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_neigh4locktime(struct device *dev, const char *val)
 320 {
 321         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime", dev->ifname, val);
 322 }
 323
 324 static void system_set_dadtransmits(struct device *dev, const char *val)
 325 {
 326         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 327 }
 328
 329 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 330 {
 331         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 332 }
 333
 334 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 335 {
 336         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 337 }
 338
 339 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 340 {
 341         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 342 }
 343
 344 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 345 {
 346         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 347                                        "/sys/class/net/%s/brport/multicast_router",
 348                               dev->ifname, val);
 349 }
 350
 351 static void system_bridge_set_robustness(struct device *dev, const char *val)
 352 {
 353         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 354                               dev->ifname, val);
 355         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 356                               dev->ifname, val);
 357 }
 358
 359 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 360 {
 361         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 362                               dev->ifname, val);
 363 }
 364
 365 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 366 {
 367         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 368                               dev->ifname, val);
 369 }
 370
 371 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 372 {
 373         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 374                               dev->ifname, val);
 375 }
 376
 377 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 378 {
 379         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 380                               dev->ifname, val);
 381 }
 382
 383 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 384 {
 385         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 386                               dev->ifname, val);
 387 }
 388
 389 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 390 {
 391         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 392                               dev->ifname, val);
 393 }
 394
 395 static void system_bridge_set_learning(struct device *dev, const char *val)
 396 {
 397         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 398 }
 399
 400 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 401 {
 402         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 403 }
 404
 405 static void system_set_sendredirects(struct device *dev, const char *val)
 406 {
 407         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 408 }
 409
 410 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 411 {
 412         int fd = -1, ret = -1;
 413
 414         fd = open(path, O_RDONLY);
 415         if (fd < 0)
 416                 goto out;
 417
 418         ssize_t len = read(fd, buf, buf_sz - 1);
 419         if (len < 0)
 420                 goto out;
 421
 422         ret = buf[len] = 0;
 423
 424 out:
 425         if (fd >= 0)
 426                 close(fd);
 427
 428         return ret;
 429 }
 430
 431 static int
 432 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 433 {
 434         snprintf(dev_buf, sizeof(dev_buf), path, device);
 435         return system_get_sysctl(dev_buf, buf, buf_sz);
 436 }
 437
 438 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 439 {
 440         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 441                         dev->ifname, buf, buf_sz);
 442 }
 443
 444 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 445 {
 446         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 447                         dev->ifname, buf, buf_sz);
 448 }
 449
 450 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 451 {
 452         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 453                         dev->ifname, buf, buf_sz);
 454 }
 455
 456 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 457 {
 458         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 459                         dev->ifname, buf, buf_sz);
 460 }
 461
 462 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 463 {
 464         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 465                         dev->ifname, buf, buf_sz);
 466 }
 467
 468 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 469 {
 470         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 471                         dev->ifname, buf, buf_sz);
 472 }
 473
 474 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 475 {
 476         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 477                         dev->ifname, buf, buf_sz);
 478 }
 479
 480 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 481 {
 482         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 483                         dev->ifname, buf, buf_sz);
 484 }
 485
 486 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 487 {
 488         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 489                         dev->ifname, buf, buf_sz);
 490 }
 491
 492 static int system_get_neigh4locktime(struct device *dev, char *buf, const size_t buf_sz)
 493 {
 494         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime",
 495                         dev->ifname, buf, buf_sz);
 496 }
 497
 498 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 499 {
 500         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 501                         dev->ifname, buf, buf_sz);
 502 }
 503
 504 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 505 {
 506         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 507                         dev->ifname, buf, buf_sz);
 508 }
 509
 510 // Evaluate netlink messages
 511 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 512 {
 513         struct nlmsghdr *nh = nlmsg_hdr(msg);
 514         struct nlattr *nla[__IFLA_MAX];
 515         int link_state = 0;
 516         char buf[10];
 517
 518         if (nh->nlmsg_type != RTM_NEWLINK)
 519                 goto out;
 520
 521         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 522         if (!nla[IFLA_IFNAME])
 523                 goto out;
 524
 525         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 526         if (!dev)
 527                 goto out;
 528
 529         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 530                 link_state = strtoul(buf, NULL, 0);
 531
 532         device_set_link(dev, link_state ? true : false);
 533
 534 out:
 535         return 0;
 536 }
 537
 538 static void
 539 handle_hotplug_msg(char *data, int size)
 540 {
 541         const char *subsystem = NULL, *interface = NULL;
 542         char *cur, *end, *sep;
 543         struct device *dev;
 544         int skip;
 545         bool add;
 546
 547         if (!strncmp(data, "add@", 4))
 548                 add = true;
 549         else if (!strncmp(data, "remove@", 7))
 550                 add = false;
 551         else
 552                 return;
 553
 554         skip = strlen(data) + 1;
 555         end = data + size;
 556
 557         for (cur = data + skip; cur < end; cur += skip) {
 558                 skip = strlen(cur) + 1;
 559
 560                 sep = strchr(cur, '=');
 561                 if (!sep)
 562                         continue;
 563
 564                 *sep = 0;
 565                 if (!strcmp(cur, "INTERFACE"))
 566                         interface = sep + 1;
 567                 else if (!strcmp(cur, "SUBSYSTEM")) {
 568                         subsystem = sep + 1;
 569                         if (strcmp(subsystem, "net") != 0)
 570                                 return;
 571                 }
 572                 if (subsystem && interface)
 573                         goto found;
 574         }
 575         return;
 576
 577 found:
 578         dev = device_find(interface);
 579         if (!dev)
 580                 return;
 581
 582         if (dev->type != &simple_device_type)
 583                 return;
 584
 585         if (add && system_if_force_external(dev->ifname))
 586                 return;
 587
 588         device_set_present(dev, add);
 589 }
 590
 591 static void
 592 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 593 {
 594         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 595         struct sockaddr_nl nla;
 596         unsigned char *buf = NULL;
 597         int size;
 598
 599         while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 600                 if (nla.nl_pid == 0)
 601                         handle_hotplug_msg((char *) buf, size);
 602
 603                 free(buf);
 604         }
 605 }
 606
 607 static int system_rtnl_call(struct nl_msg *msg)
 608 {
 609         int ret;
 610
 611         ret = nl_send_auto_complete(sock_rtnl, msg);
 612         nlmsg_free(msg);
 613
 614         if (ret < 0)
 615                 return ret;
 616
 617         return nl_wait_for_ack(sock_rtnl);
 618 }
 619
 620 int system_bridge_delbr(struct device *bridge)
 621 {
 622         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 623 }
 624
 625 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 626 {
 627         struct ifreq ifr;
 628
 629         memset(&ifr, 0, sizeof(ifr));
 630         if (dev)
 631                 ifr.ifr_ifindex = dev->ifindex;
 632         else
 633                 ifr.ifr_data = data;
 634         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name));
 635         return ioctl(sock_ioctl, cmd, &ifr);
 636 }
 637
 638 static bool system_is_bridge(const char *name, char *buf, int buflen)
 639 {
 640         struct stat st;
 641
 642         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 643         if (stat(buf, &st) < 0)
 644                 return false;
 645
 646         return true;
 647 }
 648
 649 static char *system_get_bridge(const char *name, char *buf, int buflen)
 650 {
 651         char *path;
 652         ssize_t len = -1;
 653         glob_t gl;
 654
 655         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 656         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 657                 return NULL;
 658
 659         if (gl.gl_pathc > 0)
 660                 len = readlink(gl.gl_pathv[0], buf, buflen);
 661
 662         globfree(&gl);
 663
 664         if (len < 0)
 665                 return NULL;
 666
 667         buf[len] = 0;
 668         path = strrchr(buf, '/');
 669         if (!path)
 670                 return NULL;
 671
 672         return path + 1;
 673 }
 674
 675 static void
 676 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 677 {
 678         bool mcast_to_ucast = dev->wireless_ap;
 679         bool hairpin = true;
 680
 681         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 682             !bridge->settings.multicast_to_unicast)
