src/dnsproxy.c

   1 /*
   2  *
   3  *  Connection Manager
   4  *
   5  *  Copyright (C) 2007-2012  Intel Corporation. All rights reserved.
   6  *
   7  *  This program is free software; you can redistribute it and/or modify
   8  *  it under the terms of the GNU General Public License version 2 as
   9  *  published by the Free Software Foundation.
  10  *
  11  *  This program is distributed in the hope that it will be useful,
  12  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14  *  GNU General Public License for more details.
  15  *
  16  *  You should have received a copy of the GNU General Public License
  17  *  along with this program; if not, write to the Free Software
  18  *  Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
  19  *
  20  */
  21
  22 #ifdef HAVE_CONFIG_H
  23 #include <config.h>
  24 #endif
  25
  26 #include <errno.h>
  27 #include <unistd.h>
  28 #include <string.h>
  29 #include <stdint.h>
  30 #include <arpa/inet.h>
  31 #include <netinet/in.h>
  32 #include <sys/types.h>
  33 #include <sys/socket.h>
  34 #include <netdb.h>
  35 #include <resolv.h>
  36 #include <gweb/gresolv.h>
  37
  38 #include <glib.h>
  39
  40 #include "connman.h"
  41
  42 #if __BYTE_ORDER == __LITTLE_ENDIAN
  43 struct domain_hdr {
  44         uint16_t id;
  45         uint8_t rd:1;
  46         uint8_t tc:1;
  47         uint8_t aa:1;
  48         uint8_t opcode:4;
  49         uint8_t qr:1;
  50         uint8_t rcode:4;
  51         uint8_t z:3;
  52         uint8_t ra:1;
  53         uint16_t qdcount;
  54         uint16_t ancount;
  55         uint16_t nscount;
  56         uint16_t arcount;
  57 } __attribute__ ((packed));
  58 #elif __BYTE_ORDER == __BIG_ENDIAN
  59 struct domain_hdr {
  60         uint16_t id;
  61         uint8_t qr:1;
  62         uint8_t opcode:4;
  63         uint8_t aa:1;
  64         uint8_t tc:1;
  65         uint8_t rd:1;
  66         uint8_t ra:1;
  67         uint8_t z:3;
  68         uint8_t rcode:4;
  69         uint16_t qdcount;
  70         uint16_t ancount;
  71         uint16_t nscount;
  72         uint16_t arcount;
  73 } __attribute__ ((packed));
  74 #else
  75 #error "Unknown byte order"
  76 #endif
  77
  78 struct partial_reply {
  79         uint16_t len;
  80         uint16_t received;
  81         unsigned char buf[];
  82 };
  83
  84 struct server_data {
  85         char *interface;
  86         GList *domains;
  87         char *server;
  88         int protocol;
  89         GIOChannel *channel;
  90         guint watch;
  91         guint timeout;
  92         gboolean enabled;
  93         gboolean connected;
  94         struct partial_reply *incoming_reply;
  95 };
  96
  97 struct request_data {
  98         union {
  99                 struct sockaddr_in6 __sin6; /* Only for the length */
 100                 struct sockaddr sa;
 101         };
 102         socklen_t sa_len;
 103         int client_sk;
 104         int protocol;
 105         guint16 srcid;
 106         guint16 dstid;
 107         guint16 altid;
 108         guint timeout;
 109         guint watch;
 110         guint numserv;
 111         guint numresp;
 112         gpointer request;
 113         gsize request_len;
 114         gpointer name;
 115         gpointer resp;
 116         gsize resplen;
 117         struct listener_data *ifdata;
 118         gboolean append_domain;
 119 };
 120
 121 struct listener_data {
 122         char *ifname;
 123         GIOChannel *udp_listener_channel;
 124         guint udp_listener_watch;
 125         GIOChannel *tcp_listener_channel;
 126         guint tcp_listener_watch;
 127 };
 128
 129 struct cache_data {
 130         time_t inserted;
 131         time_t valid_until;
 132         time_t cache_until;
 133         int timeout;
 134         uint16_t type;
 135         uint16_t answers;
 136         unsigned int data_len;
 137         unsigned char *data; /* contains DNS header + body */
 138 };
 139
 140 struct cache_entry {
 141         char *key;
 142         int want_refresh;
 143         int hits;
 144         struct cache_data *ipv4;
 145         struct cache_data *ipv6;
 146 };
 147
 148 struct domain_question {
 149         uint16_t type;
 150         uint16_t class;
 151 } __attribute__ ((packed));
 152
 153 struct domain_rr {
 154         uint16_t type;
 155         uint16_t class;
 156         uint32_t ttl;
 157         uint16_t rdlen;
 158 } __attribute__ ((packed));
 159
 160 /*
 161  * We limit how long the cached DNS entry stays in the cache.
 162  * By default the TTL (time-to-live) of the DNS response is used
 163  * when setting the cache entry life time. The value is in seconds.
 164  */
 165 #define MAX_CACHE_TTL (60 * 30)
 166 /*
 167  * Also limit the other end, cache at least for 30 seconds.
 168  */
 169 #define MIN_CACHE_TTL (30)
 170
 171 /*
 172  * We limit the cache size to some sane value so that cached data does
 173  * not occupy too much memory. Each cached entry occupies on average
 174  * about 100 bytes memory (depending on DNS name length).
 175  * Example: caching www.connman.net uses 97 bytes memory.
 176  * The value is the max amount of cached DNS responses (count).
 177  */
 178 #define MAX_CACHE_SIZE 256
 179
 180 static int cache_size;
 181 static GHashTable *cache;
 182 static int cache_refcount;
 183 static GSList *server_list = NULL;
 184 static GSList *request_list = NULL;
 185 static GSList *request_pending_list = NULL;
 186 static guint16 request_id = 0x0000;
 187 static GHashTable *listener_table = NULL;
 188 static time_t next_refresh;
 189
 190 static int protocol_offset(int protocol)
 191 {
 192         switch (protocol) {
 193         case IPPROTO_UDP:
 194                 return 0;
 195
 196         case IPPROTO_TCP:
 197                 return 2;
 198
 199         default:
 200                 return -EINVAL;
 201         }
 202
 203 }
 204
 205 /*
 206  * There is a power and efficiency benefit to have entries
 207  * in our cache expire at the same time. To this extend,
 208  * we round down the cache valid time to common boundaries.
 209  */
 210 static time_t round_down_ttl(time_t end_time, int ttl)
 211 {
 212         if (ttl < 15)
 213                 return end_time;
 214
 215         /* Less than 5 minutes, round to 10 second boundary */
 216         if (ttl < 300) {
 217                 end_time = end_time / 10;
 218                 end_time = end_time * 10;
 219         } else { /* 5 or more minutes, round to 30 seconds */
 220                 end_time = end_time / 30;
 221                 end_time = end_time * 30;
 222         }
 223         return end_time;
 224 }
 225
 226 static struct request_data *find_request(guint16 id)
 227 {
 228         GSList *list;
 229
 230         for (list = request_list; list; list = list->next) {
 231                 struct request_data *req = list->data;
 232
 233                 if (req->dstid == id || req->altid == id)
 234                         return req;
 235         }
 236
 237         return NULL;
 238 }
 239
 240 static struct server_data *find_server(const char *interface,
 241                                         const char *server,
 242                                                 int protocol)
 243 {
 244         GSList *list;
 245
 246         DBG("interface %s server %s", interface, server);
 247
 248         for (list = server_list; list; list = list->next) {
 249                 struct server_data *data = list->data;
 250
 251                 if (interface == NULL && data->interface == NULL &&
 252                                 g_str_equal(data->server, server) == TRUE &&
 253                                 data->protocol == protocol)
 254                         return data;
 255
 256                 if (interface == NULL ||
 257                                 data->interface == NULL || data->server == NULL)
 258                         continue;
 259
 260                 if (g_str_equal(data->interface, interface) == TRUE &&
 261                                 g_str_equal(data->server, server) == TRUE &&
 262                                 data->protocol == protocol)
 263                         return data;
 264         }
 265
 266         return NULL;
 267 }
 268
 269 /* we can keep using the same resolve's */
 270 static GResolv *ipv4_resolve;
 271 static GResolv *ipv6_resolve;
 272
 273 static void dummy_resolve_func(GResolvResultStatus status,
 274                                         char **results, gpointer user_data)
 275 {
 276 }
 277
 278 /*
 279  * Refresh a DNS entry, but also age the hit count a bit */
 280 static void refresh_dns_entry(struct cache_entry *entry, char *name)
 281 {
 282         int age = 1;
 283
 284         if (ipv4_resolve == NULL) {
 285                 ipv4_resolve = g_resolv_new(0);
 286                 g_resolv_set_address_family(ipv4_resolve, AF_INET);
 287                 g_resolv_add_nameserver(ipv4_resolve, "127.0.0.1", 53, 0);
 288         }
 289
 290         if (ipv6_resolve == NULL) {
 291                 ipv6_resolve = g_resolv_new(0);
 292                 g_resolv_set_address_family(ipv6_resolve, AF_INET6);
 293                 g_resolv_add_nameserver(ipv6_resolve, "127.0.0.1", 53, 0);
 294         }
 295
 296         if (entry->ipv4 == NULL) {
 297                 DBG("Refresing A record for %s", name);
 298                 g_resolv_lookup_hostname(ipv4_resolve, name,
 299                                         dummy_resolve_func, NULL);
 300                 age = 4;
 301         }
 302
 303         if (entry->ipv6 == NULL) {
 304                 DBG("Refresing AAAA record for %s", name);
 305                 g_resolv_lookup_hostname(ipv6_resolve, name,
 306                                         dummy_resolve_func, NULL);
 307                 age = 4;
 308         }
 309
 310         entry->hits -= age;
 311         if (entry->hits < 0)
 312                 entry->hits = 0;
 313 }
 314
 315 static int dns_name_length(unsigned char *buf)
 316 {
 317         if ((buf[0] & NS_CMPRSFLGS) == NS_CMPRSFLGS) /* compressed name */
 318                 return 2;
 319         return strlen((char *)buf);
 320 }
 321
 322 static void update_cached_ttl(unsigned char *buf, int len, int new_ttl)
 323 {
 324         unsigned char *c;
 325         uint32_t *i;
 326         uint16_t *w;
 327         int l;
 328
 329         /* skip the header */
 330         c = buf + 12;
 331         len -= 12;
 332
 333         /* skip the query, which is a name and 2 16 bit words */
 334         l = dns_name_length(c);
 335         c += l;
 336         len -= l;
 337         c += 4;
 338         len -= 4;
 339
 340         /* now we get the answer records */
 341
 342         while (len > 0) {
 343                 /* first a name */
 344                 l = dns_name_length(c);
 345                 c += l;
 346                 len -= l;
 347                 if (len < 0)
 348                         break;
 349                 /* then type + class, 2 bytes each */
 350                 c += 4;
 351                 len -= 4;
 352                 if (len < 0)
 353                         break;
 354
 355                 /* now the 4 byte TTL field */
 356                 i = (uint32_t *)c;
 357                 *i = htonl(new_ttl);
 358                 c += 4;
 359                 len -= 4;
 360                 if (len < 0)
 361                         break;
 362
 363                 /* now the 2 byte rdlen field */
 364                 w = (uint16_t *)c;
 365                 c += ntohs(*w) + 2;
 366                 len -= ntohs(*w) + 2;
 367         }
 368 }
 369
 370 static void send_cached_response(int sk, unsigned char *buf, int len,
 371                                 const struct sockaddr *to, socklen_t tolen,
 372                                 int protocol, int id, uint16_t answers, int ttl)
 373 {
 374         struct domain_hdr *hdr;
 375         int err, offset = protocol_offset(protocol);
 376
 377         if (offset < 0)
 378                 return;
 379
 380         if (len < 12)
 381                 return;
 382
 383         hdr = (void *) (buf + offset);
