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/veth.h>
  42 #include <linux/version.h>
  43
  44 #ifndef RTN_FAILED_POLICY
  45 #define RTN_FAILED_POLICY 12
  46 #endif
  47
  48 #ifndef IFA_F_NOPREFIXROUTE
  49 #define IFA_F_NOPREFIXROUTE 0x200
  50 #endif
  51
  52 #ifndef IFA_FLAGS
  53 #define IFA_FLAGS (IFA_MULTICAST + 1)
  54 #endif
  55
  56 #include <string.h>
  57 #include <fcntl.h>
  58 #include <glob.h>
  59 #include <time.h>
  60 #include <unistd.h>
  61
  62 #include <netlink/msg.h>
  63 #include <netlink/attr.h>
  64 #include <netlink/socket.h>
  65 #include <libubox/uloop.h>
  66
  67 #include "netifd.h"
  68 #include "device.h"
  69 #include "system.h"
  70
  71 struct event_socket {
  72         struct uloop_fd uloop;
  73         struct nl_sock *sock;
  74         int bufsize;
  75 };
  76
  77 static int sock_ioctl = -1;
  78 static struct nl_sock *sock_rtnl = NULL;
  79
  80 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  81 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  82
  83 static char dev_buf[256];
  84
  85 static void
  86 handler_nl_event(struct uloop_fd *u, unsigned int events)
  87 {
  88         struct event_socket *ev = container_of(u, struct event_socket, uloop);
  89         int err;
  90         socklen_t errlen = sizeof(err);
  91
  92         if (!u->error) {
  93                 nl_recvmsgs_default(ev->sock);
  94                 return;
  95         }
  96
  97         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
  98                 goto abort;
  99
 100         switch(err) {
 101         case ENOBUFS:
 102                 // Increase rx buffer size on netlink socket
 103                 ev->bufsize *= 2;
 104                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 105                         goto abort;
 106
 107                 // Request full dump since some info got dropped
 108                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 109                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 110                 break;
 111
 112         default:
 113                 goto abort;
 114         }
 115         u->error = false;
 116         return;
 117
 118 abort:
 119         uloop_fd_delete(&ev->uloop);
 120         return;
 121 }
 122
 123 static struct nl_sock *
 124 create_socket(int protocol, int groups)
 125 {
 126         struct nl_sock *sock;
 127
 128         sock = nl_socket_alloc();
 129         if (!sock)
 130                 return NULL;
 131
 132         if (groups)
 133                 nl_join_groups(sock, groups);
 134
 135         if (nl_connect(sock, protocol))
 136                 return NULL;
 137
 138         return sock;
 139 }
 140
 141 static bool
 142 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 143                         uloop_fd_handler cb, int flags)
 144 {
 145         ev->sock = create_socket(protocol, groups);
 146         if (!ev->sock)
 147                 return false;
 148
 149         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 150         ev->uloop.cb = cb;
 151         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 152                 return false;
 153
 154         return true;
 155 }
 156
 157 static bool
 158 create_event_socket(struct event_socket *ev, int protocol,
 159                     int (*cb)(struct nl_msg *msg, void *arg))
 160 {
 161         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 162                 return false;
 163
 164         // Install the valid custom callback handler
 165         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 166
 167         // Disable sequence number checking on event sockets
 168         nl_socket_disable_seq_check(ev->sock);
 169
 170         // Increase rx buffer size to 65K on event sockets
 171         ev->bufsize = 65535;
 172         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 173                 return false;
 174
 175         return true;
 176 }
 177
 178 static bool
 179 system_rtn_aton(const char *src, unsigned int *dst)
 180 {
 181         char *e;
 182         unsigned int n;
 183
 184         if (!strcmp(src, "local"))
 185                 n = RTN_LOCAL;
 186         else if (!strcmp(src, "nat"))
 187                 n = RTN_NAT;
 188         else if (!strcmp(src, "broadcast"))
 189                 n = RTN_BROADCAST;
 190         else if (!strcmp(src, "anycast"))
 191                 n = RTN_ANYCAST;
 192         else if (!strcmp(src, "multicast"))
 193                 n = RTN_MULTICAST;
 194         else if (!strcmp(src, "prohibit"))
 195                 n = RTN_PROHIBIT;
 196         else if (!strcmp(src, "unreachable"))
 197                 n = RTN_UNREACHABLE;
 198         else if (!strcmp(src, "blackhole"))
 199                 n = RTN_BLACKHOLE;
 200         else if (!strcmp(src, "xresolve"))
 201                 n = RTN_XRESOLVE;
 202         else if (!strcmp(src, "unicast"))
 203                 n = RTN_UNICAST;
 204         else if (!strcmp(src, "throw"))
 205                 n = RTN_THROW;
 206         else if (!strcmp(src, "failed_policy"))
 207                 n = RTN_FAILED_POLICY;
 208         else {
 209                 n = strtoul(src, &e, 0);
 210                 if (!e || *e || e == src || n > 255)
 211                         return false;
 212         }
 213
 214         *dst = n;
 215         return true;
 216 }
 217
 218 static bool
 219 system_tos_aton(const char *src, unsigned *dst)
 220 {
 221         char *e;
 222
 223         *dst = strtoul(src, &e, 16);
 224         if (e == src || *e || *dst > 255)
 225                 return false;
 226
 227         return true;
 228 }
 229
 230 int system_init(void)
 231 {
 232         static struct event_socket rtnl_event;
 233         static struct event_socket hotplug_event;
 234
 235         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 236         system_fd_set_cloexec(sock_ioctl);
 237
 238         // Prepare socket for routing / address control
 239         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 240         if (!sock_rtnl)
 241                 return -1;
 242
 243         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 244                 return -1;
 245
 246         if (!create_raw_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT, 1,
 247                                         handle_hotplug_event, 0))
 248                 return -1;
 249
 250         // Receive network link events form kernel
 251         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 252
 253         return 0;
 254 }
 255
 256 static void system_set_sysctl(const char *path, const char *val)
 257 {
 258         int fd;
 259
 260         fd = open(path, O_WRONLY);
 261         if (fd < 0)
 262                 return;
 263
 264         if (write(fd, val, strlen(val))) {}
 265         close(fd);
 266 }
 267
 268 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 269 {
 270         snprintf(dev_buf, sizeof(dev_buf), path, device);
 271         system_set_sysctl(dev_buf, val);
 272 }
 273
 274 static void system_set_disable_ipv6(struct device *dev, const char *val)
 275 {
 276         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 277 }
 278
 279 static void system_set_rpfilter(struct device *dev, const char *val)
 280 {
 281         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter", dev->ifname, val);
 282 }
 283
 284 static void system_set_acceptlocal(struct device *dev, const char *val)
 285 {
 286         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local", dev->ifname, val);
 287 }
 288
 289 static void system_set_igmpversion(struct device *dev, const char *val)
 290 {
 291         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version", dev->ifname, val);
 292 }
 293
 294 static void system_set_mldversion(struct device *dev, const char *val)
 295 {
 296         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version", dev->ifname, val);
 297 }
 298
 299 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 300 {
 301         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 302 }
 303
 304 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 305 {
 306         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 307 }
 308
 309 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 310 {
 311         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time", dev->ifname, val);
 312 }
 313
 314 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 315 {
 316         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time", dev->ifname, val);
 317 }
 318
 319 static void system_set_dadtransmits(struct device *dev, const char *val)
 320 {
 321         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 322 }
 323
 324 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 325 {
 326         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 327 }
 328
 329 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 330 {
 331         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 332 }
 333
 334 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 335 {
 336         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 337 }
 338
 339 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 340 {
 341         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 342                                        "/sys/class/net/%s/brport/multicast_router",
 343                               dev->ifname, val);
 344 }
 345
 346 static void system_bridge_set_robustness(struct device *dev, const char *val)
 347 {
 348         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 349                               dev->ifname, val);
 350         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 351                               dev->ifname, val);
 352 }
 353
 354 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 355 {
 356         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 357                               dev->ifname, val);
 358 }
 359
 360 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 361 {
 362         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 363                               dev->ifname, val);
 364 }
 365
 366 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 367 {
 368         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 369                               dev->ifname, val);
 370 }
 371
 372 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 373 {
 374         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 375                               dev->ifname, val);
 376 }
 377
 378 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 379 {
 380         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 381                               dev->ifname, val);
 382 }
 383
 384 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 385 {
 386         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 387                               dev->ifname, val);
 388 }
 389
 390 static void system_bridge_set_learning(struct device *dev, const char *val)
 391 {
 392         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 393 }
 394
 395 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 396 {
 397         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 398 }
 399
 400 static void system_set_sendredirects(struct device *dev, const char *val)
 401 {
 402         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 403 }
 404
 405 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 406 {
 407         int fd = -1, ret = -1;
 408
 409         fd = open(path, O_RDONLY);
 410         if (fd < 0)
 411                 goto out;
 412
 413         ssize_t len = read(fd, buf, buf_sz - 1);
 414         if (len < 0)
 415                 goto out;
 416
 417         ret = buf[len] = 0;
 418
 419 out:
 420         if (fd >= 0)
 421                 close(fd);
 422
 423         return ret;
 424 }
 425
 426 static int
 427 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 428 {
 429         snprintf(dev_buf, sizeof(dev_buf), path, device);
 430         return system_get_sysctl(dev_buf, buf, buf_sz);
 431 }
 432
 433 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 434 {
