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         int len;
 498
 499         DBG("query %s domain %s", query, domain);
 500
 501         while (query != NULL) {
 502                 const char *tmp;
 503
 504                 tmp = strchr(query, '.');
 505                 if (tmp == NULL) {
 506                         len = strlen(query);
 507                         if (len == 0)
 508                                 break;
 509                         *ptr = len;
 510                         memcpy(ptr + 1, query, len);
 511                         ptr += len + 1;
 512                         break;
 513                 }
 514
 515                 *ptr = tmp - query;
 516                 memcpy(ptr + 1, query, tmp - query);
 517                 ptr += tmp - query + 1;
 518
 519                 query = tmp + 1;
 520         }
 521
 522         while (domain != NULL) {
 523                 const char *tmp;
 524
 525                 tmp = strchr(domain, '.');
 526                 if (tmp == NULL) {
 527                         len = strlen(domain);
 528                         if (len == 0)
 529                                 break;
 530                         *ptr = len;
 531                         memcpy(ptr + 1, domain, len);
 532                         ptr += len + 1;
 533                         break;
 534                 }
 535
 536                 *ptr = tmp - domain;
 537                 memcpy(ptr + 1, domain, tmp - domain);
 538                 ptr += tmp - domain + 1;
 539
 540                 domain = tmp + 1;
 541         }
 542
 543         *ptr++ = 0x00;
 544
 545         return ptr - buf;
 546 }
 547
 548 static gboolean cache_check_is_valid(struct cache_data *data,
 549                                 time_t current_time)
 550 {
 551         if (data == NULL)
 552                 return FALSE;
 553
 554         if (data->cache_until < current_time)
 555                 return FALSE;
 556
 557         return TRUE;
 558 }
 559
 560 /*
 561  * remove stale cached entries so that they can be refreshed
 562  */
 563 static void cache_enforce_validity(struct cache_entry *entry)
 564 {
 565         time_t current_time = time(NULL);
 566
 567         if (cache_check_is_valid(entry->ipv4, current_time) == FALSE
 568                                                         && entry->ipv4) {
 569                 DBG("cache timeout \"%s\" type A", entry->key);
 570                 g_free(entry->ipv4->data);
 571                 g_free(entry->ipv4);
 572                 entry->ipv4 = NULL;
 573
 574         }
 575
 576         if (cache_check_is_valid(entry->ipv6, current_time) == FALSE
 577                                                         && entry->ipv6) {
 578                 DBG("cache timeout \"%s\" type AAAA", entry->key);
 579                 g_free(entry->ipv6->data);
 580                 g_free(entry->ipv6);
 581                 entry->ipv6 = NULL;
 582         }
 583 }
 584
 585 static uint16_t cache_check_validity(char *question, uint16_t type,
 586                                 struct cache_entry *entry)
 587 {
 588         time_t current_time = time(NULL);
 589         int want_refresh = 0;
 590
 591         /*
 592          * if we have a popular entry, we want a refresh instead of
 593          * total destruction of the entry.
 594          */
 595         if (entry->hits > 2)
 596                 want_refresh = 1;
 597
 598         cache_enforce_validity(entry);
 599
 600         switch (type) {
 601         case 1:         /* IPv4 */
 602                 if (cache_check_is_valid(entry->ipv4, current_time) == FALSE) {
 603                         DBG("cache %s \"%s\" type A", entry->ipv4 ?
 604                                         "timeout" : "entry missing", question);
 605
 606                         if (want_refresh)
 607                                 entry->want_refresh = 1;
 608
 609                         /*
 610                          * We do not remove cache entry if there is still
 611                          * valid IPv6 entry found in the cache.
 612                          */
 613                         if (cache_check_is_valid(entry->ipv6, current_time)
 614                                         == FALSE && want_refresh == FALSE) {
 615                                 g_hash_table_remove(cache, question);
 616                                 type = 0;
 617                         }
 618                 }
 619                 break;
 620
 621         case 28:        /* IPv6 */
 622                 if (cache_check_is_valid(entry->ipv6, current_time) == FALSE) {
 623                         DBG("cache %s \"%s\" type AAAA", entry->ipv6 ?
 624                                         "timeout" : "entry missing", question);
 625
 626                         if (want_refresh)
 627                                 entry->want_refresh = 1;
 628
 629                         if (cache_check_is_valid(entry->ipv4, current_time)
 630                                         == FALSE && want_refresh == FALSE) {
 631                                 g_hash_table_remove(cache, question);
 632                                 type = 0;
 633                         }
 634                 }
 635                 break;
 636         }
 637
 638         return type;
 639 }
 640
 641 static struct cache_entry *cache_check(gpointer request, int *qtype)
 642 {
 643         char *question = request + 12;
 644         struct cache_entry *entry;
 645         struct domain_question *q;
 646         uint16_t type;
 647         int offset;
 648
 649         offset = strlen(question) + 1;
 650         q = (void *) (question + offset);
 651         type = ntohs(q->type);
 652
 653         /* We only cache either A (1) or AAAA (28) requests */
 654         if (type != 1 && type != 28)
 655                 return NULL;
 656
 657         entry = g_hash_table_lookup(cache, question);
 658         if (entry == NULL)
 659                 return NULL;
 660
 661         type = cache_check_validity(question, type, entry);
 662         if (type == 0)
 663                 return NULL;
 664
 665         *qtype = type;
 666         return entry;
 667 }
 668
 669 /*
 670  * Get a label/name from DNS resource record. The function decompresses the
 671  * label if necessary. The function does not convert the name to presentation
 672  * form. This means that the result string will contain label lengths instead
 673  * of dots between labels. We intentionally do not want to convert to dotted
 674  * format so that we can cache the wire format string directly.
 675  */
 676 static int get_name(int counter,
 677                 unsigned char *pkt, unsigned char *start, unsigned char *max,
 678                 unsigned char *output, int output_max, int *output_len,
 679                 unsigned char **end, char *name, int *name_len)
 680 {
 681         unsigned char *p;
 682
 683         /* Limit recursion to 10 (this means up to 10 labels in domain name) */
 684         if (counter > 10)
 685                 return -EINVAL;
 686
 687         p = start;
 688         while (*p) {
 689                 if ((*p & NS_CMPRSFLGS) == NS_CMPRSFLGS) {
 690                         uint16_t offset = (*p & 0x3F) * 256 + *(p + 1);
 691
 692                         if (offset >= max - pkt)
 693                                 return -ENOBUFS;
 694
 695                         if (*end == NULL)
 696                                 *end = p + 2;
 697
 698                         return get_name(counter + 1, pkt, pkt + offset, max,
 699                                         output, output_max, output_len, end,
 700                                         name, name_len);
 701                 } else {
 702                         unsigned label_len = *p;
 703
 704                         if (pkt + label_len > max)
 705                                 return -ENOBUFS;
 706
 707                         if (*output_len > output_max)
 708                                 return -ENOBUFS;
 709
 710                         /*
 711                          * We need the original name in order to check
 712                          * if this answer is the correct one.
 713                          */
 714                         name[(*name_len)++] = label_len;
 715                         memcpy(name + *name_len, p + 1, label_len + 1);
 716                         *name_len += label_len;
 717
 718                         /* We compress the result */
 719                         output[0] = NS_CMPRSFLGS;
 720                         output[1] = 0x0C;
 721                         *output_len = 2;
 722
 723                         p += label_len + 1;
 724
 725                         if (*end == NULL)
 726                                 *end = p;
 727
 728                         if (p >= max)
 729                                 return -ENOBUFS;
 730                 }
 731         }
 732
 733         return 0;
 734 }
 735
 736 static int parse_rr(unsigned char *buf, unsigned char *start,
 737                         unsigned char *max,
 738                         unsigned char *response, unsigned int *response_size,
 739                         uint16_t *type, uint16_t *class, int *ttl, int *rdlen,
 740                         unsigned char **end,
 741                         char *name)
 742 {
 743         struct domain_rr *rr;
 744         int err, offset;
 745         int name_len = 0, output_len = 0, max_rsp = *response_size;
 746
 747         err = get_name(0, buf, start, max, response, max_rsp,
 748                 &output_len, end, name, &name_len);
 749         if (err < 0)
 750                 return err;
 751
 752         offset = output_len;
 753
 754         if ((unsigned int) offset > *response_size)
 755                 return -ENOBUFS;
 756
 757         rr = (void *) (*end);
 758
 759         if (rr == NULL)
 760                 return -EINVAL;
 761
 762         *type = ntohs(rr->type);
 763         *class = ntohs(rr->class);
 764         *ttl = ntohl(rr->ttl);
 765         *rdlen = ntohs(rr->rdlen);
 766
 767         if (*ttl < 0)
 768                 return -EINVAL;
 769
 770         memcpy(response + offset, *end, sizeof(struct domain_rr));
 771
 772         offset += sizeof(struct domain_rr);
 773         *end += sizeof(struct domain_rr);
 774
 775         if ((unsigned int) (offset + *rdlen) > *response_size)
 776                 return -ENOBUFS;
 777
 778         memcpy(response + offset, *end, *rdlen);
 779
 780         *end += *rdlen;
 781
 782         *response_size = offset + *rdlen;
 783
 784         return 0;
 785 }
 786
 787 static gboolean check_alias(GSList *aliases, char *name)
 788 {
 789         GSList *list;
 790
 791         if (aliases != NULL) {
 792                 for (list = aliases; list; list = list->next) {
 793                         int len = strlen((char *)list->data);
 794                         if (strncmp((char *)list->data, name, len) == 0)
 795                                 return TRUE;
 796                 }
 797         }
 798
 799         return FALSE;
 800 }
 801
 802 static int parse_response(unsigned char *buf, int buflen,
 803                         char *question, int qlen,
 804                         uint16_t *type, uint16_t *class, int *ttl,
 805                         unsigned char *response, unsigned int *response_len,
 806                         uint16_t *answers)
 807 {
 808         struct domain_hdr *hdr = (void *) buf;
 809         struct domain_question *q;
 810         unsigned char *ptr;
 811         uint16_t qdcount = ntohs(hdr->qdcount);
 812         uint16_t ancount = ntohs(hdr->ancount);
 813         int err, i;
 814         uint16_t qtype, qclass;
 815         unsigned char *next = NULL;
 816         unsigned int maxlen = *response_len;
 817         GSList *aliases = NULL, *list;
 818         char name[NS_MAXDNAME + 1];
 819
 820         if (buflen < 12)
 821                 return -EINVAL;
 822
 823         DBG("qr %d qdcount %d", hdr->qr, qdcount);
 824
 825         /* We currently only cache responses where question count is 1 */
 826         if (hdr->qr != 1 || qdcount != 1)
 827                 return -EINVAL;
 828
 829         ptr = buf + sizeof(struct domain_hdr);
 830
 831         strncpy(question, (char *) ptr, qlen);
 832         qlen = strlen(question);
 833         ptr += qlen + 1; /* skip \0 */
 834
 835         q = (void *) ptr;
 836         qtype = ntohs(q->type);
 837
 838         /* We cache only A and AAAA records */
 839         if (qtype != 1 && qtype != 28)
 840                 return -ENOMSG;
 841
 842         qclass = ntohs(q->class);
 843
 844         ptr += 2 + 2; /* ptr points now to answers */
 845
 846         err = -ENOMSG;
 847         *response_len = 0;
 848         *answers = 0;
 849
 850         /*
 851          * We have a bunch of answers (like A, AAAA, CNAME etc) to
 852          * A or AAAA question. We traverse the answers and parse the
 853          * resource records. Only A and AAAA records are cached, all
 854          * the other records in answers are skipped.