 384
 385         hdr->id = id;
 386         hdr->qr = 1;
 387         hdr->rcode = 0;
 388         hdr->ancount = htons(answers);
 389         hdr->nscount = 0;
 390         hdr->arcount = 0;
 391
 392         /* if this is a negative reply, we are authorative */
 393         if (answers == 0)
 394                 hdr->aa = 1;
 395         else
 396                 update_cached_ttl(buf, len, ttl);
 397
 398         DBG("id 0x%04x answers %d", hdr->id, answers);
 399
 400         err = sendto(sk, buf, len, 0, to, tolen);
 401         if (err < 0) {
 402                 connman_error("Cannot send cached DNS response: %s",
 403                                 strerror(errno));
 404                 return;
 405         }
 406 }
 407
 408 static void send_response(int sk, unsigned char *buf, int len,
 409                                 const struct sockaddr *to, socklen_t tolen,
 410                                 int protocol)
 411 {
 412         struct domain_hdr *hdr;
 413         int err, offset = protocol_offset(protocol);
 414
 415         DBG("");
 416
 417         if (offset < 0)
 418                 return;
 419
 420         if (len < 12)
 421                 return;
 422
 423         hdr = (void *) (buf + offset);
 424
 425         DBG("id 0x%04x qr %d opcode %d", hdr->id, hdr->qr, hdr->opcode);
 426
 427         hdr->qr = 1;
 428         hdr->rcode = 2;
 429
 430         hdr->ancount = 0;
 431         hdr->nscount = 0;
 432         hdr->arcount = 0;
 433
 434         err = sendto(sk, buf, len, 0, to, tolen);
 435         if (err < 0) {
 436                 connman_error("Failed to send DNS response: %s",
 437                                 strerror(errno));
 438                 return;
 439         }
 440 }
 441
 442 static gboolean request_timeout(gpointer user_data)
 443 {
 444         struct request_data *req = user_data;
 445         struct listener_data *ifdata;
 446
 447         DBG("id 0x%04x", req->srcid);
 448
 449         if (req == NULL)
 450                 return FALSE;
 451
 452         ifdata = req->ifdata;
 453
 454         request_list = g_slist_remove(request_list, req);
 455         req->numserv--;
 456
 457         if (req->resplen > 0 && req->resp != NULL) {
 458                 int sk, err;
 459
 460                 sk = g_io_channel_unix_get_fd(ifdata->udp_listener_channel);
 461
 462                 err = sendto(sk, req->resp, req->resplen, 0,
 463                                                 &req->sa, req->sa_len);
 464                 if (err < 0)
 465                         return FALSE;
 466         } else if (req->request && req->numserv == 0) {
 467                 struct domain_hdr *hdr;
 468
 469                 if (req->protocol == IPPROTO_TCP) {
 470                         hdr = (void *) (req->request + 2);
 471                         hdr->id = req->srcid;
 472                         send_response(req->client_sk, req->request,
 473                                 req->request_len, NULL, 0, IPPROTO_TCP);
 474
 475                 } else if (req->protocol == IPPROTO_UDP) {
 476                         int sk;
 477
 478                         hdr = (void *) (req->request);
 479                         hdr->id = req->srcid;
 480                         sk = g_io_channel_unix_get_fd(
 481                                                 ifdata->udp_listener_channel);
 482                         send_response(sk, req->request, req->request_len,
 483                                         &req->sa, req->sa_len, IPPROTO_UDP);
 484                 }
 485         }
 486
 487         g_free(req->resp);
 488         g_free(req);
 489
 490         return FALSE;
 491 }
 492
 493 static int append_query(unsigned char *buf, unsigned int size,
 494                                 const char *query, const char *domain)
 495 {
 496         unsigned char *ptr = buf;
 497         char *offset;
 498         int len;
 499
 500         DBG("query %s domain %s", query, domain);
 501
 502         offset = (char *) query;
 503         while (offset != NULL) {
 504                 char *tmp;
 505
 506                 tmp = strchr(offset, '.');
 507                 if (tmp == NULL) {
 508                         len = strlen(offset);
 509                         if (len == 0)
 510                                 break;
 511                         *ptr = len;
 512                         memcpy(ptr + 1, offset, len);
 513                         ptr += len + 1;
 514                         break;
 515                 }
 516
 517                 *ptr = tmp - offset;
 518                 memcpy(ptr + 1, offset, tmp - offset);
 519                 ptr += tmp - offset + 1;
 520
 521                 offset = tmp + 1;
 522         }
 523
 524         offset = (char *) domain;
 525         while (offset != NULL) {
 526                 char *tmp;
 527
 528                 tmp = strchr(offset, '.');
 529                 if (tmp == NULL) {
 530                         len = strlen(offset);
 531                         if (len == 0)
 532                                 break;
 533                         *ptr = len;
 534                         memcpy(ptr + 1, offset, len);
 535                         ptr += len + 1;
 536                         break;
 537                 }
 538
 539                 *ptr = tmp - offset;
 540                 memcpy(ptr + 1, offset, tmp - offset);
 541                 ptr += tmp - offset + 1;
 542
 543                 offset = tmp + 1;
 544         }
 545
 546         *ptr++ = 0x00;
 547
 548         return ptr - buf;
 549 }
 550
 551 static gboolean cache_check_is_valid(struct cache_data *data,
 552                                 time_t current_time)
 553 {
 554         if (data == NULL)
 555                 return FALSE;
 556
 557         if (data->cache_until < current_time)
 558                 return FALSE;
 559
 560         return TRUE;
 561 }
 562
 563 /*
 564  * remove stale cached entries so that they can be refreshed
 565  */
 566 static void cache_enforce_validity(struct cache_entry *entry)
 567 {
 568         time_t current_time = time(NULL);
 569
 570         if (cache_check_is_valid(entry->ipv4, current_time) == FALSE
 571                                                         && entry->ipv4) {
 572                 DBG("cache timeout \"%s\" type A", entry->key);
 573                 g_free(entry->ipv4->data);
 574                 g_free(entry->ipv4);
 575                 entry->ipv4 = NULL;
 576
 577         }
 578
 579         if (cache_check_is_valid(entry->ipv6, current_time) == FALSE
 580                                                         && entry->ipv6) {
 581                 DBG("cache timeout \"%s\" type AAAA", entry->key);
 582                 g_free(entry->ipv6->data);
 583                 g_free(entry->ipv6);
 584                 entry->ipv6 = NULL;
 585         }
 586 }
 587
 588 static uint16_t cache_check_validity(char *question, uint16_t type,
 589                                 struct cache_entry *entry)
 590 {
 591         time_t current_time = time(NULL);
 592         int want_refresh = 0;
 593
 594         /*
 595          * if we have a popular entry, we want a refresh instead of
 596          * total destruction of the entry.
 597          */
 598         if (entry->hits > 2)
 599                 want_refresh = 1;
 600
 601         cache_enforce_validity(entry);
 602
 603         switch (type) {
 604         case 1:         /* IPv4 */
 605                 if (cache_check_is_valid(entry->ipv4, current_time) == FALSE) {
 606                         DBG("cache %s \"%s\" type A", entry->ipv4 ?
 607                                         "timeout" : "entry missing", question);
 608
 609                         if (want_refresh)
 610                                 entry->want_refresh = 1;
 611
 612                         /*
 613                          * We do not remove cache entry if there is still
 614                          * valid IPv6 entry found in the cache.
 615                          */
 616                         if (cache_check_is_valid(entry->ipv6, current_time)
 617                                         == FALSE && want_refresh == FALSE) {
 618                                 g_hash_table_remove(cache, question);
 619                                 type = 0;
 620                         }
 621                 }
 622                 break;
 623
 624         case 28:        /* IPv6 */
 625                 if (cache_check_is_valid(entry->ipv6, current_time) == FALSE) {
 626                         DBG("cache %s \"%s\" type AAAA", entry->ipv6 ?
 627                                         "timeout" : "entry missing", question);
 628
 629                         if (want_refresh)
 630                                 entry->want_refresh = 1;
 631
 632                         if (cache_check_is_valid(entry->ipv4, current_time)
 633                                         == FALSE && want_refresh == FALSE) {
 634                                 g_hash_table_remove(cache, question);
 635                                 type = 0;
 636                         }
 637                 }
 638                 break;
 639         }
 640
 641         return type;
 642 }
 643
 644 static struct cache_entry *cache_check(gpointer request, int *qtype)
 645 {
 646         char *question = request + 12;
 647         struct cache_entry *entry;
 648         struct domain_question *q;
 649         uint16_t type;
 650         int offset;
 651
 652         offset = strlen(question) + 1;
 653         q = (void *) (question + offset);
 654         type = ntohs(q->type);
 655
 656         /* We only cache either A (1) or AAAA (28) requests */
 657         if (type != 1 && type != 28)
 658                 return NULL;
 659
 660         entry = g_hash_table_lookup(cache, question);
 661         if (entry == NULL)
 662                 return NULL;
 663
 664         type = cache_check_validity(question, type, entry);
 665         if (type == 0)
 666                 return NULL;
 667
 668         *qtype = type;
 669         return entry;
 670 }
 671
 672 /*
 673  * Get a label/name from DNS resource record. The function decompresses the
 674  * label if necessary. The function does not convert the name to presentation
 675  * form. This means that the result string will contain label lengths instead
 676  * of dots between labels. We intentionally do not want to convert to dotted
 677  * format so that we can cache the wire format string directly.
 678  */
 679 static int get_name(int counter,
 680                 unsigned char *pkt, unsigned char *start, unsigned char *max,
 681                 unsigned char *output, int output_max, int *output_len,
 682                 unsigned char **end, char *name, int *name_len)
 683 {
 684         unsigned char *p;
 685
 686         /* Limit recursion to 10 (this means up to 10 labels in domain name) */
 687         if (counter > 10)
 688                 return -EINVAL;
 689
 690         p = start;
 691         while (*p) {
 692                 if ((*p & NS_CMPRSFLGS) == NS_CMPRSFLGS) {
 693                         uint16_t offset = (*p & 0x3F) * 256 + *(p + 1);
 694
 695                         if (offset >= max - pkt)
 696                                 return -ENOBUFS;
 697
 698                         if (*end == NULL)
 699                                 *end = p + 2;
 700
 701                         return get_name(counter + 1, pkt, pkt + offset, max,
 702                                         output, output_max, output_len, end,
 703                                         name, name_len);
 704                 } else {
 705                         unsigned label_len = *p;
 706
 707                         if (pkt + label_len > max)
 708                                 return -ENOBUFS;
 709
 710                         if (*output_len > output_max)
 711                                 return -ENOBUFS;
 712
 713                         /*
 714                          * We need the original name in order to check
 715                          * if this answer is the correct one.
