evdns.c

   1 /* $Id: evdns.c,v 1.7 2007-11-03 20:24:42 root Exp $ */
   2
   3 /* The original version of this module was written by Adam Langley; for
   4  * a history of modifications, check out the subversion logs.
   5  *
   6  * When editing this module, try to keep it re-mergeable by Adam.  Don't
   7  * reformat the whitespace, add Tor dependencies, or so on.
   8  *
   9  * TODO:
  10  *   - Support IPv6 and PTR records.
  11  *   - Replace all externally visible magic numbers with #defined constants.
  12  *   - Write doccumentation for APIs of all external functions.
  13  */
  14
  15 /* Async DNS Library
  16  * Adam Langley <agl@imperialviolet.org>
  17  * http://www.imperialviolet.org/eventdns.html
  18  * Public Domain code
  19  *
  20  * This software is Public Domain. To view a copy of the public domain dedication,
  21  * visit http://creativecommons.org/licenses/publicdomain/ or send a letter to
  22  * Creative Commons, 559 Nathan Abbott Way, Stanford, California 94305, USA.
  23  *
  24  * I ask and expect, but do not require, that all derivative works contain an
  25  * attribution similar to:
  26  *      Parts developed by Adam Langley <agl@imperialviolet.org>
  27  *
  28  * You may wish to replace the word "Parts" with something else depending on
  29  * the amount of original code.
  30  *
  31  * (Derivative works does not include programs which link against, run or include
  32  * the source verbatim in their source distributions)
  33  *
  34  * Version: 0.1b
  35  */
  36
  37 #include <sys/types.h>
  38 #ifdef HAVE_CONFIG_H
  39 #include "config.h"
  40 #endif
  41
  42 #ifdef WIN32
  43 #include "misc.h"
  44 #endif
  45
  46 /* #define NDEBUG */
  47
  48 #ifndef DNS_USE_CPU_CLOCK_FOR_ID
  49 #ifndef DNS_USE_GETTIMEOFDAY_FOR_ID
  50 #ifndef DNS_USE_OPENSSL_FOR_ID
  51 #error Must configure at least one id generation method.
  52 #error Please see the documentation.
  53 #endif
  54 #endif
  55 #endif
  56
  57 /* #define _POSIX_C_SOURCE 200507 */
  58 #define _GNU_SOURCE
  59
  60 #ifdef DNS_USE_CPU_CLOCK_FOR_ID
  61 #ifdef DNS_USE_OPENSSL_FOR_ID
  62 #error Multiple id options selected
  63 #endif
  64 #ifdef DNS_USE_GETTIMEOFDAY_FOR_ID
  65 #error Multiple id options selected
  66 #endif
  67 #include <time.h>
  68 #endif
  69
  70 #ifdef DNS_USE_OPENSSL_FOR_ID
  71 #ifdef DNS_USE_GETTIMEOFDAY_FOR_ID
  72 #error Multiple id options selected
  73 #endif
  74 #include <openssl/rand.h>
  75 #endif
  76
  77 #define _FORTIFY_SOURCE 3
  78
  79 #include <string.h>
  80 #include <fcntl.h>
  81 #include <sys/time.h>
  82 #ifdef HAVE_STDINT_H
  83 #include <stdint.h>
  84 #endif
  85 #include <stdlib.h>
  86 #include <string.h>
  87 #include <errno.h>
  88 #include <assert.h>
  89 #include <unistd.h>
  90 #include <limits.h>
  91 #include <sys/stat.h>
  92 #include <ctype.h>
  93 #include <stdio.h>
  94 #include <stdarg.h>
  95
  96 #include "evdns.h"
  97 #include "log.h"
  98 #ifdef WIN32
  99 #include <windows.h>
 100 #include <winsock2.h>
 101 #include <iphlpapi.h>
 102 #else
 103 #include <sys/socket.h>
 104 #include <netinet/in.h>
 105 #include <arpa/inet.h>
 106 #endif
 107
 108 #ifdef HAVE_NETINET_IN6_H
 109 #include <netinet/in6.h>
 110 #endif
 111
 112 #ifdef WIN32
 113 typedef int socklen_t;
 114 #endif
 115
 116 #define EVDNS_LOG_DEBUG 0
 117 #define EVDNS_LOG_WARN 1
 118
 119 #ifndef HOST_NAME_MAX
 120 #define HOST_NAME_MAX 255
 121 #endif
 122
 123 #ifndef NDEBUG
 124 #include <stdio.h>
 125 #endif
 126
 127 #undef MIN
 128 #define MIN(a,b) ((a)<(b)?(a):(b))
 129
 130 #ifdef __USE_ISOC99B
 131 /* libevent doesn't work without this */
 132 typedef uint8_t u_char;
 133 typedef unsigned int uint;
 134 #endif
 135 #include <event.h>
 136
 137 #define u64 uint64_t
 138 #define u32 uint32_t
 139 #define u16 uint16_t
 140 #define u8  uint8_t
 141
 142 #define MAX_ADDRS 4  /* maximum number of addresses from a single packet */
 143 /* which we bother recording */
 144
 145 #define TYPE_A         EVDNS_TYPE_A
 146 #define TYPE_CNAME     5
 147 #define TYPE_PTR       EVDNS_TYPE_PTR
 148 #define TYPE_AAAA      EVDNS_TYPE_AAAA
 149
 150 #define CLASS_INET     EVDNS_CLASS_INET
 151
 152 struct request {
 153         u8 *request;  /* the dns packet data */
 154         unsigned int request_len;
 155         int reissue_count;
 156         int tx_count;  /* the number of times that this packet has been sent */
 157         unsigned int request_type; /* TYPE_PTR or TYPE_A */
 158         void *user_pointer;  /* the pointer given to us for this request */
 159         evdns_callback_type user_callback;
 160         struct nameserver *ns;  /* the server which we last sent it */
 161
 162         /* elements used by the searching code */
 163         int search_index;
 164         struct search_state *search_state;
 165         char *search_origname;  /* needs to be free()ed */
 166         int search_flags;
 167
 168         /* these objects are kept in a circular list */
 169         struct request *next, *prev;
 170
 171         struct event timeout_event;
 172
 173         u16 trans_id;  /* the transaction id */
 174         char request_appended;  /* true if the request pointer is data which follows this struct */
 175         char transmit_me;  /* needs to be transmitted */
 176 };
 177
 178 #ifndef HAVE_STRUCT_IN6_ADDR
 179 struct in6_addr {
 180         u8 s6_addr[16];
 181 };
 182 #endif
 183
 184 struct reply {
 185         unsigned int type;
 186         unsigned int have_answer;
 187         union {
 188                 struct {
 189                         u32 addrcount;
 190                         u32 addresses[MAX_ADDRS];
 191                 } a;
 192                 struct {
 193                         u32 addrcount;
 194                         struct in6_addr addresses[MAX_ADDRS];
 195                 } aaaa;
 196                 struct {
 197                         char name[HOST_NAME_MAX];
 198                 } ptr;
 199         } data;
 200 };
 201
 202 struct nameserver {
 203         int socket;  /* a connected UDP socket */
 204         u32 address;
 205         int failed_times;  /* number of times which we have given this server a chance */
 206         int timedout;  /* number of times in a row a request has timed out */
 207         struct event event;
 208         /* these objects are kept in a circular list */
 209         struct nameserver *next, *prev;
 210         struct event timeout_event;  /* used to keep the timeout for */
 211                                      /* when we next probe this server. */
 212                                      /* Valid if state == 0 */
 213         char state;  /* zero if we think that this server is down */
 214         char choked;  /* true if we have an EAGAIN from this server's socket */
 215         char write_waiting;  /* true if we are waiting for EV_WRITE events */
 216 };
 217
 218 static struct request *req_head = NULL, *req_waiting_head = NULL;
 219 static struct nameserver *server_head = NULL;
 220
 221 /* Represents a local port where we're listening for DNS requests. Right now, */
 222 /* only UDP is supported. */
 223 struct evdns_server_port {
 224         int socket; /* socket we use to read queries and write replies. */
 225         int refcnt; /* reference count. */
 226         char choked; /* Are we currently blocked from writing? */
 227         char closing; /* Are we trying to close this port, pending writes? */
 228         evdns_request_callback_fn_type user_callback; /* Fn to handle requests */
 229         void *user_data; /* Opaque pointer passed to user_callback */
 230         struct event event; /* Read/write event */
 231         /* circular list of replies that we want to write. */
 232         struct server_request *pending_replies;
 233 };
 234
 235 /* Represents part of a reply being built.      (That is, a single RR.) */
 236 struct server_reply_item {
 237         struct server_reply_item *next; /* next item in sequence. */
 238         char *name; /* name part of the RR */
 239         u16 type : 16; /* The RR type */
 240         u16 class : 16; /* The RR class (usually CLASS_INET) */
 241         u32 ttl; /* The RR TTL */
 242         char is_name; /* True iff data is a label */
 243         u16 datalen; /* Length of data; -1 if data is a label */
 244         void *data; /* The contents of the RR */
 245 };
 246
 247 /* Represents a request that we've received as a DNS server, and holds */
 248 /* the components of the reply as we're constructing it. */
 249 struct server_request {
 250         /* Pointers to the next and previous entries on the list of replies */
 251         /* that we're waiting to write.  Only set if we have tried to respond */
 252         /* and gotten EAGAIN. */
 253         struct server_request *next_pending;
 254         struct server_request *prev_pending;
 255
 256         u16 trans_id; /* Transaction id. */
 257         struct evdns_server_port *port; /* Which port received this request on? */
 258         struct sockaddr_storage addr; /* Where to send the response */
 259         socklen_t addrlen; /* length of addr */
 260
 261         int n_answer; /* how many answer RRs have been set? */
 262         int n_authority; /* how many authority RRs have been set? */
 263         int n_additional; /* how many additional RRs have been set? */
 264
 265         struct server_reply_item *answer; /* linked list of answer RRs */
 266         struct server_reply_item *authority; /* linked list of authority RRs */
 267         struct server_reply_item *additional; /* linked list of additional RRs */
 268
 269         /* Constructed response.  Only set once we're ready to send a reply. */
 270         /* Once this is set, the RR fields are cleared, and no more should be set. */
 271         char *response;
 272         size_t response_len;
 273
 274         /* Caller-visible fields: flags, questions. */
 275         struct evdns_server_request base;
 276 };
 277
 278 /* helper macro */
 279 #define OFFSET_OF(st, member) ((off_t) (((char*)&((st*)0)->member)-(char*)0))
 280
 281 /* Given a pointer to an evdns_server_request, get the corresponding */
 282 /* server_request. */
 283 #define TO_SERVER_REQUEST(base_ptr)                                                                             \
 284         ((struct server_request*)                                                                                       \
 285          (((char*)(base_ptr) - OFFSET_OF(struct server_request, base))))
 286
 287 /* The number of good nameservers that we have */
 288 static int global_good_nameservers = 0;
 289
 290 /* inflight requests are contained in the req_head list */
 291 /* and are actually going out across the network */
 292 static int global_requests_inflight = 0;
 293 /* requests which aren't inflight are in the waiting list */
 294 /* and are counted here */
 295 static int global_requests_waiting = 0;
 296
 297 static int global_max_requests_inflight = 64;
 298
 299 static struct timeval global_timeout = {5, 0};  /* 5 seconds */
 300 static int global_max_reissues = 1;  /* a reissue occurs when we get some errors from the server */
 301 static int global_max_retransmits = 3;  /* number of times we'll retransmit a request which timed out */
 302 /* number of timeouts in a row before we consider this server to be down */
 303 static int global_max_nameserver_timeout = 3;
 304
 305 /* These are the timeout values for nameservers. If we find a nameserver is down */
 306 /* we try to probe it at intervals as given below. Values are in seconds. */
 307 static const struct timeval global_nameserver_timeouts[] = {{10, 0}, {60, 0}, {300, 0}, {900, 0}, {3600, 0}};
 308 static const int global_nameserver_timeouts_length = sizeof(global_nameserver_timeouts)/sizeof(struct timeval);
 309
 310 static struct nameserver *nameserver_pick(void);
 311 static void evdns_request_insert(struct request *req, struct request **head);
 312 static void nameserver_ready_callback(int fd, short events, void *arg);
 313 static int evdns_transmit(void);
 314 static int evdns_request_transmit(struct request *req);
 315 static void nameserver_send_probe(struct nameserver *const ns);
 316 static void search_request_finished(struct request *const);
 317 static int search_try_next(struct request *const req);
 318 static int search_request_new(int type, const char *const name, int flags, evdns_callback_type user_callback, void *user_arg);
 319 static void evdns_requests_pump_waiting_queue(void);
 320 static u16 transaction_id_pick(void);
 321 static struct request *request_new(int type, const char *name, int flags, evdns_callback_type callback, void *ptr);
 322 static void request_submit(struct request *req);
 323
 324 static int server_request_free(struct server_request *req);
 325 static void server_request_free_answers(struct server_request *req);
 326 static void server_port_free(struct evdns_server_port *port);
 327 static void server_port_ready_callback(int fd, short events, void *arg);
 328
 329 static int strtoint(const char *const str);
 330
 331 #ifdef WIN32
 332 static int
 333 last_error(int sock)
 334 {
 335         int optval, optvallen=sizeof(optval);
 336         int err = WSAGetLastError();
 337         if (err == WSAEWOULDBLOCK && sock >= 0) {
 338                 if (getsockopt(sock, SOL_SOCKET, SO_ERROR, (void*)&optval,
 339                                &optvallen))
 340                         return err;
 341                 if (optval)
 342                         return optval;
 343         }
 344         return err;
 345
 346 }
 347 static int
 348 error_is_eagain(int err)
 349 {
 350         return err == EAGAIN || err == WSAEWOULDBLOCK;
 351 }
 352 static int
 353 inet_aton(const char *c, struct in_addr *addr)
 354 {
 355         uint32_t r;
 356         if (strcmp(c, "255.255.255.255") == 0) {
 357                 addr->s_addr = 0xffffffffu;
 358         } else {
 359                 r = inet_addr(c);
 360                 if (r == INADDR_NONE)
 361                         return 0;
 362                 addr->s_addr = r;
 363         }
 364         return 1;
 365 }
 366 #define CLOSE_SOCKET(x) closesocket(x)
 367 #else
 368 #define last_error(sock) (errno)
 369 #define error_is_eagain(err) ((err) == EAGAIN)
 370 #define CLOSE_SOCKET(x) close(x)
 371 #endif
 372
 373 #define ISSPACE(c) isspace((int)(unsigned char)(c))
 374 #define ISDIGIT(c) isdigit((int)(unsigned char)(c))
 375
 376 #ifndef NDEBUG
 377 static const char *
 378 debug_ntoa(u32 address)
 379 {
 380         static char buf[32];
 381         u32 a = ntohl(address);
 382         snprintf(buf, sizeof(buf), "%d.%d.%d.%d",
 383                       (int)(u8)((a>>24)&0xff),
 384                       (int)(u8)((a>>16)&0xff),
 385                       (int)(u8)((a>>8 )&0xff),
 386                       (int)(u8)((a    )&0xff));
 387         return buf;
 388 }
 389 #endif
 390
 391 static evdns_debug_log_fn_type evdns_log_fn = NULL;
 392
 393 void
 394 evdns_set_log_fn(evdns_debug_log_fn_type fn)
 395 {
 396   evdns_log_fn = fn;
 397 }
 398
 399 #ifdef __GNUC__
 400 #define EVDNS_LOG_CHECK  __attribute__ ((format(printf, 2, 3)))
 401 #else
 402 #define EVDNS_LOG_CHECK
 403 #endif
 404
 405 static void _evdns_log(int warn, const char *fmt, ...) EVDNS_LOG_CHECK;
 406 static void
 407 _evdns_log(int warn, const char *fmt, ...)
