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  * Copyright (C) 2017 Matthias Schiffer <mschiffer@universe-factory.net>
   8  * Copyright (C) 2018 Hans Dedecker <dedeckeh@gmail.com>
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License version 2
  12  * as published by the Free Software Foundation
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  */
  19 #define _GNU_SOURCE
  20
  21 #include <sys/socket.h>
  22 #include <sys/ioctl.h>
  23 #include <sys/stat.h>
  24 #include <sys/syscall.h>
  25
  26 #include <net/if.h>
  27 #include <net/if_arp.h>
  28
  29 #include <limits.h>
  30 #include <arpa/inet.h>
  31 #include <netinet/in.h>
  32 #include <netinet/ether.h>
  33
  34 #include <linux/rtnetlink.h>
  35 #include <linux/neighbour.h>
  36 #include <linux/sockios.h>
  37 #include <linux/ip.h>
  38 #include <linux/if_addr.h>
  39 #include <linux/if_link.h>
  40 #include <linux/if_vlan.h>
  41 #include <linux/if_bridge.h>
  42 #include <linux/if_tunnel.h>
  43 #include <linux/ip6_tunnel.h>
  44 #include <linux/ethtool.h>
  45 #include <linux/fib_rules.h>
  46 #include <linux/veth.h>
  47 #include <linux/version.h>
  48
  49 #include <sched.h>
  50
  51 #include "ethtool-modes.h"
  52
  53 #ifndef RTN_FAILED_POLICY
  54 #define RTN_FAILED_POLICY 12
  55 #endif
  56
  57 #ifndef IFA_F_NOPREFIXROUTE
  58 #define IFA_F_NOPREFIXROUTE 0x200
  59 #endif
  60
  61 #ifndef IFA_FLAGS
  62 #define IFA_FLAGS (IFA_MULTICAST + 1)
  63 #endif
  64
  65 #include <string.h>
  66 #include <fcntl.h>
  67 #include <glob.h>
  68 #include <time.h>
  69 #include <unistd.h>
  70
  71 #include <netlink/msg.h>
  72 #include <netlink/attr.h>
  73 #include <netlink/socket.h>
  74 #include <libubox/uloop.h>
  75
  76 #include "netifd.h"
  77 #include "device.h"
  78 #include "system.h"
  79 #include "utils.h"
  80
  81 struct event_socket {
  82         struct uloop_fd uloop;
  83         struct nl_sock *sock;
  84         int bufsize;
  85 };
  86
  87 static int sock_ioctl = -1;
  88 static struct nl_sock *sock_rtnl = NULL;
  89
  90 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  91 static void handle_hotplug_event(struct uloop_fd *u, unsigned int events);
  92 static int system_add_proto_tunnel(const char *name, const uint8_t proto,
  93                                         const unsigned int link, struct blob_attr **tb);
  94
  95 static char dev_buf[256];
  96 static const char *proc_path = "/proc";
  97 static const char *sysfs_path = "/sys";
  98
  99 struct netdev_type {
 100         unsigned short id;
 101         const char *name;
 102 };
 103
 104 static const struct netdev_type netdev_types[] = {
 105         {ARPHRD_NETROM, "netrom"},
 106         {ARPHRD_ETHER, "ethernet"},
 107         {ARPHRD_EETHER, "eethernet"},
 108         {ARPHRD_AX25, "ax25"},
 109         {ARPHRD_PRONET, "pronet"},
 110         {ARPHRD_CHAOS, "chaos"},
 111         {ARPHRD_IEEE802, "ieee802"},
 112         {ARPHRD_ARCNET, "arcnet"},
 113         {ARPHRD_APPLETLK, "appletlk"},
 114         {ARPHRD_DLCI, "dlci"},
 115         {ARPHRD_ATM, "atm"},
 116         {ARPHRD_METRICOM, "metricom"},
 117         {ARPHRD_IEEE1394, "ieee1394"},
 118         {ARPHRD_EUI64, "eui64"},
 119         {ARPHRD_INFINIBAND, "infiniband"},
 120         {ARPHRD_SLIP, "slip"},
 121         {ARPHRD_CSLIP, "cslip"},
 122         {ARPHRD_SLIP6, "slip6"},
 123         {ARPHRD_CSLIP6, "cslip6"},
 124         {ARPHRD_RSRVD, "rsrvd"},
 125         {ARPHRD_ADAPT, "adapt"},
 126         {ARPHRD_ROSE, "rose"},
 127         {ARPHRD_X25, "x25"},
 128         {ARPHRD_HWX25, "hwx25"},
 129         {ARPHRD_PPP, "ppp"},
 130         {ARPHRD_CISCO, "cisco"},
 131         {ARPHRD_LAPB, "lapb"},
 132         {ARPHRD_DDCMP, "ddcmp"},
 133         {ARPHRD_RAWHDLC, "rawhdlc"},
 134         {ARPHRD_TUNNEL, "tunnel"},
 135         {ARPHRD_TUNNEL6, "tunnel6"},
 136         {ARPHRD_FRAD, "frad"},
 137         {ARPHRD_SKIP, "skip"},
 138         {ARPHRD_LOOPBACK, "loopback"},
 139         {ARPHRD_LOCALTLK, "localtlk"},
 140         {ARPHRD_FDDI, "fddi"},
 141         {ARPHRD_BIF, "bif"},
 142         {ARPHRD_SIT, "sit"},
 143         {ARPHRD_IPDDP, "ipddp"},
 144         {ARPHRD_IPGRE, "ipgre"},
 145         {ARPHRD_PIMREG,"pimreg"},
 146         {ARPHRD_HIPPI, "hippi"},
 147         {ARPHRD_ASH, "ash"},
 148         {ARPHRD_ECONET, "econet"},
 149         {ARPHRD_IRDA, "irda"},
 150         {ARPHRD_FCPP, "fcpp"},
 151         {ARPHRD_FCAL, "fcal"},
 152         {ARPHRD_FCPL, "fcpl"},
 153         {ARPHRD_FCFABRIC, "fcfabric"},
 154         {ARPHRD_IEEE80211, "ieee80211"},
 155         {ARPHRD_IEEE80211_PRISM, "ie80211-prism"},
 156         {ARPHRD_IEEE80211_RADIOTAP, "ieee80211-radiotap"},
 157 #ifdef ARPHRD_PHONET
 158         {ARPHRD_PHONET, "phonet"},
 159 #endif
 160 #ifdef ARPHRD_PHONET_PIPE
 161         {ARPHRD_PHONET_PIPE, "phonet-pipe"},
 162 #endif
 163         {ARPHRD_IEEE802154, "ieee802154"},
 164         {ARPHRD_VOID, "void"},
 165         {ARPHRD_NONE, "none"}
 166 };
 167
 168 static void
 169 handler_nl_event(struct uloop_fd *u, unsigned int events)
 170 {
 171         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 172         int err;
 173         socklen_t errlen = sizeof(err);
 174
 175         if (!u->error) {
 176                 nl_recvmsgs_default(ev->sock);
 177                 return;
 178         }
 179
 180         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 181                 goto abort;
 182
 183         switch(err) {
 184         case ENOBUFS:
 185                 /* Increase rx buffer size on netlink socket */
 186                 ev->bufsize *= 2;
 187                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 188                         goto abort;
 189
 190                 /* Request full dump since some info got dropped */
 191                 struct rtgenmsg msg = { .rtgen_family = AF_UNSPEC };
 192                 nl_send_simple(ev->sock, RTM_GETLINK, NLM_F_DUMP, &msg, sizeof(msg));
 193                 break;
 194
 195         default:
 196                 goto abort;
 197         }
 198         u->error = false;
 199         return;
 200
 201 abort:
 202         uloop_fd_delete(&ev->uloop);
 203         return;
 204 }
 205
 206 static struct nl_sock *
 207 create_socket(int protocol, int groups)
 208 {
 209         struct nl_sock *sock;
 210
 211         sock = nl_socket_alloc();
 212         if (!sock)
 213                 return NULL;
 214
 215         if (groups)
 216                 nl_join_groups(sock, groups);
 217
 218         if (nl_connect(sock, protocol)) {
 219                 nl_socket_free(sock);
 220                 return NULL;
 221         }
 222
 223         return sock;
 224 }
 225
 226 static bool
 227 create_raw_event_socket(struct event_socket *ev, int protocol, int groups,
 228                         uloop_fd_handler cb, int flags)
 229 {
 230         ev->sock = create_socket(protocol, groups);
 231         if (!ev->sock)
 232                 return false;
 233
 234         ev->uloop.fd = nl_socket_get_fd(ev->sock);
 235         ev->uloop.cb = cb;
 236         if (uloop_fd_add(&ev->uloop, ULOOP_READ|flags))
 237                 return false;
 238
 239         return true;
 240 }
 241
 242 static bool
 243 create_event_socket(struct event_socket *ev, int protocol,
 244                     int (*cb)(struct nl_msg *msg, void *arg))
 245 {
 246         if (!create_raw_event_socket(ev, protocol, 0, handler_nl_event, ULOOP_ERROR_CB))
 247                 return false;
 248
 249         /* Install the valid custom callback handler */
 250         nl_socket_modify_cb(ev->sock, NL_CB_VALID, NL_CB_CUSTOM, cb, NULL);
 251
 252         /* Disable sequence number checking on event sockets */
 253         nl_socket_disable_seq_check(ev->sock);
 254
 255         /* Increase rx buffer size to 65K on event sockets */
 256         ev->bufsize = 65535;
 257         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 258                 return false;
 259
 260         return true;
 261 }
 262
 263 static bool
 264 create_hotplug_event_socket(struct event_socket *ev, int protocol,
 265                             void (*cb)(struct uloop_fd *u, unsigned int events))
 266 {
 267         if (!create_raw_event_socket(ev, protocol, 1, cb, ULOOP_ERROR_CB))
 268                 return false;
 269
 270         /* Increase rx buffer size to 65K on event sockets */
 271         ev->bufsize = 65535;
 272         if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 273                 return false;
 274
 275         return true;
 276 }
 277
 278 static bool
 279 system_rtn_aton(const char *src, unsigned int *dst)
 280 {
 281         char *e;
 282         unsigned int n;
 283
 284         if (!strcmp(src, "local"))
 285                 n = RTN_LOCAL;
 286         else if (!strcmp(src, "nat"))
 287                 n = RTN_NAT;
 288         else if (!strcmp(src, "broadcast"))
 289                 n = RTN_BROADCAST;
 290         else if (!strcmp(src, "anycast"))
 291                 n = RTN_ANYCAST;
 292         else if (!strcmp(src, "multicast"))
 293                 n = RTN_MULTICAST;
 294         else if (!strcmp(src, "prohibit"))
 295                 n = RTN_PROHIBIT;
 296         else if (!strcmp(src, "unreachable"))
 297                 n = RTN_UNREACHABLE;
 298         else if (!strcmp(src, "blackhole"))
 299                 n = RTN_BLACKHOLE;
 300         else if (!strcmp(src, "xresolve"))
 301                 n = RTN_XRESOLVE;
 302         else if (!strcmp(src, "unicast"))
 303                 n = RTN_UNICAST;
 304         else if (!strcmp(src, "throw"))
 305                 n = RTN_THROW;
 306         else if (!strcmp(src, "failed_policy"))
 307                 n = RTN_FAILED_POLICY;
 308         else {
 309                 n = strtoul(src, &e, 0);
 310                 if (!e || *e || e == src || n > 255)
 311                         return false;
 312         }
 313
 314         *dst = n;
 315         return true;
 316 }
 317
 318 static bool
 319 system_tos_aton(const char *src, unsigned *dst)
 320 {
 321         char *e;
 322
 323         *dst = strtoul(src, &e, 16);
 324         if (e == src || *e || *dst > 255)
 325                 return false;
 326
 327         return true;
 328 }
 329
 330 int system_init(void)
 331 {
 332         static struct event_socket rtnl_event;
 333         static struct event_socket hotplug_event;
 334
 335         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
 336         system_fd_set_cloexec(sock_ioctl);
 337
 338         /* Prepare socket for routing / address control */
 339         sock_rtnl = create_socket(NETLINK_ROUTE, 0);
 340         if (!sock_rtnl)
 341                 return -1;
 342
 343         if (!create_event_socket(&rtnl_event, NETLINK_ROUTE, cb_rtnl_event))
 344                 return -1;
 345
 346         if (!create_hotplug_event_socket(&hotplug_event, NETLINK_KOBJECT_UEVENT,
 347                                          handle_hotplug_event))
 348                 return -1;
 349
 350         /* Receive network link events form kernel */
 351         nl_socket_add_membership(rtnl_event.sock, RTNLGRP_LINK);
 352
 353         return 0;
 354 }
 355
 356 static void write_file(const char *path, const char *val)
 357 {
 358         int fd;
 359
 360         fd = open(path, O_WRONLY);
 361         if (fd < 0)
 362                 return;
 363
 364         if (write(fd, val, strlen(val))) {}
 365         close(fd);
 366 }
 367
 368 static int read_file(const char *path, char *buf, const size_t buf_sz)
 369 {
 370         int fd = -1, ret = -1;
 371
 372         fd = open(path, O_RDONLY);
 373         if (fd < 0)
 374                 goto out;
 375
 376         ssize_t len = read(fd, buf, buf_sz - 1);
 377         if (len < 0)
 378                 goto out;
 379
 380         ret = buf[len] = 0;
 381
 382 out:
 383         if (fd >= 0)
 384                 close(fd);
 385
 386         return ret;
 387 }
 388
 389
 390 static const char *
 391 dev_sysctl_path(const char *prefix, const char *ifname, const char *file)
 392 {
 393         snprintf(dev_buf, sizeof(dev_buf), "%s/sys/net/%s/%s/%s", proc_path, prefix, ifname, file);
 394
 395         return dev_buf;
 396 }
 397
 398 static const char *
 399 dev_sysfs_path(const char *ifname, const char *file)
 400 {
 401         snprintf(dev_buf, sizeof(dev_buf), "%s/class/net/%s/%s", sysfs_path, ifname, file);
 402
 403         return dev_buf;
 404 }
 405
 406 static void
 407 system_set_dev_sysctl(const char *prefix, const char *file, const char *ifname,
 408                       const char *val)
 409 {
 410         write_file(dev_sysctl_path(prefix, ifname, file), val);
 411 }
 412
 413 static int
 414 system_get_dev_sysctl(const char *prefix, const char *file, const char *ifname,
 415                       char *buf, size_t buf_sz)
 416 {
 417         return read_file(dev_sysctl_path(prefix, ifname, file), buf, buf_sz);
 418 }
 419
 420 static void
 421 system_set_dev_sysfs(const char *file, const char *ifname, const char *val)
 422 {
 423         if (!val)
 424                 return;
 425
 426         write_file(dev_sysfs_path(ifname, file), val);
 427 }
 428
 429 static void
 430 system_set_dev_sysfs_int(const char *file, const char *ifname, int val)
 431 {
 432         char buf[16];
 433
 434         snprintf(buf, sizeof(buf), "%d", val);
 435         system_set_dev_sysfs(file, ifname, buf);
 436 }
 437
 438 static int
 439 system_get_dev_sysfs(const char *file, const char *ifname, char *buf, size_t buf_sz)
 440 {
 441         return read_file(dev_sysfs_path(ifname, file), buf, buf_sz);
 442 }
 443
 444 static void system_set_disable_ipv6(struct device *dev, const char *val)
 445 {
 446         system_set_dev_sysctl("ipv6/conf", "disable_ipv6", dev->ifname, val);
 447 }
 448
 449 static void system_set_ip6segmentrouting(struct device *dev, const char *val)
 450 {
 451         system_set_dev_sysctl("ipv6/conf", "seg6_enabled", dev->ifname, val);
 452 }
 453
 454 static void system_set_rpfilter(struct device *dev, const char *val)
 455 {
 456         system_set_dev_sysctl("ipv4/conf", "rp_filter", dev->ifname, val);
 457 }
 458
 459 static void system_set_acceptlocal(struct device *dev, const char *val)
 460 {
 461         system_set_dev_sysctl("ipv4/conf", "accept_local", dev->ifname, val);
 462 }
 463
 464 static void system_set_igmpversion(struct device *dev, const char *val)
 465 {
 466         system_set_dev_sysctl("ipv4/conf", "force_igmp_version", dev->ifname, val);
 467 }
 468
 469 static void system_set_mldversion(struct device *dev, const char *val)
 470 {
 471         system_set_dev_sysctl("ipv6/conf", "force_mld_version", dev->ifname, val);
 472 }
 473
 474 static void system_set_neigh4reachabletime(struct device *dev, const char *val)
 475 {
 476         system_set_dev_sysctl("ipv4/neigh", "base_reachable_time_ms", dev->ifname, val);
 477 }
 478
 479 static void system_set_neigh6reachabletime(struct device *dev, const char *val)
 480 {
 481         system_set_dev_sysctl("ipv6/neigh", "base_reachable_time_ms", dev->ifname, val);
 482 }
 483
 484 static void system_set_neigh4gcstaletime(struct device *dev, const char *val)
 485 {
 486         system_set_dev_sysctl("ipv4/neigh", "gc_stale_time", dev->ifname, val);
 487 }
 488
 489 static void system_set_neigh6gcstaletime(struct device *dev, const char *val)
 490 {
 491         system_set_dev_sysctl("ipv6/neigh", "gc_stale_time", dev->ifname, val);
 492 }
 493
 494 static void system_set_neigh4locktime(struct device *dev, const char *val)
 495 {
 496         system_set_dev_sysctl("ipv4/neigh", "locktime", dev->ifname, val);
 497 }
 498
 499 static void system_set_dadtransmits(struct device *dev, const char *val)
 500 {
 501         system_set_dev_sysctl("ipv6/conf", "dad_transmits", dev->ifname, val);
 502 }
 503
 504 static void system_set_sendredirects(struct device *dev, const char *val)
 505 {
 506         system_set_dev_sysctl("ipv4/conf", "send_redirects", dev->ifname, val);
 507 }
 508
 509 static void system_set_drop_v4_unicast_in_l2_multicast(struct device *dev, const char *val)
 510 {
 511         system_set_dev_sysctl("ipv4/conf", "drop_unicast_in_l2_multicast", dev->ifname, val);
 512 }
 513
 514 static void system_set_drop_v6_unicast_in_l2_multicast(struct device *dev, const char *val)
 515 {
 516         system_set_dev_sysctl("ipv6/conf", "drop_unicast_in_l2_multicast", dev->ifname, val);
 517 }
 518
 519 static void system_set_drop_gratuitous_arp(struct device *dev, const char *val)
 520 {
 521         system_set_dev_sysctl("ipv4/conf", "drop_gratuitous_arp", dev->ifname, val);
 522 }
 523
 524 static void system_set_drop_unsolicited_na(struct device *dev, const char *val)
 525 {
 526         system_set_dev_sysctl("ipv6/conf", "drop_unsolicited_na", dev->ifname, val);
 527 }
 528
 529 static void system_set_arp_accept(struct device *dev, const char *val)
 530 {
 531         system_set_dev_sysctl("ipv4/conf", "arp_accept", dev->ifname, val);
 532 }
 533
 534 static void system_bridge_set_multicast_to_unicast(struct device *dev, const char *val)
 535 {
 536         system_set_dev_sysfs("brport/multicast_to_unicast", dev->ifname, val);
 537 }
 538
 539 static void system_bridge_set_multicast_fast_leave(struct device *dev, const char *val)
 540 {
 541         system_set_dev_sysfs("brport/multicast_fast_leave", dev->ifname, val);
 542 }
 543
 544 static void system_bridge_set_hairpin_mode(struct device *dev, const char *val)
 545 {
 546         system_set_dev_sysfs("brport/hairpin_mode", dev->ifname, val);
 547 }
 548
 549 static void system_bridge_set_proxyarp_wifi(struct device *dev, const char *val)
 550 {
 551         system_set_dev_sysfs("brport/proxyarp_wifi", dev->ifname, val);
 552 }
 553
 554 static void system_bridge_set_bpdu_filter(struct device *dev, const char *val)
 555 {
 556         system_set_dev_sysfs("brport/bpdu_filter", dev->ifname, val);
 557 }
 558
 559 static void system_bridge_set_isolated(struct device *dev, const char *val)
 560 {
 561         system_set_dev_sysfs("brport/isolated", dev->ifname, val);
 562 }
 563
 564 static void system_bridge_set_multicast_router(struct device *dev, const char *val)
 565 {
 566         system_set_dev_sysfs("brport/multicast_router", dev->ifname, val);
 567 }
 568
 569 void system_bridge_set_stp_state(struct device *dev, bool val)
 570 {
 571         const char *valstr = val ? "1" : "0";
 572
 573         system_set_dev_sysfs("bridge/stp_state", dev->ifname, valstr);
 574 }
 575
 576 static void system_bridge_set_learning(struct device *dev, const char *val)
 577 {
 578         system_set_dev_sysfs("brport/learning", dev->ifname, val);
 579 }
 580
 581 static void system_bridge_set_unicast_flood(struct device *dev, const char *val)
 582 {
 583         system_set_dev_sysfs("brport/unicast_flood", dev->ifname, val);
 584 }
 585
 586 static int system_get_disable_ipv6(struct device *dev, char *buf, const size_t buf_sz)
 587 {