 716                          */
 717                         name[(*name_len)++] = label_len;
 718                         memcpy(name + *name_len, p + 1, label_len + 1);
 719                         *name_len += label_len;
 720
 721                         /* We compress the result */
 722                         output[0] = NS_CMPRSFLGS;
 723                         output[1] = 0x0C;
 724                         *output_len = 2;
 725
 726                         p += label_len + 1;
 727
 728                         if (*end == NULL)
 729                                 *end = p;
 730
 731                         if (p >= max)
 732                                 return -ENOBUFS;
 733                 }
 734         }
 735
 736         return 0;
 737 }
 738
 739 static int parse_rr(unsigned char *buf, unsigned char *start,
 740                         unsigned char *max,
 741                         unsigned char *response, unsigned int *response_size,
 742                         uint16_t *type, uint16_t *class, int *ttl, int *rdlen,
 743                         unsigned char **end,
 744                         char *name)
 745 {
 746         struct domain_rr *rr;
 747         int err, offset;
 748         int name_len = 0, output_len = 0, max_rsp = *response_size;
 749
 750         err = get_name(0, buf, start, max, response, max_rsp,
 751                 &output_len, end, name, &name_len);
 752         if (err < 0)
 753                 return err;
 754
 755         offset = output_len;
 756
 757         if ((unsigned int) offset > *response_size)
 758                 return -ENOBUFS;
 759
 760         rr = (void *) (*end);
 761
 762         if (rr == NULL)
 763                 return -EINVAL;
 764
 765         *type = ntohs(rr->type);
 766         *class = ntohs(rr->class);
 767         *ttl = ntohl(rr->ttl);
 768         *rdlen = ntohs(rr->rdlen);
 769
 770         if (*ttl < 0)
 771                 return -EINVAL;
 772
 773         memcpy(response + offset, *end, sizeof(struct domain_rr));
 774
 775         offset += sizeof(struct domain_rr);
 776         *end += sizeof(struct domain_rr);
 777
 778         if ((unsigned int) (offset + *rdlen) > *response_size)
 779                 return -ENOBUFS;
 780
 781         memcpy(response + offset, *end, *rdlen);
 782
 783         *end += *rdlen;
 784
 785         *response_size = offset + *rdlen;
 786
 787         return 0;
 788 }
 789
 790 static gboolean check_alias(GSList *aliases, char *name)
 791 {
 792         GSList *list;
 793
 794         if (aliases != NULL) {
 795                 for (list = aliases; list; list = list->next) {
 796                         int len = strlen((char *)list->data);
 797                         if (strncmp((char *)list->data, name, len) == 0)
 798                                 return TRUE;
 799                 }
 800         }
 801
 802         return FALSE;
 803 }
 804
 805 static int parse_response(unsigned char *buf, int buflen,
 806                         char *question, int qlen,
 807                         uint16_t *type, uint16_t *class, int *ttl,
 808                         unsigned char *response, unsigned int *response_len,
 809                         uint16_t *answers)
 810 {
 811         struct domain_hdr *hdr = (void *) buf;
 812         struct domain_question *q;
 813         unsigned char *ptr;
 814         uint16_t qdcount = ntohs(hdr->qdcount);
 815         uint16_t ancount = ntohs(hdr->ancount);
 816         int err, i;
 817         uint16_t qtype, qclass;
 818         unsigned char *next = NULL;
 819         unsigned int maxlen = *response_len;
 820         GSList *aliases = NULL, *list;
 821         char name[NS_MAXDNAME + 1];
 822
 823         if (buflen < 12)
 824                 return -EINVAL;
 825
 826         DBG("qr %d qdcount %d", hdr->qr, qdcount);
 827
 828         /* We currently only cache responses where question count is 1 */
 829         if (hdr->qr != 1 || qdcount != 1)
 830                 return -EINVAL;
 831
 832         ptr = buf + sizeof(struct domain_hdr);
 833
 834         strncpy(question, (char *) ptr, qlen);
 835         qlen = strlen(question);
 836         ptr += qlen + 1; /* skip \0 */
 837
 838         q = (void *) ptr;
 839         qtype = ntohs(q->type);
 840
 841         /* We cache only A and AAAA records */
 842         if (qtype != 1 && qtype != 28)
 843                 return -ENOMSG;
 844
 845         qclass = ntohs(q->class);
 846
 847         ptr += 2 + 2; /* ptr points now to answers */
 848
 849         err = -ENOMSG;
 850         *response_len = 0;
 851         *answers = 0;
 852
 853         /*
 854          * We have a bunch of answers (like A, AAAA, CNAME etc) to
 855          * A or AAAA question. We traverse the answers and parse the
 856          * resource records. Only A and AAAA records are cached, all
 857          * the other records in answers are skipped.
 858          */
 859         for (i = 0; i < ancount; i++) {
 860                 /*
 861                  * Get one address at a time to this buffer.
 862                  * The max size of the answer is
 863                  *   2 (pointer) + 2 (type) + 2 (class) +
 864                  *   4 (ttl) + 2 (rdlen) + addr (16 or 4) = 28
 865                  * for A or AAAA record.
 866                  * For CNAME the size can be bigger.
 867                  */
 868                 unsigned char rsp[NS_MAXCDNAME];
 869                 unsigned int rsp_len = sizeof(rsp) - 1;
 870                 int ret, rdlen;
 871
 872                 memset(rsp, 0, sizeof(rsp));
 873
 874                 ret = parse_rr(buf, ptr, buf + buflen, rsp, &rsp_len,
 875                         type, class, ttl, &rdlen, &next, name);
 876                 if (ret != 0) {
 877                         err = ret;
 878                         goto out;
 879                 }
 880
 881                 /*
 882                  * Now rsp contains compressed or uncompressed resource
 883                  * record. Next we check if this record answers the question.
 884                  * The name var contains the uncompressed label.
 885                  * One tricky bit is the CNAME records as they alias
 886                  * the name we might be interested in.
 887                  */
 888
 889                 /*
 890                  * Go to next answer if the class is not the one we are
 891                  * looking for.
 892                  */
 893                 if (*class != qclass) {
 894                         ptr = next;
 895                         next = NULL;
 896                         continue;
 897                 }
 898
 899                 /*
 900                  * Try to resolve aliases also, type is CNAME(5).
 901                  * This is important as otherwise the aliased names would not
 902                  * be cached at all as the cache would not contain the aliased
 903                  * question.
 904                  *
 905                  * If any CNAME is found in DNS packet, then we cache the alias
 906                  * IP address instead of the question (as the server
 907                  * said that question has only an alias).
 908                  * This means in practice that if e.g., ipv6.google.com is
 909                  * queried, DNS server returns CNAME of that name which is
 910                  * ipv6.l.google.com. We then cache the address of the CNAME
 911                  * but return the question name to client. So the alias
 912                  * status of the name is not saved in cache and thus not
 913                  * returned to the client. We do not return DNS packets from
 914                  * cache to client saying that ipv6.google.com is an alias to
 915                  * ipv6.l.google.com but we return instead a DNS packet that
 916                  * says ipv6.google.com has address xxx which is in fact the
 917                  * address of ipv6.l.google.com. For caching purposes this
 918                  * should not cause any issues.
 919                  */
 920                 if (*type == 5 && strncmp(question, name, qlen) == 0) {
 921                         /*
 922                          * So now the alias answered the question. This is
 923                          * not very useful from caching point of view as
 924                          * the following A or AAAA records will not match the
 925                          * question. We need to find the real A/AAAA record
 926                          * of the alias and cache that.
 927                          */
 928                         unsigned char *end = NULL;
 929                         int name_len = 0, output_len;
 930
 931                         memset(rsp, 0, sizeof(rsp));
 932                         rsp_len = sizeof(rsp) - 1;
 933
 934                         /*
 935                          * Alias is in rdata part of the message,
 936                          * and next-rdlen points to it. So we need to get
 937                          * the real name of the alias.
 938                          */
 939                         ret = get_name(0, buf, next - rdlen, buf + buflen,
 940                                         rsp, rsp_len, &output_len, &end,
 941                                         name, &name_len);
 942                         if (ret != 0) {
 943                                 /* just ignore the error at this point */
 944                                 ptr = next;
 945                                 next = NULL;
 946                                 continue;
 947                         }
 948
 949                         /*
 950                          * We should now have the alias of the entry we might
 951                          * want to cache. Just remember it for a while.
 952                          * We check the alias list when we have parsed the
 953                          * A or AAAA record.
 954                          */
 955                         aliases = g_slist_prepend(aliases, g_strdup(name));
 956
 957                         ptr = next;
 958                         next = NULL;
 959                         continue;
 960                 }
 961
 962                 if (*type == qtype) {
 963                         /*
 964                          * We found correct type (A or AAAA)
 965                          */
 966                         if (check_alias(aliases, name) == TRUE ||
 967                                 (aliases == NULL && strncmp(question, name,
 968                                                         qlen) == 0)) {
 969                                 /*
 970                                  * We found an alias or the name of the rr
 971                                  * matches the question. If so, we append
 972                                  * the compressed label to the cache.
 973                                  * The end result is a response buffer that
 974                                  * will contain one or more cached and
 975                                  * compressed resource records.
 976                                  */
 977                                 if (*response_len + rsp_len > maxlen) {
 978                                         err = -ENOBUFS;
 979                                         goto out;
 980                                 }
 981                                 memcpy(response + *response_len, rsp, rsp_len);
 982                                 *response_len += rsp_len;
 983                                 (*answers)++;
 984                                 err = 0;
 985                         }
 986                 }
 987
 988                 ptr = next;
 989                 next = NULL;
 990         }
 991
 992 out:
 993         for (list = aliases; list; list = list->next)
 994                 g_free(list->data);
 995         g_slist_free(aliases);
 996
 997         return err;
 998 }
 999
1000 struct cache_timeout {
1001         time_t current_time;
1002         int max_timeout;
1003         int try_harder;
1004 };
1005
1006 static gboolean cache_check_entry(gpointer key, gpointer value,
1007                                         gpointer user_data)
1008 {
1009         struct cache_timeout *data = user_data;
1010         struct cache_entry *entry = value;
1011         int max_timeout;
1012
1013         /* Scale the number of hits by half as part of cache aging */
1014
1015         entry->hits /= 2;
1016
1017         /*
1018          * If either IPv4 or IPv6 cached entry has expired, we
1019          * remove both from the cache.
1020          */
1021
1022         if (entry->ipv4 != NULL && entry->ipv4->timeout > 0) {
1023                 max_timeout = entry->ipv4->cache_until;
1024                 if (max_timeout > data->max_timeout)
1025                         data->max_timeout = max_timeout;
1026
1027                 if (entry->ipv4->cache_until < data->current_time)
1028                         return TRUE;
1029         }
1030
1031         if (entry->ipv6 != NULL && entry->ipv6->timeout > 0) {
1032                 max_timeout = entry->ipv6->cache_until;
1033                 if (max_timeout > data->max_timeout)
1034                         data->max_timeout = max_timeout;
1035
1036                 if (entry->ipv6->cache_until < data->current_time)
1037                         return TRUE;
1038         }
1039
1040         /*
1041          * if we're asked to try harder, also remove entries that have
1042          * few hits
1043          */
1044         if (data->try_harder && entry->hits < 4)
1045                 return TRUE;
1046
1047         return FALSE;
1048 }
1049
1050 static void cache_cleanup(void)
1051 {
1052         static int max_timeout;
1053         struct cache_timeout data;
1054         int count = 0;
1055
1056         data.current_time = time(NULL);
1057         data.max_timeout = 0;
1058         data.try_harder = 0;
1059
1060         /*
1061          * In the first pass, we only remove entries that have timed out.
1062          * We use a cache of the first time to expire to do this only
1063          * when it makes sense.
1064          */
1065         if (max_timeout <= data.current_time) {
1066                 count = g_hash_table_foreach_remove(cache, cache_check_entry,
1067                                                 &data);
1068         }
1069         DBG("removed %d in the first pass", count);
1070
1071         /*
1072          * In the second pass, if the first pass turned up blank,
1073          * we also expire entries with a low hit count,
1074          * while aging the hit count at the same time.
