg;
1410         struct nlattr *linkinfo, *data, *qos;
1411         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1412         struct vlan_qos_mapping *elem;
1413         struct ifla_vlan_qos_mapping nl_qos_map;
1414         int rv;
1415
1416         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1417
1418         if (!msg)
1419                 return -1;
1420
1421         nlmsg_append(msg, &iim, sizeof(iim), 0);
1422         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1423         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1424
1425         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1426                 goto nla_put_failure;
1427
1428         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1429
1430         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1431                 goto nla_put_failure;
1432
1433         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1434
1435 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1436         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1437 #else
1438         if(cfg->proto == VLAN_PROTO_8021AD)
1439                 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);
1440 #endif
1441
1442         if (!(qos = nla_nest_start(msg, IFLA_VLAN_INGRESS_QOS)))
1443                 goto nla_put_failure;
1444
1445         vlist_simple_for_each_element(&cfg->ingress_qos_mapping_list, elem, node) {
1446                 nl_qos_map.from = elem->from;
1447                 nl_qos_map.to = elem->to;
1448                 nla_put(msg, IFLA_VLAN_QOS_MAPPING, sizeof(nl_qos_map), &nl_qos_map);
1449         }
1450         nla_nest_end(msg, qos);
1451
1452         if (!(qos = nla_nest_start(msg, IFLA_VLAN_EGRESS_QOS)))
1453                 goto nla_put_failure;
1454
1455         vlist_simple_for_each_element(&cfg->egress_qos_mapping_list, elem, node) {
1456                 nl_qos_map.from = elem->from;
1457                 nl_qos_map.to = elem->to;
1458                 nla_put(msg, IFLA_VLAN_QOS_MAPPING, sizeof(nl_qos_map), &nl_qos_map);
1459         }
1460         nla_nest_end(msg, qos);
1461
1462         nla_nest_end(msg, data);
1463         nla_nest_end(msg, linkinfo);
1464
1465         rv = system_rtnl_call(msg);
1466         if (rv)
1467                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1468
1469         return rv;
1470
1471 nla_put_failure:
1472         nlmsg_free(msg);
1473         return -ENOMEM;
1474 }
1475
1476 int system_vlandev_del(struct device *vlandev)
1477 {
1478         return system_link_del(vlandev->ifname);
1479 }
1480
1481 void
1482 system_if_get_settings(struct device *dev, struct device_settings *s)
1483 {
1484         struct ifreq ifr;
1485         char buf[10];
1486
1487         memset(&ifr, 0, sizeof(ifr));
1488         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1489
1490         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1491                 s->mtu = ifr.ifr_mtu;
1492                 s->flags |= DEV_OPT_MTU;
1493         }
1494
1495         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1496         if (s->mtu6 > 0)
1497                 s->flags |= DEV_OPT_MTU6;
1498
1499         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1500                 s->txqueuelen = ifr.ifr_qlen;
1501                 s->flags |= DEV_OPT_TXQUEUELEN;
1502         }
1503
1504         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1505                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1506                 s->flags |= DEV_OPT_MACADDR;
1507         }
1508
1509         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1510                 s->ipv6 = !strtoul(buf, NULL, 0);
1511                 s->flags |= DEV_OPT_IPV6;
1512         }
1513
1514         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1515                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1516                 s->flags |= DEV_OPT_PROMISC;
1517
1518                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1519                 s->flags |= DEV_OPT_MULTICAST;
1520         }
1521
1522         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1523                 s->rpfilter = strtoul(buf, NULL, 0);
1524                 s->flags |= DEV_OPT_RPFILTER;
1525         }
1526
1527         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1528                 s->acceptlocal = strtoul(buf, NULL, 0);
1529                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1530         }
1531
1532         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1533                 s->igmpversion = strtoul(buf, NULL, 0);
1534                 s->flags |= DEV_OPT_IGMPVERSION;
1535         }
1536
1537         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1538                 s->mldversion = strtoul(buf, NULL, 0);
1539                 s->flags |= DEV_OPT_MLDVERSION;
1540         }
1541
1542         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1543                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1544                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1545         }
1546
1547         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1548                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1549                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1550         }
1551
1552         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
1553                 s->neigh4locktime = strtol(buf, NULL, 0);
1554                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
1555         }
1556
1557         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1558                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1559                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1560         }
1561
1562         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1563                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1564                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1565         }
1566
1567         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1568                 s->dadtransmits = strtoul(buf, NULL, 0);
1569                 s->flags |= DEV_OPT_DADTRANSMITS;
1570         }
1571
1572         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1573                 s->sendredirects = strtoul(buf, NULL, 0);
1574                 s->flags |= DEV_OPT_SENDREDIRECTS;
1575         }
1576 }
1577
1578 void
1579 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1580 {
1581         struct ifreq ifr;
1582         char buf[12];
1583
1584         memset(&ifr, 0, sizeof(ifr));
1585         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1586         if (s->flags & DEV_OPT_MTU & apply_mask) {
1587                 ifr.ifr_mtu = s->mtu;
1588                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1589                         s->flags &= ~DEV_OPT_MTU;
1590         }
1591         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1592                 system_update_ipv6_mtu(dev, s->mtu6);
1593         }
1594         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1595                 ifr.ifr_qlen = s->txqueuelen;
1596                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1597                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1598         }
1599         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1600                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1601                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1602                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1603                         s->flags &= ~DEV_OPT_MACADDR;
1604         }
1605         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1606                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1607         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1608                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1609                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1610                         s->flags &= ~DEV_OPT_PROMISC;
1611         }
1612         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1613                 snprintf(buf, sizeof(buf), "%u", s->rpfilter);
1614                 system_set_rpfilter(dev, buf);
1615         }
1616         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1617                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1618         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1619                 snprintf(buf, sizeof(buf), "%u", s->igmpversion);
1620                 system_set_igmpversion(dev, buf);
1621         }
1622         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1623                 snprintf(buf, sizeof(buf), "%u", s->mldversion);
1624                 system_set_mldversion(dev, buf);
1625         }
1626         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1627                 snprintf(buf, sizeof(buf), "%u", s->neigh4reachabletime);
1628                 system_set_neigh4reachabletime(dev, buf);
1629                 snprintf(buf, sizeof(buf), "%u", s->neigh6reachabletime);
1630                 system_set_neigh6reachabletime(dev, buf);
1631         }
1632         if (s->flags & DEV_OPT_NEIGHLOCKTIME & apply_mask) {
1633                 snprintf(buf, sizeof(buf), "%d", s->neigh4locktime);
1634                 system_set_neigh4locktime(dev, buf);
1635         }
