system-linux.c

   1 /*
   2  * netifd - network interface daemon
   3  * Copyright (C) 2012 Felix Fietkau <nbd@openwrt.org>
   4  * Copyright (C) 2013 Jo-Philipp Wich <jow@openwrt.org>
   5  * Copyright (C) 2013 Steven Barth <steven@midlink.org>
   6  * Copyright (C) 2014 Gioacchino Mazzurco <gio@eigenlab.org>
   7  * Copyright (C) 2017 Matthias Schiffer <mschiffer@universe-factory.net>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2
  11  * as published by the Free Software Foundation
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  */
  18 #define _GNU_SOURCE
  19
  20 #include <sys/socket.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/stat.h>
  23 #include <sys/syscall.h>
  24
  25 #include <net/if.h>
  26 #include <net/if_arp.h>
  27
  28 #include <arpa/inet.h>
  29 #include <netinet/ether.h>
  30 #include <netinet/in.h>
  31
  32 #include <linux/rtnetlink.h>
  33 #include <linux/sockios.h>
  34 #include <linux/ip.h>
  35 #include <linux/if_addr.h>
  36 #include <linux/if_link.h>
  37 #include <linux/if_vlan.h>
  38 #include <linux/if_bridge.h>
  39 #include <linux/if_tunnel.h>
  40 #include <linux/ip6_tunnel.h>
  41 #include <linux/ethtool.h>
  42 #include <linux/fib_rules.h>
  43 #include <linux/veth.h>
  44 #include <linux/version.h>
  45
  46 #ifndef RTN_FAILED_POLICY
  47 #define RTN_FAILED_POLICY 12
  48 #endif
  49
  50 #ifndef IFA_F_NOPREFIXROUTE
  51 #define IFA_F_NOPREFIXROUTE 0x200
  52 #endif
  53
  54 #ifndef IFA_FLAGS
  55 #define IFA_FLAGS (IFA_MULTICAST + 1)
  56 #endif
  57
  58 #include <string.h>
  59 #include <fcntl.h>
  60 #include <glob.h>
  61 #include <time.h>
  62 #include <unistd.h>
  63
  64 #include <netlink/msg.h>
  65 #include <netlink/attr.h>
  66 #include <netlink/socket.h>
  67 #include <libubox/uloop.h>
  68
  69 #include "netifd.h"
  70 #include "device.h"
  71 #include "system.h"
  72
  73 struct event_socket {
  74         struct uloop_fd uloop;
  75         struct nl_sock *sock;
  76         int bufsize;
  77 };
  78
  79 static int sock_ioctl = -1;
  80 static struct nl_sock *sock_rtnl = NULL;
  81
  82 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  83 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  84
  85 static char dev_buf[256];
  86
  87 static void
  88 handler_nl_event(struct uloop_fd *u, unsigned int events)
  89 {
  90         struct event_socket *ev = container_of(u, struct event_socket, uloop);
  91         int err;
  92         socklen_t errlen = sizeof(err);
  93
  94         if (!u->error) {
  95                 nl_recvmsgs_default(ev->sock);
  96                 return;
  97         }
  98
  99         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 100                 goto abort;
 101
 102         switch(err) {
 103         case ENOBUFS:
 104                 // Increase rx buffer size on netlink socket
 105                 ev->bufsize *= 2;
 106                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 107                         goto abort;
 108
 109                 // Request full dump since some info got dropped
 110                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 111                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 112                 break;
 113
 114         default:
 115                 goto abort;
 116         }
 117         u->error = false;
 118         return;
 119
 120 abort:
 121         uloop_fd_delete(&ev->uloop);
 122         return;
 123 }
 124
 125 static struct nl_sock *
 126 create_socket(int protocol, int groups)
 127 {
 128         struct nl_sock *sock;
 129
 130         sock = nl_socket_alloc();
 131         if (!sock)
 132                 return NULL;
 133
 134         if (groups)
 135                 nl_join_groups(sock, groups);
 136
 137         if (nl_connect(sock, protocol))
 138                 return NULL;
 139
 140         return sock;
 141 }
 142
 143 static bool
 144 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 145                         uloop_fd_handler cb, int flags)
 146 {
 147         ev->sock = create_socket(protocol, groups);
 148         if (!ev->sock)
 149                 return false;
 150
 151         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 152         ev->uloop.cb = cb;
 153         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 154                 return false;
 155
 156         return true;
 157 }
 158
 159 static bool
 160 create_event_socket(struct event_socket *ev, int protocol,
 161                     int (*cb)(struct nl_msg *msg, void *arg))
 162 {
 163         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 164                 return false;
 165
 166         // Install the valid custom callback handler
 167         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 168
 169         // Disable sequence number checking on event sockets
 170         nl_socket_disable_seq_check(ev->sock);
 171
 172         // Increase rx buffer size to 65K on event sockets
 173         ev->bufsize = 65535;
 174         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 175                 return false;
 176
 177         return true;
 178 }
 179
 180 static bool
 181 system_rtn_aton(const char *src, unsigned int *dst)
 182 {
 183         char *e;
 184         unsigned int n;
 185
 186         if (!strcmp(src, "local"))
 187                 n = RTN_LOCAL;
 188         else if (!strcmp(src, "nat"))
 189                 n = RTN_NAT;
 190         else if (!strcmp(src, "broadcast"))
 191                 n = RTN_BROADCAST;
 192         else if (!strcmp(src, "anycast"))
 193                 n = RTN_ANYCAST;
 194         else if (!strcmp(src, "multicast"))
 195                 n = RTN_MULTICAST;
 196         else if (!strcmp(src, "prohibit"))
 197                 n = RTN_PROHIBIT;
 198         else if (!strcmp(src, "unreachable"))
 199                 n = RTN_UNREACHABLE;
 200         else if (!strcmp(src, "blackhole"))
 201                 n = RTN_BLACKHOLE;
 202         else if (!strcmp(src, "xresolve"))
 203                 n = RTN_XRESOLVE;
 204         else if (!strcmp(src, "unicast"))
 205                 n = RTN_UNICAST;
 206         else if (!strcmp(src, "throw"))
 207                 n = RTN_THROW;
 208         else if (!strcmp(src, "failed_policy"))
 209                 n = RTN_FAILED_POLICY;
 210         else {
 211                 n = strtoul(src, &e, 0);
 212                 if (!e || *e || e == src || n > 255)
 213                         return false;
 214         }
 215
 216         *dst = n;
 217         return true;
 218 }
 219
 220 static bool
 221 system_tos_aton(const char *src, unsigned *dst)
 222 {
 223         char *e;
 224
 225         *dst = strtoul(src, &e, 16);
 226         if (e == src || *e || *dst > 255)
 227                 return false;
 228
 229         return true;
 230 }
 231
 232 int system_init(void)
 233 {
 234         static struct event_socket rtnl_event;
 235         static struct event_socket hotplug_event;
 236
 237         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 238         system_fd_set_cloexec(sock_ioctl);
 239
 240         // Prepare socket for routing / address control
 241         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 242         if (!sock_rtnl)
 243                 return -1;
 244
 245         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 246                 return -1;
 247
 248         if (!create_raw_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT, 1,
 249                                         handle_hotplug_event, 0))
 250                 return -1;
 251
 252         // Receive network link events form kernel
 253         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 254
 255         return 0;
 256 }
 257
 258 static void system_set_sysctl(const char *path, const char *val)
 259 {
 260         int fd;
 261
 262         fd = open(path, O_WRONLY);
 263         if (fd < 0)
 264                 return;
 265
 266         if (write(fd, val, strlen(val))) {}
 267         close(fd);
 268 }
 269
 270 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 271 {
 272         snprintf(dev_buf, sizeof(dev_buf), path, device);
 273         system_set_sysctl(dev_buf, val);
 274 }
 275
 276 static void system_set_disable_ipv6(struct device *dev, const char *val)
 277 {
 278         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 279 }
 280
 281 static void system_set_rpfilter(struct device *dev, const char *val)
 282 {
 283         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter", dev->ifname, val);
 284 }
 285
 286 static void system_set_acceptlocal(struct device *dev, const char *val)
 287 {
 288         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local", dev->ifname, val);
 289 }
 290
 291 static void system_set_igmpversion(struct device *dev, const char *val)
 292 {
 293         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version", dev->ifname, val);
 294 }
 295
 296 static void system_set_mldversion(struct device *dev, const char *val)
 297 {
 298         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version", dev->ifname, val);
 299 }
 300
 301 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 302 {
 303         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 304 }
 305
 306 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 307 {
 308         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms", dev->ifname, val);
 309 }
 310
 311 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 312 {
 313         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time", dev->ifname, val);
 314 }
 315
 316 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 317 {
 318         system_set_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time", dev->ifname, val);
 319 }
 320
 321 static void system_set_neigh4locktime(struct device *dev, const char *val)
 322 {
 323         system_set_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime", dev->ifname, val);
 324 }
 325
 326 static void system_set_dadtransmits(struct device *dev, const char *val)
 327 {
 328         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits", dev->ifname, val);
 329 }
 330
 331 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 332 {
 333         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_to_unicast", dev->ifname, val);
 334 }
 335
 336 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 337 {
 338         system_set_dev_sysctl("/sys/class/net/%s/brport/multicast_fast_leave", dev->ifname, val);
 339 }
 340
 341 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 342 {
 343         system_set_dev_sysctl("/sys/class/net/%s/brport/hairpin_mode", dev->ifname, val);
 344 }
 345
 346 static void system_bridge_set_multicast_router(struct device *dev, const char *val, bool bridge)
 347 {
 348         system_set_dev_sysctl(bridge ? "/sys/class/net/%s/bridge/multicast_router" :
 349                                        "/sys/class/net/%s/brport/multicast_router",
 350                               dev->ifname, val);
 351 }
 352
 353 static void system_bridge_set_robustness(struct device *dev, const char *val)
 354 {
 355         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_count",
 356                               dev->ifname, val);
 357         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_count",
 358                               dev->ifname, val);
 359 }
 360
 361 static void system_bridge_set_query_interval(struct device *dev, const char *val)
 362 {
 363         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_interval",
 364                               dev->ifname, val);
 365 }
 366
 367 static void system_bridge_set_query_response_interval(struct device *dev, const char *val)
 368 {
 369         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_query_response_interval",
 370                               dev->ifname, val);
 371 }
 372
 373 static void system_bridge_set_last_member_interval(struct device *dev, const char *val)
 374 {
 375         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_last_member_interval",
 376                               dev->ifname, val);
 377 }
 378
 379 static void system_bridge_set_membership_interval(struct device *dev, const char *val)
 380 {
 381         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_membership_interval",
 382                               dev->ifname, val);
 383 }
 384
 385 static void system_bridge_set_other_querier_timeout(struct device *dev, const char *val)
 386 {
 387         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier_interval",
 388                               dev->ifname, val);
 389 }
 390
 391 static void system_bridge_set_startup_query_interval(struct device *dev, const char *val)
 392 {
 393         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_startup_query_interval",
 394                               dev->ifname, val);
 395 }
 396
 397 static void system_bridge_set_learning(struct device *dev, const char *val)
 398 {
 399         system_set_dev_sysctl("/sys/class/net/%s/brport/learning", dev->ifname, val);
 400 }
 401
 402 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 403 {
 404         system_set_dev_sysctl("/sys/class/net/%s/brport/unicast_flood", dev->ifname, val);
 405 }
 406
 407 static void system_set_sendredirects(struct device *dev, const char *val)
 408 {
 409         system_set_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects", dev->ifname, val);
 410 }
 411
 412 static int system_get_sysctl(const char *path, char *buf, const size_t buf_sz)