 588         return system_get_dev_sysctl("ipv6/conf", "disable_ipv6",
 589                                      dev->ifname, buf, buf_sz);
 590 }
 591
 592 static int system_get_ip6segmentrouting(struct device *dev, char *buf, const size_t buf_sz)
 593 {
 594         return system_get_dev_sysctl("ipv6/conf", "seg6_enabled",
 595                                      dev->ifname, buf, buf_sz);
 596 }
 597
 598 static int system_get_rpfilter(struct device *dev, char *buf, const size_t buf_sz)
 599 {
 600         return system_get_dev_sysctl("ipv4/conf", "rp_filter",
 601                                      dev->ifname, buf, buf_sz);
 602 }
 603
 604 static int system_get_acceptlocal(struct device *dev, char *buf, const size_t buf_sz)
 605 {
 606         return system_get_dev_sysctl("ipv4/conf", "accept_local",
 607                                      dev->ifname, buf, buf_sz);
 608 }
 609
 610 static int system_get_igmpversion(struct device *dev, char *buf, const size_t buf_sz)
 611 {
 612         return system_get_dev_sysctl("ipv4/conf", "force_igmp_version",
 613                                      dev->ifname, buf, buf_sz);
 614 }
 615
 616 static int system_get_mldversion(struct device *dev, char *buf, const size_t buf_sz)
 617 {
 618         return system_get_dev_sysctl("ipv6/conf", "force_mld_version",
 619                                      dev->ifname, buf, buf_sz);
 620 }
 621
 622 static int system_get_neigh4reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 623 {
 624         return system_get_dev_sysctl("ipv4/neigh", "base_reachable_time_ms",
 625                                      dev->ifname, buf, buf_sz);
 626 }
 627
 628 static int system_get_neigh6reachabletime(struct device *dev, char *buf, const size_t buf_sz)
 629 {
 630         return system_get_dev_sysctl("ipv6/neigh", "base_reachable_time_ms",
 631                                      dev->ifname, buf, buf_sz);
 632 }
 633
 634 static int system_get_neigh4gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 635 {
 636         return system_get_dev_sysctl("ipv4/neigh", "gc_stale_time",
 637                                      dev->ifname, buf, buf_sz);
 638 }
 639
 640 static int system_get_neigh6gcstaletime(struct device *dev, char *buf, const size_t buf_sz)
 641 {
 642         return system_get_dev_sysctl("ipv6/neigh", "gc_stale_time",
 643                         dev->ifname, buf, buf_sz);
 644 }
 645
 646 static int system_get_neigh4locktime(struct device *dev, char *buf, const size_t buf_sz)
 647 {
 648         return system_get_dev_sysctl("ipv4/neigh", "locktime",
 649                         dev->ifname, buf, buf_sz);
 650 }
 651
 652 static int system_get_dadtransmits(struct device *dev, char *buf, const size_t buf_sz)
 653 {
 654         return system_get_dev_sysctl("ipv6/conf", "dad_transmits",
 655                         dev->ifname, buf, buf_sz);
 656 }
 657
 658 static int system_get_sendredirects(struct device *dev, char *buf, const size_t buf_sz)
 659 {
 660         return system_get_dev_sysctl("ipv4/conf", "send_redirects",
 661                         dev->ifname, buf, buf_sz);
 662 }
 663
 664
 665 static int system_get_drop_v4_unicast_in_l2_multicast(struct device *dev, char *buf, const size_t buf_sz)
 666 {
 667         return system_get_dev_sysctl("ipv4/conf", "drop_unicast_in_l2_multicast",
 668                         dev->ifname, buf, buf_sz);
 669 }
 670
 671 static int system_get_drop_v6_unicast_in_l2_multicast(struct device *dev, char *buf, const size_t buf_sz)
 672 {
 673         return system_get_dev_sysctl("ipv6/conf", "drop_unicast_in_l2_multicast",
 674                         dev->ifname, buf, buf_sz);
 675 }
 676
 677 static int system_get_drop_gratuitous_arp(struct device *dev, char *buf, const size_t buf_sz)
 678 {
 679         return system_get_dev_sysctl("ipv4/conf", "drop_gratuitous_arp",
 680                         dev->ifname, buf, buf_sz);
 681 }
 682
 683 static int system_get_drop_unsolicited_na(struct device *dev, char *buf, const size_t buf_sz)
 684 {
 685         return system_get_dev_sysctl("ipv6/conf", "drop_unsolicited_na",
 686                         dev->ifname, buf, buf_sz);
 687 }
 688
 689 static int system_get_arp_accept(struct device *dev, char *buf, const size_t buf_sz)
 690 {
 691         return system_get_dev_sysctl("ipv4/conf", "arp_accept",
 692                         dev->ifname, buf, buf_sz);
 693 }
 694
 695 #ifndef IFF_LOWER_UP
 696 #define IFF_LOWER_UP    0x10000
 697 #endif
 698
 699 static void
 700 system_device_update_state(struct device *dev, unsigned int flags, unsigned int ifindex)
 701 {
 702         if (dev->type == &simple_device_type) {
 703                 bool present = ifindex > 0;
 704
 705                 if (dev->external)
 706                         present = present && (flags & IFF_UP);
 707
 708                 device_set_present(dev, present);
 709         }
 710         device_set_link(dev, flags & IFF_LOWER_UP ? true : false);
 711 }
 712
 713 /* Evaluate netlink messages */
 714 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 715 {
 716         struct nlmsghdr *nh = nlmsg_hdr(msg);
 717         struct ifinfomsg *ifi = NLMSG_DATA(nh);
 718         struct nlattr *nla[__IFLA_MAX];
 719         struct device *dev;
 720
 721         if (nh->nlmsg_type != RTM_NEWLINK)
 722                 return 0;
 723
 724         nlmsg_parse(nh, sizeof(struct ifinfomsg), nla, __IFLA_MAX - 1, NULL);
 725         if (!nla[IFLA_IFNAME])
 726                 return 0;
 727
 728         dev = device_find(nla_data(nla[IFLA_IFNAME]));
 729         if (!dev)
 730                 return 0;
 731
 732         system_device_update_state(dev, ifi->ifi_flags, ifi->ifi_index);
 733         return 0;
 734 }
 735
 736 static void
 737 handle_hotplug_msg(char *data, int size)
 738 {
 739         const char *subsystem = NULL, *interface = NULL, *interface_old = NULL;
 740         char *cur, *end, *sep;
 741         int skip;
 742         bool add;
 743
 744         if (!strncmp(data, "add@", 4) || !strncmp(data, "move@", 5))
 745                 add = true;
 746         else if (!strncmp(data, "remove@", 7))
 747                 add = false;
 748         else
 749                 return;
 750
 751         skip = strlen(data) + 1;
 752         end = data + size;
 753
 754         for (cur = data + skip; cur < end; cur += skip) {
 755                 skip = strlen(cur) + 1;
 756
 757                 sep = strchr(cur, '=');
 758                 if (!sep)
 759                         continue;
 760
 761                 *sep = 0;
 762                 if (!strcmp(cur, "INTERFACE"))
 763                         interface = sep + 1;
 764                 else if (!strcmp(cur, "SUBSYSTEM")) {
 765                         subsystem = sep + 1;
 766                         if (strcmp(subsystem, "net") != 0)
 767                                 return;
 768                 } else if (!strcmp(cur, "DEVPATH_OLD")) {
 769                         interface_old = strrchr(sep + 1, '/');
 770                         if (interface_old)
 771                                 interface_old++;
 772                 }
 773         }
 774
 775         if (!subsystem || !interface)
 776                 return;
 777
 778         if (interface_old)
 779                 device_hotplug_event(interface_old, false);
 780
 781         device_hotplug_event(interface, add);
 782 }
 783
 784 static void
 785 handle_hotplug_event(struct uloop_fd *u, unsigned int events)
 786 {
 787         struct event_socket *ev = container_of(u, struct event_socket, uloop);
 788         struct sockaddr_nl nla;
 789         unsigned char *buf = NULL;
 790         int size;
 791         int err;
 792         socklen_t errlen = sizeof(err);
 793
 794         if (!u->error) {
 795                 while ((size = nl_recv(ev->sock, &nla, &buf, NULL)) > 0) {
 796                         if (nla.nl_pid == 0)
 797                                 handle_hotplug_msg((char *) buf, size);
 798
 799                         free(buf);
 800                 }
 801                 return;
 802         }
 803
 804         if (getsockopt(u->fd, SOL_SOCKET, SO_ERROR, (void *)&err, &errlen))
 805                 goto abort;
 806
 807         switch(err) {
 808         case ENOBUFS:
 809                 /* Increase rx buffer size on netlink socket */
 810                 ev->bufsize *= 2;
 811                 if (nl_socket_set_buffer_size(ev->sock, ev->bufsize, 0))
 812                         goto abort;
 813                 break;
 814
 815         default:
 816                 goto abort;
 817         }
 818         u->error = false;
 819         return;
 820
 821 abort:
 822         uloop_fd_delete(&ev->uloop);
 823         return;
 824 }
 825
 826 static int system_rtnl_call(struct nl_msg *msg)
 827 {
 828         int ret;
 829
 830         ret = nl_send_auto_complete(sock_rtnl, msg);
 831         nlmsg_free(msg);
 832
 833         if (ret < 0)
 834                 return ret;
 835
 836         return nl_wait_for_ack(sock_rtnl);
 837 }
 838
 839 static struct nl_msg *__system_ifinfo_msg(int af, int index, const char *ifname, uint16_t type, uint16_t flags)
 840 {
 841         struct nl_msg *msg;
 842         struct ifinfomsg iim = {
 843                 .ifi_family = af,
 844                 .ifi_index = index,
 845         };
 846
 847         msg = nlmsg_alloc_simple(type, flags | NLM_F_REQUEST);
 848         if (!msg)
 849                 return NULL;
 850
 851         nlmsg_append(msg, &iim, sizeof(iim), 0);
 852         if (ifname)
 853                 nla_put_string(msg, IFLA_IFNAME, ifname);
 854
 855         return msg;
 856 }
 857
 858 static struct nl_msg *system_ifinfo_msg(const char *ifname, uint16_t type, uint16_t flags)
 859 {
 860         return __system_ifinfo_msg(AF_UNSPEC, 0, ifname, type, flags);
 861 }
 862
 863 static int system_link_del(const char *ifname)
 864 {
 865         struct nl_msg *msg;
 866
 867         msg = system_ifinfo_msg(ifname, RTM_DELLINK, 0);
 868         if (!msg)
 869                 return -1;
 870
 871         return system_rtnl_call(msg);
 872 }
 873
 874 int system_bridge_delbr(struct device *bridge)
 875 {
 876         return system_link_del(bridge->ifname);
 877 }
 878
 879 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 880 {
 881         struct ifreq ifr;
 882
 883         memset(&ifr, 0, sizeof(ifr));
 884         if (dev)
 885                 ifr.ifr_ifindex = dev->ifindex;
 886         else
 887                 ifr.ifr_data = data;
 888         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name) - 1);
 889         return ioctl(sock_ioctl, cmd, &ifr);
 890 }
 891
 892 static bool system_is_bridge(const char *name)
 893 {
 894         struct stat st;
 895
 896         return stat(dev_sysfs_path(name, "bridge"), &st) >= 0;
 897 }
 898
 899 static char *system_get_bridge(const char *name, char *buf, int buflen)
 900 {
 901         char *path;
 902         ssize_t len = -1;
 903         glob_t gl;
 904
 905         snprintf(buf, buflen, "%s/devices/virtual/net/*/brif/%s/bridge", sysfs_path, name);
 906         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 907                 return NULL;
 908
 909         if (gl.gl_pathc > 0)
 910                 len = readlink(gl.gl_pathv[0], buf, buflen);
 911
 912         globfree(&gl);
 913
 914         if (len < 0)
 915                 return NULL;
 916
 917         buf[len] = 0;
 918         path = strrchr(buf, '/');
 919         if (!path)
 920                 return NULL;
 921
 922         return path + 1;
 923 }
 924
 925 static void
 926 system_bridge_set_wireless(struct device *bridge, struct device *dev)
 927 {
 928         bool mcast_to_ucast = dev->wireless_ap;
 929         bool hairpin;
 930
 931         if (dev->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST)
 932                 mcast_to_ucast = dev->settings.multicast_to_unicast;
 933         else if (bridge->settings.flags & DEV_OPT_MULTICAST_TO_UNICAST &&
 934                  !bridge->settings.multicast_to_unicast)
 935                 mcast_to_ucast = false;
 936
 937         hairpin = mcast_to_ucast || dev->wireless_proxyarp;
 938         if (dev->wireless_isolate)
 939                 hairpin = false;
 940
 941         system_bridge_set_multicast_to_unicast(dev, mcast_to_ucast ? "1" : "0");
 942         system_bridge_set_hairpin_mode(dev, hairpin ? "1" : "0");
 943         system_bridge_set_proxyarp_wifi(dev, dev->wireless_proxyarp ? "1" : "0");
 944 }
 945
 946 int system_bridge_addif(struct device *bridge, struct device *dev)
 947 {
 948         char buf[64];
 949         char *oldbr;
 950         int tries = 0;
 951         int ret;
 952
 953 retry:
 954         ret = 0;
 955         oldbr = system_get_bridge(dev->ifname, dev_buf, sizeof(dev_buf));
 956         if (!oldbr || strcmp(oldbr, bridge->ifname) != 0) {
 957                 ret = system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 958                 tries++;
 959                 D(SYSTEM, "Failed to add device '%s' to bridge '%s' (tries=%d): %s\n",
 960                   dev->ifname, bridge->ifname, tries, strerror(errno));
 961                 if (tries <= 3)
 962                         goto retry;
 963         }
 964
 965         if (dev->wireless)
 966                 system_bridge_set_wireless(bridge, dev);
 967
 968         if (dev->settings.flags & DEV_OPT_MULTICAST_ROUTER) {
 969                 snprintf(buf, sizeof(buf), "%u", dev->settings.multicast_router);
 970                 system_bridge_set_multicast_router(dev, buf);
 971         }
 972
 973         if (dev->settings.flags & DEV_OPT_MULTICAST_FAST_LEAVE &&
 974             dev->settings.multicast_fast_leave)
 975                 system_bridge_set_multicast_fast_leave(dev, "1");
 976
 977         if (dev->settings.flags & DEV_OPT_LEARNING &&
 978             !dev->settings.learning)
 979                 system_bridge_set_learning(dev, "0");
 980
 981         if (dev->settings.flags & DEV_OPT_UNICAST_FLOOD &&
 982             !dev->settings.unicast_flood)
 983                 system_bridge_set_unicast_flood(dev, "0");
 984
 985         if (dev->settings.flags & DEV_OPT_ISOLATE &&
 986             dev->settings.isolate)
 987                 system_bridge_set_isolated(dev, "1");
 988
 989         if (dev->bpdu_filter)
 990                 system_bridge_set_bpdu_filter(dev, dev->bpdu_filter ? "1" : "0");
 991
 992         return ret;
 993 }
 994
 995 int system_bridge_delif(struct device *bridge, struct device *dev)
 996 {
 997         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 998 }
 999
1000 int system_bridge_vlan(const char *iface, uint16_t vid, int16_t vid_end, bool add, unsigned int vflags)
1001 {
1002         struct bridge_vlan_info vinfo = { .vid = vid, };
1003         unsigned short flags = 0;
1004         struct nlattr *afspec;
1005         struct nl_msg *nlm;
1006         int index;
1007         int ret = 0;
1008
1009         index = if_nametoindex(iface);
1010         if (!index)
1011                 return -1;
1012
1013         nlm = __system_ifinfo_msg(PF_BRIDGE, index, NULL, add ? RTM_SETLINK : RTM_DELLINK, 0);
1014         if (!nlm)
1015                 return -1;
1016
1017         if (vflags & BRVLAN_F_SELF)
1018                 flags |= BRIDGE_FLAGS_SELF;
1019
1020         if (vflags & BRVLAN_F_PVID)
1021                 vinfo.flags |= BRIDGE_VLAN_INFO_PVID;
1022
1023         if (vflags & BRVLAN_F_UNTAGGED)
1024                 vinfo.flags |= BRIDGE_VLAN_INFO_UNTAGGED;
1025
1026         afspec = nla_nest_start(nlm, IFLA_AF_SPEC);
1027         if (!afspec) {
1028                 ret = -ENOMEM;
1029                 goto failure;
1030         }
1031
1032         if (flags)
1033                 nla_put_u16(nlm, IFLA_BRIDGE_FLAGS, flags);
1034
1035         if (vid_end > vid)
1036                 vinfo.flags |= BRIDGE_VLAN_INFO_RANGE_BEGIN;
1037
1038         nla_put(nlm, IFLA_BRIDGE_VLAN_INFO, sizeof(vinfo), &vinfo);
1039
1040         if (vid_end > vid) {
1041                 vinfo.flags &= ~BRIDGE_VLAN_INFO_RANGE_BEGIN;
1042                 vinfo.flags |= BRIDGE_VLAN_INFO_RANGE_END;
1043                 vinfo.vid = vid_end;
1044                 nla_put(nlm, IFLA_BRIDGE_VLAN_INFO, sizeof(vinfo), &vinfo);
1045         }
1046
1047         nla_nest_end(nlm, afspec);
1048
1049         return system_rtnl_call(nlm);
1050
1051 failure:
1052         nlmsg_free(nlm);
1053         return ret;
1054 }
1055
1056 int system_bonding_set_device(struct device *dev, struct bonding_config *cfg)
1057 {
1058         const char *ifname = dev->ifname;
1059         struct blob_attr *cur;
1060         char op = cfg ? '+' : '-';
1061         char buf[64];
1062         size_t rem;
1063
1064         snprintf(dev_buf, sizeof(dev_buf), "%s/class/net/bonding_masters", sysfs_path);
1065         snprintf(buf, sizeof(buf), "%c%s", op, ifname);
1066         write_file(dev_buf, buf);
1067
1068         if (!cfg)
1069                 return 0;
1070
1071         system_set_dev_sysfs("bonding/mode", ifname, bonding_policy_str[cfg->policy]);
1072
1073         system_set_dev_sysfs_int("bonding/all_ports_active", ifname, cfg->all_ports_active);
1074
1075         if (cfg->policy == BONDING_MODE_BALANCE_XOR ||
1076             cfg->policy == BONDING_MODE_BALANCE_TLB ||
1077             cfg->policy == BONDING_MODE_8023AD)
1078                 system_set_dev_sysfs("bonding/xmit_hash_policy", ifname, cfg->xmit_hash_policy);
1079
1080         if (cfg->policy == BONDING_MODE_8023AD) {
1081                 system_set_dev_sysfs("bonding/ad_actor_system", ifname, cfg->ad_actor_system);
1082                 system_set_dev_sysfs_int("bonding/ad_actor_sys_prio", ifname, cfg->ad_actor_sys_prio);
1083                 system_set_dev_sysfs("bonding/ad_select", ifname, cfg->ad_select);
1084                 system_set_dev_sysfs("bonding/lacp_rate", ifname, cfg->lacp_rate);
1085                 system_set_dev_sysfs_int("bonding/min_links", ifname, cfg->min_links);
1086         }
1087
1088         if (cfg->policy == BONDING_MODE_BALANCE_RR)
1089                 system_set_dev_sysfs_int("bonding/packets_per_slave", ifname, cfg->packets_per_port);
1090
1091         if (cfg->policy == BONDING_MODE_BALANCE_TLB ||
1092             cfg->policy == BONDING_MODE_BALANCE_ALB)
1093                 system_set_dev_sysfs_int("bonding/lp_interval", ifname, cfg->lp_interval);
1094
1095         if (cfg->policy == BONDING_MODE_BALANCE_TLB)
1096                 system_set_dev_sysfs_int("bonding/tlb_dynamic_lb", ifname, cfg->dynamic_lb);
1097         system_set_dev_sysfs_int("bonding/resend_igmp", ifname, cfg->resend_igmp);
1098         system_set_dev_sysfs_int("bonding/num_grat_arp", ifname, cfg->num_peer_notif);
1099         system_set_dev_sysfs("bonding/primary_reselect", ifname, cfg->primary_reselect);
1100         system_set_dev_sysfs("bonding/fail_over_mac", ifname, cfg->failover_mac);
1101
1102         system_set_dev_sysfs_int((cfg->monitor_arp ?