1636         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1637                 snprintf(buf, sizeof(buf), "%u", s->neigh4gcstaletime);
1638                 system_set_neigh4gcstaletime(dev, buf);
1639                 snprintf(buf, sizeof(buf), "%u", s->neigh6gcstaletime);
1640                 system_set_neigh6gcstaletime(dev, buf);
1641         }
1642         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1643                 snprintf(buf, sizeof(buf), "%u", s->dadtransmits);
1644                 system_set_dadtransmits(dev, buf);
1645         }
1646         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1647                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1648                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1649                         s->flags &= ~DEV_OPT_MULTICAST;
1650         }
1651         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1652                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1653 }
1654
1655 int system_if_up(struct device *dev)
1656 {
1657         system_if_get_settings(dev, &dev->orig_settings);
1658         /* Only keep orig settings based on what needs to be set */
1659         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1660         dev->orig_settings.flags &= dev->settings.flags;
1661         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1662         return system_if_flags(dev->ifname, IFF_UP, 0);
1663 }
1664
1665 int system_if_down(struct device *dev)
1666 {
1667         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1668         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1669         return ret;
1670 }
1671
1672 struct if_check_data {
1673         struct device *dev;
1674         int pending;
1675         int ret;
1676 };
1677
1678 #ifndef IFF_LOWER_UP
1679 #define IFF_LOWER_UP    0x10000
1680 #endif
1681
1682 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1683 {
1684         struct nlmsghdr *nh = nlmsg_hdr(msg);
1685         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1686         struct if_check_data *chk = (struct if_check_data *)arg;
1687
1688         if (nh->nlmsg_type != RTM_NEWLINK)
1689                 return NL_SKIP;
1690
1691         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1692         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1693
1694         return NL_OK;
1695 }
1696
1697 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1698 {
1699         struct if_check_data *chk = (struct if_check_data *)arg;
1700         chk->pending = 0;
1701         return NL_STOP;
1702 }
1703
1704 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1705 {
1706         struct if_check_data *chk = (struct if_check_data *)arg;
1707
1708         device_set_present(chk->dev, false);
1709         device_set_link(chk->dev, false);
1710         chk->pending = err->error;
1711
1712         return NL_STOP;
1713 }
1714
1715 int system_if_check(struct device *dev)
1716 {
1717         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1718         struct nl_msg *msg;
1719         struct ifinfomsg ifi = {
1720                 .ifi_family = AF_UNSPEC,
1721                 .ifi_index = 0,
1722         };
1723         struct if_check_data chk = {
1724                 .dev = dev,
1725                 .pending = 1,
1726         };
1727         int ret = 1;
1728
1729         if (!cb)
1730                 return ret;
1731
1732         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1733         if (!msg)
1734                 goto out;
1735
1736         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1737             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1738                 goto free;
1739
1740         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1741         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1742         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1743
1744         ret = nl_send_auto_complete(sock_rtnl, msg);
1745         if (ret < 0)
1746                 goto free;
1747
1748         while (chk.pending > 0)
1749                 nl_recvmsgs(sock_rtnl, cb);
1750
1751         ret = chk.pending;
1752
1753 free:
1754         nlmsg_free(msg);
1755 out:
1756         nl_cb_put(cb);
1757         return ret;
1758 }
1759
1760 struct device *
1761 system_if_get_parent(struct device *dev)
1762 {
1763         char buf[64], *devname;
1764         int ifindex, iflink, len;
1765         FILE *f;
1766
1767         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1768         f = fopen(buf, "r");
1769         if (!f)
1770                 return NULL;
1771
1772         len = fread(buf, 1, sizeof(buf) - 1, f);
1773         fclose(f);
1774
1775         if (len <= 0)
1776                 return NULL;
1777
1778         buf[len] = 0;
1779         iflink = strtoul(buf, NULL, 0);
1780         ifindex = system_if_resolve(dev);
1781         if (!iflink || iflink == ifindex)
1782                 return NULL;
1783
1784         devname = if_indextoname(iflink, buf);
1785         if (!devname)
1786                 return NULL;
1787
1788         return device_get(devname, true);
1789 }
1790
1791 static bool
1792 read_string_file(int dir_fd, const char *file, char *buf, int len)
1793 {
1794         bool ret = false;
1795         char *c;
1796         int fd;
1797
1798         fd = openat(dir_fd, file, O_RDONLY);
1799         if (fd < 0)
1800                 return false;
1801
1802 retry:
1803         len = read(fd, buf, len - 1);
1804         if (len < 0) {
1805                 if (errno == EINTR)
1806                         goto retry;
1807         } else if (len > 0) {
1808                         buf[len] = 0;
1809
1810                         c = strchr(buf, '\n');
1811                         if (c)
1812                                 *c = 0;
1813
1814                         ret = true;
1815         }
1816
1817         close(fd);
1818
1819         return ret;
1820 }
1821
1822 static bool
1823 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1824 {
1825         char buf[64];
1826         bool ret = false;
1827
1828         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1829         if (ret)
1830                 *val = strtoull(buf, NULL, 0);
1831
1832         return ret;
1833 }
1834
1835 /* Assume advertised flags == supported flags */
1836 static const struct {
1837         uint32_t mask;
1838         const char *name;
1839 } ethtool_link_modes[] = {
1840         { ADVERTISED_10baseT_Half, "10baseT-H" },
1841         { ADVERTISED_10baseT_Full, "10baseT-F" },
1842         { ADVERTISED_100baseT_Half, "100baseT-H" },
1843         { ADVERTISED_100baseT_Full, "100baseT-F" },
1844         { ADVERTISED_1000baseT_Half, "1000baseT-H" },
1845         { ADVERTISED_1000baseT_Full, "1000baseT-F" },
1846         { ADVERTISED_1000baseKX_Full, "1000baseKX-F" },
1847         { ADVERTISED_2500baseX_Full, "2500baseX-F" },
1848         { ADVERTISED_10000baseT_Full, "10000baseT-F" },
1849         { ADVERTISED_10000baseKX4_Full, "10000baseKX4-F" },
1850         { ADVERTISED_10000baseKR_Full, "10000baseKR-F" },
1851         { ADVERTISED_20000baseMLD2_Full, "20000baseMLD2-F" },
1852         { ADVERTISED_20000baseKR2_Full, "20000baseKR2-F" },
1853         { ADVERTISED_40000baseKR4_Full, "40000baseKR4-F" },
1854         { ADVERTISED_40000baseCR4_Full, "40000baseCR4-F" },
1855         { ADVERTISED_40000baseSR4_Full, "40000baseSR4-F" },
1856         { ADVERTISED_40000baseLR4_Full, "40000baseLR4-F" },
1857 #ifdef ADVERTISED_56000baseKR4_Full
1858         { ADVERTISED_56000baseKR4_Full, "56000baseKR4-F" },
1859         { ADVERTISED_56000baseCR4_Full, "56000baseCR4-F" },
1860         { ADVERTISED_56000baseSR4_Full, "56000baseSR4-F" },
1861         { ADVERTISED_56000baseLR4_Full, "56000baseLR4-F" },
1862 #endif
1863 };
1864
1865 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1866 {
1867         int i;
1868         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1869                 if (mask & ethtool_link_modes[i].mask)
1870                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1871         }
1872 }
1873
1874 bool
1875 system_if_force_external(const char *ifname)
1876 {
1877         char buf[64];
1878         struct stat s;
1879
1880         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1881         return stat(buf, &s) == 0;
1882 }
1883
1884 int
1885 system_if_dump_info(struct device *dev, struct blob_buf *b)
1886 {
1887         struct ethtool_cmd ecmd;
1888         struct ifreq ifr;
1889         char *s;
1890         void *c;
1891
1892         memset(&ecmd, 0, sizeof(ecmd));
1893         memset(&ifr, 0, sizeof(ifr));
1894         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1895         ifr.ifr_data = (caddr_t) &ecmd;
1896         ecmd.cmd = ETHTOOL_GSET;
1897
1898         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1899                 c = blobmsg_open_array(b, "link-advertising");
1900                 system_add_link_modes(b, ecmd.advertising);
1901                 blobmsg_close_array(b, c);
1902
1903                 c = blobmsg_open_array(b, "link-partner-advertising");
1904                 system_add_link_modes(b, ecmd.lp_advertising);
1905                 blobmsg_close_array(b, c);
1906
1907                 c = blobmsg_open_array(b, "link-supported");
1908                 system_add_link_modes(b, ecmd.supported);
1909                 blobmsg_close_array(b, c);
1910
1911                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1912                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1913                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1914                 blobmsg_add_string_buffer(b);