 413 {
 414         int fd = -1, ret = -1;
 415
 416         fd = open(path, O_RDONLY);
 417         if (fd < 0)
 418                 goto out;
 419
 420         ssize_t len = read(fd, buf, buf_sz - 1);
 421         if (len < 0)
 422                 goto out;
 423
 424         ret = buf[len] = 0;
 425
 426 out:
 427         if (fd >= 0)
 428                 close(fd);
 429
 430         return ret;
 431 }
 432
 433 static int
 434 system_get_dev_sysctl(const char *path, const char *device, char *buf, const size_t buf_sz)
 435 {
 436         snprintf(dev_buf, sizeof(dev_buf), path, device);
 437         return system_get_sysctl(dev_buf, buf, buf_sz);
 438 }
 439
 440 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 441 {
 442         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6",
 443                         dev->ifname, buf, buf_sz);
 444 }
 445
 446 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 447 {
 448         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/rp_filter",
 449                         dev->ifname, buf, buf_sz);
 450 }
 451
 452 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 453 {
 454         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/accept_local",
 455                         dev->ifname, buf, buf_sz);
 456 }
 457
 458 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 459 {
 460         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/force_igmp_version",
 461                         dev->ifname, buf, buf_sz);
 462 }
 463
 464 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 465 {
 466         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/force_mld_version",
 467                         dev->ifname, buf, buf_sz);
 468 }
 469
 470 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 471 {
 472         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/base_reachable_time_ms",
 473                         dev->ifname, buf, buf_sz);
 474 }
 475
 476 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 477 {
 478         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/base_reachable_time_ms",
 479                         dev->ifname, buf, buf_sz);
 480 }
 481
 482 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 483 {
 484         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/gc_stale_time",
 485                         dev->ifname, buf, buf_sz);
 486 }
 487
 488 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 489 {
 490         return system_get_dev_sysctl("/proc/sys/net/ipv6/neigh/%s/gc_stale_time",
 491                         dev->ifname, buf, buf_sz);
 492 }
 493
 494 static int system_get_neigh4locktime(struct device *dev, char *buf, const size_t buf_sz)
 495 {
 496         return system_get_dev_sysctl("/proc/sys/net/ipv4/neigh/%s/locktime",
 497                         dev->ifname, buf, buf_sz);
 498 }
 499
 500 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 501 {
 502         return system_get_dev_sysctl("/proc/sys/net/ipv6/conf/%s/dad_transmits",
 503                         dev->ifname, buf, buf_sz);
 504 }
 505
 506 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 507 {
 508         return system_get_dev_sysctl("/proc/sys/net/ipv4/conf/%s/send_redirects",
 509                         dev->ifname, buf, buf_sz);
 510 }
 511
 512 // Evaluate netlink messages
 513 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 514 {
 515         struct nlmsghdr *nh = nlmsg_hdr(msg);
 516         struct nlattr *nla[__IFLA_MAX];
 517         int link_state = 0;
 518         char buf[10];
 519
 520         if (nh->nlmsg_type != RTM_NEWLINK)
 521                 goto out;
 522
 523         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 524         if (!nla[IFLA_IFNAME])
 525                 goto out;
 526
 527         struct device *dev = device_find(nla_data(nla[IFLA_IFNAME]));
 528         if (!dev)
 529                 goto out;
 530
 531         if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 532                 link_state = strtoul(buf, NULL, 0);
 533
 534         device_set_link(dev, link_state ? true : false);
 535
 536 out:
 537         return 0;
 538 }
 539
 540 static void
 541 handle_hotplug_msg(char *data, int size)
 542 {
 543         const char *subsystem = NULL, *interface = NULL;
 544         char *cur, *end, *sep;
 545         struct device *dev;
 546         int skip;
 547         bool add;
 548
 549         if (!strncmp(data, "add@", 4))
 550                 add = true;
 551         else if (!strncmp(data, "remove@", 7))
 552                 add = false;
 553         else
 554                 return;
 555
 556         skip = strlen(data) + 1;
 557         end = data + size;
 558
 559         for (cur = data + skip; cur < end; cur += skip) {
 560                 skip = strlen(cur) + 1;
 561
 562                 sep = strchr(cur, '=');
 563                 if (!sep)
 564                         continue;
 565
 566                 *sep = 0;
 567                 if (!strcmp(cur, "INTERFACE"))
 568                         interface = sep + 1;
 569                 else if (!strcmp(cur, "SUBSYSTEM")) {
 570                         subsystem = sep + 1;
 571                         if (strcmp(subsystem, "net") != 0)
 572                                 return;
 573                 }
 574                 if (subsystem && interface)
 575                         goto found;
 576         }
 577         return;
 578
 579 found:
 580         dev = device_find(interface);
 581         if (!dev)
 582                 return;
 583
 584         if (dev->type != &simple_device_type)
 585                 return;
 586
 587         if (add && system_if_force_external(dev->ifname))
 588                 return;
 589
 590         device_set_present(dev, add);
 591 }
 592
 593 static void
 594 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 595 {
 596         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 597         struct sockaddr_nl nla;
 598         unsigned char *buf = NULL;
 599         int size;
 600
 601         while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 602                 if (nla.nl_pid == 0)
 603                         handle_hotplug_msg((char *) buf, size);
 604
 605                 free(buf);
 606         }
 607 }
 608
 609 static int system_rtnl_call(struct nl_msg *msg)
 610 {
 611         int ret;
 612
 613         ret = nl_send_auto_complete(sock_rtnl, msg);
 614         nlmsg_free(msg);
 615
 616         if (ret < 0)
 617                 return ret;
 618
 619         return nl_wait_for_ack(sock_rtnl);
 620 }
 621
 622 int system_bridge_delbr(struct device *bridge)
 623 {
 624         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 625 }
 626
 627 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 628 {
 629         struct ifreq ifr;
 630
 631         memset(&ifr, 0, sizeof(ifr));
 632         if (dev)
 633                 ifr.ifr_ifindex = dev->ifindex;
 634         else
 635                 ifr.ifr_data = data;
 636         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name));
 637         return ioctl(sock_ioctl, cmd, &ifr);
 638 }
 639
 640 static bool system_is_bridge(const char *name, char *buf, int buflen)
 641 {
 642         struct stat st;
 643
 644         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 645         if (stat(buf, &st) < 0)
 646                 return false;
 647
 648         return true;
 649 }
 650
 651 static char *system_get_bridge(const char *name, char *buf, int buflen)
 652 {
 653         char *path;
 654         ssize_t len = -1;
 655         glob_t gl;
 656
 657         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 658         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 659                 return NULL;
 660
 661         if (gl.gl_pathc > 0)
 662                 len = readlink(gl.gl_pathv[0], buf, buflen);
 663
 664         globfree(&gl);
 665
 666         if (len < 0)
 667                 return NULL;
 668
 669         buf[len] = 0;
 670         path = strrchr(buf, '/');
 671         if (!path)
 672                 return NULL;
 673
 674         return path + 1;
 675 }
 676
 677 static void
 678 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 679 {
 680         bool mcast_to_ucast = dev->wireless_ap;
 681         bool hairpin = true;
 682
 683         if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 684             !bridge->settings.multicast_to_unicast)
 685                 mcast_to_ucast = false;
 686
 687         if (!mcast_to_ucast || dev->wireless_isolate)
 688                 hairpin = false;
 689
 690         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 691         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 692 }
 693
 694 int system_bridge_addif(struct device *bridge, struct device *dev)
 695 {
 696         char buf[64];
 697         char *oldbr;
 698         int ret = 0;
 699
 700         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 701         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0)
 702                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 703
 704         if (dev->wireless)
 705                 system_bridge_set_wireless(bridge, dev);
 706
 707         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 708                 snprintf(buf, sizeof(buf), "%i", dev->settings.multicast_router);
 709                 system_bridge_set_multicast_router(dev, buf, false);
 710         }
 711
 712         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 713             dev->settings.multicast_fast_leave)
 714                 system_bridge_set_multicast_fast_leave(dev, "1");
 715
 716         if (dev->settings.flags & DEV_OPT_LEARNING &&
 717             !dev->settings.learning)
 718                 system_bridge_set_learning(dev, "0");
 719
 720         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 721             !dev->settings.unicast_flood)
 722                 system_bridge_set_unicast_flood(dev, "0");
 723
 724         return ret;
 725 }
 726
 727 int system_bridge_delif(struct device *bridge, struct device *dev)
 728 {
 729         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 730 }
 731
 732 int system_if_resolve(struct device *dev)
 733 {
 734         struct ifreq ifr;
 735         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
 736         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 737                 return ifr.ifr_ifindex;
 738         else
 739                 return 0;
 740 }
 741
 742 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 743 {
 744         struct ifreq ifr;
 745
 746         memset(&ifr, 0, sizeof(ifr));
 747         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name));
 748         ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr);
 749         ifr.ifr_flags |= add;
 750         ifr.ifr_flags &= ~rem;
 751         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 752 }
 753
 754 struct clear_data {
 755         struct nl_msg *msg;
 756         struct device *dev;
 757         int type;
 758         int size;
 759         int af;
 760 };
 761
 762
 763 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 764 {
 765         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 766
 767         return ifa->ifa_index == ifindex;
 768 }
 769
 770 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 771 {
 772         struct rtmsg *r = NLMSG_DATA(hdr);
 773         struct nlattr *tb[__RTA_MAX];
 774
 775         if (r->rtm_protocol == RTPROT_KERNEL &&
 776             r->rtm_family == AF_INET6)
 777                 return false;
 778
 779         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 780         if (!tb[RTA_OIF])
 781                 return false;
 782
 783         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 784 }
 785
 786 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
 787 {
 788         return true;
 789 }
 790
 791 static int cb_clear_event(struct nl_msg *msg, void *arg)
 792 {
 793         struct clear_data *clr = arg;
 794         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 795         bool (*cb)(struct nlmsghdr *, int ifindex);
 796         int type;
 797
 798         switch(clr->type) {
 799         case RTM_GETADDR:
 800                 type = RTM_DELADDR;
 801                 if (hdr->nlmsg_type != RTM_NEWADDR)
 802                         return NL_SKIP;
 803
 804                 cb = check_ifaddr;
 805                 break;
 806         case RTM_GETROUTE:
 807                 type = RTM_DELROUTE;
 808                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 809                         return NL_SKIP;
 810
 811                 cb = check_route;
 812                 break;
 813         case RTM_GETRULE:
 814                 type = RTM_DELRULE;
 815                 if (hdr->nlmsg_type != RTM_NEWRULE)
 816                         return NL_SKIP;
 817
 818                 cb = check_rule;
 819                 break;
 820         default:
 821                 return NL_SKIP;
 822         }
 823
 824         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
 825                 return NL_SKIP;
 826
 827         if (type == RTM_DELRULE)
 828                 D(SYSTEM, "Remove a rule\n");
 829         else
 830                 D(SYSTEM, "Remove %s from device %s\n",
 831                   type == RTM_DELADDR ? "an address" : "a route",
 832                   clr->dev->ifname);
 833         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 834         hdr = nlmsg_hdr(clr->msg);
