src/resolve/resolved-dns-stream.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   2
   3 #include <netinet/tcp.h>
   4 #include <unistd.h>
   5
   6 #include "alloc-util.h"
   7 #include "fd-util.h"
   8 #include "io-util.h"
   9 #include "missing_network.h"
  10 #include "resolved-dns-stream.h"
  11 #include "resolved-manager.h"
  12
  13 #define DNS_STREAM_TIMEOUT_USEC (10 * USEC_PER_SEC)
  14 #define DNS_STREAMS_MAX 128
  15
  16 #define DNS_QUERIES_PER_STREAM 32
  17
  18 static void dns_stream_stop(DnsStream *s) {
  19         assert(s);
  20
  21         s->io_event_source = sd_event_source_disable_unref(s->io_event_source);
  22         s->timeout_event_source = sd_event_source_disable_unref(s->timeout_event_source);
  23         s->fd = safe_close(s->fd);
  24
  25         /* Disconnect us from the server object if we are now not usable anymore */
  26         dns_stream_detach(s);
  27 }
  28
  29 static int dns_stream_update_io(DnsStream *s) {
  30         int f = 0;
  31
  32         assert(s);
  33
  34         if (s->write_packet && s->n_written < sizeof(s->write_size) + s->write_packet->size)
  35                 f |= EPOLLOUT;
  36         else if (!ordered_set_isempty(s->write_queue)) {
  37                 dns_packet_unref(s->write_packet);
  38                 s->write_packet = ordered_set_steal_first(s->write_queue);
  39                 s->write_size = htobe16(s->write_packet->size);
  40                 s->n_written = 0;
  41                 f |= EPOLLOUT;
  42         }
  43
  44         /* Let's read a packet if we haven't queued any yet. Except if we already hit a limit of parallel
  45          * queries for this connection. */
  46         if ((!s->read_packet || s->n_read < sizeof(s->read_size) + s->read_packet->size) &&
  47                 set_size(s->queries) < DNS_QUERIES_PER_STREAM)
  48                 f |= EPOLLIN;
  49
  50 #if ENABLE_DNS_OVER_TLS
  51         /* For handshake and clean closing purposes, TLS can override requested events */
  52         if (s->dnstls_events != 0)
  53                 f = s->dnstls_events;
  54 #endif
  55
  56         return sd_event_source_set_io_events(s->io_event_source, f);
  57 }
  58
  59 static int dns_stream_complete(DnsStream *s, int error) {
  60         _cleanup_(dns_stream_unrefp) _unused_ DnsStream *ref = dns_stream_ref(s); /* Protect stream while we process it */
  61
  62         assert(s);
  63         assert(error >= 0);
  64
  65         /* Error is > 0 when the connection failed for some reason in the network stack. It's == 0 if we sent
  66          * and received exactly one packet each (in the LLMNR client case). */
  67
  68 #if ENABLE_DNS_OVER_TLS
  69         if (s->encrypted) {
  70                 int r;
  71
  72                 r = dnstls_stream_shutdown(s, error);
  73                 if (r != -EAGAIN)
  74                         dns_stream_stop(s);
  75         } else
  76 #endif
  77                 dns_stream_stop(s);
  78
  79         dns_stream_detach(s);
  80
  81         if (s->complete)
  82                 s->complete(s, error);
  83         else /* the default action if no completion function is set is to close the stream */
  84                 dns_stream_unref(s);
  85
  86         return 0;
  87 }
  88
  89 static int dns_stream_identify(DnsStream *s) {
  90         CMSG_BUFFER_TYPE(CMSG_SPACE(MAXSIZE(struct in_pktinfo, struct in6_pktinfo))
  91                          + CMSG_SPACE(int) + /* for the TTL */
  92                          + EXTRA_CMSG_SPACE /* kernel appears to require extra space */) control;
  93         struct msghdr mh = {};
  94         struct cmsghdr *cmsg;
  95         socklen_t sl;
  96         int r;
  97
  98         assert(s);
  99
 100         if (s->identified)
 101                 return 0;
 102
 103         /* Query the local side */
 104         s->local_salen = sizeof(s->local);
 105         r = getsockname(s->fd, &s->local.sa, &s->local_salen);
 106         if (r < 0)
 107                 return -errno;