1915
1916                 blobmsg_add_u8(b, "autoneg", !!ecmd.autoneg);
1917         }
1918
1919         return 0;
1920 }
1921
1922 int
1923 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1924 {
1925         const char *const counters[] = {
1926                 "collisions",     "rx_frame_errors",   "tx_compressed",
1927                 "multicast",      "rx_length_errors",  "tx_dropped",
1928                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1929                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1930                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1931                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1932                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1933                 "rx_fifo_errors", "tx_carrier_errors",
1934         };
1935         char buf[64];
1936         int stats_dir;
1937         int i;
1938         uint64_t val = 0;
1939
1940         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1941         stats_dir = open(buf, O_DIRECTORY);
1942         if (stats_dir < 0)
1943                 return -1;
1944
1945         for (i = 0; i < ARRAY_SIZE(counters); i++)
1946                 if (read_uint64_file(stats_dir, counters[i], &val))
1947                         blobmsg_add_u64(b, counters[i], val);
1948
1949         close(stats_dir);
1950         return 0;
1951 }
1952
1953 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1954 {
1955         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1956         int alen = v4 ? 4 : 16;
1957         unsigned int flags = 0;
1958         struct ifaddrmsg ifa = {
1959                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1960                 .ifa_prefixlen = addr->mask,
1961                 .ifa_index = dev->ifindex,
1962         };
1963
1964         struct nl_msg *msg;
1965         if (cmd == RTM_NEWADDR)
1966                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1967
1968         msg = nlmsg_alloc_simple(cmd, flags);
1969         if (!msg)
1970                 return -1;
1971
1972         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1973         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1974         if (v4) {
1975                 if (addr->broadcast)
1976                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1977                 if (addr->point_to_point)
1978                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1979         } else {
1980                 time_t now = system_get_rtime();
1981                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1982
1983                 if (addr->preferred_until) {
1984                         int64_t preferred = addr->preferred_until - now;
1985                         if (preferred < 0)
1986                                 preferred = 0;
1987                         else if (preferred > UINT32_MAX)
1988                                 preferred = UINT32_MAX;
1989
1990                         cinfo.ifa_prefered = preferred;
1991                 }
1992
1993                 if (addr->valid_until) {
1994                         int64_t valid = addr->valid_until - now;
1995                         if (valid <= 0) {
1996                                 nlmsg_free(msg);
1997                                 return -1;
1998                         }
1999                         else if (valid > UINT32_MAX)
2000                                 valid = UINT32_MAX;
2001
2002                         cinfo.ifa_valid = valid;
2003                 }
2004
2005                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
2006
2007                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
2008                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
2009         }
2010
2011         return system_rtnl_call(msg);
2012 }
2013
2014 int system_add_address(struct device *dev, struct device_addr *addr)
2015 {
2016         return system_addr(dev, addr, RTM_NEWADDR);
2017 }
2018
2019 int system_del_address(struct device *dev, struct device_addr *addr)
2020 {
2021         return system_addr(dev, addr, RTM_DELADDR);
2022 }
2023
2024 static int system_neigh(struct device *dev, struct device_neighbor *neighbor, int cmd)
2025 {
2026         int alen = ((neighbor->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
2027         unsigned int flags = 0;
2028         struct ndmsg ndm = {
2029                 .ndm_family = (alen == 4) ? AF_INET : AF_INET6,
2030                 .ndm_ifindex = dev->ifindex,
2031                 .ndm_state = NUD_PERMANENT,
2032                 .ndm_flags = (neighbor->proxy ? NTF_PROXY : 0) | (neighbor->router ? NTF_ROUTER : 0),
2033         };
2034         struct nl_msg *msg;
2035
2036         if (cmd == RTM_NEWNEIGH)
2037                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
2038
2039         msg = nlmsg_alloc_simple(cmd, flags);
2040
2041         if (!msg)
2042                 return -1;
2043
2044         nlmsg_append(msg, &ndm, sizeof(ndm), 0);
2045
2046         nla_put(msg, NDA_DST, alen, &neighbor->addr);
2047         if (neighbor->flags & DEVNEIGH_MAC)
2048                 nla_put(msg, NDA_LLADDR, sizeof(neighbor->macaddr), &neighbor->macaddr);
2049
2050
2051         return system_rtnl_call(msg);
2052 }
2053
2054 int system_add_neighbor(struct device *dev, struct device_neighbor *neighbor)
2055 {
2056         return system_neigh(dev, neighbor, RTM_NEWNEIGH);
2057 }
2058
2059 int system_del_neighbor(struct device *dev, struct device_neighbor *neighbor)
2060 {
2061         return system_neigh(dev, neighbor, RTM_DELNEIGH);
2062 }
2063
2064 static int system_rt(struct device *dev, struct device_route *route, int cmd)
2065 {
2066         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
2067         bool have_gw;
2068         unsigned int flags = 0;
2069
2070         if (alen == 4)
2071                 have_gw = !!route->nexthop.in.s_addr;
2072         else
2073                 have_gw = route->nexthop.in6.s6_addr32[0] ||
2074                         route->nexthop.in6.s6_addr32[1] ||
2075                         route->nexthop.in6.s6_addr32[2] ||
2076                         route->nexthop.in6.s6_addr32[3];
2077
2078         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
2079                         ? route->table : RT_TABLE_MAIN;
2080
2081         struct rtmsg rtm = {
2082                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2083                 .rtm_dst_len = route->mask,
2084                 .rtm_src_len = route->sourcemask,
2085                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
2086                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
2087                 .rtm_scope = RT_SCOPE_NOWHERE,
2088                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
2089                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
2090         };
2091         struct nl_msg *msg;
2092
2093         if (cmd == RTM_NEWROUTE) {
2094                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
2095
2096                 if (!dev) { /* Add null-route */
2097                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2098                         rtm.rtm_type = RTN_UNREACHABLE;
2099                 }
2100                 else
2101                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
2102         }
2103
2104         if (route->flags & DEVROUTE_TYPE) {
2105                 rtm.rtm_type = route->type;
2106                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
2107                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
2108                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
2109                                 rtm.rtm_table = RT_TABLE_LOCAL;
2110                 }
2111
2112                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
2113                         rtm.rtm_scope = RT_SCOPE_HOST;
2114                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
2115                                 rtm.rtm_type == RTN_ANYCAST) {
2116                         rtm.rtm_scope = RT_SCOPE_LINK;
2117                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
2118                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY ||
2119                                 rtm.rtm_type == RTN_THROW) {
2120                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
2121                         dev = NULL;
2122                 }
2123         }
2124
2125         msg = nlmsg_alloc_simple(cmd, flags);
2126         if (!msg)
2127                 return -1;
2128
2129         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2130
2131         if (route->mask)
2132                 nla_put(msg, RTA_DST, alen, &route->addr);
2133
2134         if (route->sourcemask) {
2135                 if (rtm.rtm_family == AF_INET)
2136                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
2137                 else
2138                         nla_put(msg, RTA_SRC, alen, &route->source);
2139         }
2140
2141         if (route->metric > 0)
2142                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
2143
2144         if (have_gw)
2145                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
2146
2147         if (dev)
2148                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
