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_disable_ipv6(struct device *dev, const char *val)
 103 {
 104         char buf[256];
 105
 106         snprintf(buf, sizeof(buf), "/proc/sys/net/ipv6/conf/%s/disable_ipv6", dev->ifname);
 107         system_set_sysctl(buf, "0");
 108 }
 109
 110 // Evaluate netlink messages
 111 static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 112 {
 113         struct nlmsghdr *nh = nlmsg_hdr(msg);
 114         struct ifinfomsg *ifi = NLMSG_DATA(nh);
 115         struct nlattr *nla[__IFLA_MAX];
 116
 117         if (nh->nlmsg_type != RTM_DELLINK && nh->nlmsg_type != RTM_NEWLINK)
 118                 goto out;
 119
 120         nlmsg_parse(nh, sizeof(*ifi), nla, __IFLA_MAX - 1, NULL);
 121         if (!nla[IFLA_IFNAME])
 122                 goto out;
 123
 124         struct device *dev = device_get(RTA_DATA(nla[IFLA_IFNAME]), false);
 125         if (!dev)
 126                 goto out;
 127
 128         dev->ifindex = ifi->ifi_index;
 129         device_set_present(dev, (nh->nlmsg_type == RTM_NEWLINK));
 130
 131 out:
 132         return 0;
 133 }
 134
 135 static int system_rtnl_call(struct nl_msg *msg)
 136 {
 137         int s = -(nl_send_auto_complete(sock_rtnl, msg)
 138                         || nl_wait_for_ack(sock_rtnl));
 139         nlmsg_free(msg);
 140         return s;
 141 }
 142
 143 int system_bridge_delbr(struct device *bridge)
 144 {
 145         return ioctl(sock_ioctl, SIOCBRDELBR, bridge->ifname);
 146 }
 147
 148 static int system_bridge_if(const char *bridge, struct device *dev, int cmd, void *data)
 149 {
 150         struct ifreq ifr;
 151         if (dev)
 152                 ifr.ifr_ifindex = dev->ifindex;
 153         else
 154                 ifr.ifr_data = data;
 155         strncpy(ifr.ifr_name, bridge, sizeof(ifr.ifr_name));
 156         return ioctl(sock_ioctl, cmd, &ifr);
 157 }
 158
 159 int system_bridge_addif(struct device *bridge, struct device *dev)
 160 {
 161         system_set_disable_ipv6(dev, "1");
 162         return system_bridge_if(bridge->ifname, dev, SIOCBRADDIF, NULL);
 163 }
 164
 165 int system_bridge_delif(struct device *bridge, struct device *dev)
 166 {
 167         system_set_disable_ipv6(dev, "0");
 168         return system_bridge_if(bridge->ifname, dev, SIOCBRDELIF, NULL);
 169 }
 170
 171 static bool system_is_bridge(const char *name, char *buf, int buflen)
 172 {
 173         struct stat st;
 174
 175         snprintf(buf, buflen, "/sys/devices/virtual/net/%s/bridge", name);
 176         if (stat(buf, &st) < 0)
 177                 return false;
 178
 179         return true;
 180 }
 181
 182 static char *system_get_bridge(const char *name, char *buf, int buflen)
 183 {
 184         char *path;
 185         ssize_t len;
 186         glob_t gl;
 187
 188         snprintf(buf, buflen, "/sys/devices/virtual/net/*/brif/%s/bridge", name);
 189         if (glob(buf, GLOB_NOSORT, NULL, &gl) < 0)
 190                 return NULL;
 191
 192         if (gl.gl_pathc == 0)
 193                 return NULL;
 194
 195         len = readlink(gl.gl_pathv[0], buf, buflen);
 196         if (len < 0)
 197                 return NULL;
 198
 199         buf[len] = 0;
 200         path = strrchr(buf, '/');
 201         if (!path)
 202                 return NULL;
 203
 204         return path + 1;
 205 }
 206
 207 static int system_if_resolve(struct device *dev)
 208 {
 209         struct ifreq ifr;
 210         strncpy(ifr.ifr_name, dev->ifname, sizeof(ifr.ifr_name));
 211         if (!ioctl(sock_ioctl, SIOCGIFINDEX, &ifr))