 108         if (s->local.sa.sa_family == AF_INET6 && s->ifindex <= 0)
 109                 s->ifindex = s->local.in6.sin6_scope_id;
 110
 111         /* Query the remote side */
 112         s->peer_salen = sizeof(s->peer);
 113         r = getpeername(s->fd, &s->peer.sa, &s->peer_salen);
 114         if (r < 0)
 115                 return -errno;
 116         if (s->peer.sa.sa_family == AF_INET6 && s->ifindex <= 0)
 117                 s->ifindex = s->peer.in6.sin6_scope_id;
 118
 119         /* Check consistency */
 120         assert(s->peer.sa.sa_family == s->local.sa.sa_family);
 121         assert(IN_SET(s->peer.sa.sa_family, AF_INET, AF_INET6));
 122
 123         /* Query connection meta information */
 124         sl = sizeof(control);
 125         if (s->peer.sa.sa_family == AF_INET) {
 126                 r = getsockopt(s->fd, IPPROTO_IP, IP_PKTOPTIONS, &control, &sl);
 127                 if (r < 0)
 128                         return -errno;
 129         } else if (s->peer.sa.sa_family == AF_INET6) {
 130
 131                 r = getsockopt(s->fd, IPPROTO_IPV6, IPV6_2292PKTOPTIONS, &control, &sl);
 132                 if (r < 0)
 133                         return -errno;
 134         } else
 135                 return -EAFNOSUPPORT;
 136
 137         mh.msg_control = &control;
 138         mh.msg_controllen = sl;
 139
 140         CMSG_FOREACH(cmsg, &mh) {
 141
 142                 if (cmsg->cmsg_level == IPPROTO_IPV6) {
 143                         assert(s->peer.sa.sa_family == AF_INET6);
 144
 145                         switch (cmsg->cmsg_type) {
 146
 147                         case IPV6_PKTINFO: {
 148                                 struct in6_pktinfo *i = (struct in6_pktinfo*) CMSG_DATA(cmsg);
 149
 150                                 if (s->ifindex <= 0)
 151                                         s->ifindex = i->ipi6_ifindex;
 152                                 break;
 153                         }
 154
 155                         case IPV6_HOPLIMIT:
 156                                 s->ttl = *(int *) CMSG_DATA(cmsg);
 157                                 break;
 158                         }
 159
 160                 } else if (cmsg->cmsg_level == IPPROTO_IP) {
 161                         assert(s->peer.sa.sa_family == AF_INET);
 162
 163                         switch (cmsg->cmsg_type) {
 164
 165                         case IP_PKTINFO: {
 166                                 struct in_pktinfo *i = (struct in_pktinfo*) CMSG_DATA(cmsg);
 167
 168                                 if (s->ifindex <= 0)
 169                                         s->ifindex = i->ipi_ifindex;
 170                                 break;
 171                         }
 172
 173                         case IP_TTL:
 174                                 s->ttl = *(int *) CMSG_DATA(cmsg);
 175                                 break;
 176                         }
 177                 }
 178         }
 179
 180         /* The Linux kernel sets the interface index to the loopback
 181          * device if the connection came from the local host since it
 182          * avoids the routing table in such a case. Let's unset the
 183          * interface index in such a case. */
 184         if (s->ifindex == LOOPBACK_IFINDEX)
 185                 s->ifindex = 0;
 186
 187         /* If we don't know the interface index still, we look for the
 188          * first local interface with a matching address. Yuck! */
 189         if (s->ifindex <= 0)
 190                 s->ifindex = manager_find_ifindex(s->manager, s->local.sa.sa_family, sockaddr_in_addr(&s->local.sa));
 191
 192         if (s->protocol == DNS_PROTOCOL_LLMNR && s->ifindex > 0) {
 193                 /* Make sure all packets for this connection are sent on the same interface */
 194                 r = socket_set_unicast_if(s->fd, s->local.sa.sa_family, s->ifindex);
 195                 if (r < 0)
 196                         log_debug_errno(errno, "Failed to invoke IP_UNICAST_IF/IPV6_UNICAST_IF: %m");