 683                 mcast_to_ucast = false;
 684
 685         if (!mcast_to_ucast || dev->wireless_isolate)
 686                 hairpin = false;
 687
 688         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 689         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 690 }
 691
 692 int system_bridge_addif(struct device *bridge, struct device *dev)
 693 {
 694         char buf[64];
 695         char *oldbr;
 696         int ret = 0;
 697
 698         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 699         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 700                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 701
 702         if (dev->wireless)
 703                 system_bridge_set_wireless(bridge, dev);
 704
 705         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 706                 snprintf(buf, sizeof(buf), "%i", dev->settings.multicast_router);
 707                 system_bridge_set_multicast_router(dev, buf, false);
 708         }
 709
 710         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 711             dev->settings.multicast_fast_leave)
 712                 system_bridge_set_multicast_fast_leave(dev, "1");
 713
 714         if (dev->settings.flags & DEV_OPT_LEARNING &&
 715             !dev->settings.learning)
 716                 system_bridge_set_learning(dev, "0");
 717
 718         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 719             !dev->settings.unicast_flood)
 720                 system_bridge_set_unicast_flood(dev, "0");
 721
 722         return ret;
 723 }
 724
 725 int system_bridge_delif(struct device *bridge, struct device *dev)
 726 {
 727         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 728 }
 729
 730 int system_if_resolve(struct device *dev)
 731 {
 732         struct ifreq ifr;
 733         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
 734         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 735                 return ifr.ifr_ifindex;
 736         else
 737                 return 0;
 738 }
 739
 740 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 741 {
 742         struct ifreq ifr;
 743
 744         memset(&ifr, 0, sizeof(ifr));
 745         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name));
 746         ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr);
 747         ifr.ifr_flags |= add;
 748         ifr.ifr_flags &= ~rem;
 749         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 750 }
 751
 752 struct clear_data {
 753         struct nl_msg *msg;
 754         struct device *dev;
 755         int type;
 756         int size;
 757         int af;
 758 };
 759
 760
 761 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 762 {
 763         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 764
 765         return ifa->ifa_index == ifindex;
 766 }
 767
 768 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 769 {
 770         struct rtmsg *r = NLMSG_DATA(hdr);
 771         struct nlattr *tb[__RTA_MAX];
 772
 773         if (r->rtm_protocol == RTPROT_KERNEL &&
 774             r->rtm_family == AF_INET6)
 775                 return false;
 776
 777         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 778         if (!tb[RTA_OIF])
 779                 return false;
 780
 781         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 782 }
 783
 784 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 785 {
 786         return true;
 787 }
 788
 789 static int cb_clear_event(struct nl_msg *msg, void *arg)
 790 {
 791         struct clear_data *clr = arg;
 792         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 793         bool (*cb)(struct nlmsghdr *, int ifindex);
 794         int type;
 795
 796         switch(clr->type) {
 797         case RTM_GETADDR:
 798                 type = RTM_DELADDR;
 799                 if (hdr->nlmsg_type != RTM_NEWADDR)
 800                         return NL_SKIP;
 801
 802                 cb = check_ifaddr;
 803                 break;
 804         case RTM_GETROUTE:
 805                 type = RTM_DELROUTE;
 806                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 807                         return NL_SKIP;
 808
 809                 cb = check_route;
 810                 break;
 811         case RTM_GETRULE:
 812                 type = RTM_DELRULE;
 813                 if (hdr->nlmsg_type != RTM_NEWRULE)
 814                         return NL_SKIP;
 815
 816                 cb = check_rule;
 817                 break;
 818         default:
 819                 return NL_SKIP;
 820         }
 821
 822         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 823                 return NL_SKIP;
 824
 825         if (type == RTM_DELRULE)
 826                 D(SYSTEM, "Remove a rule\n");
 827         else
 828                 D(SYSTEM, "Remove %s from device %s\n",
 829                   type == RTM_DELADDR ? "an address" : "a route",
 830                   clr->dev->ifname);
 831         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 832         hdr = nlmsg_hdr(clr->msg);
 833         hdr->nlmsg_type = type;
 834         hdr->nlmsg_flags = NLM_F_REQUEST;
 835
 836         nl_socket_disable_auto_ack(sock_rtnl);
 837         nl_send_auto_complete(sock_rtnl, clr->msg);
 838         nl_socket_enable_auto_ack(sock_rtnl);
 839
 840         return NL_SKIP;
 841 }
 842
 843 static int
 844 cb_finish_event(struct nl_msg *msg, void *arg)
 845 {
 846         int *pending = arg;
 847         *pending = 0;
 848         return NL_STOP;
 849 }
 850
 851 static int
 852 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 853 {
 854         int *pending = arg;
 855         *pending = err->error;
 856         return NL_STOP;
 857 }
 858
 859 static void
 860 system_if_clear_entries(struct device *dev, int type, int af)
 861 {
 862         struct clear_data clr;
 863         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
 864         struct rtmsg rtm = {
 865                 .rtm_family = af,
 866                 .rtm_flags = RTM_F_CLONED,
 867         };
 868         int flags = NLM_F_DUMP;
 869         int pending = 1;
 870
 871         clr.af = af;
 872         clr.dev = dev;
 873         clr.type = type;
 874         switch (type) {
 875         case RTM_GETADDR:
 876         case RTM_GETRULE:
 877                 clr.size = sizeof(struct rtgenmsg);
 878                 break;
 879         case RTM_GETROUTE:
 880                 clr.size = sizeof(struct rtmsg);
 881                 break;
 882         default:
 883                 return;
 884         }
 885
 886         if (!cb)
 887                 return;
 888
 889         clr.msg = nlmsg_alloc_simple(type, flags);
 890         if (!clr.msg)
 891                 goto out;
 892
 893         nlmsg_append(clr.msg, &rtm, clr.size, 0);
 894         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
 895         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
 896         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
 897
 898         nl_send_auto_complete(sock_rtnl, clr.msg);
 899         while (pending > 0)
 900                 nl_recvmsgs(sock_rtnl, cb);
 901
 902         nlmsg_free(clr.msg);
 903 out:
 904         nl_cb_put(cb);
 905 }
 906
 907 /*
 908  * Clear bridge (membership) state and bring down device
 909  */
 910 void system_if_clear_state(struct device *dev)
 911 {
 912         static char buf[256];
 913         char *bridge;
 914
 915         device_set_ifindex(dev, system_if_resolve(dev));
 916         if (dev->external || !dev->ifindex)
 917                 return;
 918
 919         system_if_flags(dev->ifname, 0, IFF_UP);
 920
 921         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
 922                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
 923                 system_bridge_delbr(dev);
 924                 return;
 925         }
 926
 927         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
 928         if (bridge) {
 929                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
 930                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
 931         }
 932
 933         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
 934         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
 935         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
 936         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
 937         system_set_disable_ipv6(dev, "0");
 938 }
 939
 940 static inline unsigned long
 941 sec_to_jiffies(int val)
 942 {
 943         return (unsigned long) val * 100;
 944 }
 945
 946 static void system_bridge_conf_multicast_deps(struct device *bridge,
 947                                               struct bridge_config *cfg,
 948                                               char *buf,
 949                                               int buf_len)
 950 {
 951         int val;
 952
 953         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
 954             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
 955             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
 956                 val = cfg->robustness * cfg->query_interval +
 957                         cfg->query_response_interval;
 958
 959                 snprintf(buf, buf_len, "%i", val);
 960                 system_bridge_set_membership_interval(bridge, buf);
 961
 962                 val = cfg->robustness * cfg->query_interval +
 963                         cfg->query_response_interval / 2;
 964
 965                 snprintf(buf, buf_len, "%i", val);
 966                 system_bridge_set_other_querier_timeout(bridge, buf);
 967         }
 968
 969         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
 970                 val = cfg->query_interval / 4;
 971
 972                 snprintf(buf, buf_len, "%i", val);
 973                 system_bridge_set_startup_query_interval(bridge, buf);
 974         }
 975 }
 976
 977 static void system_bridge_conf_multicast(struct device *bridge,
 978                                          struct bridge_config *cfg,
 979                                          char *buf,
 980                                          int buf_len)
 981 {
 982         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
 983                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
 984
 985         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
 986                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
 987
 988         snprintf(buf, buf_len, "%i", cfg->hash_max);
 989         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
 990                 bridge->ifname, buf);
 991
 992         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 993                 snprintf(buf, buf_len, "%i", bridge->settings.multicast_router);
 994                 system_bridge_set_multicast_router(bridge, buf, true);
 995         }
 996
 997         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
 998                 snprintf(buf, buf_len, "%i", cfg->robustness);
 999                 system_bridge_set_robustness(bridge, buf);
1000         }
1001
1002         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1003                 snprintf(buf, buf_len, "%i", cfg->query_interval);
