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