 835         hdr->nlmsg_type = type;
 836         hdr->nlmsg_flags = NLM_F_REQUEST;
 837
 838         nl_socket_disable_auto_ack(sock_rtnl);
 839         nl_send_auto_complete(sock_rtnl, clr->msg);
 840         nl_socket_enable_auto_ack(sock_rtnl);
 841
 842         return NL_SKIP;
 843 }
 844
 845 static int
 846 cb_finish_event(struct nl_msg *msg, void *arg)
 847 {
 848         int *pending = arg;
 849         *pending = 0;
 850         return NL_STOP;
 851 }
 852
 853 static int
 854 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 855 {
 856         int *pending = arg;
 857         *pending = err->error;
 858         return NL_STOP;
 859 }
 860
 861 static void
 862 system_if_clear_entries(struct device *dev, int type, int af)
 863 {
 864         struct clear_data clr;
 865         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
 866         struct rtmsg rtm = {
 867                 .rtm_family = af,
 868                 .rtm_flags = RTM_F_CLONED,
 869         };
 870         int flags = NLM_F_DUMP;
 871         int pending = 1;
 872
 873         clr.af = af;
 874         clr.dev = dev;
 875         clr.type = type;
 876         switch (type) {
 877         case RTM_GETADDR:
 878         case RTM_GETRULE:
 879                 clr.size = sizeof(struct rtgenmsg);
 880                 break;
 881         case RTM_GETROUTE:
 882                 clr.size = sizeof(struct rtmsg);
 883                 break;
 884         default:
 885                 return;
 886         }
 887
 888         if (!cb)
 889                 return;
 890
 891         clr.msg = nlmsg_alloc_simple(type, flags);
 892         if (!clr.msg)
 893                 goto out;
 894
 895         nlmsg_append(clr.msg, &rtm, clr.size, 0);
 896         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
 897         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
 898         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
 899
 900         nl_send_auto_complete(sock_rtnl, clr.msg);
 901         while (pending > 0)
 902                 nl_recvmsgs(sock_rtnl, cb);
 903
 904         nlmsg_free(clr.msg);
 905 out:
 906         nl_cb_put(cb);
 907 }
 908
 909 /*
 910  * Clear bridge (membership) state and bring down device
 911  */
 912 void system_if_clear_state(struct device *dev)
 913 {
 914         static char buf[256];
 915         char *bridge;
 916
 917         device_set_ifindex(dev, system_if_resolve(dev));
 918         if (dev->external || !dev->ifindex)
 919                 return;
 920
 921         system_if_flags(dev->ifname, 0, IFF_UP);
 922
 923         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
 924                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
 925                 system_bridge_delbr(dev);
 926                 return;
 927         }
 928
 929         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
 930         if (bridge) {
 931                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
 932                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
 933         }
 934
 935         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
 936         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
 937         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
 938         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
 939         system_set_disable_ipv6(dev, "0");
 940 }
 941
 942 static inline unsigned long
 943 sec_to_jiffies(int val)
 944 {
 945         return (unsigned long) val * 100;
 946 }
 947
 948 static void system_bridge_conf_multicast_deps(struct device *bridge,
 949                                               struct bridge_config *cfg,
 950                                               char *buf,
 951                                               int buf_len)
 952 {
 953         int val;
 954
 955         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
 956             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
 957             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
 958                 val = cfg->robustness * cfg->query_interval +
 959                         cfg->query_response_interval;
 960
 961                 snprintf(buf, buf_len, "%i", val);
 962                 system_bridge_set_membership_interval(bridge, buf);
 963
 964                 val = cfg->robustness * cfg->query_interval +
 965                         cfg->query_response_interval / 2;
 966
 967                 snprintf(buf, buf_len, "%i", val);
 968                 system_bridge_set_other_querier_timeout(bridge, buf);
 969         }
 970
 971         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
 972                 val = cfg->query_interval / 4;
 973
 974                 snprintf(buf, buf_len, "%i", val);
 975                 system_bridge_set_startup_query_interval(bridge, buf);
 976         }
 977 }
 978
 979 static void system_bridge_conf_multicast(struct device *bridge,
 980                                          struct bridge_config *cfg,
 981                                          char *buf,
 982                                          int buf_len)
 983 {
 984         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
 985                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
 986
 987         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_querier",
 988                 bridge->ifname, cfg->multicast_querier ? "1" : "0");
 989
 990         snprintf(buf, buf_len, "%i", cfg->hash_max);
 991         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/hash_max",
 992                 bridge->ifname, buf);
 993
 994         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 995                 snprintf(buf, buf_len, "%i", bridge->settings.multicast_router);
 996                 system_bridge_set_multicast_router(bridge, buf, true);
 997         }
 998
 999         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
1000                 snprintf(buf, buf_len, "%i", cfg->robustness);
1001                 system_bridge_set_robustness(bridge, buf);
1002         }
1003
1004         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1005                 snprintf(buf, buf_len, "%i", cfg->query_interval);
1006                 system_bridge_set_query_interval(bridge, buf);
1007         }
1008
1009         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1010                 snprintf(buf, buf_len, "%i", cfg->query_response_interval);
1011                 system_bridge_set_query_response_interval(bridge, buf);
1012         }
1013
1014         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL) {
1015                 snprintf(buf, buf_len, "%i", cfg->last_member_interval);
1016                 system_bridge_set_last_member_interval(bridge, buf);
1017         }
1018
1019         system_bridge_conf_multicast_deps(bridge, cfg, buf, buf_len);
1020 }
1021
1022 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1023 {
1024         char buf[64];
1025         unsigned long args[4] = {};
1026
1027         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
1028                 return -1;
1029
1030         args[0] = BRCTL_SET_BRIDGE_STP_STATE;
1031         args[1] = !!cfg->stp;
1032         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1033
1034         args[0] = BRCTL_SET_BRIDGE_FORWARD_DELAY;
1035         args[1] = sec_to_jiffies(cfg->forward_delay);
1036         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1037
1038         system_bridge_conf_multicast(bridge, cfg, buf, sizeof(buf));
1039
1040         args[0] = BRCTL_SET_BRIDGE_PRIORITY;
1041         args[1] = cfg->priority;
1042         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1043
1044         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
1045                 args[0] = BRCTL_SET_AGEING_TIME;
1046                 args[1] = sec_to_jiffies(cfg->ageing_time);
1047                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1048         }
1049
1050         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
1051                 args[0] = BRCTL_SET_BRIDGE_HELLO_TIME;
1052                 args[1] = sec_to_jiffies(cfg->hello_time);
1053                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1054         }
1055
1056         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
1057                 args[0] = BRCTL_SET_BRIDGE_MAX_AGE;
1058                 args[1] = sec_to_jiffies(cfg->max_age);
1059                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
1060         }
1061
1062         return 0;
1063 }
1064
1065 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1066 {
1067         struct nl_msg *msg;
1068         struct nlattr *linkinfo, *data;
1069         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
1070         int i, rv;
1071         static const struct {
1072                 const char *name;
1073                 enum macvlan_mode val;
1074         } modes[] = {
1075                 { "private", MACVLAN_MODE_PRIVATE },
1076                 { "vepa", MACVLAN_MODE_VEPA },
1077                 { "bridge", MACVLAN_MODE_BRIDGE },
1078                 { "passthru", MACVLAN_MODE_PASSTHRU },
1079         };
1080
1081         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1082
1083         if (!msg)
1084                 return -1;
1085
1086         nlmsg_append(msg, &iim, sizeof(iim), 0);
1087
1088         if (cfg->flags & MACVLAN_OPT_MACADDR)
1089                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1090         nla_put_string(msg, IFLA_IFNAME, macvlan->ifname);
1091         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1092
1093         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1094                 goto nla_put_failure;
1095
1096         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1097
1098         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1099                 goto nla_put_failure;
1100
1101         if (cfg->mode) {
1102                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1103                         if (strcmp(cfg->mode, modes[i].name) != 0)
1104                                 continue;
1105
1106                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1107                         break;
1108                 }
1109         }
1110
1111         nla_nest_end(msg, data);
1112         nla_nest_end(msg, linkinfo);
1113
1114         rv = system_rtnl_call(msg);
1115         if (rv)
1116                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1117
1118         return rv;
1119
1120 nla_put_failure:
1121         nlmsg_free(msg);
1122         return -ENOMEM;
1123 }
1124
1125 static int system_link_del(const char *ifname)
1126 {
1127         struct nl_msg *msg;
1128         struct ifinfomsg iim = {
1129                 .ifi_family = AF_UNSPEC,
1130                 .ifi_index = 0,
1131         };
1132
1133         msg = nlmsg_alloc_simple(RTM_DELLINK, NLM_F_REQUEST);
1134
1135         if (!msg)
1136                 return -1;
1137
1138         nlmsg_append(msg, &iim, sizeof(iim), 0);
1139         nla_put_string(msg, IFLA_IFNAME, ifname);
1140         return system_rtnl_call(msg);
1141 }
1142
1143 int system_macvlan_del(struct device *macvlan)
1144 {
1145         return system_link_del(macvlan->ifname);
1146 }
1147
1148 int system_veth_add(struct device *veth, struct veth_config *cfg)
1149 {
1150         struct nl_msg *msg;
1151         struct ifinfomsg empty_iim = {};
1152         struct nlattr *linkinfo, *data, *veth_info;
1153         int rv;
1154
1155         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1156
1157         if (!msg)
1158                 return -1;
1159
1160         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1161
1162         if (cfg->flags & VETH_OPT_MACADDR)
1163                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1164         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1165
1166         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1167                 goto nla_put_failure;
1168
1169         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1170
1171         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1172                 goto nla_put_failure;
1173
1174         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1175                 goto nla_put_failure;
1176
1177         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1178
1179         if (cfg->flags & VETH_OPT_PEER_NAME)
1180                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1181         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1182                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1183
1184         nla_nest_end(msg, veth_info);
1185         nla_nest_end(msg, data);
1186         nla_nest_end(msg, linkinfo);
1187
1188         rv = system_rtnl_call(msg);
1189         if (rv) {
1190                 if (cfg->flags & VETH_OPT_PEER_NAME)
1191                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1192                 else
1193                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1194         }
1195
1196         return rv;
1197
1198 nla_put_failure:
1199         nlmsg_free(msg);
1200         return -ENOMEM;
1201 }
1202
1203 int system_veth_del(struct device *veth)
1204 {
1205         return system_link_del(veth->ifname);
1206 }
1207
1208 static int system_vlan(struct device *dev, int id)
1209 {
1210         struct vlan_ioctl_args ifr = {
