system-linux.c

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