 435         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 436                         dev->ifname, buf, buf_sz);
 437 }
 438
 439 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 440 {
 441         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 442                         dev->ifname, buf, buf_sz);
 443 }
 444
 445 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 446 {
 447         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 448                         dev->ifname, buf, buf_sz);
 449 }
 450
 451 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 452 {
 453         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 454                         dev->ifname, buf, buf_sz);
 455 }
 456
 457 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 458 {
 459         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 460                         dev->ifname, buf, buf_sz);
 461 }
 462
 463 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 464 {
 465         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 466                         dev->ifname, buf, buf_sz);
 467 }
 468
 469 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 470 {
 471         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 472                         dev->ifname, buf, buf_sz);
 473 }
 474
 475 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 476 {
 477         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 478                         dev->ifname, buf, buf_sz);
 479 }
 480
 481 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 482 {
 483         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 484                         dev->ifname, buf, buf_sz);
 485 }
 486
 487 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 488 {
 489         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 490                         dev->ifname, buf, buf_sz);
 491 }
 492
 493 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 494 {
 495         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 496                         dev->ifname, buf, buf_sz);
 497 }
 498
 499 // Evaluate netlink messages
 500 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 501 {
 502         struct nlmsghdr *nh = nlmsg_hdr(msg);
 503         struct nlattr *nla[__IFLA_MAX];
 504         int link_state = 0;
 505         char buf[10];
 506
 507         if (nh->nlmsg_type != RTM_NEWLINK)
 508                 goto out;
 509
 510         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 511         if (!nla[IFLA_IFNAME])
 512                 goto out;
 513
 514         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 515         if (!dev)
 516                 goto out;
 517
 518         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 519                 link_state = strtoul(buf, NULL, 0);
 520
 521         device_set_link(dev, link_state ? true : false);
 522
 523 out:
 524         return 0;
 525 }
 526
 527 static void
 528 handle_hotplug_msg(char *data, int size)
 529 {
 530         const char *subsystem = NULL, *interface = NULL;
 531         char *cur, *end, *sep;
 532         struct device *dev;
 533         int skip;
 534         bool add;
 535
 536         if (!strncmp(data, "add@", 4))
 537                 add = true;
 538         else if (!strncmp(data, "remove@", 7))
 539                 add = false;
 540         else
 541                 return;
 542
 543         skip = strlen(data) + 1;
 544         end = data + size;
 545
 546         for (cur = data + skip; cur < end; cur += skip) {
 547                 skip = strlen(cur) + 1;
 548
 549                 sep = strchr(cur, '=');
 550                 if (!sep)
 551                         continue;
 552
 553                 *sep = 0;
 554                 if (!strcmp(cur, "INTERFACE"))
 555                         interface = sep + 1;
 556                 else if (!strcmp(cur, "SUBSYSTEM")) {
 557                         subsystem = sep + 1;
 558                         if (strcmp(subsystem, "net") != 0)
 559                                 return;
 560                 }
 561                 if (subsystem && interface)
 562                         goto found;
 563         }
 564         return;
 565
 566 found:
 567         dev = device_find(interface);
 568         if (!dev)
 569                 return;
 570
 571         if (dev->type != &simple_device_type)
 572                 return;
 573
 574         if (add && system_if_force_external(dev->ifname))
 575                 return;
 576
 577         device_set_present(dev, add);
 578 }
 579
 580 static void
 581 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 582 {
 583         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 584         struct sockaddr_nl nla;
 585         unsigned char *buf = NULL;
 586         int size;
 587
 588         while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 589                 if (nla.nl_pid == 0)
 590                         handle_hotplug_msg((char *) buf, size);
 591
 592                 free(buf);
 593         }
 594 }
 595
 596 static int system_rtnl_call(struct nl_msg *msg)
 597 {
 598         int ret;
 599
 600         ret = nl_send_auto_complete(sock_rtnl, msg);
 601         nlmsg_free(msg);
 602
 603         if (ret < 0)
 604                 return ret;
 605
 606         return nl_wait_for_ack(sock_rtnl);
 607 }
 608
 609 int system_bridge_delbr(struct device *bridge)
 610 {
 611         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 612 }
 613
 614 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 615 {
 616         struct ifreq ifr;
 617
 618         memset(&ifr, 0, sizeof(ifr));
 619         if (dev)
 620                 ifr.ifr_ifindex = dev->ifindex;
 621         else
 622                 ifr.ifr_data = data;
 623         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name));
 624         return ioctl(sock_ioctl, cmd, &ifr);
 625 }
 626
 627 static bool system_is_bridge(const char *name, char *buf, int buflen)
 628 {
 629         struct stat st;
 630
 631         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 632         if (stat(buf, &st) < 0)
 633                 return false;
 634
 635         return true;
 636 }
 637
 638 static char *system_get_bridge(const char *name, char *buf, int buflen)
 639 {
 640         char *path;
 641         ssize_t len = -1;
 642         glob_t gl;
 643
 644         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 645         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 646                 return NULL;
 647
 648         if (gl.gl_pathc > 0)
 649                 len = readlink(gl.gl_pathv[0], buf, buflen);
 650
 651         globfree(&gl);
 652
 653         if (len < 0)
 654                 return NULL;
 655
 656         buf[len] = 0;
 657         path = strrchr(buf, '/');
 658         if (!path)
 659                 return NULL;
 660
 661         return path + 1;
 662 }
 663
 664 static void
 665 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 666 {
 667         bool mcast_to_ucast = dev->wireless_ap;
 668         bool hairpin = true;
 669
 670         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 671             !bridge->settings.multicast_to_unicast)
 672                 mcast_to_ucast = false;
 673
 674         if (!mcast_to_ucast || dev->wireless_isolate)
 675                 hairpin = false;
 676
 677         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 678         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 679 }
 680
 681 int system_bridge_addif(struct device *bridge, struct device *dev)
 682 {
 683         char buf[64];
 684         char *oldbr;
 685         int ret = 0;
 686
 687         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 688         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 689                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 690
 691         if (dev->wireless)
 692                 system_bridge_set_wireless(bridge, dev);
 693
 694         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 695                 snprintf(buf, sizeof(buf), "%i", dev->settings.multicast_router);
 696                 system_bridge_set_multicast_router(dev, buf, false);
 697         }
 698
 699         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 700             dev->settings.multicast_fast_leave)
 701                 system_bridge_set_multicast_fast_leave(dev, "1");
 702
 703         if (dev->settings.flags & DEV_OPT_LEARNING &&
 704             !dev->settings.learning)
 705                 system_bridge_set_learning(dev, "0");
 706
 707         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 708             !dev->settings.unicast_flood)
 709                 system_bridge_set_unicast_flood(dev, "0");
 710
 711         return ret;
 712 }
 713
 714 int system_bridge_delif(struct device *bridge, struct device *dev)
 715 {
 716         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 717 }
 718
 719 int system_if_resolve(struct device *dev)
 720 {
 721         struct ifreq ifr;
 722         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
 723         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 724                 return ifr.ifr_ifindex;
 725         else
 726                 return 0;
 727 }
 728
 729 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 730 {
 731         struct ifreq ifr;
 732
 733         memset(&ifr, 0, sizeof(ifr));
 734         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name));
 735         ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr);
 736         ifr.ifr_flags |= add;
 737         ifr.ifr_flags &= ~rem;
 738         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 739 }
 740
 741 struct clear_data {
 742         struct nl_msg *msg;
 743         struct device *dev;
 744         int type;
 745         int size;
 746         int af;
 747 };
 748
 749
 750 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 751 {
 752         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 753
 754         return ifa->ifa_index == ifindex;
 755 }
 756
 757 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 758 {
 759         struct rtmsg *r = NLMSG_DATA(hdr);
 760         struct nlattr *tb[__RTA_MAX];
 761
 762         if (r->rtm_protocol == RTPROT_KERNEL &&
 763             r->rtm_family == AF_INET6)
 764                 return false;
 765
 766         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 767         if (!tb[RTA_OIF])
 768                 return false;
 769
 770         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 771 }
 772
 773 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 774 {
 775         return true;
 776 }
 777
 778 static int cb_clear_event(struct nl_msg *msg, void *arg)
 779 {
 780         struct clear_data *clr = arg;
 781         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 782         bool (*cb)(struct nlmsghdr *, int ifindex);
 783         int type;
 784
 785         switch(clr->type) {
 786         case RTM_GETADDR:
 787                 type = RTM_DELADDR;
 788                 if (hdr->nlmsg_type != RTM_NEWADDR)
 789                         return NL_SKIP;
 790
 791                 cb = check_ifaddr;
 792                 break;
 793         case RTM_GETROUTE:
 794                 type = RTM_DELROUTE;
 795                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 796                         return NL_SKIP;
 797
 798                 cb = check_route;
 799                 break;
 800         case RTM_GETRULE:
 801                 type = RTM_DELRULE;
 802                 if (hdr->nlmsg_type != RTM_NEWRULE)
 803                         return NL_SKIP;
 804
 805                 cb = check_rule;
 806                 break;
 807         default:
 808                 return NL_SKIP;
 809         }
 810
 811         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 812                 return NL_SKIP;
 813
 814         if (type == RTM_DELRULE)
 815                 D(SYSTEM, "Remove a rule\n");
 816         else
 817                 D(SYSTEM, "Remove %s from device %s\n",
 818                   type == RTM_DELADDR ? "an address" : "a route",