 408 {
 409   va_list args;
 410   static char buf[512];
 411   if (!evdns_log_fn)
 412     return;
 413   va_start(args,fmt);
 414 #ifdef WIN32
 415   _vsnprintf(buf, sizeof(buf), fmt, args);
 416 #else
 417   vsnprintf(buf, sizeof(buf), fmt, args);
 418 #endif
 419   buf[sizeof(buf)-1] = '\0';
 420   evdns_log_fn(warn, buf);
 421   va_end(args);
 422 }
 423
 424 #define log _evdns_log
 425
 426 /* This walks the list of inflight requests to find the */
 427 /* one with a matching transaction id. Returns NULL on */
 428 /* failure */
 429 static struct request *
 430 request_find_from_trans_id(u16 trans_id) {
 431         struct request *req = req_head, *const started_at = req_head;
 432
 433         if (req) {
 434                 do {
 435                         if (req->trans_id == trans_id) return req;
 436                         req = req->next;
 437                 } while (req != started_at);
 438         }
 439
 440         return NULL;
 441 }
 442
 443 /* a libevent callback function which is called when a nameserver */
 444 /* has gone down and we want to test if it has came back to life yet */
 445 static void
 446 nameserver_prod_callback(int fd, short events, void *arg) {
 447         struct nameserver *const ns = (struct nameserver *) arg;
 448         (void)fd;
 449         (void)events;
 450
 451         nameserver_send_probe(ns);
 452 }
 453
 454 /* a libevent callback which is called when a nameserver probe (to see if */
 455 /* it has come back to life) times out. We increment the count of failed_times */
 456 /* and wait longer to send the next probe packet. */
 457 static void
 458 nameserver_probe_failed(struct nameserver *const ns) {
 459         const struct timeval * timeout;
 460         (void) evtimer_del(&ns->timeout_event);
 461         if (ns->state == 1) {
 462                 /* This can happen if the nameserver acts in a way which makes us mark */
 463                 /* it as bad and then starts sending good replies. */
 464                 return;
 465         }
 466
 467         timeout =
 468           &global_nameserver_timeouts[MIN(ns->failed_times,
 469                                           global_nameserver_timeouts_length - 1)];
 470         ns->failed_times++;
 471
 472         evtimer_set(&ns->timeout_event, nameserver_prod_callback, ns);
 473         if (evtimer_add(&ns->timeout_event, (struct timeval *) timeout) < 0) {
 474           log(EVDNS_LOG_WARN,
 475               "Error from libevent when adding timer event for %s",
 476               debug_ntoa(ns->address));
 477           /* ???? Do more? */
 478         }
 479 }
 480
 481 /* called when a nameserver has been deemed to have failed. For example, too */
 482 /* many packets have timed out etc */
 483 static void
 484 nameserver_failed(struct nameserver *const ns, const char *msg) {
 485         struct request *req, *started_at;
 486         /* if this nameserver has already been marked as failed */
 487         /* then don't do anything */
 488         if (!ns->state) return;
 489
 490         log(EVDNS_LOG_WARN, "Nameserver %s has failed: %s",
 491             debug_ntoa(ns->address), msg);
 492         global_good_nameservers--;
 493         assert(global_good_nameservers >= 0);
 494         if (global_good_nameservers == 0) {
 495                 log(EVDNS_LOG_WARN, "All nameservers have failed");
 496         }
 497
 498         ns->state = 0;
 499         ns->failed_times = 1;
 500
 501         evtimer_set(&ns->timeout_event, nameserver_prod_callback, ns);
 502         if (evtimer_add(&ns->timeout_event, (struct timeval *) &global_nameserver_timeouts[0]) < 0) {
 503                 log(EVDNS_LOG_WARN,
 504                     "Error from libevent when adding timer event for %s",
 505                     debug_ntoa(ns->address));
 506                 /* ???? Do more? */
 507         }
 508
 509         /* walk the list of inflight requests to see if any can be reassigned to */
 510         /* a different server. Requests in the waiting queue don't have a */
 511         /* nameserver assigned yet */
 512
 513         /* if we don't have *any* good nameservers then there's no point */
 514         /* trying to reassign requests to one */
 515         if (!global_good_nameservers) return;
 516
 517         req = req_head;
 518         started_at = req_head;
 519         if (req) {
 520                 do {
 521                         if (req->tx_count == 0 && req->ns == ns) {
 522                                 /* still waiting to go out, can be moved */
 523                                 /* to another server */
 524                                 req->ns = nameserver_pick();
 525                         }
 526                         req = req->next;
 527                 } while (req != started_at);
 528         }
 529 }
 530
 531 static void
 532 nameserver_up(struct nameserver *const ns) {
 533         if (ns->state) return;
 534         log(EVDNS_LOG_WARN, "Nameserver %s is back up",
 535             debug_ntoa(ns->address));
 536         evtimer_del(&ns->timeout_event);
 537         ns->state = 1;
 538         ns->failed_times = 0;
 539         ns->timedout = 0;
 540         global_good_nameservers++;
 541 }
 542
 543 static void
 544 request_trans_id_set(struct request *const req, const u16 trans_id) {
 545         req->trans_id = trans_id;
 546         *((u16 *) req->request) = htons(trans_id);
 547 }
 548
 549 /* Called to remove a request from a list and dealloc it. */
 550 /* head is a pointer to the head of the list it should be */
 551 /* removed from or NULL if the request isn't in a list. */
 552 static void
 553 request_finished(struct request *const req, struct request **head) {
 554         if (head) {
 555                 if (req->next == req) {
 556                         /* only item in the list */
 557                         *head = NULL;
 558                 } else {
 559                         req->next->prev = req->prev;
 560                         req->prev->next = req->next;
 561                         if (*head == req) *head = req->next;
 562                 }
 563         }
 564
 565         log(EVDNS_LOG_DEBUG, "Removing timeout for request %lx",
 566             (unsigned long) req);
 567         evtimer_del(&req->timeout_event);
 568
 569         search_request_finished(req);
 570         global_requests_inflight--;
 571
 572         if (!req->request_appended) {
 573                 /* need to free the request data on it's own */
 574                 free(req->request);
 575         } else {
 576                 /* the request data is appended onto the header */
 577                 /* so everything gets free()ed when we: */
 578         }
 579
 580         free(req);
 581
 582         evdns_requests_pump_waiting_queue();
 583 }
 584
 585 /* This is called when a server returns a funny error code. */
 586 /* We try the request again with another server. */
 587 /* */
 588 /* return: */
 589 /*   0 ok */
 590 /*   1 failed/reissue is pointless */
 591 static int
 592 request_reissue(struct request *req) {
 593         const struct nameserver *const last_ns = req->ns;
 594         /* the last nameserver should have been marked as failing */
 595         /* by the caller of this function, therefore pick will try */
 596         /* not to return it */
 597         req->ns = nameserver_pick();
 598         if (req->ns == last_ns) {
 599                 /* ... but pick did return it */
 600                 /* not a lot of point in trying again with the */
 601                 /* same server */
 602                 return 1;
 603         }
 604
 605         req->reissue_count++;
 606         req->tx_count = 0;
 607         req->transmit_me = 1;
 608
 609         return 0;
 610 }
 611
 612 /* this function looks for space on the inflight queue and promotes */
 613 /* requests from the waiting queue if it can. */
 614 static void
 615 evdns_requests_pump_waiting_queue(void) {
 616         while (global_requests_inflight < global_max_requests_inflight &&
 617             global_requests_waiting) {
 618                 struct request *req;
 619                 /* move a request from the waiting queue to the inflight queue */
 620                 assert(req_waiting_head);
 621                 if (req_waiting_head->next == req_waiting_head) {
 622                         /* only one item in the queue */
 623                         req = req_waiting_head;
 624                         req_waiting_head = NULL;
 625                 } else {
 626                         req = req_waiting_head;
 627                         req->next->prev = req->prev;
 628                         req->prev->next = req->next;
 629                         req_waiting_head = req->next;
 630                 }
 631
 632                 global_requests_waiting--;
 633                 global_requests_inflight++;
 634
 635                 req->ns = nameserver_pick();
 636                 request_trans_id_set(req, transaction_id_pick());
 637
 638                 evdns_request_insert(req, &req_head);
 639                 evdns_request_transmit(req);
 640                 evdns_transmit();
 641         }
 642 }
 643
 644 static void
 645 reply_callback(struct request *const req, u32 ttl, u32 err, struct reply *reply) {
 646         switch (req->request_type) {
 647         case TYPE_A:
 648                 if (reply)
 649                         req->user_callback(DNS_ERR_NONE, DNS_IPv4_A,
 650                                                            reply->data.a.addrcount, ttl,
 651                                                  reply->data.a.addresses,
 652                                                            req->user_pointer);
 653                 else
 654                         req->user_callback(err, 0, 0, 0, NULL, req->user_pointer);
 655                 return;
 656         case TYPE_PTR:
 657                 if (reply) {
 658                         char *name = reply->data.ptr.name;
 659                         req->user_callback(DNS_ERR_NONE, DNS_PTR, 1, ttl,
 660                                                            &name, req->user_pointer);
 661                 } else {
 662                         req->user_callback(err, 0, 0, 0, NULL,
 663                                                            req->user_pointer);
 664                 }
 665                 return;
 666         case TYPE_AAAA:
 667                 if (reply)
 668                         req->user_callback(DNS_ERR_NONE, DNS_IPv6_AAAA,
 669                                                            reply->data.aaaa.addrcount, ttl,
 670                                                            reply->data.aaaa.addresses,
 671                                                            req->user_pointer);
 672                 else
 673                         req->user_callback(err, 0, 0, 0, NULL, req->user_pointer);
 674                 return;
 675         }
 676         assert(0);
 677 }
 678
 679 /* this processes a parsed reply packet */
 680 static void
 681 reply_handle(struct request *const req, u16 flags, u32 ttl, struct reply *reply) {
 682         int error;
 683         static const int error_codes[] = {DNS_ERR_FORMAT, DNS_ERR_SERVERFAILED, DNS_ERR_NOTEXIST, DNS_ERR_NOTIMPL, DNS_ERR_REFUSED};
 684
 685         if (flags & 0x020f || !reply || !reply->have_answer) {
 686                 /* there was an error */
 687                 if (flags & 0x0200) {
 688                         error = DNS_ERR_TRUNCATED;
 689                 } else {
 690                         u16 error_code = (flags & 0x000f) - 1;
 691                         if (error_code > 4) {
 692                                 error = DNS_ERR_UNKNOWN;
 693                         } else {
 694                                 error = error_codes[error_code];
 695                         }
 696                 }
 697
 698                 switch(error) {
 699                 case DNS_ERR_NOTIMPL:
 700                 case DNS_ERR_REFUSED:
 701                         /* we regard these errors as marking a bad nameserver */
 702                         if (req->reissue_count < global_max_reissues) {
 703                                 char msg[64];
 704                                 snprintf(msg, sizeof(msg), "Bad response %d (%s)",
 705                                          error, evdns_err_to_string(error));
 706                                 nameserver_failed(req->ns, msg);
 707                                 if (!request_reissue(req)) return;
 708                         }
 709                         break;
 710                 case DNS_ERR_SERVERFAILED:
 711                         /* rcode 2 (servfailed) sometimes means "we are broken" and
 712                          * sometimes (with some binds) means "that request was very
 713                          * confusing."  Treat this as a timeout, not a failure.