1004                 system_bridge_set_query_interval(bridge, buf);
1005         }
1006
1007         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1008                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
1009                 system_bridge_set_query_response_interval(bridge, buf);
1010         }
1011
1012         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1013                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1014                 system_bridge_set_last_member_interval(bridge, buf);
1015         }
1016
1017         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1018 }
1019
1020 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1021 {
1022         char buf[64];
1023         unsigned long args[4] = {};
1024
1025         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1026                 return -1;
1027
1028         args[0] = BRCTL_SET_BRIDGE_STP_STATE;
1029         args[1] = !!cfg->stp;
1030         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1031
1032         args[0] = BRCTL_SET_BRIDGE_FORWARD_DELAY;
1033         args[1] = sec_to_jiffies(cfg->forward_delay);
1034         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1035
1036         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1037
1038         args[0] = BRCTL_SET_BRIDGE_PRIORITY;
1039         args[1] = cfg->priority;
1040         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1041
1042         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1043                 args[0] = BRCTL_SET_AGEING_TIME;
1044                 args[1] = sec_to_jiffies(cfg->ageing_time);
1045                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1046         }
1047
1048         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1049                 args[0] = BRCTL_SET_BRIDGE_HELLO_TIME;
1050                 args[1] = sec_to_jiffies(cfg->hello_time);
1051                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1052         }
1053
1054         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1055                 args[0] = BRCTL_SET_BRIDGE_MAX_AGE;
1056                 args[1] = sec_to_jiffies(cfg->max_age);
1057                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1058         }
1059
1060         return 0;
1061 }
1062
1063 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1064 {
1065         struct nl_msg *msg;
1066         struct nlattr *linkinfo, *data;
1067         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1068         int i, rv;
1069         static const struct {
1070                 const char *name;
1071                 enum macvlan_mode val;
1072         } modes[] = {
1073                 { "private", MACVLAN_MODE_PRIVATE },
1074                 { "vepa", MACVLAN_MODE_VEPA },
1075                 { "bridge", MACVLAN_MODE_BRIDGE },
1076                 { "passthru", MACVLAN_MODE_PASSTHRU },
1077         };
1078
1079         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1080
1081         if (!msg)
1082                 return -1;
1083
1084         nlmsg_append(msg, &iim, sizeof(iim), 0);
1085
1086         if (cfg->flags & MACVLAN_OPT_MACADDR)
1087                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1088         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1089         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1090
1091         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1092                 goto nla_put_failure;
1093
1094         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1095
1096         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1097                 goto nla_put_failure;
1098
1099         if (cfg->mode) {
1100                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1101                         if (strcmp(cfg->mode, modes[i].name) != 0)
1102                                 continue;
1103
1104                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1105                         break;
1106                 }
1107         }
1108
1109         nla_nest_end(msg, data);
1110         nla_nest_end(msg, linkinfo);
1111
1112         rv = system_rtnl_call(msg);
1113         if (rv)
1114                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1115
1116         return rv;
1117
1118 nla_put_failure:
1119         nlmsg_free(msg);
1120         return -ENOMEM;
1121 }
1122
1123 static int system_link_del(const char *ifname)
1124 {
1125         struct nl_msg *msg;
1126         struct ifinfomsg iim = {
1127                 .ifi_family = AF_UNSPEC,
1128                 .ifi_index = 0,
1129         };
1130
1131         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1132
1133         if (!msg)
1134                 return -1;
1135
1136         nlmsg_append(msg, &iim, sizeof(iim), 0);
1137         nla_put_string(msg, IFLA_IFNAME, ifname);
1138         return system_rtnl_call(msg);
1139 }
1140
1141 int system_macvlan_del(struct device *macvlan)
1142 {
1143         return system_link_del(macvlan->ifname);
1144 }
1145
1146 int system_veth_add(struct device *veth, struct veth_config *cfg)
1147 {
1148         struct nl_msg *msg;
1149         struct ifinfomsg empty_iim = {};
1150         struct nlattr *linkinfo, *data, *veth_info;
1151         int rv;
1152
1153         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1154
1155         if (!msg)
1156                 return -1;
1157
1158         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1159
1160         if (cfg->flags & VETH_OPT_MACADDR)
1161                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1162         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1163
1164         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1165                 goto nla_put_failure;
1166
1167         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1168
1169         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1170                 goto nla_put_failure;
1171
1172         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1173                 goto nla_put_failure;
1174
1175         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1176
1177         if (cfg->flags & VETH_OPT_PEER_NAME)
1178                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1179         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1180                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1181
1182         nla_nest_end(msg, veth_info);
1183         nla_nest_end(msg, data);
1184         nla_nest_end(msg, linkinfo);
1185
1186         rv = system_rtnl_call(msg);
1187         if (rv) {
1188                 if (cfg->flags & VETH_OPT_PEER_NAME)
1189                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1190                 else
1191                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1192         }
1193
1194         return rv;
1195
1196 nla_put_failure:
1197         nlmsg_free(msg);
1198         return -ENOMEM;
1199 }
1200
1201 int system_veth_del(struct device *veth)
1202 {
1203         return system_link_del(veth->ifname);
1204 }
1205
1206 static int system_vlan(struct device *dev, int id)
1207 {
1208         struct vlan_ioctl_args ifr = {
1209                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1210                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1211         };
1212
1213         ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1214
1215         if (id < 0) {
1216                 ifr.cmd = DEL_VLAN_CMD;
1217                 ifr.u.VID = 0;
1218         } else {
1219                 ifr.cmd = ADD_VLAN_CMD;
1220                 ifr.u.VID = id;
1221         }
1222         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1223         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1224 }
1225
1226 int system_vlan_add(struct device *dev, int id)
1227 {
1228         return system_vlan(dev, id);
1229 }
1230
1231 int system_vlan_del(struct device *dev)
1232 {
1233         return system_vlan(dev, -1);
1234 }
1235
1236 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1237 {
1238         struct nl_msg *msg;
1239         struct nlattr *linkinfo, *data;
1240         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1241         int rv;
1242
1243         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1244
1245         if (!msg)
1246                 return -1;
1247
1248         nlmsg_append(msg, &iim, sizeof(iim), 0);
1249         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1250         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1251
1252         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1253                 goto nla_put_failure;
1254
1255         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1256
1257         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1258                 goto nla_put_failure;
1259
1260         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1261
1262 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1263         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1264 #else
1265         if(cfg->proto == VLAN_PROTO_8021AD)
1266                 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);
1267 #endif
1268
1269         nla_nest_end(msg, data);
1270         nla_nest_end(msg, linkinfo);
1271
1272         rv = system_rtnl_call(msg);
1273         if (rv)
1274                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1275
1276         return rv;
1277
1278 nla_put_failure:
1279         nlmsg_free(msg);
1280         return -ENOMEM;
1281 }
1282
1283 int system_vlandev_del(struct device *vlandev)
1284 {
1285         return system_link_del(vlandev->ifname);
1286 }
1287
1288 void
1289 system_if_get_settings(struct device *dev, struct device_settings *s)
1290 {
1291         struct ifreq ifr;
1292         char buf[10];
1293
1294         memset(&ifr, 0, sizeof(ifr));
1295         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1296
1297         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1298                 s->mtu = ifr.ifr_mtu;
1299                 s->flags |= DEV_OPT_MTU;
1300         }
1301
1302         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1303         if (s->mtu6 > 0)
1304                 s->flags |= DEV_OPT_MTU6;
1305
1306         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1307                 s->txqueuelen = ifr.ifr_qlen;
1308                 s->flags |= DEV_OPT_TXQUEUELEN;
1309         }
1310
1311         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1312                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1313                 s->flags |= DEV_OPT_MACADDR;
1314         }
1315
1316         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1317                 s->ipv6 = !strtoul(buf, NULL, 0);
1318                 s->flags |= DEV_OPT_IPV6;
1319         }
1320
1321         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1322                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1323                 s->flags |= DEV_OPT_PROMISC;
1324
1325                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1326                 s->flags |= DEV_OPT_MULTICAST;
1327         }
1328