 819                   clr->dev->ifname);
 820         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 821         hdr = nlmsg_hdr(clr->msg);
 822         hdr->nlmsg_type = type;
 823         hdr->nlmsg_flags = NLM_F_REQUEST;
 824
 825         nl_socket_disable_auto_ack(sock_rtnl);
 826         nl_send_auto_complete(sock_rtnl, clr->msg);
 827         nl_socket_enable_auto_ack(sock_rtnl);
 828
 829         return NL_SKIP;
 830 }
 831
 832 static int
 833 cb_finish_event(struct nl_msg *msg, void *arg)
 834 {
 835         int *pending = arg;
 836         *pending = 0;
 837         return NL_STOP;
 838 }
 839
 840 static int
 841 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 842 {
 843         int *pending = arg;
 844         *pending = err->error;
 845         return NL_STOP;
 846 }
 847
 848 static void
 849 system_if_clear_entries(struct device *dev, int type, int af)
 850 {
 851         struct clear_data clr;
 852         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
 853         struct rtmsg rtm = {
 854                 .rtm_family = af,
 855                 .rtm_flags = RTM_F_CLONED,
 856         };
 857         int flags = NLM_F_DUMP;
 858         int pending = 1;
 859
 860         clr.af = af;
 861         clr.dev = dev;
 862         clr.type = type;
 863         switch (type) {
 864         case RTM_GETADDR:
 865         case RTM_GETRULE:
 866                 clr.size = sizeof(struct rtgenmsg);
 867                 break;
 868         case RTM_GETROUTE:
 869                 clr.size = sizeof(struct rtmsg);
 870                 break;
 871         default:
 872                 return;
 873         }
 874
 875         if (!cb)
 876                 return;
 877
 878         clr.msg = nlmsg_alloc_simple(type, flags);
 879         if (!clr.msg)
 880                 goto out;
 881
 882         nlmsg_append(clr.msg, &rtm, clr.size, 0);
 883         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
 884         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
 885         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
 886
 887         nl_send_auto_complete(sock_rtnl, clr.msg);
 888         while (pending > 0)
 889                 nl_recvmsgs(sock_rtnl, cb);
 890
 891         nlmsg_free(clr.msg);
 892 out:
 893         nl_cb_put(cb);
 894 }
 895
 896 /*
 897  * Clear bridge (membership) state and bring down device
 898  */
 899 void system_if_clear_state(struct device *dev)
 900 {
 901         static char buf[256];
 902         char *bridge;
 903
 904         device_set_ifindex(dev, system_if_resolve(dev));
 905         if (dev->external || !dev->ifindex)
 906                 return;
 907
 908         system_if_flags(dev->ifname, 0, IFF_UP);
 909
 910         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
 911                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
 912                 system_bridge_delbr(dev);
 913                 return;
 914         }
 915
 916         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
 917         if (bridge) {
 918                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
 919                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
 920         }
 921
 922         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
 923         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
 924         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
 925         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
 926         system_set_disable_ipv6(dev, "0");
 927 }
 928
 929 static inline unsigned long
 930 sec_to_jiffies(int val)
 931 {
 932         return (unsigned long) val * 100;
 933 }
 934
 935 static void system_bridge_conf_multicast_deps(struct device *bridge,
 936                                               struct bridge_config *cfg,
 937                                               char *buf,
 938                                               int buf_len)
 939 {
 940         int val;
 941
 942         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
 943             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
 944             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
 945                 val = cfg->robustness * cfg->query_interval +
 946                         cfg->query_response_interval;
 947
 948                 snprintf(buf, buf_len, "%i", val);
 949                 system_bridge_set_membership_interval(bridge, buf);
 950
 951                 val = cfg->robustness * cfg->query_interval +
 952                         cfg->query_response_interval / 2;
 953
 954                 snprintf(buf, buf_len, "%i", val);
 955                 system_bridge_set_other_querier_timeout(bridge, buf);
 956         }
 957
 958         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
 959                 val = cfg->query_interval / 4;
 960
 961                 snprintf(buf, buf_len, "%i", val);
 962                 system_bridge_set_startup_query_interval(bridge, buf);
 963         }
 964 }
 965
 966 static void system_bridge_conf_multicast(struct device *bridge,
 967                                          struct bridge_config *cfg,
 968                                          char *buf,
 969                                          int buf_len)
 970 {
 971         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
 972                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
 973
 974         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
 975                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
 976
 977         snprintf(buf, buf_len, "%i", cfg->hash_max);
 978         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
 979                 bridge->ifname, buf);
 980
 981         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 982                 snprintf(buf, buf_len, "%i", bridge->settings.multicast_router);
 983                 system_bridge_set_multicast_router(bridge, buf, true);
 984         }
 985
 986         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
 987                 snprintf(buf, buf_len, "%i", cfg->robustness);
 988                 system_bridge_set_robustness(bridge, buf);
 989         }
 990
 991         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
 992                 snprintf(buf, buf_len, "%i", cfg->query_interval);
 993                 system_bridge_set_query_interval(bridge, buf);
 994         }
 995
 996         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
 997                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
 998                 system_bridge_set_query_response_interval(bridge, buf);
 999         }
1000
1001         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1002                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1003                 system_bridge_set_last_member_interval(bridge, buf);
1004         }
1005
1006         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1007 }
1008
1009 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1010 {
1011         char buf[64];
1012         unsigned long args[4] = {};
1013
1014         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1015                 return -1;
1016
1017         args[0] = BRCTL_SET_BRIDGE_STP_STATE;
1018         args[1] = !!cfg->stp;
1019         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1020
1021         args[0] = BRCTL_SET_BRIDGE_FORWARD_DELAY;
1022         args[1] = sec_to_jiffies(cfg->forward_delay);
1023         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1024
1025         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1026
1027         args[0] = BRCTL_SET_BRIDGE_PRIORITY;
1028         args[1] = cfg->priority;
1029         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1030
1031         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1032                 args[0] = BRCTL_SET_AGEING_TIME;
1033                 args[1] = sec_to_jiffies(cfg->ageing_time);
1034                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1035         }
1036
1037         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1038                 args[0] = BRCTL_SET_BRIDGE_HELLO_TIME;
1039                 args[1] = sec_to_jiffies(cfg->hello_time);
1040                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1041         }
1042
1043         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1044                 args[0] = BRCTL_SET_BRIDGE_MAX_AGE;
1045                 args[1] = sec_to_jiffies(cfg->max_age);
1046                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1047         }
1048
1049         return 0;
1050 }
1051
1052 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1053 {
1054         struct nl_msg *msg;
1055         struct nlattr *linkinfo, *data;
1056         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1057         int i, rv;
1058         static const struct {
1059                 const char *name;
1060                 enum macvlan_mode val;
1061         } modes[] = {
1062                 { "private", MACVLAN_MODE_PRIVATE },
1063                 { "vepa", MACVLAN_MODE_VEPA },
1064                 { "bridge", MACVLAN_MODE_BRIDGE },
1065                 { "passthru", MACVLAN_MODE_PASSTHRU },
1066         };
1067
1068         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1069
1070         if (!msg)
1071                 return -1;
1072
1073         nlmsg_append(msg, &iim, sizeof(iim), 0);
1074
1075         if (cfg->flags & MACVLAN_OPT_MACADDR)
1076                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1077         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1078         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1079
1080         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1081                 goto nla_put_failure;
1082
1083         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1084
1085         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1086                 goto nla_put_failure;
1087
1088         if (cfg->mode) {
1089                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1090                         if (strcmp(cfg->mode, modes[i].name) != 0)
1091                                 continue;
1092
1093                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1094                         break;
1095                 }
1096         }
1097
1098         nla_nest_end(msg, data);
1099         nla_nest_end(msg, linkinfo);
1100
1101         rv = system_rtnl_call(msg);
1102         if (rv)
1103                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1104
1105         return rv;
1106
1107 nla_put_failure:
1108         nlmsg_free(msg);
1109         return -ENOMEM;
1110 }
1111
1112 static int system_link_del(const char *ifname)
1113 {
1114         struct nl_msg *msg;
1115         struct ifinfomsg iim = {
1116                 .ifi_family = AF_UNSPEC,
1117                 .ifi_index = 0,
1118         };
1119
1120         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1121
1122         if (!msg)
1123                 return -1;
1124
1125         nlmsg_append(msg, &iim, sizeof(iim), 0);
1126         nla_put_string(msg, IFLA_IFNAME, ifname);
1127         return system_rtnl_call(msg);
1128 }
1129
1130 int system_macvlan_del(struct device *macvlan)
1131 {
1132         return system_link_del(macvlan->ifname);
1133 }
1134
1135 int system_veth_add(struct device *veth, struct veth_config *cfg)
1136 {
1137         struct nl_msg *msg;
1138         struct ifinfomsg empty_iim = {};
1139         struct nlattr *linkinfo, *data, *veth_info;
1140         int rv;
1141
1142         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1143
1144         if (!msg)
1145                 return -1;
1146
1147         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1148
1149         if (cfg->flags & VETH_OPT_MACADDR)
1150                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1151         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1152
1153         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1154                 goto nla_put_failure;
1155