 714                          */
 715                         log(EVDNS_LOG_DEBUG, "Got a SERVERFAILED from nameserver %s; "
 716                                 "will allow the request to time out.",
 717                                 debug_ntoa(req->ns->address));
 718                         break;
 719                 default:
 720                         /* we got a good reply from the nameserver */
 721                         nameserver_up(req->ns);
 722                 }
 723
 724                 if (req->search_state && req->request_type != TYPE_PTR) {
 725                         /* if we have a list of domains to search in, try the next one */
 726                         if (!search_try_next(req)) {
 727                                 /* a new request was issued so this request is finished and */
 728                                 /* the user callback will be made when that request (or a */
 729                                 /* child of it) finishes. */
 730                                 request_finished(req, &req_head);
 731                                 return;
 732                         }
 733                 }
 734
 735                 /* all else failed. Pass the failure up */
 736                 reply_callback(req, 0, error, NULL);
 737                 request_finished(req, &req_head);
 738         } else {
 739                 /* all ok, tell the user */
 740                 reply_callback(req, ttl, 0, reply);
 741                 nameserver_up(req->ns);
 742                 request_finished(req, &req_head);
 743         }
 744 }
 745
 746 static int
 747 name_parse(u8 *packet, int length, int *idx, char *name_out, int name_out_len) {
 748         int name_end = -1;
 749         int j = *idx;
 750         int ptr_count = 0;
 751 #define GET32(x) do { if (j + 4 > length) goto err; memcpy(&_t32, packet + j, 4); j += 4; x = ntohl(_t32); } while(0)
 752 #define GET16(x) do { if (j + 2 > length) goto err; memcpy(&_t, packet + j, 2); j += 2; x = ntohs(_t); } while(0)
 753 #define GET8(x) do { if (j >= length) goto err; x = packet[j++]; } while(0)
 754
 755         char *cp = name_out;
 756         const char *const end = name_out + name_out_len;
 757
 758         /* Normally, names are a series of length prefixed strings terminated */
 759         /* with a length of 0 (the lengths are u8's < 63). */
 760         /* However, the length can start with a pair of 1 bits and that */
 761         /* means that the next 14 bits are a pointer within the current */
 762         /* packet. */
 763
 764         for(;;) {
 765                 u8 label_len;
 766                 if (j >= length) return -1;
 767                 GET8(label_len);
 768                 if (!label_len) break;
 769                 if (label_len & 0xc0) {
 770                         u8 ptr_low;
 771                         GET8(ptr_low);
 772                         if (name_end < 0) name_end = j;
 773                         j = (((int)label_len & 0x3f) << 8) + ptr_low;
 774                         /* Make sure that the target offset is in-bounds. */
 775                         if (j < 0 || j >= length) return -1;
 776                         /* If we've jumped more times than there are characters in the
 777                          * message, we must have a loop. */
 778                         if (++ptr_count > length) return -1;
 779                         continue;
 780                 }
 781                 if (label_len > 63) return -1;
 782                 if (cp != name_out) {
 783                         if (cp + 1 >= end) return -1;
 784                         *cp++ = '.';
 785                 }
 786                 if (cp + label_len >= end) return -1;
 787                 memcpy(cp, packet + j, label_len);
 788                 cp += label_len;
 789                 j += label_len;
 790         }
 791         if (cp >= end) return -1;
 792         *cp = '\0';
 793         if (name_end < 0)
 794                 *idx = j;
 795         else
 796                 *idx = name_end;
 797         return 0;
 798  err:
 799         return -1;
 800 }
 801
 802 /* parses a raw request from a nameserver */
 803 static int
 804 reply_parse(u8 *packet, int length) {
 805         int j = 0;  /* index into packet */
 806         u16 _t;  /* used by the macros */
 807         u32 _t32;  /* used by the macros */
 808         char tmp_name[256]; /* used by the macros */
 809
 810         u16 trans_id, questions, answers, authority, additional, datalength;
 811         u16 flags = 0;
 812         u32 ttl, ttl_r = 0xffffffff;
 813         struct reply reply;
 814         struct request *req = NULL;
 815         unsigned int i;
 816
 817         GET16(trans_id);
 818         GET16(flags);
 819         GET16(questions);
 820         GET16(answers);
 821         GET16(authority);
 822         GET16(additional);
 823         (void) authority; /* suppress "unused variable" warnings. */
 824         (void) additional; /* suppress "unused variable" warnings. */
 825
 826         req = request_find_from_trans_id(trans_id);
 827         if (!req) return -1;
 828
 829         memset(&reply, 0, sizeof(reply));
 830
 831         /* If it's not an answer, it doesn't correspond to any request. */
 832         if (!(flags & 0x8000)) return -1;  /* must be an answer */
 833         if (flags & 0x020f) {
 834                 /* there was an error */
 835                 goto err;
 836         }
 837         /* if (!answers) return; */  /* must have an answer of some form */
 838
 839         /* This macro skips a name in the DNS reply. */
 840 #define SKIP_NAME \
 841         do { tmp_name[0] = '\0';                                        \
 842                 if (name_parse(packet, length, &j, tmp_name, sizeof(tmp_name))<0) \
 843                         goto err;                                                                                                       \
 844         } while(0);
 845
 846         reply.type = req->request_type;
 847
 848         /* skip over each question in the reply */
 849         for (i = 0; i < questions; ++i) {
 850                 /* the question looks like
 851                  *   <label:name><u16:type><u16:class>
 852                  */
 853                 SKIP_NAME;
 854                 j += 4;
 855                 if (j >= length) goto err;
 856         }
 857
 858         /* now we have the answer section which looks like
 859          * <label:name><u16:type><u16:class><u32:ttl><u16:len><data...>
 860          */
 861
 862         for (i = 0; i < answers; ++i) {
 863                 u16 type, class;
 864
 865                 SKIP_NAME;
 866                 GET16(type);
 867                 GET16(class);
 868                 GET32(ttl);
 869                 GET16(datalength);
 870
 871                 if (type == TYPE_A && class == CLASS_INET) {
 872                         int addrcount, addrtocopy;
 873                         if (req->request_type != TYPE_A) {
 874                                 j += datalength; continue;
 875                         }
 876                         if ((datalength & 3) != 0) /* not an even number of As. */
 877                             goto err;
 878                         addrcount = datalength >> 2;
 879                         addrtocopy = MIN(MAX_ADDRS - reply.data.a.addrcount, (unsigned)addrcount);
 880
 881                         ttl_r = MIN(ttl_r, ttl);
 882                         /* we only bother with the first four addresses. */
 883                         if (j + 4*addrtocopy > length) goto err;
 884                         memcpy(&reply.data.a.addresses[reply.data.a.addrcount],
 885                                    packet + j, 4*addrtocopy);
 886                         j += 4*addrtocopy;
 887                         reply.data.a.addrcount += addrtocopy;
 888                         reply.have_answer = 1;
 889                         if (reply.data.a.addrcount == MAX_ADDRS) break;
 890                 } else if (type == TYPE_PTR && class == CLASS_INET) {
 891                         if (req->request_type != TYPE_PTR) {
 892                                 j += datalength; continue;
 893                         }
 894                         if (name_parse(packet, length, &j, reply.data.ptr.name,
 895                                                    sizeof(reply.data.ptr.name))<0)
 896                                 goto err;
 897                         ttl_r = MIN(ttl_r, ttl);
 898                         reply.have_answer = 1;
 899                         break;
 900                 } else if (type == TYPE_AAAA && class == CLASS_INET) {
 901                         int addrcount, addrtocopy;
 902                         if (req->request_type != TYPE_AAAA) {
 903                                 j += datalength; continue;
 904                         }
 905                         if ((datalength & 15) != 0) /* not an even number of AAAAs. */
 906                                 goto err;
 907                         addrcount = datalength >> 4;  /* each address is 16 bytes long */
 908                         addrtocopy = MIN(MAX_ADDRS - reply.data.aaaa.addrcount, (unsigned)addrcount);
 909                         ttl_r = MIN(ttl_r, ttl);
 910
 911                         /* we only bother with the first four addresses. */
 912                         if (j + 16*addrtocopy > length) goto err;
 913                         memcpy(&reply.data.aaaa.addresses[reply.data.aaaa.addrcount],
 914                                    packet + j, 16*addrtocopy);
 915                         reply.data.aaaa.addrcount += addrtocopy;
 916                         j += 16*addrtocopy;
 917                         reply.have_answer = 1;
 918                         if (reply.data.aaaa.addrcount == MAX_ADDRS) break;
 919                 } else {
 920                         /* skip over any other type of resource */
 921                         j += datalength;
 922                 }
 923         }
 924
 925         reply_handle(req, flags, ttl_r, &reply);
 926         return 0;
 927  err:
 928         if (req)
 929                 reply_handle(req, flags, 0, NULL);
 930         return -1;
 931 }
 932
 933 /* Parse a raw request (packet,length) sent to a nameserver port (port) from */
 934 /* a DNS client (addr,addrlen), and if it's well-formed, call the corresponding */
 935 /* callback. */
 936 static int
 937 request_parse(u8 *packet, int length, struct evdns_server_port *port, struct sockaddr *addr, socklen_t addrlen)
 938 {
 939         int j = 0;      /* index into packet */
 940         u16 _t;  /* used by the macros */
 941         char tmp_name[256]; /* used by the macros */
 942
 943         int i;
 944         u16 trans_id, flags, questions, answers, authority, additional;
 945         struct server_request *server_req = NULL;
 946
 947         /* Get the header fields */
 948         GET16(trans_id);
 949         GET16(flags);
 950         GET16(questions);
 951         GET16(answers);
 952         GET16(authority);
 953         GET16(additional);
 954
 955         if (flags & 0x8000) return -1; /* Must not be an answer. */
 956         if (flags & 0x7800) return -1; /* only standard queries are supported */
 957         flags &= 0x0300; /* Only TC and RD get preserved. */
 958
 959         server_req = malloc(sizeof(struct server_request));
 960         if (server_req == NULL) return -1;
 961         memset(server_req, 0, sizeof(struct server_request));
 962
 963         server_req->trans_id = trans_id;
 964         memcpy(&server_req->addr, addr, addrlen);
 965         server_req->addrlen = addrlen;
 966
 967         server_req->base.flags = flags;
 968         server_req->base.nquestions = 0;
 969         server_req->base.questions = malloc(sizeof(struct evdns_server_question *) * questions);
 970         if (server_req->base.questions == NULL)
 971                 goto err;
 972
 973         for (i = 0; i < questions; ++i) {
 974                 u16 type, class;
 975                 struct evdns_server_question *q;
 976                 int namelen;
 977                 if (name_parse(packet, length, &j, tmp_name, sizeof(tmp_name))<0)
 978                         goto err;
 979                 GET16(type);
 980                 GET16(class);
 981                 namelen = strlen(tmp_name);
 982                 q = malloc(sizeof(struct evdns_server_question) + namelen);
 983                 if (!q)
 984                         goto err;
 985                 q->type = type;
 986                 q->class = class;
 987                 memcpy(q->name, tmp_name, namelen+1);
 988                 server_req->base.questions[server_req->base.nquestions++] = q;
 989         }
 990
 991         /* Ignore answers, authority, and additional. */
 992
 993         server_req->port = port;
 994         port->refcnt++;
 995         port->user_callback(&(server_req->base), port->user_data);
 996
 997         return 0;
 998 err:
 999         if (server_req) {
1000                 if (server_req->base.questions) {
1001                         for (i = 0; i < server_req->base.nquestions; ++i)
1002                                 free(server_req->base.questions[i]);