 855          */
 856         for (i = 0; i < ancount; i++) {
 857                 /*
 858                  * Get one address at a time to this buffer.
 859                  * The max size of the answer is
 860                  *   2 (pointer) + 2 (type) + 2 (class) +
 861                  *   4 (ttl) + 2 (rdlen) + addr (16 or 4) = 28
 862                  * for A or AAAA record.
 863                  * For CNAME the size can be bigger.
 864                  */
 865                 unsigned char rsp[NS_MAXCDNAME];
 866                 unsigned int rsp_len = sizeof(rsp) - 1;
 867                 int ret, rdlen;
 868
 869                 memset(rsp, 0, sizeof(rsp));
 870
 871                 ret = parse_rr(buf, ptr, buf + buflen, rsp, &rsp_len,
 872                         type, class, ttl, &rdlen, &next, name);
 873                 if (ret != 0) {
 874                         err = ret;
 875                         goto out;
 876                 }
 877
 878                 /*
 879                  * Now rsp contains compressed or uncompressed resource
 880                  * record. Next we check if this record answers the question.
 881                  * The name var contains the uncompressed label.
 882                  * One tricky bit is the CNAME records as they alias
 883                  * the name we might be interested in.
 884                  */
 885
 886                 /*
 887                  * Go to next answer if the class is not the one we are
 888                  * looking for.
 889                  */
 890                 if (*class != qclass) {
 891                         ptr = next;
 892                         next = NULL;
 893                         continue;
 894                 }
 895
 896                 /*
 897                  * Try to resolve aliases also, type is CNAME(5).
 898                  * This is important as otherwise the aliased names would not
 899                  * be cached at all as the cache would not contain the aliased
 900                  * question.
 901                  *
 902                  * If any CNAME is found in DNS packet, then we cache the alias
 903                  * IP address instead of the question (as the server
 904                  * said that question has only an alias).
 905                  * This means in practice that if e.g., ipv6.google.com is
 906                  * queried, DNS server returns CNAME of that name which is
 907                  * ipv6.l.google.com. We then cache the address of the CNAME
 908                  * but return the question name to client. So the alias
 909                  * status of the name is not saved in cache and thus not
 910                  * returned to the client. We do not return DNS packets from
 911                  * cache to client saying that ipv6.google.com is an alias to
 912                  * ipv6.l.google.com but we return instead a DNS packet that
 913                  * says ipv6.google.com has address xxx which is in fact the
 914                  * address of ipv6.l.google.com. For caching purposes this
 915                  * should not cause any issues.
 916                  */
 917                 if (*type == 5 && strncmp(question, name, qlen) == 0) {
 918                         /*
 919                          * So now the alias answered the question. This is
 920                          * not very useful from caching point of view as
 921                          * the following A or AAAA records will not match the
 922                          * question. We need to find the real A/AAAA record
 923                          * of the alias and cache that.
 924                          */
 925                         unsigned char *end = NULL;
 926                         int name_len = 0, output_len;
 927
 928                         memset(rsp, 0, sizeof(rsp));
 929                         rsp_len = sizeof(rsp) - 1;
 930
 931                         /*
 932                          * Alias is in rdata part of the message,
 933                          * and next-rdlen points to it. So we need to get
 934                          * the real name of the alias.
 935                          */
 936                         ret = get_name(0, buf, next - rdlen, buf + buflen,
 937                                         rsp, rsp_len, &output_len, &end,
 938                                         name, &name_len);
 939                         if (ret != 0) {
 940                                 /* just ignore the error at this point */
 941                                 ptr = next;
 942                                 next = NULL;
 943                                 continue;
 944                         }
 945
 946                         /*
 947                          * We should now have the alias of the entry we might
 948                          * want to cache. Just remember it for a while.
 949                          * We check the alias list when we have parsed the
 950                          * A or AAAA record.
 951                          */
 952                         aliases = g_slist_prepend(aliases, g_strdup(name));
 953
 954                         ptr = next;
 955                         next = NULL;
 956                         continue;
 957                 }
 958
 959                 if (*type == qtype) {
 960                         /*
 961                          * We found correct type (A or AAAA)
 962                          */
 963                         if (check_alias(aliases, name) == TRUE ||
 964                                 (aliases == NULL && strncmp(question, name,
 965                                                         qlen) == 0)) {
 966                                 /*
 967                                  * We found an alias or the name of the rr
 968                                  * matches the question. If so, we append
 969                                  * the compressed label to the cache.
 970                                  * The end result is a response buffer that
 971                                  * will contain one or more cached and
 972                                  * compressed resource records.
 973                                  */
 974                                 if (*response_len + rsp_len > maxlen) {
 975                                         err = -ENOBUFS;
 976                                         goto out;
 977                                 }
 978                                 memcpy(response + *response_len, rsp, rsp_len);
 979                                 *response_len += rsp_len;
 980                                 (*answers)++;
 981                                 err = 0;
 982                         }
 983                 }
 984
 985                 ptr = next;
 986                 next = NULL;
 987         }
 988
 989 out:
 990         for (list = aliases; list; list = list->next)
 991                 g_free(list->data);
 992         g_slist_free(aliases);
 993
 994         return err;
 995 }
 996
 997 struct cache_timeout {
 998         time_t current_time;
 999         int max_timeout;
1000         int try_harder;
1001 };
1002
1003 static gboolean cache_check_entry(gpointer key, gpointer value,
1004                                         gpointer user_data)
1005 {
1006         struct cache_timeout *data = user_data;
1007         struct cache_entry *entry = value;
1008         int max_timeout;
1009
1010         /* Scale the number of hits by half as part of cache aging */
1011
1012         entry->hits /= 2;
1013
1014         /*
1015          * If either IPv4 or IPv6 cached entry has expired, we
1016          * remove both from the cache.
1017          */
1018
1019         if (entry->ipv4 != NULL && entry->ipv4->timeout > 0) {
1020                 max_timeout = entry->ipv4->cache_until;
1021                 if (max_timeout > data->max_timeout)
1022                         data->max_timeout = max_timeout;
1023
1024                 if (entry->ipv4->cache_until < data->current_time)
1025                         return TRUE;
1026         }
1027
1028         if (entry->ipv6 != NULL && entry->ipv6->timeout > 0) {
1029                 max_timeout = entry->ipv6->cache_until;
1030                 if (max_timeout > data->max_timeout)
1031                         data->max_timeout = max_timeout;
1032
1033                 if (entry->ipv6->cache_until < data->current_time)
1034                         return TRUE;
1035         }
1036
1037         /*
1038          * if we're asked to try harder, also remove entries that have
1039          * few hits
1040          */
1041         if (data->try_harder && entry->hits < 4)
1042                 return TRUE;
1043
1044         return FALSE;
1045 }
1046
1047 static void cache_cleanup(void)
1048 {
1049         static int max_timeout;
1050         struct cache_timeout data;
1051         int count = 0;
1052
1053         data.current_time = time(NULL);
1054         data.max_timeout = 0;
1055         data.try_harder = 0;
1056
1057         /*
1058          * In the first pass, we only remove entries that have timed out.
1059          * We use a cache of the first time to expire to do this only
1060          * when it makes sense.
1061          */
1062         if (max_timeout <= data.current_time) {
1063                 count = g_hash_table_foreach_remove(cache, cache_check_entry,
1064                                                 &data);
1065         }
1066         DBG("removed %d in the first pass", count);
1067
1068         /*
1069          * In the second pass, if the first pass turned up blank,
1070          * we also expire entries with a low hit count,
1071          * while aging the hit count at the same time.