 197         }
 198
 199         s->identified = true;
 200
 201         return 0;
 202 }
 203
 204 ssize_t dns_stream_writev(DnsStream *s, const struct iovec *iov, size_t iovcnt, int flags) {
 205         ssize_t m;
 206
 207         assert(s);
 208         assert(iov);
 209
 210 #if ENABLE_DNS_OVER_TLS
 211         if (s->encrypted && !(flags & DNS_STREAM_WRITE_TLS_DATA)) {
 212                 ssize_t ss;
 213                 size_t i;
 214
 215                 m = 0;
 216                 for (i = 0; i < iovcnt; i++) {
 217                         ss = dnstls_stream_write(s, iov[i].iov_base, iov[i].iov_len);
 218                         if (ss < 0)
 219                                 return ss;
 220
 221                         m += ss;
 222                         if (ss != (ssize_t) iov[i].iov_len)
 223                                 continue;
 224                 }
 225         } else
 226 #endif
 227         if (s->tfo_salen > 0) {
 228                 struct msghdr hdr = {
 229                         .msg_iov = (struct iovec*) iov,
 230                         .msg_iovlen = iovcnt,
 231                         .msg_name = &s->tfo_address.sa,
 232                         .msg_namelen = s->tfo_salen
 233                 };
 234
 235                 m = sendmsg(s->fd, &hdr, MSG_FASTOPEN);
 236                 if (m < 0) {
 237                         if (errno == EOPNOTSUPP) {
 238                                 s->tfo_salen = 0;
 239                                 if (connect(s->fd, &s->tfo_address.sa, s->tfo_salen) < 0)
 240                                         return -errno;
 241
 242                                 return -EAGAIN;
 243                         }
 244                         if (errno == EINPROGRESS)
 245                                 return -EAGAIN;
 246
 247                         return -errno;
 248                 } else
 249                         s->tfo_salen = 0; /* connection is made */
 250         } else {
 251                 m = writev(s->fd, iov, iovcnt);
 252                 if (m < 0)
 253                         return -errno;
 254         }
 255
 256         return m;
 257 }
 258
 259 static ssize_t dns_stream_read(DnsStream *s, void *buf, size_t count) {
 260         ssize_t ss;
 261
 262 #if ENABLE_DNS_OVER_TLS
 263         if (s->encrypted)
 264                 ss = dnstls_stream_read(s, buf, count);
 265         else
 266 #endif
 267         {
 268                 ss = read(s->fd, buf, count);
 269                 if (ss < 0)
 270                         return -errno;
 271         }
 272
 273         return ss;
 274 }
 275
 276 static int on_stream_timeout(sd_event_source *es, usec_t usec, void *userdata) {
 277         DnsStream *s = userdata;
 278
 279         assert(s);
 280
 281         return dns_stream_complete(s, ETIMEDOUT);
 282 }
 283
 284 static int on_stream_io(sd_event_source *es, int fd, uint32_t revents, void *userdata) {
 285         _cleanup_(dns_stream_unrefp) DnsStream *s = dns_stream_ref(userdata); /* Protect stream while we process it */
 286         bool progressed = false;
 287         int r;
 288
 289         assert(s);
 290
 291 #if ENABLE_DNS_OVER_TLS
 292         if (s->encrypted) {
 293                 r = dnstls_stream_on_io(s, revents);
 294                 if (r == DNSTLS_STREAM_CLOSED)
 295                         return 0;
 296                 if (r == -EAGAIN)
 297                         return dns_stream_update_io(s);
 298                 if (r < 0)
 299                         return dns_stream_complete(s, -r);
 300
 301                 r = dns_stream_update_io(s);
 302                 if (r < 0)
 303                         return r;
 304         }
 305 #endif
 306
 307         /* only identify after connecting */
 308         if (s->tfo_salen == 0) {
 309                 r = dns_stream_identify(s);
 310                 if (r < 0)
 311                         return dns_stream_complete(s, -r);
 312         }
 313
 314         if ((revents & EPOLLOUT) &&
