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