1072          */
1073         data.try_harder = 1;
1074         if (count == 0)
1075                 count = g_hash_table_foreach_remove(cache, cache_check_entry,
1076                                                 &data);
1077
1078         if (count == 0)
1079                 /*
1080                  * If we could not remove anything, then remember
1081                  * what is the max timeout and do nothing if we
1082                  * have not yet reached it. This will prevent
1083                  * constant traversal of the cache if it is full.
1084                  */
1085                 max_timeout = data.max_timeout;
1086         else
1087                 max_timeout = 0;
1088 }
1089
1090 static gboolean cache_invalidate_entry(gpointer key, gpointer value,
1091                                         gpointer user_data)
1092 {
1093         struct cache_entry *entry = value;
1094
1095         /* first, delete any expired elements */
1096         cache_enforce_validity(entry);
1097
1098         /* if anything is not expired, mark the entry for refresh */
1099         if (entry->hits > 0 && (entry->ipv4 || entry->ipv6))
1100                 entry->want_refresh = 1;
1101
1102         /* delete the cached data */
1103         if (entry->ipv4) {
1104                 g_free(entry->ipv4->data);
1105                 g_free(entry->ipv4);
1106                 entry->ipv4 = NULL;
1107         }
1108
1109         if (entry->ipv6) {
1110                 g_free(entry->ipv6->data);
1111                 g_free(entry->ipv6);
1112                 entry->ipv6 = NULL;
1113         }
1114
1115         /* keep the entry if we want it refreshed, delete it otherwise */
1116         if (entry->want_refresh)
1117                 return FALSE;
1118         else
1119                 return TRUE;
1120 }
1121
1122 /*
1123  * cache_invalidate is called from places where the DNS landscape
1124  * has changed, say because connections are added or we entered a VPN.
1125  * The logic is to wipe all cache data, but mark all non-expired
1126  * parts of the cache for refresh rather than deleting the whole cache.
1127  */
1128 static void cache_invalidate(void)
1129 {
1130         DBG("Invalidating the DNS cache %p", cache);
1131
1132         if (cache == NULL)
1133                 return;
1134
1135         g_hash_table_foreach_remove(cache, cache_invalidate_entry, NULL);
1136 }
1137
1138 static void cache_refresh_entry(struct cache_entry *entry)
1139 {
1140
1141         cache_enforce_validity(entry);
1142
1143         if (entry->hits > 2 && entry->ipv4 == NULL)
1144                 entry->want_refresh = 1;
1145         if (entry->hits > 2 && entry->ipv6 == NULL)
1146                 entry->want_refresh = 1;
1147
1148         if (entry->want_refresh) {
1149                 char *c;
1150                 char dns_name[NS_MAXDNAME + 1];
1151                 entry->want_refresh = 0;
1152
1153                 /* turn a DNS name into a hostname with dots */
1154                 strncpy(dns_name, entry->key, NS_MAXDNAME);
1155                 c = dns_name;
1156                 while (c && *c) {
1157                         int jump;
1158                         jump = *c;
1159                         *c = '.';
1160                         c += jump + 1;
1161                 }
1162                 DBG("Refreshing %s\n", dns_name);
1163                 /* then refresh the hostname */
1164                 refresh_dns_entry(entry, &dns_name[1]);
1165         }
1166 }
1167
1168 static void cache_refresh_iterator(gpointer key, gpointer value,
1169                                         gpointer user_data)
1170 {
1171         struct cache_entry *entry = value;
1172
1173         cache_refresh_entry(entry);
1174 }
1175
1176 static void cache_refresh(void)
1177 {
1178         if (cache == NULL)
1179                 return;
1180
1181         g_hash_table_foreach(cache, cache_refresh_iterator, NULL);
1182 }
1183
1184 static int reply_query_type(unsigned char *msg, int len)
1185 {
1186         unsigned char *c;
1187         uint16_t *w;
1188         int l;
1189         int type;
1190
1191         /* skip the header */
1192         c = msg + sizeof(struct domain_hdr);
1193         len -= sizeof(struct domain_hdr);
1194
1195         if (len < 0)
1196                 return 0;
1197
1198         /* now the query, which is a name and 2 16 bit words */
1199         l = dns_name_length(c) + 1;
1200         c += l;
1201         w = (uint16_t *) c;
1202         type = ntohs(*w);
1203
1204         return type;
1205 }
1206
1207 static int cache_update(struct server_data *srv, unsigned char *msg,
1208                         unsigned int msg_len)
1209 {
1210         int offset = protocol_offset(srv->protocol);
1211         int err, qlen, ttl = 0;
1212         uint16_t answers = 0, type = 0, class = 0;
1213         struct domain_question *q;
1214         struct cache_entry *entry;
1215         struct cache_data *data;
1216         char question[NS_MAXDNAME + 1];
1217         unsigned char response[NS_MAXDNAME + 1];
1218         unsigned char *ptr;
1219         unsigned int rsplen;
1220         gboolean new_entry = TRUE;
1221         time_t current_time;
1222
1223         if (cache_size >= MAX_CACHE_SIZE) {
1224                 cache_cleanup();
1225                 if (cache_size >= MAX_CACHE_SIZE)
1226                         return 0;
1227         }
1228
1229         current_time = time(NULL);
1230
1231         /* don't do a cache refresh more than twice a minute */
1232         if (next_refresh < current_time) {
1233                 cache_refresh();
1234                 next_refresh = current_time + 30;
1235         }
1236
1237
1238         /* Continue only if response code is 0 (=ok) */
1239         if (msg[3] & 0x0f)
1240                 return 0;
1241
1242         if (offset < 0)
1243                 return 0;
1244
1245         rsplen = sizeof(response) - 1;
1246         question[sizeof(question) - 1] = '\0';
1247
1248         err = parse_response(msg + offset, msg_len - offset,
1249                                 question, sizeof(question) - 1,
1250                                 &type, &class, &ttl,
1251                                 response, &rsplen, &answers);
1252
1253         /*
1254          * special case: if we do a ipv6 lookup and get no result
1255          * for a record that's already in our ipv4 cache.. we want
1256          * to cache the negative response.
1257          */
1258         if ((err == -ENOMSG || err == -ENOBUFS) &&
1259                         reply_query_type(msg, msg_len) == 28) {
1260                 entry = g_hash_table_lookup(cache, question);
1261                 if (entry && entry->ipv4 && entry->ipv6 == NULL) {
1262                         data = g_try_new(struct cache_data, 1);
1263                         if (data == NULL)
1264                                 return -ENOMEM;
1265                         data->inserted = entry->ipv4->inserted;
1266                         data->type = type;
1267                         data->answers = msg[5];
1268                         data->timeout = entry->ipv4->timeout;
1269                         data->data_len = msg_len;
1270                         data->data = ptr = g_malloc(msg_len);
1271                         data->valid_until = entry->ipv4->valid_until;
1272                         data->cache_until = entry->ipv4->cache_until;
1273                         memcpy(data->data, msg, msg_len);
1274                         entry->ipv6 = data;
1275                         /*
1276                          * we will get a "hit" when we serve the response
1277                          * out of the cache
1278                          */
1279                         entry->hits--;
1280                         if (entry->hits < 0)
1281                                 entry->hits = 0;
1282                         return 0;
1283                 }
1284         }
1285
1286         if (err < 0 || ttl == 0)
1287                 return 0;
1288
1289         qlen = strlen(question);
1290
1291         /*
1292          * If the cache contains already data, check if the
1293          * type of the cached data is the same and do not add
1294          * to cache if data is already there.
1295          * This is needed so that we can cache both A and AAAA
1296          * records for the same name.
1297          */
1298         entry = g_hash_table_lookup(cache, question);
1299         if (entry == NULL) {
1300                 entry = g_try_new(struct cache_entry, 1);
1301                 if (entry == NULL)
1302                         return -ENOMEM;
1303
1304                 data = g_try_new(struct cache_data, 1);
1305                 if (data == NULL) {
1306                         g_free(entry);
1307                         return -ENOMEM;
1308                 }
1309
1310                 entry->key = g_strdup(question);
1311                 entry->ipv4 = entry->ipv6 = NULL;
1312                 entry->want_refresh = 0;
1313                 entry->hits = 0;
1314
1315                 if (type == 1)
1316                         entry->ipv4 = data;
1317                 else
1318                         entry->ipv6 = data;
1319         } else {
1320                 if (type == 1 && entry->ipv4 != NULL)
1321                         return 0;
1322
1323                 if (type == 28 && entry->ipv6 != NULL)
1324                         return 0;
1325
1326                 data = g_try_new(struct cache_data, 1);
1327                 if (data == NULL)
1328                         return -ENOMEM;
1329
1330                 if (type == 1)
1331                         entry->ipv4 = data;
1332                 else
1333                         entry->ipv6 = data;
1334
1335                 /*
1336                  * compensate for the hit we'll get for serving
1337                  * the response out of the cache
1338                  */
1339                 entry->hits--;
1340                 if (entry->hits < 0)
1341                         entry->hits = 0;
1342
1343                 new_entry = FALSE;
1344         }
1345
1346         if (ttl < MIN_CACHE_TTL)
1347                 ttl = MIN_CACHE_TTL;
1348
1349         data->inserted = current_time;
1350         data->type = type;
1351         data->answers = answers;
1352         data->timeout = ttl;
1353         data->data_len = 12 + qlen + 1 + 2 + 2 + rsplen;
1354         data->data = ptr = g_malloc(data->data_len);
1355         data->valid_until = current_time + ttl;
1356
1357         /*
1358          * Restrict the cached DNS record TTL to some sane value
1359          * in order to prevent data staying in the cache too long.