1211                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1212                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1213         };
1214
1215         ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1216
1217         if (id < 0) {
1218                 ifr.cmd = DEL_VLAN_CMD;
1219                 ifr.u.VID = 0;
1220         } else {
1221                 ifr.cmd = ADD_VLAN_CMD;
1222                 ifr.u.VID = id;
1223         }
1224         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1225         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1226 }
1227
1228 int system_vlan_add(struct device *dev, int id)
1229 {
1230         return system_vlan(dev, id);
1231 }
1232
1233 int system_vlan_del(struct device *dev)
1234 {
1235         return system_vlan(dev, -1);
1236 }
1237
1238 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1239 {
1240         struct nl_msg *msg;
1241         struct nlattr *linkinfo, *data;
1242         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1243         int rv;
1244
1245         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1246
1247         if (!msg)
1248                 return -1;
1249
1250         nlmsg_append(msg, &iim, sizeof(iim), 0);
1251         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1252         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1253
1254         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1255                 goto nla_put_failure;
1256
1257         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1258
1259         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1260                 goto nla_put_failure;
1261
1262         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1263
1264 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1265         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1266 #else
1267         if(cfg->proto == VLAN_PROTO_8021AD)
1268                 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);
1269 #endif
1270
1271         nla_nest_end(msg, data);
1272         nla_nest_end(msg, linkinfo);
1273
1274         rv = system_rtnl_call(msg);
1275         if (rv)
1276                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1277
1278         return rv;
1279
1280 nla_put_failure:
1281         nlmsg_free(msg);
1282         return -ENOMEM;
1283 }
1284
1285 int system_vlandev_del(struct device *vlandev)
1286 {
1287         return system_link_del(vlandev->ifname);
1288 }
1289
1290 void
1291 system_if_get_settings(struct device *dev, struct device_settings *s)
1292 {
1293         struct ifreq ifr;
1294         char buf[10];
1295
1296         memset(&ifr, 0, sizeof(ifr));
1297         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1298
1299         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
1300                 s->mtu = ifr.ifr_mtu;
1301                 s->flags |= DEV_OPT_MTU;
1302         }
1303
1304         s->mtu6 = system_update_ipv6_mtu(dev, 0);
1305         if (s->mtu6 > 0)
1306                 s->flags |= DEV_OPT_MTU6;
1307
1308         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
1309                 s->txqueuelen = ifr.ifr_qlen;
1310                 s->flags |= DEV_OPT_TXQUEUELEN;
1311         }
1312
1313         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
1314                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
1315                 s->flags |= DEV_OPT_MACADDR;
1316         }
1317
1318         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
1319                 s->ipv6 = !strtoul(buf, NULL, 0);
1320                 s->flags |= DEV_OPT_IPV6;
1321         }
1322
1323         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
1324                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
1325                 s->flags |= DEV_OPT_PROMISC;
1326
1327                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
1328                 s->flags |= DEV_OPT_MULTICAST;
1329         }
1330
1331         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
1332                 s->rpfilter = strtoul(buf, NULL, 0);
1333                 s->flags |= DEV_OPT_RPFILTER;
1334         }
1335
1336         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
1337                 s->acceptlocal = strtoul(buf, NULL, 0);
1338                 s->flags |= DEV_OPT_ACCEPTLOCAL;
1339         }
1340
1341         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
1342                 s->igmpversion = strtoul(buf, NULL, 0);
1343                 s->flags |= DEV_OPT_IGMPVERSION;
1344         }
1345
1346         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
1347                 s->mldversion = strtoul(buf, NULL, 0);
1348                 s->flags |= DEV_OPT_MLDVERSION;
1349         }
1350
1351         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
1352                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
1353                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1354         }
1355
1356         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
1357                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
1358                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
1359         }
1360
1361         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
1362                 s->neigh4locktime = strtoul(buf, NULL, 0);
1363                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
1364         }
1365
1366         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
1367                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
1368                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1369         }
1370
1371         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
1372                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
1373                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
1374         }
1375
1376         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
1377                 s->dadtransmits = strtoul(buf, NULL, 0);
1378                 s->flags |= DEV_OPT_DADTRANSMITS;
1379         }
1380
1381         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
1382                 s->sendredirects = strtoul(buf, NULL, 0);
1383                 s->flags |= DEV_OPT_SENDREDIRECTS;
1384         }
1385 }
1386
1387 static void
1388 system_if_set_rps_xps_val(const char *path, int val)
1389 {
1390         char val_buf[8];
1391         glob_t gl;
1392         int i;
1393
1394         if (glob(path, 0, NULL, &gl))
1395                 return;
1396
1397         snprintf(val_buf, sizeof(val_buf), "%x", val);
1398         for (i = 0; i < gl.gl_pathc; i++)
1399                 system_set_sysctl(gl.gl_pathv[i], val_buf);
1400
1401         globfree(&gl);
1402 }
1403
1404 static void
1405 system_if_apply_rps_xps(struct device *dev, struct device_settings *s)
1406 {
1407         long n_cpus = sysconf(_SC_NPROCESSORS_ONLN);
1408         int val;
1409
1410         if (n_cpus < 2)
1411                 return;
1412
1413         val = (1 << n_cpus) - 1;
1414         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/rps_cpus", dev->ifname);
1415         system_if_set_rps_xps_val(dev_buf, s->rps ? val : 0);
1416
1417         snprintf(dev_buf, sizeof(dev_buf), "/sys/class/net/%s/queues/*/xps_cpus", dev->ifname);
1418         system_if_set_rps_xps_val(dev_buf, s->xps ? val : 0);
1419 }
1420
1421 void
1422 system_if_apply_settings(struct device *dev, struct device_settings *s, unsigned int apply_mask)
1423 {
1424         struct ifreq ifr;
1425         char buf[12];
1426
1427         memset(&ifr, 0, sizeof(ifr));
1428         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
1429         if (s->flags & DEV_OPT_MTU & apply_mask) {
1430                 ifr.ifr_mtu = s->mtu;
1431                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
1432                         s->flags &= ~DEV_OPT_MTU;
1433         }
1434         if (s->flags & DEV_OPT_MTU6 & apply_mask) {
1435                 system_update_ipv6_mtu(dev, s->mtu6);
1436         }
1437         if (s->flags & DEV_OPT_TXQUEUELEN & apply_mask) {
1438                 ifr.ifr_qlen = s->txqueuelen;
1439                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
1440                         s->flags &= ~DEV_OPT_TXQUEUELEN;
1441         }
1442         if ((s->flags & DEV_OPT_MACADDR & apply_mask) && !dev->external) {
1443                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
1444                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
1445                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
1446                         s->flags &= ~DEV_OPT_MACADDR;
1447         }
1448         if (s->flags & DEV_OPT_IPV6 & apply_mask)
1449                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
1450         if (s->flags & DEV_OPT_PROMISC & apply_mask) {
1451                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
1452                                     !s->promisc ? IFF_PROMISC : 0) < 0)
1453                         s->flags &= ~DEV_OPT_PROMISC;
1454         }
1455         if (s->flags & DEV_OPT_RPFILTER & apply_mask) {
1456                 snprintf(buf, sizeof(buf), "%d", s->rpfilter);
1457                 system_set_rpfilter(dev, buf);
1458         }
1459         if (s->flags & DEV_OPT_ACCEPTLOCAL & apply_mask)
1460                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
1461         if (s->flags & DEV_OPT_IGMPVERSION & apply_mask) {
1462                 snprintf(buf, sizeof(buf), "%d", s->igmpversion);
1463                 system_set_igmpversion(dev, buf);
1464         }
1465         if (s->flags & DEV_OPT_MLDVERSION & apply_mask) {
1466                 snprintf(buf, sizeof(buf), "%d", s->mldversion);
1467                 system_set_mldversion(dev, buf);
1468         }
1469         if (s->flags & DEV_OPT_NEIGHREACHABLETIME & apply_mask) {
1470                 snprintf(buf, sizeof(buf), "%d", s->neigh4reachabletime);
1471                 system_set_neigh4reachabletime(dev, buf);
1472                 snprintf(buf, sizeof(buf), "%d", s->neigh6reachabletime);
1473                 system_set_neigh6reachabletime(dev, buf);
1474         }
1475         if (s->flags & DEV_OPT_NEIGHLOCKTIME & apply_mask) {
1476                 snprintf(buf, sizeof(buf), "%d", s->neigh4locktime);
1477                 system_set_neigh4locktime(dev, buf);
1478         }
1479         if (s->flags & DEV_OPT_NEIGHGCSTALETIME & apply_mask) {
1480                 snprintf(buf, sizeof(buf), "%d", s->neigh4gcstaletime);
1481                 system_set_neigh4gcstaletime(dev, buf);
1482                 snprintf(buf, sizeof(buf), "%d", s->neigh6gcstaletime);
1483                 system_set_neigh6gcstaletime(dev, buf);
1484         }
1485         if (s->flags & DEV_OPT_DADTRANSMITS & apply_mask) {
1486                 snprintf(buf, sizeof(buf), "%d", s->dadtransmits);
1487                 system_set_dadtransmits(dev, buf);
1488         }
1489         if (s->flags & DEV_OPT_MULTICAST & apply_mask) {
1490                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
1491                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
1492                         s->flags &= ~DEV_OPT_MULTICAST;
1493         }
1494         if (s->flags & DEV_OPT_SENDREDIRECTS & apply_mask)
1495                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
1496
1497         system_if_apply_rps_xps(dev, s);
1498 }
1499
1500 int system_if_up(struct device *dev)
1501 {
1502         system_if_get_settings(dev, &dev->orig_settings);
1503         /* Only keep orig settings based on what needs to be set */
1504         dev->orig_settings.valid_flags = dev->orig_settings.flags;
1505         dev->orig_settings.flags &= dev->settings.flags;
1506         system_if_apply_settings(dev, &dev->settings, dev->settings.flags);
1507         return system_if_flags(dev->ifname, IFF_UP, 0);
1508 }
1509
1510 int system_if_down(struct device *dev)
1511 {
1512         int ret = system_if_flags(dev->ifname, 0, IFF_UP);
1513         system_if_apply_settings(dev, &dev->orig_settings, dev->orig_settings.flags);
1514         return ret;
1515 }
1516
1517 struct if_check_data {
1518         struct device *dev;
1519         int pending;
1520         int ret;
1521 };
1522
1523 #ifndef IFF_LOWER_UP
1524 #define IFF_LOWER_UP    0x10000
1525 #endif
1526
1527 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
1528 {
1529         struct nlmsghdr *nh = nlmsg_hdr(msg);
1530         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1531         struct if_check_data *chk = (struct if_check_data *)arg;
1532
1533         if (nh->nlmsg_type != RTM_NEWLINK)
1534                 return NL_SKIP;
1535
1536         device_set_present(chk->dev, ifi->ifi_index > 0 ? true : false);
1537         device_set_link(chk->dev, ifi->ifi_flags & IFF_LOWER_UP ? true : false);
1538
1539         return NL_OK;
1540 }
1541
1542 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
1543 {
1544         struct if_check_data *chk = (struct if_check_data *)arg;
1545         chk->pending = 0;
1546         return NL_STOP;
1547 }
1548
1549 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1550 {
1551         struct if_check_data *chk = (struct if_check_data *)arg;
1552
1553         device_set_present(chk->dev, false);
1554         device_set_link(chk->dev, false);
1555         chk->pending = err->error;
1556
1557         return NL_STOP;
1558 }
1559