1003                         free(server_req->base.questions);
1004                 }
1005                 free(server_req);
1006         }
1007         return -1;
1008
1009 #undef SKIP_NAME
1010 #undef GET32
1011 #undef GET16
1012 #undef GET8
1013 }
1014
1015 /* Try to choose a strong transaction id which isn't already in flight */
1016 static u16
1017 transaction_id_pick(void) {
1018         for (;;) {
1019                 const struct request *req = req_head, *started_at;
1020 #ifdef DNS_USE_CPU_CLOCK_FOR_ID
1021                 struct timespec ts;
1022                 u16 trans_id;
1023 #ifdef CLOCK_MONOTONIC
1024                 if (clock_gettime(CLOCK_MONOTONIC, &ts) == -1)
1025 #else
1026                 if (clock_gettime(CLOCK_REALTIME, &ts) == -1)
1027 #endif
1028                         event_err(1, "clock_gettime");
1029                 trans_id = ts.tv_nsec & 0xffff;
1030 #endif
1031
1032 #ifdef DNS_USE_GETTIMEOFDAY_FOR_ID
1033                 struct timeval tv;
1034                 u16 trans_id;
1035                 gettimeofday(&tv, NULL);
1036                 trans_id = tv.tv_usec & 0xffff;
1037 #endif
1038
1039 #ifdef DNS_USE_OPENSSL_FOR_ID
1040                 u16 trans_id;
1041                 if (RAND_pseudo_bytes((u8 *) &trans_id, 2) == -1) {
1042                         /* in the case that the RAND call fails we back */
1043                         /* down to using gettimeofday. */
1044                         struct timeval tv;
1045                         gettimeofday(&tv, NULL);
1046                         trans_id = tv.tv_usec & 0xffff; */
1047                         abort();
1048                 }
1049 #endif
1050
1051                 if (trans_id == 0xffff) continue;
1052                 /* now check to see if that id is already inflight */
1053                 req = started_at = req_head;
1054                 if (req) {
1055                         do {
1056                                 if (req->trans_id == trans_id) break;
1057                                 req = req->next;
1058                         } while (req != started_at);
1059                 }
1060                 /* we didn't find it, so this is a good id */
1061                 if (req == started_at) return trans_id;
1062         }
1063 }
1064
1065 /* choose a namesever to use. This function will try to ignore */
1066 /* nameservers which we think are down and load balance across the rest */
1067 /* by updating the server_head global each time. */
1068 static struct nameserver *
1069 nameserver_pick(void) {
1070         struct nameserver *started_at = server_head, *picked;
1071         if (!server_head) return NULL;
1072
1073         /* if we don't have any good nameservers then there's no */
1074         /* point in trying to find one. */
1075         if (!global_good_nameservers) {
1076                 server_head = server_head->next;
1077                 return server_head;
1078         }
1079
1080         /* remember that nameservers are in a circular list */
1081         for (;;) {
1082                 if (server_head->state) {
1083                         /* we think this server is currently good */
1084                         picked = server_head;
1085                         server_head = server_head->next;
1086                         return picked;
1087                 }
1088
1089                 server_head = server_head->next;
1090                 if (server_head == started_at) {
1091                         /* all the nameservers seem to be down */
1092                         /* so we just return this one and hope for the */
1093                         /* best */
1094                         assert(global_good_nameservers == 0);
1095                         picked = server_head;
1096                         server_head = server_head->next;
1097                         return picked;
1098                 }
1099         }
1100 }
1101
1102 /* this is called when a namesever socket is ready for reading */
1103 static void
1104 nameserver_read(struct nameserver *ns) {
1105         u8 packet[1500];
1106
1107         for (;;) {
1108                 const int r = recv(ns->socket, packet, sizeof(packet), 0);
1109                 if (r < 0) {
1110                         int err = last_error(ns->socket);
1111                         if (error_is_eagain(err)) return;
1112                         nameserver_failed(ns, strerror(err));
1113                         return;
1114                 }
1115                 ns->timedout = 0;
1116                 reply_parse(packet, r);
1117         }
1118 }
1119
1120 /* Read a packet from a DNS client on a server port s, parse it, and */
1121 /* act accordingly. */
1122 static void
1123 server_port_read(struct evdns_server_port *s) {
1124         u8 packet[1500];
1125         struct sockaddr_storage addr;
1126         socklen_t addrlen;
1127         int r;
1128
1129         for (;;) {
1130                 addrlen = sizeof(struct sockaddr_storage);
1131                 r = recvfrom(s->socket, packet, sizeof(packet), 0,
1132                                          (struct sockaddr*) &addr, &addrlen);
1133                 if (r < 0) {
1134                         int err = last_error(s->socket);
1135                         if (error_is_eagain(err)) return;
1136                         log(EVDNS_LOG_WARN, "Error %s (%d) while reading request.",
1137                                 strerror(err), err);
1138                         return;
1139                 }
1140                 request_parse(packet, r, s, (struct sockaddr*) &addr, addrlen);
1141         }
1142 }
1143
1144 /* Try to write all pending replies on a given DNS server port. */
1145 static void
1146 server_port_flush(struct evdns_server_port *port)
1147 {
1148         while (port->pending_replies) {
1149                 struct server_request *req = port->pending_replies;
1150                 int r = sendto(port->socket, req->response, req->response_len, 0,
1151                            (struct sockaddr*) &req->addr, req->addrlen);
1152                 if (r < 0) {
1153                         int err = last_error(port->socket);
1154                         if (error_is_eagain(err))
1155                                 return;
1156                         log(EVDNS_LOG_WARN, "Error %s (%d) while writing response to port; dropping", strerror(err), err);
1157                 }
1158                 if (server_request_free(req)) {
1159                         /* we released the last reference to req->port. */
1160                         return;
1161                 }
1162         }
1163
1164         /* We have no more pending requests; stop listening for 'writeable' events. */
1165         (void) event_del(&port->event);
1166         event_set(&port->event, port->socket, EV_READ | EV_PERSIST,
1167                           server_port_ready_callback, port);
1168         if (event_add(&port->event, NULL) < 0) {
1169                 log(EVDNS_LOG_WARN, "Error from libevent when adding event for DNS server.");
1170                 /* ???? Do more? */
1171         }
1172 }
1173
1174 /* set if we are waiting for the ability to write to this server. */
1175 /* if waiting is true then we ask libevent for EV_WRITE events, otherwise */
1176 /* we stop these events. */
1177 static void
1178 nameserver_write_waiting(struct nameserver *ns, char waiting) {
1179         if (ns->write_waiting == waiting) return;
1180
1181         ns->write_waiting = waiting;
1182         (void) event_del(&ns->event);
1183         event_set(&ns->event, ns->socket, EV_READ | (waiting ? EV_WRITE : 0) | EV_PERSIST,
1184                         nameserver_ready_callback, ns);
1185         if (event_add(&ns->event, NULL) < 0) {
1186           log(EVDNS_LOG_WARN, "Error from libevent when adding event for %s",
1187               debug_ntoa(ns->address));
1188           /* ???? Do more? */
1189         }
1190 }
1191
1192 /* a callback function. Called by libevent when the kernel says that */
1193 /* a nameserver socket is ready for writing or reading */
1194 static void
1195 nameserver_ready_callback(int fd, short events, void *arg) {
1196         struct nameserver *ns = (struct nameserver *) arg;
1197         (void)fd;
1198
1199         if (events & EV_WRITE) {
1200                 ns->choked = 0;
1201                 if (!evdns_transmit()) {
1202                         nameserver_write_waiting(ns, 0);
1203                 }
1204         }
1205         if (events & EV_READ) {
1206                 nameserver_read(ns);
1207         }
1208 }
1209
1210 /* a callback function. Called by libevent when the kernel says that */
1211 /* a server socket is ready for writing or reading. */
1212 static void
1213 server_port_ready_callback(int fd, short events, void *arg) {
1214         struct evdns_server_port *port = (struct evdns_server_port *) arg;
1215         (void) fd;
1216
1217         if (events & EV_WRITE) {
1218                 port->choked = 0;
1219                 server_port_flush(port);
1220         }
1221         if (events & EV_READ) {
1222                 server_port_read(port);
1223         }
1224 }
1225
1226 /* This is an inefficient representation; only use it via the dnslabel_table_*
1227  * functions, so that is can be safely replaced with something smarter later. */
1228 #define MAX_LABELS 128
1229 /* Structures used to implement name compression */
1230 struct dnslabel_entry { char *v; off_t pos; };
1231 struct dnslabel_table {
1232         int n_labels; /* number of current entries */
1233         /* map from name to position in message */
1234         struct dnslabel_entry labels[MAX_LABELS];
1235 };
1236
1237 /* Initialize dnslabel_table. */
1238 static void
1239 dnslabel_table_init(struct dnslabel_table *table)
1240 {
1241         table->n_labels = 0;
1242 }
1243
1244 /* Free all storage held by table, but not the table itself. */
1245 static void
1246 dnslabel_clear(struct dnslabel_table *table)
1247 {
1248         int i;
1249         for (i = 0; i < table->n_labels; ++i)
1250                 free(table->labels[i].v);
1251         table->n_labels = 0;
1252 }
1253
1254 /* return the position of the label in the current message, or -1 if the label */
1255 /* hasn't been used yet. */
1256 static int
1257 dnslabel_table_get_pos(const struct dnslabel_table *table, const char *label)
1258 {
1259         int i;
1260         for (i = 0; i < table->n_labels; ++i) {
1261                 if (!strcmp(label, table->labels[i].v))
1262                         return table->labels[i].pos;
1263         }
1264         return -1;
1265 }
1266
1267 /* remember that we've used the label at position pos */
1268 static int
1269 dnslabel_table_add(struct dnslabel_table *table, const char *label, off_t pos)
1270 {
1271         char *v;
1272         int p;
1273         if (table->n_labels == MAX_LABELS)
1274                 return (-1);
1275         v = strdup(label);
1276         if (v == NULL)
1277                 return (-1);
1278         p = table->n_labels++;
1279         table->labels[p].v = v;
1280         table->labels[p].pos = pos;
1281
1282         return (0);
1283 }
1284
1285 /* Converts a string to a length-prefixed set of DNS labels, starting */
1286 /* at buf[j]. name and buf must not overlap. name_len should be the length */
1287 /* of name.      table is optional, and is used for compression. */
1288 /* */
1289 /* Input: abc.def */
1290 /* Output: <3>abc<3>def<0> */
1291 /* */
1292 /* Returns the first index after the encoded name, or negative on error. */
1293 /*       -1      label was > 63 bytes */
1294 /*       -2      name too long to fit in buffer. */
1295 /* */
1296 static off_t
1297 dnsname_to_labels(u8 *const buf, size_t buf_len, off_t j,
1298                                   const char *name, const int name_len,
1299                                   struct dnslabel_table *table) {
1300         const char *end = name + name_len;
1301         int ref = 0;
1302         u16 _t;
1303
1304 #define APPEND16(x) do {                                                   \
1305                 if (j + 2 > (off_t)buf_len)                                \
1306                         goto overflow;                                             \
1307                 _t = htons(x);                                                     \
1308                 memcpy(buf + j, &_t, 2);                                   \
1309                 j += 2;                                                                    \
1310         } while (0)
1311 #define APPEND32(x) do {                                                   \
1312                 if (j + 4 > (off_t)buf_len)                                \
1313                         goto overflow;                                             \
1314                 _t32 = htonl(x);                                                   \
1315                 memcpy(buf + j, &_t32, 4);                                 \
1316                 j += 4;                                                                    \
1317         } while (0)
1318
1319         if (name_len > 255) return -2;
1320
1321         for (;;) {
1322                 const char *const start = name;
1323                 if (table && (ref = dnslabel_table_get_pos(table, name)) >= 0) {