1156         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1157
1158         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1159                 goto nla_put_failure;
1160
1161         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1162                 goto nla_put_failure;
1163
1164         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1165
1166         if (cfg->flags & VETH_OPT_PEER_NAME)
1167                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1168         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1169                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1170
1171         nla_nest_end(msg, veth_info);
1172         nla_nest_end(msg, data);
1173         nla_nest_end(msg, linkinfo);
1174
1175         rv = system_rtnl_call(msg);
1176         if (rv) {
1177                 if (cfg->flags & VETH_OPT_PEER_NAME)
1178                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1179                 else
1180                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1181         }
1182
1183         return rv;
1184
1185 nla_put_failure:
1186         nlmsg_free(msg);
1187         return -ENOMEM;
1188 }
1189
1190 int system_veth_del(struct device *veth)
1191 {
1192         return system_link_del(veth->ifname);
1193 }
1194
1195 static int system_vlan(struct device *dev, int id)
1196 {
1197         struct vlan_ioctl_args ifr = {
1198                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1199                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1200         };
1201
1202         ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1203
1204         if (id < 0) {
1205                 ifr.cmd = DEL_VLAN_CMD;
1206                 ifr.u.VID = 0;
1207         } else {
1208                 ifr.cmd = ADD_VLAN_CMD;
1209                 ifr.u.VID = id;
1210         }
1211         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1212         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1213 }
1214
1215 int system_vlan_add(struct device *dev, int id)
1216 {
1217         return system_vlan(dev, id);
1218 }
1219
1220 int system_vlan_del(struct device *dev)
1221 {
1222         return system_vlan(dev, -1);
1223 }
1224
1225 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1226 {
1227         struct nl_msg *msg;
1228         struct nlattr *linkinfo, *data;
1229         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1230         int rv;
1231
1232         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1233
1234         if (!msg)
1235                 return -1;
1236
1237         nlmsg_append(msg, &iim, sizeof(iim), 0);
1238         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1239         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1240
1241         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1242                 goto nla_put_failure;
1243
1244         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1245
1246         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1247                 goto nla_put_failure;
1248
1249         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1250
1251 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1252         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1253 #else
1254         if(cfg->proto == VLAN_PROTO_8021AD)
1255                 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);
1256 #endif
1257
1258         nla_nest_end(msg, data);
1259         nla_nest_end(msg, linkinfo);
1260
1261         rv = system_rtnl_call(msg);
1262         if (rv)
1263                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1264
1265         return rv;
1266
1267 nla_put_failure:
1268         nlmsg_free(msg);
1269         return -ENOMEM;
1270 }
1271
1272 int system_vlandev_del(struct device *vlandev)
1273 {
1274         return system_link_del(vlandev->ifname);
1275 }
1276
1277 void
1278 system_if_get_settings(struct device *dev, struct device_settings *s)
1279 {
1280         struct ifreq ifr;
1281         char buf[10];
1282
1283         memset(&ifr, 0, sizeof(ifr));
1284         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1285
1286         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1287                 s->mtu = ifr.ifr_mtu;
1288                 s->flags |= DEV_OPT_MTU;
1289         }
1290
1291         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1292         if (s->mtu6 > 0)
1293                 s->flags |= DEV_OPT_MTU6;
1294
1295         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1296                 s->txqueuelen = ifr.ifr_qlen;
1297                 s->flags |= DEV_OPT_TXQUEUELEN;
1298         }
1299
1300         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1301                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1302                 s->flags |= DEV_OPT_MACADDR;
1303         }
1304
1305         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1306                 s->ipv6 = !strtoul(buf, NULL, 0);
1307                 s->flags |= DEV_OPT_IPV6;
1308         }
1309
1310         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1311                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1312                 s->flags |= DEV_OPT_PROMISC;
1313
1314                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1315                 s->flags |= DEV_OPT_MULTICAST;
1316         }
1317
1318         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1319                 s->rpfilter = strtoul(buf, NULL, 0);
1320                 s->flags |= DEV_OPT_RPFILTER;
1321         }
1322
1323         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1324                 s->acceptlocal = strtoul(buf, NULL, 0);
1325                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1326         }
1327
1328         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1329                 s->igmpversion = strtoul(buf, NULL, 0);
1330                 s->flags |= DEV_OPT_IGMPVERSION;
1331         }
1332
1333         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1334                 s->mldversion = strtoul(buf, NULL, 0);
1335                 s->flags |= DEV_OPT_MLDVERSION;
1336         }
1337
1338         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1339                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1340                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1341         }
1342
1343         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1344                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1345                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1346         }
1347
1348         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1349                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1350                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1351         }
1352
1353         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1354                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1355                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1356         }
1357
1358         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1359                 s->dadtransmits = strtoul(buf, NULL, 0);
1360                 s->flags |= DEV_OPT_DADTRANSMITS;
1361         }
1362
1363         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1364                 s->sendredirects = strtoul(buf, NULL, 0);
1365                 s->flags |= DEV_OPT_SENDREDIRECTS;
1366         }
1367 }
1368
1369 static void
1370 system_if_set_rps_xps_val(const char *path, int val)
1371 {
1372         char val_buf[8];
1373         glob_t gl;
1374         int i;
1375
1376         if (glob(path, 0, NULL, &gl))
1377                 return;
1378
1379         snprintf(val_buf, sizeof(val_buf), "%x", val);
1380         for (i = 0; i < gl.gl_pathc; i++)
1381                 system_set_sysctl(gl.gl_pathv[i], val_buf);
1382
1383         globfree(&gl);
1384 }
1385
1386 static void
1387 system_if_apply_rps_xps(struct device *dev, struct device_settings *s)
1388 {
1389         long n_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1390         int val;
1391
1392         if (n_cpus < 2)
1393                 return;
1394
1395         val = (1 << n_cpus) - 1;
1396         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/rps_cpus", dev->ifname);
1397         system_if_set_rps_xps_val(dev_buf, s->rps ? val : 0);
1398
1399         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/xps_cpus", dev->ifname);
1400         system_if_set_rps_xps_val(dev_buf, s->xps ? val : 0);
1401 }
1402
1403 void
1404 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1405 {
1406         struct ifreq ifr;
1407         char buf[12];
1408
1409         memset(&ifr, 0, sizeof(ifr));
1410         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1411         if (s->flags & DEV_OPT_MTU & apply_mask) {
1412                 ifr.ifr_mtu = s->mtu;
1413                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1414                         s->flags &= ~DEV_OPT_MTU;
1415         }
1416         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1417                 system_update_ipv6_mtu(dev, s->mtu6);
1418         }
1419         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1420                 ifr.ifr_qlen = s->txqueuelen;
1421                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1422                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1423         }
1424         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1425                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1426                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1427                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1428                         s->flags &= ~DEV_OPT_MACADDR;
1429         }
1430         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1431                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1432         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1433                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1434                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1435                         s->flags &= ~DEV_OPT_PROMISC;
1436         }
1437         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1438                 snprintf(buf, sizeof(buf), "%d", s->rpfilter);
1439                 system_set_rpfilter(dev, buf);
1440         }
1441         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1442                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1443         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1444                 snprintf(buf, sizeof(buf), "%d", s->igmpversion);
1445                 system_set_igmpversion(dev, buf);
1446         }
1447         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1448                 snprintf(buf, sizeof(buf), "%d", s->mldversion);
1449                 system_set_mldversion(dev, buf);
1450         }
1451         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1452                 snprintf(buf, sizeof(buf), "%d", s->neigh4reachabletime);
1453                 system_set_neigh4reachabletime(dev, buf);
1454                 snprintf(buf, sizeof(buf), "%d", s->neigh6reachabletime);
1455                 system_set_neigh6reachabletime(dev, buf);
1456         }
1457         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1458                 snprintf(buf, sizeof(buf), "%d", s->neigh4gcstaletime);
1459                 system_set_neigh4gcstaletime(dev, buf);
1460                 snprintf(buf, sizeof(buf), "%d", s->neigh6gcstaletime);
1461                 system_set_neigh6gcstaletime(dev, buf);
1462         }
1463         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1464                 snprintf(buf, sizeof(buf), "%d", s->dadtransmits);
1465                 system_set_dadtransmits(dev, buf);
1466         }
1467         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1468                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1469                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1470                         s->flags &= ~DEV_OPT_MULTICAST;