 212                 return ifr.ifr_ifindex;
 213         else
 214                 return 0;
 215 }
 216
 217 static int system_if_flags(const char *ifname, unsigned add, unsigned rem)
 218 {
 219         struct ifreq ifr;
 220         strncpy(ifr.ifr_name, ifname, sizeof(ifr.ifr_name));
 221         ioctl(sock_ioctl, SIOCGIFFLAGS, &ifr);
 222         ifr.ifr_flags |= add;
 223         ifr.ifr_flags &= ~rem;
 224         return ioctl(sock_ioctl, SIOCSIFFLAGS, &ifr);
 225 }
 226
 227 struct clear_data {
 228         struct nl_msg *msg;
 229         struct device *dev;
 230         int type;
 231         int size;
 232         int af;
 233 };
 234
 235
 236 static bool check_ifaddr(struct nlmsghdr *hdr, int ifindex)
 237 {
 238         struct ifaddrmsg *ifa = NLMSG_DATA(hdr);
 239
 240         return ifa->ifa_index == ifindex;
 241 }
 242
 243 static bool check_route(struct nlmsghdr *hdr, int ifindex)
 244 {
 245         struct nlattr *tb[__RTA_MAX];
 246
 247         nlmsg_parse(hdr, sizeof(struct rtmsg), tb, __RTA_MAX - 1, NULL);
 248         if (!tb[RTA_OIF])
 249                 return false;
 250
 251         return *(int *)RTA_DATA(tb[RTA_OIF]) == ifindex;
 252 }
 253
 254 static int cb_clear_event(struct nl_msg *msg, void *arg)
 255 {
 256         struct clear_data *clr = arg;
 257         struct nlmsghdr *hdr = nlmsg_hdr(msg);
 258         bool (*cb)(struct nlmsghdr *, int ifindex);
 259         int type;
 260
 261         switch(clr->type) {
 262         case RTM_GETADDR:
 263                 type = RTM_DELADDR;
 264                 if (hdr->nlmsg_type != RTM_NEWADDR)
 265                         return NL_SKIP;
 266
 267                 cb = check_ifaddr;
 268                 break;
 269         case RTM_GETROUTE:
 270                 type = RTM_DELROUTE;
 271                 if (hdr->nlmsg_type != RTM_NEWROUTE)
 272                         return NL_SKIP;
 273
 274                 cb = check_route;
 275                 break;
 276         default:
 277                 return NL_SKIP;
 278         }
 279
 280         if (!cb(hdr, clr->dev->ifindex))
 281                 return NL_SKIP;
 282
 283         D(SYSTEM, "Remove %s from device %s\n",
 284           type == RTM_DELADDR ? "an address" : "a route",
 285           clr->dev->ifname);
 286         memcpy(nlmsg_hdr(clr->msg), hdr, hdr->nlmsg_len);
 287         hdr = nlmsg_hdr(clr->msg);
 288         hdr->nlmsg_type = type;
 289         hdr->nlmsg_flags = NLM_F_REQUEST;
 290
 291         if (!nl_send_auto_complete(sock_rtnl, clr->msg))
 292                 nl_wait_for_ack(sock_rtnl);
 293
 294         return NL_SKIP;
 295 }
 296
 297 static int
 298 cb_finish_event(struct nl_msg *msg, void *arg)
 299 {
 300         int *pending = arg;
 301         *pending = 0;
 302         return NL_STOP;
 303 }
 304
 305 static int
 306 error_handler(struct sockaddr_nl *nla, struct nlmsgerr *err, void *arg)
 307 {
 308         int *pending = arg;
 309         *pending = err->error;
 310         return NL_STOP;
 311 }
 312
 313 static void
 314 system_if_clear_entries(struct device *dev, int type, int af)
 315 {
 316         struct clear_data clr;
 317         struct nl_cb *cb = nl_cb_alloc(NL_CB_DEFAULT);
 318         struct rtmsg rtm = {
 319                 .rtm_family = af,
 320                 .rtm_flags = RTM_F_CLONED,
 321         };
 322         int flags = NLM_F_DUMP;
 323         int pending = 1;
 324
 325         clr.af = af;
 326         clr.dev = dev;
 327         clr.type = type;
 328         switch (type) {
 329         case RTM_GETADDR:
 330                 clr.size = sizeof(struct rtgenmsg);