1324                         APPEND16(ref | 0xc000);
1325                         return j;
1326                 }
1327                 name = strchr(name, '.');
1328                 if (!name) {
1329                         const unsigned int label_len = end - start;
1330                         if (label_len > 63) return -1;
1331                         if ((size_t)(j+label_len+1) > buf_len) return -2;
1332                         if (table) dnslabel_table_add(table, start, j);
1333                         buf[j++] = label_len;
1334
1335                         memcpy(buf + j, start, end - start);
1336                         j += end - start;
1337                         break;
1338                 } else {
1339                         /* append length of the label. */
1340                         const unsigned int label_len = name - start;
1341                         if (label_len > 63) return -1;
1342                         if ((size_t)(j+label_len+1) > buf_len) return -2;
1343                         if (table) dnslabel_table_add(table, start, j);
1344                         buf[j++] = label_len;
1345
1346                         memcpy(buf + j, start, name - start);
1347                         j += name - start;
1348                         /* hop over the '.' */
1349                         name++;
1350                 }
1351         }
1352
1353         /* the labels must be terminated by a 0. */
1354         /* It's possible that the name ended in a . */
1355         /* in which case the zero is already there */
1356         if (!j || buf[j-1]) buf[j++] = 0;
1357         return j;
1358  overflow:
1359         return (-2);
1360 }
1361
1362 /* Finds the length of a dns request for a DNS name of the given */
1363 /* length. The actual request may be smaller than the value returned */
1364 /* here */
1365 static int
1366 evdns_request_len(const int name_len) {
1367         return 96 + /* length of the DNS standard header */
1368                 name_len + 2 +
1369                 4;  /* space for the resource type */
1370 }
1371
1372 /* build a dns request packet into buf. buf should be at least as long */
1373 /* as evdns_request_len told you it should be. */
1374 /* */
1375 /* Returns the amount of space used. Negative on error. */
1376 static int
1377 evdns_request_data_build(const char *const name, const int name_len,
1378     const u16 trans_id, const u16 type, const u16 class,
1379     u8 *const buf, size_t buf_len) {
1380         off_t j = 0;  /* current offset into buf */
1381         u16 _t;  /* used by the macros */
1382
1383         APPEND16(trans_id);
1384         APPEND16(0x0100);  /* standard query, recusion needed */
1385         APPEND16(1);  /* one question */
1386         APPEND16(0);  /* no answers */
1387         APPEND16(0);  /* no authority */
1388         APPEND16(0);  /* no additional */
1389
1390         j = dnsname_to_labels(buf, buf_len, j, name, name_len, NULL);
1391         if (j < 0) {
1392                 return (int)j;
1393         }
1394
1395         APPEND16(type);
1396         APPEND16(class);
1397
1398         return (int)j;
1399  overflow:
1400         return (-1);
1401 }
1402
1403 /* exported function */
1404 struct evdns_server_port *
1405 evdns_add_server_port(int socket, int is_tcp, evdns_request_callback_fn_type cb, void *user_data)
1406 {
1407         struct evdns_server_port *port;
1408         if (!(port = malloc(sizeof(struct evdns_server_port))))
1409                 return NULL;
1410         memset(port, 0, sizeof(struct evdns_server_port));
1411
1412         assert(!is_tcp); /* TCP sockets not yet implemented */
1413         port->socket = socket;
1414         port->refcnt = 1;
1415         port->choked = 0;
1416         port->closing = 0;
1417         port->user_callback = cb;
1418         port->user_data = user_data;
1419         port->pending_replies = NULL;
1420
1421         event_set(&port->event, port->socket, EV_READ | EV_PERSIST,
1422                           server_port_ready_callback, port);
1423         event_add(&port->event, NULL); /* check return. */
1424         return port;
1425 }
1426
1427 /* exported function */
1428 void
1429 evdns_close_server_port(struct evdns_server_port *port)
1430 {
1431         if (--port->refcnt == 0)
1432                 server_port_free(port);
1433         port->closing = 1;
1434 }
1435
1436 /* exported function */
1437 int
1438 evdns_server_request_add_reply(struct evdns_server_request *_req, int section, const char *name, int type, int class, int ttl, int datalen, int is_name, const char *data)
1439 {
1440         struct server_request *req = TO_SERVER_REQUEST(_req);
1441         struct server_reply_item **itemp, *item;
1442         int *countp;
1443
1444         if (req->response) /* have we already answered? */
1445                 return (-1);
1446
1447         switch (section) {
1448         case EVDNS_ANSWER_SECTION:
1449                 itemp = &req->answer;
1450                 countp = &req->n_answer;
1451                 break;
1452         case EVDNS_AUTHORITY_SECTION:
1453                 itemp = &req->authority;
1454                 countp = &req->n_authority;
1455                 break;
1456         case EVDNS_ADDITIONAL_SECTION:
1457                 itemp = &req->additional;
1458                 countp = &req->n_additional;
1459                 break;
1460         default:
1461                 return (-1);
1462         }
1463         while (*itemp) {
1464                 itemp = &((*itemp)->next);
1465         }
1466         item = malloc(sizeof(struct server_reply_item));
1467         if (!item)
1468                 return -1;
1469         item->next = NULL;
1470         if (!(item->name = strdup(name))) {
1471                 free(item);
1472                 return -1;
1473         }
1474         item->type = type;
1475         item->class = class;
1476         item->ttl = ttl;
1477         item->is_name = is_name != 0;
1478         item->datalen = 0;
1479         item->data = NULL;
1480         if (data) {
1481                 if (item->is_name) {
1482                         if (!(item->data = strdup(data))) {
1483                                 free(item->name);
1484                                 free(item);
1485                                 return -1;
1486                         }
1487                         item->datalen = (u16)-1;
1488                 } else {
1489                         if (!(item->data = malloc(datalen))) {
1490                                 free(item->name);
1491                                 free(item);
1492                                 return -1;
1493                         }
1494                         item->datalen = datalen;
1495                         memcpy(item->data, data, datalen);
1496                 }
1497         }
1498
1499         *itemp = item;
1500         ++(*countp);
1501         return 0;
1502 }
1503
1504 /* exported function */
1505 int
1506 evdns_server_request_add_a_reply(struct evdns_server_request *req, const char *name, int n, void *addrs, int ttl)
1507 {
1508         return evdns_server_request_add_reply(
1509                   req, EVDNS_ANSWER_SECTION, name, TYPE_A, CLASS_INET,
1510                   ttl, n*4, 0, addrs);
1511 }
1512
1513 /* exported function */
1514 int
1515 evdns_server_request_add_aaaa_reply(struct evdns_server_request *req, const char *name, int n, void *addrs, int ttl)
1516 {
1517         return evdns_server_request_add_reply(
1518                   req, EVDNS_ANSWER_SECTION, name, TYPE_AAAA, CLASS_INET,
1519                   ttl, n*16, 0, addrs);
1520 }
1521
1522 /* exported function */
1523 int
1524 evdns_server_request_add_ptr_reply(struct evdns_server_request *req, struct in_addr *in, const char *inaddr_name, const char *hostname, int ttl)
1525 {
1526         u32 a;
1527         char buf[32];
1528         assert(in || inaddr_name);
1529         assert(!(in && inaddr_name));
1530         if (in) {
1531                 a = ntohl(in->s_addr);
1532                 snprintf(buf, sizeof(buf), "%d.%d.%d.%d.in-addr.arpa",
1533                                 (int)(u8)((a    )&0xff),
1534                                 (int)(u8)((a>>8 )&0xff),
1535                                 (int)(u8)((a>>16)&0xff),
1536                                 (int)(u8)((a>>24)&0xff));
1537                 inaddr_name = buf;
1538         }
1539         return evdns_server_request_add_reply(
1540                   req, EVDNS_ANSWER_SECTION, inaddr_name, TYPE_PTR, CLASS_INET,
1541                   ttl, -1, 1, hostname);
1542 }
1543
1544 /* exported function */
1545 int
1546 evdns_server_request_add_cname_reply(struct evdns_server_request *req, const char *name, const char *cname, int ttl)
1547 {
1548         return evdns_server_request_add_reply(
1549                   req, EVDNS_ANSWER_SECTION, name, TYPE_A, CLASS_INET,
1550                   ttl, -1, 1, cname);
1551 }
1552
1553
1554 static int
1555 evdns_server_request_format_response(struct server_request *req, int err)
1556 {
1557         unsigned char buf[1500];
1558         size_t buf_len = sizeof(buf);
1559         off_t j = 0, r;
1560         u16 _t;
1561         u32 _t32;
1562         int i;
1563         u16 flags;
1564         struct dnslabel_table table;
1565
1566         if (err < 0 || err > 15) return -1;
1567
1568         /* Set response bit and error code; copy OPCODE and RD fields from
1569          * question; copy RA and AA if set by caller. */
1570         flags = req->base.flags;
1571         flags |= (0x8000 | err);
1572
1573         dnslabel_table_init(&table);
1574         APPEND16(req->trans_id);
1575         APPEND16(flags);
1576         APPEND16(req->base.nquestions);
1577         APPEND16(req->n_answer);
1578         APPEND16(req->n_authority);
1579         APPEND16(req->n_additional);
1580
1581         /* Add questions. */
1582         for (i=0; i < req->base.nquestions; ++i) {
1583                 const char *s = req->base.questions[i]->name;
1584                 j = dnsname_to_labels(buf, buf_len, j, s, strlen(s), &table);
1585                 if (j < 0) {
1586                         dnslabel_clear(&table);
1587                         return (int) j;
1588                 }
1589                 APPEND16(req->base.questions[i]->type);
1590                 APPEND16(req->base.questions[i]->class);
1591         }
1592
1593         /* Add answer, authority, and additional sections. */
1594         for (i=0; i<3; ++i) {
1595                 struct server_reply_item *item;
1596                 if (i==0)
1597                         item = req->answer;
1598                 else if (i==1)
1599                         item = req->authority;
1600                 else
1601                         item = req->additional;
1602                 while (item) {
1603                         r = dnsname_to_labels(buf, buf_len, j, item->name, strlen(item->name), &table);
1604                         if (r < 0)
1605                                 goto overflow;
1606                         j = r;
1607
1608                         APPEND16(item->type);
1609                         APPEND16(item->class);
1610                         APPEND32(item->ttl);
1611                         if (item->is_name) {
1612                                 off_t len_idx = j, name_start;
1613                                 j += 2;
1614                                 name_start = j;
1615                                 r = dnsname_to_labels(buf, buf_len, j, item->data, strlen(item->data), &table);
1616                                 if (r < 0)
1617                                         goto overflow;
1618                                 j = r;
1619                                 _t = htons( (j-name_start) );
1620                                 memcpy(buf+len_idx, &_t, 2);
1621                         } else {
1622                                 APPEND16(item->datalen);
1623                                 if (j+item->datalen > (off_t)buf_len)
1624                                         goto overflow;
1625                                 memcpy(buf+j, item->data, item->datalen);
1626                                 j += item->datalen;
1627                         }
1628                         item = item->next;
1629                 }
1630         }
1631
1632         if (j > 512) {
1633 overflow:
1634                 j = 512;
1635                 buf[3] |= 0x02; /* set the truncated bit. */
1636         }
1637
1638         req->response_len = j;
1639
1640         if (!(req->response = malloc(req->response_len))) {
1641                 server_request_free_answers(req);
1642                 dnslabel_clear(&table);
1643                 return (-1);
1644         }
1645         memcpy(req->response, buf, req->response_len);
1646         server_request_free_answers(req);
1647         dnslabel_clear(&table);
1648         return (0);
1649 }
1650
1651 /* exported function */
1652 int
1653 evdns_server_request_respond(struct evdns_server_request *_req, int err)
1654 {
1655         struct server_request *req = TO_SERVER_REQUEST(_req);
1656         struct evdns_server_port *port = req->port;
1657         int r;
1658         if (!req->response) {
1659                 if ((r = evdns_server_request_format_response(req, err))<0)