1075          */
1076         data.try_harder = 1;
1077         if (count == 0)
1078                 count = g_hash_table_foreach_remove(cache, cache_check_entry,
1079                                                 &data);
1080
1081         if (count == 0)
1082                 /*
1083                  * If we could not remove anything, then remember
1084                  * what is the max timeout and do nothing if we
1085                  * have not yet reached it. This will prevent
1086                  * constant traversal of the cache if it is full.
1087                  */
1088                 max_timeout = data.max_timeout;
1089         else
1090                 max_timeout = 0;
1091 }
1092
1093 static gboolean cache_invalidate_entry(gpointer key, gpointer value,
1094                                         gpointer user_data)
1095 {
1096         struct cache_entry *entry = value;
1097
1098         /* first, delete any expired elements */
1099         cache_enforce_validity(entry);
1100
1101         /* if anything is not expired, mark the entry for refresh */
1102         if (entry->hits > 0 && (entry->ipv4 || entry->ipv6))
1103                 entry->want_refresh = 1;
1104
1105         /* delete the cached data */
1106         if (entry->ipv4) {
1107                 g_free(entry->ipv4->data);
1108                 g_free(entry->ipv4);
1109                 entry->ipv4 = NULL;
1110         }
1111
1112         if (entry->ipv6) {
1113                 g_free(entry->ipv6->data);
1114                 g_free(entry->ipv6);
1115                 entry->ipv6 = NULL;
1116         }
1117
1118         /* keep the entry if we want it refreshed, delete it otherwise */
1119         if (entry->want_refresh)
1120                 return FALSE;
1121         else
1122                 return TRUE;
1123 }
1124
1125 /*
1126  * cache_invalidate is called from places where the DNS landscape
1127  * has changed, say because connections are added or we entered a VPN.
1128  * The logic is to wipe all cache data, but mark all non-expired
1129  * parts of the cache for refresh rather than deleting the whole cache.
1130  */
1131 static void cache_invalidate(void)
1132 {
1133         DBG("Invalidating the DNS cache %p", cache);
1134
1135         if (cache == NULL)
1136                 return;
1137
1138         g_hash_table_foreach_remove(cache, cache_invalidate_entry, NULL);
1139 }
1140
1141 static void cache_refresh_entry(struct cache_entry *entry)
1142 {
1143
1144         cache_enforce_validity(entry);
1145
1146         if (entry->hits > 2 && entry->ipv4 == NULL)
1147                 entry->want_refresh = 1;
1148         if (entry->hits > 2 && entry->ipv6 == NULL)
1149                 entry->want_refresh = 1;
1150
1151         if (entry->want_refresh) {
1152                 char *c;
1153                 char dns_name[NS_MAXDNAME + 1];
1154                 entry->want_refresh = 0;
1155
1156                 /* turn a DNS name into a hostname with dots */
1157                 strncpy(dns_name, entry->key, NS_MAXDNAME);
1158                 c = dns_name;
1159                 while (c && *c) {
1160                         int jump;
1161                         jump = *c;
1162                         *c = '.';
1163                         c += jump + 1;
1164                 }
1165                 DBG("Refreshing %s\n", dns_name);
1166                 /* then refresh the hostname */
1167                 refresh_dns_entry(entry, &dns_name[1]);
1168         }
1169 }
1170
1171 static void cache_refresh_iterator(gpointer key, gpointer value,
1172                                         gpointer user_data)
1173 {
1174         struct cache_entry *entry = value;
1175
1176         cache_refresh_entry(entry);
1177 }
1178
1179 static void cache_refresh(void)
1180 {
1181         if (cache == NULL)
1182                 return;
1183
1184         g_hash_table_foreach(cache, cache_refresh_iterator, NULL);
1185 }
1186
1187 static int reply_query_type(unsigned char *msg, int len)
1188 {
1189         unsigned char *c;
1190         uint16_t *w;
1191         int l;
1192         int type;
1193
1194         /* skip the header */
1195         c = msg + sizeof(struct domain_hdr);
1196         len -= sizeof(struct domain_hdr);
1197
1198         if (len < 0)
1199                 return 0;
1200
1201         /* now the query, which is a name and 2 16 bit words */
1202         l = dns_name_length(c) + 1;
1203         c += l;
1204         w = (uint16_t *) c;
1205         type = ntohs(*w);
1206
1207         return type;
1208 }
1209
1210 static int cache_update(struct server_data *srv, unsigned char *msg,
1211                         unsigned int msg_len)
1212 {
1213         int offset = protocol_offset(srv->protocol);
1214         int err, qlen, ttl = 0;
1215         uint16_t answers = 0, type = 0, class = 0;
1216         struct domain_question *q;
1217         struct cache_entry *entry;
1218         struct cache_data *data;
1219         char question[NS_MAXDNAME + 1];
1220         unsigned char response[NS_MAXDNAME + 1];
1221         unsigned char *ptr;
1222         unsigned int rsplen;
1223         gboolean new_entry = TRUE;
1224         time_t current_time;
1225
1226         if (cache_size >= MAX_CACHE_SIZE) {
1227                 cache_cleanup();
1228                 if (cache_size >= MAX_CACHE_SIZE)
1229                         return 0;
1230         }
1231
1232         current_time = time(NULL);
1233
1234         /* don't do a cache refresh more than twice a minute */
1235         if (next_refresh < current_time) {
1236                 cache_refresh();
1237                 next_refresh = current_time + 30;
1238         }
1239
1240
1241         /* Continue only if response code is 0 (=ok) */
1242         if (msg[3] & 0x0f)
1243                 return 0;
1244
1245         if (offset < 0)
1246                 return 0;
1247
1248         rsplen = sizeof(response) - 1;
1249         question[sizeof(question) - 1] = '\0';
1250
1251         err = parse_response(msg + offset, msg_len - offset,
1252                                 question, sizeof(question) - 1,
1253                                 &type, &class, &ttl,
1254                                 response, &rsplen, &answers);
1255
1256         /*
1257          * special case: if we do a ipv6 lookup and get no result
1258          * for a record that's already in our ipv4 cache.. we want
1259          * to cache the negative response.
1260          */
1261         if ((err == -ENOMSG || err == -ENOBUFS) &&
1262                         reply_query_type(msg, msg_len) == 28) {
1263                 entry = g_hash_table_lookup(cache, question);
1264                 if (entry && entry->ipv4 && entry->ipv6 == NULL) {
1265                         data = g_try_new(struct cache_data, 1);
1266                         if (data == NULL)
1267                                 return -ENOMEM;
1268                         data->inserted = entry->ipv4->inserted;
1269                         data->type = type;
1270                         data->answers = msg[5];
1271                         data->timeout = entry->ipv4->timeout;
1272                         data->data_len = msg_len;
1273                         data->data = ptr = g_malloc(msg_len);
1274                         data->valid_until = entry->ipv4->valid_until;
1275                         data->cache_until = entry->ipv4->cache_until;
1276                         memcpy(data->data, msg, msg_len);
1277                         entry->ipv6 = data;
1278                         /*
1279                          * we will get a "hit" when we serve the response
1280                          * out of the cache
1281                          */
1282                         entry->hits--;
1283                         if (entry->hits < 0)
1284                                 entry->hits = 0;
1285                         return 0;
1286                 }
1287         }
1288
1289         if (err < 0 || ttl == 0)
1290                 return 0;
1291
1292         qlen = strlen(question);
1293
1294         /*
1295          * If the cache contains already data, check if the
1296          * type of the cached data is the same and do not add
1297          * to cache if data is already there.
1298          * This is needed so that we can cache both A and AAAA
1299          * records for the same name.
1300          */
1301         entry = g_hash_table_lookup(cache, question);
1302         if (entry == NULL) {
1303                 entry = g_try_new(struct cache_entry, 1);
1304                 if (entry == NULL)
1305                         return -ENOMEM;
1306
1307                 data = g_try_new(struct cache_data, 1);
1308                 if (data == NULL) {
1309                         g_free(entry);
1310                         return -ENOMEM;
1311                 }
1312
1313                 entry->key = g_strdup(question);
1314                 entry->ipv4 = entry->ipv6 = NULL;
1315                 entry->want_refresh = 0;
1316                 entry->hits = 0;
1317
1318                 if (type == 1)
1319                         entry->ipv4 = data;
1320                 else
1321                         entry->ipv6 = data;
1322         } else {
1323                 if (type == 1 && entry->ipv4 != NULL)
1324                         return 0;
1325
1326                 if (type == 28 && entry->ipv6 != NULL)
1327                         return 0;
1328
1329                 data = g_try_new(struct cache_data, 1);
1330                 if (data == NULL)
1331                         return -ENOMEM;
1332
1333                 if (type == 1)
1334                         entry->ipv4 = data;
1335                 else
1336                         entry->ipv6 = data;
1337
1338                 /*
1339                  * compensate for the hit we'll get for serving
1340                  * the response out of the cache
1341                  */
1342                 entry->hits--;
1343                 if (entry->hits < 0)
1344                         entry->hits = 0;
1345
1346                 new_entry = FALSE;
1347         }
1348
1349         if (ttl < MIN_CACHE_TTL)
1350                 ttl = MIN_CACHE_TTL;
1351
1352         data->inserted = current_time;
1353         data->type = type;
1354         data->answers = answers;
1355         data->timeout = ttl;
1356         data->data_len = 12 + qlen + 1 + 2 + 2 + rsplen;
1357         data->data = ptr = g_malloc(data->data_len);
1358         data->valid_until = current_time + ttl;
1359
1360         /*
1361          * Restrict the cached DNS record TTL to some sane value
1362          * in order to prevent data staying in the cache too long.