1360          */
1361         if (ttl > MAX_CACHE_TTL)
1362                 ttl = MAX_CACHE_TTL;
1363
1364         data->cache_until = round_down_ttl(current_time + ttl, ttl);
1365
1366         if (data->data == NULL) {
1367                 g_free(entry->key);
1368                 g_free(data);
1369                 g_free(entry);
1370                 return -ENOMEM;
1371         }
1372
1373         memcpy(ptr, msg, 12);
1374         memcpy(ptr + 12, question, qlen + 1); /* copy also the \0 */
1375
1376         q = (void *) (ptr + 12 + qlen + 1);
1377         q->type = htons(type);
1378         q->class = htons(class);
1379         memcpy(ptr + 12 + qlen + 1 + sizeof(struct domain_question),
1380                 response, rsplen);
1381
1382         if (new_entry == TRUE) {
1383                 g_hash_table_replace(cache, entry->key, entry);
1384                 cache_size++;
1385         }
1386
1387         DBG("cache %d %squestion \"%s\" type %d ttl %d size %zd",
1388                 cache_size, new_entry ? "new " : "old ",
1389                 question, type, ttl,
1390                 sizeof(*entry) + sizeof(*data) + data->data_len + qlen);
1391
1392         return 0;
1393 }
1394
1395 static int ns_resolv(struct server_data *server, struct request_data *req,
1396                                 gpointer request, gpointer name)
1397 {
1398         GList *list;
1399         int sk, err, type = 0;
1400         char *dot, *lookup = (char *) name;
1401         struct cache_entry *entry;
1402
1403         entry = cache_check(request, &type);
1404         if (entry != NULL) {
1405                 int ttl_left = 0;
1406                 struct cache_data *data;
1407
1408                 DBG("cache hit %s type %s", lookup, type == 1 ? "A" : "AAAA");
1409                 if (type == 1)
1410                         data = entry->ipv4;
1411                 else
1412                         data = entry->ipv6;
1413
1414                 if (data) {
1415                         ttl_left = data->valid_until - time(NULL);
1416                         entry->hits++;
1417                 }
1418
1419                 if (data != NULL && req->protocol == IPPROTO_TCP) {
1420                         send_cached_response(req->client_sk, data->data,
1421                                         data->data_len, NULL, 0, IPPROTO_TCP,
1422                                         req->srcid, data->answers, ttl_left);
1423                         return 1;
1424                 }
1425
1426                 if (data != NULL && req->protocol == IPPROTO_UDP) {
1427                         int sk;
1428                         sk = g_io_channel_unix_get_fd(
1429                                         req->ifdata->udp_listener_channel);
1430
1431                         send_cached_response(sk, data->data,
1432                                 data->data_len, &req->sa, req->sa_len,
1433                                 IPPROTO_UDP, req->srcid, data->answers,
1434                                 ttl_left);
1435                         return 1;
1436                 }
1437         }
1438
1439         sk = g_io_channel_unix_get_fd(server->channel);
1440
1441         err = send(sk, request, req->request_len, 0);
1442         if (err < 0)
1443                 return -EIO;
1444
1445         req->numserv++;
1446
1447         /* If we have more than one dot, we don't add domains */
1448         dot = strchr(lookup, '.');
1449         if (dot != NULL && dot != lookup + strlen(lookup) - 1)
1450                 return 0;
1451
1452         if (server->domains != NULL && server->domains->data != NULL)
1453                 req->append_domain = TRUE;
1454
1455         for (list = server->domains; list; list = list->next) {
1456                 char *domain;
1457                 unsigned char alt[1024];
1458                 struct domain_hdr *hdr = (void *) &alt;
1459                 int altlen, domlen, offset;
1460
1461                 domain = list->data;
1462
1463                 if (domain == NULL)
1464                         continue;
1465
1466                 offset = protocol_offset(server->protocol);
1467                 if (offset < 0)
1468                         return offset;
1469
1470                 domlen = strlen(domain) + 1;
1471                 if (domlen < 5)
1472                         return -EINVAL;
1473
1474                 alt[offset] = req->altid & 0xff;
1475                 alt[offset + 1] = req->altid >> 8;
1476
1477                 memcpy(alt + offset + 2, request + offset + 2, 10);
1478                 hdr->qdcount = htons(1);
1479
1480                 altlen = append_query(alt + offset + 12, sizeof(alt) - 12,
1481                                         name, domain);
1482                 if (altlen < 0)
1483                         return -EINVAL;
1484
1485                 altlen += 12;
1486
1487                 memcpy(alt + offset + altlen,
1488                         request + offset + altlen - domlen,
1489                                 req->request_len - altlen - offset + domlen);
1490
1491                 if (server->protocol == IPPROTO_TCP) {
1492                         int req_len = req->request_len + domlen - 2;
1493
1494                         alt[0] = (req_len >> 8) & 0xff;
1495                         alt[1] = req_len & 0xff;
1496                 }
1497
1498                 err = send(sk, alt, req->request_len + domlen, 0);
1499                 if (err < 0)
1500                         return -EIO;
1501
1502                 req->numserv++;
1503         }
1504
1505         return 0;
1506 }
1507
1508 static int forward_dns_reply(unsigned char *reply, int reply_len, int protocol,
1509                                 struct server_data *data)
1510 {
1511         struct domain_hdr *hdr;
1512         struct request_data *req;
1513         int dns_id, sk, err, offset = protocol_offset(protocol);
1514         struct listener_data *ifdata;
1515
1516         if (offset < 0)
1517                 return offset;
1518
1519         hdr = (void *)(reply + offset);
1520         dns_id = reply[offset] | reply[offset + 1] << 8;
1521
1522         DBG("Received %d bytes (id 0x%04x)", reply_len, dns_id);
1523
1524         req = find_request(dns_id);
1525         if (req == NULL)
1526                 return -EINVAL;
1527
1528         DBG("id 0x%04x rcode %d", hdr->id, hdr->rcode);
1529
1530         ifdata = req->ifdata;
1531
1532         reply[offset] = req->srcid & 0xff;
1533         reply[offset + 1] = req->srcid >> 8;
1534
1535         req->numresp++;
1536
1537         if (hdr->rcode == 0 || req->resp == NULL) {
1538
1539                 /*
1540                  * If the domain name was append
1541                  * remove it before forwarding the reply.
1542                  */
1543                 if (req->append_domain == TRUE) {
1544                         unsigned char *ptr;
1545                         uint8_t host_len;
1546                         unsigned int domain_len;
1547
1548                         /*
1549                          * ptr points to the first char of the hostname.
1550                          * ->hostname.domain.net
1551                          */
1552                         ptr = reply + offset + sizeof(struct domain_hdr);
1553                         host_len = *ptr;
1554                         domain_len = strlen((const char *)ptr + host_len + 1);
1555
1556                         /*
1557                          * Remove the domain name and replace it by the end
1558                          * of reply. Check if the domain is really there
1559                          * before trying to copy the data. The domain_len can
1560                          * be 0 because if the original query did not contain
1561                          * a domain name, then we are sending two packets,
1562                          * first without the domain name and the second packet
1563                          * with domain name. The append_domain is set to true
1564                          * even if we sent the first packet without domain
1565                          * name. In this case we end up in this branch.
1566                          */
1567                         if (domain_len > 0) {
1568                                 memcpy(ptr + host_len + 1,
1569                                         ptr + host_len + domain_len + 1,
1570                                         reply_len - (ptr - reply + domain_len));
1571
1572                                 reply_len = reply_len - domain_len;
1573                         }
1574                 }
1575
1576                 g_free(req->resp);
1577                 req->resplen = 0;
1578
1579                 req->resp = g_try_malloc(reply_len);
1580                 if (req->resp == NULL)
1581                         return -ENOMEM;
1582
1583                 memcpy(req->resp, reply, reply_len);
1584                 req->resplen = reply_len;
1585
1586                 cache_update(data, reply, reply_len);
1587         }
1588
1589         if (hdr->rcode > 0 && req->numresp < req->numserv)
1590                 return -EINVAL;
1591
1592         if (req->timeout > 0)
1593                 g_source_remove(req->timeout);
1594
1595         request_list = g_slist_remove(request_list, req);
1596
1597         if (protocol == IPPROTO_UDP) {
1598                 sk = g_io_channel_unix_get_fd(ifdata->udp_listener_channel);
1599                 err = sendto(sk, req->resp, req->resplen, 0,
1600                              &req->sa, req->sa_len);
1601         } else {
1602                 sk = req->client_sk;
1603                 err = send(sk, req->resp, req->resplen, 0);
1604                 close(sk);
1605         }
1606
1607         g_free(req->resp);
1608         g_free(req);
1609
1610         return err;
1611 }
1612
1613 static void cache_element_destroy(gpointer value)
1614 {
1615         struct cache_entry *entry = value;
1616
1617         if (entry == NULL)
1618                 return;
1619
1620         if (entry->ipv4 != NULL) {
1621                 g_free(entry->ipv4->data);
1622                 g_free(entry->ipv4);
1623         }
1624
1625         if (entry->ipv6 != NULL) {
1626                 g_free(entry->ipv6->data);
1627                 g_free(entry->ipv6);
1628         }
1629
1630         g_free(entry->key);
1631         g_free(entry);
1632
1633         if (--cache_size < 0)
1634                 cache_size = 0;
1635 }
1636
1637 static gboolean try_remove_cache(gpointer user_data)
1638 {
1639         if (__sync_fetch_and_sub(&cache_refcount, 1) == 1) {
1640                 DBG("No cache users, removing it.");
1641
1642                 g_hash_table_destroy(cache);
1643                 cache = NULL;
1644         }
1645
1646         return FALSE;
1647 }
1648
1649 static void destroy_server(struct server_data *server)
1650 {
1651         GList *list;
1652
1653         DBG("interface %s server %s", server->interface, server->server);
1654
1655         server_list = g_slist_remove(server_list, server);
1656
1657         if (server->watch > 0)
1658                 g_source_remove(server->watch);
1659
1660         if (server->timeout > 0)
1661                 g_source_remove(server->timeout);
1662
1663         g_io_channel_unref(server->channel);
1664
1665         if (server->protocol == IPPROTO_UDP)
1666                 connman_info("Removing DNS server %s", server->server);
1667
1668         g_free(server->incoming_reply);
1669         g_free(server->server);
1670         for (list = server->domains; list; list = list->next) {
1671                 char *domain = list->data;
1672
1673                 server->domains = g_list_remove(server->domains, domain);
1674                 g_free(domain);
1675         }
1676         g_free(server->interface);
1677
1678         /*
1679          * We do not remove cache right away but delay it few seconds.