1660                         return r;
1661         }
1662
1663         r = sendto(port->socket, req->response, req->response_len, 0,
1664                            (struct sockaddr*) &req->addr, req->addrlen);
1665         if (r<0) {
1666                 int err = last_error(port->socket);
1667                 if (! error_is_eagain(err))
1668                         return -1;
1669
1670                 if (port->pending_replies) {
1671                         req->prev_pending = port->pending_replies->prev_pending;
1672                         req->next_pending = port->pending_replies;
1673                         req->prev_pending->next_pending =
1674                                 req->next_pending->prev_pending = req;
1675                 } else {
1676                         req->prev_pending = req->next_pending = req;
1677                         port->pending_replies = req;
1678                         port->choked = 1;
1679
1680                         (void) event_del(&port->event);
1681                         event_set(&port->event, port->socket, (port->closing?0:EV_READ) | EV_WRITE | EV_PERSIST, server_port_ready_callback, port);
1682
1683                         if (event_add(&port->event, NULL) < 0) {
1684                                 log(EVDNS_LOG_WARN, "Error from libevent when adding event for DNS server");
1685                         }
1686
1687                 }
1688
1689                 return 1;
1690         }
1691         if (server_request_free(req))
1692                 return 0;
1693
1694         if (port->pending_replies)
1695                 server_port_flush(port);
1696
1697         return 0;
1698 }
1699
1700 /* Free all storage held by RRs in req. */
1701 static void
1702 server_request_free_answers(struct server_request *req)
1703 {
1704         struct server_reply_item *victim, *next, **list;
1705         int i;
1706         for (i = 0; i < 3; ++i) {
1707                 if (i==0)
1708                         list = &req->answer;
1709                 else if (i==1)
1710                         list = &req->authority;
1711                 else
1712                         list = &req->additional;
1713
1714                 victim = *list;
1715                 while (victim) {
1716                         next = victim->next;
1717                         free(victim->name);
1718                         if (victim->data)
1719                                 free(victim->data);
1720                         free(victim);
1721                         victim = next;
1722                 }
1723                 *list = NULL;
1724         }
1725 }
1726
1727 /* Free all storage held by req, and remove links to it. */
1728 /* return true iff we just wound up freeing the server_port. */
1729 static int
1730 server_request_free(struct server_request *req)
1731 {
1732         int i, rc=1;
1733         if (req->base.questions) {
1734                 for (i = 0; i < req->base.nquestions; ++i)
1735                         free(req->base.questions[i]);
1736                 free(req->base.questions);
1737         }
1738
1739         if (req->port) {
1740                 if (req->port->pending_replies == req) {
1741                         if (req->next_pending)
1742                                 req->port->pending_replies = req->next_pending;
1743                         else
1744                                 req->port->pending_replies = NULL;
1745                 }
1746                 rc = --req->port->refcnt;
1747         }
1748
1749         if (req->response) {
1750                 free(req->response);
1751         }
1752
1753         server_request_free_answers(req);
1754
1755         if (req->next_pending && req->next_pending != req) {
1756                 req->next_pending->prev_pending = req->prev_pending;
1757                 req->prev_pending->next_pending = req->next_pending;
1758         }
1759
1760         if (rc == 0) {
1761                 server_port_free(req->port);
1762                 free(req);
1763                 return (1);
1764         }
1765         free(req);
1766         return (0);
1767 }
1768
1769 /* Free all storage held by an evdns_server_port.  Only called when  */
1770 static void
1771 server_port_free(struct evdns_server_port *port)
1772 {
1773         assert(port);
1774         assert(!port->refcnt);
1775         assert(!port->pending_replies);
1776         if (port->socket > 0) {
1777                 CLOSE_SOCKET(port->socket);
1778                 port->socket = -1;
1779         }
1780         (void) event_del(&port->event);
1781         /* XXXX actually free the port? -NM */
1782 }
1783
1784 /* exported function */
1785 int
1786 evdns_server_request_drop(struct evdns_server_request *_req)
1787 {
1788         struct server_request *req = TO_SERVER_REQUEST(_req);
1789         server_request_free(req);
1790         return 0;
1791 }
1792
1793 /* exported function */
1794 int
1795 evdns_server_request_get_requesting_addr(struct evdns_server_request *_req, struct sockaddr *sa, int addr_len)
1796 {
1797         struct server_request *req = TO_SERVER_REQUEST(_req);
1798         if (addr_len < (int)req->addrlen)
1799                 return -1;
1800         memcpy(sa, &(req->addr), req->addrlen);
1801         return req->addrlen;
1802 }
1803
1804 #undef APPEND16
1805 #undef APPEND32
1806
1807 /* this is a libevent callback function which is called when a request */
1808 /* has timed out. */
1809 static void
1810 evdns_request_timeout_callback(int fd, short events, void *arg) {
1811         struct request *const req = (struct request *) arg;
1812         (void) fd;
1813         (void) events;
1814
1815         log(EVDNS_LOG_DEBUG, "Request %lx timed out", (unsigned long) arg);
1816
1817         req->ns->timedout++;
1818         if (req->ns->timedout > global_max_nameserver_timeout) {
1819                 req->ns->timedout = 0;
1820                 nameserver_failed(req->ns, "request timed out.");
1821         }
1822
1823         (void) evtimer_del(&req->timeout_event);
1824         if (req->tx_count >= global_max_retransmits) {
1825                 /* this request has failed */
1826                 reply_callback(req, 0, DNS_ERR_TIMEOUT, NULL);
1827                 request_finished(req, &req_head);
1828         } else {
1829                 /* retransmit it */
1830                 evdns_request_transmit(req);
1831         }
1832 }
1833
1834 /* try to send a request to a given server. */
1835 /* */
1836 /* return: */
1837 /*   0 ok */
1838 /*   1 temporary failure */
1839 /*   2 other failure */
1840 static int
1841 evdns_request_transmit_to(struct request *req, struct nameserver *server) {
1842         const int r = send(server->socket, req->request, req->request_len, 0);
1843         if (r < 0) {
1844                 int err = last_error(server->socket);
1845                 if (error_is_eagain(err)) return 1;
1846                 nameserver_failed(req->ns, strerror(err));
1847                 return 2;
1848         } else if (r != (int)req->request_len) {
1849                 return 1;  /* short write */
1850         } else {
1851                 return 0;
1852         }
1853 }
1854
1855 /* try to send a request, updating the fields of the request */
1856 /* as needed */
1857 /* */
1858 /* return: */
1859 /*   0 ok */
1860 /*   1 failed */
1861 static int
1862 evdns_request_transmit(struct request *req) {
1863         int retcode = 0, r;
1864
1865         /* if we fail to send this packet then this flag marks it */
1866         /* for evdns_transmit */
1867         req->transmit_me = 1;
1868         if (req->trans_id == 0xffff) abort();
1869
1870         if (req->ns->choked) {
1871                 /* don't bother trying to write to a socket */
1872                 /* which we have had EAGAIN from */
1873                 return 1;
1874         }
1875
1876         r = evdns_request_transmit_to(req, req->ns);
1877         switch (r) {
1878         case 1:
1879                 /* temp failure */
1880                 req->ns->choked = 1;
1881                 nameserver_write_waiting(req->ns, 1);
1882                 return 1;
1883         case 2:
1884                 /* failed in some other way */
1885                 retcode = 1;
1886                 /* fall through */
1887         default:
1888                 /* all ok */
1889                 log(EVDNS_LOG_DEBUG,
1890                     "Setting timeout for request %lx", (unsigned long) req);
1891                 evtimer_set(&req->timeout_event, evdns_request_timeout_callback, req);
1892                 if (evtimer_add(&req->timeout_event, &global_timeout) < 0) {
1893                   log(EVDNS_LOG_WARN,
1894                       "Error from libevent when adding timer for request %lx",
1895                       (unsigned long) req);
1896                   /* ???? Do more? */
1897                 }
1898                 req->tx_count++;
1899                 req->transmit_me = 0;
1900                 return retcode;
1901         }
1902 }
1903
1904 static void
1905 nameserver_probe_callback(int result, char type, int count, int ttl, void *addresses, void *arg) {
1906         struct nameserver *const ns = (struct nameserver *) arg;
1907         (void) type;
1908         (void) count;
1909         (void) ttl;
1910         (void) addresses;
1911
1912         if (result == DNS_ERR_NONE || result == DNS_ERR_NOTEXIST) {
1913                 /* this is a good reply */
1914                 nameserver_up(ns);
1915         } else nameserver_probe_failed(ns);
1916 }
1917
1918 static void
1919 nameserver_send_probe(struct nameserver *const ns) {
1920         struct request *req;
1921         /* here we need to send a probe to a given nameserver */
1922         /* in the hope that it is up now. */
1923
1924         log(EVDNS_LOG_DEBUG, "Sending probe to %s", debug_ntoa(ns->address));
1925
1926         req = request_new(TYPE_A, "www.google.com", DNS_QUERY_NO_SEARCH, nameserver_probe_callback, ns);
1927         if (!req) return;
1928         /* we force this into the inflight queue no matter what */
1929         request_trans_id_set(req, transaction_id_pick());
1930         req->ns = ns;
1931         request_submit(req);
1932 }
1933
1934 /* returns: */
1935 /*   0 didn't try to transmit anything */
1936 /*   1 tried to transmit something */
1937 static int
1938 evdns_transmit(void) {
1939         char did_try_to_transmit = 0;
1940
1941         if (req_head) {
1942                 struct request *const started_at = req_head, *req = req_head;
1943                 /* first transmit all the requests which are currently waiting */
1944                 do {
1945                         if (req->transmit_me) {
1946                                 did_try_to_transmit = 1;
1947                                 evdns_request_transmit(req);
1948                         }
1949
1950                         req = req->next;
1951                 } while (req != started_at);
1952         }
1953
1954         return did_try_to_transmit;
1955 }
1956
1957 /* exported function */
1958 int
1959 evdns_count_nameservers(void)
1960 {
1961         const struct nameserver *server = server_head;
1962         int n = 0;
1963         if (!server)
1964                 return 0;
1965         do {
1966                 ++n;
1967                 server = server->next;
1968         } while (server != server_head);
1969         return n;
1970 }
1971
1972 /* exported function */
1973 int
1974 evdns_clear_nameservers_and_suspend(void)
1975 {
1976         struct nameserver *server = server_head, *started_at = server_head;
1977         struct request *req = req_head, *req_started_at = req_head;
1978
1979         if (!server)
1980                 return 0;
1981         while (1) {
1982                 struct nameserver *next = server->next;
1983                 (void) event_del(&server->event);
1984                 (void) evtimer_del(&server->timeout_event);
1985                 if (server->socket >= 0)
1986                         CLOSE_SOCKET(server->socket);
1987                 free(server);
1988                 if (next == started_at)
1989                         break;
1990                 server = next;
1991         }
1992         server_head = NULL;
1993         global_good_nameservers = 0;
1994
1995         while (req) {
1996                 struct request *next = req->next;
1997                 req->tx_count = req->reissue_count = 0;
1998                 req->ns = NULL;
1999                 /* ???? What to do about searches? */
2000                 (void) evtimer_del(&req->timeout_event);
2001                 req->trans_id = 0;
2002                 req->transmit_me = 0;
2003
2004                 global_requests_waiting++;
2005                 evdns_request_insert(req, &req_waiting_head);
2006                 /* We want to insert these suspended elements at the front of
2007                  * the waiting queue, since they were pending before any of
2008                  * the waiting entries were added.  This is a circular list,
2009                  * so we can just shift the start back by one.*/
2010                 req_waiting_head = req_waiting_head->prev;
2011
2012                 if (next == req_started_at)
2013                         break;