1103                                   "bonding/arp_interval" :
1104                                   "bonding/miimon"), ifname, cfg->monitor_interval);
1105
1106         blobmsg_for_each_attr(cur, cfg->arp_target, rem) {
1107                 snprintf(buf, sizeof(buf), "+%s", blobmsg_get_string(cur));
1108                 system_set_dev_sysfs("bonding/arp_ip_target", ifname, buf);
1109         }
1110
1111         system_set_dev_sysfs_int("bonding/arp_all_targets", ifname, cfg->arp_all_targets);
1112         if (cfg->policy < BONDING_MODE_8023AD)
1113                 system_set_dev_sysfs("bonding/arp_validate", ifname, cfg->arp_validate);
1114         system_set_dev_sysfs_int("bonding/use_carrier", ifname, cfg->use_carrier);
1115         if (!cfg->monitor_arp && cfg->monitor_interval) {
1116                 system_set_dev_sysfs_int("bonding/updelay", ifname, cfg->updelay);
1117                 system_set_dev_sysfs_int("bonding/downdelay", ifname, cfg->downdelay);
1118         }
1119
1120         return 0;
1121 }
1122
1123 int system_bonding_set_port(struct device *dev, struct device *port, bool add, bool primary)
1124 {
1125         const char *port_name = port->ifname;
1126         const char op_ch = add ? '+' : '-';
1127         char buf[IFNAMSIZ + 1];
1128
1129         snprintf(buf, sizeof(buf), "%c%s", op_ch, port_name);
1130         system_if_down(port);
1131         system_set_dev_sysfs("bonding/slaves", dev->ifname, buf);
1132         system_if_up(port);
1133
1134         if (primary)
1135                 system_set_dev_sysfs("bonding/primary", dev->ifname,
1136                                      add ? port_name : "");
1137
1138         return 0;
1139 }
1140
1141 int system_if_resolve(struct device *dev)
1142 {
1143         struct ifreq ifr;
1144
1145         memset(&ifr, 0, sizeof(ifr));
1146         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1147         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
1148                 return ifr.ifr_ifindex;
1149         else
1150                 return 0;
1151 }
1152
1153 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
1154 {
1155         struct ifreq ifr;
1156
1157         memset(&ifr, 0, sizeof(ifr));
1158         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
1159         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) < 0)
1160                 return -1;
1161
1162         ifr.ifr_flags |= add;
1163         ifr.ifr_flags &= ~rem;
1164         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
1165 }
1166
1167 struct clear_data {
1168         struct nl_msg *msg;
1169         struct device *dev;
1170         int type;
1171         int size;
1172         int af;
1173 };
1174
1175
1176 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
1177 {
1178         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
1179
1180         return (long)ifa->ifa_index == ifindex;
1181 }
1182
1183 static bool check_route(struct nlmsghdr *hdr, int ifindex)
1184 {
1185         struct rtmsg *r = NLMSG_DATA(hdr);
1186         struct nlattr *tb[__RTA_MAX];
1187
1188         if (r->rtm_protocol == RTPROT_KERNEL &&
1189             r->rtm_family == AF_INET6)
1190                 return false;
1191
1192         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
1193         if (!tb[RTA_OIF])
1194                 return false;
1195
1196         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
1197 }
1198
1199 static bool check_rule(struct nlmsghdr *hdr, int ifindex)
1200 {
1201         return true;
1202 }
1203
1204 static int cb_clear_event(struct nl_msg *msg, void *arg)
1205 {
1206         struct clear_data *clr = arg;
1207         struct nlmsghdr *hdr = nlmsg_hdr(msg);
1208         bool (*cb)(struct nlmsghdr *, int ifindex);
1209         int type, ret;
1210
1211         switch(clr->type) {
1212         case RTM_GETADDR:
1213                 type = RTM_DELADDR;
1214                 if (hdr->nlmsg_type != RTM_NEWADDR)
1215                         return NL_SKIP;
1216
1217                 cb = check_ifaddr;
1218                 break;
1219         case RTM_GETROUTE:
1220                 type = RTM_DELROUTE;
1221                 if (hdr->nlmsg_type != RTM_NEWROUTE)
1222                         return NL_SKIP;
1223
1224                 cb = check_route;
1225                 break;
1226         case RTM_GETRULE:
1227                 type = RTM_DELRULE;
1228                 if (hdr->nlmsg_type != RTM_NEWRULE)
1229                         return NL_SKIP;
1230
1231                 cb = check_rule;
1232                 break;
1233         default:
1234                 return NL_SKIP;
1235         }
1236
1237         if (!cb(hdr, clr->dev ? clr->dev->ifindex : 0))
1238                 return NL_SKIP;
1239
1240         if (type == RTM_DELRULE)
1241                 D(SYSTEM, "Remove a rule\n");
1242         else
1243                 D(SYSTEM, "Remove %s from device %s\n",
1244                   type == RTM_DELADDR ? "an address" : "a route",
1245                   clr->dev->ifname);
1246
1247         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
1248         hdr = nlmsg_hdr(clr->msg);
1249         hdr->nlmsg_type = type;
1250         hdr->nlmsg_flags = NLM_F_REQUEST;
1251
1252         nl_socket_disable_auto_ack(sock_rtnl);
1253         ret = nl_send_auto_complete(sock_rtnl, clr->msg);
1254         if (ret < 0) {
1255                 if (type == RTM_DELRULE)
1256                         D(SYSTEM, "Error deleting a rule: %d\n", ret);
1257                 else
1258                         D(SYSTEM, "Error deleting %s from device '%s': %d\n",
1259                                 type == RTM_DELADDR ? "an address" : "a route",
1260                                 clr->dev->ifname, ret);
1261         }
1262
1263         nl_socket_enable_auto_ack(sock_rtnl);
1264
1265         return NL_SKIP;
1266 }
1267
1268 static int
1269 cb_finish_event(struct nl_msg *msg, void *arg)
1270 {
1271         int *pending = arg;
1272         *pending = 0;
1273         return NL_STOP;
1274 }
1275
1276 static int
1277 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1278 {
1279         int *pending = arg;
1280         *pending = err->error;
1281         return NL_STOP;
1282 }
1283
1284 static void
1285 system_if_clear_entries(struct device *dev, int type, int af)
1286 {
1287         struct clear_data clr;
1288         struct nl_cb *cb;
1289         struct rtmsg rtm = {
1290                 .rtm_family = af,
1291                 .rtm_flags = RTM_F_CLONED,
1292         };
1293         int flags = NLM_F_DUMP;
1294         int pending = 1;
1295
1296         clr.af = af;
1297         clr.dev = dev;
1298         clr.type = type;
1299         switch (type) {
1300         case RTM_GETADDR:
1301         case RTM_GETRULE:
1302                 clr.size = sizeof(struct rtgenmsg);
1303                 break;
1304         case RTM_GETROUTE:
1305                 clr.size = sizeof(struct rtmsg);
1306                 break;
1307         default:
1308                 return;
1309         }
1310
1311         cb = nl_cb_alloc(NL_CB_DEFAULT);
1312         if (!cb)
1313                 return;
1314
1315         clr.msg = nlmsg_alloc_simple(type, flags);
1316         if (!clr.msg)
1317                 goto out;
1318
1319         nlmsg_append(clr.msg, &rtm, clr.size, 0);
1320         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
1321         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
1322         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
1323
1324         if (nl_send_auto_complete(sock_rtnl, clr.msg) < 0)
1325                 goto free;
1326
1327         while (pending > 0)
1328                 nl_recvmsgs(sock_rtnl, cb);
1329
1330 free:
1331         nlmsg_free(clr.msg);
1332 out:
1333         nl_cb_put(cb);
1334 }
1335
1336 /*
1337  * Clear bridge (membership) state and bring down device
1338  */
1339 void system_if_clear_state(struct device *dev)
1340 {
1341         static char buf[256];
1342         char *bridge;
1343         device_set_ifindex(dev, system_if_resolve(dev));
1344
1345         if (dev->external || !dev->ifindex)
1346                 return;
1347
1348         system_if_flags(dev->ifname, 0, IFF_UP);
1349
1350         if (system_is_bridge(dev->ifname)) {
1351                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
1352                 system_bridge_delbr(dev);
1353                 return;
1354         }
1355
1356         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
1357         if (bridge) {
1358                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
1359                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
1360         }
1361
1362         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
1363         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
1364         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
1365         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
1366         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET);
1367         system_if_clear_entries(dev, RTM_GETNEIGH, AF_INET6);
1368         system_set_disable_ipv6(dev, "0");
1369 }
1370
1371 static inline unsigned long
1372 sec_to_jiffies(int val)
1373 {
1374         return (unsigned long) val * 100;
1375 }
1376
1377 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
1378 {
1379         struct nlattr *linkinfo, *data;
1380         struct nl_msg *msg;
1381         uint64_t val;
1382         int rv;
1383
1384         msg = system_ifinfo_msg(bridge->ifname, RTM_NEWLINK, NLM_F_CREATE | NLM_F_EXCL);
1385         if (!msg)
1386                 return -1;
1387
1388         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1389                 goto nla_put_failure;
1390
1391         nla_put_string(msg, IFLA_INFO_KIND, "bridge");
1392
1393         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1394                 goto nla_put_failure;
1395
1396         nla_put_u32(msg, IFLA_BR_STP_STATE, cfg->stp);
1397         nla_put_u32(msg, IFLA_BR_FORWARD_DELAY, sec_to_jiffies(cfg->forward_delay));
1398         nla_put_u8(msg, IFLA_BR_MCAST_SNOOPING, !!cfg->igmp_snoop);
1399         nla_put_u8(msg, IFLA_BR_MCAST_QUERIER, !!cfg->multicast_querier);
1400         nla_put_u32(msg, IFLA_BR_MCAST_HASH_MAX, cfg->hash_max);
1401
1402         if (bridge->settings.flags & DEV_OPT_MULTICAST_ROUTER)
1403                 nla_put_u8(msg, IFLA_BR_MCAST_ROUTER, !!bridge->settings.multicast_router);
1404
1405         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS) {
1406                 nla_put_u32(msg, IFLA_BR_MCAST_STARTUP_QUERY_CNT, cfg->robustness);
1407                 nla_put_u32(msg, IFLA_BR_MCAST_LAST_MEMBER_CNT, cfg->robustness);
1408         }
1409
1410         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL)
1411                 nla_put_u64(msg, IFLA_BR_MCAST_QUERY_INTVL, cfg->query_interval);
1412
1413         if (cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL)
1414                 nla_put_u64(msg, IFLA_BR_MCAST_QUERY_RESPONSE_INTVL, cfg->query_response_interval);
1415
1416         if (cfg->flags & BRIDGE_OPT_LAST_MEMBER_INTERVAL)
1417                 nla_put_u64(msg, IFLA_BR_MCAST_LAST_MEMBER_INTVL, cfg->last_member_interval);
1418
1419         if (cfg->flags & BRIDGE_OPT_ROBUSTNESS ||
1420             cfg->flags & BRIDGE_OPT_QUERY_INTERVAL ||
1421             cfg->flags & BRIDGE_OPT_QUERY_RESPONSE_INTERVAL) {
1422                 val = cfg->robustness * cfg->query_interval +
1423                         cfg->query_response_interval;
1424
1425                 nla_put_u64(msg, IFLA_BR_MCAST_MEMBERSHIP_INTVL, val);
1426
1427                 val -= cfg->query_response_interval / 2;
1428
1429                 nla_put_u64(msg, IFLA_BR_MCAST_QUERIER_INTVL, val);
1430         }
1431
1432         if (cfg->flags & BRIDGE_OPT_QUERY_INTERVAL) {
1433                 val = cfg->query_interval / 4;
1434
1435                 nla_put_u64(msg, IFLA_BR_MCAST_STARTUP_QUERY_INTVL, val);
1436         }
1437
1438         nla_put_u8(msg, IFLA_BR_VLAN_FILTERING, !!cfg->vlan_filtering);
1439         nla_put_u16(msg, IFLA_BR_PRIORITY, cfg->priority);
1440         nla_put_u32(msg, IFLA_BR_HELLO_TIME, sec_to_jiffies(cfg->hello_time));
1441         nla_put_u32(msg, IFLA_BR_MAX_AGE, sec_to_jiffies(cfg->max_age));
1442
1443         if (cfg->flags & BRIDGE_OPT_AGEING_TIME)
1444                 nla_put_u32(msg, IFLA_BR_AGEING_TIME, sec_to_jiffies(cfg->ageing_time));
1445
1446         nla_nest_end(msg, data);
1447         nla_nest_end(msg, linkinfo);
1448
1449         rv = system_rtnl_call(msg);
1450         if (rv)
1451                 D(SYSTEM, "Error adding bridge '%s': %d\n", bridge->ifname, rv);
1452
1453         return rv;
1454
1455 nla_put_failure:
1456         nlmsg_free(msg);
1457         return -ENOMEM;
1458 }
1459
1460 int system_macvlan_add(struct device *macvlan, struct device *dev, struct macvlan_config *cfg)
1461 {
1462         struct nl_msg *msg;
1463         struct nlattr *linkinfo, *data;
1464         size_t i;
1465         int rv;
1466         static const struct {
1467                 const char *name;
1468                 enum macvlan_mode val;
1469         } modes[] = {
1470                 { "private", MACVLAN_MODE_PRIVATE },
1471                 { "vepa", MACVLAN_MODE_VEPA },
1472                 { "bridge", MACVLAN_MODE_BRIDGE },
1473                 { "passthru", MACVLAN_MODE_PASSTHRU },
1474         };
1475
1476         msg = system_ifinfo_msg(macvlan->ifname, RTM_NEWLINK, NLM_F_CREATE | NLM_F_EXCL);
1477         if (!msg)
1478                 return -1;
1479
1480         if (cfg->flags & MACVLAN_OPT_MACADDR)
1481                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1482         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1483
1484         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1485                 goto nla_put_failure;
1486
1487         nla_put_string(msg, IFLA_INFO_KIND, "macvlan");
1488
1489         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1490                 goto nla_put_failure;
1491
1492         if (cfg->mode) {
1493                 for (i = 0; i < ARRAY_SIZE(modes); i++) {
1494                         if (strcmp(cfg->mode, modes[i].name) != 0)
1495                                 continue;
1496
1497                         nla_put_u32(msg, IFLA_MACVLAN_MODE, modes[i].val);
1498                         break;
1499                 }
1500         }
1501
1502         nla_nest_end(msg, data);
1503         nla_nest_end(msg, linkinfo);
1504
1505         rv = system_rtnl_call(msg);
1506         if (rv)
1507                 D(SYSTEM, "Error adding macvlan '%s' over '%s': %d\n", macvlan->ifname, dev->ifname, rv);
1508
1509         return rv;
1510
1511 nla_put_failure:
1512         nlmsg_free(msg);
1513         return -ENOMEM;
1514 }
1515
1516 int system_link_netns_move(struct device *dev, int netns_fd, const char *target_ifname)
1517 {
1518         struct nl_msg *msg;
1519         int index;
1520
1521         if (!dev)
1522                 return -1;
1523
1524         index = system_if_resolve(dev);
1525         msg = __system_ifinfo_msg(AF_UNSPEC, index, target_ifname, RTM_NEWLINK, 0);
1526         if (!msg)
1527                 return -1;
1528
1529         nla_put_u32(msg, IFLA_NET_NS_FD, netns_fd);
1530         return system_rtnl_call(msg);
1531 }
1532
1533 int system_macvlan_del(struct device *macvlan)
1534 {
1535         return system_link_del(macvlan->ifname);
1536 }
1537
1538 int system_netns_open(const pid_t target_ns)
1539 {
1540         char pid_net_path[PATH_MAX];
1541
1542         snprintf(pid_net_path, sizeof(pid_net_path), "/proc/%u/ns/net", target_ns);
1543
1544         return open(pid_net_path, O_RDONLY);
1545 }
1546
1547 int system_netns_set(int netns_fd)
1548 {
1549         return setns(netns_fd, CLONE_NEWNET);
1550 }
1551
1552 int system_veth_add(struct device *veth, struct veth_config *cfg)
1553 {
1554         struct nl_msg *msg;
1555         struct ifinfomsg empty_iim = {0,};
1556         struct nlattr *linkinfo, *data, *veth_info;
1557         int rv;
1558
1559         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1560
1561         if (!msg)
1562                 return -1;
1563
1564         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1565
1566         if (cfg->flags & VETH_OPT_MACADDR)
1567                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->macaddr), cfg->macaddr);
1568         nla_put_string(msg, IFLA_IFNAME, veth->ifname);
1569
1570         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1571                 goto nla_put_failure;
1572
1573         nla_put_string(msg, IFLA_INFO_KIND, "veth");
1574
1575         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1576                 goto nla_put_failure;
1577
1578         if (!(veth_info = nla_nest_start(msg, VETH_INFO_PEER)))
1579                 goto nla_put_failure;
1580
1581         nlmsg_append(msg, &empty_iim, sizeof(empty_iim), 0);
1582
1583         if (cfg->flags & VETH_OPT_PEER_NAME)
1584                 nla_put_string(msg, IFLA_IFNAME, cfg->peer_name);
1585         if (cfg->flags & VETH_OPT_PEER_MACADDR)
1586                 nla_put(msg, IFLA_ADDRESS, sizeof(cfg->peer_macaddr), cfg->peer_macaddr);
1587
1588         nla_nest_end(msg, veth_info);
1589         nla_nest_end(msg, data);
1590         nla_nest_end(msg, linkinfo);
1591
1592         rv = system_rtnl_call(msg);
1593         if (rv) {
1594                 if (cfg->flags & VETH_OPT_PEER_NAME)
1595                         D(SYSTEM, "Error adding veth '%s' with peer '%s': %d\n", veth->ifname, cfg->peer_name, rv);
1596                 else
1597                         D(SYSTEM, "Error adding veth '%s': %d\n", veth->ifname, rv);
1598         }
1599
1600         return rv;
1601
1602 nla_put_failure:
1603         nlmsg_free(msg);
1604         return -ENOMEM;
1605 }
1606
1607 int system_veth_del(struct device *veth)
1608 {
1609         return system_link_del(veth->ifname);
1610 }
1611
1612 static int system_vlan(struct device *dev, int id)
1613 {
1614         struct vlan_ioctl_args ifr = {
1615                 .cmd = SET_VLAN_NAME_TYPE_CMD,
1616                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
1617         };
1618
1619         if (ioctl(sock_ioctl, SIOCSIFVLAN, &ifr) < 0)
1620                 return -1;
1621
1622         if (id < 0) {
1623                 ifr.cmd = DEL_VLAN_CMD;
1624                 ifr.u.VID = 0;
1625         } else {
1626                 ifr.cmd = ADD_VLAN_CMD;
1627                 ifr.u.VID = id;
1628         }
1629         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
1630         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
1631 }
1632
1633 int system_vlan_add(struct device *dev, int id)
1634 {
1635         return system_vlan(dev, id);
1636 }
1637
1638 int system_vlan_del(struct device *dev)
1639 {
1640         return system_vlan(dev, -1);
1641 }
1642
1643 int system_vlandev_add(struct device *vlandev, struct device *dev, struct vlandev_config *cfg)
1644 {
1645         struct nl_msg *msg;
1646         struct nlattr *linkinfo, *data, *qos;
1647         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC };
1648         struct vlan_qos_mapping *elem;
1649         struct ifla_vlan_qos_mapping nl_qos_map;
1650         int rv;
1651
1652         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
1653
1654         if (!msg)
1655                 return -1;
1656
1657         nlmsg_append(msg, &iim, sizeof(iim), 0);
1658         nla_put_string(msg, IFLA_IFNAME, vlandev->ifname);
1659         nla_put_u32(msg, IFLA_LINK, dev->ifindex);
1660
1661         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO)))
1662                 goto nla_put_failure;
1663
1664         nla_put_string(msg, IFLA_INFO_KIND, "vlan");
1665
1666         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA)))
1667                 goto nla_put_failure;
1668
1669         nla_put_u16(msg, IFLA_VLAN_ID, cfg->vid);
1670
1671 #if LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)
1672         nla_put_u16(msg, IFLA_VLAN_PROTOCOL, htons(cfg->proto));
1673 #else
1674         if(cfg->proto == VLAN_PROTO_8021AD)
1675                 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);
1676 #endif
1677
1678         if (!(qos = nla_nest_start(msg, IFLA_VLAN_INGRESS_QOS)))
1679                 goto nla_put_failure;
1680
1681         vlist_simple_for_each_element(&cfg->ingress_qos_mapping_list, elem, node) {
1682                 nl_qos_map.from = elem->from;
1683                 nl_qos_map.to = elem->to;
1684                 nla_put(msg, IFLA_VLAN_QOS_MAPPING, sizeof(nl_qos_map), &nl_qos_map);
1685         }
1686         nla_nest_end(msg, qos);
1687
1688         if (!(qos = nla_nest_start(msg, IFLA_VLAN_EGRESS_QOS)))
1689                 goto nla_put_failure;
1690
1691         vlist_simple_for_each_element(&cfg->egress_qos_mapping_list, elem, node) {
1692                 nl_qos_map.from = elem->from;
1693                 nl_qos_map.to = elem->to;
1694                 nla_put(msg, IFLA_VLAN_QOS_MAPPING, sizeof(nl_qos_map), &nl_qos_map);
1695         }
1696         nla_nest_end(msg, qos);
1697
1698         nla_nest_end(msg, data);
1699         nla_nest_end(msg, linkinfo);
1700
1701         rv = system_rtnl_call(msg);
1702         if (rv)
1703                 D(SYSTEM, "Error adding vlandev '%s' over '%s': %d\n", vlandev->ifname, dev->ifname, rv);
1704
1705         return rv;
1706
1707 nla_put_failure:
1708         nlmsg_free(msg);
1709         return -ENOMEM;
1710 }
1711
1712 int system_vlandev_del(struct device *vlandev)
1713 {
1714         return system_link_del(vlandev->ifname);
1715 }