 331                 break;
 332         case RTM_GETROUTE:
 333                 clr.size = sizeof(struct rtmsg);
 334                 break;
 335         default:
 336                 return;
 337         }
 338
 339         if (!cb)
 340                 return;
 341
 342         clr.msg = nlmsg_alloc_simple(type, flags);
 343         if (!clr.msg)
 344                 goto out;
 345
 346         nlmsg_append(clr.msg, &rtm, clr.size, 0);
 347         nl_cb_set(cb, NL_CB_VALID, NL_CB_CUSTOM, cb_clear_event, &clr);
 348         nl_cb_set(cb, NL_CB_FINISH, NL_CB_CUSTOM, cb_finish_event, &pending);
 349         nl_cb_err(cb, NL_CB_CUSTOM, error_handler, &pending);
 350
 351         nl_send_auto_complete(sock_rtnl, clr.msg);
 352         while (pending > 0)
 353                 nl_recvmsgs(sock_rtnl, cb);
 354
 355         nlmsg_free(clr.msg);
 356 out:
 357         nl_cb_put(cb);
 358 }
 359
 360 /*
 361  * Clear bridge (membership) state and bring down device
 362  */
 363 void system_if_clear_state(struct device *dev)
 364 {
 365         char buf[256];
 366         char *bridge;
 367
 368         dev->ifindex = system_if_resolve(dev);
 369         if (!dev->ifindex)
 370                 return;
 371
 372         system_if_flags(dev->ifname, 0, IFF_UP);
 373
 374         if (system_is_bridge(dev->ifname, buf, sizeof(buf))) {
 375                 D(SYSTEM, "Delete existing bridge named '%s'\n", dev->ifname);
 376                 system_bridge_delbr(dev);
 377                 return;
 378         }
 379
 380         bridge = system_get_bridge(dev->ifname, buf, sizeof(buf));
 381         if (bridge) {
 382                 D(SYSTEM, "Remove device '%s' from bridge '%s'\n", dev->ifname, bridge);
 383                 system_bridge_if(bridge, dev, SIOCBRDELIF, NULL);
 384         }
 385
 386         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET);
 387         system_if_clear_entries(dev, RTM_GETADDR, AF_INET);
 388         system_if_clear_entries(dev, RTM_GETROUTE, AF_INET6);
 389         system_if_clear_entries(dev, RTM_GETADDR, AF_INET6);
 390         system_set_disable_ipv6(dev, "0");
 391 }
 392
 393 static inline unsigned long
 394 sec_to_jiffies(int val)
 395 {
 396         return (unsigned long) val * 100;
 397 }
 398
 399 int system_bridge_addbr(struct device *bridge, struct bridge_config *cfg)
 400 {
 401         unsigned long args[4] = {};
 402
 403         if (ioctl(sock_ioctl, SIOCBRADDBR, bridge->ifname) < 0)
 404                 return -1;
 405
 406         args[0] = BRCTL_SET_BRIDGE_STP_STATE;
 407         args[1] = !!cfg->stp;
 408         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 409
 410         args[0] = BRCTL_SET_BRIDGE_FORWARD_DELAY;
 411         args[1] = sec_to_jiffies(cfg->forward_delay);
 412         system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 413
 414         if (cfg->flags & BRIDGE_OPT_AGEING_TIME) {
 415                 args[0] = BRCTL_SET_AGEING_TIME;
 416                 args[1] = sec_to_jiffies(cfg->ageing_time);
 417                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 418         }
 419
 420         if (cfg->flags & BRIDGE_OPT_HELLO_TIME) {
 421                 args[0] = BRCTL_SET_BRIDGE_HELLO_TIME;
 422                 args[1] = sec_to_jiffies(cfg->hello_time);
 423                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 424         }
 425
 426         if (cfg->flags & BRIDGE_OPT_MAX_AGE) {
 427                 args[0] = BRCTL_SET_BRIDGE_MAX_AGE;
 428                 args[1] = sec_to_jiffies(cfg->max_age);