1560 int system_if_check(struct device *dev)
1561 {
1562         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1563         struct nl_msg *msg;
1564         struct ifinfomsg ifi = {
1565                 .ifi_family = AF_UNSPEC,
1566                 .ifi_index = 0,
1567         };
1568         struct if_check_data chk = {
1569                 .dev = dev,
1570                 .pending = 1,
1571         };
1572         int ret = 1;
1573
1574         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
1575         if (!msg)
1576                 goto out;
1577
1578         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1579             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1580                 goto free;
1581
1582         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
1583         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
1584         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
1585
1586         nl_send_auto_complete(sock_rtnl, msg);
1587         while (chk.pending > 0)
1588                 nl_recvmsgs(sock_rtnl, cb);
1589
1590         ret = chk.pending;
1591
1592 free:
1593         nlmsg_free(msg);
1594 out:
1595         nl_cb_put(cb);
1596         return ret;
1597 }
1598
1599 struct device *
1600 system_if_get_parent(struct device *dev)
1601 {
1602         char buf[64], *devname;
1603         int ifindex, iflink, len;
1604         FILE *f;
1605
1606         snprintf(buf, sizeof(buf), "/sys/class/net/%s/iflink", dev->ifname);
1607         f = fopen(buf, "r");
1608         if (!f)
1609                 return NULL;
1610
1611         len = fread(buf, 1, sizeof(buf) - 1, f);
1612         fclose(f);
1613
1614         if (len <= 0)
1615                 return NULL;
1616
1617         buf[len] = 0;
1618         iflink = strtoul(buf, NULL, 0);
1619         ifindex = system_if_resolve(dev);
1620         if (!iflink || iflink == ifindex)
1621                 return NULL;
1622
1623         devname = if_indextoname(iflink, buf);
1624         if (!devname)
1625                 return NULL;
1626
1627         return device_get(devname, true);
1628 }
1629
1630 static bool
1631 read_string_file(int dir_fd, const char *file, char *buf, int len)
1632 {
1633         bool ret = false;
1634         char *c;
1635         int fd;
1636
1637         fd = openat(dir_fd, file, O_RDONLY);
1638         if (fd < 0)
1639                 return false;
1640
1641 retry:
1642         len = read(fd, buf, len - 1);
1643         if (len < 0) {
1644                 if (errno == EINTR)
1645                         goto retry;
1646         } else if (len > 0) {
1647                         buf[len] = 0;
1648
1649                         c = strchr(buf, '\n');
1650                         if (c)
1651                                 *c = 0;
1652
1653                         ret = true;
1654         }
1655
1656         close(fd);
1657
1658         return ret;
1659 }
1660
1661 static bool
1662 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
1663 {
1664         char buf[64];
1665         bool ret = false;
1666
1667         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
1668         if (ret)
1669                 *val = strtoull(buf, NULL, 0);
1670
1671         return ret;
1672 }
1673
1674 /* Assume advertised flags == supported flags */
1675 static const struct {
1676         uint32_t mask;
1677         const char *name;
1678 } ethtool_link_modes[] = {
1679         { ADVERTISED_10baseT_Half, "10H" },
1680         { ADVERTISED_10baseT_Full, "10F" },
1681         { ADVERTISED_100baseT_Half, "100H" },
1682         { ADVERTISED_100baseT_Full, "100F" },
1683         { ADVERTISED_1000baseT_Half, "1000H" },
1684         { ADVERTISED_1000baseT_Full, "1000F" },
1685 };
1686
1687 static void system_add_link_modes(struct blob_buf *b, __u32 mask)
1688 {
1689         int i;
1690         for (i = 0; i < ARRAY_SIZE(ethtool_link_modes); i++) {
1691                 if (mask & ethtool_link_modes[i].mask)
1692                         blobmsg_add_string(b, NULL, ethtool_link_modes[i].name);
1693         }
1694 }
1695
1696 bool
1697 system_if_force_external(const char *ifname)
1698 {
1699         char buf[64];
1700         struct stat s;
1701
1702         snprintf(buf, sizeof(buf), "/sys/class/net/%s/phy80211", ifname);
1703         return stat(buf, &s) == 0;
1704 }
1705
1706 int
1707 system_if_dump_info(struct device *dev, struct blob_buf *b)
1708 {
1709         struct ethtool_cmd ecmd;
1710         struct ifreq ifr;
1711         char buf[64], *s;
1712         void *c;
1713         int dir_fd;
1714
1715         snprintf(buf, sizeof(buf), "/sys/class/net/%s", dev->ifname);
1716         dir_fd = open(buf, O_DIRECTORY);
1717
1718         memset(&ecmd, 0, sizeof(ecmd));
1719         memset(&ifr, 0, sizeof(ifr));
1720         strcpy(ifr.ifr_name, dev->ifname);
1721         ifr.ifr_data = (caddr_t) &ecmd;
1722         ecmd.cmd = ETHTOOL_GSET;
1723
1724         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == 0) {
1725                 c = blobmsg_open_array(b, "link-advertising");
1726                 system_add_link_modes(b, ecmd.advertising);
1727                 blobmsg_close_array(b, c);
1728
1729                 c = blobmsg_open_array(b, "link-supported");
1730                 system_add_link_modes(b, ecmd.supported);
1731                 blobmsg_close_array(b, c);
1732
1733                 s = blobmsg_alloc_string_buffer(b, "speed", 8);
1734                 snprintf(s, 8, "%d%c", ethtool_cmd_speed(&ecmd),
1735                         ecmd.duplex == DUPLEX_HALF ? 'H' : 'F');
1736                 blobmsg_add_string_buffer(b);
1737         }
1738
1739         close(dir_fd);
1740         return 0;
1741 }
1742
1743 int
1744 system_if_dump_stats(struct device *dev, struct blob_buf *b)
1745 {
1746         const char *const counters[] = {
1747                 "collisions",     "rx_frame_errors",   "tx_compressed",
1748                 "multicast",      "rx_length_errors",  "tx_dropped",
1749                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
1750                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
1751                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
1752                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
1753                 "rx_errors",      "tx_bytes",          "tx_window_errors",
1754                 "rx_fifo_errors", "tx_carrier_errors",
1755         };
1756         char buf[64];
1757         int stats_dir;
1758         int i;
1759         uint64_t val = 0;
1760
1761         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
1762         stats_dir = open(buf, O_DIRECTORY);
1763         if (stats_dir < 0)
1764                 return -1;
1765
1766         for (i = 0; i < ARRAY_SIZE(counters); i++)
1767                 if (read_uint64_file(stats_dir, counters[i], &val))
1768                         blobmsg_add_u64(b, counters[i], val);
1769
1770         close(stats_dir);
1771         return 0;
1772 }
1773
1774 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
1775 {
1776         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
1777         int alen = v4 ? 4 : 16;
1778         unsigned int flags = 0;
1779         struct ifaddrmsg ifa = {
1780                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
1781                 .ifa_prefixlen = addr->mask,
1782                 .ifa_index = dev->ifindex,
1783         };
1784
1785         struct nl_msg *msg;
1786         if (cmd == RTM_NEWADDR)
1787                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1788
1789         msg = nlmsg_alloc_simple(cmd, flags);
1790         if (!msg)
1791                 return -1;
1792
1793         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
1794         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
1795         if (v4) {
1796                 if (addr->broadcast)
1797                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
1798                 if (addr->point_to_point)
1799                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
1800         } else {
1801                 time_t now = system_get_rtime();
1802                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
1803
1804                 if (addr->preferred_until) {
1805                         int64_t preferred = addr->preferred_until - now;
1806                         if (preferred < 0)
1807                                 preferred = 0;
1808                         else if (preferred > UINT32_MAX)
1809                                 preferred = UINT32_MAX;
1810
1811                         cinfo.ifa_prefered = preferred;
1812                 }
1813
1814                 if (addr->valid_until) {
1815                         int64_t valid = addr->valid_until - now;
1816                         if (valid <= 0) {
1817                                 nlmsg_free(msg);
1818                                 return -1;
1819                         }
1820                         else if (valid > UINT32_MAX)
1821                                 valid = UINT32_MAX;
1822
1823                         cinfo.ifa_valid = valid;
1824                 }
1825
1826                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
1827
1828                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
1829                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
1830         }
1831
1832         return system_rtnl_call(msg);
1833 }
1834
1835 int system_add_address(struct device *dev, struct device_addr *addr)
1836 {
1837         return system_addr(dev, addr, RTM_NEWADDR);
1838 }
1839
1840 int system_del_address(struct device *dev, struct device_addr *addr)
1841 {
1842         return system_addr(dev, addr, RTM_DELADDR);
1843 }
1844
1845 static int system_rt(struct device *dev, struct device_route *route, int cmd)
1846 {
1847         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
1848         bool have_gw;
1849         unsigned int flags = 0;
1850
1851         if (alen == 4)
1852                 have_gw = !!route->nexthop.in.s_addr;
1853         else
1854                 have_gw = route->nexthop.in6.s6_addr32[0] ||
1855                         route->nexthop.in6.s6_addr32[1] ||
1856                         route->nexthop.in6.s6_addr32[2] ||
1857                         route->nexthop.in6.s6_addr32[3];
1858
1859         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
1860                         ? route->table : RT_TABLE_MAIN;
1861
1862         struct rtmsg rtm = {
1863                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
1864                 .rtm_dst_len = route->mask,
1865                 .rtm_src_len = route->sourcemask,
1866                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
1867                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
1868                 .rtm_scope = RT_SCOPE_NOWHERE,
1869                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
1870                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
1871         };
1872         struct nl_msg *msg;
1873
1874         if (cmd == RTM_NEWROUTE) {
1875                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
1876
1877                 if (!dev) { // Add null-route
1878                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1879                         rtm.rtm_type = RTN_UNREACHABLE;
1880                 }
1881                 else
1882                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
1883         }
1884
1885         if (route->flags & DEVROUTE_TYPE) {
1886                 rtm.rtm_type = route->type;
1887                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
1888                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
1889                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
1890                                 rtm.rtm_table = RT_TABLE_LOCAL;
1891                 }
1892
1893                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
1894                         rtm.rtm_scope = RT_SCOPE_HOST;
1895                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
1896                                 rtm.rtm_type == RTN_ANYCAST) {
1897                         rtm.rtm_scope = RT_SCOPE_LINK;
1898                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
1899                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY ||
1900                                 rtm.rtm_type == RTN_THROW) {
1901                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
1902                         dev = NULL;
1903                 }
1904         }
1905
1906         msg = nlmsg_alloc_simple(cmd, flags);
1907         if (!msg)
1908                 return -1;
1909
1910         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
1911
1912         if (route->mask)
1913                 nla_put(msg, RTA_DST, alen, &route->addr);
1914
1915         if (route->sourcemask) {
1916                 if (rtm.rtm_family == AF_INET)
1917                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
1918                 else
1919                         nla_put(msg, RTA_SRC, alen, &route->source);
1920         }
1921
1922         if (route->metric > 0)
1923                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
1924
1925         if (have_gw)
1926                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
1927
1928         if (dev)
1929                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
