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