1716
1717 #if LINUX_VERSION_CODE >= KERNEL_VERSION(6,1,0)
1718 struct if_get_master_data {
1719         int ifindex;
1720         int master_ifindex;
1721         int pending;
1722 };
1723
1724 static void if_get_master_dsa_linkinfo_attr(struct if_get_master_data *data,
1725                                struct rtattr *attr)
1726 {
1727         struct rtattr *cur;
1728         int rem = RTA_PAYLOAD(attr);
1729
1730         for (cur = RTA_DATA(attr); RTA_OK(cur, rem); cur = RTA_NEXT(cur, rem)) {
1731                 if (cur->rta_type != IFLA_DSA_MASTER)
1732                         continue;
1733
1734                 data->master_ifindex = *(__u32 *)RTA_DATA(cur);
1735         }
1736 }
1737
1738 static void if_get_master_linkinfo_attr(struct if_get_master_data *data,
1739                                struct rtattr *attr)
1740 {
1741         struct rtattr *cur;
1742         int rem = RTA_PAYLOAD(attr);
1743
1744         for (cur = RTA_DATA(attr); RTA_OK(cur, rem); cur = RTA_NEXT(cur, rem)) {
1745                 if (cur->rta_type != IFLA_INFO_KIND && cur->rta_type != IFLA_INFO_DATA)
1746                         continue;
1747
1748                 if (cur->rta_type == IFLA_INFO_KIND && strcmp("dsa", (char *)RTA_DATA(cur)))
1749                         break;
1750
1751                 if (cur->rta_type == IFLA_INFO_DATA)
1752                         if_get_master_dsa_linkinfo_attr(data, cur);
1753         }
1754 }
1755
1756 static int cb_if_get_master_valid(struct nl_msg *msg, void *arg)
1757 {
1758         struct nlmsghdr *nh = nlmsg_hdr(msg);
1759         struct ifinfomsg *ifi = NLMSG_DATA(nh);
1760         struct if_get_master_data *data = (struct if_get_master_data *)arg;
1761         struct rtattr *attr;
1762         int rem;
1763
1764         if (nh->nlmsg_type != RTM_NEWLINK)
1765                 return NL_SKIP;
1766
1767         if (ifi->ifi_family != AF_UNSPEC)
1768                 return NL_SKIP;
1769
1770         if (ifi->ifi_index != data->ifindex)
1771                 return NL_SKIP;
1772
1773         attr = IFLA_RTA(ifi);
1774         rem = nh->nlmsg_len - NLMSG_LENGTH(sizeof(*ifi));
1775
1776         while (RTA_OK(attr, rem)) {
1777                 if (attr->rta_type == IFLA_LINKINFO)
1778                         if_get_master_linkinfo_attr(data, attr);
1779
1780                 attr = RTA_NEXT(attr, rem);
1781         }
1782
1783         return NL_OK;
1784 }
1785
1786 static int cb_if_get_master_ack(struct nl_msg *msg, void *arg)
1787 {
1788         struct if_get_master_data *data = (struct if_get_master_data *)arg;
1789         data->pending = 0;
1790         return NL_STOP;
1791 }
1792
1793 static int cb_if_get_master_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
1794 {
1795         struct if_get_master_data *data = (struct if_get_master_data *)arg;
1796         data->pending = 0;
1797         return NL_STOP;
1798 }
1799
1800 static int system_if_get_master_ifindex(struct device *dev)
1801 {
1802         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
1803         struct nl_msg *msg;
1804         struct ifinfomsg ifi = {
1805                 .ifi_family = AF_UNSPEC,
1806                 .ifi_index = 0,
1807         };
1808         struct if_get_master_data data = {
1809                 .ifindex = if_nametoindex(dev->ifname),
1810                 .master_ifindex = -1,
1811                 .pending = 1,
1812         };
1813         int ret = -1;
1814
1815         if (!cb)
1816                 return ret;
1817
1818         msg = nlmsg_alloc_simple(RTM_GETLINK, NLM_F_REQUEST);
1819         if (!msg)
1820                 goto out;
1821
1822         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
1823             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
1824                 goto free;
1825
1826         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_get_master_valid, &data);
1827         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_get_master_ack, &data);
1828         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_get_master_error, &data);
1829
1830         ret = nl_send_auto_complete(sock_rtnl, msg);
1831         if (ret < 0)
1832                 goto free;
1833
1834         while (data.pending > 0)
1835                 nl_recvmsgs(sock_rtnl, cb);
1836
1837         if (data.master_ifindex >= 0)
1838                 ret = data.master_ifindex;
1839
1840 free:
1841         nlmsg_free(msg);
1842 out:
1843         nl_cb_put(cb);
1844         return ret;
1845 }
1846
1847 static void system_set_master(struct device *dev, int master_ifindex)
1848 {
1849         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
1850         struct nl_msg *nlm;
1851
1852         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST);
1853         if (!nlm)
1854                 return;
1855
1856         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
1857         nla_put_string(nlm, IFLA_IFNAME, dev->ifname);
1858
1859         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
1860         if (!linkinfo)
1861                 goto failure;
1862
1863         nla_put_string(nlm, IFLA_INFO_KIND, "dsa");
1864         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
1865         if (!infodata)
1866                 goto failure;
1867
1868         nla_put_u32(nlm, IFLA_DSA_MASTER, master_ifindex);
1869
1870         nla_nest_end(nlm, infodata);
1871         nla_nest_end(nlm, linkinfo);
1872
1873         system_rtnl_call(nlm);
1874
1875         return;
1876
1877 failure:
1878         nlmsg_free(nlm);
1879 }
1880 #endif
1881
1882 static void ethtool_link_mode_clear_bit(__s8 nwords, int nr, __u32 *mask)
1883 {
1884         if (nr < 0)
1885                 return;
1886
1887         if (nr >= (nwords * 32))
1888                 return;
1889
1890         mask[nr / 32] &= ~(1U << (nr % 32));
1891 }
1892
1893 static bool ethtool_link_mode_test_bit(__s8 nwords, int nr, const __u32 *mask)
1894 {
1895         if (nr < 0)
1896                 return false;
1897
1898         if (nr >= (nwords * 32))
1899                 return false;
1900
1901         return !!(mask[nr / 32] & (1U << (nr % 32)));
1902 }
1903
1904 static int
1905 system_get_ethtool_gro(struct device *dev)
1906 {
1907         struct ethtool_value ecmd;
1908         struct ifreq ifr = {
1909                 .ifr_data = (caddr_t)&ecmd,
1910         };
1911
1912         memset(&ecmd, 0, sizeof(ecmd));
1913         ecmd.cmd = ETHTOOL_GGRO;
1914         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1915
1916         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr))
1917                 return -1;
1918
1919         return ecmd.data;
1920 }
1921
1922 static void
1923 system_set_ethtool_gro(struct device *dev, struct device_settings *s)
1924 {
1925         struct ethtool_value ecmd;
1926         struct ifreq ifr = {
1927                 .ifr_data = (caddr_t)&ecmd,
1928         };
1929
1930         memset(&ecmd, 0, sizeof(ecmd));
1931         ecmd.cmd = ETHTOOL_SGRO;
1932         ecmd.data = s->gro;
1933         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1934
1935         ioctl(sock_ioctl, SIOCETHTOOL, &ifr);
1936 }
1937
1938 static void
1939 system_set_ethtool_pause(struct device *dev, struct device_settings *s)
1940 {
1941         struct ethtool_pauseparam pp;
1942         struct ifreq ifr = {
1943                 .ifr_data = (caddr_t)&pp,
1944         };
1945
1946         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
1947         memset(&pp, 0, sizeof(pp));
1948         pp.cmd = ETHTOOL_GPAUSEPARAM;
1949         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr))
1950                 return;
1951
1952         if (s->flags & DEV_OPT_RXPAUSE || s->flags & DEV_OPT_TXPAUSE) {
1953                 pp.autoneg = AUTONEG_DISABLE;
1954
1955                 if (s->flags & DEV_OPT_PAUSE) {
1956                         if (s->flags & DEV_OPT_RXPAUSE)
1957                                 pp.rx_pause = s->rxpause && s->pause;
1958                         else
1959                                 pp.rx_pause = s->pause;
1960
1961                         if (s->flags & DEV_OPT_TXPAUSE)
1962                                 pp.tx_pause = s->txpause && s->pause;
1963                         else
1964                                 pp.tx_pause = s->pause;
1965                 } else {
1966                         if (s->flags & DEV_OPT_RXPAUSE)
1967                                 pp.rx_pause = s->rxpause;
1968
1969                         if (s->flags & DEV_OPT_TXPAUSE)
1970                                 pp.tx_pause = s->txpause;
1971                 }
1972
1973                 if (s->flags & DEV_OPT_ASYM_PAUSE &&
1974                     !s->asym_pause && (pp.rx_pause != pp.tx_pause))
1975                         pp.rx_pause = pp.tx_pause = false;
1976         } else {
1977                 pp.autoneg = AUTONEG_ENABLE;
1978                 /* Pause and Asym_Pause advertising bits will be set via
1979                  * ETHTOOL_SLINKSETTINGS in system_set_ethtool_settings()
1980                  */
1981         }
1982
1983         pp.cmd = ETHTOOL_SPAUSEPARAM;
1984         ioctl(sock_ioctl, SIOCETHTOOL, &ifr);
1985 }
1986
1987 static void
1988 system_set_ethtool_settings(struct device *dev, struct device_settings *s)
1989 {
1990         struct {
1991                 struct ethtool_link_settings req;
1992                 __u32 link_mode_data[3 * 127];
1993         } ecmd;
1994         struct ifreq ifr = {
1995                 .ifr_data = (caddr_t)&ecmd,
1996         };
1997         size_t i;
1998         __s8 nwords;
1999         __u32 *supported, *advertising;
2000
2001         system_set_ethtool_pause(dev, s);
2002
2003         memset(&ecmd, 0, sizeof(ecmd));
2004         ecmd.req.cmd = ETHTOOL_GLINKSETTINGS;
2005         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
2006
2007         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) < 0 ||
2008             ecmd.req.link_mode_masks_nwords >= 0 ||
2009             ecmd.req.cmd != ETHTOOL_GLINKSETTINGS)
2010                 return;
2011
2012         ecmd.req.link_mode_masks_nwords = -ecmd.req.link_mode_masks_nwords;
2013
2014         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) < 0 ||
2015             ecmd.req.link_mode_masks_nwords <= 0 ||
2016             ecmd.req.cmd != ETHTOOL_GLINKSETTINGS)
2017                 return;
2018
2019         nwords = ecmd.req.link_mode_masks_nwords;
2020         supported = &ecmd.link_mode_data[0];
2021         advertising = &ecmd.link_mode_data[nwords];
2022         memcpy(advertising, supported, sizeof(__u32) * nwords);
2023
2024         for (i = 0; i < ARRAY_SIZE(ethtool_modes); i++) {
2025                 if (s->flags & DEV_OPT_DUPLEX) {
2026                         if (s->duplex)
2027                                 ethtool_link_mode_clear_bit(nwords, ethtool_modes[i].bit_half, advertising);
2028                         else
2029                                 ethtool_link_mode_clear_bit(nwords, ethtool_modes[i].bit_full, advertising);
2030                 }
2031                 if (!(s->flags & DEV_OPT_SPEED) ||
2032                     s->speed == ethtool_modes[i].speed)
2033                         continue;
2034
2035                 ethtool_link_mode_clear_bit(nwords, ethtool_modes[i].bit_full, advertising);
2036                 ethtool_link_mode_clear_bit(nwords, ethtool_modes[i].bit_half, advertising);
2037         }
2038
2039         if (s->flags & DEV_OPT_PAUSE)
2040                 if (!s->pause)
2041                         ethtool_link_mode_clear_bit(nwords, ETHTOOL_LINK_MODE_Pause_BIT, advertising);
2042
2043         if (s->flags & DEV_OPT_ASYM_PAUSE)
2044                 if (!s->asym_pause)
2045                         ethtool_link_mode_clear_bit(nwords, ETHTOOL_LINK_MODE_Asym_Pause_BIT, advertising);
2046
2047         if (s->flags & DEV_OPT_AUTONEG) {
2048                 ecmd.req.autoneg = s->autoneg ? AUTONEG_ENABLE : AUTONEG_DISABLE;
2049                 if (!s->autoneg) {
2050                         if (s->flags & DEV_OPT_SPEED)
2051                                 ecmd.req.speed = s->speed;
2052
2053                         if (s->flags & DEV_OPT_DUPLEX)
2054                                 ecmd.req.duplex = s->duplex ? DUPLEX_FULL : DUPLEX_HALF;
2055                 }
2056         }
2057
2058         ecmd.req.cmd = ETHTOOL_SLINKSETTINGS;
2059         ioctl(sock_ioctl, SIOCETHTOOL, &ifr);
2060 }
2061
2062 static void
2063 system_set_ethtool_settings_after_up(struct device *dev, struct device_settings *s)
2064 {
2065         if (s->flags & DEV_OPT_GRO)
2066                 system_set_ethtool_gro(dev, s);
2067 }
2068
2069 void
2070 system_if_get_settings(struct device *dev, struct device_settings *s)
2071 {
2072         struct ifreq ifr;
2073         char buf[10];
2074         int ret;
2075
2076         memset(&ifr, 0, sizeof(ifr));
2077         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
2078
2079         if (ioctl(sock_ioctl, SIOCGIFMTU, &ifr) == 0) {
2080                 s->mtu = ifr.ifr_mtu;
2081                 s->flags |= DEV_OPT_MTU;
2082         }
2083
2084         s->mtu6 = system_update_ipv6_mtu(dev, 0);
2085         if (s->mtu6 > 0)
2086                 s->flags |= DEV_OPT_MTU6;
2087
2088         if (ioctl(sock_ioctl, SIOCGIFTXQLEN, &ifr) == 0) {
2089                 s->txqueuelen = ifr.ifr_qlen;
2090                 s->flags |= DEV_OPT_TXQUEUELEN;
2091         }
2092
2093         if (ioctl(sock_ioctl, SIOCGIFHWADDR, &ifr) == 0) {
2094                 memcpy(s->macaddr, &ifr.ifr_hwaddr.sa_data, sizeof(s->macaddr));
2095                 s->flags |= DEV_OPT_MACADDR;
2096         }
2097
2098         if (!system_get_disable_ipv6(dev, buf, sizeof(buf))) {
2099                 s->ipv6 = !strtoul(buf, NULL, 0);
2100                 s->flags |= DEV_OPT_IPV6;
2101         }
2102
2103         if (!system_get_ip6segmentrouting(dev, buf, sizeof(buf))) {
2104                 s->ip6segmentrouting = strtoul(buf, NULL, 0);
2105                 s->flags |= DEV_OPT_IP6SEGMENTROUTING;
2106         }
2107
2108         if (ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr) == 0) {
2109                 s->promisc = ifr.ifr_flags & IFF_PROMISC;
2110                 s->flags |= DEV_OPT_PROMISC;
2111
2112                 s->multicast = ifr.ifr_flags & IFF_MULTICAST;
2113                 s->flags |= DEV_OPT_MULTICAST;
2114         }
2115
2116         if (!system_get_rpfilter(dev, buf, sizeof(buf))) {
2117                 s->rpfilter = strtoul(buf, NULL, 0);
2118                 s->flags |= DEV_OPT_RPFILTER;
2119         }
2120
2121         if (!system_get_acceptlocal(dev, buf, sizeof(buf))) {
2122                 s->acceptlocal = strtoul(buf, NULL, 0);
2123                 s->flags |= DEV_OPT_ACCEPTLOCAL;
2124         }
2125
2126         if (!system_get_igmpversion(dev, buf, sizeof(buf))) {
2127                 s->igmpversion = strtoul(buf, NULL, 0);
2128                 s->flags |= DEV_OPT_IGMPVERSION;
2129         }
2130
2131         if (!system_get_mldversion(dev, buf, sizeof(buf))) {
2132                 s->mldversion = strtoul(buf, NULL, 0);
2133                 s->flags |= DEV_OPT_MLDVERSION;
2134         }
2135
2136         if (!system_get_neigh4reachabletime(dev, buf, sizeof(buf))) {
2137                 s->neigh4reachabletime = strtoul(buf, NULL, 0);
2138                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
2139         }
2140
2141         if (!system_get_neigh6reachabletime(dev, buf, sizeof(buf))) {
2142                 s->neigh6reachabletime = strtoul(buf, NULL, 0);
2143                 s->flags |= DEV_OPT_NEIGHREACHABLETIME;
2144         }
2145
2146         if (!system_get_neigh4locktime(dev, buf, sizeof(buf))) {
2147                 s->neigh4locktime = strtol(buf, NULL, 0);
2148                 s->flags |= DEV_OPT_NEIGHLOCKTIME;
2149         }
2150
2151         if (!system_get_neigh4gcstaletime(dev, buf, sizeof(buf))) {
2152                 s->neigh4gcstaletime = strtoul(buf, NULL, 0);
2153                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
2154         }
2155
2156         if (!system_get_neigh6gcstaletime(dev, buf, sizeof(buf))) {
2157                 s->neigh6gcstaletime = strtoul(buf, NULL, 0);
2158                 s->flags |= DEV_OPT_NEIGHGCSTALETIME;
2159         }
2160
2161         if (!system_get_dadtransmits(dev, buf, sizeof(buf))) {
2162                 s->dadtransmits = strtoul(buf, NULL, 0);
2163                 s->flags |= DEV_OPT_DADTRANSMITS;
2164         }
2165
2166         if (!system_get_sendredirects(dev, buf, sizeof(buf))) {
2167                 s->sendredirects = strtoul(buf, NULL, 0);
2168                 s->flags |= DEV_OPT_SENDREDIRECTS;
2169         }
2170
2171         if (!system_get_drop_v4_unicast_in_l2_multicast(dev, buf, sizeof(buf))) {
2172                 s->drop_v4_unicast_in_l2_multicast = strtoul(buf, NULL, 0);
2173                 s->flags |= DEV_OPT_DROP_V4_UNICAST_IN_L2_MULTICAST;
2174         }
2175
2176         if (!system_get_drop_v6_unicast_in_l2_multicast(dev, buf, sizeof(buf))) {
2177                 s->drop_v6_unicast_in_l2_multicast = strtoul(buf, NULL, 0);
2178                 s->flags |= DEV_OPT_DROP_V6_UNICAST_IN_L2_MULTICAST;
2179         }
2180
2181         if (!system_get_drop_gratuitous_arp(dev, buf, sizeof(buf))) {
2182                 s->drop_gratuitous_arp = strtoul(buf, NULL, 0);
2183                 s->flags |= DEV_OPT_DROP_GRATUITOUS_ARP;
2184         }
2185
2186         if (!system_get_drop_unsolicited_na(dev, buf, sizeof(buf))) {
2187                 s->drop_unsolicited_na = strtoul(buf, NULL, 0);
2188                 s->flags |= DEV_OPT_DROP_UNSOLICITED_NA;
2189         }
2190
2191         if (!system_get_arp_accept(dev, buf, sizeof(buf))) {
2192                 s->arp_accept = strtoul(buf, NULL, 0);
2193                 s->flags |= DEV_OPT_ARP_ACCEPT;
2194         }
2195
2196         ret = system_get_ethtool_gro(dev);
2197         if (ret >= 0) {
2198                 s->gro = ret;
2199                 s->flags |= DEV_OPT_GRO;
2200         }
2201
2202 #if LINUX_VERSION_CODE >= KERNEL_VERSION(6,1,0)
2203         ret = system_if_get_master_ifindex(dev);
2204         if (ret >= 0) {
2205                 s->master_ifindex = ret;
2206                 s->flags |= DEV_OPT_MASTER;
2207         }
2208 #endif
2209 }
2210
2211 void
2212 system_if_apply_settings(struct device *dev, struct device_settings *s, uint64_t apply_mask)
2213 {
2214         struct ifreq ifr;
2215         char buf[12];
2216
2217         apply_mask &= s->flags;
2218
2219         memset(&ifr, 0, sizeof(ifr));
2220         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
2221         if (apply_mask & DEV_OPT_MTU) {
2222                 ifr.ifr_mtu = s->mtu;
2223                 if (ioctl(sock_ioctl, SIOCSIFMTU, &ifr) < 0)
2224                         s->flags &= ~DEV_OPT_MTU;
2225         }
2226         if (apply_mask & DEV_OPT_MTU6) {
2227                 system_update_ipv6_mtu(dev, s->mtu6);
2228         }
2229         if (apply_mask & DEV_OPT_TXQUEUELEN) {
2230                 ifr.ifr_qlen = s->txqueuelen;
2231                 if (ioctl(sock_ioctl, SIOCSIFTXQLEN, &ifr) < 0)
2232                         s->flags &= ~DEV_OPT_TXQUEUELEN;
2233         }
2234         if ((apply_mask & (DEV_OPT_MACADDR | DEV_OPT_DEFAULT_MACADDR)) && !dev->external) {
2235                 ifr.ifr_hwaddr.sa_family = ARPHRD_ETHER;
2236                 memcpy(&ifr.ifr_hwaddr.sa_data, s->macaddr, sizeof(s->macaddr));
2237                 if (ioctl(sock_ioctl, SIOCSIFHWADDR, &ifr) < 0)
2238                         s->flags &= ~DEV_OPT_MACADDR;
2239         }
2240         if (apply_mask & DEV_OPT_IPV6)
2241                 system_set_disable_ipv6(dev, s->ipv6 ? "0" : "1");
2242         if (s->flags & DEV_OPT_IP6SEGMENTROUTING & apply_mask) {
2243                 struct device dummy = {
2244                         .ifname = "all",
2245                 };
2246                 bool ip6segmentrouting = device_check_ip6segmentrouting();
2247
2248                 system_set_ip6segmentrouting(dev, s->ip6segmentrouting ? "1" : "0");
2249                 system_set_ip6segmentrouting(&dummy, ip6segmentrouting ? "1" : "0");
2250         }
2251         if (apply_mask & DEV_OPT_PROMISC) {
2252                 if (system_if_flags(dev->ifname, s->promisc ? IFF_PROMISC : 0,
2253                                     !s->promisc ? IFF_PROMISC : 0) < 0)
2254                         s->flags &= ~DEV_OPT_PROMISC;
2255         }
2256         if (apply_mask & DEV_OPT_RPFILTER) {
2257                 snprintf(buf, sizeof(buf), "%u", s->rpfilter);
2258                 system_set_rpfilter(dev, buf);
2259         }
2260         if (apply_mask & DEV_OPT_ACCEPTLOCAL)
2261                 system_set_acceptlocal(dev, s->acceptlocal ? "1" : "0");
2262         if (apply_mask & DEV_OPT_IGMPVERSION) {
2263                 snprintf(buf, sizeof(buf), "%u", s->igmpversion);
2264                 system_set_igmpversion(dev, buf);
2265         }
2266         if (apply_mask & DEV_OPT_MLDVERSION) {
2267                 snprintf(buf, sizeof(buf), "%u", s->mldversion);
2268                 system_set_mldversion(dev, buf);
2269         }
2270         if (apply_mask & DEV_OPT_NEIGHREACHABLETIME) {
2271                 snprintf(buf, sizeof(buf), "%u", s->neigh4reachabletime);
2272                 system_set_neigh4reachabletime(dev, buf);
2273                 snprintf(buf, sizeof(buf), "%u", s->neigh6reachabletime);
2274                 system_set_neigh6reachabletime(dev, buf);
2275         }
2276         if (apply_mask & DEV_OPT_NEIGHLOCKTIME) {
2277                 snprintf(buf, sizeof(buf), "%d", s->neigh4locktime);