2014                 req = next;
2015         }
2016         req_head = NULL;
2017         global_requests_inflight = 0;
2018
2019         return 0;
2020 }
2021
2022
2023 /* exported function */
2024 int
2025 evdns_resume(void)
2026 {
2027         evdns_requests_pump_waiting_queue();
2028         return 0;
2029 }
2030
2031 static int
2032 _evdns_nameserver_add_impl(unsigned long int address, int port) {
2033         /* first check to see if we already have this nameserver */
2034
2035         const struct nameserver *server = server_head, *const started_at = server_head;
2036         struct nameserver *ns;
2037         struct sockaddr_in sin;
2038         int err = 0;
2039         if (server) {
2040                 do {
2041                         if (server->address == address) return 3;
2042                         server = server->next;
2043                 } while (server != started_at);
2044         }
2045
2046         ns = (struct nameserver *) malloc(sizeof(struct nameserver));
2047         if (!ns) return -1;
2048
2049         memset(ns, 0, sizeof(struct nameserver));
2050
2051         ns->socket = socket(PF_INET, SOCK_DGRAM, 0);
2052         if (ns->socket < 0) { err = 1; goto out1; }
2053 #ifdef WIN32
2054         {
2055                 u_long nonblocking = 1;
2056                 ioctlsocket(ns->socket, FIONBIO, &nonblocking);
2057         }
2058 #else
2059         fcntl(ns->socket, F_SETFL, O_NONBLOCK);
2060 #endif
2061         sin.sin_addr.s_addr = address;
2062         sin.sin_port = htons(port);
2063         sin.sin_family = AF_INET;
2064         if (connect(ns->socket, (struct sockaddr *) &sin, sizeof(sin)) != 0) {
2065                 err = 2;
2066                 goto out2;
2067         }
2068
2069         ns->address = address;
2070         ns->state = 1;
2071         event_set(&ns->event, ns->socket, EV_READ | EV_PERSIST, nameserver_ready_callback, ns);
2072         if (event_add(&ns->event, NULL) < 0) {
2073           err = 2;
2074           goto out2;
2075         }
2076
2077         log(EVDNS_LOG_DEBUG, "Added nameserver %s", debug_ntoa(address));
2078
2079         /* insert this nameserver into the list of them */
2080         if (!server_head) {
2081                 ns->next = ns->prev = ns;
2082                 server_head = ns;
2083         } else {
2084                 ns->next = server_head->next;
2085                 ns->prev = server_head;
2086                 server_head->next = ns;
2087                 if (server_head->prev == server_head) {
2088                         server_head->prev = ns;
2089                 }
2090         }
2091
2092         global_good_nameservers++;
2093
2094         return 0;
2095
2096 out2:
2097         CLOSE_SOCKET(ns->socket);
2098 out1:
2099         free(ns);
2100         log(EVDNS_LOG_WARN, "Unable to add nameserver %s: error %d", debug_ntoa(address), err);
2101         return err;
2102 }
2103
2104 /* exported function */
2105 int
2106 evdns_nameserver_add(unsigned long int address) {
2107         return _evdns_nameserver_add_impl(address, 53);
2108 }
2109
2110 /* exported function */
2111 int
2112 evdns_nameserver_ip_add(const char *ip_as_string) {
2113         struct in_addr ina;
2114         int port;
2115         char buf[20];
2116         const char *cp;
2117         cp = strchr(ip_as_string, ':');
2118         if (! cp) {
2119                 cp = ip_as_string;
2120                 port = 53;
2121         } else {
2122                 port = strtoint(cp+1);
2123                 if (port < 0 || port > 65535) {
2124                         return 4;
2125                 }
2126                 if ((cp-ip_as_string) >= (int)sizeof(buf)) {
2127                         return 4;
2128                 }
2129                 memcpy(buf, ip_as_string, cp-ip_as_string);
2130                 buf[cp-ip_as_string] = '\0';
2131                 cp = buf;
2132         }
2133         if (!inet_aton(cp, &ina)) {
2134                 return 4;
2135         }
2136         return _evdns_nameserver_add_impl(ina.s_addr, port);
2137 }
2138
2139 /* insert into the tail of the queue */
2140 static void
2141 evdns_request_insert(struct request *req, struct request **head) {
2142         if (!*head) {
2143                 *head = req;
2144                 req->next = req->prev = req;
2145                 return;
2146         }
2147
2148         req->prev = (*head)->prev;
2149         req->prev->next = req;
2150         req->next = *head;
2151         (*head)->prev = req;
2152 }
2153
2154 static int
2155 string_num_dots(const char *s) {
2156         int count = 0;
2157         while ((s = strchr(s, '.'))) {
2158                 s++;
2159                 count++;
2160         }
2161         return count;
2162 }
2163
2164 static struct request *
2165 request_new(int type, const char *name, int flags,
2166     evdns_callback_type callback, void *user_ptr) {
2167         const char issuing_now =
2168             (global_requests_inflight < global_max_requests_inflight) ? 1 : 0;
2169
2170         const int name_len = strlen(name);
2171         const int request_max_len = evdns_request_len(name_len);
2172         const u16 trans_id = issuing_now ? transaction_id_pick() : 0xffff;
2173         /* the request data is alloced in a single block with the header */
2174         struct request *const req =
2175             (struct request *) malloc(sizeof(struct request) + request_max_len);
2176         int rlen;
2177         (void) flags;
2178
2179         if (!req) return NULL;
2180         memset(req, 0, sizeof(struct request));
2181
2182         /* request data lives just after the header */
2183         req->request = ((u8 *) req) + sizeof(struct request);
2184         /* denotes that the request data shouldn't be free()ed */
2185         req->request_appended = 1;
2186         rlen = evdns_request_data_build(name, name_len, trans_id,
2187             type, CLASS_INET, req->request, request_max_len);
2188         if (rlen < 0)
2189                 goto err1;
2190         req->request_len = rlen;
2191         req->trans_id = trans_id;
2192         req->tx_count = 0;
2193         req->request_type = type;
2194         req->user_pointer = user_ptr;
2195         req->user_callback = callback;
2196         req->ns = issuing_now ? nameserver_pick() : NULL;
2197         req->next = req->prev = NULL;
2198
2199         return req;
2200 err1:
2201         free(req);
2202         return NULL;
2203 }
2204
2205 static void
2206 request_submit(struct request *const req) {
2207         if (req->ns) {
2208                 /* if it has a nameserver assigned then this is going */
2209                 /* straight into the inflight queue */
2210                 evdns_request_insert(req, &req_head);
2211                 global_requests_inflight++;
2212                 evdns_request_transmit(req);
2213         } else {
2214                 evdns_request_insert(req, &req_waiting_head);
2215                 global_requests_waiting++;
2216         }
2217 }
2218
2219 /* exported function */
2220 int evdns_resolve_ipv4(const char *name, int flags,
2221     evdns_callback_type callback, void *ptr) {
2222         log(EVDNS_LOG_DEBUG, "Resolve requested for %s", name);
2223         if (flags & DNS_QUERY_NO_SEARCH) {
2224                 struct request *const req =
2225                         request_new(TYPE_A, name, flags, callback, ptr);
2226                 if (req == NULL)
2227                         return (1);
2228                 request_submit(req);
2229                 return (0);
2230         } else {
2231                 return (search_request_new(TYPE_A, name, flags, callback, ptr));
2232         }
2233 }
2234
2235 /* exported function */
2236 int evdns_resolve_ipv6(const char *name, int flags,
2237                                            evdns_callback_type callback, void *ptr) {
2238         log(EVDNS_LOG_DEBUG, "Resolve requested for %s", name);
2239         if (flags & DNS_QUERY_NO_SEARCH) {
2240                 struct request *const req =
2241                         request_new(TYPE_AAAA, name, flags, callback, ptr);
2242                 if (req == NULL)
2243                         return (1);
2244                 request_submit(req);
2245                 return (0);
2246         } else {
2247                 return (search_request_new(TYPE_AAAA, name, flags, callback, ptr));
2248         }
2249 }
2250
2251 int evdns_resolve_reverse(struct in_addr *in, int flags, evdns_callback_type callback, void *ptr) {
2252         char buf[32];
2253         struct request *req;
2254         u32 a;
2255         assert(in);
2256         a = ntohl(in->s_addr);
2257         snprintf(buf, sizeof(buf), "%d.%d.%d.%d.in-addr.arpa",
2258                         (int)(u8)((a    )&0xff),
2259                         (int)(u8)((a>>8 )&0xff),
2260                         (int)(u8)((a>>16)&0xff),
2261                         (int)(u8)((a>>24)&0xff));
2262         log(EVDNS_LOG_DEBUG, "Resolve requested for %s (reverse)", buf);
2263         req = request_new(TYPE_PTR, buf, flags, callback, ptr);
2264         if (!req) return 1;
2265         request_submit(req);
2266         return 0;
2267 }
2268
2269 int evdns_resolve_reverse_ipv6(struct in6_addr *in, int flags, evdns_callback_type callback, void *ptr) {
2270         char buf[96];
2271         char *cp;
2272         struct request *req;
2273         int i;
2274         assert(in);
2275         cp = buf;
2276         for (i=15; i >= 0; --i) {
2277                 u8 byte = in->s6_addr[i];
2278                 *cp++ = "0123456789abcdef"[byte & 0x0f];
2279                 *cp++ = '.';
2280                 *cp++ = "0123456789abcdef"[byte >> 4];
2281                 *cp++ = '.';
2282         }
2283         assert(cp + strlen(".ip6.arpa") < buf+sizeof(buf));
2284         memcpy(cp, ".ip6.arpa", strlen(".ip6.arpa")+1);
2285         log(EVDNS_LOG_DEBUG, "Resolve requested for %s (reverse)", buf);
2286         req = request_new(TYPE_PTR, buf, flags, callback, ptr);
2287         if (!req) return 1;
2288         request_submit(req);
2289         return 0;
2290 }
2291
2292 /*/////////////////////////////////////////////////////////////////// */
2293 /* Search support */
2294 /* */
2295 /* the libc resolver has support for searching a number of domains */
2296 /* to find a name. If nothing else then it takes the single domain */
2297 /* from the gethostname() call. */
2298 /* */
2299 /* It can also be configured via the domain and search options in a */
2300 /* resolv.conf. */
2301 /* */
2302 /* The ndots option controls how many dots it takes for the resolver */
2303 /* to decide that a name is non-local and so try a raw lookup first. */
2304
2305 struct search_domain {
2306         int len;
2307         struct search_domain *next;
2308         /* the text string is appended to this structure */
2309 };
2310
2311 struct search_state {
2312         int refcount;
2313         int ndots;
2314         int num_domains;
2315         struct search_domain *head;
2316 };
2317
2318 static struct search_state *global_search_state = NULL;
2319
2320 static void
2321 search_state_decref(struct search_state *const state) {
2322         if (!state) return;
2323         state->refcount--;
2324         if (!state->refcount) {
2325                 struct search_domain *next, *dom;
2326                 for (dom = state->head; dom; dom = next) {
2327                         next = dom->next;
2328                         free(dom);
2329                 }
2330                 free(state);
2331         }
2332 }
2333
2334 static struct search_state *
2335 search_state_new(void) {
2336         struct search_state *state = (struct search_state *) malloc(sizeof(struct search_state));
2337         if (!state) return NULL;
2338         memset(state, 0, sizeof(struct search_state));
2339         state->refcount = 1;
2340         state->ndots = 1;
2341
2342         return state;
2343 }
2344
2345 static void
2346 search_postfix_clear(void) {
2347         search_state_decref(global_search_state);
2348
2349         global_search_state = search_state_new();
2350 }
2351
2352 /* exported function */
2353 void
2354 evdns_search_clear(void) {
2355         search_postfix_clear();
2356 }
2357
2358 static void
2359 search_postfix_add(const char *domain) {
2360         int domain_len;
2361         struct search_domain *sdomain;
2362         while (domain[0] == '.') domain++;
2363         domain_len = strlen(domain);
2364
2365         if (!global_search_state) global_search_state = search_state_new();
2366         if (!global_search_state) return;
2367         global_search_state->num_domains++;
2368
2369         sdomain = (struct search_domain *) malloc(sizeof(struct search_domain) + domain_len);
2370         if (!sdomain) return;
2371         memcpy( ((u8 *) sdomain) + sizeof(struct search_domain), domain, domain_len);
2372         sdomain->next = global_search_state->head;
2373         sdomain->len = domain_len;
2374
2375         global_search_state->head = sdomain;
2376 }
2377
2378 /* reverse the order of members in the postfix list. This is needed because, */
2379 /* when parsing resolv.conf we push elements in the wrong order */
2380 static void
2381 search_reverse(void) {
2382         struct search_domain *cur, *prev = NULL, *next;