1329         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1330                 s->rpfilter = strtoul(buf, NULL, 0);
1331                 s->flags |= DEV_OPT_RPFILTER;
1332         }
1333
1334         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1335                 s->acceptlocal = strtoul(buf, NULL, 0);
1336                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1337         }
1338
1339         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1340                 s->igmpversion = strtoul(buf, NULL, 0);
1341                 s->flags |= DEV_OPT_IGMPVERSION;
1342         }
1343
1344         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1345                 s->mldversion = strtoul(buf, NULL, 0);
1346                 s->flags |= DEV_OPT_MLDVERSION;
1347         }
1348
1349         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1350                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1351                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1352         }
1353
1354         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1355                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1356                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1357         }
1358
1359         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
1360                 s->neigh4locktime = strtoul(buf, NULL, 0);
1361                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
1362         }
1363
1364         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1365                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1366                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1367         }
1368
1369         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1370                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1371                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1372         }
1373
1374         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1375                 s->dadtransmits = strtoul(buf, NULL, 0);
1376                 s->flags |= DEV_OPT_DADTRANSMITS;
1377         }
1378
1379         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1380                 s->sendredirects = strtoul(buf, NULL, 0);
1381                 s->flags |= DEV_OPT_SENDREDIRECTS;
1382         }
1383 }
1384
1385 static void
1386 system_if_set_rps_xps_val(const char *path, int val)
1387 {
1388         char val_buf[8];
1389         glob_t gl;
1390         int i;
1391
1392         if (glob(path, 0, NULL, &gl))
1393                 return;
1394
1395         snprintf(val_buf, sizeof(val_buf), "%x", val);
1396         for (i = 0; i < gl.gl_pathc; i++)
1397                 system_set_sysctl(gl.gl_pathv[i], val_buf);
1398
1399         globfree(&gl);
1400 }
1401
1402 static void
1403 system_if_apply_rps_xps(struct device *dev, struct device_settings *s)
1404 {
1405         long n_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1406         int val;
1407
1408         if (n_cpus < 2)
1409                 return;
1410
1411         val = (1 << n_cpus) - 1;
1412         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/rps_cpus", dev->ifname);
1413         system_if_set_rps_xps_val(dev_buf, s->rps ? val : 0);
1414
1415         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/xps_cpus", dev->ifname);
1416         system_if_set_rps_xps_val(dev_buf, s->xps ? val : 0);
1417 }
1418
1419 void
1420 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1421 {
1422         struct ifreq ifr;
1423         char buf[12];
1424
1425         memset(&ifr, 0, sizeof(ifr));
1426         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1427         if (s->flags & DEV_OPT_MTU & apply_mask) {
1428                 ifr.ifr_mtu = s->mtu;
1429                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1430                         s->flags &= ~DEV_OPT_MTU;
1431         }
1432         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1433                 system_update_ipv6_mtu(dev, s->mtu6);
1434         }
1435         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1436                 ifr.ifr_qlen = s->txqueuelen;
1437                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1438                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1439         }
1440         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1441                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1442                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1443                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1444                         s->flags &= ~DEV_OPT_MACADDR;
1445         }
1446         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1447                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1448         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1449                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1450                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1451                         s->flags &= ~DEV_OPT_PROMISC;
1452         }
1453         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1454                 snprintf(buf, sizeof(buf), "%d", s->rpfilter);
1455                 system_set_rpfilter(dev, buf);
1456         }
1457         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1458                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1459         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1460                 snprintf(buf, sizeof(buf), "%d", s->igmpversion);
1461                 system_set_igmpversion(dev, buf);
1462         }
1463         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1464                 snprintf(buf, sizeof(buf), "%d", s->mldversion);
1465                 system_set_mldversion(dev, buf);
1466         }
1467         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1468                 snprintf(buf, sizeof(buf), "%d", s->neigh4reachabletime);
1469                 system_set_neigh4reachabletime(dev, buf);
1470                 snprintf(buf, sizeof(buf), "%d", s->neigh6reachabletime);
1471                 system_set_neigh6reachabletime(dev, buf);
1472         }
1473         if (s->flags & DEV_OPT_NEIGHLOCKTIME & apply_mask) {
1474                 snprintf(buf, sizeof(buf), "%d", s->neigh4locktime);
1475                 system_set_neigh4locktime(dev, buf);
1476         }
1477         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1478                 snprintf(buf, sizeof(buf), "%d", s->neigh4gcstaletime);
1479                 system_set_neigh4gcstaletime(dev, buf);
1480                 snprintf(buf, sizeof(buf), "%d", s->neigh6gcstaletime);
1481                 system_set_neigh6gcstaletime(dev, buf);
1482         }
1483         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1484                 snprintf(buf, sizeof(buf), "%d", s->dadtransmits);
1485                 system_set_dadtransmits(dev, buf);
1486         }
1487         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1488                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1489                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1490                         s->flags &= ~DEV_OPT_MULTICAST;
1491         }
1492         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1493                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1494
1495         system_if_apply_rps_xps(dev, s);
1496 }
1497
1498 int system_if_up(struct device *dev)
1499 {
1500         system_if_get_settings(dev, &dev->orig_settings);
1501         /* Only keep orig settings based on what needs to be set */
1502         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1503         dev->orig_settings.flags &= dev->settings.flags;
1504         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1505         return system_if_flags(dev->ifname, IFF_UP, 0);
1506 }
1507
1508 int system_if_down(struct device *dev)
1509 {
1510         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1511         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1512         return ret;
1513 }
1514
1515 struct if_check_data {
1516         struct device *dev;
1517         int pending;
1518         int ret;
1519 };
1520
1521 #ifndef IFF_LOWER_UP
1522 #define IFF_LOWER_UP    0x10000
1523 #endif
1524
1525 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1526 {
1527         struct nlmsghdr *nh = nlmsg_hdr(msg);
1528         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1529         struct if_check_data *chk = (struct if_check_data *)arg;
1530
1531         if (nh->nlmsg_type != RTM_NEWLINK)
1532                 return NL_SKIP;
1533
1534         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1535         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1536
1537         return NL_OK;
1538 }
1539
1540 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1541 {
1542         struct if_check_data *chk = (struct if_check_data *)arg;
1543         chk->pending = 0;
1544         return NL_STOP;
1545 }
1546
1547 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1548 {
1549         struct if_check_data *chk = (struct if_check_data *)arg;
1550
1551         device_set_present(chk->dev, false);
1552         device_set_link(chk->dev, false);
1553         chk->pending = err->error;
1554
1555         return NL_STOP;
1556 }
1557
1558 int system_if_check(struct device *dev)
1559 {
1560         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1561         struct nl_msg *msg;
1562         struct ifinfomsg ifi = {
1563                 .ifi_family = AF_UNSPEC,
1564                 .ifi_index = 0,
1565         };
1566         struct if_check_data chk = {
1567                 .dev = dev,
1568                 .pending = 1,
1569         };
1570         int ret = 1;
1571
1572         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1573         if (!msg)
1574                 goto out;
1575
1576         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1577             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1578                 goto free;
1579
1580         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1581         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1582         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1583
1584         nl_send_auto_complete(sock_rtnl, msg);
1585         while (chk.pending > 0)
1586                 nl_recvmsgs(sock_rtnl, cb);
1587
1588         ret = chk.pending;
1589
1590 free:
1591         nlmsg_free(msg);
1592 out:
1593         nl_cb_put(cb);
1594         return ret;
1595 }
1596
1597 struct device *
1598 system_if_get_parent(struct device *dev)
1599 {
1600         char buf[64], *devname;
1601         int ifindex, iflink, len;
1602         FILE *f;
1603
1604         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1605         f = fopen(buf, "r");
1606         if (!f)
1607                 return NULL;
1608
1609         len = fread(buf, 1, sizeof(buf) - 1, f);
1610         fclose(f);
1611
1612         if (len <= 0)
1613                 return NULL;
1614
1615         buf[len] = 0;
1616         iflink = strtoul(buf, NULL, 0);
1617         ifindex = system_if_resolve(dev);
1618         if (!iflink || iflink == ifindex)
1619                 return NULL;