2278                 system_set_neigh4locktime(dev, buf);
2279         }
2280         if (apply_mask & DEV_OPT_NEIGHGCSTALETIME) {
2281                 snprintf(buf, sizeof(buf), "%u", s->neigh4gcstaletime);
2282                 system_set_neigh4gcstaletime(dev, buf);
2283                 snprintf(buf, sizeof(buf), "%u", s->neigh6gcstaletime);
2284                 system_set_neigh6gcstaletime(dev, buf);
2285         }
2286         if (apply_mask & DEV_OPT_DADTRANSMITS) {
2287                 snprintf(buf, sizeof(buf), "%u", s->dadtransmits);
2288                 system_set_dadtransmits(dev, buf);
2289         }
2290         if (apply_mask & DEV_OPT_MULTICAST) {
2291                 if (system_if_flags(dev->ifname, s->multicast ? IFF_MULTICAST : 0,
2292                                     !s->multicast ? IFF_MULTICAST : 0) < 0)
2293                         s->flags &= ~DEV_OPT_MULTICAST;
2294         }
2295         if (apply_mask & DEV_OPT_SENDREDIRECTS)
2296                 system_set_sendredirects(dev, s->sendredirects ? "1" : "0");
2297         if (apply_mask & DEV_OPT_DROP_V4_UNICAST_IN_L2_MULTICAST)
2298                 system_set_drop_v4_unicast_in_l2_multicast(dev, s->drop_v4_unicast_in_l2_multicast ? "1" : "0");
2299         if (apply_mask & DEV_OPT_DROP_V6_UNICAST_IN_L2_MULTICAST)
2300                 system_set_drop_v6_unicast_in_l2_multicast(dev, s->drop_v6_unicast_in_l2_multicast ? "1" : "0");
2301         if (apply_mask & DEV_OPT_DROP_GRATUITOUS_ARP)
2302                 system_set_drop_gratuitous_arp(dev, s->drop_gratuitous_arp ? "1" : "0");
2303         if (apply_mask & DEV_OPT_DROP_UNSOLICITED_NA)
2304                 system_set_drop_unsolicited_na(dev, s->drop_unsolicited_na ? "1" : "0");
2305         if (apply_mask & DEV_OPT_ARP_ACCEPT)
2306                 system_set_arp_accept(dev, s->arp_accept ? "1" : "0");
2307         if (apply_mask & DEV_OPT_MASTER)
2308 #if LINUX_VERSION_CODE >= KERNEL_VERSION(6,1,0)
2309                 system_set_master(dev, s->master_ifindex);
2310 #else
2311                 netifd_log_message(L_WARNING, "%s Your kernel is older than linux 6.1.0, changing DSA port conduit is not supported!", dev->ifname);
2312 #endif
2313         system_set_ethtool_settings(dev, s);
2314 }
2315
2316 void system_if_apply_settings_after_up(struct device *dev, struct device_settings *s)
2317 {
2318         system_set_ethtool_settings_after_up(dev, s);
2319 }
2320
2321 int system_if_up(struct device *dev)
2322 {
2323         return system_if_flags(dev->ifname, IFF_UP, 0);
2324 }
2325
2326 int system_if_down(struct device *dev)
2327 {
2328         return system_if_flags(dev->ifname, 0, IFF_UP);
2329 }
2330
2331 struct if_check_data {
2332         struct device *dev;
2333         int pending;
2334         int ret;
2335 };
2336
2337 static int cb_if_check_valid(struct nl_msg *msg, void *arg)
2338 {
2339         struct nlmsghdr *nh = nlmsg_hdr(msg);
2340         struct ifinfomsg *ifi = NLMSG_DATA(nh);
2341         struct if_check_data *chk = (struct if_check_data *)arg;
2342
2343         if (nh->nlmsg_type != RTM_NEWLINK)
2344                 return NL_SKIP;
2345
2346         system_device_update_state(chk->dev, ifi->ifi_flags, ifi->ifi_index);
2347         return NL_OK;
2348 }
2349
2350 static int cb_if_check_ack(struct nl_msg *msg, void *arg)
2351 {
2352         struct if_check_data *chk = (struct if_check_data *)arg;
2353         chk->pending = 0;
2354         return NL_STOP;
2355 }
2356
2357 static int cb_if_check_error(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
2358 {
2359         struct if_check_data *chk = (struct if_check_data *)arg;
2360
2361         if (chk->dev->type == &simple_device_type)
2362                 device_set_present(chk->dev, false);
2363         device_set_link(chk->dev, false);
2364         chk->pending = err->error;
2365
2366         return NL_STOP;
2367 }
2368
2369 struct bridge_vlan_check_data {
2370         struct device *check_dev;
2371         int ifindex;
2372         int ret;
2373         bool pending;
2374 };
2375
2376 static void bridge_vlan_check_port(struct bridge_vlan_check_data *data,
2377                                    struct bridge_vlan_port *port,
2378                                    struct bridge_vlan_info *vinfo)
2379 {
2380         uint16_t flags = 0, diff, mask;
2381
2382         if (port->flags & BRVLAN_F_PVID)
2383                 flags |= BRIDGE_VLAN_INFO_PVID;
2384         if (port->flags & BRVLAN_F_UNTAGGED)
2385                 flags |= BRIDGE_VLAN_INFO_UNTAGGED;
2386
2387         diff = vinfo->flags ^ flags;
2388         mask = BRVLAN_F_UNTAGGED | (flags & BRIDGE_VLAN_INFO_PVID);
2389         if (diff & mask) {
2390                 data->ret = 1;
2391                 data->pending = false;
2392         }
2393
2394         port->check = 1;
2395 }
2396
2397 static void bridge_vlan_check_attr(struct bridge_vlan_check_data *data,
2398                                    struct rtattr *attr)
2399 {
2400         struct bridge_vlan_hotplug_port *port;
2401         struct bridge_vlan_info *vinfo;
2402         struct bridge_vlan *vlan;
2403         struct rtattr *cur;
2404         int rem = RTA_PAYLOAD(attr);
2405         int i;
2406
2407         for (cur = RTA_DATA(attr); RTA_OK(cur, rem); cur = RTA_NEXT(cur, rem)) {
2408                 if (cur->rta_type != IFLA_BRIDGE_VLAN_INFO)
2409                         continue;
2410
2411                 vinfo = RTA_DATA(cur);
2412                 vlan = vlist_find(&data->check_dev->vlans, &vinfo->vid, vlan, node);
2413                 if (!vlan) {
2414                         data->ret = 1;
2415                         data->pending = false;
2416                         return;
2417                 }
2418
2419                 for (i = 0; i < vlan->n_ports; i++)
2420                         if (!vlan->ports[i].check)
2421                                 bridge_vlan_check_port(data, &vlan->ports[i], vinfo);
2422
2423                 list_for_each_entry(port, &vlan->hotplug_ports, list)
2424                         if (!port->port.check)
2425                                 bridge_vlan_check_port(data, &port->port, vinfo);
2426         }
2427 }
2428
2429 static int bridge_vlan_check_cb(struct nl_msg *msg, void *arg)
2430 {
2431         struct bridge_vlan_check_data *data = arg;
2432         struct nlmsghdr *nh = nlmsg_hdr(msg);
2433         struct ifinfomsg *ifi = NLMSG_DATA(nh);
2434         struct rtattr *attr;
2435         int rem;
2436
2437         if (nh->nlmsg_type != RTM_NEWLINK)
2438                 return NL_SKIP;
2439
2440         if (ifi->ifi_family != AF_BRIDGE)
2441                 return NL_SKIP;
2442
2443         if (ifi->ifi_index != data->ifindex)
2444                 return NL_SKIP;
2445
2446         attr = IFLA_RTA(ifi);
2447         rem = nh->nlmsg_len - NLMSG_LENGTH(sizeof(*ifi));
2448         while (RTA_OK(attr, rem)) {
2449                 if (attr->rta_type == IFLA_AF_SPEC)
2450                         bridge_vlan_check_attr(data, attr);
2451
2452                 attr = RTA_NEXT(attr, rem);
2453         }
2454
2455         return NL_SKIP;
2456 }
2457
2458 static int bridge_vlan_ack_cb(struct nl_msg *msg, void *arg)
2459 {
2460         struct bridge_vlan_check_data *data = arg;
2461         data->pending = false;
2462         return NL_STOP;
2463 }
2464
2465 static int bridge_vlan_error_cb(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
2466 {
2467         struct bridge_vlan_check_data *data = arg;
2468         data->pending = false;
2469         return NL_STOP;
2470 }
2471
2472 int system_bridge_vlan_check(struct device *dev, char *ifname)
2473 {
2474         struct bridge_vlan_check_data data = {
2475                 .check_dev = dev,
2476                 .ifindex = if_nametoindex(ifname),
2477                 .ret = -1,
2478                 .pending = true,
2479         };
2480         static struct ifinfomsg ifi = {
2481                 .ifi_family = AF_BRIDGE
2482         };
2483         static struct rtattr ext_req = {
2484                 .rta_type = IFLA_EXT_MASK,
2485                 .rta_len = RTA_LENGTH(sizeof(uint32_t)),
2486         };
2487         uint32_t filter = RTEXT_FILTER_BRVLAN;
2488         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
2489         struct bridge_vlan *vlan;
2490         struct nl_msg *msg;
2491         int i;
2492
2493         if (!data.ifindex)
2494                 return 0;
2495
2496         msg = nlmsg_alloc_simple(RTM_GETLINK, NLM_F_DUMP);
2497
2498         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
2499                 nlmsg_append(msg, &ext_req, sizeof(ext_req), NLMSG_ALIGNTO) ||
2500                 nlmsg_append(msg, &filter, sizeof(filter), 0))
2501                 goto free;
2502
2503         vlist_for_each_element(&dev->vlans, vlan, node) {
2504                 struct bridge_vlan_hotplug_port *port;
2505
2506                 for (i = 0; i < vlan->n_ports; i++) {
2507                         if (!strcmp(vlan->ports[i].ifname, ifname))
2508                                 vlan->ports[i].check = 0;
2509                         else
2510                                 vlan->ports[i].check = -1;
2511                 }
2512
2513                 list_for_each_entry(port, &vlan->hotplug_ports, list) {
2514                         if (!strcmp(port->port.ifname, ifname))
2515                                 port->port.check = 0;
2516                         else
2517                                 port->port.check = -1;
2518                 }
2519         }
2520
2521         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, bridge_vlan_check_cb, &data);
2522         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, bridge_vlan_ack_cb, &data);
2523         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, bridge_vlan_ack_cb, &data);
2524         nl_cb_err(cb, NL_CB_CUSTOM, bridge_vlan_error_cb, &data);
2525
2526         if (nl_send_auto_complete(sock_rtnl, msg) < 0)
2527                 goto free;
2528
2529         data.ret = 0;
2530         while (data.pending)
2531                 nl_recvmsgs(sock_rtnl, cb);
2532
2533         vlist_for_each_element(&dev->vlans, vlan, node) {
2534                 struct bridge_vlan_hotplug_port *port;
2535
2536                 for (i = 0; i < vlan->n_ports; i++) {
2537                         if (!vlan->ports[i].check) {
2538                                 data.ret = 1;
2539                                 break;
2540                         }
2541                 }
2542
2543                 list_for_each_entry(port, &vlan->hotplug_ports, list) {
2544                         if (!port->port.check) {
2545                                 data.ret = 1;
2546                                 break;
2547                         }
2548                 }
2549         }
2550
2551 free:
2552         nlmsg_free(msg);
2553         nl_cb_put(cb);
2554         return data.ret;
2555 }
2556
2557 int system_if_check(struct device *dev)
2558 {
2559         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
2560         struct nl_msg *msg;
2561         struct ifinfomsg ifi = {
2562                 .ifi_family = AF_UNSPEC,
2563                 .ifi_index = 0,
2564         };
2565         struct if_check_data chk = {
2566                 .dev = dev,
2567                 .pending = 1,
2568         };
2569         int ret = 1;
2570
2571         if (!cb)
2572                 return ret;
2573
2574         msg = nlmsg_alloc_simple(RTM_GETLINK, 0);
2575         if (!msg)
2576                 goto out;
2577
2578         if (nlmsg_append(msg, &ifi, sizeof(ifi), 0) ||
2579             nla_put_string(msg, IFLA_IFNAME, dev->ifname))
2580                 goto free;
2581
2582         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_if_check_valid, &chk);
2583         nl_cb_set(cb, NL_CB_ACK, NL_CB_CUSTOM, cb_if_check_ack, &chk);
2584         nl_cb_err(cb, NL_CB_CUSTOM, cb_if_check_error, &chk);
2585
2586         ret = nl_send_auto_complete(sock_rtnl, msg);
2587         if (ret < 0)
2588                 goto free;
2589
2590         while (chk.pending > 0)
2591                 nl_recvmsgs(sock_rtnl, cb);
2592
2593         ret = chk.pending;
2594
2595 free:
2596         nlmsg_free(msg);
2597 out:
2598         nl_cb_put(cb);
2599         return ret;
2600 }
2601
2602 struct device *
2603 system_if_get_parent(struct device *dev)
2604 {
2605         char buf[64], *devname;
2606         int ifindex, iflink;
2607
2608         if (system_get_dev_sysfs("iflink", dev->ifname, buf, sizeof(buf)) < 0)
2609                 return NULL;
2610
2611         iflink = strtoul(buf, NULL, 0);
2612         ifindex = system_if_resolve(dev);
2613         if (!iflink || iflink == ifindex)
2614                 return NULL;
2615
2616         devname = if_indextoname(iflink, buf);
2617         if (!devname)
2618                 return NULL;
2619
2620         return device_get(devname, true);
2621 }
2622
2623 static bool
2624 read_string_file(int dir_fd, const char *file, char *buf, int len)
2625 {
2626         bool ret = false;
2627         char *c;
2628         int fd;
2629
2630         fd = openat(dir_fd, file, O_RDONLY);
2631         if (fd < 0)
2632                 return false;
2633
2634 retry:
2635         len = read(fd, buf, len - 1);
2636         if (len < 0) {
2637                 if (errno == EINTR)
2638                         goto retry;
2639         } else if (len > 0) {
2640                         buf[len] = 0;
2641
2642                         c = strchr(buf, '\n');
2643                         if (c)
2644                                 *c = 0;
2645
2646                         ret = true;
2647         }
2648
2649         close(fd);
2650
2651         return ret;
2652 }
2653
2654 static bool
2655 read_uint64_file(int dir_fd, const char *file, uint64_t *val)
2656 {
2657         char buf[64];
2658         bool ret = false;
2659
2660         ret = read_string_file(dir_fd, file, buf, sizeof(buf));
2661         if (ret)
2662                 *val = strtoull(buf, NULL, 0);
2663
2664         return ret;
2665 }
2666
2667 bool
2668 system_if_force_external(const char *ifname)
2669 {
2670         struct stat s;
2671
2672         return stat(dev_sysfs_path(ifname, "phy80211"), &s) == 0;
2673 }
2674
2675 static const char *
2676 system_netdevtype_name(unsigned short dev_type)
2677 {
2678         size_t i;
2679
2680         for (i = 0; i < ARRAY_SIZE(netdev_types); i++) {
2681                 if (netdev_types[i].id == dev_type)
2682                         return netdev_types[i].name;
2683         }
2684
2685         /* the last key is used by default */
2686         i = ARRAY_SIZE(netdev_types) - 1;
2687
2688         return netdev_types[i].name;
2689 }
2690
2691 static void
2692 system_add_devtype(struct blob_buf *b, const char *ifname)
2693 {
2694         char buf[100];
2695         bool found = false;
2696
2697         if (!system_get_dev_sysfs("uevent", ifname, buf, sizeof(buf))) {
2698                 const char *info = "DEVTYPE=";
2699                 char *context = NULL;
2700                 const char *line = strtok_r(buf, "\r\n", &context);
2701
2702                 while (line != NULL) {
2703                         char *index = strstr(line, info);
2704
2705                         if (index != NULL) {
2706                                 blobmsg_add_string(b, "devtype", index + strlen(info));
2707                                 found = true;
2708                                 break;
2709                         }
2710
2711                         line = strtok_r(NULL, "\r\n", &context);
2712                 }
2713         }
2714
2715         if (!found) {
2716                 unsigned short number = 0;
2717                 const char *name = NULL;
2718
2719                 if (!system_get_dev_sysfs("type", ifname, buf, sizeof(buf))) {
2720                         number = strtoul(buf, NULL, 0);
2721                         name = system_netdevtype_name(number);
2722                         blobmsg_add_string(b, "devtype", name);
2723                 }
2724         }
2725 }
2726
2727 #define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
2728
2729 static int32_t
2730 ethtool_feature_count(const char *ifname)
2731 {
2732         struct {
2733                 struct ethtool_sset_info hdr;
2734                 uint32_t buf;
2735         } req = {
2736                 .hdr = {
2737                         .cmd = ETHTOOL_GSSET_INFO,
2738                         .sset_mask = 1 << ETH_SS_FEATURES
2739                 }
2740         };
2741
2742         struct ifreq ifr = {
2743                 .ifr_data = (void *)&req
2744         };
2745
2746         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
2747
2748         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) != 0)
2749                 return -1;
2750
2751         if (!req.hdr.sset_mask)
2752                 return 0;
2753
2754         return req.buf;
2755 }
2756
2757 static int32_t
2758 ethtool_feature_index(const char *ifname, const char *keyname)
2759 {
2760         struct ethtool_gstrings *feature_names;
2761         struct ifreq ifr = { 0 };
2762         int32_t n_features;
2763         uint32_t i;
2764
2765         n_features = ethtool_feature_count(ifname);
2766
2767         if (n_features <= 0)
2768                 return -1;
2769
2770         feature_names = calloc(1, sizeof(*feature_names) + n_features * ETH_GSTRING_LEN);
2771
2772         if (!feature_names)
2773                 return -1;
2774
2775         feature_names->cmd = ETHTOOL_GSTRINGS;
2776         feature_names->string_set = ETH_SS_FEATURES;
2777         feature_names->len = n_features;
2778
2779         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
2780         ifr.ifr_data = (void *)feature_names;
2781
2782         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) != 0) {
2783                 free(feature_names);
2784
2785                 return -1;
2786         }
2787
2788         for (i = 0; i < feature_names->len; i++)
2789                 if (!strcmp((char *)&feature_names->data[i * ETH_GSTRING_LEN], keyname))
2790                         break;
2791
2792         if (i >= feature_names->len)
2793                 i = -1;
2794
2795         free(feature_names);
2796
2797         return i;
2798 }
2799
2800 static bool
2801 ethtool_feature_value(const char *ifname, const char *keyname)
2802 {
2803         struct ethtool_get_features_block *feature_block;
2804         struct ethtool_gfeatures *feature_values;
2805         struct ifreq ifr = { 0 };
2806         int32_t feature_idx;
2807         bool active;
2808
2809         feature_idx = ethtool_feature_index(ifname, keyname);
2810
2811         if (feature_idx < 0)
2812                 return false;
2813
2814         feature_values = calloc(1,
2815                 sizeof(*feature_values) +
2816                 sizeof(feature_values->features[0]) * DIV_ROUND_UP(feature_idx, 32));
2817
2818         if (!feature_values)
2819                 return false;
2820
2821         feature_values->cmd = ETHTOOL_GFEATURES;
2822         feature_values->size = DIV_ROUND_UP(feature_idx, 32);
2823
2824         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name) - 1);
2825         ifr.ifr_data = (void *)feature_values;
2826
2827         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) != 0) {
2828                 free(feature_values);
2829
2830                 return false;
2831         }
2832
2833         feature_block = &feature_values->features[feature_idx / 32];
2834         active = feature_block->active & (1U << feature_idx % 32);
2835
2836         free(feature_values);
2837
2838         return active;
2839 }
2840
2841 static void
2842 system_add_link_mode_name(struct blob_buf *b, int i, bool half)
2843 {
2844         char *buf;
2845
2846         /* allocate string buffer large enough for the mode name and a suffix
2847          * "-F" or "-H" indicating full duplex or half duplex.
2848          */
2849         buf = blobmsg_alloc_string_buffer(b, NULL, strlen(ethtool_modes[i].name) + 3);
2850         if (!buf)
2851                 return;
2852
2853         strcpy(buf, ethtool_modes[i].name);
2854         if (half)
2855                 strcat(buf, "-H");
2856         else
2857                 strcat(buf, "-F");
2858
2859         blobmsg_add_string_buffer(b);
2860 }
2861
2862 static void
2863 system_add_link_modes(__s8 nwords, struct blob_buf *b, __u32 *mask)
2864 {
2865         size_t i;
2866
2867         for (i = 0; i < ARRAY_SIZE(ethtool_modes); i++) {
2868                 if (ethtool_link_mode_test_bit(nwords, ethtool_modes[i].bit_half, mask))
2869                         system_add_link_mode_name(b, i, true);
2870
2871                 if (ethtool_link_mode_test_bit(nwords, ethtool_modes[i].bit_full, mask))
2872                         system_add_link_mode_name(b, i, false);
2873         }
2874 }
2875
2876 static void
2877 system_add_pause_modes(__s8 nwords, struct blob_buf *b, __u32 *mask)
2878 {
2879         if (ethtool_link_mode_test_bit(nwords, ETHTOOL_LINK_MODE_Pause_BIT, mask))
2880                 blobmsg_add_string(b, NULL, "pause");
2881
2882         if (ethtool_link_mode_test_bit(nwords, ETHTOOL_LINK_MODE_Asym_Pause_BIT, mask))
2883                 blobmsg_add_string(b, NULL, "asym_pause");
2884 }
2885
2886
2887 static void
2888 system_add_ethtool_pause_an(struct blob_buf *b, __s8 nwords,
2889                             __u32 *advertising, __u32 *lp_advertising)
2890 {
2891         bool an_rx = false, an_tx = false;
2892         void *d;
2893
2894         d = blobmsg_open_array(b, "negotiated");
2895
2896         /* Work out negotiated pause frame usage per
2897          * IEEE 802.3-2005 table 28B-3.
