system-linux.c

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