1620
1621         devname = if_indextoname(iflink, buf);
1622         if (!devname)
1623                 return NULL;
1624
1625         return device_get(devname, true);
1626 }
1627
1628 static bool
1629 read_string_file(int dir_fd, const char *file, char *buf, int len)
1630 {
1631         bool ret = false;
1632         char *c;
1633         int fd;
1634
1635         fd = openat(dir_fd, file, O_RDONLY);
1636         if (fd < 0)
1637                 return false;
1638
1639 retry:
1640         len = read(fd, buf, len - 1);
1641         if (len < 0) {
1642                 if (errno == EINTR)
1643                         goto retry;
1644         } else if (len > 0) {
1645                         buf[len] = 0;
1646
1647                         c = strchr(buf, '\n');
1648                         if (c)
1649                                 *c = 0;
1650
1651                         ret = true;
1652         }
1653
1654         close(fd);
1655
1656         return ret;
1657 }
1658
1659 static bool
1660 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1661 {
1662         char buf[64];
1663         bool ret = false;
1664
1665         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1666         if (ret)
1667                 *val = strtoull(buf, NULL, 0);
1668
1669         return ret;
1670 }
1671
1672 /* Assume advertised flags == supported flags */
1673 static const struct {
1674         uint32_t mask;
1675         const char *name;
1676 } ethtool_link_modes[] = {
1677         { ADVERTISED_10baseT_Half, "10H" },
1678         { ADVERTISED_10baseT_Full, "10F" },
1679         { ADVERTISED_100baseT_Half, "100H" },
1680         { ADVERTISED_100baseT_Full, "100F" },
1681         { ADVERTISED_1000baseT_Half, "1000H" },
1682         { ADVERTISED_1000baseT_Full, "1000F" },
1683 };
1684
1685 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1686 {
1687         int i;
1688         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1689                 if (mask & ethtool_link_modes[i].mask)
1690                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1691         }
1692 }
1693
1694 bool
1695 system_if_force_external(const char *ifname)
1696 {
1697         char buf[64];
1698         struct stat s;
1699
1700         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1701         return stat(buf, &s) == 0;
1702 }
1703
1704 int
1705 system_if_dump_info(struct device *dev, struct blob_buf *b)
1706 {
1707         struct ethtool_cmd ecmd;
1708         struct ifreq ifr;
1709         char buf[64], *s;
1710         void *c;
1711         int dir_fd;
1712
1713         snprintf(buf, sizeof(buf), "/sys/class/net/%s", dev->ifname);
1714         dir_fd = open(buf, O_DIRECTORY);
1715
1716         memset(&ecmd, 0, sizeof(ecmd));
1717         memset(&ifr, 0, sizeof(ifr));
1718         strcpy(ifr.ifr_name, dev->ifname);
1719         ifr.ifr_data = (caddr_t) &ecmd;
1720         ecmd.cmd = ETHTOOL_GSET;
1721
1722         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1723                 c = blobmsg_open_array(b, "link-advertising");
1724                 system_add_link_modes(b, ecmd.advertising);
1725                 blobmsg_close_array(b, c);
1726
1727                 c = blobmsg_open_array(b, "link-supported");
1728                 system_add_link_modes(b, ecmd.supported);
1729                 blobmsg_close_array(b, c);
1730
1731                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1732                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1733                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1734                 blobmsg_add_string_buffer(b);
1735         }
1736
1737         close(dir_fd);
1738         return 0;
1739 }
1740
1741 int
1742 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1743 {
1744         const char *const counters[] = {
1745                 "collisions",     "rx_frame_errors",   "tx_compressed",
1746                 "multicast",      "rx_length_errors",  "tx_dropped",
1747                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1748                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1749                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1750                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1751                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1752                 "rx_fifo_errors", "tx_carrier_errors",
1753         };
1754         char buf[64];
1755         int stats_dir;
1756         int i;
1757         uint64_t val = 0;
1758
1759         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1760         stats_dir = open(buf, O_DIRECTORY);
1761         if (stats_dir < 0)
1762                 return -1;
1763
1764         for (i = 0; i < ARRAY_SIZE(counters); i++)
1765                 if (read_uint64_file(stats_dir, counters[i], &val))
1766                         blobmsg_add_u64(b, counters[i], val);
1767
1768         close(stats_dir);
1769         return 0;
1770 }
1771
1772 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1773 {
1774         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1775         int alen = v4 ? 4 : 16;
1776         unsigned int flags = 0;
1777         struct ifaddrmsg ifa = {
1778                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1779                 .ifa_prefixlen = addr->mask,
1780                 .ifa_index = dev->ifindex,
1781         };
1782
1783         struct nl_msg *msg;
1784         if (cmd == RTM_NEWADDR)
1785                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1786
1787         msg = nlmsg_alloc_simple(cmd, flags);
1788         if (!msg)
1789                 return -1;
1790
1791         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1792         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1793         if (v4) {
1794                 if (addr->broadcast)
1795                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1796                 if (addr->point_to_point)
1797                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1798         } else {
1799                 time_t now = system_get_rtime();
1800                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1801
1802                 if (addr->preferred_until) {
1803                         int64_t preferred = addr->preferred_until - now;
1804                         if (preferred < 0)
1805                                 preferred = 0;
1806                         else if (preferred > UINT32_MAX)
1807                                 preferred = UINT32_MAX;
1808
1809                         cinfo.ifa_prefered = preferred;
1810                 }
1811
1812                 if (addr->valid_until) {
1813                         int64_t valid = addr->valid_until - now;
1814                         if (valid <= 0) {
1815                                 nlmsg_free(msg);
1816                                 return -1;
1817                         }
1818                         else if (valid > UINT32_MAX)
1819                                 valid = UINT32_MAX;
1820
1821                         cinfo.ifa_valid = valid;
1822                 }
1823
1824                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
1825
1826                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
1827                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
1828         }
1829
1830         return system_rtnl_call(msg);
1831 }
1832
1833 int system_add_address(struct device *dev, struct device_addr *addr)
1834 {
1835         return system_addr(dev, addr, RTM_NEWADDR);
1836 }
1837
1838 int system_del_address(struct device *dev, struct device_addr *addr)
1839 {
1840         return system_addr(dev, addr, RTM_DELADDR);
1841 }
1842
1843 static int system_rt(struct device *dev, struct device_route *route, int cmd)
1844 {
1845         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1846         bool have_gw;
1847         unsigned int flags = 0;
1848
1849         if (alen == 4)
1850                 have_gw = !!route->nexthop.in.s_addr;
1851         else
1852                 have_gw = route->nexthop.in6.s6_addr32[0] ||
1853                         route->nexthop.in6.s6_addr32[1] ||
1854                         route->nexthop.in6.s6_addr32[2] ||
1855                         route->nexthop.in6.s6_addr32[3];
1856
1857         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
1858                         ? route->table : RT_TABLE_MAIN;
1859
1860         struct rtmsg rtm = {
1861                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
1862                 .rtm_dst_len = route->mask,
1863                 .rtm_src_len = route->sourcemask,
1864                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
1865                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
1866                 .rtm_scope = RT_SCOPE_NOWHERE,
1867                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
1868                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
1869         };
1870         struct nl_msg *msg;
1871
1872         if (cmd == RTM_NEWROUTE) {
1873                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1874
1875                 if (!dev) { // Add null-route
1876                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1877                         rtm.rtm_type = RTN_UNREACHABLE;
1878                 }
1879                 else
1880                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
1881         }
1882
1883         if (route->flags & DEVROUTE_TYPE) {
1884                 rtm.rtm_type = route->type;
1885                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
1886                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
1887                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
1888                                 rtm.rtm_table = RT_TABLE_LOCAL;
1889                 }
1890
1891                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
1892                         rtm.rtm_scope = RT_SCOPE_HOST;
1893                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
1894                                 rtm.rtm_type == RTN_ANYCAST) {
1895                         rtm.rtm_scope = RT_SCOPE_LINK;
1896                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
1897                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY) {
1898                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1899                         dev = NULL;
1900                 }
1901         }
1902
1903         msg = nlmsg_alloc_simple(cmd, flags);
1904         if (!msg)
1905                 return -1;
1906
1907         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
1908
1909         if (route->mask)
1910                 nla_put(msg, RTA_DST, alen, &route->addr);
1911
1912         if (route->sourcemask) {
1913                 if (rtm.rtm_family == AF_INET)
1914                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
1915                 else
1916                         nla_put(msg, RTA_SRC, alen, &route->source);
1917         }
1918
1919         if (route->metric > 0)
1920                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
1921
1922         if (have_gw)