 315             s->write_packet &&
 316             s->n_written < sizeof(s->write_size) + s->write_packet->size) {
 317
 318                 struct iovec iov[2];
 319                 ssize_t ss;
 320
 321                 iov[0] = IOVEC_MAKE(&s->write_size, sizeof(s->write_size));
 322                 iov[1] = IOVEC_MAKE(DNS_PACKET_DATA(s->write_packet), s->write_packet->size);
 323
 324                 IOVEC_INCREMENT(iov, 2, s->n_written);
 325
 326                 ss = dns_stream_writev(s, iov, 2, 0);
 327                 if (ss < 0) {
 328                         if (!IN_SET(-ss, EINTR, EAGAIN))
 329                                 return dns_stream_complete(s, -ss);
 330                 } else {
 331                         progressed = true;
 332                         s->n_written += ss;
 333                 }
 334
 335                 /* Are we done? If so, disable the event source for EPOLLOUT */
 336                 if (s->n_written >= sizeof(s->write_size) + s->write_packet->size) {
 337                         r = dns_stream_update_io(s);
 338                         if (r < 0)
 339                                 return dns_stream_complete(s, -r);
 340                 }
 341         }
 342
 343         if ((revents & (EPOLLIN|EPOLLHUP|EPOLLRDHUP)) &&
 344             (!s->read_packet ||
 345              s->n_read < sizeof(s->read_size) + s->read_packet->size)) {
 346
 347                 if (s->n_read < sizeof(s->read_size)) {
 348                         ssize_t ss;
 349
 350                         ss = dns_stream_read(s, (uint8_t*) &s->read_size + s->n_read, sizeof(s->read_size) - s->n_read);
 351                         if (ss < 0) {
 352                                 if (!IN_SET(-ss, EINTR, EAGAIN))
 353                                         return dns_stream_complete(s, -ss);
 354                         } else if (ss == 0)
 355                                 return dns_stream_complete(s, ECONNRESET);
 356                         else {
 357                                 progressed = true;
 358                                 s->n_read += ss;
 359                         }
 360                 }
 361
 362                 if (s->n_read >= sizeof(s->read_size)) {
 363
 364                         if (be16toh(s->read_size) < DNS_PACKET_HEADER_SIZE)
 365                                 return dns_stream_complete(s, EBADMSG);
 366
 367                         if (s->n_read < sizeof(s->read_size) + be16toh(s->read_size)) {
 368                                 ssize_t ss;
 369
 370                                 if (!s->read_packet) {
 371                                         r = dns_packet_new(&s->read_packet, s->protocol, be16toh(s->read_size), DNS_PACKET_SIZE_MAX);
 372                                         if (r < 0)
 373                                                 return dns_stream_complete(s, -r);
 374
 375                                         s->read_packet->size = be16toh(s->read_size);
 376                                         s->read_packet->ipproto = IPPROTO_TCP;
 377                                         s->read_packet->family = s->peer.sa.sa_family;
 378                                         s->read_packet->ttl = s->ttl;
 379                                         s->read_packet->ifindex = s->ifindex;
 380                                         s->read_packet->timestamp = now(clock_boottime_or_monotonic());
 381
 382                                         if (s->read_packet->family == AF_INET) {
 383                                                 s->read_packet->sender.in = s->peer.in.sin_addr;
 384                                                 s->read_packet->sender_port = be16toh(s->peer.in.sin_port);
 385                                                 s->read_packet->destination.in = s->local.in.sin_addr;
 386                                                 s->read_packet->destination_port = be16toh(s->local.in.sin_port);
 387                                         } else {
