system-linux.c

   1 #define _GNU_SOURCE
   2
   3 #include <sys/socket.h>
   4 #include <sys/ioctl.h>
   5 #include <sys/stat.h>
   6 #include <sys/syscall.h>
   7
   8 #include <linux/rtnetlink.h>
   9 #include <linux/sockios.h>
  10 #include <linux/if_vlan.h>
  11 #include <linux/if_bridge.h>
  12
  13 #include <unistd.h>
  14 #include <string.h>
  15 #include <fcntl.h>
  16 #include <glob.h>
  17
  18 #include <netlink/msg.h>
  19 #include <netlink/attr.h>
  20 #include <netlink/socket.h>
  21 #include <libubox/uloop.h>
  22
  23 #include "netifd.h"
  24 #include "device.h"
  25 #include "system.h"
  26
  27 static int sock_ioctl = -1;
  28 static struct nl_sock *sock_rtnl = NULL;
  29 static struct nl_sock *sock_rtnl_event = NULL;
  30
  31 static void handler_rtnl_event(struct uloop_fd *u, unsigned int events);
  32 static int cb_rtnl_event(struct nl_msg *msg, void *arg);
  33 static struct uloop_fd rtnl_event = {.cb = handler_rtnl_event};
  34 static struct nl_cb *nl_cb_rtnl_event;
  35
  36 int system_init(void)
  37 {
  38         sock_ioctl = socket(AF_LOCAL, SOCK_DGRAM, 0);
  39         fcntl(sock_ioctl, F_SETFD, fcntl(sock_ioctl, F_GETFD) | FD_CLOEXEC);
  40
  41         // Prepare socket for routing / address control
  42         sock_rtnl = nl_socket_alloc();
  43         if (!sock_rtnl)
  44                 return -1;
  45
  46         if (nl_connect(sock_rtnl, NETLINK_ROUTE))
  47                 goto error_free_sock;
  48
  49         // Prepare socket for link events
  50         nl_cb_rtnl_event = nl_cb_alloc(NL_CB_DEFAULT);
  51         if (!nl_cb_rtnl_event)
  52                 goto error_free_sock;
  53
  54         nl_cb_set(nl_cb_rtnl_event, NL_CB_VALID, NL_CB_CUSTOM,
  55                   cb_rtnl_event, NULL);
  56
  57         sock_rtnl_event = nl_socket_alloc();
  58         if (!sock_rtnl_event)
  59                 goto error_free_cb;
  60
  61         if (nl_connect(sock_rtnl_event, NETLINK_ROUTE))
  62                 goto error_free_event;
  63
  64         // Receive network link events form kernel
  65         nl_socket_add_membership(sock_rtnl_event, RTNLGRP_LINK);
  66
  67         rtnl_event.fd = nl_socket_get_fd(sock_rtnl_event);
  68         uloop_fd_add(&rtnl_event, ULOOP_READ | ULOOP_EDGE_TRIGGER);
  69
  70         return 0;
  71
  72 error_free_event:
  73         nl_socket_free(sock_rtnl_event);
  74         sock_rtnl_event = NULL;
  75 error_free_cb:
  76         nl_cb_put(nl_cb_rtnl_event);
  77         nl_cb_rtnl_event = NULL;
  78 error_free_sock:
  79         nl_socket_free(sock_rtnl);
  80         sock_rtnl = NULL;
  81         return -1;
  82 }
  83
  84 // If socket is ready for reading parse netlink events
  85 static void handler_rtnl_event(struct uloop_fd *u, unsigned int events)
  86 {
  87         nl_recvmsgs(sock_rtnl_event, nl_cb_rtnl_event);
  88 }
  89
  90 static void system_set_sysctl(const char *path, const char *val)
  91 {
  92         int fd;
  93
  94         fd = open(path, O_WRONLY);
  95         if (fd < 0)
  96                 return;
  97
  98         write(fd, val, strlen(val));
  99         close(fd);
 100 }
 101
 102 static void system_set_dev_sysctl(const char *path, const char *device, const char *val)
 103 {
 104         char buf[256];
 105
 106         snprintf(buf, sizeof(buf), path, val);
 107         system_set_sysctl(buf, val);
 108 }
 109
 110 static void system_set_disable_ipv6(struct device *dev, const char *val)
 111 {
 112         system_set_dev_sysctl("/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname, val);
 113 }
 114
 115 // Evaluate netlink messages
 116 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 117 {
 118         struct nlmsghdr *nh = nlmsg_hdr(msg);
 119         struct ifinfomsg *ifi = NLMSG_DATA(nh);
 120         struct nlattr *nla[__IFLA_MAX];
 121
 122         if (nh->nlmsg_type != RTM_DELLINK && nh->nlmsg_type != RTM_NEWLINK)
 123                 goto out;
 124
 125         nlmsg_parse(nh, sizeof(*ifi), nla, __IFLA_MAX - 1, NULL);
 126         if (!nla[IFLA_IFNAME])
 127                 goto out;
 128
 129         struct device *dev = device_get(RTA_DATA(nla[IFLA_IFNAME]), false);
 130         if (!dev)
 131                 goto out;
 132
 133         dev->ifindex = ifi->ifi_index;