1680          * The idea is that when IPv6 DNS server is added via RDNSS, it has a
1681          * lifetime. When the lifetime expires we decrease the refcount so it
1682          * is possible that the cache is then removed. Because a new DNS server
1683          * is usually created almost immediately we would then loose the cache
1684          * without any good reason. The small delay allows the new RDNSS to
1685          * create a new DNS server instance and the refcount does not go to 0.
1686          */
1687         g_timeout_add_seconds(3, try_remove_cache, NULL);
1688
1689         g_free(server);
1690 }
1691
1692 static gboolean udp_server_event(GIOChannel *channel, GIOCondition condition,
1693                                                         gpointer user_data)
1694 {
1695         unsigned char buf[4096];
1696         int sk, err, len;
1697         struct server_data *data = user_data;
1698
1699         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
1700                 connman_error("Error with UDP server %s", data->server);
1701                 data->watch = 0;
1702                 return FALSE;
1703         }
1704
1705         sk = g_io_channel_unix_get_fd(channel);
1706
1707         len = recv(sk, buf, sizeof(buf), 0);
1708         if (len < 12)
1709                 return TRUE;
1710
1711         err = forward_dns_reply(buf, len, IPPROTO_UDP, data);
1712         if (err < 0)
1713                 return TRUE;
1714
1715         return TRUE;
1716 }
1717
1718 static gboolean tcp_server_event(GIOChannel *channel, GIOCondition condition,
1719                                                         gpointer user_data)
1720 {
1721         int sk;
1722         struct server_data *server = user_data;
1723
1724         sk = g_io_channel_unix_get_fd(channel);
1725         if (sk == 0)
1726                 return FALSE;
1727
1728         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
1729                 GSList *list;
1730 hangup:
1731                 DBG("TCP server channel closed");
1732
1733                 /*
1734                  * Discard any partial response which is buffered; better
1735                  * to get a proper response from a working server.
1736                  */
1737                 g_free(server->incoming_reply);
1738                 server->incoming_reply = NULL;
1739
1740                 for (list = request_list; list; list = list->next) {
1741                         struct request_data *req = list->data;
1742                         struct domain_hdr *hdr;
1743
1744                         if (req->protocol == IPPROTO_UDP)
1745                                 continue;
1746
1747                         if (req->request == NULL)
1748                                 continue;
1749
1750                         /*
1751                          * If we're not waiting for any further response
1752                          * from another name server, then we send an error
1753                          * response to the client.
1754                          */
1755                         if (req->numserv && --(req->numserv))
1756                                 continue;
1757
1758                         hdr = (void *) (req->request + 2);
1759                         hdr->id = req->srcid;
1760                         send_response(req->client_sk, req->request,
1761                                 req->request_len, NULL, 0, IPPROTO_TCP);
1762
1763                         request_list = g_slist_remove(request_list, req);
1764                 }
1765
1766                 destroy_server(server);
1767
1768                 return FALSE;
1769         }
1770
1771         if ((condition & G_IO_OUT) && !server->connected) {
1772                 GSList *list;
1773                 GList *domains;
1774                 struct server_data *udp_server;
1775
1776                 udp_server = find_server(server->interface, server->server,
1777                                                                 IPPROTO_UDP);
1778                 if (udp_server != NULL) {
1779                         for (domains = udp_server->domains; domains;
1780                                                 domains = domains->next) {
1781                                 char *dom = domains->data;
1782
1783                                 DBG("Adding domain %s to %s",
1784                                                 dom, server->server);
1785
1786                                 server->domains = g_list_append(server->domains,
1787                                                                 g_strdup(dom));
1788                         }
1789                 }
1790
1791                 server->connected = TRUE;
1792                 server_list = g_slist_append(server_list, server);
1793
1794                 if (server->timeout > 0) {
1795                         g_source_remove(server->timeout);
1796                         server->timeout = 0;
1797                 }
1798
1799                 for (list = request_list; list; list = list->next) {
1800                         struct request_data *req = list->data;
1801
1802                         if (req->protocol == IPPROTO_UDP)
1803                                 continue;
1804
1805                         DBG("Sending req %s over TCP", (char *)req->name);
1806
1807                         if (req->timeout > 0)
1808                                 g_source_remove(req->timeout);
1809
1810                         req->timeout = g_timeout_add_seconds(30,
1811                                                 request_timeout, req);
1812                         if (ns_resolv(server, req, req->request,
1813                                         req->name) > 0) {
1814                                 /* We sent cached result so no need for timeout
1815                                  * handler.
1816                                  */
1817                                 if (req->timeout > 0) {
1818                                         g_source_remove(req->timeout);
1819                                         req->timeout = 0;
1820                                 }
1821                         }
1822                 }
1823
1824         } else if (condition & G_IO_IN) {
1825                 struct partial_reply *reply = server->incoming_reply;
1826                 int bytes_recv;
1827
1828                 if (!reply) {
1829                         unsigned char reply_len_buf[2];
1830                         uint16_t reply_len;
1831
1832                         bytes_recv = recv(sk, reply_len_buf, 2, MSG_PEEK);
1833                         if (!bytes_recv) {
1834                                 goto hangup;
1835                         } else if (bytes_recv < 0) {
1836                                 if (errno == EAGAIN || errno == EWOULDBLOCK)
1837                                         return TRUE;
1838
1839                                 connman_error("DNS proxy error %s",
1840                                                 strerror(errno));
1841                                 goto hangup;
1842                         } else if (bytes_recv < 2)
1843                                 return TRUE;
1844
1845                         reply_len = reply_len_buf[1] | reply_len_buf[0] << 8;
1846                         reply_len += 2;
1847
1848                         DBG("TCP reply %d bytes", reply_len);
1849
1850                         reply = g_try_malloc(sizeof(*reply) + reply_len + 2);
1851                         if (!reply)
1852                                 return TRUE;
1853
1854                         reply->len = reply_len;
1855                         reply->received = 0;
1856
1857                         server->incoming_reply = reply;
1858                 }
1859
1860                 while (reply->received < reply->len) {
1861                         bytes_recv = recv(sk, reply->buf + reply->received,
1862                                         reply->len - reply->received, 0);
1863                         if (!bytes_recv) {
1864                                 connman_error("DNS proxy TCP disconnect");
1865                                 break;
1866                         } else if (bytes_recv < 0) {
1867                                 if (errno == EAGAIN || errno == EWOULDBLOCK)
1868                                         return TRUE;
1869
1870                                 connman_error("DNS proxy error %s",
1871                                                 strerror(errno));
1872                                 break;
1873                         }
1874                         reply->received += bytes_recv;
1875                 }
1876
1877                 forward_dns_reply(reply->buf, reply->received, IPPROTO_TCP,
1878                                         server);
1879
1880                 g_free(reply);
1881                 server->incoming_reply = NULL;
1882
1883                 destroy_server(server);
1884
1885                 return FALSE;
1886         }
1887
1888         return TRUE;
1889 }
1890
1891 static gboolean tcp_idle_timeout(gpointer user_data)
1892 {
1893         struct server_data *server = user_data;
1894
1895         DBG("");
1896
1897         if (server == NULL)
1898                 return FALSE;
1899
1900         destroy_server(server);
1901
1902         return FALSE;
1903 }
1904
1905 static struct server_data *create_server(const char *interface,
1906                                         const char *domain, const char *server,
1907                                         int protocol)
1908 {
1909         struct addrinfo hints, *rp;
1910         struct server_data *data;
1911         int sk, ret;
1912
1913         DBG("interface %s server %s", interface, server);
1914
1915         memset(&hints, 0, sizeof(hints));
1916
1917         switch (protocol) {
1918         case IPPROTO_UDP:
1919                 hints.ai_socktype = SOCK_DGRAM;
1920                 break;
1921
1922         case IPPROTO_TCP:
1923                 hints.ai_socktype = SOCK_STREAM;
1924                 break;
1925
1926         default:
1927                 return NULL;
1928         }
1929         hints.ai_family = AF_UNSPEC;
1930         hints.ai_flags = AI_PASSIVE | AI_NUMERICSERV | AI_NUMERICHOST;
1931
1932         ret = getaddrinfo(server, "53", &hints, &rp);
1933         if (ret) {
1934                 connman_error("Failed to parse server %s address: %s\n",
1935                               server, gai_strerror(ret));
1936                 return NULL;
1937         }
1938         /* Do not blindly copy this code elsewhere; it doesn't loop over the
1939            results using ->ai_next as it should. That's OK in *this* case
1940            because it was a numeric lookup; we *know* there's only one. */
1941
1942         sk = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
1943         if (sk < 0) {
1944                 connman_error("Failed to create server %s socket", server);
1945                 freeaddrinfo(rp);
1946                 return NULL;
1947         }
1948
1949         if (interface != NULL) {
1950                 if (setsockopt(sk, SOL_SOCKET, SO_BINDTODEVICE,
1951                                 interface, strlen(interface) + 1) < 0) {
1952                         connman_error("Failed to bind server %s "
1953                                                 "to interface %s",
1954                                                         server, interface);
1955                         freeaddrinfo(rp);
1956                         close(sk);
1957                         return NULL;
1958                 }
1959         }
1960
1961         data = g_try_new0(struct server_data, 1);
1962         if (data == NULL) {
1963                 connman_error("Failed to allocate server %s data", server);
1964                 freeaddrinfo(rp);
1965                 close(sk);
1966                 return NULL;
1967         }
1968
1969         data->channel = g_io_channel_unix_new(sk);
1970         if (data->channel == NULL) {
1971                 connman_error("Failed to create server %s channel", server);
1972                 freeaddrinfo(rp);
1973                 close(sk);
1974                 g_free(data);
1975                 return NULL;
1976         }
1977
1978         g_io_channel_set_close_on_unref(data->channel, TRUE);