2149
2150         if (table >= 256)
2151                 nla_put_u32(msg, RTA_TABLE, table);
2152
2153         if (route->flags & DEVROUTE_MTU) {
2154                 struct nlattr *metrics;
2155
2156                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
2157                         goto nla_put_failure;
2158
2159                 nla_put_u32(msg, RTAX_MTU, route->mtu);
2160
2161                 nla_nest_end(msg, metrics);
2162         }
2163
2164         return system_rtnl_call(msg);
2165
2166 nla_put_failure:
2167         nlmsg_free(msg);
2168         return -ENOMEM;
2169 }
2170
2171 int system_add_route(struct device *dev, struct device_route *route)
2172 {
2173         return system_rt(dev, route, RTM_NEWROUTE);
2174 }
2175
2176 int system_del_route(struct device *dev, struct device_route *route)
2177 {
2178         return system_rt(dev, route, RTM_DELROUTE);
2179 }
2180
2181 int system_flush_routes(void)
2182 {
2183         const char *names[] = {
2184                 "/proc/sys/net/ipv4/route/flush",
2185                 "/proc/sys/net/ipv6/route/flush"
2186         };
2187         int fd, i;
2188
2189         for (i = 0; i < ARRAY_SIZE(names); i++) {
2190                 fd = open(names[i], O_WRONLY);
2191                 if (fd < 0)
2192                         continue;
2193
2194                 if (write(fd, "-1", 2)) {}
2195                 close(fd);
2196         }
2197         return 0;
2198 }
2199
2200 bool system_resolve_rt_type(const char *type, unsigned int *id)
2201 {
2202         return system_rtn_aton(type, id);
2203 }
2204
2205 bool system_resolve_rt_proto(const char *type, unsigned int *id)
2206 {
2207         FILE *f;
2208         char *e, buf[128];
2209         unsigned int n, proto = 256;
2210         n = strtoul(type, &e, 0);
2211         if (!*e && e != type)
2212                 proto = n;
2213         else if (!strcmp(type, "unspec"))
2214                 proto = RTPROT_UNSPEC;
2215         else if (!strcmp(type, "kernel"))
2216                 proto = RTPROT_KERNEL;
2217         else if (!strcmp(type, "boot"))
2218                 proto = RTPROT_BOOT;
2219         else if (!strcmp(type, "static"))
2220                 proto = RTPROT_STATIC;
2221         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
2222                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
2223                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2224                                 continue;
2225
2226                         n = strtoul(e, NULL, 10);
2227                         e = strtok(NULL, " \t\n");
2228
2229                         if (e && !strcmp(e, type)) {
2230                                 proto = n;
2231                                 break;
2232                         }
2233                 }
2234                 fclose(f);
2235         }
2236
2237         if (proto > 255)
2238                 return false;
2239
2240         *id = proto;
2241         return true;
2242 }
2243
2244 bool system_resolve_rt_table(const char *name, unsigned int *id)
2245 {
2246         FILE *f;
2247         char *e, buf[128];
2248         unsigned int n, table = RT_TABLE_UNSPEC;
2249
2250         /* first try to parse table as number */
2251         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2252                 table = n;
2253
2254         /* handle well known aliases */
2255         else if (!strcmp(name, "default"))
2256                 table = RT_TABLE_DEFAULT;
2257         else if (!strcmp(name, "main"))
2258                 table = RT_TABLE_MAIN;
2259         else if (!strcmp(name, "local"))
2260                 table = RT_TABLE_LOCAL;
2261
2262         /* try to look up name in /etc/iproute2/rt_tables */
2263         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2264         {
2265                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2266                 {
2267                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2268                                 continue;
2269
2270                         n = strtoul(e, NULL, 10);
2271                         e = strtok(NULL, " \t\n");
2272
2273                         if (e && !strcmp(e, name))
2274                         {
2275                                 table = n;
2276                                 break;
2277                         }
2278                 }
2279
2280                 fclose(f);
2281         }
2282
2283         if (table == RT_TABLE_UNSPEC)
2284                 return false;
2285
2286         *id = table;
2287         return true;
2288 }
2289
2290 bool system_is_default_rt_table(unsigned int id)
2291 {
2292         return (id == RT_TABLE_MAIN);
2293 }
2294
2295 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2296 {
2297         char *e;
2298         unsigned int n;
2299
2300         if (!strcmp(filter, "strict"))
2301                 n = 1;
2302         else if (!strcmp(filter, "loose"))
2303                 n = 2;
2304         else {
2305                 n = strtoul(filter, &e, 0);
2306                 if (*e || e == filter || n > 2)
2307                         return false;
2308         }
2309
2310         *id = n;
2311         return true;
2312 }
2313
2314 static int system_iprule(struct iprule *rule, int cmd)
2315 {
2316         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2317
2318         struct nl_msg *msg;
2319         struct rtmsg rtm = {
2320                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2321                 .rtm_protocol = RTPROT_STATIC,
2322                 .rtm_scope = RT_SCOPE_UNIVERSE,
2323                 .rtm_table = RT_TABLE_UNSPEC,
2324                 .rtm_type = RTN_UNSPEC,
2325                 .rtm_flags = 0,
2326         };
2327
2328         if (cmd == RTM_NEWRULE)
2329                 rtm.rtm_type = RTN_UNICAST;
2330
2331         if (rule->invert)
2332                 rtm.rtm_flags |= FIB_RULE_INVERT;
2333
2334         if (rule->flags & IPRULE_SRC)
2335                 rtm.rtm_src_len = rule->src_mask;
2336
2337         if (rule->flags & IPRULE_DEST)
2338                 rtm.rtm_dst_len = rule->dest_mask;
2339
2340         if (rule->flags & IPRULE_TOS)
2341                 rtm.rtm_tos = rule->tos;
2342
2343         if (rule->flags & IPRULE_LOOKUP) {
2344                 if (rule->lookup < 256)
2345                         rtm.rtm_table = rule->lookup;
2346         }
2347
2348         if (rule->flags & IPRULE_ACTION)
2349                 rtm.rtm_type = rule->action;
2350         else if (rule->flags & IPRULE_GOTO)
2351                 rtm.rtm_type = FR_ACT_GOTO;
2352         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2353                 rtm.rtm_type = FR_ACT_NOP;
2354
2355         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2356
2357         if (!msg)
2358                 return -1;
2359
2360         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2361
2362         if (rule->flags & IPRULE_IN)
2363                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2364
2365         if (rule->flags & IPRULE_OUT)
2366                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2367
2368         if (rule->flags & IPRULE_SRC)
2369                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2370
2371         if (rule->flags & IPRULE_DEST)
2372                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2373
2374         if (rule->flags & IPRULE_PRIORITY)
2375                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2376         else if (cmd == RTM_NEWRULE)
2377                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2378
2379         if (rule->flags & IPRULE_FWMARK)
2380                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2381
2382         if (rule->flags & IPRULE_FWMASK)
2383                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2384
2385         if (rule->flags & IPRULE_LOOKUP) {
2386                 if (rule->lookup >= 256)
2387                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2388         }
2389
2390         if (rule->flags & IPRULE_SUP_PREFIXLEN)
2391                 nla_put_u32(msg, FRA_SUPPRESS_PREFIXLEN, rule->sup_prefixlen);
2392
2393         if (rule->flags & IPRULE_GOTO)
2394                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2395
2396         return system_rtnl_call(msg);
2397 }
2398
2399 int system_add_iprule(struct iprule *rule)
2400 {
2401         return system_iprule(rule, RTM_NEWRULE);
2402 }
2403
2404 int system_del_iprule(struct iprule *rule)
2405 {
2406         return system_iprule(rule, RTM_DELRULE);
2407 }
2408
2409 int system_flush_iprules(void)
2410 {
2411         int rv = 0;
2412         struct iprule rule;
2413
2414         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2415         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2416
2417         memset(&rule, 0, sizeof(rule));
2418
2419
2420         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2421
2422         rule.priority = 0;
2423         rule.lookup = RT_TABLE_LOCAL;
2424         rv |= system_iprule(&rule, RTM_NEWRULE);
2425
2426         rule.priority = 32766;
2427         rule.lookup = RT_TABLE_MAIN;
2428         rv |= system_iprule(&rule, RTM_NEWRULE);
2429
2430         rule.priority = 32767;
2431         rule.lookup = RT_TABLE_DEFAULT;
2432         rv |= system_iprule(&rule, RTM_NEWRULE);
2433
2434
2435         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2436
2437         rule.priority = 0;