1930
1931         if (table >= 256)
1932                 nla_put_u32(msg, RTA_TABLE, table);
1933
1934         if (route->flags & DEVROUTE_MTU) {
1935                 struct nlattr *metrics;
1936
1937                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
1938                         goto nla_put_failure;
1939
1940                 nla_put_u32(msg, RTAX_MTU, route->mtu);
1941
1942                 nla_nest_end(msg, metrics);
1943         }
1944
1945         return system_rtnl_call(msg);
1946
1947 nla_put_failure:
1948         nlmsg_free(msg);
1949         return -ENOMEM;
1950 }
1951
1952 int system_add_route(struct device *dev, struct device_route *route)
1953 {
1954         return system_rt(dev, route, RTM_NEWROUTE);
1955 }
1956
1957 int system_del_route(struct device *dev, struct device_route *route)
1958 {
1959         return system_rt(dev, route, RTM_DELROUTE);
1960 }
1961
1962 int system_flush_routes(void)
1963 {
1964         const char *names[] = {
1965                 "/proc/sys/net/ipv4/route/flush",
1966                 "/proc/sys/net/ipv6/route/flush"
1967         };
1968         int fd, i;
1969
1970         for (i = 0; i < ARRAY_SIZE(names); i++) {
1971                 fd = open(names[i], O_WRONLY);
1972                 if (fd < 0)
1973                         continue;
1974
1975                 if (write(fd, "-1", 2)) {}
1976                 close(fd);
1977         }
1978         return 0;
1979 }
1980
1981 bool system_resolve_rt_type(const char *type, unsigned int *id)
1982 {
1983         return system_rtn_aton(type, id);
1984 }
1985
1986 bool system_resolve_rt_proto(const char *type, unsigned int *id)
1987 {
1988         FILE *f;
1989         char *e, buf[128];
1990         unsigned int n, proto = 256;
1991
1992         if ((n = strtoul(type, &e, 0)) >= 0 && !*e && e != type)
1993                 proto = n;
1994         else if (!strcmp(type, "unspec"))
1995                 proto = RTPROT_UNSPEC;
1996         else if (!strcmp(type, "kernel"))
1997                 proto = RTPROT_KERNEL;
1998         else if (!strcmp(type, "boot"))
1999                 proto = RTPROT_BOOT;
2000         else if (!strcmp(type, "static"))
2001                 proto = RTPROT_STATIC;
2002         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
2003                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
2004                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2005                                 continue;
2006
2007                         n = strtoul(e, NULL, 10);
2008                         e = strtok(NULL, " \t\n");
2009
2010                         if (e && !strcmp(e, type)) {
2011                                 proto = n;
2012                                 break;
2013                         }
2014                 }
2015                 fclose(f);
2016         }
2017
2018         if (proto > 255)
2019                 return false;
2020
2021         *id = proto;
2022         return true;
2023 }
2024
2025 bool system_resolve_rt_table(const char *name, unsigned int *id)
2026 {
2027         FILE *f;
2028         char *e, buf[128];
2029         unsigned int n, table = RT_TABLE_UNSPEC;
2030
2031         /* first try to parse table as number */
2032         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
2033                 table = n;
2034
2035         /* handle well known aliases */
2036         else if (!strcmp(name, "default"))
2037                 table = RT_TABLE_DEFAULT;
2038         else if (!strcmp(name, "main"))
2039                 table = RT_TABLE_MAIN;
2040         else if (!strcmp(name, "local"))
2041                 table = RT_TABLE_LOCAL;
2042
2043         /* try to look up name in /etc/iproute2/rt_tables */
2044         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
2045         {
2046                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
2047                 {
2048                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
2049                                 continue;
2050
2051                         n = strtoul(e, NULL, 10);
2052                         e = strtok(NULL, " \t\n");
2053
2054                         if (e && !strcmp(e, name))
2055                         {
2056                                 table = n;
2057                                 break;
2058                         }
2059                 }
2060
2061                 fclose(f);
2062         }
2063
2064         if (table == RT_TABLE_UNSPEC)
2065                 return false;
2066
2067         *id = table;
2068         return true;
2069 }
2070
2071 bool system_is_default_rt_table(unsigned int id)
2072 {
2073         return (id == RT_TABLE_MAIN);
2074 }
2075
2076 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
2077 {
2078         char *e;
2079         unsigned int n;
2080
2081         if (!strcmp(filter, "strict"))
2082                 n = 1;
2083         else if (!strcmp(filter, "loose"))
2084                 n = 2;
2085         else {
2086                 n = strtoul(filter, &e, 0);
2087                 if (*e || e == filter || n > 2)
2088                         return false;
2089         }
2090
2091         *id = n;
2092         return true;
2093 }
2094
2095 static int system_iprule(struct iprule *rule, int cmd)
2096 {
2097         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
2098
2099         struct nl_msg *msg;
2100         struct rtmsg rtm = {
2101                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
2102                 .rtm_protocol = RTPROT_STATIC,
2103                 .rtm_scope = RT_SCOPE_UNIVERSE,
2104                 .rtm_table = RT_TABLE_UNSPEC,
2105                 .rtm_type = RTN_UNSPEC,
2106                 .rtm_flags = 0,
2107         };
2108
2109         if (cmd == RTM_NEWRULE)
2110                 rtm.rtm_type = RTN_UNICAST;
2111
2112         if (rule->invert)
2113                 rtm.rtm_flags |= FIB_RULE_INVERT;
2114
2115         if (rule->flags & IPRULE_SRC)
2116                 rtm.rtm_src_len = rule->src_mask;
2117
2118         if (rule->flags & IPRULE_DEST)
2119                 rtm.rtm_dst_len = rule->dest_mask;
2120
2121         if (rule->flags & IPRULE_TOS)
2122                 rtm.rtm_tos = rule->tos;
2123
2124         if (rule->flags & IPRULE_LOOKUP) {
2125                 if (rule->lookup < 256)
2126                         rtm.rtm_table = rule->lookup;
2127         }
2128
2129         if (rule->flags & IPRULE_ACTION)
2130                 rtm.rtm_type = rule->action;
2131         else if (rule->flags & IPRULE_GOTO)
2132                 rtm.rtm_type = FR_ACT_GOTO;
2133         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
2134                 rtm.rtm_type = FR_ACT_NOP;
2135
2136         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
2137
2138         if (!msg)
2139                 return -1;
2140
2141         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
2142
2143         if (rule->flags & IPRULE_IN)
2144                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
2145
2146         if (rule->flags & IPRULE_OUT)
2147                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
2148
2149         if (rule->flags & IPRULE_SRC)
2150                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
2151
2152         if (rule->flags & IPRULE_DEST)
2153                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
2154
2155         if (rule->flags & IPRULE_PRIORITY)
2156                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
2157         else if (cmd == RTM_NEWRULE)
2158                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
2159
2160         if (rule->flags & IPRULE_FWMARK)
2161                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
2162
2163         if (rule->flags & IPRULE_FWMASK)
2164                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
2165
2166         if (rule->flags & IPRULE_LOOKUP) {
2167                 if (rule->lookup >= 256)
2168                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
2169         }
2170
2171         if (rule->flags & IPRULE_GOTO)
2172                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
2173
2174         return system_rtnl_call(msg);
2175 }
2176
2177 int system_add_iprule(struct iprule *rule)
2178 {
2179         return system_iprule(rule, RTM_NEWRULE);
2180 }
2181
2182 int system_del_iprule(struct iprule *rule)
2183 {
2184         return system_iprule(rule, RTM_DELRULE);
2185 }
2186
2187 int system_flush_iprules(void)
2188 {
2189         int rv = 0;
2190         struct iprule rule;
2191
2192         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
2193         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
2194
2195         memset(&rule, 0, sizeof(rule));
2196
2197
2198         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2199
2200         rule.priority = 0;
2201         rule.lookup = RT_TABLE_LOCAL;
2202         rv |= system_iprule(&rule, RTM_NEWRULE);
2203
2204         rule.priority = 32766;
2205         rule.lookup = RT_TABLE_MAIN;
2206         rv |= system_iprule(&rule, RTM_NEWRULE);
2207
2208         rule.priority = 32767;
2209         rule.lookup = RT_TABLE_DEFAULT;
2210         rv |= system_iprule(&rule, RTM_NEWRULE);
2211
2212
2213         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
2214
2215         rule.priority = 0;
2216         rule.lookup = RT_TABLE_LOCAL;
2217         rv |= system_iprule(&rule, RTM_NEWRULE);
2218
2219         rule.priority = 32766;
2220         rule.lookup = RT_TABLE_MAIN;
2221         rv |= system_iprule(&rule, RTM_NEWRULE);
2222
2223         return rv;
2224 }
2225
2226 bool system_resolve_iprule_action(const char *action, unsigned int *id)
2227 {
2228         return system_rtn_aton(action, id);
2229 }
2230
2231 time_t system_get_rtime(void)
2232 {
2233         struct timespec ts;
2234         struct timeval tv;
2235
2236         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
2237                 return ts.tv_sec;
2238
2239         if (gettimeofday(&tv, NULL) == 0)
2240                 return tv.tv_sec;
2241
2242         return 0;
2243 }
2244
2245 #ifndef IP_DF
2246 #define IP_DF       0x4000
2247 #endif
2248
2249 static int tunnel_ioctl(const char *name, int cmd, void *p)
2250 {
2251         struct ifreq ifr;
2252
2253         memset(&ifr, 0, sizeof(ifr));
2254         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name));
2255         ifr.ifr_ifru.ifru_data = p;
2256         return ioctl(sock_ioctl, cmd, &ifr);
2257 }
2258
2259 #ifdef IFLA_IPTUN_MAX
2260 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
2261 static int system_add_gre_tunnel(const char *name, const char *kind,
2262                                  const unsigned int link, struct blob_attr **tb, bool v6)
2263 {
2264         struct nl_msg *nlm;
2265         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2266         struct blob_attr *cur;
2267         uint32_t ikey = 0, okey = 0, flags = 0, flowinfo = 0;
2268         uint16_t iflags = 0, oflags = 0;
2269         uint8_t tos = 0;
2270         int ret = 0, ttl = 0;
2271
2272         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2273         if (!nlm)
2274                 return -1;
2275
2276         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2277         nla_put_string(nlm, IFLA_IFNAME, name);
2278
2279         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2280         if (!linkinfo) {
2281                 ret = -ENOMEM;
2282                 goto failure;
2283         }
2284
2285         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2286         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2287         if (!infodata) {
2288                 ret = -ENOMEM;
2289                 goto failure;
2290         }
2291
2292         if (link)
2293                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
2294
2295         if ((cur = tb[TUNNEL_ATTR_TTL]))
2296                 ttl = blobmsg_get_u32(cur);
2297
2298         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2299                 char *str = blobmsg_get_string(cur);
2300                 if (strcmp(str, "inherit")) {
2301                         unsigned uval;
2302
2303                         if (!system_tos_aton(str, &uval)) {
2304                                 ret = -EINVAL;
2305                                 goto failure;
2306                         }
2307
2308                         if (v6)
2309                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
2310                         else
2311                                 tos = uval;
2312                 } else {
2313                         if (v6)
2314                                 flags |= IP6_TNL_F_USE_ORIG_TCLASS;
2315                         else
2316                                 tos = 1;
2317                 }
2318         }
2319
2320         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2321                 struct blob_attr *tb_data[__GRE_DATA_ATTR_MAX];
2322
2323                 blobmsg_parse(gre_data_attr_list.params, __GRE_DATA_ATTR_MAX, tb_data,
2324                         blobmsg_data(cur), blobmsg_len(cur));
2325
2326                 if ((cur = tb_data[GRE_DATA_IKEY])) {
2327                         if ((ikey = blobmsg_get_u32(cur)))
2328                                 iflags |= GRE_KEY;
2329                 }
2330
2331                 if ((cur = tb_data[GRE_DATA_OKEY])) {
2332                         if ((okey = blobmsg_get_u32(cur)))