1979
1980         if (protocol == IPPROTO_TCP) {
1981                 g_io_channel_set_flags(data->channel, G_IO_FLAG_NONBLOCK, NULL);
1982                 data->watch = g_io_add_watch(data->channel,
1983                         G_IO_OUT | G_IO_IN | G_IO_HUP | G_IO_NVAL | G_IO_ERR,
1984                                                 tcp_server_event, data);
1985                 data->timeout = g_timeout_add_seconds(30, tcp_idle_timeout,
1986                                                                 data);
1987         } else
1988                 data->watch = g_io_add_watch(data->channel,
1989                         G_IO_IN | G_IO_NVAL | G_IO_ERR | G_IO_HUP,
1990                                                 udp_server_event, data);
1991
1992         data->interface = g_strdup(interface);
1993         if (domain)
1994                 data->domains = g_list_append(data->domains, g_strdup(domain));
1995         data->server = g_strdup(server);
1996         data->protocol = protocol;
1997
1998         ret = connect(sk, rp->ai_addr, rp->ai_addrlen);
1999         freeaddrinfo(rp);
2000         if (ret < 0) {
2001                 if ((protocol == IPPROTO_TCP && errno != EINPROGRESS) ||
2002                                 protocol == IPPROTO_UDP) {
2003                         GList *list;
2004
2005                         connman_error("Failed to connect to server %s", server);
2006                         if (data->watch > 0)
2007                                 g_source_remove(data->watch);
2008                         if (data->timeout > 0)
2009                                 g_source_remove(data->timeout);
2010
2011                         g_io_channel_unref(data->channel);
2012                         close(sk);
2013
2014                         g_free(data->server);
2015                         g_free(data->interface);
2016                         for (list = data->domains; list; list = list->next) {
2017                                 char *domain = list->data;
2018
2019                                 data->domains = g_list_remove(data->domains,
2020                                                                         domain);
2021                                 g_free(domain);
2022                         }
2023                         g_free(data);
2024                         return NULL;
2025                 }
2026         }
2027
2028         if (__sync_fetch_and_add(&cache_refcount, 1) == 0)
2029                 cache = g_hash_table_new_full(g_str_hash,
2030                                         g_str_equal,
2031                                         NULL,
2032                                         cache_element_destroy);
2033
2034         if (protocol == IPPROTO_UDP) {
2035                 /* Enable new servers by default */
2036                 data->enabled = TRUE;
2037                 connman_info("Adding DNS server %s", data->server);
2038
2039                 server_list = g_slist_append(server_list, data);
2040
2041                 return data;
2042         }
2043
2044         return NULL;
2045 }
2046
2047 static gboolean resolv(struct request_data *req,
2048                                 gpointer request, gpointer name)
2049 {
2050         GSList *list;
2051
2052         for (list = server_list; list; list = list->next) {
2053                 struct server_data *data = list->data;
2054
2055                 DBG("server %s enabled %d", data->server, data->enabled);
2056
2057                 if (data->enabled == FALSE)
2058                         continue;
2059
2060                 if (data->watch == 0 && data->protocol == IPPROTO_UDP)
2061                         data->watch = g_io_add_watch(data->channel,
2062                                 G_IO_IN | G_IO_NVAL | G_IO_ERR | G_IO_HUP,
2063                                                 udp_server_event, data);
2064
2065                 if (ns_resolv(data, req, request, name) > 0)
2066                         return TRUE;
2067         }
2068
2069         return FALSE;
2070 }
2071
2072 static void append_domain(const char *interface, const char *domain)
2073 {
2074         GSList *list;
2075
2076         DBG("interface %s domain %s", interface, domain);
2077
2078         if (domain == NULL)
2079                 return;
2080
2081         for (list = server_list; list; list = list->next) {
2082                 struct server_data *data = list->data;
2083                 GList *dom_list;
2084                 char *dom;
2085                 gboolean dom_found = FALSE;
2086
2087                 if (data->interface == NULL)
2088                         continue;
2089
2090                 if (g_str_equal(data->interface, interface) == FALSE)
2091                         continue;
2092
2093                 for (dom_list = data->domains; dom_list;
2094                                 dom_list = dom_list->next) {
2095                         dom = dom_list->data;
2096
2097                         if (g_str_equal(dom, domain)) {
2098                                 dom_found = TRUE;
2099                                 break;
2100                         }
2101                 }
2102
2103                 if (dom_found == FALSE) {
2104                         data->domains =
2105                                 g_list_append(data->domains, g_strdup(domain));
2106                 }
2107         }
2108 }
2109
2110 int __connman_dnsproxy_append(const char *interface, const char *domain,
2111                                                         const char *server)
2112 {
2113         struct server_data *data;
2114
2115         DBG("interface %s server %s", interface, server);
2116
2117         if (server == NULL && domain == NULL)
2118                 return -EINVAL;
2119
2120         if (server == NULL) {
2121                 append_domain(interface, domain);
2122
2123                 return 0;
2124         }
2125
2126         if (g_str_equal(server, "127.0.0.1") == TRUE)
2127                 return -ENODEV;
2128
2129         data = find_server(interface, server, IPPROTO_UDP);
2130         if (data != NULL) {
2131                 append_domain(interface, domain);
2132                 return 0;
2133         }
2134
2135         data = create_server(interface, domain, server, IPPROTO_UDP);
2136         if (data == NULL)
2137                 return -EIO;
2138
2139         return 0;
2140 }
2141
2142 static void remove_server(const char *interface, const char *domain,
2143                         const char *server, int protocol)
2144 {
2145         struct server_data *data;
2146
2147         data = find_server(interface, server, protocol);
2148         if (data == NULL)
2149                 return;
2150
2151         destroy_server(data);
2152 }
2153
2154 int __connman_dnsproxy_remove(const char *interface, const char *domain,
2155                                                         const char *server)
2156 {
2157         DBG("interface %s server %s", interface, server);
2158
2159         if (server == NULL)
2160                 return -EINVAL;
2161
2162         if (g_str_equal(server, "127.0.0.1") == TRUE)
2163                 return -ENODEV;
2164
2165         remove_server(interface, domain, server, IPPROTO_UDP);
2166         remove_server(interface, domain, server, IPPROTO_TCP);
2167
2168         return 0;
2169 }
2170
2171 void __connman_dnsproxy_flush(void)
2172 {
2173         GSList *list;
2174
2175         list = request_pending_list;
2176         while (list) {
2177                 struct request_data *req = list->data;
2178
2179                 list = list->next;
2180
2181                 request_pending_list =
2182                                 g_slist_remove(request_pending_list, req);
2183                 resolv(req, req->request, req->name);
2184                 g_free(req->request);
2185                 g_free(req->name);
2186         }
2187 }
2188
2189 static void dnsproxy_offline_mode(connman_bool_t enabled)
2190 {
2191         GSList *list;
2192
2193         DBG("enabled %d", enabled);
2194
2195         for (list = server_list; list; list = list->next) {
2196                 struct server_data *data = list->data;
2197
2198                 if (enabled == FALSE) {
2199                         connman_info("Enabling DNS server %s", data->server);
2200                         data->enabled = TRUE;
2201                         cache_invalidate();
2202                         cache_refresh();
2203                 } else {
2204                         connman_info("Disabling DNS server %s", data->server);
2205                         data->enabled = FALSE;
2206                         cache_invalidate();
2207                 }
2208         }
2209 }
2210
2211 static void dnsproxy_default_changed(struct connman_service *service)
2212 {
2213         GSList *list;
2214         char *interface;
2215
2216         DBG("service %p", service);
2217
2218         /* DNS has changed, invalidate the cache */
2219         cache_invalidate();
2220
2221         if (service == NULL) {
2222                 /* When no services are active, then disable DNS proxying */
2223                 dnsproxy_offline_mode(TRUE);
2224                 return;
2225         }
2226
2227         interface = connman_service_get_interface(service);
2228         if (interface == NULL)
2229                 return;
2230
2231         for (list = server_list; list; list = list->next) {
2232                 struct server_data *data = list->data;
2233
2234                 if (g_strcmp0(data->interface, interface) == 0) {
2235                         connman_info("Enabling DNS server %s", data->server);
2236                         data->enabled = TRUE;
2237                 } else {
2238                         connman_info("Disabling DNS server %s", data->server);
2239                         data->enabled = FALSE;
2240                 }
2241         }
2242
2243         g_free(interface);
2244         cache_refresh();
2245 }
2246
2247 static struct connman_notifier dnsproxy_notifier = {
2248         .name                   = "dnsproxy",
2249         .default_changed        = dnsproxy_default_changed,
2250         .offline_mode           = dnsproxy_offline_mode,
2251 };
2252
2253 static unsigned char opt_edns0_type[2] = { 0x00, 0x29 };
2254
2255 static int parse_request(unsigned char *buf, int len,
2256                                         char *name, unsigned int size)
2257 {
2258         struct domain_hdr *hdr = (void *) buf;
2259         uint16_t qdcount = ntohs(hdr->qdcount);
2260         uint16_t arcount = ntohs(hdr->arcount);
2261         unsigned char *ptr;
2262         char *last_label = NULL;
2263         unsigned int remain, used = 0;
2264
2265         if (len < 12)
2266                 return -EINVAL;
2267
2268         DBG("id 0x%04x qr %d opcode %d qdcount %d arcount %d",
2269                                         hdr->id, hdr->qr, hdr->opcode,
2270                                                         qdcount, arcount);
2271
2272         if (hdr->qr != 0 || qdcount != 1)
2273                 return -EINVAL;
2274
2275         name[0] = '\0';
2276
2277         ptr = buf + sizeof(struct domain_hdr);
2278         remain = len - sizeof(struct domain_hdr);
2279
2280         while (remain > 0) {
2281                 uint8_t len = *ptr;
2282
2283                 if (len == 0x00) {
2284                         last_label = (char *) (ptr + 1);
2285                         break;
2286                 }
2287
2288                 if (used + len + 1 > size)
2289                         return -ENOBUFS;
2290
2291                 strncat(name, (char *) (ptr + 1), len);
2292                 strcat(name, ".");
2293
2294                 used += len + 1;
2295
2296                 ptr += len + 1;
2297                 remain -= len + 1;
2298         }
2299
2300         if (last_label && arcount && remain >= 9 && last_label[4] == 0 &&
2301                                 !memcmp(last_label + 5, opt_edns0_type, 2)) {
2302                 uint16_t edns0_bufsize;
2303
2304                 edns0_bufsize = last_label[7] << 8 | last_label[8];
2305
2306                 DBG("EDNS0 buffer size %u", edns0_bufsize);
2307
2308                 /* This is an evil hack until full TCP support has been
2309                  * implemented.