2898          */
2899         if (ethtool_link_mode_test_bit(nwords,
2900                                        ETHTOOL_LINK_MODE_Pause_BIT,
2901                                        advertising) &&
2902             ethtool_link_mode_test_bit(nwords,
2903                                        ETHTOOL_LINK_MODE_Pause_BIT,
2904                                        lp_advertising)) {
2905                 an_tx = true;
2906                 an_rx = true;
2907         } else if (ethtool_link_mode_test_bit(nwords,
2908                                               ETHTOOL_LINK_MODE_Asym_Pause_BIT,
2909                                               advertising) &&
2910                    ethtool_link_mode_test_bit(nwords,
2911                                               ETHTOOL_LINK_MODE_Asym_Pause_BIT,
2912                                               lp_advertising)) {
2913                 if (ethtool_link_mode_test_bit(nwords,
2914                                                ETHTOOL_LINK_MODE_Pause_BIT,
2915                                                advertising))
2916                         an_rx = true;
2917                 else if (ethtool_link_mode_test_bit(nwords,
2918                                                     ETHTOOL_LINK_MODE_Pause_BIT,
2919                                                     lp_advertising))
2920                         an_tx = true;
2921         }
2922         if (an_tx)
2923                 blobmsg_add_string(b, NULL, "rx");
2924
2925         if (an_rx)
2926                 blobmsg_add_string(b, NULL, "tx");
2927
2928         blobmsg_close_array(b, d);
2929 }
2930
2931 static void
2932 system_get_ethtool_pause(struct device *dev, bool *rx_pause, bool *tx_pause, bool *pause_autoneg)
2933 {
2934         struct ethtool_pauseparam pp;
2935         struct ifreq ifr = {
2936                 .ifr_data = (caddr_t)&pp,
2937         };
2938
2939         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
2940         memset(&pp, 0, sizeof(pp));
2941         pp.cmd = ETHTOOL_GPAUSEPARAM;
2942
2943         /* may fail */
2944         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) == -1) {
2945                 *pause_autoneg = true;
2946                 return;
2947         }
2948
2949         *rx_pause = pp.rx_pause;
2950         *tx_pause = pp.tx_pause;
2951         *pause_autoneg = pp.autoneg;
2952 }
2953
2954 int
2955 system_if_dump_info(struct device *dev, struct blob_buf *b)
2956 {
2957         __u32 *supported, *advertising, *lp_advertising;
2958         bool rx_pause, tx_pause, pause_autoneg;
2959         struct {
2960                 struct ethtool_link_settings req;
2961                 __u32 link_mode_data[3 * 127];
2962         } ecmd;
2963         struct ifreq ifr = {
2964                 .ifr_data = (caddr_t)&ecmd,
2965         };
2966         __s8 nwords;
2967         void *c, *d;
2968         char *s;
2969
2970         system_get_ethtool_pause(dev, &rx_pause, &tx_pause, &pause_autoneg);
2971
2972         memset(&ecmd, 0, sizeof(ecmd));
2973         ecmd.req.cmd = ETHTOOL_GLINKSETTINGS;
2974         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name) - 1);
2975
2976         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) < 0 ||
2977             ecmd.req.link_mode_masks_nwords >= 0 ||
2978             ecmd.req.cmd != ETHTOOL_GLINKSETTINGS)
2979                 return -EOPNOTSUPP;
2980
2981         ecmd.req.link_mode_masks_nwords = -ecmd.req.link_mode_masks_nwords;
2982
2983         if (ioctl(sock_ioctl, SIOCETHTOOL, &ifr) < 0 ||
2984             ecmd.req.link_mode_masks_nwords <= 0 ||
2985             ecmd.req.cmd != ETHTOOL_GLINKSETTINGS)
2986                 return -EIO;
2987
2988         nwords = ecmd.req.link_mode_masks_nwords;
2989         supported = &ecmd.link_mode_data[0];
2990         advertising = &ecmd.link_mode_data[nwords];
2991         lp_advertising = &ecmd.link_mode_data[2 * nwords];
2992
2993         c = blobmsg_open_array(b, "link-advertising");
2994         system_add_link_modes(nwords, b, advertising);
2995         blobmsg_close_array(b, c);
2996
2997         c = blobmsg_open_array(b, "link-partner-advertising");
2998         system_add_link_modes(nwords, b, lp_advertising);
2999         blobmsg_close_array(b, c);
3000
3001         c = blobmsg_open_array(b, "link-supported");
3002         system_add_link_modes(nwords, b, supported);
3003         blobmsg_close_array(b, c);
3004
3005         if (ethtool_validate_speed(ecmd.req.speed) &&
3006             (ecmd.req.speed != (__u32)SPEED_UNKNOWN) &&
3007             (ecmd.req.speed != 0)) {
3008                 s = blobmsg_alloc_string_buffer(b, "speed", 10);
3009                 snprintf(s, 8, "%d%c", ecmd.req.speed,
3010                         ecmd.req.duplex == DUPLEX_HALF ? 'H' : 'F');
3011                 blobmsg_add_string_buffer(b);
3012         }
3013         blobmsg_add_u8(b, "autoneg", !!ecmd.req.autoneg);
3014
3015         c = blobmsg_open_table(b, "flow-control");
3016         blobmsg_add_u8(b, "autoneg", pause_autoneg);
3017
3018         d = blobmsg_open_array(b, "supported");
3019         system_add_pause_modes(nwords, b, supported);
3020         blobmsg_close_array(b, d);
3021
3022         if (pause_autoneg) {
3023                 d = blobmsg_open_array(b, "link-advertising");
3024                 system_add_pause_modes(nwords, b, advertising);
3025                 blobmsg_close_array(b, d);
3026         }
3027
3028         d = blobmsg_open_array(b, "link-partner-advertising");
3029         system_add_pause_modes(nwords, b, lp_advertising);
3030         blobmsg_close_array(b, d);
3031
3032         if (pause_autoneg) {
3033                 system_add_ethtool_pause_an(b, nwords, advertising,
3034                                             lp_advertising);
3035         } else {
3036                 d = blobmsg_open_array(b, "selected");
3037                 if (rx_pause)
3038                         blobmsg_add_string(b, NULL, "rx");
3039
3040                 if (tx_pause)
3041                         blobmsg_add_string(b, NULL, "tx");
3042
3043                 blobmsg_close_array(b, d);
3044         }
3045
3046         blobmsg_close_table(b, c);
3047
3048         blobmsg_add_u8(b, "hw-tc-offload",
3049                 ethtool_feature_value(dev->ifname, "hw-tc-offload"));
3050
3051         system_add_devtype(b, dev->ifname);
3052
3053         return 0;
3054 }
3055
3056 int
3057 system_if_dump_stats(struct device *dev, struct blob_buf *b)
3058 {
3059         const char *const counters[] = {
3060                 "collisions",     "rx_frame_errors",   "tx_compressed",
3061                 "multicast",      "rx_length_errors",  "tx_dropped",
3062                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
3063                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
3064                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
3065                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
3066                 "rx_errors",      "tx_bytes",          "tx_window_errors",
3067                 "rx_fifo_errors", "tx_carrier_errors",
3068         };
3069         int stats_dir;
3070         size_t i;
3071         uint64_t val = 0;
3072
3073         stats_dir = open(dev_sysfs_path(dev->ifname, "statistics"), O_DIRECTORY);
3074         if (stats_dir < 0)
3075                 return -1;
3076
3077         for (i = 0; i < ARRAY_SIZE(counters); i++)
3078                 if (read_uint64_file(stats_dir, counters[i], &val))
3079                         blobmsg_add_u64(b, counters[i], val);
3080
3081         close(stats_dir);
3082         return 0;
3083 }
3084
3085 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
3086 {
3087         bool v4 = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4);
3088         int alen = v4 ? 4 : 16;
3089         unsigned int flags = 0;
3090         struct ifaddrmsg ifa = {
3091                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
3092                 .ifa_prefixlen = addr->mask,
3093                 .ifa_index = dev->ifindex,
3094         };
3095
3096         struct nl_msg *msg;
3097         if (cmd == RTM_NEWADDR)
3098                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
3099
3100         msg = nlmsg_alloc_simple(cmd, flags);
3101         if (!msg)
3102                 return -1;
3103
3104         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
3105         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
3106         if (v4) {
3107                 if (addr->broadcast)
3108                         nla_put_u32(msg, IFA_BROADCAST, addr->broadcast);
3109                 if (addr->point_to_point)
3110                         nla_put_u32(msg, IFA_ADDRESS, addr->point_to_point);
3111         } else {
3112                 time_t now = system_get_rtime();
3113                 struct ifa_cacheinfo cinfo = {0xffffffffU, 0xffffffffU, 0, 0};
3114
3115                 if (addr->preferred_until) {
3116                         int64_t preferred = addr->preferred_until - now;
3117                         if (preferred < 0)
3118                                 preferred = 0;
3119                         else if (preferred > UINT32_MAX)
3120                                 preferred = UINT32_MAX;
3121
3122                         cinfo.ifa_prefered = preferred;
3123                 }
3124
3125                 if (addr->valid_until) {
3126                         int64_t valid = addr->valid_until - now;
3127                         if (valid <= 0) {
3128                                 nlmsg_free(msg);
3129                                 return -1;
3130                         }
3131                         else if (valid > UINT32_MAX)
3132                                 valid = UINT32_MAX;
3133
3134                         cinfo.ifa_valid = valid;
3135                 }
3136
3137                 nla_put(msg, IFA_CACHEINFO, sizeof(cinfo), &cinfo);
3138
3139                 if (cmd == RTM_NEWADDR && (addr->flags & DEVADDR_OFFLINK))
3140                         nla_put_u32(msg, IFA_FLAGS, IFA_F_NOPREFIXROUTE);
3141         }
3142
3143         return system_rtnl_call(msg);
3144 }
3145
3146 int system_add_address(struct device *dev, struct device_addr *addr)
3147 {
3148         return system_addr(dev, addr, RTM_NEWADDR);
3149 }
3150
3151 int system_del_address(struct device *dev, struct device_addr *addr)
3152 {
3153         return system_addr(dev, addr, RTM_DELADDR);
3154 }
3155
3156 static int system_neigh(struct device *dev, struct device_neighbor *neighbor, int cmd)
3157 {
3158         int alen = ((neighbor->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
3159         unsigned int flags = 0;
3160         struct ndmsg ndm = {
3161                 .ndm_family = (alen == 4) ? AF_INET : AF_INET6,
3162                 .ndm_ifindex = dev->ifindex,
3163                 .ndm_state = NUD_PERMANENT,
3164                 .ndm_flags = (neighbor->proxy ? NTF_PROXY : 0) | (neighbor->router ? NTF_ROUTER : 0),
3165         };
3166         struct nl_msg *msg;
3167
3168         if (cmd == RTM_NEWNEIGH)
3169                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
3170
3171         msg = nlmsg_alloc_simple(cmd, flags);
3172
3173         if (!msg)
3174                 return -1;
3175
3176         nlmsg_append(msg, &ndm, sizeof(ndm), 0);
3177
3178         nla_put(msg, NDA_DST, alen, &neighbor->addr);
3179         if (neighbor->flags & DEVNEIGH_MAC)
3180                 nla_put(msg, NDA_LLADDR, sizeof(neighbor->macaddr), &neighbor->macaddr);
3181
3182
3183         return system_rtnl_call(msg);
3184 }
3185
3186 int system_add_neighbor(struct device *dev, struct device_neighbor *neighbor)
3187 {
3188         return system_neigh(dev, neighbor, RTM_NEWNEIGH);
3189 }
3190
3191 int system_del_neighbor(struct device *dev, struct device_neighbor *neighbor)
3192 {
3193         return system_neigh(dev, neighbor, RTM_DELNEIGH);
3194 }
3195
3196 static int system_rt(struct device *dev, struct device_route *route, int cmd)
3197 {
3198         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
3199         bool have_gw;
3200         unsigned int flags = 0;
3201
3202         if (alen == 4)
3203                 have_gw = !!route->nexthop.in.s_addr;
3204         else
3205                 have_gw = route->nexthop.in6.s6_addr32[0] ||
3206                         route->nexthop.in6.s6_addr32[1] ||
3207                         route->nexthop.in6.s6_addr32[2] ||
3208                         route->nexthop.in6.s6_addr32[3];
3209
3210         unsigned int table = (route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))
3211                         ? route->table : RT_TABLE_MAIN;
3212
3213         struct rtmsg rtm = {
3214                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
3215                 .rtm_dst_len = route->mask,
3216                 .rtm_src_len = route->sourcemask,
3217                 .rtm_table = (table < 256) ? table : RT_TABLE_UNSPEC,
3218                 .rtm_protocol = (route->flags & DEVROUTE_PROTO) ? route->proto : RTPROT_STATIC,
3219                 .rtm_scope = RT_SCOPE_NOWHERE,
3220                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
3221                 .rtm_flags = (route->flags & DEVROUTE_ONLINK) ? RTNH_F_ONLINK : 0,
3222         };
3223         struct nl_msg *msg;
3224
3225         if (cmd == RTM_NEWROUTE) {
3226                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
3227
3228                 if (!dev) { /* Add null-route */
3229                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
3230                         rtm.rtm_type = RTN_UNREACHABLE;
3231                 }
3232                 else
3233                         rtm.rtm_scope = (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
3234         }
3235
3236         if (route->flags & DEVROUTE_TYPE) {
3237                 rtm.rtm_type = route->type;
3238                 if (!(route->flags & (DEVROUTE_TABLE | DEVROUTE_SRCTABLE))) {
3239                         if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_BROADCAST ||
3240                             rtm.rtm_type == RTN_NAT || rtm.rtm_type == RTN_ANYCAST)
3241                                 rtm.rtm_table = RT_TABLE_LOCAL;
3242                 }
3243
3244                 if (rtm.rtm_type == RTN_LOCAL || rtm.rtm_type == RTN_NAT) {
3245                         rtm.rtm_scope = RT_SCOPE_HOST;
3246                 } else if (rtm.rtm_type == RTN_BROADCAST || rtm.rtm_type == RTN_MULTICAST ||
3247                                 rtm.rtm_type == RTN_ANYCAST) {
3248                         rtm.rtm_scope = RT_SCOPE_LINK;
3249                 } else if (rtm.rtm_type == RTN_BLACKHOLE || rtm.rtm_type == RTN_UNREACHABLE ||
3250                                 rtm.rtm_type == RTN_PROHIBIT || rtm.rtm_type == RTN_FAILED_POLICY ||
3251                                 rtm.rtm_type == RTN_THROW) {
3252                         rtm.rtm_scope = RT_SCOPE_UNIVERSE;
3253                         dev = NULL;
3254                 }
3255         }
3256
3257         if (route->flags & DEVROUTE_NODEV)
3258                 dev = NULL;
3259
3260         msg = nlmsg_alloc_simple(cmd, flags);
3261         if (!msg)
3262                 return -1;
3263
3264         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
3265
3266         if (route->mask)
3267                 nla_put(msg, RTA_DST, alen, &route->addr);
3268
3269         if (route->sourcemask) {
3270                 if (rtm.rtm_family == AF_INET)
3271                         nla_put(msg, RTA_PREFSRC, alen, &route->source);
3272                 else
3273                         nla_put(msg, RTA_SRC, alen, &route->source);
3274         }
3275
3276         if (route->metric > 0)
3277                 nla_put_u32(msg, RTA_PRIORITY, route->metric);
3278
3279         if (have_gw)
3280                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
3281
3282         if (dev)
3283                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
3284
3285         if (table >= 256)
3286                 nla_put_u32(msg, RTA_TABLE, table);
3287
3288         if (route->flags & DEVROUTE_MTU) {
3289                 struct nlattr *metrics;
3290
3291                 if (!(metrics = nla_nest_start(msg, RTA_METRICS)))
3292                         goto nla_put_failure;
3293
3294                 nla_put_u32(msg, RTAX_MTU, route->mtu);
3295
3296                 nla_nest_end(msg, metrics);
3297         }
3298
3299         return system_rtnl_call(msg);
3300
3301 nla_put_failure:
3302         nlmsg_free(msg);
3303         return -ENOMEM;
3304 }
3305
3306 int system_add_route(struct device *dev, struct device_route *route)
3307 {
3308         return system_rt(dev, route, RTM_NEWROUTE);
3309 }
3310
3311 int system_del_route(struct device *dev, struct device_route *route)
3312 {
3313         return system_rt(dev, route, RTM_DELROUTE);
3314 }
3315
3316 int system_flush_routes(void)
3317 {
3318         const char *names[] = { "ipv4", "ipv6" };
3319         size_t i;
3320         int fd;
3321
3322         for (i = 0; i < ARRAY_SIZE(names); i++) {
3323                 snprintf(dev_buf, sizeof(dev_buf), "%s/sys/net/%s/route/flush", proc_path, names[i]);
3324                 fd = open(dev_buf, O_WRONLY);
3325                 if (fd < 0)
3326                         continue;
3327
3328                 if (write(fd, "-1", 2)) {}
3329                 close(fd);
3330         }
3331         return 0;
3332 }
3333
3334 bool system_resolve_rt_type(const char *type, unsigned int *id)
3335 {
3336         return system_rtn_aton(type, id);
3337 }
3338
3339 bool system_resolve_rt_proto(const char *type, unsigned int *id)
3340 {
3341         FILE *f;
3342         char *e, buf[128];
3343         unsigned int n, proto = 256;
3344         n = strtoul(type, &e, 0);
3345         if (!*e && e != type)
3346                 proto = n;
3347         else if (!strcmp(type, "unspec"))
3348                 proto = RTPROT_UNSPEC;
3349         else if (!strcmp(type, "kernel"))
3350                 proto = RTPROT_KERNEL;
3351         else if (!strcmp(type, "boot"))
3352                 proto = RTPROT_BOOT;
3353         else if (!strcmp(type, "static"))
3354                 proto = RTPROT_STATIC;
3355         else if ((f = fopen("/etc/iproute2/rt_protos", "r")) != NULL) {
3356                 while (fgets(buf, sizeof(buf) - 1, f) != NULL) {
3357                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
3358                                 continue;
3359
3360                         n = strtoul(e, NULL, 10);
3361                         e = strtok(NULL, " \t\n");
3362
3363                         if (e && !strcmp(e, type)) {
3364                                 proto = n;
3365                                 break;
3366                         }
3367                 }
3368                 fclose(f);
3369         }
3370
3371         if (proto > 255)
3372                 return false;
3373
3374         *id = proto;
3375         return true;
3376 }
3377
3378 bool system_resolve_rt_table(const char *name, unsigned int *id)
3379 {
3380         FILE *f;
3381         char *e, buf[128];
3382         unsigned int n, table = RT_TABLE_UNSPEC;
3383
3384         /* first try to parse table as number */
3385         if ((n = strtoul(name, &e, 0)) > 0 && !*e)
3386                 table = n;
3387
3388         /* handle well known aliases */
3389         else if (!strcmp(name, "default"))
3390                 table = RT_TABLE_DEFAULT;
3391         else if (!strcmp(name, "main"))
3392                 table = RT_TABLE_MAIN;
3393         else if (!strcmp(name, "local"))
3394                 table = RT_TABLE_LOCAL;
3395
3396         /* try to look up name in /etc/iproute2/rt_tables */
3397         else if ((f = fopen("/etc/iproute2/rt_tables", "r")) != NULL)
3398         {
3399                 while (fgets(buf, sizeof(buf) - 1, f) != NULL)
3400                 {
3401                         if ((e = strtok(buf, " \t\n")) == NULL || *e == '#')
3402                                 continue;
3403
3404                         n = strtoul(e, NULL, 10);
3405                         e = strtok(NULL, " \t\n");
3406
3407                         if (e && !strcmp(e, name))
3408                         {
3409                                 table = n;
3410                                 break;
3411                         }
3412                 }
3413
3414                 fclose(f);
3415         }
3416
3417         if (table == RT_TABLE_UNSPEC)
3418                 return false;
3419
3420         *id = table;
3421         return true;
3422 }
3423
3424 bool system_is_default_rt_table(unsigned int id)
3425 {
3426         return (id == RT_TABLE_MAIN);
3427 }
3428
3429 bool system_resolve_rpfilter(const char *filter, unsigned int *id)
3430 {
3431         char *e;
3432         unsigned int n;
3433
3434         if (!strcmp(filter, "strict"))
3435                 n = 1;
3436         else if (!strcmp(filter, "loose"))
3437                 n = 2;
3438         else {
3439                 n = strtoul(filter, &e, 0);
3440                 if (*e || e == filter || n > 2)
3441                         return false;
3442         }
3443
3444         *id = n;
3445         return true;
3446 }
3447
3448 static int system_iprule(struct iprule *rule, int cmd)
3449 {
3450         int alen = ((rule->flags & IPRULE_FAMILY) == IPRULE_INET4) ? 4 : 16;
3451
3452         struct nl_msg *msg;
3453         struct rtmsg rtm = {
3454                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
3455                 .rtm_protocol = RTPROT_STATIC,
3456                 .rtm_scope = RT_SCOPE_UNIVERSE,
3457                 .rtm_table = RT_TABLE_UNSPEC,
3458                 .rtm_type = RTN_UNSPEC,
3459                 .rtm_flags = 0,
3460         };
3461
3462         if (cmd == RTM_NEWRULE)
3463                 rtm.rtm_type = RTN_UNICAST;
3464
3465         if (rule->invert)