2438         rule.lookup = RT_TABLE_LOCAL;
2439         rv |= system_iprule(&rule, RTM_NEWRULE);
2440
2441         rule.priority = 32766;
2442         rule.lookup = RT_TABLE_MAIN;
2443         rv |= system_iprule(&rule, RTM_NEWRULE);
2444
2445         return rv;
2446 }
2447
2448 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2449 {
2450         return system_rtn_aton(action, id);
2451 }
2452
2453 time_t system_get_rtime(void)
2454 {
2455         struct timespec ts;
2456         struct timeval tv;
2457
2458         if (clock_gettime(CLOCK_MONOTONIC, &ts) == 0)
2459                 return ts.tv_sec;
2460
2461         if (gettimeofday(&tv, NULL) == 0)
2462                 return tv.tv_sec;
2463
2464         return 0;
2465 }
2466
2467 #ifndef IP_DF
2468 #define IP_DF       0x4000
2469 #endif
2470
2471 static int tunnel_ioctl(const char *name, int cmd, void *p)
2472 {
2473         struct ifreq ifr;
2474
2475         memset(&ifr, 0, sizeof(ifr));
2476         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name) - 1);
2477         ifr.ifr_ifru.ifru_data = p;
2478         return ioctl(sock_ioctl, cmd, &ifr);
2479 }
2480
2481 #ifdef IFLA_IPTUN_MAX
2482 static int system_add_ip6_tunnel(const char *name, const unsigned int link,
2483                                  struct blob_attr **tb)
2484 {
2485         struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2486                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2487         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2488         struct blob_attr *cur;
2489         int ret = 0, ttl = 0;
2490
2491         if (!nlm)
2492                 return -1;
2493
2494         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2495         nla_put_string(nlm, IFLA_IFNAME, name);
2496
2497         if (link)
2498                 nla_put_u32(nlm, IFLA_LINK, link);
2499
2500         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2501         if (!linkinfo) {
2502                 ret = -ENOMEM;
2503                 goto failure;
2504         }
2505
2506         nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2507         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2508         if (!infodata) {
2509                 ret = -ENOMEM;
2510                 goto failure;
2511         }
2512
2513         if (link)
2514                 nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2515
2516         if ((cur = tb[TUNNEL_ATTR_TTL]))
2517                 ttl = blobmsg_get_u32(cur);
2518
2519         nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2520         nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2521
2522         struct in6_addr in6buf;
2523         if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2524                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2525                         ret = -EINVAL;
2526                         goto failure;
2527                 }
2528                 nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2529         }
2530
2531         if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2532                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2533                         ret = -EINVAL;
2534                         goto failure;
2535                 }
2536                 nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2537         }
2538
2539         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2540                 struct blob_attr *tb_data[__IPIP6_DATA_ATTR_MAX];
2541                 uint32_t tun_flags = IP6_TNL_F_IGN_ENCAP_LIMIT;
2542
2543                 blobmsg_parse(ipip6_data_attr_list.params, __IPIP6_DATA_ATTR_MAX, tb_data,
2544                         blobmsg_data(cur), blobmsg_len(cur));
2545
2546                 if ((cur = tb_data[IPIP6_DATA_ENCAPLIMIT])) {
2547                         char *str = blobmsg_get_string(cur);
2548
2549                         if (strcmp(str, "ignore")) {
2550                                 char *e;
2551                                 unsigned encap_limit = strtoul(str, &e, 0);
2552
2553                                 if (e == str || *e || encap_limit > 255) {
2554                                         ret = -EINVAL;
2555                                         goto failure;
2556                                 }
2557
2558                                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, encap_limit);
2559                                 tun_flags &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2560                         }
2561                 }
2562
2563 #ifdef IFLA_IPTUN_FMR_MAX
2564                 if ((cur = tb_data[IPIP6_DATA_FMRS])) {
2565                         struct blob_attr *rcur;
2566                         unsigned rrem, fmrcnt = 0;
2567                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2568
2569                         if (!fmrs) {
2570                                 ret = -ENOMEM;
2571                                 goto failure;
2572                         }
2573
2574                         blobmsg_for_each_attr(rcur, cur, rrem) {
2575                                 struct blob_attr *tb_fmr[__FMR_DATA_ATTR_MAX], *tb_cur;
2576                                 struct in6_addr ip6prefix;
2577                                 struct in_addr ip4prefix;
2578                                 unsigned ip4len, ip6len, ealen, offset;
2579
2580                                 blobmsg_parse(fmr_data_attr_list.params, __FMR_DATA_ATTR_MAX, tb_fmr,
2581                                                 blobmsg_data(rcur), blobmsg_len(rcur));
2582
2583                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX6]) ||
2584                                                 !parse_ip_and_netmask(AF_INET6,
2585                                                         blobmsg_data(tb_cur), &ip6prefix,
2586                                                         &ip6len)) {
2587                                         ret = -EINVAL;
2588                                         goto failure;
2589                                 }
2590
2591                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX4]) ||
2592                                                 !parse_ip_and_netmask(AF_INET,
2593                                                         blobmsg_data(tb_cur), &ip4prefix,
2594                                                         &ip4len)) {
2595                                         ret = -EINVAL;
2596                                         goto failure;
2597                                 }
2598
2599                                 if (!(tb_cur = tb_fmr[FMR_DATA_EALEN])) {
2600                                         ret = -EINVAL;
2601                                         goto failure;
2602                                 }
2603                                 ealen = blobmsg_get_u32(tb_cur);
2604
2605                                 if (!(tb_cur = tb_fmr[FMR_DATA_OFFSET])) {
2606                                         ret = -EINVAL;
2607                                         goto failure;
2608                                 }
2609                                 offset = blobmsg_get_u32(tb_cur);
2610
2611                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2612                                 if (!rule) {
2613                                         ret = -ENOMEM;
2614                                         goto failure;
2615                                 }
2616
2617                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2618                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2619                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2620                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2621                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2622                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2623
2624                                 nla_nest_end(nlm, rule);
2625                         }
2626
2627                         nla_nest_end(nlm, fmrs);
2628                 }
2629 #endif
2630                 if (tun_flags)
2631                         nla_put_u32(nlm, IFLA_IPTUN_FLAGS, tun_flags);
2632         }
2633
2634         nla_nest_end(nlm, infodata);
2635         nla_nest_end(nlm, linkinfo);
2636
2637         return system_rtnl_call(nlm);
2638
2639 failure:
2640         nlmsg_free(nlm);
2641         return ret;
2642 }
2643 #endif
2644
2645 #ifdef IFLA_IPTUN_MAX
2646 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2647 static int system_add_gre_tunnel(const char *name, const char *kind,
2648                                  const unsigned int link, struct blob_attr **tb, bool v6)
2649 {
2650         struct nl_msg *nlm;
2651         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2652         struct blob_attr *cur;
2653         uint32_t ikey = 0, okey = 0, flowinfo = 0, flags6 = IP6_TNL_F_IGN_ENCAP_LIMIT;
2654         uint16_t iflags = 0, oflags = 0;
2655         uint8_t tos = 0;
2656         int ret = 0, ttl = 0;
2657         unsigned encap_limit = 0;
2658
2659         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2660         if (!nlm)
2661                 return -1;
2662
2663         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2664         nla_put_string(nlm, IFLA_IFNAME, name);
2665
2666         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2667         if (!linkinfo) {
2668                 ret = -ENOMEM;
2669                 goto failure;