 134         device_set_present(dev, (nh->nlmsg_type == RTM_NEWLINK));
 135
 136 out:
 137         return 0;
 138 }
 139
 140 static int system_rtnl_call(struct nl_msg *msg)
 141 {
 142         int s = -(nl_send_auto_complete(sock_rtnl, msg)
 143                         || nl_wait_for_ack(sock_rtnl));
 144         nlmsg_free(msg);
 145         return s;
 146 }
 147
 148 int system_bridge_delbr(struct device *bridge)
 149 {
 150         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 151 }
 152
 153 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 154 {
 155         struct ifreq ifr;
 156         if (dev)
 157                 ifr.ifr_ifindex = dev->ifindex;
 158         else
 159                 ifr.ifr_data = data;
 160         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name));
 161         return ioctl(sock_ioctl, cmd, &ifr);
 162 }
 163
 164 int system_bridge_addif(struct device *bridge, struct device *dev)
 165 {
 166         system_set_disable_ipv6(dev, "1");
 167         return system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 168 }
 169
 170 int system_bridge_delif(struct device *bridge, struct device *dev)
 171 {
 172         system_set_disable_ipv6(dev, "0");
 173         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 174 }
 175
 176 static bool system_is_bridge(const char *name, char *buf, int buflen)
 177 {
 178         struct stat st;
 179
 180         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 181         if (stat(buf, &st) < 0)
 182                 return false;
 183
 184         return true;
 185 }
 186
 187 static char *system_get_bridge(const char *name, char *buf, int buflen)
 188 {
 189         char *path;
 190         ssize_t len;
 191         glob_t gl;
 192
 193         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 194         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 195                 return NULL;
 196
 197         if (gl.gl_pathc == 0)
 198                 return NULL;
 199
 200         len = readlink(gl.gl_pathv[0], buf, buflen);
 201         if (len < 0)
 202                 return NULL;
 203
 204         buf[len] = 0;
 205         path = strrchr(buf, '/');
 206         if (!path)
 207                 return NULL;
 208
 209         return path + 1;
 210 }
 211
 212 static int system_if_resolve(struct device *dev)
 213 {
 214         struct ifreq ifr;
 215         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
 216         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 217                 return ifr.ifr_ifindex;
 218         else
 219                 return 0;
 220 }
 221
 222 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 223 {
 224         struct ifreq ifr;
 225         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name));
 226         ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr);
 227         ifr.ifr_flags |= add;
 228         ifr.ifr_flags &= ~rem;
 229         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 230 }
 231
 232 struct clear_data {
 233         struct nl_msg *msg;
 234         struct device *dev;
 235         int type;
 236         int size;
 237         int af;
 238 };
 239
 240
 241 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 242 {
 243         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 244
 245         return ifa->ifa_index == ifindex;
 246 }
 247
 248 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 249 {
 250         struct nlattr *tb[__RTA_MAX];
 251
 252         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 253         if (!tb[RTA_OIF])
 254                 return false;
 255
 256         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 257 }
 258
 259 static int cb_clear_event(struct nl_msg *msg, void *arg)
 260 {
 261         struct clear_data *clr = arg;
 262         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 263         bool (*cb)(struct nlmsghdr *, int ifindex);
 264         int type;
 265
 266         switch(clr->type) {
 267         case RTM_GETADDR:
 268                 type = RTM_DELADDR;
 269                 if (hdr->nlmsg_type != RTM_NEWADDR)
 270                         return NL_SKIP;
 271
 272                 cb = check_ifaddr;
 273                 break;