1363          */
1364         if (ttl > MAX_CACHE_TTL)
1365                 ttl = MAX_CACHE_TTL;
1366
1367         data->cache_until = round_down_ttl(current_time + ttl, ttl);
1368
1369         if (data->data == NULL) {
1370                 g_free(entry->key);
1371                 g_free(data);
1372                 g_free(entry);
1373                 return -ENOMEM;
1374         }
1375
1376         memcpy(ptr, msg, 12);
1377         memcpy(ptr + 12, question, qlen + 1); /* copy also the \0 */
1378
1379         q = (void *) (ptr + 12 + qlen + 1);
1380         q->type = htons(type);
1381         q->class = htons(class);
1382         memcpy(ptr + 12 + qlen + 1 + sizeof(struct domain_question),
1383                 response, rsplen);
1384
1385         if (new_entry == TRUE) {
1386                 g_hash_table_replace(cache, entry->key, entry);
1387                 cache_size++;
1388         }
1389
1390         DBG("cache %d %squestion \"%s\" type %d ttl %d size %zd",
1391                 cache_size, new_entry ? "new " : "old ",
1392                 question, type, ttl,
1393                 sizeof(*entry) + sizeof(*data) + data->data_len + qlen);
1394
1395         return 0;
1396 }
1397
1398 static int ns_resolv(struct server_data *server, struct request_data *req,
1399                                 gpointer request, gpointer name)
1400 {
1401         GList *list;
1402         int sk, err, type = 0;
1403         char *dot, *lookup = (char *) name;
1404         struct cache_entry *entry;
1405
1406         entry = cache_check(request, &type);
1407         if (entry != NULL) {
1408                 int ttl_left = 0;
1409                 struct cache_data *data;
1410
1411                 DBG("cache hit %s type %s", lookup, type == 1 ? "A" : "AAAA");
1412                 if (type == 1)
1413                         data = entry->ipv4;
1414                 else
1415                         data = entry->ipv6;
1416
1417                 if (data) {
1418                         ttl_left = data->valid_until - time(NULL);
1419                         entry->hits++;
1420                 }
1421
1422                 if (data != NULL && req->protocol == IPPROTO_TCP) {
1423                         send_cached_response(req->client_sk, data->data,
1424                                         data->data_len, NULL, 0, IPPROTO_TCP,
1425                                         req->srcid, data->answers, ttl_left);
1426                         return 1;
1427                 }
1428
1429                 if (data != NULL && req->protocol == IPPROTO_UDP) {
1430                         int sk;
1431                         sk = g_io_channel_unix_get_fd(
1432                                         req->ifdata->udp_listener_channel);
1433
1434                         send_cached_response(sk, data->data,
1435                                 data->data_len, &req->sa, req->sa_len,
1436                                 IPPROTO_UDP, req->srcid, data->answers,
1437                                 ttl_left);
1438                         return 1;
1439                 }
1440         }
1441
1442         sk = g_io_channel_unix_get_fd(server->channel);
1443
1444         err = send(sk, request, req->request_len, 0);
1445         if (err < 0)
1446                 return -EIO;
1447
1448         req->numserv++;
1449
1450         /* If we have more than one dot, we don't add domains */
1451         dot = strchr(lookup, '.');
1452         if (dot != NULL && dot != lookup + strlen(lookup) - 1)
1453                 return 0;
1454
1455         if (server->domains != NULL && server->domains->data != NULL)
1456                 req->append_domain = TRUE;
1457
1458         for (list = server->domains; list; list = list->next) {
1459                 char *domain;
1460                 unsigned char alt[1024];
1461                 struct domain_hdr *hdr = (void *) &alt;
1462                 int altlen, domlen, offset;
1463
1464                 domain = list->data;
1465
1466                 if (domain == NULL)
1467                         continue;
1468
1469                 offset = protocol_offset(server->protocol);
1470                 if (offset < 0)
1471                         return offset;
1472
1473                 domlen = strlen(domain) + 1;
1474                 if (domlen < 5)
1475                         return -EINVAL;
1476
1477                 alt[offset] = req->altid & 0xff;
1478                 alt[offset + 1] = req->altid >> 8;
1479
1480                 memcpy(alt + offset + 2, request + offset + 2, 10);
1481                 hdr->qdcount = htons(1);
1482
1483                 altlen = append_query(alt + offset + 12, sizeof(alt) - 12,
1484                                         name, domain);
1485                 if (altlen < 0)
1486                         return -EINVAL;
1487
1488                 altlen += 12;
1489
1490                 memcpy(alt + offset + altlen,
1491                         request + offset + altlen - domlen,
1492                                 req->request_len - altlen - offset + domlen);
1493
1494                 if (server->protocol == IPPROTO_TCP) {
1495                         int req_len = req->request_len + domlen - 2;
1496
1497                         alt[0] = (req_len >> 8) & 0xff;
1498                         alt[1] = req_len & 0xff;
1499                 }
1500
1501                 err = send(sk, alt, req->request_len + domlen, 0);
1502                 if (err < 0)
1503                         return -EIO;
1504
1505                 req->numserv++;
1506         }
1507
1508         return 0;
1509 }
1510
1511 static int forward_dns_reply(unsigned char *reply, int reply_len, int protocol,
1512                                 struct server_data *data)
1513 {
1514         struct domain_hdr *hdr;
1515         struct request_data *req;
1516         int dns_id, sk, err, offset = protocol_offset(protocol);
1517         struct listener_data *ifdata;
1518
1519         if (offset < 0)
1520                 return offset;
1521
1522         hdr = (void *)(reply + offset);
1523         dns_id = reply[offset] | reply[offset + 1] << 8;
1524
1525         DBG("Received %d bytes (id 0x%04x)", reply_len, dns_id);
1526
1527         req = find_request(dns_id);
1528         if (req == NULL)
1529                 return -EINVAL;
1530
1531         DBG("id 0x%04x rcode %d", hdr->id, hdr->rcode);
1532
1533         ifdata = req->ifdata;
1534
1535         reply[offset] = req->srcid & 0xff;
1536         reply[offset + 1] = req->srcid >> 8;
1537
1538         req->numresp++;
1539
1540         if (hdr->rcode == 0 || req->resp == NULL) {
1541
1542                 /*
1543                  * If the domain name was append
1544                  * remove it before forwarding the reply.
1545                  */
1546                 if (req->append_domain == TRUE) {
1547                         unsigned char *ptr;
1548                         uint8_t host_len;
1549                         unsigned int domain_len;
1550
1551                         /*
1552                          * ptr points to the first char of the hostname.
1553                          * ->hostname.domain.net
1554                          */
1555                         ptr = reply + offset + sizeof(struct domain_hdr);
1556                         host_len = *ptr;
1557                         domain_len = strlen((const char *)ptr) - host_len - 1;
1558
1559                         /*
1560                          * remove the domain name and replaced it by the end
1561                          * of reply.
1562                          */
1563                         memmove(ptr + host_len + 1,
1564                                 ptr + host_len + domain_len + 1,
1565                                 reply_len - (ptr - reply + domain_len));
1566
1567                         reply_len = reply_len - domain_len;
1568                 }
1569
1570                 g_free(req->resp);
1571                 req->resplen = 0;
1572
1573                 req->resp = g_try_malloc(reply_len);
1574                 if (req->resp == NULL)
1575                         return -ENOMEM;
1576
1577                 memcpy(req->resp, reply, reply_len);
1578                 req->resplen = reply_len;
1579
1580                 cache_update(data, reply, reply_len);
1581         }
1582
1583         if (hdr->rcode > 0 && req->numresp < req->numserv)
1584                 return -EINVAL;
1585
1586         if (req->timeout > 0)
1587                 g_source_remove(req->timeout);
1588
1589         request_list = g_slist_remove(request_list, req);
1590
1591         if (protocol == IPPROTO_UDP) {
1592                 sk = g_io_channel_unix_get_fd(ifdata->udp_listener_channel);
1593                 err = sendto(sk, req->resp, req->resplen, 0,
1594                              &req->sa, req->sa_len);
1595         } else {
1596                 sk = req->client_sk;
1597                 err = send(sk, req->resp, req->resplen, 0);
1598                 close(sk);
1599         }
1600
1601         g_free(req->resp);
1602         g_free(req);
1603
1604         return err;
1605 }
1606
1607 static void cache_element_destroy(gpointer value)
1608 {
1609         struct cache_entry *entry = value;
1610
1611         if (entry == NULL)
1612                 return;
1613
1614         if (entry->ipv4 != NULL) {
1615                 g_free(entry->ipv4->data);
1616                 g_free(entry->ipv4);
1617         }
1618
1619         if (entry->ipv6 != NULL) {
1620                 g_free(entry->ipv6->data);
1621                 g_free(entry->ipv6);
1622         }
1623
1624         g_free(entry->key);
1625         g_free(entry);
1626
1627         if (--cache_size < 0)
1628                 cache_size = 0;
1629 }
1630
1631 static gboolean try_remove_cache(gpointer user_data)
1632 {
1633         if (__sync_fetch_and_sub(&cache_refcount, 1) == 1) {
1634                 DBG("No cache users, removing it.");
1635
1636                 g_hash_table_destroy(cache);
1637                 cache = NULL;
1638         }
1639
1640         return FALSE;
1641 }
1642
1643 static void destroy_server(struct server_data *server)
1644 {
1645         GList *list;
1646
1647         DBG("interface %s server %s", server->interface, server->server);
1648
1649         server_list = g_slist_remove(server_list, server);
1650
1651         if (server->watch > 0)
1652                 g_source_remove(server->watch);
1653
1654         if (server->timeout > 0)
1655                 g_source_remove(server->timeout);
1656
1657         g_io_channel_unref(server->channel);
1658
1659         if (server->protocol == IPPROTO_UDP)
1660                 connman_info("Removing DNS server %s", server->server);
1661
1662         g_free(server->incoming_reply);
1663         g_free(server->server);
1664         for (list = server->domains; list; list = list->next) {
1665                 char *domain = list->data;
1666
1667                 server->domains = g_list_remove(server->domains, domain);
1668                 g_free(domain);
1669         }
1670         g_free(server->interface);
1671
1672         /*
1673          * We do not remove cache right away but delay it few seconds.
1674          * The idea is that when IPv6 DNS server is added via RDNSS, it has a
1675          * lifetime. When the lifetime expires we decrease the refcount so it
1676          * is possible that the cache is then removed. Because a new DNS server
1677          * is usually created almost immediately we would then loose the cache
1678          * without any good reason. The small delay allows the new RDNSS to
1679          * create a new DNS server instance and the refcount does not go to 0.
1680          */
1681         g_timeout_add_seconds(3, try_remove_cache, NULL);
1682
1683         g_free(server);
1684 }
1685
1686 static gboolean udp_server_event(GIOChannel *channel, GIOCondition condition,
1687                                                         gpointer user_data)
1688 {
1689         unsigned char buf[4096];
1690         int sk, err, len;
1691         struct server_data *data = user_data;
1692
1693         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
1694                 connman_error("Error with UDP server %s", data->server);
1695                 data->watch = 0;
1696                 return FALSE;
1697         }
1698
1699         sk = g_io_channel_unix_get_fd(channel);
1700
1701         len = recv(sk, buf, sizeof(buf), 0);
1702         if (len < 12)
1703                 return TRUE;
1704
1705         err = forward_dns_reply(buf, len, IPPROTO_UDP, data);
1706         if (err < 0)
1707                 return TRUE;
1708
1709         return TRUE;
1710 }
1711
1712 static gboolean tcp_server_event(GIOChannel *channel, GIOCondition condition,
1713                                                         gpointer user_data)
1714 {
1715         int sk;
1716         struct server_data *server = user_data;
1717
1718         sk = g_io_channel_unix_get_fd(channel);
1719         if (sk == 0)
1720                 return FALSE;
1721
1722         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
1723                 GSList *list;
1724 hangup:
1725                 DBG("TCP server channel closed");
1726
1727                 /*
1728                  * Discard any partial response which is buffered; better
1729                  * to get a proper response from a working server.
1730                  */
1731                 g_free(server->incoming_reply);
1732                 server->incoming_reply = NULL;
1733
1734                 for (list = request_list; list; list = list->next) {
1735                         struct request_data *req = list->data;
1736                         struct domain_hdr *hdr;
1737
1738                         if (req->protocol == IPPROTO_UDP)
1739                                 continue;
1740
1741                         if (req->request == NULL)
1742                                 continue;
1743
1744                         /*
1745                          * If we're not waiting for any further response
1746                          * from another name server, then we send an error
1747                          * response to the client.