2670         }
2671
2672         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2673         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2674         if (!infodata) {
2675                 ret = -ENOMEM;
2676                 goto failure;
2677         }
2678
2679         if (link)
2680                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2681
2682         if ((cur = tb[TUNNEL_ATTR_TTL]))
2683                 ttl = blobmsg_get_u32(cur);
2684
2685         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2686                 char *str = blobmsg_get_string(cur);
2687                 if (strcmp(str, "inherit")) {
2688                         unsigned uval;
2689
2690                         if (!system_tos_aton(str, &uval)) {
2691                                 ret = -EINVAL;
2692                                 goto failure;
2693                         }
2694
2695                         if (v6)
2696                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2697                         else
2698                                 tos = uval;
2699                 } else {
2700                         if (v6)
2701                                 flags6 |= IP6_TNL_F_USE_ORIG_TCLASS;
2702                         else
2703                                 tos = 1;
2704                 }
2705         }
2706
2707         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2708                 struct blob_attr *tb_data[__GRE_DATA_ATTR_MAX];
2709
2710                 blobmsg_parse(gre_data_attr_list.params, __GRE_DATA_ATTR_MAX, tb_data,
2711                         blobmsg_data(cur), blobmsg_len(cur));
2712
2713                 if ((cur = tb_data[GRE_DATA_IKEY])) {
2714                         if ((ikey = blobmsg_get_u32(cur)))
2715                                 iflags |= GRE_KEY;
2716                 }
2717
2718                 if ((cur = tb_data[GRE_DATA_OKEY])) {
2719                         if ((okey = blobmsg_get_u32(cur)))
2720                                 oflags |= GRE_KEY;
2721                 }
2722
2723                 if ((cur = tb_data[GRE_DATA_ICSUM])) {
2724                         if (blobmsg_get_bool(cur))
2725                                 iflags |= GRE_CSUM;
2726                 }
2727
2728                 if ((cur = tb_data[GRE_DATA_OCSUM])) {
2729                         if (blobmsg_get_bool(cur))
2730                                 oflags |= GRE_CSUM;
2731                 }
2732
2733                 if ((cur = tb_data[GRE_DATA_ISEQNO])) {
2734                         if (blobmsg_get_bool(cur))
2735                                 iflags |= GRE_SEQ;
2736                 }
2737
2738                 if ((cur = tb_data[GRE_DATA_OSEQNO])) {
2739                         if (blobmsg_get_bool(cur))
2740                                 oflags |= GRE_SEQ;
2741                 }
2742
2743                 if ((cur = tb_data[GRE_DATA_ENCAPLIMIT])) {
2744                         char *str = blobmsg_get_string(cur);
2745
2746                         if (strcmp(str, "ignore")) {
2747                                 char *e;
2748
2749                                 encap_limit = strtoul(str, &e, 0);
2750
2751                                 if (e == str || *e || encap_limit > 255) {
2752                                         ret = -EINVAL;
2753                                         goto failure;
2754                                 }
2755
2756                                 flags6 &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
2757                         }
2758                 }
2759         }
2760
2761         if (v6) {
2762                 struct in6_addr in6buf;
2763                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2764                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2765                                 ret = -EINVAL;
2766                                 goto failure;
2767                         }
2768                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2769                 }
2770
2771                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2772                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2773                                 ret = -EINVAL;
2774                                 goto failure;
2775                         }
2776                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2777                 }
2778
2779                 if (!(flags6 & IP6_TNL_F_IGN_ENCAP_LIMIT))
2780                         nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, encap_limit);
2781
2782                 if (flowinfo)
2783                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2784
2785                 if (flags6)
2786                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags6);
2787
2788                 if (!ttl)
2789                         ttl = 64;
2790         } else {
2791                 struct in_addr inbuf;
2792                 bool set_df = true;
2793
2794                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2795                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2796                                 ret = -EINVAL;
2797                                 goto failure;
2798                         }
2799                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2800                 }
2801
2802                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2803                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2804                                 ret = -EINVAL;
2805                                 goto failure;
2806                         }
2807                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2808
2809                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2810                                 if (!okey) {
2811                                         okey = inbuf.s_addr;
2812                                         oflags |= GRE_KEY;
2813                                 }
2814
2815                                 if (!ikey) {
2816                                         ikey = inbuf.s_addr;
2817                                         iflags |= GRE_KEY;
2818                                 }
2819                         }
2820                 }
2821
2822                 if ((cur = tb[TUNNEL_ATTR_DF]))
2823                         set_df = blobmsg_get_bool(cur);
2824
2825                 if (!set_df) {
2826                         /* ttl != 0 and nopmtudisc are incompatible */
2827                         if (ttl) {
2828                                 ret = -EINVAL;
2829                                 goto failure;
2830                         }
2831                 } else if (!ttl)
2832                         ttl = 64;
2833
2834                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2835
2836                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2837         }
2838
2839         if (ttl)
2840                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2841
2842         if (oflags)
2843                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2844
2845         if (iflags)
2846                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2847
2848         if (okey)
2849                 nla_put_u32(nlm, IFLA_GRE_OKEY, htonl(okey));
2850
2851         if (ikey)
2852                 nla_put_u32(nlm, IFLA_GRE_IKEY, htonl(ikey));
2853
2854         nla_nest_end(nlm, infodata);
2855         nla_nest_end(nlm, linkinfo);
2856
2857         return system_rtnl_call(nlm);
2858
2859 failure:
2860         nlmsg_free(nlm);
2861         return ret;
2862 }
2863 #endif
2864
2865 #ifdef IFLA_VTI_MAX
2866 static int system_add_vti_tunnel(const char *name, const char *kind,
2867                                  const unsigned int link, struct blob_attr **tb, bool v6)
2868 {
2869         struct nl_msg *nlm;
2870         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2871         struct blob_attr *cur;
2872         int ret = 0;
2873
2874         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2875         if (!nlm)
2876                 return -1;
2877
2878         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2879         nla_put_string(nlm, IFLA_IFNAME, name);
2880
2881         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2882         if (!linkinfo) {
2883                 ret = -ENOMEM;
2884                 goto failure;
2885         }
2886
2887         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2888         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2889         if (!infodata) {
2890                 ret = -ENOMEM;
2891                 goto failure;
2892         }
2893
2894         if (link)
2895                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2896
2897         if (v6) {
2898                 struct in6_addr in6buf;
2899                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2900                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2901                                 ret = -EINVAL;
2902                                 goto failure;
2903                         }
2904                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2905                 }
2906
2907                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2908                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2909                                 ret = -EINVAL;