 388                                                 assert(s->read_packet->family == AF_INET6);
 389                                                 s->read_packet->sender.in6 = s->peer.in6.sin6_addr;
 390                                                 s->read_packet->sender_port = be16toh(s->peer.in6.sin6_port);
 391                                                 s->read_packet->destination.in6 = s->local.in6.sin6_addr;
 392                                                 s->read_packet->destination_port = be16toh(s->local.in6.sin6_port);
 393
 394                                                 if (s->read_packet->ifindex == 0)
 395                                                         s->read_packet->ifindex = s->peer.in6.sin6_scope_id;
 396                                                 if (s->read_packet->ifindex == 0)
 397                                                         s->read_packet->ifindex = s->local.in6.sin6_scope_id;
 398                                         }
 399                                 }
 400
 401                                 ss = dns_stream_read(s,
 402                                           (uint8_t*) DNS_PACKET_DATA(s->read_packet) + s->n_read - sizeof(s->read_size),
 403                                           sizeof(s->read_size) + be16toh(s->read_size) - s->n_read);
 404                                 if (ss < 0) {
 405                                         if (!IN_SET(-ss, EINTR, EAGAIN))
 406                                                 return dns_stream_complete(s, -ss);
 407                                 } else if (ss == 0)
 408                                         return dns_stream_complete(s, ECONNRESET);
 409                                 else
 410                                         s->n_read += ss;
 411                         }
 412
 413                         /* Are we done? If so, disable the event source for EPOLLIN */
 414                         if (s->n_read >= sizeof(s->read_size) + be16toh(s->read_size)) {
 415                                 /* If there's a packet handler
 416                                  * installed, call that. Note that
 417                                  * this is optional... */
 418                                 if (s->on_packet) {
 419                                         r = s->on_packet(s);
 420                                         if (r < 0)
 421                                                 return r;
 422                                 }
 423
 424                                 r = dns_stream_update_io(s);
 425                                 if (r < 0)
 426                                         return dns_stream_complete(s, -r);
 427                         }
 428                 }
 429         }
 430
 431         /* Call "complete" callback if finished reading and writing one packet, and there's nothing else left
 432          * to write. */
 433         if (s->type == DNS_STREAM_LLMNR_SEND &&
 434             (s->write_packet && s->n_written >= sizeof(s->write_size) + s->write_packet->size) &&
 435             ordered_set_isempty(s->write_queue) &&
 436             (s->read_packet && s->n_read >= sizeof(s->read_size) + s->read_packet->size))
 437                 return dns_stream_complete(s, 0);
 438
 439         /* If we did something, let's restart the timeout event source */
 440         if (progressed && s->timeout_event_source) {
 441                 r = sd_event_source_set_time_relative(s->timeout_event_source, DNS_STREAM_TIMEOUT_USEC);
 442                 if (r < 0)
 443                         log_warning_errno(errno, "Couldn't restart TCP connection timeout, ignoring: %m");
 444         }
 445
 446         return 0;
 447 }
 448
 449 static DnsStream *dns_stream_free(DnsStream *s) {
 450         DnsPacket *p;
 451
 452         assert(s);
 453
 454         dns_stream_stop(s);
 455
 456         if (s->manager) {
 457                 LIST_REMOVE(streams, s->manager->dns_streams, s);
 458                 s->manager->n_dns_streams[s->type]--;