1748                          */
1749                         if (req->numserv && --(req->numserv))
1750                                 continue;
1751
1752                         hdr = (void *) (req->request + 2);
1753                         hdr->id = req->srcid;
1754                         send_response(req->client_sk, req->request,
1755                                 req->request_len, NULL, 0, IPPROTO_TCP);
1756
1757                         request_list = g_slist_remove(request_list, req);
1758                 }
1759
1760                 destroy_server(server);
1761
1762                 return FALSE;
1763         }
1764
1765         if ((condition & G_IO_OUT) && !server->connected) {
1766                 GSList *list;
1767                 GList *domains;
1768                 struct server_data *udp_server;
1769
1770                 udp_server = find_server(server->interface, server->server,
1771                                                                 IPPROTO_UDP);
1772                 if (udp_server != NULL) {
1773                         for (domains = udp_server->domains; domains;
1774                                                 domains = domains->next) {
1775                                 char *dom = domains->data;
1776
1777                                 DBG("Adding domain %s to %s",
1778                                                 dom, server->server);
1779
1780                                 server->domains = g_list_append(server->domains,
1781                                                                 g_strdup(dom));
1782                         }
1783                 }
1784
1785                 server->connected = TRUE;
1786                 server_list = g_slist_append(server_list, server);
1787
1788                 if (server->timeout > 0) {
1789                         g_source_remove(server->timeout);
1790                         server->timeout = 0;
1791                 }
1792
1793                 for (list = request_list; list; list = list->next) {
1794                         struct request_data *req = list->data;
1795
1796                         if (req->protocol == IPPROTO_UDP)
1797                                 continue;
1798
1799                         DBG("Sending req %s over TCP", (char *)req->name);
1800
1801                         if (req->timeout > 0)
1802                                 g_source_remove(req->timeout);
1803
1804                         req->timeout = g_timeout_add_seconds(30,
1805                                                 request_timeout, req);
1806                         if (ns_resolv(server, req, req->request,
1807                                         req->name) > 0) {
1808                                 /* We sent cached result so no need for timeout
1809                                  * handler.
1810                                  */
1811                                 if (req->timeout > 0) {
1812                                         g_source_remove(req->timeout);
1813                                         req->timeout = 0;
1814                                 }
1815                         }
1816                 }
1817
1818         } else if (condition & G_IO_IN) {
1819                 struct partial_reply *reply = server->incoming_reply;
1820                 int bytes_recv;
1821
1822                 if (!reply) {
1823                         unsigned char reply_len_buf[2];
1824                         uint16_t reply_len;
1825
1826                         bytes_recv = recv(sk, reply_len_buf, 2, MSG_PEEK);
1827                         if (!bytes_recv) {
1828                                 goto hangup;
1829                         } else if (bytes_recv < 0) {
1830                                 if (errno == EAGAIN || errno == EWOULDBLOCK)
1831                                         return TRUE;
1832
1833                                 connman_error("DNS proxy error %s",
1834                                                 strerror(errno));
1835                                 goto hangup;
1836                         } else if (bytes_recv < 2)
1837                                 return TRUE;
1838
1839                         reply_len = reply_len_buf[1] | reply_len_buf[0] << 8;
1840                         reply_len += 2;
1841
1842                         DBG("TCP reply %d bytes", reply_len);
1843
1844                         reply = g_try_malloc(sizeof(*reply) + reply_len + 2);
1845                         if (!reply)
1846                                 return TRUE;
1847
1848                         reply->len = reply_len;
1849                         reply->received = 0;
1850
1851                         server->incoming_reply = reply;
1852                 }
1853
1854                 while (reply->received < reply->len) {
1855                         bytes_recv = recv(sk, reply->buf + reply->received,
1856                                         reply->len - reply->received, 0);
1857                         if (!bytes_recv) {
1858                                 connman_error("DNS proxy TCP disconnect");
1859                                 break;
1860                         } else if (bytes_recv < 0) {
1861                                 if (errno == EAGAIN || errno == EWOULDBLOCK)
1862                                         return TRUE;
1863
1864                                 connman_error("DNS proxy error %s",
1865                                                 strerror(errno));
1866                                 break;
1867                         }
1868                         reply->received += bytes_recv;
1869                 }
1870
1871                 forward_dns_reply(reply->buf, reply->received, IPPROTO_TCP,
1872                                         server);
1873
1874                 g_free(reply);
1875                 server->incoming_reply = NULL;
1876
1877                 destroy_server(server);
1878
1879                 return FALSE;
1880         }
1881
1882         return TRUE;
1883 }
1884
1885 static gboolean tcp_idle_timeout(gpointer user_data)
1886 {
1887         struct server_data *server = user_data;
1888
1889         DBG("");
1890
1891         if (server == NULL)
1892                 return FALSE;
1893
1894         destroy_server(server);
1895
1896         return FALSE;
1897 }
1898
1899 static struct server_data *create_server(const char *interface,
1900                                         const char *domain, const char *server,
1901                                         int protocol)
1902 {
1903         struct addrinfo hints, *rp;
1904         struct server_data *data;
1905         int sk, ret;
1906
1907         DBG("interface %s server %s", interface, server);
1908
1909         memset(&hints, 0, sizeof(hints));
1910
1911         switch (protocol) {
1912         case IPPROTO_UDP:
1913                 hints.ai_socktype = SOCK_DGRAM;
1914                 break;
1915
1916         case IPPROTO_TCP:
1917                 hints.ai_socktype = SOCK_STREAM;
1918                 break;
1919
1920         default:
1921                 return NULL;
1922         }
1923         hints.ai_family = AF_UNSPEC;
1924         hints.ai_flags = AI_PASSIVE | AI_NUMERICSERV | AI_NUMERICHOST;
1925
1926         ret = getaddrinfo(server, "53", &hints, &rp);
1927         if (ret) {
1928                 connman_error("Failed to parse server %s address: %s\n",
1929                               server, gai_strerror(ret));
1930                 return NULL;
1931         }
1932         /* Do not blindly copy this code elsewhere; it doesn't loop over the
1933            results using ->ai_next as it should. That's OK in *this* case
1934            because it was a numeric lookup; we *know* there's only one. */
1935
1936         sk = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
1937         if (sk < 0) {
1938                 connman_error("Failed to create server %s socket", server);
1939                 freeaddrinfo(rp);
1940                 return NULL;
1941         }
1942
1943         if (interface != NULL) {
1944                 if (setsockopt(sk, SOL_SOCKET, SO_BINDTODEVICE,
1945                                 interface, strlen(interface) + 1) < 0) {
1946                         connman_error("Failed to bind server %s "
1947                                                 "to interface %s",
1948                                                         server, interface);
1949                         freeaddrinfo(rp);
1950                         close(sk);
1951                         return NULL;
1952                 }
1953         }
1954
1955         data = g_try_new0(struct server_data, 1);
1956         if (data == NULL) {
1957                 connman_error("Failed to allocate server %s data", server);
1958                 freeaddrinfo(rp);
1959                 close(sk);
1960                 return NULL;
1961         }
1962
1963         data->channel = g_io_channel_unix_new(sk);
1964         if (data->channel == NULL) {
1965                 connman_error("Failed to create server %s channel", server);
1966                 freeaddrinfo(rp);
1967                 close(sk);
1968                 g_free(data);
1969                 return NULL;
1970         }
1971
1972         g_io_channel_set_close_on_unref(data->channel, TRUE);
1973
1974         if (protocol == IPPROTO_TCP) {
1975                 g_io_channel_set_flags(data->channel, G_IO_FLAG_NONBLOCK, NULL);
1976                 data->watch = g_io_add_watch(data->channel,
1977                         G_IO_OUT | G_IO_IN | G_IO_HUP | G_IO_NVAL | G_IO_ERR,
1978                                                 tcp_server_event, data);
1979                 data->timeout = g_timeout_add_seconds(30, tcp_idle_timeout,
1980                                                                 data);
1981         } else
1982                 data->watch = g_io_add_watch(data->channel,
1983                         G_IO_IN | G_IO_NVAL | G_IO_ERR | G_IO_HUP,
1984                                                 udp_server_event, data);
1985
1986         data->interface = g_strdup(interface);
1987         if (domain)
1988                 data->domains = g_list_append(data->domains, g_strdup(domain));
1989         data->server = g_strdup(server);
1990         data->protocol = protocol;
1991
1992         ret = connect(sk, rp->ai_addr, rp->ai_addrlen);
1993         freeaddrinfo(rp);
1994         if (ret < 0) {
1995                 if ((protocol == IPPROTO_TCP && errno != EINPROGRESS) ||
1996                                 protocol == IPPROTO_UDP) {
1997                         GList *list;
1998
1999                         connman_error("Failed to connect to server %s", server);
2000                         if (data->watch > 0)
2001                                 g_source_remove(data->watch);
2002                         if (data->timeout > 0)
2003                                 g_source_remove(data->timeout);
2004
2005                         g_io_channel_unref(data->channel);
2006                         close(sk);
2007
2008                         g_free(data->server);
2009                         g_free(data->interface);
2010                         for (list = data->domains; list; list = list->next) {
2011                                 char *domain = list->data;
2012
2013                                 data->domains = g_list_remove(data->domains,
2014                                                                         domain);
2015                                 g_free(domain);
2016                         }
2017                         g_free(data);
2018                         return NULL;
2019                 }
2020         }
2021
2022         if (__sync_fetch_and_add(&cache_refcount, 1) == 0)
2023                 cache = g_hash_table_new_full(g_str_hash,
2024                                         g_str_equal,
2025                                         NULL,
2026                                         cache_element_destroy);
2027
2028         if (protocol == IPPROTO_UDP) {
2029                 /* Enable new servers by default */
2030                 data->enabled = TRUE;
2031                 connman_info("Adding DNS server %s", data->server);
2032
2033                 server_list = g_slist_append(server_list, data);
2034
2035                 return data;
2036         }
2037
2038         return NULL;
2039 }
2040
2041 static gboolean resolv(struct request_data *req,
2042                                 gpointer request, gpointer name)
2043 {
2044         GSList *list;
2045         int status;
2046
2047         for (list = server_list; list; list = list->next) {
2048                 struct server_data *data = list->data;
2049
2050                 DBG("server %s enabled %d", data->server, data->enabled);
2051
2052                 if (data->enabled == FALSE)
2053                         continue;
2054
2055                 if (data->watch == 0 && data->protocol == IPPROTO_UDP)
2056                         data->watch = g_io_add_watch(data->channel,