1471         }
1472         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1473                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1474
1475         system_if_apply_rps_xps(dev, s);
1476 }
1477
1478 int system_if_up(struct device *dev)
1479 {
1480         system_if_get_settings(dev, &dev->orig_settings);
1481         /* Only keep orig settings based on what needs to be set */
1482         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1483         dev->orig_settings.flags &= dev->settings.flags;
1484         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1485         return system_if_flags(dev->ifname, IFF_UP, 0);
1486 }
1487
1488 int system_if_down(struct device *dev)
1489 {
1490         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1491         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1492         return ret;
1493 }
1494
1495 struct if_check_data {
1496         struct device *dev;
1497         int pending;
1498         int ret;
1499 };
1500
1501 #ifndef IFF_LOWER_UP
1502 #define IFF_LOWER_UP    0x10000
1503 #endif
1504
1505 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1506 {
1507         struct nlmsghdr *nh = nlmsg_hdr(msg);
1508         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1509         struct if_check_data *chk = (struct if_check_data *)arg;
1510
1511         if (nh->nlmsg_type != RTM_NEWLINK)
1512                 return NL_SKIP;
1513
1514         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1515         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1516
1517         return NL_OK;
1518 }
1519
1520 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1521 {
1522         struct if_check_data *chk = (struct if_check_data *)arg;
1523         chk->pending = 0;
1524         return NL_STOP;
1525 }
1526
1527 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1528 {
1529         struct if_check_data *chk = (struct if_check_data *)arg;
1530
1531         device_set_present(chk->dev, false);
1532         device_set_link(chk->dev, false);
1533         chk->pending = err->error;
1534
1535         return NL_STOP;
1536 }
1537
1538 int system_if_check(struct device *dev)
1539 {
1540         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1541         struct nl_msg *msg;
1542         struct ifinfomsg ifi = {
1543                 .ifi_family = AF_UNSPEC,
1544                 .ifi_index = 0,
1545         };
1546         struct if_check_data chk = {
1547                 .dev = dev,
1548                 .pending = 1,
1549         };
1550         int ret = 1;
1551
1552         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1553         if (!msg)
1554                 goto out;
1555
1556         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1557             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1558                 goto free;
1559
1560         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1561         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1562         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1563
1564         nl_send_auto_complete(sock_rtnl, msg);
1565         while (chk.pending > 0)
1566                 nl_recvmsgs(sock_rtnl, cb);
1567
1568         ret = chk.pending;
1569
1570 free:
1571         nlmsg_free(msg);
1572 out:
1573         nl_cb_put(cb);
1574         return ret;
1575 }
1576
1577 struct device *
1578 system_if_get_parent(struct device *dev)
1579 {
1580         char buf[64], *devname;
1581         int ifindex, iflink, len;
1582         FILE *f;
1583
1584         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1585         f = fopen(buf, "r");
1586         if (!f)
1587                 return NULL;
1588
1589         len = fread(buf, 1, sizeof(buf) - 1, f);
1590         fclose(f);
1591
1592         if (len <= 0)
1593                 return NULL;
1594
1595         buf[len] = 0;
1596         iflink = strtoul(buf, NULL, 0);
1597         ifindex = system_if_resolve(dev);
1598         if (!iflink || iflink == ifindex)
1599                 return NULL;
1600
1601         devname = if_indextoname(iflink, buf);
1602         if (!devname)
1603                 return NULL;
1604
1605         return device_get(devname, true);
1606 }
1607
1608 static bool
1609 read_string_file(int dir_fd, const char *file, char *buf, int len)
1610 {
1611         bool ret = false;
1612         char *c;
1613         int fd;
1614
1615         fd = openat(dir_fd, file, O_RDONLY);
1616         if (fd < 0)
1617                 return false;
1618
1619 retry:
1620         len = read(fd, buf, len - 1);
1621         if (len < 0) {
1622                 if (errno == EINTR)
1623                         goto retry;
1624         } else if (len > 0) {
1625                         buf[len] = 0;
1626
1627                         c = strchr(buf, '\n');
1628                         if (c)
1629                                 *c = 0;
1630
1631                         ret = true;
1632         }
1633
1634         close(fd);
1635
1636         return ret;
1637 }
1638
1639 static bool
1640 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1641 {
1642         char buf[64];
1643         bool ret = false;
1644
1645         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1646         if (ret)
1647                 *val = strtoull(buf, NULL, 0);
1648
1649         return ret;
1650 }
1651
1652 /* Assume advertised flags == supported flags */
1653 static const struct {
1654         uint32_t mask;
1655         const char *name;
1656 } ethtool_link_modes[] = {
1657         { ADVERTISED_10baseT_Half, "10H" },
1658         { ADVERTISED_10baseT_Full, "10F" },
1659         { ADVERTISED_100baseT_Half, "100H" },
1660         { ADVERTISED_100baseT_Full, "100F" },
1661         { ADVERTISED_1000baseT_Half, "1000H" },
1662         { ADVERTISED_1000baseT_Full, "1000F" },
1663 };
1664
1665 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1666 {
1667         int i;
1668         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1669                 if (mask & ethtool_link_modes[i].mask)
1670                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1671         }
1672 }
1673
1674 bool
1675 system_if_force_external(const char *ifname)
1676 {
1677         char buf[64];
1678         struct stat s;
1679
1680         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1681         return stat(buf, &s) == 0;
1682 }
1683
1684 int
1685 system_if_dump_info(struct device *dev, struct blob_buf *b)
1686 {
1687         struct ethtool_cmd ecmd;
1688         struct ifreq ifr;
1689         char buf[64], *s;
1690         void *c;
1691         int dir_fd;
1692
1693         snprintf(buf, sizeof(buf), "/sys/class/net/%s", dev->ifname);
1694         dir_fd = open(buf, O_DIRECTORY);
1695
1696         memset(&ecmd, 0, sizeof(ecmd));
1697         memset(&ifr, 0, sizeof(ifr));
1698         strcpy(ifr.ifr_name, dev->ifname);
1699         ifr.ifr_data = (caddr_t) &ecmd;
1700         ecmd.cmd = ETHTOOL_GSET;
1701
1702         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1703                 c = blobmsg_open_array(b, "link-advertising");
1704                 system_add_link_modes(b, ecmd.advertising);
1705                 blobmsg_close_array(b, c);
1706
1707                 c = blobmsg_open_array(b, "link-supported");
1708                 system_add_link_modes(b, ecmd.supported);
1709                 blobmsg_close_array(b, c);
1710
1711                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1712                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1713                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1714                 blobmsg_add_string_buffer(b);
1715         }
1716
1717         close(dir_fd);
1718         return 0;
1719 }
1720
1721 int
1722 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1723 {
1724         const char *const counters[] = {
1725                 "collisions",     "rx_frame_errors",   "tx_compressed",
1726                 "multicast",      "rx_length_errors",  "tx_dropped",
1727                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1728                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1729                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1730                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1731                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1732                 "rx_fifo_errors", "tx_carrier_errors",
1733         };
1734         char buf[64];
1735         int stats_dir;
1736         int i;
1737         uint64_t val = 0;
1738
1739         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1740         stats_dir = open(buf, O_DIRECTORY);
1741         if (stats_dir < 0)
1742                 return -1;
1743
1744         for (i = 0; i < ARRAY_SIZE(counters); i++)
1745                 if (read_uint64_file(stats_dir, counters[i], &val))
1746                         blobmsg_add_u64(b, counters[i], val);
1747
1748         close(stats_dir);
1749         return 0;
1750 }
1751
1752 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1753 {
1754         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1755         int alen = v4 ? 4 : 16;
1756         unsigned int flags = 0;
1757         struct ifaddrmsg ifa = {
1758                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1759                 .ifa_prefixlen = addr->mask,
1760                 .ifa_index = dev->ifindex,
1761         };
1762
1763         struct nl_msg *msg;
1764         if (cmd == RTM_NEWADDR)
1765                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1766
1767         msg = nlmsg_alloc_simple(cmd, flags);
1768         if (!msg)
1769                 return -1;
1770
1771         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1772         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1773         if (v4) {
1774                 if (addr->broadcast)
1775                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1776                 if (addr->point_to_point)
1777                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1778         } else {
1779                 time_t now = system_get_rtime();
1780                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1781
1782                 if (addr->preferred_until) {
1783                         int64_t preferred = addr->preferred_until - now;
1784                         if (preferred < 0)
1785                                 preferred = 0;
1786                         else if (preferred > UINT32_MAX)
1787                                 preferred = UINT32_MAX;
1788
1789                         cinfo.ifa_prefered = preferred;
1790                 }
1791
1792                 if (addr->valid_until) {
1793                         int64_t valid = addr->valid_until - now;
1794                         if (valid <= 0) {
1795                                 nlmsg_free(msg);
1796                                 return -1;
1797                         }
1798                         else if (valid > UINT32_MAX)
1799                                 valid = UINT32_MAX;
1800
1801                         cinfo.ifa_valid = valid;
1802                 }
1803
1804                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
1805
1806                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
1807                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
1808         }
1809
1810         return system_rtnl_call(msg);
1811 }
1812
1813 int system_add_address(struct device *dev, struct device_addr *addr)
1814 {
1815         return system_addr(dev, addr, RTM_NEWADDR);
1816 }
1817
1818 int system_del_address(struct device *dev, struct device_addr *addr)
1819 {
1820         return system_addr(dev, addr, RTM_DELADDR);
1821 }
1822
1823 static int system_rt(struct device *dev, struct device_route *route, int cmd)
1824 {
1825         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1826         bool have_gw;
1827         unsigned int flags = 0;
1828
1829         if (alen == 4)
1830                 have_gw = !!route->nexthop.in.s_addr;