2310                  *
2311                  * Somtimes the EDNS0 request gets send with a too-small
2312                  * buffer size. Since glibc doesn't seem to crash when it
2313                  * gets a response biffer then it requested, just bump
2314                  * the buffer size up to 4KiB.
2315                  */
2316                 if (edns0_bufsize < 0x1000) {
2317                         last_label[7] = 0x10;
2318                         last_label[8] = 0x00;
2319                 }
2320         }
2321
2322         DBG("query %s", name);
2323
2324         return 0;
2325 }
2326
2327 static gboolean tcp_listener_event(GIOChannel *channel, GIOCondition condition,
2328                                                         gpointer user_data)
2329 {
2330         unsigned char buf[768];
2331         char query[512];
2332         struct request_data *req;
2333         struct server_data *server;
2334         int sk, client_sk, len, err;
2335         struct sockaddr_in6 client_addr;
2336         socklen_t client_addr_len = sizeof(client_addr);
2337         GSList *list;
2338         struct listener_data *ifdata = user_data;
2339
2340         DBG("condition 0x%x", condition);
2341
2342         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
2343                 if (ifdata->tcp_listener_watch > 0)
2344                         g_source_remove(ifdata->tcp_listener_watch);
2345                 ifdata->tcp_listener_watch = 0;
2346
2347                 connman_error("Error with TCP listener channel");
2348
2349                 return FALSE;
2350         }
2351
2352         sk = g_io_channel_unix_get_fd(channel);
2353
2354         client_sk = accept(sk, (void *)&client_addr, &client_addr_len);
2355         if (client_sk < 0) {
2356                 connman_error("Accept failure on TCP listener");
2357                 ifdata->tcp_listener_watch = 0;
2358                 return FALSE;
2359         }
2360
2361         len = recv(client_sk, buf, sizeof(buf), 0);
2362         if (len < 2)
2363                 return TRUE;
2364
2365         DBG("Received %d bytes (id 0x%04x)", len, buf[2] | buf[3] << 8);
2366
2367         err = parse_request(buf + 2, len - 2, query, sizeof(query));
2368         if (err < 0 || (g_slist_length(server_list) == 0)) {
2369                 send_response(client_sk, buf, len, NULL, 0, IPPROTO_TCP);
2370                 return TRUE;
2371         }
2372
2373         req = g_try_new0(struct request_data, 1);
2374         if (req == NULL)
2375                 return TRUE;
2376
2377         memcpy(&req->sa, &client_addr, client_addr_len);
2378         req->sa_len = client_addr_len;
2379         req->client_sk = client_sk;
2380         req->protocol = IPPROTO_TCP;
2381
2382         request_id += 2;
2383         if (request_id == 0x0000 || request_id == 0xffff)
2384                 request_id += 2;
2385
2386         req->srcid = buf[2] | (buf[3] << 8);
2387         req->dstid = request_id;
2388         req->altid = request_id + 1;
2389         req->request_len = len;
2390
2391         buf[2] = req->dstid & 0xff;
2392         buf[3] = req->dstid >> 8;
2393
2394         req->numserv = 0;
2395         req->ifdata = (struct listener_data *) ifdata;
2396         req->append_domain = FALSE;
2397         request_list = g_slist_append(request_list, req);
2398
2399         for (list = server_list; list; list = list->next) {
2400                 struct server_data *data = list->data;
2401                 GList *domains;
2402
2403                 if (data->protocol != IPPROTO_UDP || data->enabled == FALSE)
2404                         continue;
2405
2406                 server = create_server(data->interface, NULL,
2407                                         data->server, IPPROTO_TCP);
2408
2409                 /*
2410                  * If server is NULL, we're not connected yet.
2411                  * Copy the relevant buffers and continue with
2412                  * the next nameserver.
2413                  * The request will actually be sent once we're
2414                  * properly connected over TCP to this nameserver.
2415                  */
2416                 if (server == NULL) {
2417                         req->request = g_try_malloc0(req->request_len);
2418                         if (req->request == NULL)
2419                                 return TRUE;
2420
2421                         memcpy(req->request, buf, req->request_len);
2422
2423                         req->name = g_try_malloc0(sizeof(query));
2424                         if (req->name == NULL) {
2425                                 g_free(req->request);
2426                                 return TRUE;
2427                         }
2428                         memcpy(req->name, query, sizeof(query));
2429
2430                         continue;
2431                 }
2432
2433                 if (req->timeout > 0)
2434                         g_source_remove(req->timeout);
2435
2436                 for (domains = data->domains; domains;
2437                                 domains = domains->next) {
2438                         char *dom = domains->data;
2439
2440                         DBG("Adding domain %s to %s", dom, server->server);
2441
2442                         server->domains = g_list_append(server->domains,
2443                                                 g_strdup(dom));
2444                 }
2445
2446                 req->timeout = g_timeout_add_seconds(30, request_timeout, req);
2447                 if (ns_resolv(server, req, buf, query) > 0) {
2448                         if (req->timeout > 0) {
2449                                 g_source_remove(req->timeout);
2450                                 req->timeout = 0;
2451                         }
2452                 }
2453         }
2454
2455         return TRUE;
2456 }
2457
2458 static gboolean udp_listener_event(GIOChannel *channel, GIOCondition condition,
2459                                                         gpointer user_data)
2460 {
2461         unsigned char buf[768];
2462         char query[512];
2463         struct request_data *req;
2464         struct sockaddr_in6 client_addr;
2465         socklen_t client_addr_len = sizeof(client_addr);
2466         int sk, err, len;
2467         struct listener_data *ifdata = user_data;
2468
2469         if (condition & (G_IO_NVAL | G_IO_ERR | G_IO_HUP)) {
2470                 connman_error("Error with UDP listener channel");
2471                 ifdata->udp_listener_watch = 0;
2472                 return FALSE;
2473         }
2474
2475         sk = g_io_channel_unix_get_fd(channel);
2476
2477         memset(&client_addr, 0, client_addr_len);
2478         len = recvfrom(sk, buf, sizeof(buf), 0, (void *)&client_addr,
2479                        &client_addr_len);
2480         if (len < 2)
2481                 return TRUE;
2482
2483         DBG("Received %d bytes (id 0x%04x)", len, buf[0] | buf[1] << 8);
2484
2485         err = parse_request(buf, len, query, sizeof(query));
2486         if (err < 0 || (g_slist_length(server_list) == 0)) {
2487                 send_response(sk, buf, len, (void *)&client_addr,
2488                                 client_addr_len, IPPROTO_UDP);
2489                 return TRUE;
2490         }
2491
2492         req = g_try_new0(struct request_data, 1);
2493         if (req == NULL)
2494                 return TRUE;
2495
2496         memcpy(&req->sa, &client_addr, client_addr_len);
2497         req->sa_len = client_addr_len;
2498         req->client_sk = 0;
2499         req->protocol = IPPROTO_UDP;
2500
2501         request_id += 2;
2502         if (request_id == 0x0000 || request_id == 0xffff)
2503                 request_id += 2;
2504
2505         req->srcid = buf[0] | (buf[1] << 8);
2506         req->dstid = request_id;
2507         req->altid = request_id + 1;
2508         req->request_len = len;
2509
2510         buf[0] = req->dstid & 0xff;
2511         buf[1] = req->dstid >> 8;
2512
2513         req->numserv = 0;
2514         req->ifdata = (struct listener_data *) ifdata;
2515         req->append_domain = FALSE;
2516
2517         if (resolv(req, buf, query) == TRUE) {
2518                 /* a cached result was sent, so the request can be released */
2519                 g_free(req);
2520                 return TRUE;
2521         }
2522
2523         req->timeout = g_timeout_add_seconds(5, request_timeout, req);
2524         request_list = g_slist_append(request_list, req);
2525
2526         return TRUE;
2527 }
2528
2529 static int create_dns_listener(int protocol, struct listener_data *ifdata)
2530 {
2531         GIOChannel *channel;
2532         const char *proto;
2533         union {
2534                 struct sockaddr sa;
2535                 struct sockaddr_in6 sin6;
2536                 struct sockaddr_in sin;
2537         } s;
2538         socklen_t slen;
2539         int sk, type, v6only = 0;
2540         int family = AF_INET6;
2541
2542
2543         DBG("interface %s", ifdata->ifname);
2544
2545         switch (protocol) {
2546         case IPPROTO_UDP:
2547                 proto = "UDP";
2548                 type = SOCK_DGRAM | SOCK_CLOEXEC;