3466                 rtm.rtm_flags |= FIB_RULE_INVERT;
3467
3468         if (rule->flags & IPRULE_SRC)
3469                 rtm.rtm_src_len = rule->src_mask;
3470
3471         if (rule->flags & IPRULE_DEST)
3472                 rtm.rtm_dst_len = rule->dest_mask;
3473
3474         if (rule->flags & IPRULE_TOS)
3475                 rtm.rtm_tos = rule->tos;
3476
3477         if (rule->flags & IPRULE_LOOKUP) {
3478                 if (rule->lookup < 256)
3479                         rtm.rtm_table = rule->lookup;
3480         }
3481
3482         if (rule->flags & IPRULE_ACTION)
3483                 rtm.rtm_type = rule->action;
3484         else if (rule->flags & IPRULE_GOTO)
3485                 rtm.rtm_type = FR_ACT_GOTO;
3486         else if (!(rule->flags & (IPRULE_LOOKUP | IPRULE_ACTION | IPRULE_GOTO)))
3487                 rtm.rtm_type = FR_ACT_NOP;
3488
3489         msg = nlmsg_alloc_simple(cmd, NLM_F_REQUEST);
3490
3491         if (!msg)
3492                 return -1;
3493
3494         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
3495
3496         if (rule->flags & IPRULE_IN)
3497                 nla_put(msg, FRA_IFNAME, strlen(rule->in_dev) + 1, rule->in_dev);
3498
3499         if (rule->flags & IPRULE_OUT)
3500                 nla_put(msg, FRA_OIFNAME, strlen(rule->out_dev) + 1, rule->out_dev);
3501
3502         if (rule->flags & IPRULE_SRC)
3503                 nla_put(msg, FRA_SRC, alen, &rule->src_addr);
3504
3505         if (rule->flags & IPRULE_DEST)
3506                 nla_put(msg, FRA_DST, alen, &rule->dest_addr);
3507
3508         if (rule->flags & IPRULE_PRIORITY)
3509                 nla_put_u32(msg, FRA_PRIORITY, rule->priority);
3510         else if (cmd == RTM_NEWRULE)
3511                 nla_put_u32(msg, FRA_PRIORITY, rule->order);
3512
3513         if (rule->flags & IPRULE_FWMARK)
3514                 nla_put_u32(msg, FRA_FWMARK, rule->fwmark);
3515
3516         if (rule->flags & IPRULE_FWMASK)
3517                 nla_put_u32(msg, FRA_FWMASK, rule->fwmask);
3518
3519         if (rule->flags & IPRULE_LOOKUP) {
3520                 if (rule->lookup >= 256)
3521                         nla_put_u32(msg, FRA_TABLE, rule->lookup);
3522         }
3523
3524         if (rule->flags & IPRULE_SUP_PREFIXLEN)
3525                 nla_put_u32(msg, FRA_SUPPRESS_PREFIXLEN, rule->sup_prefixlen);
3526
3527         if (rule->flags & IPRULE_UIDRANGE) {
3528                 struct fib_rule_uid_range uidrange = {
3529                         .start = rule->uidrange_start,
3530                         .end = rule->uidrange_end
3531                 };
3532
3533                 nla_put(msg, FRA_UID_RANGE, sizeof(uidrange), &uidrange);
3534         }
3535
3536         if (rule->flags & IPRULE_GOTO)
3537                 nla_put_u32(msg, FRA_GOTO, rule->gotoid);
3538
3539         return system_rtnl_call(msg);
3540 }
3541
3542 int system_add_iprule(struct iprule *rule)
3543 {
3544         return system_iprule(rule, RTM_NEWRULE);
3545 }
3546
3547 int system_del_iprule(struct iprule *rule)
3548 {
3549         return system_iprule(rule, RTM_DELRULE);
3550 }
3551
3552 int system_flush_iprules(void)
3553 {
3554         int rv = 0;
3555         struct iprule rule;
3556
3557         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET);
3558         system_if_clear_entries(NULL, RTM_GETRULE, AF_INET6);
3559
3560         memset(&rule, 0, sizeof(rule));
3561
3562
3563         rule.flags = IPRULE_INET4 | IPRULE_PRIORITY | IPRULE_LOOKUP;
3564
3565         rule.priority = 0;
3566         rule.lookup = RT_TABLE_LOCAL;
3567         rv |= system_iprule(&rule, RTM_NEWRULE);
3568
3569         rule.priority = 32766;
3570         rule.lookup = RT_TABLE_MAIN;
3571         rv |= system_iprule(&rule, RTM_NEWRULE);
3572
3573         rule.priority = 32767;
3574         rule.lookup = RT_TABLE_DEFAULT;
3575         rv |= system_iprule(&rule, RTM_NEWRULE);
3576
3577
3578         rule.flags = IPRULE_INET6 | IPRULE_PRIORITY | IPRULE_LOOKUP;
3579
3580         rule.priority = 0;
3581         rule.lookup = RT_TABLE_LOCAL;
3582         rv |= system_iprule(&rule, RTM_NEWRULE);
3583
3584         rule.priority = 32766;
3585         rule.lookup = RT_TABLE_MAIN;
3586         rv |= system_iprule(&rule, RTM_NEWRULE);
3587
3588         return rv;
3589 }
3590
3591 bool system_resolve_iprule_action(const char *action, unsigned int *id)
3592 {
3593         return system_rtn_aton(action, id);
3594 }
3595
3596 time_t system_get_rtime(void)
3597 {
3598         struct timespec ts;
3599         struct timeval tv;
3600
3601         if (clock_gettime(CLOCK_MONOTONIC, &ts) == 0)
3602                 return ts.tv_sec;
3603
3604         if (gettimeofday(&tv, NULL) == 0)
3605                 return tv.tv_sec;
3606
3607         return 0;
3608 }
3609
3610 #ifndef IP_DF
3611 #define IP_DF       0x4000
3612 #endif
3613
3614 static int tunnel_ioctl(const char *name, int cmd, void *p)
3615 {
3616         struct ifreq ifr;
3617
3618         memset(&ifr, 0, sizeof(ifr));
3619         strncpy(ifr.ifr_name, name, sizeof(ifr.ifr_name) - 1);
3620         ifr.ifr_ifru.ifru_data = p;
3621         return ioctl(sock_ioctl, cmd, &ifr);
3622 }
3623
3624 #ifdef IFLA_IPTUN_MAX
3625 static int system_add_ip6_tunnel(const char *name, const unsigned int link,
3626                                  struct blob_attr **tb)
3627 {
3628         struct nl_msg *nlm = nlmsg_alloc_simple(RTM_NEWLINK,
3629                                 NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
3630         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC };
3631         struct blob_attr *cur;
3632         int ret = 0, ttl = 0;
3633
3634         if (!nlm)
3635                 return -1;
3636
3637         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
3638         nla_put_string(nlm, IFLA_IFNAME, name);
3639
3640         if (link)
3641                 nla_put_u32(nlm, IFLA_LINK, link);
3642
3643         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
3644         if (!linkinfo) {
3645                 ret = -ENOMEM;
3646                 goto failure;
3647         }
3648
3649         nla_put_string(nlm, IFLA_INFO_KIND, "ip6tnl");
3650         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
3651         if (!infodata) {
3652                 ret = -ENOMEM;
3653                 goto failure;
3654         }
3655
3656         if (link)
3657                 nla_put_u32(nlm, IFLA_IPTUN_LINK, link);
3658
3659         if ((cur = tb[TUNNEL_ATTR_TTL]))
3660                 ttl = blobmsg_get_u32(cur);
3661
3662         nla_put_u8(nlm, IFLA_IPTUN_PROTO, IPPROTO_IPIP);
3663         nla_put_u8(nlm, IFLA_IPTUN_TTL, (ttl) ? ttl : 64);
3664
3665         struct in6_addr in6buf;
3666         if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3667                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3668                         ret = -EINVAL;
3669                         goto failure;
3670                 }
3671                 nla_put(nlm, IFLA_IPTUN_LOCAL, sizeof(in6buf), &in6buf);
3672         }
3673
3674         if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3675                 if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3676                         ret = -EINVAL;
3677                         goto failure;
3678                 }
3679                 nla_put(nlm, IFLA_IPTUN_REMOTE, sizeof(in6buf), &in6buf);
3680         }
3681
3682         if ((cur = tb[TUNNEL_ATTR_DATA])) {
3683                 struct blob_attr *tb_data[__IPIP6_DATA_ATTR_MAX];
3684                 uint32_t tun_flags = IP6_TNL_F_IGN_ENCAP_LIMIT;
3685
3686                 blobmsg_parse(ipip6_data_attr_list.params, __IPIP6_DATA_ATTR_MAX, tb_data,
3687                         blobmsg_data(cur), blobmsg_len(cur));
3688
3689                 if ((cur = tb_data[IPIP6_DATA_ENCAPLIMIT])) {
3690                         char *str = blobmsg_get_string(cur);
3691
3692                         if (strcmp(str, "ignore")) {
3693                                 char *e;
3694                                 unsigned encap_limit = strtoul(str, &e, 0);
3695
3696                                 if (e == str || *e || encap_limit > 255) {
3697                                         ret = -EINVAL;
3698                                         goto failure;
3699                                 }
3700
3701                                 nla_put_u8(nlm, IFLA_IPTUN_ENCAP_LIMIT, encap_limit);
3702                                 tun_flags &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
3703                         }
3704                 }
3705
3706 #ifdef IFLA_IPTUN_FMR_MAX
3707                 if ((cur = tb_data[IPIP6_DATA_FMRS])) {
3708                         struct blob_attr *rcur;
3709                         unsigned rrem, fmrcnt = 0;
3710                         struct nlattr *fmrs = nla_nest_start(nlm, IFLA_IPTUN_FMRS);
3711
3712                         if (!fmrs) {
3713                                 ret = -ENOMEM;
3714                                 goto failure;
3715                         }
3716
3717                         blobmsg_for_each_attr(rcur, cur, rrem) {
3718                                 struct blob_attr *tb_fmr[__FMR_DATA_ATTR_MAX], *tb_cur;
3719                                 struct in6_addr ip6prefix;
3720                                 struct in_addr ip4prefix;
3721                                 unsigned ip4len, ip6len, ealen, offset;
3722
3723                                 blobmsg_parse(fmr_data_attr_list.params, __FMR_DATA_ATTR_MAX, tb_fmr,
3724                                                 blobmsg_data(rcur), blobmsg_len(rcur));
3725
3726                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX6]) ||
3727                                                 !parse_ip_and_netmask(AF_INET6,
3728                                                         blobmsg_data(tb_cur), &ip6prefix,
3729                                                         &ip6len)) {
3730                                         ret = -EINVAL;
3731                                         goto failure;
3732                                 }
3733
3734                                 if (!(tb_cur = tb_fmr[FMR_DATA_PREFIX4]) ||
3735                                                 !parse_ip_and_netmask(AF_INET,
3736                                                         blobmsg_data(tb_cur), &ip4prefix,
3737                                                         &ip4len)) {
3738                                         ret = -EINVAL;
3739                                         goto failure;
3740                                 }
3741
3742                                 if (!(tb_cur = tb_fmr[FMR_DATA_EALEN])) {
3743                                         ret = -EINVAL;
3744                                         goto failure;
3745                                 }
3746                                 ealen = blobmsg_get_u32(tb_cur);
3747
3748                                 if (!(tb_cur = tb_fmr[FMR_DATA_OFFSET])) {
3749                                         ret = -EINVAL;
3750                                         goto failure;
3751                                 }
3752                                 offset = blobmsg_get_u32(tb_cur);
3753
3754                                 struct nlattr *rule = nla_nest_start(nlm, ++fmrcnt);
3755                                 if (!rule) {
3756                                         ret = -ENOMEM;
3757                                         goto failure;
3758                                 }
3759
3760                                 nla_put(nlm, IFLA_IPTUN_FMR_IP6_PREFIX, sizeof(ip6prefix), &ip6prefix);
3761                                 nla_put(nlm, IFLA_IPTUN_FMR_IP4_PREFIX, sizeof(ip4prefix), &ip4prefix);
3762                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP6_PREFIX_LEN, ip6len);
3763                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_IP4_PREFIX_LEN, ip4len);
3764                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_EA_LEN, ealen);
3765                                 nla_put_u8(nlm, IFLA_IPTUN_FMR_OFFSET, offset);
3766
3767                                 nla_nest_end(nlm, rule);
3768                         }
3769
3770                         nla_nest_end(nlm, fmrs);
3771                 }
3772 #endif
3773                 if (tun_flags)
3774                         nla_put_u32(nlm, IFLA_IPTUN_FLAGS, tun_flags);
3775         }
3776
3777         nla_nest_end(nlm, infodata);
3778         nla_nest_end(nlm, linkinfo);
3779
3780         return system_rtnl_call(nlm);
3781
3782 failure:
3783         nlmsg_free(nlm);
3784         return ret;
3785 }
3786 #endif
3787
3788 #ifdef IFLA_IPTUN_MAX
3789 #define IP6_FLOWINFO_TCLASS     htonl(0x0FF00000)
3790 static int system_add_gre_tunnel(const char *name, const char *kind,
3791                                  const unsigned int link, struct blob_attr **tb, bool v6)
3792 {
3793         struct nl_msg *nlm;
3794         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
3795         struct blob_attr *cur;
3796         uint32_t ikey = 0, okey = 0, flowinfo = 0, flags6 = IP6_TNL_F_IGN_ENCAP_LIMIT;
3797         uint16_t iflags = 0, oflags = 0;
3798         uint8_t tos = 0;
3799         int ret = 0, ttl = 0;
3800         unsigned encap_limit = 0;
3801
3802         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
3803         if (!nlm)
3804                 return -1;
3805
3806         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
3807         nla_put_string(nlm, IFLA_IFNAME, name);
3808
3809         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
3810         if (!linkinfo) {
3811                 ret = -ENOMEM;
3812                 goto failure;
3813         }
3814
3815         nla_put_string(nlm, IFLA_INFO_KIND, kind);
3816         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
3817         if (!infodata) {
3818                 ret = -ENOMEM;
3819                 goto failure;
3820         }
3821
3822         if (link)
3823                 nla_put_u32(nlm, IFLA_GRE_LINK, link);
3824
3825         if ((cur = tb[TUNNEL_ATTR_TTL]))
3826                 ttl = blobmsg_get_u32(cur);
3827
3828         if ((cur = tb[TUNNEL_ATTR_TOS])) {
3829                 char *str = blobmsg_get_string(cur);
3830                 if (strcmp(str, "inherit")) {
3831                         unsigned uval;
3832
3833                         if (!system_tos_aton(str, &uval)) {
3834                                 ret = -EINVAL;
3835                                 goto failure;
3836                         }
3837
3838                         if (v6)
3839                                 flowinfo |= htonl(uval << 20) & IP6_FLOWINFO_TCLASS;
3840                         else
3841                                 tos = uval;
3842                 } else {
3843                         if (v6)
3844                                 flags6 |= IP6_TNL_F_USE_ORIG_TCLASS;
3845                         else
3846                                 tos = 1;
3847                 }
3848         }
3849
3850         if ((cur = tb[TUNNEL_ATTR_DATA])) {
3851                 struct blob_attr *tb_data[__GRE_DATA_ATTR_MAX];
3852
3853                 blobmsg_parse(gre_data_attr_list.params, __GRE_DATA_ATTR_MAX, tb_data,
3854                         blobmsg_data(cur), blobmsg_len(cur));
3855
3856                 if ((cur = tb_data[GRE_DATA_IKEY])) {
3857                         if ((ikey = blobmsg_get_u32(cur)))
3858                                 iflags |= GRE_KEY;
3859                 }
3860
3861                 if ((cur = tb_data[GRE_DATA_OKEY])) {
3862                         if ((okey = blobmsg_get_u32(cur)))
3863                                 oflags |= GRE_KEY;
3864                 }
3865
3866                 if ((cur = tb_data[GRE_DATA_ICSUM])) {
3867                         if (blobmsg_get_bool(cur))
3868                                 iflags |= GRE_CSUM;
3869                 }
3870
3871                 if ((cur = tb_data[GRE_DATA_OCSUM])) {
3872                         if (blobmsg_get_bool(cur))
3873                                 oflags |= GRE_CSUM;
3874                 }
3875
3876                 if ((cur = tb_data[GRE_DATA_ISEQNO])) {
3877                         if (blobmsg_get_bool(cur))
3878                                 iflags |= GRE_SEQ;
3879                 }
3880
3881                 if ((cur = tb_data[GRE_DATA_OSEQNO])) {
3882                         if (blobmsg_get_bool(cur))
3883                                 oflags |= GRE_SEQ;
3884                 }
3885
3886                 if ((cur = tb_data[GRE_DATA_ENCAPLIMIT])) {
3887                         char *str = blobmsg_get_string(cur);
3888
3889                         if (strcmp(str, "ignore")) {
3890                                 char *e;
3891
3892                                 encap_limit = strtoul(str, &e, 0);
3893
3894                                 if (e == str || *e || encap_limit > 255) {
3895                                         ret = -EINVAL;
3896                                         goto failure;
3897                                 }
3898
3899                                 flags6 &= ~IP6_TNL_F_IGN_ENCAP_LIMIT;
3900                         }
3901                 }
3902         }
3903
3904         if (v6) {
3905                 struct in6_addr in6buf;
3906                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3907                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3908                                 ret = -EINVAL;
3909                                 goto failure;
3910                         }
3911                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(in6buf), &in6buf);
3912                 }
3913
3914                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3915                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
3916                                 ret = -EINVAL;
3917                                 goto failure;
3918                         }
3919                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(in6buf), &in6buf);
3920                 }
3921
3922                 if (!(flags6 & IP6_TNL_F_IGN_ENCAP_LIMIT))
3923                         nla_put_u8(nlm, IFLA_GRE_ENCAP_LIMIT, encap_limit);
3924
3925                 if (flowinfo)
3926                         nla_put_u32(nlm, IFLA_GRE_FLOWINFO, flowinfo);
3927
3928                 if (flags6)
3929                         nla_put_u32(nlm, IFLA_GRE_FLAGS, flags6);
3930
3931                 if (!ttl)
3932                         ttl = 64;
3933         } else {
3934                 struct in_addr inbuf;
3935                 bool set_df = true;
3936
3937                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
3938                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3939                                 ret = -EINVAL;
3940                                 goto failure;
3941                         }
3942                         nla_put(nlm, IFLA_GRE_LOCAL, sizeof(inbuf), &inbuf);
3943                 }
3944
3945                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
3946                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
3947                                 ret = -EINVAL;
3948                                 goto failure;
3949                         }
3950                         nla_put(nlm, IFLA_GRE_REMOTE, sizeof(inbuf), &inbuf);
3951
3952                         if (IN_MULTICAST(ntohl(inbuf.s_addr))) {
3953                                 if (!okey) {
3954                                         okey = inbuf.s_addr;
3955                                         oflags |= GRE_KEY;
3956                                 }
3957
3958                                 if (!ikey) {
3959                                         ikey = inbuf.s_addr;
3960                                         iflags |= GRE_KEY;
3961                                 }
3962                         }
3963                 }
3964
3965                 if ((cur = tb[TUNNEL_ATTR_DF]))
3966                         set_df = blobmsg_get_bool(cur);
3967
3968                 if (!set_df) {
3969                         /* ttl != 0 and nopmtudisc are incompatible */
3970                         if (ttl) {
3971                                 ret = -EINVAL;
3972                                 goto failure;
3973                         }
3974                 } else if (!ttl)
3975                         ttl = 64;
3976
3977                 nla_put_u8(nlm, IFLA_GRE_PMTUDISC, set_df ? 1 : 0);
3978
3979                 nla_put_u8(nlm, IFLA_GRE_TOS, tos);
3980         }
3981
3982         if (ttl)
3983                 nla_put_u8(nlm, IFLA_GRE_TTL, ttl);
3984
3985         if (oflags)
3986                 nla_put_u16(nlm, IFLA_GRE_OFLAGS, oflags);
3987
3988         if (iflags)
3989                 nla_put_u16(nlm, IFLA_GRE_IFLAGS, iflags);
3990
3991         if (okey)
3992                 nla_put_u32(nlm, IFLA_GRE_OKEY, htonl(okey));
3993
3994         if (ikey)
3995                 nla_put_u32(nlm, IFLA_GRE_IKEY, htonl(ikey));
3996
3997         nla_nest_end(nlm, infodata);
3998         nla_nest_end(nlm, linkinfo);
3999
4000         return system_rtnl_call(nlm);
4001
4002 failure:
4003         nlmsg_free(nlm);
4004         return ret;
4005 }
4006 #endif
4007
4008 #ifdef IFLA_VTI_MAX
4009 static int system_add_vti_tunnel(const char *name, const char *kind,
4010                                  const unsigned int link, struct blob_attr **tb, bool v6)
4011 {
4012         struct nl_msg *nlm;
4013         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
4014         struct blob_attr *cur;
4015         int ret = 0;
4016
4017         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
4018         if (!nlm)
4019                 return -1;
4020
4021         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
4022         nla_put_string(nlm, IFLA_IFNAME, name);
4023