2333                                 oflags |= GRE_KEY;
2334                 }
2335
2336                 if ((cur = tb_data[GRE_DATA_ICSUM])) {
2337                         if (blobmsg_get_bool(cur))
2338                                 iflags |= GRE_CSUM;
2339                 }
2340
2341                 if ((cur = tb_data[GRE_DATA_OCSUM])) {
2342                         if (blobmsg_get_bool(cur))
2343                                 oflags |= GRE_CSUM;
2344                 }
2345
2346                 if ((cur = tb_data[GRE_DATA_ISEQNO])) {
2347                         if (blobmsg_get_bool(cur))
2348                                 iflags |= GRE_SEQ;
2349                 }
2350
2351                 if ((cur = tb[GRE_DATA_OSEQNO])) {
2352                         if (blobmsg_get_bool(cur))
2353                                 oflags |= GRE_SEQ;
2354                 }
2355         }
2356
2357         if (v6) {
2358                 struct in6_addr in6buf;
2359                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2360                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2361                                 ret = -EINVAL;
2362                                 goto failure;
2363                         }
2364                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
2365                 }
2366
2367                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2368                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2369                                 ret = -EINVAL;
2370                                 goto failure;
2371                         }
2372                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
2373                 }
2374                 nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, 4);
2375
2376                 if (flowinfo)
2377                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
2378
2379                 if (flags)
2380                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags);
2381
2382                 if (!ttl)
2383                         ttl = 64;
2384         } else {
2385                 struct in_addr inbuf;
2386                 bool set_df = true;
2387
2388                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2389                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2390                                 ret = -EINVAL;
2391                                 goto failure;
2392                         }
2393                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
2394                 }
2395
2396                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2397                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2398                                 ret = -EINVAL;
2399                                 goto failure;
2400                         }
2401                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
2402
2403                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
2404                                 if (!okey) {
2405                                         okey = inbuf.s_addr;
2406                                         oflags |= GRE_KEY;
2407                                 }
2408
2409                                 if (!ikey) {
2410                                         ikey = inbuf.s_addr;
2411                                         iflags |= GRE_KEY;
2412                                 }
2413                         }
2414                 }
2415
2416                 if ((cur = tb[TUNNEL_ATTR_DF]))
2417                         set_df = blobmsg_get_bool(cur);
2418
2419                 if (!set_df) {
2420                         /* ttl != 0 and nopmtudisc are incompatible */
2421                         if (ttl) {
2422                                 ret = -EINVAL;
2423                                 goto failure;
2424                         }
2425                 } else if (!ttl)
2426                         ttl = 64;
2427
2428                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
2429
2430                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
2431         }
2432
2433         if (ttl)
2434                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
2435
2436         if (oflags)
2437                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
2438
2439         if (iflags)
2440                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
2441
2442         if (okey)
2443                 nla_put_u32(nlm, IFLA_GRE_OKEY, okey);
2444
2445         if (ikey)
2446                 nla_put_u32(nlm, IFLA_GRE_IKEY, ikey);
2447
2448         nla_nest_end(nlm, infodata);
2449         nla_nest_end(nlm, linkinfo);
2450
2451         return system_rtnl_call(nlm);
2452
2453 failure:
2454         nlmsg_free(nlm);
2455         return ret;
2456 }
2457 #endif
2458
2459 #ifdef IFLA_VTI_MAX
2460 static int system_add_vti_tunnel(const char *name, const char *kind,
2461                                  const unsigned int link, struct blob_attr **tb, bool v6)
2462 {
2463         struct nl_msg *nlm;
2464         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
2465         struct blob_attr *cur;
2466         int ret = 0;
2467
2468         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2469         if (!nlm)
2470                 return -1;
2471
2472         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2473         nla_put_string(nlm, IFLA_IFNAME, name);
2474
2475         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2476         if (!linkinfo) {
2477                 ret = -ENOMEM;
2478                 goto failure;
2479         }
2480
2481         nla_put_string(nlm, IFLA_INFO_KIND, kind);
2482         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2483         if (!infodata) {
2484                 ret = -ENOMEM;
2485                 goto failure;
2486         }
2487
2488         if (link)
2489                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
2490
2491         if (v6) {
2492                 struct in6_addr in6buf;
2493                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2494                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2495                                 ret = -EINVAL;
2496                                 goto failure;
2497                         }
2498                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
2499                 }
2500
2501                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2502                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2503                                 ret = -EINVAL;
2504                                 goto failure;
2505                         }
2506                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
2507                 }
2508
2509         } else {
2510                 struct in_addr inbuf;
2511
2512                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2513                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2514                                 ret = -EINVAL;
2515                                 goto failure;
2516                         }
2517                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
2518                 }
2519
2520                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2521                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2522                                 ret = -EINVAL;
2523                                 goto failure;
2524                         }
2525                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
2526                 }
2527
2528         }
2529
2530         if ((cur = tb[TUNNEL_ATTR_DATA])) {
2531                 struct blob_attr *tb_data[__VTI_DATA_ATTR_MAX];
2532                 uint32_t ikey = 0, okey = 0;
2533
2534                 blobmsg_parse(vti_data_attr_list.params, __VTI_DATA_ATTR_MAX, tb_data,
2535                         blobmsg_data(cur), blobmsg_len(cur));
2536
2537                 if ((cur = tb_data[VTI_DATA_IKEY])) {
2538                         if ((ikey = blobmsg_get_u32(cur)))
2539                                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
2540                 }
2541
2542                 if ((cur = tb_data[VTI_DATA_OKEY])) {
2543                         if ((okey = blobmsg_get_u32(cur)))
2544                                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
2545                 }
2546         }
2547
2548         nla_nest_end(nlm, infodata);
2549         nla_nest_end(nlm, linkinfo);
2550
2551         return system_rtnl_call(nlm);
2552
2553 failure:
2554         nlmsg_free(nlm);
2555         return ret;
2556 }
2557 #endif
2558
2559 #ifdef IFLA_VXLAN_MAX
2560 static int system_add_vxlan(const char *name, const unsigned int link, struct blob_attr **tb, bool v6)
2561 {
2562         struct blob_attr *tb_data[__VXLAN_DATA_ATTR_MAX];
2563         struct nl_msg *msg;
2564         struct nlattr *linkinfo, *data;
2565         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
2566         struct blob_attr *cur;
2567         int ret = 0;
2568
2569         if ((cur = tb[TUNNEL_ATTR_DATA]))
2570                 blobmsg_parse(vxlan_data_attr_list.params, __VXLAN_DATA_ATTR_MAX, tb_data,
2571                         blobmsg_data(cur), blobmsg_len(cur));
2572         else
2573                 return -EINVAL;
2574
2575         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
2576
2577         if (!msg)
2578                 return -1;
2579
2580         nlmsg_append(msg, &iim, sizeof(iim), 0);
2581
2582         nla_put_string(msg, IFLA_IFNAME, name);
2583
2584         if ((cur = tb_data[VXLAN_DATA_ATTR_MACADDR])) {
2585                 struct ether_addr *ea = ether_aton(blobmsg_get_string(cur));
2586                 if (!ea) {
2587                         ret = -EINVAL;
2588                         goto failure;
2589                 }
2590
2591                 nla_put(msg, IFLA_ADDRESS, ETH_ALEN, ea);
2592         }
2593
2594         if ((cur = tb[TUNNEL_ATTR_MTU])) {
2595                 uint32_t mtu = blobmsg_get_u32(cur);
2596                 nla_put_u32(msg, IFLA_MTU, mtu);
2597         }
2598
2599         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO))) {
2600                 ret = -ENOMEM;
2601                 goto failure;
2602         }
2603
2604         nla_put_string(msg, IFLA_INFO_KIND, "vxlan");
2605
2606         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA))) {
2607                 ret = -ENOMEM;
2608                 goto failure;
2609         }
2610
2611         if (link)
2612                 nla_put_u32(msg, IFLA_VXLAN_LINK, link);
2613
2614         if ((cur = tb_data[VXLAN_DATA_ATTR_ID])) {
2615                 uint32_t id = blobmsg_get_u32(cur);
2616                 if (id >= (1u << 24) - 1) {
2617                         ret = -EINVAL;
2618                         goto failure;
2619                 }
2620
2621                 nla_put_u32(msg, IFLA_VXLAN_ID, id);
2622         }
2623
2624         if (v6) {
2625                 struct in6_addr in6buf;
2626                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2627                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2628                                 ret = -EINVAL;
2629                                 goto failure;
2630                         }
2631                         nla_put(msg, IFLA_VXLAN_LOCAL6, sizeof(in6buf), &in6buf);
2632                 }
2633
2634                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2635                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2636                                 ret = -EINVAL;
2637                                 goto failure;
2638                         }
2639                         nla_put(msg, IFLA_VXLAN_GROUP6, sizeof(in6buf), &in6buf);
2640                 }
2641         } else {
2642                 struct in_addr inbuf;
2643
2644                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2645                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2646                                 ret = -EINVAL;
2647                                 goto failure;
2648                         }
2649                         nla_put(msg, IFLA_VXLAN_LOCAL, sizeof(inbuf), &inbuf);
2650                 }
2651
2652                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2653                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
2654                                 ret = -EINVAL;
2655                                 goto failure;
2656                         }
2657                         nla_put(msg, IFLA_VXLAN_GROUP, sizeof(inbuf), &inbuf);
2658                 }
2659         }
2660
2661         uint32_t port = 4789;
2662         if ((cur = tb_data[VXLAN_DATA_ATTR_PORT])) {
2663                 port = blobmsg_get_u32(cur);
2664                 if (port < 1 || port > 65535) {
2665                         ret = -EINVAL;
2666                         goto failure;
2667                 }
2668         }
2669         nla_put_u16(msg, IFLA_VXLAN_PORT, htons(port));
2670
2671         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2672                 char *str = blobmsg_get_string(cur);
2673                 unsigned tos = 1;
2674
2675                 if (strcmp(str, "inherit")) {
2676                         if (!system_tos_aton(str, &tos))
2677                                 return -EINVAL;