 429                 system_bridge_if(bridge->ifname, NULL, SIOCDEVPRIVATE, &args);
 430         }
 431
 432         return 0;
 433 }
 434
 435 static int system_vlan(struct device *dev, int id)
 436 {
 437         struct vlan_ioctl_args ifr = {
 438                 .cmd = SET_VLAN_NAME_TYPE_CMD,
 439                 .u.name_type = VLAN_NAME_TYPE_RAW_PLUS_VID_NO_PAD,
 440         };
 441
 442         ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
 443
 444         if (id < 0) {
 445                 ifr.cmd = DEL_VLAN_CMD;
 446                 ifr.u.VID = 0;
 447         } else {
 448                 ifr.cmd = ADD_VLAN_CMD;
 449                 ifr.u.VID = id;
 450         }
 451         strncpy(ifr.device1, dev->ifname, sizeof(ifr.device1));
 452         return ioctl(sock_ioctl, SIOCSIFVLAN, &ifr);
 453 }
 454
 455 int system_vlan_add(struct device *dev, int id)
 456 {
 457         return system_vlan(dev, id);
 458 }
 459
 460 int system_vlan_del(struct device *dev)
 461 {
 462         return system_vlan(dev, -1);
 463 }
 464
 465 int system_if_up(struct device *dev)
 466 {
 467         dev->ifindex = system_if_resolve(dev);
 468         return system_if_flags(dev->ifname, IFF_UP, 0);
 469 }
 470
 471 int system_if_down(struct device *dev)
 472 {
 473         return system_if_flags(dev->ifname, 0, IFF_UP);
 474 }
 475
 476 int system_if_check(struct device *dev)
 477 {
 478         device_set_present(dev, (system_if_resolve(dev) >= 0));
 479         return 0;
 480 }
 481
 482 int system_if_dump_stats(struct device *dev, struct blob_buf *b)
 483 {
 484         const char *const counters[] = {
 485                 "collisions",     "rx_frame_errors",   "tx_compressed",
 486                 "multicast",      "rx_length_errors",  "tx_dropped",
 487                 "rx_bytes",       "rx_missed_errors",  "tx_errors",
 488                 "rx_compressed",  "rx_over_errors",    "tx_fifo_errors",
 489                 "rx_crc_errors",  "rx_packets",        "tx_heartbeat_errors",
 490                 "rx_dropped",     "tx_aborted_errors", "tx_packets",
 491                 "rx_errors",      "tx_bytes",          "tx_window_errors",
 492                 "rx_fifo_errors", "tx_carrier_errors",
 493         };
 494         char buf[64];
 495         int stats_dir;
 496         int i, fd, len;
 497
 498         snprintf(buf, sizeof(buf), "/sys/class/net/%s/statistics", dev->ifname);
 499         stats_dir = open(buf, O_DIRECTORY);
 500         if (stats_dir < 0)
 501                 return -1;
 502
 503         for (i = 0; i < ARRAY_SIZE(counters); i++) {
 504                 fd = openat(stats_dir, counters[i], O_RDONLY);
 505                 if (fd < 0)
 506                         continue;
 507
 508 retry:
 509                 len = read(fd, buf, sizeof(buf));
 510                 if (len < 0) {
 511                         if (errno == EINTR)
 512                                 goto retry;
 513                         continue;
 514                 }
 515
 516                 buf[len] = 0;
 517                 blobmsg_add_u32(b, counters[i], strtoul(buf, NULL, 0));
 518                 close(fd);
 519         }
 520
 521         close(stats_dir);
 522         return 0;
 523 }
 524
 525 static int system_addr(struct device *dev, struct device_addr *addr, int cmd)
 526 {
 527         int alen = ((addr->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
 528         struct ifaddrmsg ifa = {
 529                 .ifa_family = (alen == 4) ? AF_INET : AF_INET6,
 530                 .ifa_prefixlen = addr->mask,
 531                 .ifa_index = dev->ifindex,
 532         };