1923                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
1924
1925         if (dev)
1926                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
1927
1928         if (table >= 256)
1929                 nla_put_u32(msg, RTA_TABLE, table);
1930
1931         if (route->flags & DEVROUTE_MTU) {
1932                 struct nlattr *metrics;
1933
1934                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
1935                         goto nla_put_failure;
1936
1937                 nla_put_u32(msg, RTAX_MTU, route->mtu);
1938
1939                 nla_nest_end(msg, metrics);
1940         }
1941
1942         return system_rtnl_call(msg);
1943
1944 nla_put_failure:
1945         nlmsg_free(msg);
1946         return -ENOMEM;
1947 }
1948
1949 int system_add_route(struct device *dev, struct device_route *route)
1950 {
1951         return system_rt(dev, route, RTM_NEWROUTE);
1952 }
1953
1954 int system_del_route(struct device *dev, struct device_route *route)
1955 {
1956         return system_rt(dev, route, RTM_DELROUTE);
1957 }
1958
1959 int system_flush_routes(void)
1960 {
1961         const char *names[] = {
1962                 "/proc/sys/net/ipv4/route/flush",
1963                 "/proc/sys/net/ipv6/route/flush"
1964         };
1965         int fd, i;
1966
1967         for (i = 0; i < ARRAY_SIZE(names); i++) {
1968                 fd = open(names[i], O_WRONLY);
1969                 if (fd < 0)
1970                         continue;
1971
1972                 if (write(fd, "-1", 2)) {}
1973                 close(fd);
1974         }
1975         return 0;
1976 }
1977
1978 bool system_resolve_rt_type(const char *type, unsigned int *id)
1979 {
1980         return system_rtn_aton(type, id);
1981 }
1982
1983 bool system_resolve_rt_proto(const char *type, unsigned int *id)
1984 {
1985         FILE *f;
1986         char *e, buf[128];
1987         unsigned int n, proto = 256;
1988
1989         if ((n = strtoul(type, &e, 0)) >= 0 && !*e && e != type)
1990                 proto = n;
1991         else if (!strcmp(type, "unspec"))
1992                 proto = RTPROT_UNSPEC;
1993         else if (!strcmp(type, "kernel"))
1994                 proto = RTPROT_KERNEL;
1995         else if (!strcmp(type, "boot"))
1996                 proto = RTPROT_BOOT;
1997         else if (!strcmp(type, "static"))
1998                 proto = RTPROT_STATIC;
1999         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
2000                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
2001                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2002                                 continue;
2003
2004                         n = strtoul(e, NULL, 10);
2005                         e = strtok(NULL, " \t\n");
2006
2007                         if (e && !strcmp(e, type)) {
2008                                 proto = n;
2009                                 break;
2010                         }
2011                 }
2012                 fclose(f);
2013         }
2014
2015         if (proto > 255)
2016                 return false;
2017
2018         *id = proto;
2019         return true;
2020 }
2021
2022 bool system_resolve_rt_table(const char *name, unsigned int *id)
2023 {
2024         FILE *f;
2025         char *e, buf[128];
2026         unsigned int n, table = RT_TABLE_UNSPEC;
2027
2028         /* first try to parse table as number */
2029         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2030                 table = n;
2031
2032         /* handle well known aliases */
2033         else if (!strcmp(name, "default"))
2034                 table = RT_TABLE_DEFAULT;
2035         else if (!strcmp(name, "main"))
2036                 table = RT_TABLE_MAIN;
2037         else if (!strcmp(name, "local"))
2038                 table = RT_TABLE_LOCAL;
2039
2040         /* try to look up name in /etc/iproute2/rt_tables */
2041         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2042         {
2043                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2044                 {
2045                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2046                                 continue;
2047
2048                         n = strtoul(e, NULL, 10);
2049                         e = strtok(NULL, " \t\n");
2050
2051                         if (e && !strcmp(e, name))
2052                         {
2053                                 table = n;
2054                                 break;
2055                         }
2056                 }
2057
2058                 fclose(f);
2059         }
2060
2061         if (table == RT_TABLE_UNSPEC)
2062                 return false;
2063
2064         *id = table;
2065         return true;
2066 }
2067
2068 bool system_is_default_rt_table(unsigned int id)
2069 {
2070         return (id == RT_TABLE_MAIN);
2071 }
2072
2073 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2074 {
2075         char *e;
2076         unsigned int n;
2077
2078         if (!strcmp(filter, "strict"))
2079                 n = 1;
2080         else if (!strcmp(filter, "loose"))
2081                 n = 2;
2082         else {
2083                 n = strtoul(filter, &e, 0);
2084                 if (*e || e == filter || n > 2)
2085                         return false;
2086         }
2087
2088         *id = n;
2089         return true;
2090 }
2091
2092 static int system_iprule(struct iprule *rule, int cmd)
2093 {
2094         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2095
2096         struct nl_msg *msg;
2097         struct rtmsg rtm = {
2098                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2099                 .rtm_protocol = RTPROT_STATIC,
2100                 .rtm_scope = RT_SCOPE_UNIVERSE,
2101                 .rtm_table = RT_TABLE_UNSPEC,
2102                 .rtm_type = RTN_UNSPEC,
2103                 .rtm_flags = 0,
2104         };
2105
2106         if (cmd == RTM_NEWRULE)
2107                 rtm.rtm_type = RTN_UNICAST;
2108
2109         if (rule->invert)
2110                 rtm.rtm_flags |= FIB_RULE_INVERT;
2111
2112         if (rule->flags & IPRULE_SRC)
2113                 rtm.rtm_src_len = rule->src_mask;
2114
2115         if (rule->flags & IPRULE_DEST)
2116                 rtm.rtm_dst_len = rule->dest_mask;
2117
2118         if (rule->flags & IPRULE_TOS)
2119                 rtm.rtm_tos = rule->tos;
2120
2121         if (rule->flags & IPRULE_LOOKUP) {
2122                 if (rule->lookup < 256)
2123                         rtm.rtm_table = rule->lookup;
2124         }
2125
2126         if (rule->flags & IPRULE_ACTION)
2127                 rtm.rtm_type = rule->action;
2128         else if (rule->flags & IPRULE_GOTO)
2129                 rtm.rtm_type = FR_ACT_GOTO;
2130         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2131                 rtm.rtm_type = FR_ACT_NOP;
2132
2133         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2134
2135         if (!msg)
2136                 return -1;
2137
2138         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2139
2140         if (rule->flags & IPRULE_IN)
2141                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2142
2143         if (rule->flags & IPRULE_OUT)
2144                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2145
2146         if (rule->flags & IPRULE_SRC)
2147                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2148
2149         if (rule->flags & IPRULE_DEST)
2150                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2151
2152         if (rule->flags & IPRULE_PRIORITY)
2153                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2154         else if (cmd == RTM_NEWRULE)
2155                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2156
2157         if (rule->flags & IPRULE_FWMARK)
2158                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2159
2160         if (rule->flags & IPRULE_FWMASK)
2161                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2162
2163         if (rule->flags & IPRULE_LOOKUP) {
2164                 if (rule->lookup >= 256)
2165                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2166         }
2167
2168         if (rule->flags & IPRULE_GOTO)
2169                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2170
2171         return system_rtnl_call(msg);
2172 }
2173
2174 int system_add_iprule(struct iprule *rule)
2175 {
2176         return system_iprule(rule, RTM_NEWRULE);
2177 }
2178
2179 int system_del_iprule(struct iprule *rule)
2180 {
2181         return system_iprule(rule, RTM_DELRULE);
2182 }
2183
2184 int system_flush_iprules(void)
2185 {
2186         int rv = 0;
2187         struct iprule rule;
2188
2189         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2190         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2191
2192         memset(&rule, 0, sizeof(rule));
2193
2194
2195         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2196
2197         rule.priority = 0;
2198         rule.lookup = RT_TABLE_LOCAL;
2199         rv |= system_iprule(&rule, RTM_NEWRULE);
2200
2201         rule.priority = 32766;
2202         rule.lookup = RT_TABLE_MAIN;
2203         rv |= system_iprule(&rule, RTM_NEWRULE);
2204
2205         rule.priority = 32767;
2206         rule.lookup = RT_TABLE_DEFAULT;
2207         rv |= system_iprule(&rule, RTM_NEWRULE);
2208
2209
2210         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2211
2212         rule.priority = 0;
2213         rule.lookup = RT_TABLE_LOCAL;
2214         rv |= system_iprule(&rule, RTM_NEWRULE);
2215
2216         rule.priority = 32766;
2217         rule.lookup = RT_TABLE_MAIN;
2218         rv |= system_iprule(&rule, RTM_NEWRULE);
2219
2220         return rv;
2221 }
2222
2223 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2224 {
2225         return system_rtn_aton(action, id);
2226 }
2227
2228 time_t system_get_rtime(void)
2229 {
2230         struct timespec ts;
2231         struct timeval tv;
2232
2233         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
2234                 return ts.tv_sec;
2235
2236         if (gettimeofday(&tv, NULL) == 0)
2237                 return tv.tv_sec;
2238
2239         return 0;
2240 }
2241
2242 #ifndef IP_DF
2243 #define IP_DF       0x4000
2244 #endif
2245
2246 static int tunnel_ioctl(const char *name, int cmd, void *p)
2247 {
2248         struct ifreq ifr;
2249
2250         memset(&ifr, 0, sizeof(ifr));
2251         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name));
2252         ifr.ifr_ifru.ifru_data = p;
2253         return ioctl(sock_ioctl, cmd, &ifr);
2254 }
2255
2256 #ifdef IFLA_IPTUN_MAX
2257 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2258 static int system_add_gre_tunnel(const char *name, const char *kind,
2259                                  const unsigned int link, struct blob_attr **tb, bool v6)
2260 {
2261         struct nl_msg *nlm;
2262         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2263         struct blob_attr *cur;
2264         uint32_t ikey = 0, okey = 0, flags = 0, flowinfo = 0;
2265         uint16_t iflags = 0, oflags = 0;
2266         uint8_t tos = 0;
2267         int ret = 0, ttl = 0;
2268
2269         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2270         if (!nlm)
