system-linux.c

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