 459         }
 460
 461 #if ENABLE_DNS_OVER_TLS
 462         if (s->encrypted)
 463                 dnstls_stream_free(s);
 464 #endif
 465
 466         ORDERED_SET_FOREACH(p, s->write_queue)
 467                 dns_packet_unref(ordered_set_remove(s->write_queue, p));
 468
 469         dns_packet_unref(s->write_packet);
 470         dns_packet_unref(s->read_packet);
 471         dns_server_unref(s->server);
 472
 473         ordered_set_free(s->write_queue);
 474
 475         return mfree(s);
 476 }
 477
 478 DEFINE_TRIVIAL_REF_UNREF_FUNC(DnsStream, dns_stream, dns_stream_free);
 479
 480 int dns_stream_new(
 481                 Manager *m,
 482                 DnsStream **ret,
 483                 DnsStreamType type,
 484                 DnsProtocol protocol,
 485                 int fd,
 486                 const union sockaddr_union *tfo_address) {
 487
 488         _cleanup_(dns_stream_unrefp) DnsStream *s = NULL;
 489         int r;
 490
 491         assert(m);
 492         assert(ret);
 493         assert(type >= 0);
 494         assert(type < _DNS_STREAM_TYPE_MAX);
 495         assert(protocol >= 0);
 496         assert(protocol < _DNS_PROTOCOL_MAX);
 497         assert(fd >= 0);
 498
 499         if (m->n_dns_streams[type] > DNS_STREAMS_MAX)
 500                 return -EBUSY;
 501
 502         s = new(DnsStream, 1);
 503         if (!s)
 504                 return -ENOMEM;
 505
 506         *s = (DnsStream) {
 507                 .n_ref = 1,
 508                 .fd = -1,
 509                 .protocol = protocol,
 510                 .type = type,
 511         };
 512
 513         r = ordered_set_ensure_allocated(&s->write_queue, &dns_packet_hash_ops);
 514         if (r < 0)
 515                 return r;
 516
 517         r = sd_event_add_io(m->event, &s->io_event_source, fd, EPOLLIN, on_stream_io, s);
 518         if (r < 0)
 519                 return r;
 520
 521         (void) sd_event_source_set_description(s->io_event_source, "dns-stream-io");
 522
 523         r = sd_event_add_time_relative(
 524                         m->event,
 525                         &s->timeout_event_source,
 526                         clock_boottime_or_monotonic(),
 527                         DNS_STREAM_TIMEOUT_USEC, 0,
 528                         on_stream_timeout, s);
 529         if (r < 0)
 530                 return r;
 531
 532         (void) sd_event_source_set_description(s->timeout_event_source, "dns-stream-timeout");
 533
 534         LIST_PREPEND(streams, m->dns_streams, s);
 535         m->n_dns_streams[type]++;
 536         s->manager = m;
 537
 538         s->fd = fd;
 539
 540         if (tfo_address) {
 541                 s->tfo_address = *tfo_address;
 542                 s->tfo_salen = tfo_address->sa.sa_family == AF_INET6 ? sizeof(tfo_address->in6) : sizeof(tfo_address->in);
 543         }
 544
 545         *ret = TAKE_PTR(s);
 546
 547         return 0;
 548 }
 549
 550 int dns_stream_write_packet(DnsStream *s, DnsPacket *p) {
 551         int r;
 552
 553         assert(s);
 554         assert(p);
 555
 556         r = ordered_set_put(s->write_queue, p);
 557         if (r < 0)
 558                 return r;
 559
 560         dns_packet_ref(p);
 561
 562         return dns_stream_update_io(s);
 563 }
 564
 565 DnsPacket *dns_stream_take_read_packet(DnsStream *s) {
 566         assert(s);
 567
 568         if (!s->read_packet)
 569                 return NULL;
 570
 571         if (s->n_read < sizeof(s->read_size))
 572                 return NULL;
 573
 574         if (s->n_read < sizeof(s->read_size) + be16toh(s->read_size))
 575                 return NULL;
 576
 577         s->n_read = 0;
 578         return TAKE_PTR(s->read_packet);
 579 }
 580
 581 void dns_stream_detach(DnsStream *s) {
 582         assert(s);
 583
 584         if (!s->server)
 585                 return;
 586
 587         if (s->server->stream != s)
 588                 return;
 589
 590         dns_server_unref_stream(s->server);
 591 }