2271                 return -1;
2272
2273         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2274         nla_put_string(nlm, IFLA_IFNAME, name);
2275
2276         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2277         if (!linkinfo) {
2278                 ret = -ENOMEM;
2279                 goto failure;
2280         }
2281
2282         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2283         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2284         if (!infodata) {
2285                 ret = -ENOMEM;
2286                 goto failure;
2287         }
2288
2289         if (link)
2290                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2291
2292         if ((cur = tb[TUNNEL_ATTR_TTL]))
2293                 ttl = blobmsg_get_u32(cur);
2294
2295         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2296                 char *str = blobmsg_get_string(cur);
2297                 if (strcmp(str, "inherit")) {
2298                         unsigned uval;
2299
2300                         if (!system_tos_aton(str, &uval)) {
2301                                 ret = -EINVAL;
2302                                 goto failure;
2303                         }
2304
2305                         if (v6)
2306                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2307                         else
2308                                 tos = uval;
2309                 } else {
2310                         if (v6)
2311                                 flags |= IP6_TNL_F_USE_ORIG_TCLASS;
2312                         else
2313                                 tos = 1;
2314                 }
2315         }
2316
2317         if ((cur = tb[TUNNEL_ATTR_INFO]) && (blobmsg_type(cur) == BLOBMSG_TYPE_STRING)) {
2318                 uint8_t icsum, ocsum, iseqno, oseqno;
2319                 if (sscanf(blobmsg_get_string(cur), "%u,%u,%hhu,%hhu,%hhu,%hhu",
2320                         &ikey, &okey, &icsum, &ocsum, &iseqno, &oseqno) < 6) {
2321                         ret = -EINVAL;
2322                         goto failure;
2323                 }
2324
2325                 if (ikey)
2326                         iflags |= GRE_KEY;
2327
2328                 if (okey)
2329                         oflags |= GRE_KEY;
2330
2331                 if (icsum)
2332                         iflags |= GRE_CSUM;
2333
2334                 if (ocsum)
2335                         oflags |= GRE_CSUM;
2336
2337                 if (iseqno)
2338                         iflags |= GRE_SEQ;
2339
2340                 if (oseqno)
2341                         oflags |= GRE_SEQ;
2342         }
2343
2344         if (v6) {
2345                 struct in6_addr in6buf;
2346                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2347                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2348                                 ret = -EINVAL;
2349                                 goto failure;
2350                         }
2351                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2352                 }
2353
2354                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2355                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2356                                 ret = -EINVAL;
2357                                 goto failure;
2358                         }
2359                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2360                 }
2361                 nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, 4);
2362
2363                 if (flowinfo)
2364                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2365
2366                 if (flags)
2367                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags);
2368
2369                 if (!ttl)
2370                         ttl = 64;
2371         } else {
2372                 struct in_addr inbuf;
2373                 bool set_df = true;
2374
2375                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2376                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2377                                 ret = -EINVAL;
2378                                 goto failure;
2379                         }
2380                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2381                 }
2382
2383                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2384                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2385                                 ret = -EINVAL;
2386                                 goto failure;
2387                         }
2388                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2389
2390                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2391                                 if (!okey) {
2392                                         okey = inbuf.s_addr;
2393                                         oflags |= GRE_KEY;
2394                                 }
2395
2396                                 if (!ikey) {
2397                                         ikey = inbuf.s_addr;
2398                                         iflags |= GRE_KEY;
2399                                 }
2400                         }
2401                 }
2402
2403                 if ((cur = tb[TUNNEL_ATTR_DF]))
2404                         set_df = blobmsg_get_bool(cur);
2405
2406                 if (!set_df) {
2407                         /* ttl != 0 and nopmtudisc are incompatible */
2408                         if (ttl) {
2409                                 ret = -EINVAL;
2410                                 goto failure;
2411                         }
2412                 } else if (!ttl)
2413                         ttl = 64;
2414
2415                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2416
2417                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2418         }
2419
2420         if (ttl)
2421                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2422
2423         if (oflags)
2424                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2425
2426         if (iflags)
2427                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2428
2429         if (okey)
2430                 nla_put_u32(nlm, IFLA_GRE_OKEY, okey);
2431
2432         if (ikey)
2433                 nla_put_u32(nlm, IFLA_GRE_IKEY, ikey);
2434
2435         nla_nest_end(nlm, infodata);
2436         nla_nest_end(nlm, linkinfo);
2437
2438         return system_rtnl_call(nlm);
2439
2440 failure:
2441         nlmsg_free(nlm);
2442         return ret;
2443 }
2444 #endif
2445
2446 #ifdef IFLA_VTI_MAX
2447 static int system_add_vti_tunnel(const char *name, const char *kind,
2448                                  const unsigned int link, struct blob_attr **tb, bool v6)
2449 {
2450         struct nl_msg *nlm;
2451         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2452         struct blob_attr *cur;
2453         uint32_t ikey = 0, okey = 0;
2454         int ret = 0;
2455
2456         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2457         if (!nlm)
2458                 return -1;
2459
2460         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2461         nla_put_string(nlm, IFLA_IFNAME, name);
2462
2463         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2464         if (!linkinfo) {
2465                 ret = -ENOMEM;
2466                 goto failure;
2467         }
2468
2469         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2470         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2471         if (!infodata) {
2472                 ret = -ENOMEM;
2473                 goto failure;
2474         }
2475
2476         if (link)
2477                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2478
2479         if ((cur = tb[TUNNEL_ATTR_INFO]) && (blobmsg_type(cur) == BLOBMSG_TYPE_STRING)) {
2480                 if (sscanf(blobmsg_get_string(cur), "%u,%u",
2481                         &ikey, &okey) < 2) {
2482                         ret = -EINVAL;
2483                         goto failure;
2484                 }
2485         }
2486
2487         if (v6) {
2488                 struct in6_addr in6buf;
2489                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2490                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2491                                 ret = -EINVAL;
2492                                 goto failure;
2493                         }
2494                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2495                 }
2496
2497                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2498                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2499                                 ret = -EINVAL;
2500                                 goto failure;
2501                         }
2502                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2503                 }
2504
2505         } else {
2506                 struct in_addr inbuf;
2507
2508                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2509                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2510                                 ret = -EINVAL;
2511                                 goto failure;
2512                         }
2513                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2514                 }
2515
2516                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2517                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2518                                 ret = -EINVAL;
2519                                 goto failure;
2520                         }
2521                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2522                 }
2523
2524         }
2525
2526         if (okey)
2527                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2528
2529         if (ikey)
2530                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2531
2532         nla_nest_end(nlm, infodata);
2533         nla_nest_end(nlm, linkinfo);
2534
2535         return system_rtnl_call(nlm);
2536
2537 failure:
2538         nlmsg_free(nlm);
2539         return ret;
2540 }
2541 #endif
2542
2543 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
2544 {
2545         struct blob_attr *cur;
2546         bool set_df = true;
2547         struct ip_tunnel_parm p  = {
2548                 .link = link,
2549                 .iph = {
2550                         .version = 4,
2551                         .ihl = 5,
2552                         .protocol = proto,
2553                 }
2554         };
2555
2556         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
2557                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
2558                 return -EINVAL;
2559
2560         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
2561                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
2562                 return -EINVAL;
2563
2564         if ((cur = tb[TUNNEL_ATTR_DF]))
2565                 set_df = blobmsg_get_bool(cur);
2566
2567         if ((cur = tb[TUNNEL_ATTR_TTL]))
2568                 p.iph.ttl = blobmsg_get_u32(cur);
2569