1831         else
1832                 have_gw = route->nexthop.in6.s6_addr32[0] ||
1833                         route->nexthop.in6.s6_addr32[1] ||
1834                         route->nexthop.in6.s6_addr32[2] ||
1835                         route->nexthop.in6.s6_addr32[3];
1836
1837         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
1838                         ? route->table : RT_TABLE_MAIN;
1839
1840         struct rtmsg rtm = {
1841                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
1842                 .rtm_dst_len = route->mask,
1843                 .rtm_src_len = route->sourcemask,
1844                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
1845                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
1846                 .rtm_scope = RT_SCOPE_NOWHERE,
1847                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
1848                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
1849         };
1850         struct nl_msg *msg;
1851
1852         if (cmd == RTM_NEWROUTE) {
1853                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1854
1855                 if (!dev) { // Add null-route
1856                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1857                         rtm.rtm_type = RTN_UNREACHABLE;
1858                 }
1859                 else
1860                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
1861         }
1862
1863         if (route->flags & DEVROUTE_TYPE) {
1864                 rtm.rtm_type = route->type;
1865                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
1866                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
1867                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
1868                                 rtm.rtm_table = RT_TABLE_LOCAL;
1869                 }
1870
1871                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
1872                         rtm.rtm_scope = RT_SCOPE_HOST;
1873                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
1874                                 rtm.rtm_type == RTN_ANYCAST) {
1875                         rtm.rtm_scope = RT_SCOPE_LINK;
1876                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
1877                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY) {
1878                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1879                         dev = NULL;
1880                 }
1881         }
1882
1883         msg = nlmsg_alloc_simple(cmd, flags);
1884         if (!msg)
1885                 return -1;
1886
1887         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
1888
1889         if (route->mask)
1890                 nla_put(msg, RTA_DST, alen, &route->addr);
1891
1892         if (route->sourcemask) {
1893                 if (rtm.rtm_family == AF_INET)
1894                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
1895                 else
1896                         nla_put(msg, RTA_SRC, alen, &route->source);
1897         }
1898
1899         if (route->metric > 0)
1900                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
1901
1902         if (have_gw)
1903                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
1904
1905         if (dev)
1906                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
1907
1908         if (table >= 256)
1909                 nla_put_u32(msg, RTA_TABLE, table);
1910
1911         if (route->flags & DEVROUTE_MTU) {
1912                 struct nlattr *metrics;
1913
1914                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
1915                         goto nla_put_failure;
1916
1917                 nla_put_u32(msg, RTAX_MTU, route->mtu);
1918
1919                 nla_nest_end(msg, metrics);
1920         }
1921
1922         return system_rtnl_call(msg);
1923
1924 nla_put_failure:
1925         nlmsg_free(msg);
1926         return -ENOMEM;
1927 }
1928
1929 int system_add_route(struct device *dev, struct device_route *route)
1930 {
1931         return system_rt(dev, route, RTM_NEWROUTE);
1932 }
1933
1934 int system_del_route(struct device *dev, struct device_route *route)
1935 {
1936         return system_rt(dev, route, RTM_DELROUTE);
1937 }
1938
1939 int system_flush_routes(void)
1940 {
1941         const char *names[] = {
1942                 "/proc/sys/net/ipv4/route/flush",
1943                 "/proc/sys/net/ipv6/route/flush"
1944         };
1945         int fd, i;
1946
1947         for (i = 0; i < ARRAY_SIZE(names); i++) {
1948                 fd = open(names[i], O_WRONLY);
1949                 if (fd < 0)
1950                         continue;
1951
1952                 if (write(fd, "-1", 2)) {}
1953                 close(fd);
1954         }
1955         return 0;
1956 }
1957
1958 bool system_resolve_rt_type(const char *type, unsigned int *id)
1959 {
1960         return system_rtn_aton(type, id);
1961 }
1962
1963 bool system_resolve_rt_proto(const char *type, unsigned int *id)
1964 {
1965         FILE *f;
1966         char *e, buf[128];
1967         unsigned int n, proto = 256;
1968
1969         if ((n = strtoul(type, &e, 0)) >= 0 && !*e && e != type)
1970                 proto = n;
1971         else if (!strcmp(type, "unspec"))
1972                 proto = RTPROT_UNSPEC;
1973         else if (!strcmp(type, "kernel"))
1974                 proto = RTPROT_KERNEL;
1975         else if (!strcmp(type, "boot"))
1976                 proto = RTPROT_BOOT;
1977         else if (!strcmp(type, "static"))
1978                 proto = RTPROT_STATIC;
1979         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
1980                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
1981                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
1982                                 continue;
1983
1984                         n = strtoul(e, NULL, 10);
1985                         e = strtok(NULL, " \t\n");
1986
1987                         if (e && !strcmp(e, type)) {
1988                                 proto = n;
1989                                 break;
1990                         }
1991                 }
1992                 fclose(f);
1993         }
1994
1995         if (proto > 255)
1996                 return false;
1997
1998         *id = proto;
1999         return true;
2000 }
2001
2002 bool system_resolve_rt_table(const char *name, unsigned int *id)
2003 {
2004         FILE *f;
2005         char *e, buf[128];
2006         unsigned int n, table = RT_TABLE_UNSPEC;
2007
2008         /* first try to parse table as number */
2009         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2010                 table = n;
2011
2012         /* handle well known aliases */
2013         else if (!strcmp(name, "default"))
2014                 table = RT_TABLE_DEFAULT;
2015         else if (!strcmp(name, "main"))
2016                 table = RT_TABLE_MAIN;
2017         else if (!strcmp(name, "local"))
2018                 table = RT_TABLE_LOCAL;
2019
2020         /* try to look up name in /etc/iproute2/rt_tables */
2021         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2022         {
2023                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2024                 {
2025                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2026                                 continue;
2027
2028                         n = strtoul(e, NULL, 10);
2029                         e = strtok(NULL, " \t\n");
2030
2031                         if (e && !strcmp(e, name))
2032                         {
2033                                 table = n;
2034                                 break;
2035                         }
2036                 }
2037
2038                 fclose(f);
2039         }
2040
2041         if (table == RT_TABLE_UNSPEC)
2042                 return false;
2043
2044         *id = table;
2045         return true;
2046 }
2047
2048 bool system_is_default_rt_table(unsigned int id)
2049 {
2050         return (id == RT_TABLE_MAIN);
2051 }
2052
2053 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2054 {
2055         char *e;
2056         unsigned int n;
2057
2058         if (!strcmp(filter, "strict"))
2059                 n = 1;
2060         else if (!strcmp(filter, "loose"))
2061                 n = 2;
2062         else {
2063                 n = strtoul(filter, &e, 0);
2064                 if (*e || e == filter || n > 2)
2065                         return false;
2066         }
2067
2068         *id = n;
2069         return true;
2070 }
2071
2072 static int system_iprule(struct iprule *rule, int cmd)
2073 {
2074         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2075
2076         struct nl_msg *msg;
2077         struct rtmsg rtm = {
2078                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2079                 .rtm_protocol = RTPROT_STATIC,
2080                 .rtm_scope = RT_SCOPE_UNIVERSE,
2081                 .rtm_table = RT_TABLE_UNSPEC,
2082                 .rtm_type = RTN_UNSPEC,
2083                 .rtm_flags = 0,
2084         };
2085
2086         if (cmd == RTM_NEWRULE)
2087                 rtm.rtm_type = RTN_UNICAST;
2088
2089         if (rule->invert)
2090                 rtm.rtm_flags |= FIB_RULE_INVERT;
2091
2092         if (rule->flags & IPRULE_SRC)
2093                 rtm.rtm_src_len = rule->src_mask;
2094
2095         if (rule->flags & IPRULE_DEST)
2096                 rtm.rtm_dst_len = rule->dest_mask;
2097
2098         if (rule->flags & IPRULE_TOS)
2099                 rtm.rtm_tos = rule->tos;
2100
2101         if (rule->flags & IPRULE_LOOKUP) {
2102                 if (rule->lookup < 256)
2103                         rtm.rtm_table = rule->lookup;
2104         }
2105
2106         if (rule->flags & IPRULE_ACTION)
2107                 rtm.rtm_type = rule->action;
2108         else if (rule->flags & IPRULE_GOTO)
2109                 rtm.rtm_type = FR_ACT_GOTO;
2110         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2111                 rtm.rtm_type = FR_ACT_NOP;
2112
2113         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2114
2115         if (!msg)
2116                 return -1;
2117
2118         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2119
2120         if (rule->flags & IPRULE_IN)
2121                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2122
2123         if (rule->flags & IPRULE_OUT)
2124                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2125
2126         if (rule->flags & IPRULE_SRC)
2127                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2128
2129         if (rule->flags & IPRULE_DEST)
2130                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2131
2132         if (rule->flags & IPRULE_PRIORITY)
2133                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2134         else if (cmd == RTM_NEWRULE)
2135                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2136
2137         if (rule->flags & IPRULE_FWMARK)
2138                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2139
2140         if (rule->flags & IPRULE_FWMASK)
2141                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2142
2143         if (rule->flags & IPRULE_LOOKUP) {
2144                 if (rule->lookup >= 256)
2145                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2146         }
2147
2148         if (rule->flags & IPRULE_GOTO)
2149                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2150
2151         return system_rtnl_call(msg);
2152 }
2153
2154 int system_add_iprule(struct iprule *rule)
2155 {
2156         return system_iprule(rule, RTM_NEWRULE);
2157 }
2158
2159 int system_del_iprule(struct iprule *rule)
2160 {
2161         return system_iprule(rule, RTM_DELRULE);
2162 }
2163
2164 int system_flush_iprules(void)
2165 {
2166         int rv = 0;
2167         struct iprule rule;
2168
2169         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2170         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2171
2172         memset(&rule, 0, sizeof(rule));
2173