 274         case RTM_GETROUTE:
 275                 type = RTM_DELROUTE;
 276                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 277                         return NL_SKIP;
 278
 279                 cb = check_route;
 280                 break;
 281         default:
 282                 return NL_SKIP;
 283         }
 284
 285         if (!cb(hdr, clr->dev->ifindex))
 286                 return NL_SKIP;
 287
 288         D(SYSTEM, "Remove %s from device %s\n",
 289           type == RTM_DELADDR ? "an address" : "a route",
 290           clr->dev->ifname);
 291         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 292         hdr = nlmsg_hdr(clr->msg);
 293         hdr->nlmsg_type = type;
 294         hdr->nlmsg_flags = NLM_F_REQUEST;
 295
 296         if (!nl_send_auto_complete(sock_rtnl, clr->msg))
 297                 nl_wait_for_ack(sock_rtnl);
 298
 299         return NL_SKIP;
 300 }
 301
 302 static int
 303 cb_finish_event(struct nl_msg *msg, void *arg)
 304 {
 305         int *pending = arg;
 306         *pending = 0;
 307         return NL_STOP;
 308 }
 309
 310 static int
 311 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 312 {
 313         int *pending = arg;
 314         *pending = err->error;
 315         return NL_STOP;
 316 }
 317
 318 static void
 319 system_if_clear_entries(struct device *dev, int type, int af)
 320 {
 321         struct clear_data clr;
 322         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
 323         struct rtmsg rtm = {
 324                 .rtm_family = af,
 325                 .rtm_flags = RTM_F_CLONED,
 326         };
 327         int flags = NLM_F_DUMP;
 328         int pending = 1;
 329
 330         clr.af = af;
 331         clr.dev = dev;
 332         clr.type = type;
 333         switch (type) {
 334         case RTM_GETADDR:
 335                 clr.size = sizeof(struct rtgenmsg);
 336                 break;
 337         case RTM_GETROUTE:
 338                 clr.size = sizeof(struct rtmsg);
 339                 break;
 340         default:
 341                 return;
 342         }
 343
 344         if (!cb)
 345                 return;
 346
 347         clr.msg = nlmsg_alloc_simple(type, flags);
 348         if (!clr.msg)
 349                 goto out;
 350
 351         nlmsg_append(clr.msg, &rtm, clr.size, 0);
 352         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
 353         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
 354         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
 355
 356         nl_send_auto_complete(sock_rtnl, clr.msg);
 357         while (pending > 0)
 358                 nl_recvmsgs(sock_rtnl, cb);
 359
 360         nlmsg_free(clr.msg);
 361 out:
 362         nl_cb_put(cb);
 363 }
 364
 365 /*
 366  * Clear bridge (membership) state and bring down device
 367  */
 368 void system_if_clear_state(struct device *dev)
 369 {
 370         char buf[256];
 371         char *bridge;
 372
 373         dev->ifindex = system_if_resolve(dev);
 374         if (!dev->ifindex)
 375                 return;
 376
 377         system_if_flags(dev->ifname, 0, IFF_UP);
 378
 379         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
 380                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
 381                 system_bridge_delbr(dev);
 382                 return;
 383         }
 384
 385         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
 386         if (bridge) {
 387                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
 388                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
 389         }
 390
 391         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
 392         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
 393         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
 394         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
 395         system_set_disable_ipv6(dev, "0");
 396 }
 397
 398 static inline unsigned long
 399 sec_to_jiffies(int val)
 400 {
 401         return (unsigned long) val * 100;
 402 }
 403
 404 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
 405 {
 406         unsigned long args[4] = {};
 407
 408         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