2678                 }
2679
2680                 nla_put_u8(msg, IFLA_VXLAN_TOS, tos);
2681         }
2682
2683         if ((cur = tb[TUNNEL_ATTR_TTL])) {
2684                 uint32_t ttl = blobmsg_get_u32(cur);
2685                 if (ttl < 1 || ttl > 255) {
2686                         ret = -EINVAL;
2687                         goto failure;
2688                 }
2689
2690                 nla_put_u8(msg, IFLA_VXLAN_TTL, ttl);
2691         }
2692
2693         nla_nest_end(msg, data);
2694         nla_nest_end(msg, linkinfo);
2695
2696         ret = system_rtnl_call(msg);
2697         if (ret)
2698                 D(SYSTEM, "Error adding vxlan '%s': %d\n", name, ret);
2699
2700         return ret;
2701
2702 failure:
2703         nlmsg_free(msg);
2704         return ret;
2705 }
2706 #endif
2707
2708 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
2709 {
2710         struct blob_attr *cur;
2711         bool set_df = true;
2712         struct ip_tunnel_parm p  = {
2713                 .link = link,
2714                 .iph = {
2715                         .version = 4,
2716                         .ihl = 5,
2717                         .protocol = proto,
2718                 }
2719         };
2720
2721         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
2722                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
2723                 return -EINVAL;
2724
2725         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
2726                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
2727                 return -EINVAL;
2728
2729         if ((cur = tb[TUNNEL_ATTR_DF]))
2730                 set_df = blobmsg_get_bool(cur);
2731
2732         if ((cur = tb[TUNNEL_ATTR_TTL]))
2733                 p.iph.ttl = blobmsg_get_u32(cur);
2734
2735         if ((cur = tb[TUNNEL_ATTR_TOS])) {
2736                 char *str = blobmsg_get_string(cur);
2737                 if (strcmp(str, "inherit")) {
2738                         unsigned uval;
2739
2740                         if (!system_tos_aton(str, &uval))
2741                                 return -EINVAL;
2742
2743                         p.iph.tos = uval;
2744                 } else
2745                         p.iph.tos = 1;
2746         }
2747
2748         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
2749         /* ttl !=0 and nopmtudisc are incompatible */
2750         if (p.iph.ttl && p.iph.frag_off == 0)
2751                 return -EINVAL;
2752
2753         strncpy(p.name, name, sizeof(p.name));
2754
2755         switch (p.iph.protocol) {
2756         case IPPROTO_IPIP:
2757                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
2758         case IPPROTO_IPV6:
2759                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
2760         default:
2761                 break;
2762         }
2763         return -1;
2764 }
2765
2766 static int __system_del_ip_tunnel(const char *name, struct blob_attr **tb)
2767 {
2768         struct blob_attr *cur;
2769         const char *str;
2770
2771         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2772                 return -EINVAL;
2773         str = blobmsg_data(cur);
2774
2775         if (!strcmp(str, "greip") || !strcmp(str, "gretapip") ||
2776             !strcmp(str, "greip6") || !strcmp(str, "gretapip6") ||
2777             !strcmp(str, "vtiip") || !strcmp(str, "vtiip6") ||
2778             !strcmp(str, "vxlan") || !strcmp(str, "vxlan6"))
2779                 return system_link_del(name);
2780         else
2781                 return tunnel_ioctl(name, SIOCDELTUNNEL, NULL);
2782 }
2783
2784 int system_del_ip_tunnel(const char *name, struct blob_attr *attr)
2785 {
2786         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2787
2788         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2789                 blob_data(attr), blob_len(attr));
2790
2791         return __system_del_ip_tunnel(name, tb);
2792 }
2793
2794 int system_update_ipv6_mtu(struct device *dev, int mtu)
2795 {
2796         int ret = -1;
2797         char buf[64];
2798         int fd;
2799
2800         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/mtu",
2801                         dev->ifname);
2802
2803         fd = open(buf, O_RDWR);
2804         if (fd < 0)
2805                 return ret;
2806
2807         if (!mtu) {
2808                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
2809                 if (len < 0)
2810                         goto out;
2811
2812                 buf[len] = 0;
2813                 ret = atoi(buf);
2814         } else {
2815                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
2816                         ret = mtu;
2817         }
2818
2819 out:
2820         close(fd);
2821         return ret;
2822 }
2823
2824 int system_add_ip_tunnel(const char *name, struct blob_attr *attr)
2825 {
2826         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
2827         struct blob_attr *cur;
2828         const char *str;
2829
2830         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
2831                 blob_data(attr), blob_len(attr));
2832
2833         __system_del_ip_tunnel(name, tb);
2834
2835         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
2836                 return -EINVAL;
2837         str = blobmsg_data(cur);
2838
2839         unsigned int ttl = 0;
2840         if ((cur = tb[TUNNEL_ATTR_TTL])) {
2841                 ttl = blobmsg_get_u32(cur);
2842                 if (ttl > 255)
2843                         return -EINVAL;
2844         }
2845
2846         unsigned int link = 0;
2847         if ((cur = tb[TUNNEL_ATTR_LINK])) {
2848                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
2849                 if (!iface)
2850                         return -EINVAL;
2851
2852                 if (iface->l3_dev.dev)
2853                         link = iface->l3_dev.dev->ifindex;
2854         }
2855
2856         if (!strcmp(str, "sit")) {
2857                 if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
2858                         return -1;
2859
2860 #ifdef SIOCADD6RD
2861                 if ((cur = tb[TUNNEL_ATTR_6RD_PREFIX])) {
2862                         unsigned int mask;
2863                         struct ip_tunnel_6rd p6;
2864
2865                         memset(&p6, 0, sizeof(p6));
2866
2867                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
2868                                                 &p6.prefix, &mask) || mask > 128)
2869                                 return -EINVAL;
2870                         p6.prefixlen = mask;
2871
2872                         if ((cur = tb[TUNNEL_ATTR_6RD_RELAY_PREFIX])) {
2873                                 if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
2874                                                         &p6.relay_prefix, &mask) || mask > 32)
2875                                         return -EINVAL;
2876                                 p6.relay_prefixlen = mask;
2877                         }
2878
2879                         if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
2880                                 __system_del_ip_tunnel(name, tb);
2881                                 return -1;
2882                         }
2883                 }
2884 #endif
2885 #ifdef IFLA_IPTUN_MAX
2886         } else if (!strcmp(str, "ipip6")) {
2887                 struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
2888                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
2889                 struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
2890                 int ret = 0;
2891
2892                 if (!nlm)
2893                         return -1;
2894
2895                 nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
2896                 nla_put_string(nlm, IFLA_IFNAME, name);
2897
2898                 if (link)
2899                         nla_put_u32(nlm, IFLA_LINK, link);
2900
2901                 struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
2902                 if (!linkinfo) {
2903                         ret = -ENOMEM;
2904                         goto failure;
2905                 }
2906                 nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
2907                 struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
2908                 if (!infodata) {
2909                         ret = -ENOMEM;
2910                         goto failure;
2911                 }
2912
2913                 if (link)
2914                         nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
2915
2916                 nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
2917                 nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
2918                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, 4);
2919
2920                 struct in6_addr in6buf;
2921                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
2922                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2923                                 ret = -EINVAL;
2924                                 goto failure;
2925                         }
2926                         nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
2927                 }
2928
2929                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
2930                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
2931                                 ret = -EINVAL;
2932                                 goto failure;
2933                         }
2934                         nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
2935                 }
2936
2937 #ifdef IFLA_IPTUN_FMR_MAX
2938                 if ((cur = tb[TUNNEL_ATTR_FMRS])) {
2939                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
2940
2941                         struct blob_attr *fmr;
2942                         unsigned rem, fmrcnt = 0;
2943                         blobmsg_for_each_attr(fmr, cur, rem) {
2944                                 if (blobmsg_type(fmr) != BLOBMSG_TYPE_STRING)
2945                                         continue;
2946
2947                                 unsigned ip4len, ip6len, ealen, offset = 6;
2948                                 char ip6buf[48];
2949                                 char ip4buf[16];
2950
2951                                 if (sscanf(blobmsg_get_string(fmr), "%47[^/]/%u,%15[^/]/%u,%u,%u",
2952                                                 ip6buf, &ip6len, ip4buf, &ip4len, &ealen, &offset) < 5) {
2953                                         ret = -EINVAL;
2954                                         goto failure;
2955                                 }
2956
2957                                 struct in6_addr ip6prefix;
2958                                 struct in_addr ip4prefix;
2959                                 if (inet_pton(AF_INET6, ip6buf, &ip6prefix) != 1 ||
2960                                                 inet_pton(AF_INET, ip4buf, &ip4prefix) != 1) {
2961                                         ret = -EINVAL;
2962                                         goto failure;
2963                                 }
2964
2965                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
2966
2967                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
2968                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
2969                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
2970                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
2971                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
2972                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
2973
2974                                 nla_nest_end(nlm, rule);
2975                         }
2976
2977                         nla_nest_end(nlm, fmrs);
2978                 }
2979 #endif
2980
2981                 nla_nest_end(nlm, infodata);
2982                 nla_nest_end(nlm, linkinfo);
2983
2984                 return system_rtnl_call(nlm);
2985 failure:
2986                 nlmsg_free(nlm);
2987                 return ret;
2988         } else if (!strcmp(str, "greip")) {
2989                 return system_add_gre_tunnel(name, "gre", link, tb, false);
2990         } else if (!strcmp(str, "gretapip"))  {
2991                 return system_add_gre_tunnel(name, "gretap", link, tb, false);
2992         } else if (!strcmp(str, "greip6")) {
2993                 return system_add_gre_tunnel(name, "ip6gre", link, tb, true);
2994         } else if (!strcmp(str, "gretapip6")) {
2995                 return system_add_gre_tunnel(name, "ip6gretap", link, tb, true);
2996 #ifdef IFLA_VTI_MAX
2997         } else if (!strcmp(str, "vtiip")) {
2998                 return system_add_vti_tunnel(name, "vti", link, tb, false);
2999         } else if (!strcmp(str, "vtiip6")) {
3000                 return system_add_vti_tunnel(name, "vti6", link, tb, true);
3001 #endif
3002 #ifdef IFLA_VXLAN_MAX
3003         } else if(!strcmp(str, "vxlan")) {
3004                 return system_add_vxlan(name, link, tb, false);
3005         } else if(!strcmp(str, "vxlan6")) {
3006                 return system_add_vxlan(name, link, tb, true);
3007 #endif
3008 #endif
3009         } else if (!strcmp(str, "ipip")) {
3010                 return system_add_proto_tunnel(name, IPPROTO_IPIP, link, tb);
3011         }
3012         else
3013                 return -EINVAL;
3014
3015         return 0;
3016 }