2910                                 goto failure;
2911                         }
2912                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2913                 }
2914
2915         } else {
2916                 struct in_addr inbuf;
2917
2918                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2919                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2920                                 ret = -EINVAL;
2921                                 goto failure;
2922                         }
2923                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2924                 }
2925
2926                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2927                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2928                                 ret = -EINVAL;
2929                                 goto failure;
2930                         }
2931                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2932                 }
2933
2934         }
2935
2936         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2937                 struct blob_attr *tb_data[__VTI_DATA_ATTR_MAX];
2938                 uint32_t ikey = 0, okey = 0;
2939
2940                 blobmsg_parse(vti_data_attr_list.params, __VTI_DATA_ATTR_MAX, tb_data,
2941                         blobmsg_data(cur), blobmsg_len(cur));
2942
2943                 if ((cur = tb_data[VTI_DATA_IKEY])) {
2944                         if ((ikey = blobmsg_get_u32(cur)))
2945                                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2946                 }
2947
2948                 if ((cur = tb_data[VTI_DATA_OKEY])) {
2949                         if ((okey = blobmsg_get_u32(cur)))
2950                                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2951                 }
2952         }
2953
2954         nla_nest_end(nlm, infodata);
2955         nla_nest_end(nlm, linkinfo);
2956
2957         return system_rtnl_call(nlm);
2958
2959 failure:
2960         nlmsg_free(nlm);
2961         return ret;
2962 }
2963 #endif
2964
2965 #ifdef IFLA_XFRM_MAX
2966 static int system_add_xfrm_tunnel(const char *name, const char *kind,
2967                                  const unsigned int link, struct blob_attr **tb)
2968 {
2969         struct nl_msg *nlm;
2970         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2971         struct blob_attr *cur;
2972         int ret = 0;
2973
2974         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2975         if (!nlm)
2976                 return -1;
2977
2978         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2979         nla_put_string(nlm, IFLA_IFNAME, name);
2980
2981         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2982         if (!linkinfo) {
2983                 ret = -ENOMEM;
2984                 goto failure;
2985         }
2986
2987         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2988         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2989         if (!infodata) {
2990                 ret = -ENOMEM;
2991                 goto failure;
2992         }
2993
2994         if (link)
2995                 nla_put_u32(nlm, IFLA_XFRM_LINK, link);
2996
2997         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2998                 struct blob_attr *tb_data[__XFRM_DATA_ATTR_MAX];
2999                 uint32_t if_id = 0;
3000
3001                 blobmsg_parse(xfrm_data_attr_list.params, __XFRM_DATA_ATTR_MAX, tb_data,
3002                         blobmsg_data(cur), blobmsg_len(cur));
3003
3004                 if ((cur = tb_data[XFRM_DATA_IF_ID])) {
3005                         if ((if_id = blobmsg_get_u32(cur)))
3006                                 nla_put_u32(nlm, IFLA_XFRM_IF_ID, if_id);
3007                 }
3008
3009         }
3010
3011         nla_nest_end(nlm, infodata);
3012         nla_nest_end(nlm, linkinfo);
3013
3014         return system_rtnl_call(nlm);
3015
3016 failure:
3017         nlmsg_free(nlm);
3018         return ret;
3019 }
3020 #endif
3021
3022 #ifdef IFLA_VXLAN_MAX
3023 static int system_add_vxlan(const char *name, const unsigned int link, struct blob_attr **tb, bool v6)
3024 {
3025         struct blob_attr *tb_data[__VXLAN_DATA_ATTR_MAX];
3026         struct nl_msg *msg;
3027         struct nlattr *linkinfo, *data;
3028         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
3029         struct blob_attr *cur;
3030         int ret = 0;
3031
3032         if ((cur = tb[TUNNEL_ATTR_DATA]))
3033                 blobmsg_parse(vxlan_data_attr_list.params, __VXLAN_DATA_ATTR_MAX, tb_data,
3034                         blobmsg_data(cur), blobmsg_len(cur));
3035         else
3036                 return -EINVAL;
3037
3038         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
3039
3040         if (!msg)
3041                 return -1;
3042
3043         nlmsg_append(msg, &iim, sizeof(iim), 0);
3044
3045         nla_put_string(msg, IFLA_IFNAME, name);
3046
3047         if ((cur = tb_data[VXLAN_DATA_ATTR_MACADDR])) {
3048                 struct ether_addr *ea = ether_aton(blobmsg_get_string(cur));
3049                 if (!ea) {
3050                         ret = -EINVAL;
3051                         goto failure;
3052                 }
3053
3054                 nla_put(msg, IFLA_ADDRESS, ETH_ALEN, ea);
3055         }
3056
3057         if ((cur = tb[TUNNEL_ATTR_MTU])) {
3058                 uint32_t mtu = blobmsg_get_u32(cur);
3059                 nla_put_u32(msg, IFLA_MTU, mtu);
3060         }
3061
3062         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO))) {
3063                 ret = -ENOMEM;
3064                 goto failure;
3065         }
3066
3067         nla_put_string(msg, IFLA_INFO_KIND, "vxlan");
3068
3069         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA))) {
3070                 ret = -ENOMEM;
3071                 goto failure;
3072         }
3073
3074         if (link)
3075                 nla_put_u32(msg, IFLA_VXLAN_LINK, link);
3076
3077         if ((cur = tb_data[VXLAN_DATA_ATTR_ID])) {
3078                 uint32_t id = blobmsg_get_u32(cur);
3079                 if (id >= (1u << 24) - 1) {
3080                         ret = -EINVAL;
3081                         goto failure;
3082                 }
3083
3084                 nla_put_u32(msg, IFLA_VXLAN_ID, id);
3085         }
3086
3087         if (v6) {
3088                 struct in6_addr in6buf;
3089                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3090                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3091                                 ret = -EINVAL;
3092                                 goto failure;
3093                         }
3094                         nla_put(msg, IFLA_VXLAN_LOCAL6, sizeof(in6buf), &in6buf);
3095                 }
3096
3097                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3098                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3099                                 ret = -EINVAL;
3100                                 goto failure;
3101                         }
3102                         nla_put(msg, IFLA_VXLAN_GROUP6, sizeof(in6buf), &in6buf);
3103                 }
3104         } else {
3105                 struct in_addr inbuf;
3106
3107                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3108                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3109                                 ret = -EINVAL;
3110                                 goto failure;
3111                         }
3112                         nla_put(msg, IFLA_VXLAN_LOCAL, sizeof(inbuf), &inbuf);
3113                 }
3114
3115                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3116                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3117                                 ret = -EINVAL;
3118                                 goto failure;
3119                         }
3120                         nla_put(msg, IFLA_VXLAN_GROUP, sizeof(inbuf), &inbuf);
3121                 }
3122         }
3123
3124         uint32_t port = 4789;
3125         if ((cur = tb_data[VXLAN_DATA_ATTR_PORT])) {
3126                 port = blobmsg_get_u32(cur);
3127                 if (port < 1 || port > 65535) {
3128                         ret = -EINVAL;
3129                         goto failure;
3130                 }
3131         }
3132         nla_put_u16(msg, IFLA_VXLAN_PORT, htons(port));
3133
3134         if ((cur = tb_data[VXLAN_DATA_ATTR_RXCSUM])) {
3135                 bool rxcsum = blobmsg_get_bool(cur);
3136                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_RX, !rxcsum);
3137         }
3138
3139         if ((cur = tb_data[VXLAN_DATA_ATTR_TXCSUM])) {
3140                 bool txcsum = blobmsg_get_bool(cur);
3141                 nla_put_u8(msg, IFLA_VXLAN_UDP_CSUM, txcsum);
3142                 nla_put_u8(msg, IFLA_VXLAN_UDP_ZERO_CSUM6_TX, !txcsum);
3143         }
3144
3145         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3146                 char *str = blobmsg_get_string(cur);
3147                 unsigned tos = 1;
3148
3149                 if (strcmp(str, "inherit")) {
3150                         if (!system_tos_aton(str, &tos)) {
3151                                 ret = -EINVAL;
3152                                 goto failure;
3153                         }
3154                 }
3155
3156                 nla_put_u8(msg, IFLA_VXLAN_TOS, tos);
3157         }
3158