2174
2175         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2176
2177         rule.priority = 0;
2178         rule.lookup = RT_TABLE_LOCAL;
2179         rv |= system_iprule(&rule, RTM_NEWRULE);
2180
2181         rule.priority = 32766;
2182         rule.lookup = RT_TABLE_MAIN;
2183         rv |= system_iprule(&rule, RTM_NEWRULE);
2184
2185         rule.priority = 32767;
2186         rule.lookup = RT_TABLE_DEFAULT;
2187         rv |= system_iprule(&rule, RTM_NEWRULE);
2188
2189
2190         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2191
2192         rule.priority = 0;
2193         rule.lookup = RT_TABLE_LOCAL;
2194         rv |= system_iprule(&rule, RTM_NEWRULE);
2195
2196         rule.priority = 32766;
2197         rule.lookup = RT_TABLE_MAIN;
2198         rv |= system_iprule(&rule, RTM_NEWRULE);
2199
2200         return rv;
2201 }
2202
2203 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2204 {
2205         return system_rtn_aton(action, id);
2206 }
2207
2208 time_t system_get_rtime(void)
2209 {
2210         struct timespec ts;
2211         struct timeval tv;
2212
2213         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
2214                 return ts.tv_sec;
2215
2216         if (gettimeofday(&tv, NULL) == 0)
2217                 return tv.tv_sec;
2218
2219         return 0;
2220 }
2221
2222 #ifndef IP_DF
2223 #define IP_DF       0x4000
2224 #endif
2225
2226 static int tunnel_ioctl(const char *name, int cmd, void *p)
2227 {
2228         struct ifreq ifr;
2229
2230         memset(&ifr, 0, sizeof(ifr));
2231         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name));
2232         ifr.ifr_ifru.ifru_data = p;
2233         return ioctl(sock_ioctl, cmd, &ifr);
2234 }
2235
2236 #ifdef IFLA_IPTUN_MAX
2237 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2238 static int system_add_gre_tunnel(const char *name, const char *kind,
2239                                  const unsigned int link, struct blob_attr **tb, bool v6)
2240 {
2241         struct nl_msg *nlm;
2242         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2243         struct blob_attr *cur;
2244         uint32_t ikey = 0, okey = 0, flags = 0, flowinfo = 0;
2245         uint16_t iflags = 0, oflags = 0;
2246         uint8_t tos = 0;
2247         int ret = 0, ttl = 0;
2248
2249         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2250         if (!nlm)
2251                 return -1;
2252
2253         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2254         nla_put_string(nlm, IFLA_IFNAME, name);
2255
2256         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2257         if (!linkinfo) {
2258                 ret = -ENOMEM;
2259                 goto failure;
2260         }
2261
2262         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2263         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2264         if (!infodata) {
2265                 ret = -ENOMEM;
2266                 goto failure;
2267         }
2268
2269         if (link)
2270                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2271
2272         if ((cur = tb[TUNNEL_ATTR_TTL]))
2273                 ttl = blobmsg_get_u32(cur);
2274
2275         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2276                 char *str = blobmsg_get_string(cur);
2277                 if (strcmp(str, "inherit")) {
2278                         unsigned uval;
2279
2280                         if (!system_tos_aton(str, &uval)) {
2281                                 ret = -EINVAL;
2282                                 goto failure;
2283                         }
2284
2285                         if (v6)
2286                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2287                         else
2288                                 tos = uval;
2289                 } else {
2290                         if (v6)
2291                                 flags |= IP6_TNL_F_USE_ORIG_TCLASS;
2292                         else
2293                                 tos = 1;
2294                 }
2295         }
2296
2297         if ((cur = tb[TUNNEL_ATTR_INFO]) && (blobmsg_type(cur) == BLOBMSG_TYPE_STRING)) {
2298                 uint8_t icsum, ocsum, iseqno, oseqno;
2299                 if (sscanf(blobmsg_get_string(cur), "%u,%u,%hhu,%hhu,%hhu,%hhu",
2300                         &ikey, &okey, &icsum, &ocsum, &iseqno, &oseqno) < 6) {
2301                         ret = -EINVAL;
2302                         goto failure;
2303                 }
2304
2305                 if (ikey)
2306                         iflags |= GRE_KEY;
2307
2308                 if (okey)
2309                         oflags |= GRE_KEY;
2310
2311                 if (icsum)
2312                         iflags |= GRE_CSUM;
2313
2314                 if (ocsum)
2315                         oflags |= GRE_CSUM;
2316
2317                 if (iseqno)
2318                         iflags |= GRE_SEQ;
2319
2320                 if (oseqno)
2321                         oflags |= GRE_SEQ;
2322         }
2323
2324         if (v6) {
2325                 struct in6_addr in6buf;
2326                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2327                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2328                                 ret = -EINVAL;
2329                                 goto failure;
2330                         }
2331                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2332                 }
2333
2334                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2335                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2336                                 ret = -EINVAL;
2337                                 goto failure;
2338                         }
2339                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2340                 }
2341                 nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, 4);
2342
2343                 if (flowinfo)
2344                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2345
2346                 if (flags)
2347                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags);
2348
2349                 if (!ttl)
2350                         ttl = 64;
2351         } else {
2352                 struct in_addr inbuf;
2353                 bool set_df = true;
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_GRE_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_GRE_REMOTE, sizeof(inbuf), &inbuf);
2369
2370                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2371                                 if (!okey) {
2372                                         okey = inbuf.s_addr;
2373                                         oflags |= GRE_KEY;
2374                                 }
2375
2376                                 if (!ikey) {
2377                                         ikey = inbuf.s_addr;
2378                                         iflags |= GRE_KEY;
2379                                 }
2380                         }
2381                 }
2382
2383                 if ((cur = tb[TUNNEL_ATTR_DF]))
2384                         set_df = blobmsg_get_bool(cur);
2385
2386                 if (!set_df) {
2387                         /* ttl != 0 and nopmtudisc are incompatible */
2388                         if (ttl) {
2389                                 ret = -EINVAL;
2390                                 goto failure;
2391                         }
2392                 } else if (!ttl)
2393                         ttl = 64;
2394
2395                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2396
2397                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2398         }
2399
2400         if (ttl)
2401                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2402
2403         if (oflags)
2404                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2405
2406         if (iflags)
2407                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2408
2409         if (okey)
2410                 nla_put_u32(nlm, IFLA_GRE_OKEY, okey);
2411
2412         if (ikey)
2413                 nla_put_u32(nlm, IFLA_GRE_IKEY, ikey);
2414
2415         nla_nest_end(nlm, infodata);
2416         nla_nest_end(nlm, linkinfo);
2417
2418         return system_rtnl_call(nlm);
2419
2420 failure:
2421         nlmsg_free(nlm);
2422         return ret;
2423 }
2424 #endif
2425
2426 #ifdef IFLA_VTI_MAX
2427 static int system_add_vti_tunnel(const char *name, const char *kind,
2428                                  const unsigned int link, struct blob_attr **tb, bool v6)
2429 {
2430         struct nl_msg *nlm;
2431         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2432         struct blob_attr *cur;
2433         uint32_t ikey = 0, okey = 0;
2434         int ret = 0;
2435
2436         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2437         if (!nlm)
2438                 return -1;
2439
2440         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2441         nla_put_string(nlm, IFLA_IFNAME, name);
2442
2443         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2444         if (!linkinfo) {
2445                 ret = -ENOMEM;
2446                 goto failure;
2447         }
2448
2449         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2450         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2451         if (!infodata) {
2452                 ret = -ENOMEM;
2453                 goto failure;
2454         }
2455
2456         if (link)
2457                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2458
2459         if ((cur = tb[TUNNEL_ATTR_INFO]) && (blobmsg_type(cur) == BLOBMSG_TYPE_STRING)) {
2460                 if (sscanf(blobmsg_get_string(cur), "%u,%u",
2461                         &ikey, &okey) < 2) {
2462                         ret = -EINVAL;
2463                         goto failure;
2464                 }
2465         }
2466
2467         if (v6) {
2468                 struct in6_addr in6buf;
2469                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2470                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2471                                 ret = -EINVAL;
2472                                 goto failure;
2473                         }
2474                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2475                 }
2476
2477                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2478                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2479                                 ret = -EINVAL;
2480                                 goto failure;
2481                         }
2482                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2483                 }
2484
2485         } else {
2486                 struct in_addr inbuf;
2487
2488                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2489                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2490                                 ret = -EINVAL;
2491                                 goto failure;
2492                         }
2493                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2494                 }
2495
2496                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2497                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2498                                 ret = -EINVAL;
2499                                 goto failure;
2500                         }
2501                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2502                 }
2503
2504         }
2505
2506         if (okey)
2507                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2508
2509         if (ikey)
2510                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2511
2512         nla_nest_end(nlm, infodata);
2513         nla_nest_end(nlm, linkinfo);
2514
2515         return system_rtnl_call(nlm);
2516
2517 failure:
2518         nlmsg_free(nlm);
2519         return ret;
2520 }
2521 #endif
2522
2523 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
2524 {
2525         struct blob_attr *cur;