2549                 break;
2550
2551         case IPPROTO_TCP:
2552                 proto = "TCP";
2553                 type = SOCK_STREAM | SOCK_CLOEXEC;
2554                 break;
2555
2556         default:
2557                 return -EINVAL;
2558         }
2559
2560         sk = socket(family, type, protocol);
2561         if (sk < 0 && family == AF_INET6 && errno == EAFNOSUPPORT) {
2562                 connman_error("No IPv6 support; DNS proxy listening only on Legacy IP");
2563                 family = AF_INET;
2564                 sk = socket(family, type, protocol);
2565         }
2566         if (sk < 0) {
2567                 connman_error("Failed to create %s listener socket", proto);
2568                 return -EIO;
2569         }
2570
2571         if (setsockopt(sk, SOL_SOCKET, SO_BINDTODEVICE,
2572                                         ifdata->ifname,
2573                                         strlen(ifdata->ifname) + 1) < 0) {
2574                 connman_error("Failed to bind %s listener interface", proto);
2575                 close(sk);
2576                 return -EIO;
2577         }
2578         /* Ensure it accepts Legacy IP connections too */
2579         if (family == AF_INET6 &&
2580                         setsockopt(sk, SOL_IPV6, IPV6_V6ONLY,
2581                                         &v6only, sizeof(v6only)) < 0) {
2582                 connman_error("Failed to clear V6ONLY on %s listener socket",
2583                               proto);
2584                 close(sk);
2585                 return -EIO;
2586         }
2587
2588         if (family == AF_INET) {
2589                 memset(&s.sin, 0, sizeof(s.sin));
2590                 s.sin.sin_family = AF_INET;
2591                 s.sin.sin_port = htons(53);
2592                 s.sin.sin_addr.s_addr = htonl(INADDR_ANY);
2593                 slen = sizeof(s.sin);
2594         } else {
2595                 memset(&s.sin6, 0, sizeof(s.sin6));
2596                 s.sin6.sin6_family = AF_INET6;
2597                 s.sin6.sin6_port = htons(53);
2598                 s.sin6.sin6_addr = in6addr_any;
2599                 slen = sizeof(s.sin6);
2600         }
2601
2602         if (bind(sk, &s.sa, slen) < 0) {
2603                 connman_error("Failed to bind %s listener socket", proto);
2604                 close(sk);
2605                 return -EIO;
2606         }
2607
2608         if (protocol == IPPROTO_TCP && listen(sk, 10) < 0) {
2609                 connman_error("Failed to listen on TCP socket");
2610                 close(sk);
2611                 return -EIO;
2612         }
2613
2614         channel = g_io_channel_unix_new(sk);
2615         if (channel == NULL) {
2616                 connman_error("Failed to create %s listener channel", proto);
2617                 close(sk);
2618                 return -EIO;
2619         }
2620
2621         g_io_channel_set_close_on_unref(channel, TRUE);
2622
2623         if (protocol == IPPROTO_TCP) {
2624                 ifdata->tcp_listener_channel = channel;
2625                 ifdata->tcp_listener_watch = g_io_add_watch(channel,
2626                                 G_IO_IN, tcp_listener_event, (gpointer) ifdata);
2627         } else {
2628                 ifdata->udp_listener_channel = channel;
2629                 ifdata->udp_listener_watch = g_io_add_watch(channel,
2630                                 G_IO_IN, udp_listener_event, (gpointer) ifdata);
2631         }
2632
2633         return 0;
2634 }
2635
2636 static void destroy_udp_listener(struct listener_data *ifdata)
2637 {
2638         DBG("interface %s", ifdata->ifname);
2639
2640         if (ifdata->udp_listener_watch > 0)
2641                 g_source_remove(ifdata->udp_listener_watch);
2642
2643         g_io_channel_unref(ifdata->udp_listener_channel);
2644 }
2645
2646 static void destroy_tcp_listener(struct listener_data *ifdata)
2647 {
2648         DBG("interface %s", ifdata->ifname);
2649
2650         if (ifdata->tcp_listener_watch > 0)
2651                 g_source_remove(ifdata->tcp_listener_watch);
2652
2653         g_io_channel_unref(ifdata->tcp_listener_channel);
2654 }
2655
2656 static int create_listener(struct listener_data *ifdata)
2657 {
2658         int err;
2659
2660         err = create_dns_listener(IPPROTO_UDP, ifdata);
2661         if (err < 0)
2662                 return err;
2663
2664         err = create_dns_listener(IPPROTO_TCP, ifdata);
2665         if (err < 0) {
2666                 destroy_udp_listener(ifdata);
2667                 return err;
2668         }
2669
2670         if (g_strcmp0(ifdata->ifname, "lo") == 0)
2671                 __connman_resolvfile_append("lo", NULL, "127.0.0.1");
2672
2673         return 0;
2674 }
2675
2676 static void destroy_request_data(struct request_data *req)
2677 {
2678         if (req->timeout > 0)
2679                 g_source_remove(req->timeout);
2680
2681         g_free(req->resp);
2682         g_free(req->request);
2683         g_free(req->name);
2684         g_free(req);
2685 }
2686
2687 static void destroy_listener(struct listener_data *ifdata)
2688 {
2689         GSList *list;
2690
2691         if (g_strcmp0(ifdata->ifname, "lo") == 0)
2692                 __connman_resolvfile_remove("lo", NULL, "127.0.0.1");
2693
2694         for (list = request_pending_list; list; list = list->next) {
2695                 struct request_data *req = list->data;
2696
2697                 DBG("Dropping pending request (id 0x%04x -> 0x%04x)",
2698                                                 req->srcid, req->dstid);
2699                 destroy_request_data(req);
2700                 list->data = NULL;
2701         }
2702
2703         g_slist_free(request_pending_list);
2704         request_pending_list = NULL;
2705
2706         for (list = request_list; list; list = list->next) {
2707                 struct request_data *req = list->data;
2708
2709                 DBG("Dropping request (id 0x%04x -> 0x%04x)",
2710                                                 req->srcid, req->dstid);
2711                 destroy_request_data(req);
2712                 list->data = NULL;
2713         }
2714
2715         g_slist_free(request_list);
2716         request_list = NULL;
2717
2718         destroy_tcp_listener(ifdata);
2719         destroy_udp_listener(ifdata);
2720 }
2721
2722 int __connman_dnsproxy_add_listener(const char *interface)
2723 {
2724         struct listener_data *ifdata;
2725         int err;
2726
2727         DBG("interface %s", interface);
2728
2729         if (g_hash_table_lookup(listener_table, interface) != NULL)
2730                 return 0;
2731
2732         ifdata = g_try_new0(struct listener_data, 1);
2733         if (ifdata == NULL)
2734                 return -ENOMEM;
2735
2736         ifdata->ifname = g_strdup(interface);
2737         ifdata->udp_listener_channel = NULL;
2738         ifdata->udp_listener_watch = 0;
2739         ifdata->tcp_listener_channel = NULL;
2740         ifdata->tcp_listener_watch = 0;
2741
2742         err = create_listener(ifdata);
2743         if (err < 0) {
2744                 connman_error("Couldn't create listener for %s err %d",
2745                                 interface, err);
2746                 g_free(ifdata->ifname);
2747                 g_free(ifdata);
2748                 return err;
2749         }
2750         g_hash_table_insert(listener_table, ifdata->ifname, ifdata);
2751         return 0;
2752 }
2753
2754 void __connman_dnsproxy_remove_listener(const char *interface)
2755 {
2756         struct listener_data *ifdata;
2757
2758         DBG("interface %s", interface);
2759
2760         ifdata = g_hash_table_lookup(listener_table, interface);
2761         if (ifdata == NULL)
2762                 return;
2763
2764         destroy_listener(ifdata);
2765
2766         g_hash_table_remove(listener_table, interface);
2767 }
2768
2769 static void remove_listener(gpointer key, gpointer value, gpointer user_data)
2770 {
2771         const char *interface = key;
2772         struct listener_data *ifdata = value;
2773
2774         DBG("interface %s", interface);
2775
2776         destroy_listener(ifdata);
2777 }
2778
2779 int __connman_dnsproxy_init(void)
2780 {
2781         int err;
2782
2783         DBG("");
2784
2785         listener_table = g_hash_table_new_full(g_str_hash, g_str_equal,
2786                                                         g_free, g_free);
2787         err = __connman_dnsproxy_add_listener("lo");
2788         if (err < 0)
2789                 return err;
2790
2791         err = connman_notifier_register(&dnsproxy_notifier);
2792         if (err < 0)
2793                 goto destroy;
2794
2795         return 0;
2796
2797 destroy:
2798         __connman_dnsproxy_remove_listener("lo");
2799         g_hash_table_destroy(listener_table);
2800
2801         return err;
2802 }
2803
2804 void __connman_dnsproxy_cleanup(void)
2805 {
2806         DBG("");
2807
2808         connman_notifier_unregister(&dnsproxy_notifier);
2809
2810         g_hash_table_foreach(listener_table, remove_listener, NULL);
2811
2812         g_hash_table_destroy(listener_table);
2813 }