 409                 return -1;
 410
 411         args[0] = BRCTL_SET_BRIDGE_STP_STATE;
 412         args[1] = !!cfg->stp;
 413         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 414
 415         args[0] = BRCTL_SET_BRIDGE_FORWARD_DELAY;
 416         args[1] = sec_to_jiffies(cfg->forward_delay);
 417         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 418
 419         system_set_dev_sysctl("/sys/devices/virtual/net/%s/bridge/multicast_snooping",
 420                 bridge->ifname, cfg->igmp_snoop ? "1" : "0");
 421
 422         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
 423                 args[0] = BRCTL_SET_AGEING_TIME;
 424                 args[1] = sec_to_jiffies(cfg->ageing_time);
 425                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 426         }
 427
 428         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
 429                 args[0] = BRCTL_SET_BRIDGE_HELLO_TIME;
 430                 args[1] = sec_to_jiffies(cfg->hello_time);
 431                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 432         }
 433
 434         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
 435                 args[0] = BRCTL_SET_BRIDGE_MAX_AGE;
 436                 args[1] = sec_to_jiffies(cfg->max_age);
 437                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 438         }
 439
 440         return 0;
 441 }
 442
 443 static int system_vlan(struct device *dev, int id)
 444 {
 445         struct vlan_ioctl_args ifr = {
 446                 .cmd = SET_VLAN_NAME_TYPE_CMD,
 447                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
 448         };
 449
 450         ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
 451
 452         if (id < 0) {
 453                 ifr.cmd = DEL_VLAN_CMD;
 454                 ifr.u.VID = 0;
 455         } else {
 456                 ifr.cmd = ADD_VLAN_CMD;
 457                 ifr.u.VID = id;
 458         }
 459         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
 460         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
 461 }
 462
 463 int system_vlan_add(struct device *dev, int id)
 464 {
 465         return system_vlan(dev, id);
 466 }
 467
 468 int system_vlan_del(struct device *dev)
 469 {
 470         return system_vlan(dev, -1);
 471 }
 472
 473 int system_if_up(struct device *dev)
 474 {
 475         dev->ifindex = system_if_resolve(dev);
 476         return system_if_flags(dev->ifname, IFF_UP, 0);
 477 }
 478
 479 int system_if_down(struct device *dev)
 480 {
 481         return system_if_flags(dev->ifname, 0, IFF_UP);
 482 }
 483
 484 int system_if_check(struct device *dev)
 485 {
 486         device_set_present(dev, (system_if_resolve(dev) >= 0));
 487         return 0;
 488 }
 489
 490 int system_if_dump_stats(struct device *dev, struct blob_buf *b)
 491 {
 492         const char *const counters[] = {
 493                 "collisions",     "rx_frame_errors",   "tx_compressed",
 494                 "multicast",      "rx_length_errors",  "tx_dropped",
 495                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
 496                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
 497                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
 498                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
 499                 "rx_errors",      "tx_bytes",          "tx_window_errors",
 500                 "rx_fifo_errors", "tx_carrier_errors",
 501         };
 502         char buf[64];
 503         int stats_dir;
 504         int i, fd, len;
 505
 506         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
 507         stats_dir = open(buf, O_DIRECTORY);
 508         if (stats_dir < 0)
 509                 return -1;
 510
 511         for (i = 0; i < ARRAY_SIZE(counters); i++) {
 512                 fd = openat(stats_dir, counters[i], O_RDONLY);
 513                 if (fd < 0)
 514                         continue;
 515
 516 retry:
 517                 len = read(fd, buf, sizeof(buf));
 518                 if (len < 0) {
 519                         if (errno == EINTR)
 520                                 goto retry;
 521                         continue;
