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