system-linux.c

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