 522                 }
 523
 524                 buf[len] = 0;
 525                 blobmsg_add_u32(b, counters[i], strtoul(buf, NULL, 0));
 526                 close(fd);
 527         }
 528
 529         close(stats_dir);
 530         return 0;
 531 }
 532
 533 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
 534 {
 535         int alen = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
 536         struct ifaddrmsg ifa = {
 537                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
 538                 .ifa_prefixlen = addr->mask,
 539                 .ifa_index = dev->ifindex,
 540         };
 541
 542         struct nl_msg *msg = nlmsg_alloc_simple(cmd, 0);
 543         if (!msg)
 544                 return -1;
 545
 546         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
 547         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
 548         return system_rtnl_call(msg);
 549 }
 550
 551 int system_add_address(struct device *dev, struct device_addr *addr)
 552 {
 553         return system_addr(dev, addr, RTM_NEWADDR);
 554 }
 555
 556 int system_del_address(struct device *dev, struct device_addr *addr)
 557 {
 558         return system_addr(dev, addr, RTM_DELADDR);
 559 }
 560
 561 static int system_rt(struct device *dev, struct device_route *route, int cmd)
 562 {
 563         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
 564         bool have_gw;
 565         unsigned int flags = 0;
 566
 567         if (alen == 4)
 568                 have_gw = !!route->nexthop.in.s_addr;
 569         else
 570                 have_gw = route->nexthop.in6.s6_addr32[0] ||
 571                         route->nexthop.in6.s6_addr32[1] ||
 572                         route->nexthop.in6.s6_addr32[2] ||
 573                         route->nexthop.in6.s6_addr32[3];
 574
 575         unsigned char scope = (cmd == RTM_DELROUTE) ? RT_SCOPE_NOWHERE :
 576                         (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
 577
 578         struct rtmsg rtm = {
 579                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
 580                 .rtm_dst_len = route->mask,
 581                 .rtm_table = RT_TABLE_MAIN,
 582                 .rtm_protocol = RTPROT_BOOT,
 583                 .rtm_scope = scope,
 584                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
 585         };
 586
 587         if (cmd == RTM_NEWROUTE)
 588                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
 589
 590         struct nl_msg *msg = nlmsg_alloc_simple(cmd, flags);
 591         if (!msg)
 592                 return -1;
 593
 594         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
 595
 596         if (route->mask)
 597                 nla_put(msg, RTA_DST, alen, &route->addr);
 598
 599         if (have_gw)
 600                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
 601
 602         if (route->flags & DEVADDR_DEVICE)
 603                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
 604
 605         return system_rtnl_call(msg);
 606 }
 607
 608 int system_add_route(struct device *dev, struct device_route *route)
 609 {
 610         return system_rt(dev, route, RTM_NEWROUTE);
 611 }
 612
 613 int system_del_route(struct device *dev, struct device_route *route)
 614 {
 615         return system_rt(dev, route, RTM_DELROUTE);
 616 }
 617
 618 int system_flush_routes(void)
 619 {
 620         const char *names[] = {
 621                 "/proc/sys/net/ipv4/route/flush",
 622                 "/proc/sys/net/ipv6/route/flush"
 623         };
 624         int fd, i;
 625
 626         for (i = 0; i < ARRAY_SIZE(names); i++) {
 627                 fd = open(names[i], O_WRONLY);
 628                 if (fd < 0)
 629                         continue;
 630
 631                 write(fd, "-1", 2);
 632                 close(fd);
 633         }
 634         return 0;
 635 }
 636
 637 time_t system_get_rtime(void)
 638 {
 639         struct timespec ts;
 640         struct timeval tv;
 641
 642         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
 643                 return ts.tv_sec;
 644
 645         if (gettimeofday(&tv, NULL) == 0)
 646                 return tv.tv_sec;
 647
 648         return 0;
 649 }