 533
 534         struct nl_msg *msg = nlmsg_alloc_simple(cmd, 0);
 535         if (!msg)
 536                 return -1;
 537
 538         nlmsg_append(msg, &ifa, sizeof(ifa), 0);
 539         nla_put(msg, IFA_LOCAL, alen, &addr->addr);
 540         return system_rtnl_call(msg);
 541 }
 542
 543 int system_add_address(struct device *dev, struct device_addr *addr)
 544 {
 545         return system_addr(dev, addr, RTM_NEWADDR);
 546 }
 547
 548 int system_del_address(struct device *dev, struct device_addr *addr)
 549 {
 550         return system_addr(dev, addr, RTM_DELADDR);
 551 }
 552
 553 static int system_rt(struct device *dev, struct device_route *route, int cmd)
 554 {
 555         int alen = ((route->flags & DEVADDR_FAMILY) == DEVADDR_INET4) ? 4 : 16;
 556         bool have_gw;
 557         unsigned int flags = 0;
 558
 559         if (alen == 4)
 560                 have_gw = !!route->nexthop.in.s_addr;
 561         else
 562                 have_gw = route->nexthop.in6.s6_addr32[0] ||
 563                         route->nexthop.in6.s6_addr32[1] ||
 564                         route->nexthop.in6.s6_addr32[2] ||
 565                         route->nexthop.in6.s6_addr32[3];
 566
 567         unsigned char scope = (cmd == RTM_DELROUTE) ? RT_SCOPE_NOWHERE :
 568                         (have_gw) ? RT_SCOPE_UNIVERSE : RT_SCOPE_LINK;
 569
 570         struct rtmsg rtm = {
 571                 .rtm_family = (alen == 4) ? AF_INET : AF_INET6,
 572                 .rtm_dst_len = route->mask,
 573                 .rtm_table = RT_TABLE_MAIN,
 574                 .rtm_protocol = RTPROT_BOOT,
 575                 .rtm_scope = scope,
 576                 .rtm_type = (cmd == RTM_DELROUTE) ? 0: RTN_UNICAST,
 577         };
 578
 579         if (cmd == RTM_NEWROUTE)
 580                 flags |= NLM_F_CREATE | NLM_F_REPLACE;
 581
 582         struct nl_msg *msg = nlmsg_alloc_simple(cmd, flags);
 583         if (!msg)
 584                 return -1;
 585
 586         nlmsg_append(msg, &rtm, sizeof(rtm), 0);
 587
 588         if (route->mask)
 589                 nla_put(msg, RTA_DST, alen, &route->addr);
 590
 591         if (have_gw)
 592                 nla_put(msg, RTA_GATEWAY, alen, &route->nexthop);
 593
 594         if (route->flags & DEVADDR_DEVICE)
 595                 nla_put_u32(msg, RTA_OIF, dev->ifindex);
 596
 597         return system_rtnl_call(msg);
 598 }
 599
 600 int system_add_route(struct device *dev, struct device_route *route)
 601 {
 602         return system_rt(dev, route, RTM_NEWROUTE);
 603 }
 604
 605 int system_del_route(struct device *dev, struct device_route *route)
 606 {
 607         return system_rt(dev, route, RTM_DELROUTE);
 608 }
 609
 610 int system_flush_routes(void)
 611 {
 612         const char *names[] = {
 613                 "/proc/sys/net/ipv4/route/flush",
 614                 "/proc/sys/net/ipv6/route/flush"
 615         };
 616         int fd, i;
 617
 618         for (i = 0; i < ARRAY_SIZE(names); i++) {
 619                 fd = open(names[i], O_WRONLY);
 620                 if (fd < 0)
 621                         continue;
 622
 623                 write(fd, "-1", 2);
 624                 close(fd);
 625         }
 626         return 0;
 627 }
 628
 629 time_t system_get_rtime(void)
 630 {
 631         struct timespec ts;
 632         struct timeval tv;
 633
 634         if (syscall(__NR_clock_gettime, CLOCK_MONOTONIC, &ts) == 0)
 635                 return ts.tv_sec;
 636
 637         if (gettimeofday(&tv, NULL) == 0)
 638                 return tv.tv_sec;
 639
 640         return 0;
 641 }