3159         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3160                 uint32_t ttl = blobmsg_get_u32(cur);
3161                 if (ttl < 1 || ttl > 255) {
3162                         ret = -EINVAL;
3163                         goto failure;
3164                 }
3165
3166                 nla_put_u8(msg, IFLA_VXLAN_TTL, ttl);
3167         }
3168
3169         nla_nest_end(msg, data);
3170         nla_nest_end(msg, linkinfo);
3171
3172         ret = system_rtnl_call(msg);
3173         if (ret)
3174                 D(SYSTEM, "Error adding vxlan '%s': %d\n", name, ret);
3175
3176         return ret;
3177
3178 failure:
3179         nlmsg_free(msg);
3180         return ret;
3181 }
3182 #endif
3183
3184 static int system_add_sit_tunnel(const char *name, const unsigned int link, struct blob_attr **tb)
3185 {
3186         struct blob_attr *cur;
3187         int ret = 0;
3188
3189         if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
3190                 return -1;
3191
3192 #ifdef SIOCADD6RD
3193         if ((cur = tb[TUNNEL_ATTR_DATA])) {
3194                 struct blob_attr *tb_data[__SIXRD_DATA_ATTR_MAX];
3195                 unsigned int mask;
3196                 struct ip_tunnel_6rd p6;
3197
3198                 blobmsg_parse(sixrd_data_attr_list.params, __SIXRD_DATA_ATTR_MAX, tb_data,
3199                         blobmsg_data(cur), blobmsg_len(cur));
3200
3201                 memset(&p6, 0, sizeof(p6));
3202
3203                 if ((cur = tb_data[SIXRD_DATA_PREFIX])) {
3204                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
3205                                                 &p6.prefix, &mask) || mask > 128) {
3206                                 ret = -EINVAL;
3207                                 goto failure;
3208                         }
3209
3210                         p6.prefixlen = mask;
3211                 }
3212
3213                 if ((cur = tb_data[SIXRD_DATA_RELAY_PREFIX])) {
3214                         if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
3215                                                 &p6.relay_prefix, &mask) || mask > 32) {
3216                                 ret = -EINVAL;
3217                                 goto failure;
3218                         }
3219
3220                         p6.relay_prefixlen = mask;
3221                 }
3222
3223                 if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
3224                         ret = -1;
3225                         goto failure;
3226                 }
3227         }
3228 #endif
3229
3230         return ret;
3231
3232 failure:
3233         __system_del_ip_tunnel(name, tb);
3234         return ret;
3235 }
3236
3237 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
3238 {
3239         struct blob_attr *cur;
3240         bool set_df = true;
3241         struct ip_tunnel_parm p  = {
3242                 .link = link,
3243                 .iph = {
3244                         .version = 4,
3245                         .ihl = 5,
3246                         .protocol = proto,
3247                 }
3248         };
3249
3250         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
3251                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
3252                 return -EINVAL;
3253
3254         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
3255                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
3256                 return -EINVAL;
3257
3258         if ((cur = tb[TUNNEL_ATTR_DF]))
3259                 set_df = blobmsg_get_bool(cur);
3260
3261         if ((cur = tb[TUNNEL_ATTR_TTL]))
3262                 p.iph.ttl = blobmsg_get_u32(cur);
3263
3264         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3265                 char *str = blobmsg_get_string(cur);
3266                 if (strcmp(str, "inherit")) {
3267                         unsigned uval;
3268
3269                         if (!system_tos_aton(str, &uval))
3270                                 return -EINVAL;
3271
3272                         p.iph.tos = uval;
3273                 } else
3274                         p.iph.tos = 1;
3275         }
3276
3277         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
3278         /* ttl !=0 and nopmtudisc are incompatible */
3279         if (p.iph.ttl && p.iph.frag_off == 0)
3280                 return -EINVAL;
3281
3282         strncpy(p.name, name, sizeof(p.name) - 1);
3283
3284         switch (p.iph.protocol) {
3285         case IPPROTO_IPIP:
3286                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
3287         case IPPROTO_IPV6:
3288                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
3289         default:
3290                 break;
3291         }
3292         return -1;
3293 }
3294
3295 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
3296 {
3297         struct blob_attr *cur;
3298         const char *str;
3299
3300         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3301                 return -EINVAL;
3302         str = blobmsg_data(cur);
3303
3304         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
3305             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
3306             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6") ||
3307             !strcmp(str, "vxlan") || !strcmp(str, "vxlan6") ||
3308             !strcmp(str, "xfrm"))
3309                 return system_link_del(name);
3310         else
3311                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
3312 }
3313
3314 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
3315 {
3316         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3317
3318         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3319                 blob_data(attr), blob_len(attr));
3320
3321         return __system_del_ip_tunnel(name, tb);
3322 }
3323
3324 int system_update_ipv6_mtu(struct device *dev, int mtu)
3325 {
3326         int ret = -1;
3327         char buf[64];
3328         int fd;
3329
3330         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
3331                         dev->ifname);
3332
3333         fd = open(buf, O_RDWR);
3334         if (fd < 0)
3335                 return ret;
3336
3337         if (!mtu) {
3338                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
3339                 if (len < 0)
3340                         goto out;
3341
3342                 buf[len] = 0;
3343                 ret = atoi(buf);
3344         } else {
3345                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
3346                         ret = mtu;
3347         }
3348
3349 out:
3350         close(fd);
3351         return ret;
3352 }
3353
3354 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
3355 {
3356         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
3357         struct blob_attr *cur;
3358         const char *str;
3359
3360         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
3361                 blob_data(attr), blob_len(attr));
3362
3363         __system_del_ip_tunnel(name, tb);
3364
3365         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
3366                 return -EINVAL;
3367         str = blobmsg_data(cur);
3368
3369         unsigned int ttl = 0;
3370         if ((cur = tb[TUNNEL_ATTR_TTL])) {
3371                 ttl = blobmsg_get_u32(cur);
3372                 if (ttl > 255)
3373                         return -EINVAL;
3374         }
3375
3376         unsigned int link = 0;
3377         if ((cur = tb[TUNNEL_ATTR_LINK])) {
3378                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
3379                 if (!iface)
3380                         return -EINVAL;
3381
3382                 if (iface->l3_dev.dev)
3383                         link = iface->l3_dev.dev->ifindex;
3384         }
3385
3386         if (!strcmp(str, "sit"))
3387                 return system_add_sit_tunnel(name, link, tb);
3388 #ifdef IFLA_IPTUN_MAX
3389         else if (!strcmp(str, "ipip6")) {
3390                 return system_add_ip6_tunnel(name, link, tb);
3391         } else if (!strcmp(str, "greip")) {
3392                 return system_add_gre_tunnel(name, "gre", link, tb, false);
3393         } else if (!strcmp(str, "gretapip"))  {
3394                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
3395         } else if (!strcmp(str, "greip6")) {
3396                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
3397         } else if (!strcmp(str, "gretapip6")) {
3398                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
3399 #ifdef IFLA_VTI_MAX
3400         } else if (!strcmp(str, "vtiip")) {
3401                 return system_add_vti_tunnel(name, "vti", link, tb, false);
3402         } else if (!strcmp(str, "vtiip6")) {
3403                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
3404 #endif
3405 #ifdef IFLA_XFRM_MAX
3406         } else if (!strcmp(str, "xfrm")) {
3407                 return system_add_xfrm_tunnel(name, "xfrm", link, tb);
3408 #endif
3409 #ifdef IFLA_VXLAN_MAX
3410         } else if(!strcmp(str, "vxlan")) {
3411                 return system_add_vxlan(name, link, tb, false);
3412         } else if(!strcmp(str, "vxlan6")) {
3413                 return system_add_vxlan(name, link, tb, true);
3414 #endif
3415 #endif
3416         } else if (!strcmp(str, "ipip")) {
3417                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
3418         }
3419         else
3420                 return -EINVAL;
3421
3422         return 0;
3423 }