2570         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2571                 char *str = blobmsg_get_string(cur);
2572                 if (strcmp(str, "inherit")) {
2573                         unsigned uval;
2574
2575                         if (!system_tos_aton(str, &uval))
2576                                 return -EINVAL;
2577
2578                         p.iph.tos = uval;
2579                 } else
2580                         p.iph.tos = 1;
2581         }
2582
2583         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
2584         /* ttl !=0 and nopmtudisc are incompatible */
2585         if (p.iph.ttl && p.iph.frag_off == 0)
2586                 return -EINVAL;
2587
2588         strncpy(p.name, name, sizeof(p.name));
2589
2590         switch (p.iph.protocol) {
2591         case IPPROTO_IPIP:
2592                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
2593         case IPPROTO_IPV6:
2594                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
2595         default:
2596                 break;
2597         }
2598         return -1;
2599 }
2600
2601 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
2602 {
2603         struct blob_attr *cur;
2604         const char *str;
2605
2606         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2607                 return -EINVAL;
2608         str = blobmsg_data(cur);
2609
2610         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
2611             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
2612             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6"))
2613                 return system_link_del(name);
2614         else
2615                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
2616 }
2617
2618 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
2619 {
2620         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2621
2622         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2623                 blob_data(attr), blob_len(attr));
2624
2625         return __system_del_ip_tunnel(name, tb);
2626 }
2627
2628 int system_update_ipv6_mtu(struct device *dev, int mtu)
2629 {
2630         int ret = -1;
2631         char buf[64];
2632         int fd;
2633
2634         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
2635                         dev->ifname);
2636
2637         fd = open(buf, O_RDWR);
2638         if (fd < 0)
2639                 return ret;
2640
2641         if (!mtu) {
2642                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
2643                 if (len < 0)
2644                         goto out;
2645
2646                 buf[len] = 0;
2647                 ret = atoi(buf);
2648         } else {
2649                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
2650                         ret = mtu;
2651         }
2652
2653 out:
2654         close(fd);
2655         return ret;
2656 }
2657
2658 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
2659 {
2660         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2661         struct blob_attr *cur;
2662         const char *str;
2663
2664         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2665                 blob_data(attr), blob_len(attr));
2666
2667         __system_del_ip_tunnel(name, tb);
2668
2669         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2670                 return -EINVAL;
2671         str = blobmsg_data(cur);
2672
2673         unsigned int ttl = 0;
2674         if ((cur = tb[TUNNEL_ATTR_TTL])) {
2675                 ttl = blobmsg_get_u32(cur);
2676                 if (ttl > 255)
2677                         return -EINVAL;
2678         }
2679
2680         unsigned int link = 0;
2681         if ((cur = tb[TUNNEL_ATTR_LINK])) {
2682                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
2683                 if (!iface)
2684                         return -EINVAL;
2685
2686                 if (iface->l3_dev.dev)
2687                         link = iface->l3_dev.dev->ifindex;
2688         }
2689
2690         if (!strcmp(str, "sit")) {
2691                 if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
2692                         return -1;
2693
2694 #ifdef SIOCADD6RD
2695                 if ((cur = tb[TUNNEL_ATTR_6RD_PREFIX])) {
2696                         unsigned int mask;
2697                         struct ip_tunnel_6rd p6;
2698
2699                         memset(&p6, 0, sizeof(p6));
2700
2701                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
2702                                                 &p6.prefix, &mask) || mask > 128)
2703                                 return -EINVAL;
2704                         p6.prefixlen = mask;
2705
2706                         if ((cur = tb[TUNNEL_ATTR_6RD_RELAY_PREFIX])) {
2707                                 if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
2708                                                         &p6.relay_prefix, &mask) || mask > 32)
2709                                         return -EINVAL;
2710                                 p6.relay_prefixlen = mask;
2711                         }
2712
2713                         if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
2714                                 __system_del_ip_tunnel(name, tb);
2715                                 return -1;
2716                         }
2717                 }
2718 #endif
2719 #ifdef IFLA_IPTUN_MAX
2720         } else if (!strcmp(str, "ipip6")) {
2721                 struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2722                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2723                 struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2724                 int ret = 0;
2725
2726                 if (!nlm)
2727                         return -1;
2728
2729                 nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2730                 nla_put_string(nlm, IFLA_IFNAME, name);
2731
2732                 if (link)
2733                         nla_put_u32(nlm, IFLA_LINK, link);
2734
2735                 struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2736                 if (!linkinfo) {
2737                         ret = -ENOMEM;
2738                         goto failure;
2739                 }
2740                 nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2741                 struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2742                 if (!infodata) {
2743                         ret = -ENOMEM;
2744                         goto failure;
2745                 }
2746
2747                 if (link)
2748                         nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2749
2750                 nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2751                 nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2752                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, 4);
2753
2754                 struct in6_addr in6buf;
2755                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2756                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2757                                 ret = -EINVAL;
2758                                 goto failure;
2759                         }
2760                         nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2761                 }
2762
2763                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2764                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2765                                 ret = -EINVAL;
2766                                 goto failure;
2767                         }
2768                         nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2769                 }
2770
2771 #ifdef IFLA_IPTUN_FMR_MAX
2772                 if ((cur = tb[TUNNEL_ATTR_FMRS])) {
2773                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2774
2775                         struct blob_attr *fmr;
2776                         unsigned rem, fmrcnt = 0;
2777                         blobmsg_for_each_attr(fmr, cur, rem) {
2778                                 if (blobmsg_type(fmr) != BLOBMSG_TYPE_STRING)
2779                                         continue;
2780
2781                                 unsigned ip4len, ip6len, ealen, offset = 6;
2782                                 char ip6buf[48];
2783                                 char ip4buf[16];
2784
2785                                 if (sscanf(blobmsg_get_string(fmr), "%47[^/]/%u,%15[^/]/%u,%u,%u",
2786                                                 ip6buf, &ip6len, ip4buf, &ip4len, &ealen, &offset) < 5) {
2787                                         ret = -EINVAL;
2788                                         goto failure;
2789                                 }
2790
2791                                 struct in6_addr ip6prefix;
2792                                 struct in_addr ip4prefix;
2793                                 if (inet_pton(AF_INET6, ip6buf, &ip6prefix) != 1 ||
2794                                                 inet_pton(AF_INET, ip4buf, &ip4prefix) != 1) {
2795                                         ret = -EINVAL;
2796                                         goto failure;
2797                                 }
2798
2799                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2800
2801                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2802                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2803                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2804                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2805                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2806                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2807
2808                                 nla_nest_end(nlm, rule);
2809                         }
2810
2811                         nla_nest_end(nlm, fmrs);
2812                 }
2813 #endif
2814
2815                 nla_nest_end(nlm, infodata);
2816                 nla_nest_end(nlm, linkinfo);
2817
2818                 return system_rtnl_call(nlm);
2819 failure:
2820                 nlmsg_free(nlm);
2821                 return ret;
2822         } else if (!strcmp(str, "greip")) {
2823                 return system_add_gre_tunnel(name, "gre", link, tb, false);
2824         } else if (!strcmp(str, "gretapip"))  {
2825                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
2826         } else if (!strcmp(str, "greip6")) {
2827                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
2828         } else if (!strcmp(str, "gretapip6")) {
2829                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
2830 #ifdef IFLA_VTI_MAX
2831         } else if (!strcmp(str, "vtiip")) {
2832                 return system_add_vti_tunnel(name, "vti", link, tb, false);
2833         } else if (!strcmp(str, "vtiip6")) {
2834                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
2835 #endif
2836 #endif
2837         } else if (!strcmp(str, "ipip")) {
2838                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
2839         }
2840         else
2841                 return -EINVAL;
2842
2843         return 0;
2844 }