2526         bool set_df = true;
2527         struct ip_tunnel_parm p  = {
2528                 .link = link,
2529                 .iph = {
2530                         .version = 4,
2531                         .ihl = 5,
2532                         .protocol = proto,
2533                 }
2534         };
2535
2536         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
2537                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
2538                 return -EINVAL;
2539
2540         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
2541                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
2542                 return -EINVAL;
2543
2544         if ((cur = tb[TUNNEL_ATTR_DF]))
2545                 set_df = blobmsg_get_bool(cur);
2546
2547         if ((cur = tb[TUNNEL_ATTR_TTL]))
2548                 p.iph.ttl = blobmsg_get_u32(cur);
2549
2550         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2551                 char *str = blobmsg_get_string(cur);
2552                 if (strcmp(str, "inherit")) {
2553                         unsigned uval;
2554
2555                         if (!system_tos_aton(str, &uval))
2556                                 return -EINVAL;
2557
2558                         p.iph.tos = uval;
2559                 } else
2560                         p.iph.tos = 1;
2561         }
2562
2563         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
2564         /* ttl !=0 and nopmtudisc are incompatible */
2565         if (p.iph.ttl && p.iph.frag_off == 0)
2566                 return -EINVAL;
2567
2568         strncpy(p.name, name, sizeof(p.name));
2569
2570         switch (p.iph.protocol) {
2571         case IPPROTO_IPIP:
2572                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
2573         case IPPROTO_IPV6:
2574                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
2575         default:
2576                 break;
2577         }
2578         return -1;
2579 }
2580
2581 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
2582 {
2583         struct blob_attr *cur;
2584         const char *str;
2585
2586         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2587                 return -EINVAL;
2588         str = blobmsg_data(cur);
2589
2590         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
2591             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
2592             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6"))
2593                 return system_link_del(name);
2594         else
2595                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
2596 }
2597
2598 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
2599 {
2600         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2601
2602         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2603                 blob_data(attr), blob_len(attr));
2604
2605         return __system_del_ip_tunnel(name, tb);
2606 }
2607
2608 int system_update_ipv6_mtu(struct device *dev, int mtu)
2609 {
2610         int ret = -1;
2611         char buf[64];
2612         int fd;
2613
2614         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
2615                         dev->ifname);
2616
2617         fd = open(buf, O_RDWR);
2618         if (fd < 0)
2619                 return ret;
2620
2621         if (!mtu) {
2622                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
2623                 if (len < 0)
2624                         goto out;
2625
2626                 buf[len] = 0;
2627                 ret = atoi(buf);
2628         } else {
2629                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
2630                         ret = mtu;
2631         }
2632
2633 out:
2634         close(fd);
2635         return ret;
2636 }
2637
2638 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
2639 {
2640         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2641         struct blob_attr *cur;
2642         const char *str;
2643
2644         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2645                 blob_data(attr), blob_len(attr));
2646
2647         __system_del_ip_tunnel(name, tb);
2648
2649         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2650                 return -EINVAL;
2651         str = blobmsg_data(cur);
2652
2653         unsigned int ttl = 0;
2654         if ((cur = tb[TUNNEL_ATTR_TTL])) {
2655                 ttl = blobmsg_get_u32(cur);
2656                 if (ttl > 255)
2657                         return -EINVAL;
2658         }
2659
2660         unsigned int link = 0;
2661         if ((cur = tb[TUNNEL_ATTR_LINK])) {
2662                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
2663                 if (!iface)
2664                         return -EINVAL;
2665
2666                 if (iface->l3_dev.dev)
2667                         link = iface->l3_dev.dev->ifindex;
2668         }
2669
2670         if (!strcmp(str, "sit")) {
2671                 if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
2672                         return -1;
2673
2674 #ifdef SIOCADD6RD
2675                 if ((cur = tb[TUNNEL_ATTR_6RD_PREFIX])) {
2676                         unsigned int mask;
2677                         struct ip_tunnel_6rd p6;
2678
2679                         memset(&p6, 0, sizeof(p6));
2680
2681                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
2682                                                 &p6.prefix, &mask) || mask > 128)
2683                                 return -EINVAL;
2684                         p6.prefixlen = mask;
2685
2686                         if ((cur = tb[TUNNEL_ATTR_6RD_RELAY_PREFIX])) {
2687                                 if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
2688                                                         &p6.relay_prefix, &mask) || mask > 32)
2689                                         return -EINVAL;
2690                                 p6.relay_prefixlen = mask;
2691                         }
2692
2693                         if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
2694                                 __system_del_ip_tunnel(name, tb);
2695                                 return -1;
2696                         }
2697                 }
2698 #endif
2699 #ifdef IFLA_IPTUN_MAX
2700         } else if (!strcmp(str, "ipip6")) {
2701                 struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2702                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2703                 struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2704                 int ret = 0;
2705
2706                 if (!nlm)
2707                         return -1;
2708
2709                 nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2710                 nla_put_string(nlm, IFLA_IFNAME, name);
2711
2712                 if (link)
2713                         nla_put_u32(nlm, IFLA_LINK, link);
2714
2715                 struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2716                 if (!linkinfo) {
2717                         ret = -ENOMEM;
2718                         goto failure;
2719                 }
2720                 nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2721                 struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2722                 if (!infodata) {
2723                         ret = -ENOMEM;
2724                         goto failure;
2725                 }
2726
2727                 if (link)
2728                         nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2729
2730                 nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2731                 nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2732                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, 4);
2733
2734                 struct in6_addr in6buf;
2735                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2736                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2737                                 ret = -EINVAL;
2738                                 goto failure;
2739                         }
2740                         nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2741                 }
2742
2743                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2744                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2745                                 ret = -EINVAL;
2746                                 goto failure;
2747                         }
2748                         nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2749                 }
2750
2751 #ifdef IFLA_IPTUN_FMR_MAX
2752                 if ((cur = tb[TUNNEL_ATTR_FMRS])) {
2753                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2754
2755                         struct blob_attr *fmr;
2756                         unsigned rem, fmrcnt = 0;
2757                         blobmsg_for_each_attr(fmr, cur, rem) {
2758                                 if (blobmsg_type(fmr) != BLOBMSG_TYPE_STRING)
2759                                         continue;
2760
2761                                 unsigned ip4len, ip6len, ealen, offset = 6;
2762                                 char ip6buf[48];
2763                                 char ip4buf[16];
2764
2765                                 if (sscanf(blobmsg_get_string(fmr), "%47[^/]/%u,%15[^/]/%u,%u,%u",
2766                                                 ip6buf, &ip6len, ip4buf, &ip4len, &ealen, &offset) < 5) {
2767                                         ret = -EINVAL;
2768                                         goto failure;
2769                                 }
2770
2771                                 struct in6_addr ip6prefix;
2772                                 struct in_addr ip4prefix;
2773                                 if (inet_pton(AF_INET6, ip6buf, &ip6prefix) != 1 ||
2774                                                 inet_pton(AF_INET, ip4buf, &ip4prefix) != 1) {
2775                                         ret = -EINVAL;
2776                                         goto failure;
2777                                 }
2778
2779                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2780
2781                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2782                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2783                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2784                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2785                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2786                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2787
2788                                 nla_nest_end(nlm, rule);
2789                         }
2790
2791                         nla_nest_end(nlm, fmrs);
2792                 }
2793 #endif
2794
2795                 nla_nest_end(nlm, infodata);
2796                 nla_nest_end(nlm, linkinfo);
2797
2798                 return system_rtnl_call(nlm);
2799 failure:
2800                 nlmsg_free(nlm);
2801                 return ret;
2802         } else if (!strcmp(str, "greip")) {
2803                 return system_add_gre_tunnel(name, "gre", link, tb, false);
2804         } else if (!strcmp(str, "gretapip"))  {
2805                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
2806         } else if (!strcmp(str, "greip6")) {
2807                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
2808         } else if (!strcmp(str, "gretapip6")) {
2809                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
2810 #ifdef IFLA_VTI_MAX
2811         } else if (!strcmp(str, "vtiip")) {
2812                 return system_add_vti_tunnel(name, "vti", link, tb, false);
2813         } else if (!strcmp(str, "vtiip6")) {
2814                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
2815 #endif
2816 #endif
2817         } else if (!strcmp(str, "ipip")) {
2818                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
2819         }
2820         else
2821                 return -EINVAL;
2822
2823         return 0;
2824 }