2057                                 G_IO_IN | G_IO_NVAL | G_IO_ERR | G_IO_HUP,
2058                                                 udp_server_event, data);
2059
2060                 status = ns_resolv(data, req, request, name);
2061                 if (status < 0)
2062                         continue;
2063
2064                 if (status > 0) {
2065                         if (req->timeout > 0) {
2066                                 g_source_remove(req->timeout);
2067                                 req->timeout = 0;
2068                         }
2069                 }
2070         }
2071
2072         return TRUE;
2073 }
2074
2075 static void append_domain(const char *interface, const char *domain)
2076 {
2077         GSList *list;
2078
2079         DBG("interface %s domain %s", interface, domain);
2080
2081         if (domain == NULL)
2082                 return;
2083
2084         for (list = server_list; list; list = list->next) {
2085                 struct server_data *data = list->data;
2086                 GList *dom_list;
2087                 char *dom;
2088                 gboolean dom_found = FALSE;
2089
2090                 if (data->interface == NULL)
2091                         continue;
2092
2093                 if (g_str_equal(data->interface, interface) == FALSE)
2094                         continue;
2095
2096                 for (dom_list = data->domains; dom_list;
2097                                 dom_list = dom_list->next) {
2098                         dom = dom_list->data;
2099
2100                         if (g_str_equal(dom, domain)) {
2101                                 dom_found = TRUE;
2102                                 break;
2103                         }
2104                 }
2105
2106                 if (dom_found == FALSE) {
2107                         data->domains =
2108                                 g_list_append(data->domains, g_strdup(domain));
2109                 }
2110         }
2111 }
2112
2113 int __connman_dnsproxy_append(const char *interface, const char *domain,
2114                                                         const char *server)
2115 {
2116         struct server_data *data;
2117
2118         DBG("interface %s server %s", interface, server);
2119
2120         if (server == NULL && domain == NULL)
2121                 return -EINVAL;
2122
2123         if (server == NULL) {
2124                 append_domain(interface, domain);
2125
2126                 return 0;
2127         }
2128
2129         if (g_str_equal(server, "127.0.0.1") == TRUE)
2130                 return -ENODEV;
2131
2132         data = find_server(interface, server, IPPROTO_UDP);
2133         if (data != NULL) {
2134                 append_domain(interface, domain);
2135                 return 0;
2136         }
2137
2138         data = create_server(interface, domain, server, IPPROTO_UDP);
2139         if (data == NULL)
2140                 return -EIO;
2141
2142         return 0;
2143 }
2144
2145 static void remove_server(const char *interface, const char *domain,
2146                         const char *server, int protocol)
2147 {
2148         struct server_data *data;
2149
2150         data = find_server(interface, server, protocol);
2151         if (data == NULL)
2152                 return;
2153
2154         destroy_server(data);
2155 }
2156
2157 int __connman_dnsproxy_remove(const char *interface, const char *domain,
2158                                                         const char *server)
2159 {
2160         DBG("interface %s server %s", interface, server);
2161
2162         if (server == NULL)
2163                 return -EINVAL;
2164
2165         if (g_str_equal(server, "127.0.0.1") == TRUE)
2166                 return -ENODEV;
2167
2168         remove_server(interface, domain, server, IPPROTO_UDP);
2169         remove_server(interface, domain, server, IPPROTO_TCP);
2170
2171         return 0;
2172 }
2173
2174 void __connman_dnsproxy_flush(void)
2175 {
2176         GSList *list;
2177
2178         list = request_pending_list;
2179         while (list) {
2180                 struct request_data *req = list->data;
2181
2182                 list = list->next;
2183
2184                 request_pending_list =
2185                                 g_slist_remove(request_pending_list, req);
2186                 resolv(req, req->request, req->name);
2187                 g_free(req->request);
2188                 g_free(req->name);
2189         }
2190 }
2191
2192 static void dnsproxy_offline_mode(connman_bool_t enabled)
2193 {
2194         GSList *list;
2195
2196         DBG("enabled %d", enabled);
2197
2198         for (list = server_list; list; list = list->next) {
2199                 struct server_data *data = list->data;
2200
2201                 if (enabled == FALSE) {
2202                         connman_info("Enabling DNS server %s", data->server);
2203                         data->enabled = TRUE;
2204                         cache_invalidate();
2205                         cache_refresh();
2206                 } else {
2207                         connman_info("Disabling DNS server %s", data->server);
2208                         data->enabled = FALSE;
2209                         cache_invalidate();
2210                 }
2211         }
2212 }
2213
2214 static void dnsproxy_default_changed(struct connman_service *service)
2215 {
2216         GSList *list;
2217         char *interface;
2218
2219         DBG("service %p", service);
2220
2221         /* DNS has changed, invalidate the cache */
2222         cache_invalidate();
2223
2224         if (service == NULL) {
2225                 /* When no services are active, then disable DNS proxying */
2226                 dnsproxy_offline_mode(TRUE);
2227                 return;
2228         }
2229
2230         interface = connman_service_get_interface(service);
2231         if (interface == NULL)
2232                 return;
2233
2234         for (list = server_list; list; list = list->next) {
2235                 struct server_data *data = list->data;
2236
2237                 if (g_strcmp0(data->interface, interface) == 0) {
2238                         connman_info("Enabling DNS server %s", data->server);
2239                         data->enabled = TRUE;
2240                 } else {
2241                         connman_info("Disabling DNS server %s", data->server);
2242                         data->enabled = FALSE;
2243                 }
2244         }
2245
2246         g_free(interface);
2247         cache_refresh();
2248 }
2249
2250 static struct connman_notifier dnsproxy_notifier = {
2251         .name                   = "dnsproxy",
2252         .default_changed        = dnsproxy_default_changed,
2253         .offline_mode           = dnsproxy_offline_mode,
2254 };
2255
2256 static unsigned char opt_edns0_type[2] = { 0x00, 0x29 };
2257
2258 static int parse_request(unsigned char *buf, int len,
2259                                         char *name, unsigned int size)
2260 {
2261         struct domain_hdr *hdr = (void *) buf;
2262         uint16_t qdcount = ntohs(hdr->qdcount);
2263         uint16_t arcount = ntohs(hdr->arcount);
2264         unsigned char *ptr;
2265         char *last_label = NULL;
2266         unsigned int remain, used = 0;
2267
2268         if (len < 12)
2269                 return -EINVAL;
2270
2271         DBG("id 0x%04x qr %d opcode %d qdcount %d arcount %d",
2272                                         hdr->id, hdr->qr, hdr->opcode,
2273                                                         qdcount, arcount);
2274
2275         if (hdr->qr != 0 || qdcount != 1)
2276                 return -EINVAL;
2277
2278         memset(name, 0, size);
2279
2280         ptr = buf + sizeof(struct domain_hdr);
2281         remain = len - sizeof(struct domain_hdr);
2282
2283         while (remain > 0) {
2284                 uint8_t len = *ptr;
2285
2286                 if (len == 0x00) {
2287                         last_label = (char *) (ptr + 1);
2288                         break;
2289                 }
2290
2291                 if (used + len + 1 > size)
2292                         return -ENOBUFS;
2293
2294                 strncat(name, (char *) (ptr + 1), len);
2295                 strcat(name, ".");
2296
2297                 used += len + 1;
2298
2299                 ptr += len + 1;
2300                 remain -= len + 1;
2301         }
2302
2303         if (last_label && arcount && remain >= 9 && last_label[4] == 0 &&
2304                                 !memcmp(last_label + 5, opt_edns0_type, 2)) {
2305                 uint16_t edns0_bufsize;
2306
2307                 edns0_bufsize = last_label[7] << 8 | last_label[8];
2308
2309                 DBG("EDNS0 buffer size %u", edns0_bufsize);
2310
2311                 /* This is an evil hack until full TCP support has been
2312                  * implemented.
2313                  *
2314                  * Somtimes the EDNS0 request gets send with a too-small
2315                  * buffer size. Since glibc doesn't seem to crash when it
2316                  * gets a response biffer then it requested, just bump
2317                  * the buffer size up to 4KiB.
2318                  */
2319                 if (edns0_bufsize < 0x1000) {
2320                         last_label[7] = 0x10;
2321                         last_label[8] = 0x00;
2322                 }
2323         }
2324
2325         DBG("query %s", name);
2326
2327         return 0;
2328 }
2329
2330 static gboolean tcp_listener_event(GIOChannel *channel, GIOCondition condition,
2331                                                         gpointer user_data)
2332 {
2333         unsigned char buf[768];
2334         char query[512];
2335         struct request_data *req;
2336         struct server_data *server;
2337         int sk, client_sk, len, err;
2338         struct sockaddr_in6 client_addr;
2339         socklen_t client_addr_len = sizeof(client_addr);
2340         GSList *list;
2341         struct listener_data *ifdata = user_data;
2342
2343         DBG("condition 0x%x", condition);
2344
2345         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
2346                 if (ifdata->tcp_listener_watch > 0)
2347                         g_source_remove(ifdata->tcp_listener_watch);
2348                 ifdata->tcp_listener_watch = 0;
2349
2350                 connman_error("Error with TCP listener channel");
2351
2352                 return FALSE;
2353         }
2354
2355         sk = g_io_channel_unix_get_fd(channel);
2356
2357         client_sk = accept(sk, (void *)&client_addr, &client_addr_len);
2358         if (client_sk < 0) {
2359                 connman_error("Accept failure on TCP listener");
2360                 ifdata->tcp_listener_watch = 0;
2361                 return FALSE;
2362         }
2363
2364         len = recv(client_sk, buf, sizeof(buf), 0);
2365         if (len < 2)
2366                 return TRUE;
2367
2368         DBG("Received %d bytes (id 0x%04x)", len, buf[2] | buf[3] << 8);
2369
2370         err = parse_request(buf + 2, len - 2, query, sizeof(query));
2371         if (err < 0 || (g_slist_length(server_list) == 0)) {
2372                 send_response(client_sk, buf, len, NULL, 0, IPPROTO_TCP);
2373                 return TRUE;
2374         }
2375
2376         req = g_try_new0(struct request_data, 1);
2377         if (req == NULL)
2378                 return TRUE;
2379
2380         memcpy(&req->sa, &client_addr, client_addr_len);
2381         req->sa_len = client_addr_len;
2382         req->client_sk = client_sk;
2383         req->protocol = IPPROTO_TCP;
2384
2385         request_id += 2;
2386         if (request_id == 0x0000 || request_id == 0xffff)
2387                 request_id += 2;
2388
2389         req->srcid = buf[2] | (buf[3] << 8);
2390         req->dstid = request_id;
2391         req->altid = request_id + 1;
2392         req->request_len = len;
2393
2394         buf[2] = req->dstid & 0xff;
2395         buf[3] = req->dstid >> 8;
2396
2397         req->numserv = 0;
2398         req->ifdata = (struct listener_data *) ifdata;
2399         req->append_domain = FALSE;
2400         request_list = g_slist_append(request_list, req);
2401
2402         for (list = server_list; list; list = list->next) {
2403                 struct server_data *data = list->data;
2404                 GList *domains;
2405
2406                 if (data->protocol != IPPROTO_UDP || data->enabled == FALSE)
2407                         continue;
2408
2409                 server = create_server(data->interface, NULL,
2410                                         data->server, IPPROTO_TCP);
2411
2412                 /*
2413                  * If server is NULL, we're not connected yet.
2414                  * Copy the relevant buffers and continue with
2415                  * the next nameserver.
2416                  * The request will actually be sent once we're
2417                  * properly connected over TCP to this nameserver.