4024         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
4025         if (!linkinfo) {
4026                 ret = -ENOMEM;
4027                 goto failure;
4028         }
4029
4030         nla_put_string(nlm, IFLA_INFO_KIND, kind);
4031         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
4032         if (!infodata) {
4033                 ret = -ENOMEM;
4034                 goto failure;
4035         }
4036
4037         if (link)
4038                 nla_put_u32(nlm, IFLA_VTI_LINK, link);
4039
4040         if (v6) {
4041                 struct in6_addr in6buf;
4042                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
4043                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
4044                                 ret = -EINVAL;
4045                                 goto failure;
4046                         }
4047                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(in6buf), &in6buf);
4048                 }
4049
4050                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
4051                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
4052                                 ret = -EINVAL;
4053                                 goto failure;
4054                         }
4055                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(in6buf), &in6buf);
4056                 }
4057
4058         } else {
4059                 struct in_addr inbuf;
4060
4061                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
4062                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
4063                                 ret = -EINVAL;
4064                                 goto failure;
4065                         }
4066                         nla_put(nlm, IFLA_VTI_LOCAL, sizeof(inbuf), &inbuf);
4067                 }
4068
4069                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
4070                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
4071                                 ret = -EINVAL;
4072                                 goto failure;
4073                         }
4074                         nla_put(nlm, IFLA_VTI_REMOTE, sizeof(inbuf), &inbuf);
4075                 }
4076
4077         }
4078
4079         if ((cur = tb[TUNNEL_ATTR_DATA])) {
4080                 struct blob_attr *tb_data[__VTI_DATA_ATTR_MAX];
4081                 uint32_t ikey = 0, okey = 0;
4082
4083                 blobmsg_parse(vti_data_attr_list.params, __VTI_DATA_ATTR_MAX, tb_data,
4084                         blobmsg_data(cur), blobmsg_len(cur));
4085
4086                 if ((cur = tb_data[VTI_DATA_IKEY])) {
4087                         if ((ikey = blobmsg_get_u32(cur)))
4088                                 nla_put_u32(nlm, IFLA_VTI_IKEY, htonl(ikey));
4089                 }
4090
4091                 if ((cur = tb_data[VTI_DATA_OKEY])) {
4092                         if ((okey = blobmsg_get_u32(cur)))
4093                                 nla_put_u32(nlm, IFLA_VTI_OKEY, htonl(okey));
4094                 }
4095         }
4096
4097         nla_nest_end(nlm, infodata);
4098         nla_nest_end(nlm, linkinfo);
4099
4100         return system_rtnl_call(nlm);
4101
4102 failure:
4103         nlmsg_free(nlm);
4104         return ret;
4105 }
4106 #endif
4107
4108 #ifdef IFLA_XFRM_MAX
4109 static int system_add_xfrm_tunnel(const char *name, const char *kind,
4110                                  const unsigned int link, struct blob_attr **tb)
4111 {
4112         struct nl_msg *nlm;
4113         struct ifinfomsg ifi = { .ifi_family = AF_UNSPEC, };
4114         struct blob_attr *cur;
4115         int ret = 0;
4116
4117         nlm = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_REPLACE | NLM_F_CREATE);
4118         if (!nlm)
4119                 return -1;
4120
4121         nlmsg_append(nlm, &ifi, sizeof(ifi), 0);
4122         nla_put_string(nlm, IFLA_IFNAME, name);
4123
4124         struct nlattr *linkinfo = nla_nest_start(nlm, IFLA_LINKINFO);
4125         if (!linkinfo) {
4126                 ret = -ENOMEM;
4127                 goto failure;
4128         }
4129
4130         nla_put_string(nlm, IFLA_INFO_KIND, kind);
4131         struct nlattr *infodata = nla_nest_start(nlm, IFLA_INFO_DATA);
4132         if (!infodata) {
4133                 ret = -ENOMEM;
4134                 goto failure;
4135         }
4136
4137         if (link)
4138                 nla_put_u32(nlm, IFLA_XFRM_LINK, link);
4139
4140         if ((cur = tb[TUNNEL_ATTR_DATA])) {
4141                 struct blob_attr *tb_data[__XFRM_DATA_ATTR_MAX];
4142                 uint32_t if_id = 0;
4143
4144                 blobmsg_parse(xfrm_data_attr_list.params, __XFRM_DATA_ATTR_MAX, tb_data,
4145                         blobmsg_data(cur), blobmsg_len(cur));
4146
4147                 if ((cur = tb_data[XFRM_DATA_IF_ID])) {
4148                         if ((if_id = blobmsg_get_u32(cur)))
4149                                 nla_put_u32(nlm, IFLA_XFRM_IF_ID, if_id);
4150                 }
4151
4152         }
4153
4154         nla_nest_end(nlm, infodata);
4155         nla_nest_end(nlm, linkinfo);
4156
4157         return system_rtnl_call(nlm);
4158
4159 failure:
4160         nlmsg_free(nlm);
4161         return ret;
4162 }
4163 #endif
4164
4165 #ifdef IFLA_VXLAN_MAX
4166 static void system_vxlan_map_bool_attr(struct nl_msg *msg, struct blob_attr **tb_data, int attrtype, int vxlandatatype, bool invert) {
4167         struct blob_attr *cur;
4168         if ((cur = tb_data[vxlandatatype])) {
4169                 bool val = blobmsg_get_bool(cur);
4170                 if (invert)
4171                         val = !val;
4172
4173                 if ((attrtype == IFLA_VXLAN_GBP) && val)
4174                         nla_put_flag(msg, attrtype);
4175                 else
4176                         nla_put_u8(msg, attrtype, val);
4177
4178         }
4179 }
4180
4181 static int system_add_vxlan(const char *name, const unsigned int link, struct blob_attr **tb, bool v6)
4182 {
4183         struct blob_attr *tb_data[__VXLAN_DATA_ATTR_MAX];
4184         struct nl_msg *msg;
4185         struct nlattr *linkinfo, *data;
4186         struct ifinfomsg iim = { .ifi_family = AF_UNSPEC, };
4187         struct blob_attr *cur;
4188         int ret = 0;
4189
4190         if ((cur = tb[TUNNEL_ATTR_DATA]))
4191                 blobmsg_parse(vxlan_data_attr_list.params, __VXLAN_DATA_ATTR_MAX, tb_data,
4192                         blobmsg_data(cur), blobmsg_len(cur));
4193         else
4194                 return -EINVAL;
4195
4196         msg = nlmsg_alloc_simple(RTM_NEWLINK, NLM_F_REQUEST | NLM_F_CREATE | NLM_F_EXCL);
4197
4198         if (!msg)
4199                 return -1;
4200
4201         nlmsg_append(msg, &iim, sizeof(iim), 0);
4202
4203         nla_put_string(msg, IFLA_IFNAME, name);
4204
4205         if ((cur = tb_data[VXLAN_DATA_ATTR_MACADDR])) {
4206                 struct ether_addr *ea = ether_aton(blobmsg_get_string(cur));
4207                 if (!ea) {
4208                         ret = -EINVAL;
4209                         goto failure;
4210                 }
4211
4212                 nla_put(msg, IFLA_ADDRESS, ETH_ALEN, ea);
4213         }
4214
4215         if ((cur = tb[TUNNEL_ATTR_MTU])) {
4216                 uint32_t mtu = blobmsg_get_u32(cur);
4217                 nla_put_u32(msg, IFLA_MTU, mtu);
4218         }
4219
4220         if (!(linkinfo = nla_nest_start(msg, IFLA_LINKINFO))) {
4221                 ret = -ENOMEM;
4222                 goto failure;
4223         }
4224
4225         nla_put_string(msg, IFLA_INFO_KIND, "vxlan");
4226
4227         if (!(data = nla_nest_start(msg, IFLA_INFO_DATA))) {
4228                 ret = -ENOMEM;
4229                 goto failure;
4230         }
4231
4232         if (link)
4233                 nla_put_u32(msg, IFLA_VXLAN_LINK, link);
4234
4235         if ((cur = tb_data[VXLAN_DATA_ATTR_ID])) {
4236                 uint32_t id = blobmsg_get_u32(cur);
4237                 if (id >= (1u << 24) - 1) {
4238                         ret = -EINVAL;
4239                         goto failure;
4240                 }
4241
4242                 nla_put_u32(msg, IFLA_VXLAN_ID, id);
4243         }
4244
4245         if (v6) {
4246                 struct in6_addr in6buf;
4247                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
4248                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
4249                                 ret = -EINVAL;
4250                                 goto failure;
4251                         }
4252                         nla_put(msg, IFLA_VXLAN_LOCAL6, sizeof(in6buf), &in6buf);
4253                 }
4254
4255                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
4256                         if (inet_pton(AF_INET6, blobmsg_data(cur), &in6buf) < 1) {
4257                                 ret = -EINVAL;
4258                                 goto failure;
4259                         }
4260                         nla_put(msg, IFLA_VXLAN_GROUP6, sizeof(in6buf), &in6buf);
4261                 }
4262         } else {
4263                 struct in_addr inbuf;
4264
4265                 if ((cur = tb[TUNNEL_ATTR_LOCAL])) {
4266                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
4267                                 ret = -EINVAL;
4268                                 goto failure;
4269                         }
4270                         nla_put(msg, IFLA_VXLAN_LOCAL, sizeof(inbuf), &inbuf);
4271                 }
4272
4273                 if ((cur = tb[TUNNEL_ATTR_REMOTE])) {
4274                         if (inet_pton(AF_INET, blobmsg_data(cur), &inbuf) < 1) {
4275                                 ret = -EINVAL;
4276                                 goto failure;
4277                         }
4278                         nla_put(msg, IFLA_VXLAN_GROUP, sizeof(inbuf), &inbuf);
4279                 }
4280         }
4281
4282         uint32_t port = 4789;
4283         if ((cur = tb_data[VXLAN_DATA_ATTR_PORT])) {
4284                 port = blobmsg_get_u32(cur);
4285                 if (port < 1 || port > 65535) {
4286                         ret = -EINVAL;
4287                         goto failure;
4288                 }
4289         }
4290         nla_put_u16(msg, IFLA_VXLAN_PORT, htons(port));
4291
4292         if ((cur = tb_data[VXLAN_DATA_ATTR_SRCPORTMIN])) {
4293                 struct ifla_vxlan_port_range srcports = {0,0};
4294
4295                 uint32_t low = blobmsg_get_u32(cur);
4296                 if (low < 1 || low > 65535 - 1) {
4297                         ret = -EINVAL;
4298                         goto failure;
4299                 }
4300
4301                 srcports.low = htons((uint16_t) low);
4302                 srcports.high = htons((uint16_t) (low+1));
4303
4304                 if ((cur = tb_data[VXLAN_DATA_ATTR_SRCPORTMAX])) {
4305                         uint32_t high = blobmsg_get_u32(cur);
4306                         if (high < 1 || high > 65535) {
4307                                 ret = -EINVAL;
4308                                 goto failure;
4309                         }
4310
4311                         if (high > low)
4312                                 srcports.high = htons((uint16_t) high);
4313                 }
4314
4315                 nla_put(msg, IFLA_VXLAN_PORT_RANGE, sizeof(srcports), &srcports);
4316         }
4317
4318         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_UDP_CSUM, VXLAN_DATA_ATTR_TXCSUM, false);
4319         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_UDP_ZERO_CSUM6_RX, VXLAN_DATA_ATTR_RXCSUM, true);
4320         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_UDP_ZERO_CSUM6_TX, VXLAN_DATA_ATTR_TXCSUM, true);
4321         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_LEARNING, VXLAN_DATA_ATTR_LEARNING, false);
4322         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_RSC , VXLAN_DATA_ATTR_RSC, false);
4323         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_PROXY , VXLAN_DATA_ATTR_PROXY, false);
4324         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_L2MISS , VXLAN_DATA_ATTR_L2MISS, false);
4325         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_L3MISS , VXLAN_DATA_ATTR_L3MISS, false);
4326         system_vxlan_map_bool_attr(msg, tb_data, IFLA_VXLAN_GBP , VXLAN_DATA_ATTR_GBP, false);
4327
4328         if ((cur = tb_data[VXLAN_DATA_ATTR_AGEING])) {
4329                 uint32_t ageing = blobmsg_get_u32(cur);
4330                 nla_put_u32(msg, IFLA_VXLAN_AGEING, ageing);
4331         }
4332
4333         if ((cur = tb_data[VXLAN_DATA_ATTR_LIMIT])) {
4334                 uint32_t maxaddress = blobmsg_get_u32(cur);
4335                 nla_put_u32(msg, IFLA_VXLAN_LIMIT, maxaddress);
4336         }
4337
4338         if ((cur = tb[TUNNEL_ATTR_TOS])) {
4339                 char *str = blobmsg_get_string(cur);
4340                 unsigned tos = 1;
4341
4342                 if (strcmp(str, "inherit")) {
4343                         if (!system_tos_aton(str, &tos)) {
4344                                 ret = -EINVAL;
4345                                 goto failure;
4346                         }
4347                 }
4348
4349                 nla_put_u8(msg, IFLA_VXLAN_TOS, tos);
4350         }
4351
4352         if ((cur = tb[TUNNEL_ATTR_TTL])) {
4353                 uint32_t ttl = blobmsg_get_u32(cur);
4354                 if (ttl < 1 || ttl > 255) {
4355                         ret = -EINVAL;
4356                         goto failure;
4357                 }
4358
4359                 nla_put_u8(msg, IFLA_VXLAN_TTL, ttl);
4360         }
4361
4362         nla_nest_end(msg, data);
4363         nla_nest_end(msg, linkinfo);
4364
4365         ret = system_rtnl_call(msg);
4366         if (ret)
4367                 D(SYSTEM, "Error adding vxlan '%s': %d\n", name, ret);
4368
4369         return ret;
4370
4371 failure:
4372         nlmsg_free(msg);
4373         return ret;
4374 }
4375 #endif
4376
4377 static int system_add_sit_tunnel(const char *name, const unsigned int link, struct blob_attr **tb)
4378 {
4379         struct blob_attr *cur;
4380         int ret = 0;
4381
4382         if (system_add_proto_tunnel(name, IPPROTO_IPV6, link, tb) < 0)
4383                 return -1;
4384
4385 #ifdef SIOCADD6RD
4386         if ((cur = tb[TUNNEL_ATTR_DATA])) {
4387                 struct blob_attr *tb_data[__SIXRD_DATA_ATTR_MAX];
4388                 unsigned int mask;
4389                 struct ip_tunnel_6rd p6;
4390
4391                 blobmsg_parse(sixrd_data_attr_list.params, __SIXRD_DATA_ATTR_MAX, tb_data,
4392                         blobmsg_data(cur), blobmsg_len(cur));
4393
4394                 memset(&p6, 0, sizeof(p6));
4395
4396                 if ((cur = tb_data[SIXRD_DATA_PREFIX])) {
4397                         if (!parse_ip_and_netmask(AF_INET6, blobmsg_data(cur),
4398                                                 &p6.prefix, &mask) || mask > 128) {
4399                                 ret = -EINVAL;
4400                                 goto failure;
4401                         }
4402
4403                         p6.prefixlen = mask;
4404                 }
4405
4406                 if ((cur = tb_data[SIXRD_DATA_RELAY_PREFIX])) {
4407                         if (!parse_ip_and_netmask(AF_INET, blobmsg_data(cur),
4408                                                 &p6.relay_prefix, &mask) || mask > 32) {
4409                                 ret = -EINVAL;
4410                                 goto failure;
4411                         }
4412
4413                         p6.relay_prefixlen = mask;
4414                 }
4415
4416                 if (tunnel_ioctl(name, SIOCADD6RD, &p6) < 0) {
4417                         ret = -1;
4418                         goto failure;
4419                 }
4420         }
4421 #endif
4422
4423         return ret;
4424
4425 failure:
4426         system_link_del(name);
4427         return ret;
4428 }
4429
4430 static int system_add_proto_tunnel(const char *name, const uint8_t proto, const unsigned int link, struct blob_attr **tb)
4431 {
4432         struct blob_attr *cur;
4433         bool set_df = true;
4434         struct ip_tunnel_parm p  = {
4435                 .link = link,
4436                 .iph = {
4437                         .version = 4,
4438                         .ihl = 5,
4439                         .protocol = proto,
4440                 }
4441         };
4442
4443         if ((cur = tb[TUNNEL_ATTR_LOCAL]) &&
4444                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.saddr) < 1)
4445                 return -EINVAL;
4446
4447         if ((cur = tb[TUNNEL_ATTR_REMOTE]) &&
4448                         inet_pton(AF_INET, blobmsg_data(cur), &p.iph.daddr) < 1)
4449                 return -EINVAL;
4450
4451         if ((cur = tb[TUNNEL_ATTR_DF]))
4452                 set_df = blobmsg_get_bool(cur);
4453
4454         if ((cur = tb[TUNNEL_ATTR_TTL]))
4455                 p.iph.ttl = blobmsg_get_u32(cur);
4456
4457         if ((cur = tb[TUNNEL_ATTR_TOS])) {
4458                 char *str = blobmsg_get_string(cur);
4459                 if (strcmp(str, "inherit")) {
4460                         unsigned uval;
4461
4462                         if (!system_tos_aton(str, &uval))
4463                                 return -EINVAL;
4464
4465                         p.iph.tos = uval;
4466                 } else
4467                         p.iph.tos = 1;
4468         }
4469
4470         p.iph.frag_off = set_df ? htons(IP_DF) : 0;
4471         /* ttl !=0 and nopmtudisc are incompatible */
4472         if (p.iph.ttl && p.iph.frag_off == 0)
4473                 return -EINVAL;
4474
4475         strncpy(p.name, name, sizeof(p.name) - 1);
4476
4477         switch (p.iph.protocol) {
4478         case IPPROTO_IPIP:
4479                 return tunnel_ioctl("tunl0", SIOCADDTUNNEL, &p);
4480         case IPPROTO_IPV6:
4481                 return tunnel_ioctl("sit0", SIOCADDTUNNEL, &p);
4482         default:
4483                 break;
4484         }
4485         return -1;
4486 }
4487
4488 int system_del_ip_tunnel(const struct device *dev)
4489 {
4490         return system_link_del(dev->ifname);
4491 }
4492
4493 int system_update_ipv6_mtu(struct device *dev, int mtu)
4494 {
4495         int ret = -1;
4496         char buf[64];
4497         int fd;
4498
4499         fd = open(dev_sysctl_path("ipv6/conf", dev->ifname, "mtu"), O_RDWR);
4500         if (fd < 0)
4501                 return ret;
4502
4503         if (!mtu) {
4504                 ssize_t len = read(fd, buf, sizeof(buf) - 1);
4505                 if (len < 0)
4506                         goto out;
4507
4508                 buf[len] = 0;
4509                 ret = atoi(buf);
4510         } else {
4511                 if (write(fd, buf, snprintf(buf, sizeof(buf), "%i", mtu)) > 0)
4512                         ret = mtu;
4513         }
4514
4515 out:
4516         close(fd);
4517         return ret;
4518 }
4519
4520 int system_add_ip_tunnel(const struct device *dev, struct blob_attr *attr)
4521 {
4522         struct blob_attr *tb[__TUNNEL_ATTR_MAX];
4523         struct blob_attr *cur;
4524         const char *str;
4525
4526         blobmsg_parse(tunnel_attr_list.params, __TUNNEL_ATTR_MAX, tb,
4527                 blob_data(attr), blob_len(attr));
4528
4529         system_link_del(dev->ifname);
4530
4531         if (!(cur = tb[TUNNEL_ATTR_TYPE]))
4532                 return -EINVAL;
4533         str = blobmsg_data(cur);
4534
4535         unsigned int ttl = 0;
4536         if ((cur = tb[TUNNEL_ATTR_TTL])) {
4537                 ttl = blobmsg_get_u32(cur);
4538                 if (ttl > 255)
4539                         return -EINVAL;
4540         }
4541
4542         unsigned int link = 0;
4543         if ((cur = tb[TUNNEL_ATTR_LINK])) {
4544                 struct interface *iface = vlist_find(&interfaces, blobmsg_data(cur), iface, node);
4545                 if (!iface)
4546                         return -EINVAL;
4547
4548                 if (iface->l3_dev.dev)
4549                         link = iface->l3_dev.dev->ifindex;
4550         }
4551
4552         if (!strcmp(str, "sit"))
4553                 return system_add_sit_tunnel(dev->ifname, link, tb);
4554 #ifdef IFLA_IPTUN_MAX
4555         else if (!strcmp(str, "ipip6")) {
4556                 return system_add_ip6_tunnel(dev->ifname, link, tb);
4557         } else if (!strcmp(str, "greip")) {
4558                 return system_add_gre_tunnel(dev->ifname, "gre", link, tb, false);
4559         } else if (!strcmp(str, "gretapip"))  {
4560                 return system_add_gre_tunnel(dev->ifname, "gretap", link, tb, false);
4561         } else if (!strcmp(str, "greip6")) {
4562                 return system_add_gre_tunnel(dev->ifname, "ip6gre", link, tb, true);
4563         } else if (!strcmp(str, "gretapip6")) {
4564                 return system_add_gre_tunnel(dev->ifname, "ip6gretap", link, tb, true);
4565 #ifdef IFLA_VTI_MAX
4566         } else if (!strcmp(str, "vtiip")) {
4567                 return system_add_vti_tunnel(dev->ifname, "vti", link, tb, false);
4568         } else if (!strcmp(str, "vtiip6")) {
4569                 return system_add_vti_tunnel(dev->ifname, "vti6", link, tb, true);
4570 #endif
4571 #ifdef IFLA_XFRM_MAX
4572         } else if (!strcmp(str, "xfrm")) {
4573                 return system_add_xfrm_tunnel(dev->ifname, "xfrm", link, tb);
4574 #endif
4575 #ifdef IFLA_VXLAN_MAX
4576         } else if(!strcmp(str, "vxlan")) {
4577                 return system_add_vxlan(dev->ifname, link, tb, false);
4578         } else if(!strcmp(str, "vxlan6")) {
4579                 return system_add_vxlan(dev->ifname, link, tb, true);
4580 #endif
4581 #endif
4582         } else if (!strcmp(str, "ipip")) {
4583                 return system_add_proto_tunnel(dev->ifname, IPPROTO_IPIP, link, tb);
4584         }
4585         else
4586                 return -EINVAL;
4587
4588         return 0;
4589 }