2418                  */
2419                 if (server == NULL) {
2420                         req->request = g_try_malloc0(req->request_len);
2421                         if (req->request == NULL)
2422                                 return TRUE;
2423
2424                         memcpy(req->request, buf, req->request_len);
2425
2426                         req->name = g_try_malloc0(sizeof(query));
2427                         if (req->name == NULL) {
2428                                 g_free(req->request);
2429                                 return TRUE;
2430                         }
2431                         memcpy(req->name, query, sizeof(query));
2432
2433                         continue;
2434                 }
2435
2436                 if (req->timeout > 0)
2437                         g_source_remove(req->timeout);
2438
2439                 for (domains = data->domains; domains;
2440                                 domains = domains->next) {
2441                         char *dom = domains->data;
2442
2443                         DBG("Adding domain %s to %s", dom, server->server);
2444
2445                         server->domains = g_list_append(server->domains,
2446                                                 g_strdup(dom));
2447                 }
2448
2449                 req->timeout = g_timeout_add_seconds(30, request_timeout, req);
2450                 if (ns_resolv(server, req, buf, query) > 0) {
2451                         if (req->timeout > 0) {
2452                                 g_source_remove(req->timeout);
2453                                 req->timeout = 0;
2454                         }
2455                 }
2456         }
2457
2458         return TRUE;
2459 }
2460
2461 static gboolean udp_listener_event(GIOChannel *channel, GIOCondition condition,
2462                                                         gpointer user_data)
2463 {
2464         unsigned char buf[768];
2465         char query[512];
2466         struct request_data *req;
2467         struct sockaddr_in6 client_addr;
2468         socklen_t client_addr_len = sizeof(client_addr);
2469         int sk, err, len;
2470         struct listener_data *ifdata = user_data;
2471
2472         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
2473                 connman_error("Error with UDP listener channel");
2474                 ifdata->udp_listener_watch = 0;
2475                 return FALSE;
2476         }
2477
2478         sk = g_io_channel_unix_get_fd(channel);
2479
2480         memset(&client_addr, 0, client_addr_len);
2481         len = recvfrom(sk, buf, sizeof(buf), 0, (void *)&client_addr,
2482                        &client_addr_len);
2483         if (len < 2)
2484                 return TRUE;
2485
2486         DBG("Received %d bytes (id 0x%04x)", len, buf[0] | buf[1] << 8);
2487
2488         err = parse_request(buf, len, query, sizeof(query));
2489         if (err < 0 || (g_slist_length(server_list) == 0)) {
2490                 send_response(sk, buf, len, (void *)&client_addr,
2491                                 client_addr_len, IPPROTO_UDP);
2492                 return TRUE;
2493         }
2494
2495         req = g_try_new0(struct request_data, 1);
2496         if (req == NULL)
2497                 return TRUE;
2498
2499         memcpy(&req->sa, &client_addr, client_addr_len);
2500         req->sa_len = client_addr_len;
2501         req->client_sk = 0;
2502         req->protocol = IPPROTO_UDP;
2503
2504         request_id += 2;
2505         if (request_id == 0x0000 || request_id == 0xffff)
2506                 request_id += 2;
2507
2508         req->srcid = buf[0] | (buf[1] << 8);
2509         req->dstid = request_id;
2510         req->altid = request_id + 1;
2511         req->request_len = len;
2512
2513         buf[0] = req->dstid & 0xff;
2514         buf[1] = req->dstid >> 8;
2515
2516         req->numserv = 0;
2517         req->ifdata = (struct listener_data *) ifdata;
2518         req->timeout = g_timeout_add_seconds(5, request_timeout, req);
2519         req->append_domain = FALSE;
2520         request_list = g_slist_append(request_list, req);
2521
2522         return resolv(req, buf, query);
2523 }
2524
2525 static int create_dns_listener(int protocol, struct listener_data *ifdata)
2526 {
2527         GIOChannel *channel;
2528         const char *proto;
2529         union {
2530                 struct sockaddr sa;
2531                 struct sockaddr_in6 sin6;
2532                 struct sockaddr_in sin;
2533         } s;
2534         socklen_t slen;
2535         int sk, type, v6only = 0;
2536         int family = AF_INET6;
2537
2538
2539         DBG("interface %s", ifdata->ifname);
2540
2541         switch (protocol) {
2542         case IPPROTO_UDP:
2543                 proto = "UDP";
2544                 type = SOCK_DGRAM | SOCK_CLOEXEC;
2545                 break;
2546
2547         case IPPROTO_TCP:
2548                 proto = "TCP";
2549                 type = SOCK_STREAM | SOCK_CLOEXEC;
2550                 break;
2551
2552         default:
2553                 return -EINVAL;
2554         }
2555
2556         sk = socket(family, type, protocol);
2557         if (sk < 0 && family == AF_INET6 && errno == EAFNOSUPPORT) {
2558                 connman_error("No IPv6 support; DNS proxy listening only on Legacy IP");
2559                 family = AF_INET;
2560                 sk = socket(family, type, protocol);
2561         }
2562         if (sk < 0) {
2563                 connman_error("Failed to create %s listener socket", proto);
2564                 return -EIO;
2565         }
2566
2567         if (setsockopt(sk, SOL_SOCKET, SO_BINDTODEVICE,
2568                                         ifdata->ifname,
2569                                         strlen(ifdata->ifname) + 1) < 0) {
2570                 connman_error("Failed to bind %s listener interface", proto);
2571                 close(sk);
2572                 return -EIO;
2573         }
2574         /* Ensure it accepts Legacy IP connections too */
2575         if (family == AF_INET6 &&
2576                         setsockopt(sk, SOL_IPV6, IPV6_V6ONLY,
2577                                         &v6only, sizeof(v6only)) < 0) {
2578                 connman_error("Failed to clear V6ONLY on %s listener socket",
2579                               proto);
2580                 close(sk);
2581                 return -EIO;
2582         }
2583
2584         if (family == AF_INET) {
2585                 memset(&s.sin, 0, sizeof(s.sin));
2586                 s.sin.sin_family = AF_INET;
2587                 s.sin.sin_port = htons(53);
2588                 s.sin.sin_addr.s_addr = htonl(INADDR_ANY);
2589                 slen = sizeof(s.sin);
2590         } else {
2591                 memset(&s.sin6, 0, sizeof(s.sin6));
2592                 s.sin6.sin6_family = AF_INET6;
2593                 s.sin6.sin6_port = htons(53);
2594                 s.sin6.sin6_addr = in6addr_any;
2595                 slen = sizeof(s.sin6);
2596         }
2597
2598         if (bind(sk, &s.sa, slen) < 0) {
2599                 connman_error("Failed to bind %s listener socket", proto);
2600                 close(sk);
2601                 return -EIO;
2602         }
2603
2604         if (protocol == IPPROTO_TCP && listen(sk, 10) < 0) {
2605                 connman_error("Failed to listen on TCP socket");
2606                 close(sk);
2607                 return -EIO;
2608         }
2609
2610         channel = g_io_channel_unix_new(sk);
2611         if (channel == NULL) {
2612                 connman_error("Failed to create %s listener channel", proto);
2613                 close(sk);
2614                 return -EIO;
2615         }
2616
2617         g_io_channel_set_close_on_unref(channel, TRUE);
2618
2619         if (protocol == IPPROTO_TCP) {
2620                 ifdata->tcp_listener_channel = channel;
2621                 ifdata->tcp_listener_watch = g_io_add_watch(channel,
2622                                 G_IO_IN, tcp_listener_event, (gpointer) ifdata);
2623         } else {
2624                 ifdata->udp_listener_channel = channel;
2625                 ifdata->udp_listener_watch = g_io_add_watch(channel,
2626                                 G_IO_IN, udp_listener_event, (gpointer) ifdata);
2627         }
2628
2629         return 0;
2630 }
2631
2632 static void destroy_udp_listener(struct listener_data *ifdata)
2633 {
2634         DBG("interface %s", ifdata->ifname);
2635
2636         if (ifdata->udp_listener_watch > 0)
2637                 g_source_remove(ifdata->udp_listener_watch);
2638
2639         g_io_channel_unref(ifdata->udp_listener_channel);
2640 }
2641
2642 static void destroy_tcp_listener(struct listener_data *ifdata)
2643 {
2644         DBG("interface %s", ifdata->ifname);
2645
2646         if (ifdata->tcp_listener_watch > 0)
2647                 g_source_remove(ifdata->tcp_listener_watch);
2648
2649         g_io_channel_unref(ifdata->tcp_listener_channel);
2650 }
2651
2652 static int create_listener(struct listener_data *ifdata)
2653 {
2654         int err;
2655
2656         err = create_dns_listener(IPPROTO_UDP, ifdata);
2657         if (err < 0)
2658                 return err;
2659
2660         err = create_dns_listener(IPPROTO_TCP, ifdata);
2661         if (err < 0) {
2662                 destroy_udp_listener(ifdata);
2663                 return err;
2664         }
2665
2666         if (g_strcmp0(ifdata->ifname, "lo") == 0)
2667                 __connman_resolvfile_append("lo", NULL, "127.0.0.1");
2668
2669         return 0;
2670 }
2671
2672 static void destroy_request_data(struct request_data *req)
2673 {
2674         if (req->timeout > 0)
2675                 g_source_remove(req->timeout);
2676
2677         g_free(req->resp);
2678         g_free(req->request);
2679         g_free(req->name);
2680         g_free(req);
2681 }
2682
2683 static void destroy_listener(struct listener_data *ifdata)
2684 {
2685         GSList *list;
2686
2687         if (g_strcmp0(ifdata->ifname, "lo") == 0)
2688                 __connman_resolvfile_remove("lo", NULL, "127.0.0.1");
2689
2690         for (list = request_pending_list; list; list = list->next) {
2691                 struct request_data *req = list->data;
2692
2693                 DBG("Dropping pending request (id 0x%04x -> 0x%04x)",
2694                                                 req->srcid, req->dstid);
2695                 destroy_request_data(req);
2696                 list->data = NULL;
2697         }
2698
2699         g_slist_free(request_pending_list);
2700         request_pending_list = NULL;
2701
2702         for (list = request_list; list; list = list->next) {
2703                 struct request_data *req = list->data;
2704
2705                 DBG("Dropping request (id 0x%04x -> 0x%04x)",
2706                                                 req->srcid, req->dstid);
2707                 destroy_request_data(req);
2708                 list->data = NULL;
2709         }
2710
2711         g_slist_free(request_list);
2712         request_list = NULL;
2713
2714         destroy_tcp_listener(ifdata);
2715         destroy_udp_listener(ifdata);
2716 }
2717
2718 int __connman_dnsproxy_add_listener(const char *interface)
2719 {
2720         struct listener_data *ifdata;
2721         int err;
2722
2723         DBG("interface %s", interface);
2724
2725         if (g_hash_table_lookup(listener_table, interface) != NULL)
2726                 return 0;
2727
2728         ifdata = g_try_new0(struct listener_data, 1);
2729         if (ifdata == NULL)
2730                 return -ENOMEM;
2731
2732         ifdata->ifname = g_strdup(interface);
2733         ifdata->udp_listener_channel = NULL;
2734         ifdata->udp_listener_watch = 0;
2735         ifdata->tcp_listener_channel = NULL;
2736         ifdata->tcp_listener_watch = 0;
2737
2738         err = create_listener(ifdata);
2739         if (err < 0) {
2740                 connman_error("Couldn't create listener for %s err %d",
2741                                 interface, err);
2742                 g_free(ifdata->ifname);
2743                 g_free(ifdata);
2744                 return err;
2745         }
2746         g_hash_table_insert(listener_table, ifdata->ifname, ifdata);
2747         return 0;
2748 }
2749
2750 void __connman_dnsproxy_remove_listener(const char *interface)
2751 {
2752         struct listener_data *ifdata;
2753
2754         DBG("interface %s", interface);
2755
2756         ifdata = g_hash_table_lookup(listener_table, interface);
2757         if (ifdata == NULL)
2758                 return;
2759
2760         destroy_listener(ifdata);
2761
2762         g_hash_table_remove(listener_table, interface);
2763 }
2764
2765 static void remove_listener(gpointer key, gpointer value, gpointer user_data)
2766 {
2767         const char *interface = key;
2768         struct listener_data *ifdata = value;
2769
2770         DBG("interface %s", interface);
2771
2772         destroy_listener(ifdata);
2773 }
2774
2775 int __connman_dnsproxy_init(void)
2776 {
2777         int err;
2778
2779         DBG("");
2780
2781         listener_table = g_hash_table_new_full(g_str_hash, g_str_equal,
2782                                                         g_free, g_free);
2783         err = __connman_dnsproxy_add_listener("lo");
2784         if (err < 0)
2785                 return err;
2786
2787         err = connman_notifier_register(&dnsproxy_notifier);
2788         if (err < 0)
2789                 goto destroy;
2790
2791         return 0;
2792
2793 destroy:
2794         __connman_dnsproxy_remove_listener("lo");
2795         g_hash_table_destroy(listener_table);
2796
2797         return err;
2798 }
2799
2800 void __connman_dnsproxy_cleanup(void)
2801 {
2802         DBG("");
2803
2804         connman_notifier_unregister(&dnsproxy_notifier);
2805
2806         g_hash_table_foreach(listener_table, remove_listener, NULL);
2807
2808         g_hash_table_destroy(listener_table);
2809 }