2383         cur = global_search_state->head;
2384         while (cur) {
2385                 next = cur->next;
2386                 cur->next = prev;
2387                 prev = cur;
2388                 cur = next;
2389         }
2390
2391         global_search_state->head = prev;
2392 }
2393
2394 /* exported function */
2395 void
2396 evdns_search_add(const char *domain) {
2397         search_postfix_add(domain);
2398 }
2399
2400 /* exported function */
2401 void
2402 evdns_search_ndots_set(const int ndots) {
2403         if (!global_search_state) global_search_state = search_state_new();
2404         if (!global_search_state) return;
2405         global_search_state->ndots = ndots;
2406 }
2407
2408 static void
2409 search_set_from_hostname(void) {
2410         char hostname[HOST_NAME_MAX + 1], *domainname;
2411
2412         search_postfix_clear();
2413         if (gethostname(hostname, sizeof(hostname))) return;
2414         domainname = strchr(hostname, '.');
2415         if (!domainname) return;
2416         search_postfix_add(domainname);
2417 }
2418
2419 /* warning: returns malloced string */
2420 static char *
2421 search_make_new(const struct search_state *const state, int n, const char *const base_name) {
2422         const int base_len = strlen(base_name);
2423         const char need_to_append_dot = base_name[base_len - 1] == '.' ? 0 : 1;
2424         struct search_domain *dom;
2425
2426         for (dom = state->head; dom; dom = dom->next) {
2427                 if (!n--) {
2428                         /* this is the postfix we want */
2429                         /* the actual postfix string is kept at the end of the structure */
2430                         const u8 *const postfix = ((u8 *) dom) + sizeof(struct search_domain);
2431                         const int postfix_len = dom->len;
2432                         char *const newname = (char *) malloc(base_len + need_to_append_dot + postfix_len + 1);
2433                         if (!newname) return NULL;
2434                         memcpy(newname, base_name, base_len);
2435                         if (need_to_append_dot) newname[base_len] = '.';
2436                         memcpy(newname + base_len + need_to_append_dot, postfix, postfix_len);
2437                         newname[base_len + need_to_append_dot + postfix_len] = 0;
2438                         return newname;
2439                 }
2440         }
2441
2442         /* we ran off the end of the list and still didn't find the requested string */
2443         abort();
2444         return NULL; /* unreachable; stops warnings in some compilers. */
2445 }
2446
2447 static int
2448 search_request_new(int type, const char *const name, int flags, evdns_callback_type user_callback, void *user_arg) {
2449         assert(type == TYPE_A || type == TYPE_AAAA);
2450         if ( ((flags & DNS_QUERY_NO_SEARCH) == 0) &&
2451              global_search_state &&
2452                  global_search_state->num_domains) {
2453                 /* we have some domains to search */
2454                 struct request *req;
2455                 if (string_num_dots(name) >= global_search_state->ndots) {
2456                         req = request_new(type, name, flags, user_callback, user_arg);
2457                         if (!req) return 1;
2458                         req->search_index = -1;
2459                 } else {
2460                         char *const new_name = search_make_new(global_search_state, 0, name);
2461                         if (!new_name) return 1;
2462                         req = request_new(type, new_name, flags, user_callback, user_arg);
2463                         free(new_name);
2464                         if (!req) return 1;
2465                         req->search_index = 0;
2466                 }
2467                 req->search_origname = strdup(name);
2468                 req->search_state = global_search_state;
2469                 req->search_flags = flags;
2470                 global_search_state->refcount++;
2471                 request_submit(req);
2472                 return 0;
2473         } else {
2474                 struct request *const req = request_new(type, name, flags, user_callback, user_arg);
2475                 if (!req) return 1;
2476                 request_submit(req);
2477                 return 0;
2478         }
2479 }
2480
2481 /* this is called when a request has failed to find a name. We need to check */
2482 /* if it is part of a search and, if so, try the next name in the list */
2483 /* returns: */
2484 /*   0 another request has been submitted */
2485 /*   1 no more requests needed */
2486 static int
2487 search_try_next(struct request *const req) {
2488         if (req->search_state) {
2489                 /* it is part of a search */
2490                 char *new_name;
2491                 struct request *newreq;
2492                 req->search_index++;
2493                 if (req->search_index >= req->search_state->num_domains) {
2494                         /* no more postfixes to try, however we may need to try */
2495                         /* this name without a postfix */
2496                         if (string_num_dots(req->search_origname) < req->search_state->ndots) {
2497                                 /* yep, we need to try it raw */
2498                                 struct request *const newreq = request_new(req->request_type, req->search_origname, req->search_flags, req->user_callback, req->user_pointer);
2499                                 log(EVDNS_LOG_DEBUG, "Search: trying raw query %s", req->search_origname);
2500                                 if (newreq) {
2501                                         request_submit(newreq);
2502                                         return 0;
2503                                 }
2504                         }
2505                         return 1;
2506                 }
2507
2508                 new_name = search_make_new(req->search_state, req->search_index, req->search_origname);
2509                 if (!new_name) return 1;
2510                 log(EVDNS_LOG_DEBUG, "Search: now trying %s (%d)", new_name, req->search_index);
2511                 newreq = request_new(req->request_type, new_name, req->search_flags, req->user_callback, req->user_pointer);
2512                 free(new_name);
2513                 if (!newreq) return 1;
2514                 newreq->search_origname = req->search_origname;
2515                 req->search_origname = NULL;
2516                 newreq->search_state = req->search_state;
2517                 newreq->search_flags = req->search_flags;
2518                 newreq->search_index = req->search_index;
2519                 newreq->search_state->refcount++;
2520                 request_submit(newreq);
2521                 return 0;
2522         }
2523         return 1;
2524 }
2525
2526 static void
2527 search_request_finished(struct request *const req) {
2528         if (req->search_state) {
2529                 search_state_decref(req->search_state);
2530                 req->search_state = NULL;
2531         }
2532         if (req->search_origname) {
2533                 free(req->search_origname);
2534                 req->search_origname = NULL;
2535         }
2536 }
2537
2538 /*/////////////////////////////////////////////////////////////////// */
2539 /* Parsing resolv.conf files */
2540
2541 static void
2542 evdns_resolv_set_defaults(int flags) {
2543         /* if the file isn't found then we assume a local resolver */
2544         if (flags & DNS_OPTION_SEARCH) search_set_from_hostname();
2545         if (flags & DNS_OPTION_NAMESERVERS) evdns_nameserver_ip_add("127.0.0.1");
2546 }
2547
2548 #ifndef HAVE_STRTOK_R
2549 static char *
2550 strtok_r(char *s, const char *delim, char **state) {
2551         return strtok(s, delim);
2552 }
2553 #endif
2554
2555 /* helper version of atoi which returns -1 on error */
2556 static int
2557 strtoint(const char *const str) {
2558         char *endptr;
2559         const int r = strtol(str, &endptr, 10);
2560         if (*endptr) return -1;
2561         return r;
2562 }
2563
2564 /* helper version of atoi that returns -1 on error and clips to bounds. */
2565 static int
2566 strtoint_clipped(const char *const str, int min, int max)
2567 {
2568         int r = strtoint(str);
2569         if (r == -1)
2570                 return r;
2571         else if (r<min)
2572                 return min;
2573         else if (r>max)
2574                 return max;
2575         else
2576                 return r;
2577 }
2578
2579 /* exported function */
2580 int
2581 evdns_set_option(const char *option, const char *val, int flags)
2582 {
2583         if (!strncmp(option, "ndots:", 6)) {
2584                 const int ndots = strtoint(val);
2585                 if (ndots == -1) return -1;
2586                 if (!(flags & DNS_OPTION_SEARCH)) return 0;
2587                 log(EVDNS_LOG_DEBUG, "Setting ndots to %d", ndots);
2588                 if (!global_search_state) global_search_state = search_state_new();
2589                 if (!global_search_state) return -1;
2590                 global_search_state->ndots = ndots;
2591         } else if (!strncmp(option, "timeout:", 8)) {
2592                 const int timeout = strtoint(val);
2593                 if (timeout == -1) return -1;
2594                 if (!(flags & DNS_OPTION_MISC)) return 0;
2595                 log(EVDNS_LOG_DEBUG, "Setting timeout to %d", timeout);
2596                 global_timeout.tv_sec = timeout;
2597         } else if (!strncmp(option, "max-timeouts:", 12)) {
2598                 const int maxtimeout = strtoint_clipped(val, 1, 255);
2599                 if (maxtimeout == -1) return -1;
2600                 if (!(flags & DNS_OPTION_MISC)) return 0;
2601                 log(EVDNS_LOG_DEBUG, "Setting maximum allowed timeouts to %d",
2602                         maxtimeout);
2603                 global_max_nameserver_timeout = maxtimeout;
2604         } else if (!strncmp(option, "max-inflight:", 13)) {
2605                 const int maxinflight = strtoint_clipped(val, 1, 65000);
2606                 if (maxinflight == -1) return -1;
2607                 if (!(flags & DNS_OPTION_MISC)) return 0;
2608                 log(EVDNS_LOG_DEBUG, "Setting maximum inflight requests to %d",
2609                         maxinflight);
2610                 global_max_requests_inflight = maxinflight;
2611         } else if (!strncmp(option, "attempts:", 9)) {
2612                 int retries = strtoint(val);
2613                 if (retries == -1) return -1;
2614                 if (retries > 255) retries = 255;
2615                 if (!(flags & DNS_OPTION_MISC)) return 0;
2616                 log(EVDNS_LOG_DEBUG, "Setting retries to %d", retries);
2617                 global_max_retransmits = retries;
2618         }
2619         return 0;
2620 }
2621
2622 static void
2623 resolv_conf_parse_line(char *const start, int flags) {
2624         char *strtok_state;
2625         static const char *const delims = " \t";
2626 #define NEXT_TOKEN strtok_r(NULL, delims, &strtok_state)
2627
2628         char *const first_token = strtok_r(start, delims, &strtok_state);
2629         if (!first_token) return;
2630
2631         if (!strcmp(first_token, "nameserver") && (flags & DNS_OPTION_NAMESERVERS)) {
2632                 const char *const nameserver = NEXT_TOKEN;
2633                 struct in_addr ina;
2634
2635                 if (inet_aton(nameserver, &ina)) {
2636                         /* address is valid */
2637                         evdns_nameserver_add(ina.s_addr);
2638                 }
2639         } else if (!strcmp(first_token, "domain") && (flags & DNS_OPTION_SEARCH)) {
2640                 const char *const domain = NEXT_TOKEN;
2641                 if (domain) {
2642                         search_postfix_clear();
2643                         search_postfix_add(domain);
2644                 }
2645         } else if (!strcmp(first_token, "search") && (flags & DNS_OPTION_SEARCH)) {
2646                 const char *domain;
2647                 search_postfix_clear();
2648
2649                 while ((domain = NEXT_TOKEN)) {
2650                         search_postfix_add(domain);
2651                 }
2652                 search_reverse();
2653         } else if (!strcmp(first_token, "options")) {
2654                 const char *option;
2655                 while ((option = NEXT_TOKEN)) {
2656                         const char *val = strchr(option, ':');
2657                         evdns_set_option(option, val ? val+1 : "", flags);
2658                 }
2659         }
2660 #undef NEXT_TOKEN
2661 }
2662
2663 /* exported function */
2664 /* returns: */
2665 /*   0 no errors */
2666 /*   1 failed to open file */
2667 /*   2 failed to stat file */
2668 /*   3 file too large */
2669 /*   4 out of memory */
2670 /*   5 short read from file */
2671 int
2672 evdns_resolv_conf_parse(int flags, const char *const filename) {
2673         struct stat st;
2674         int fd, n, r;
2675         u8 *resolv;
2676         char *start;
2677         int err = 0;
2678
2679         log(EVDNS_LOG_DEBUG, "Parsing resolv.conf file %s", filename);
2680
2681         fd = open(filename, O_RDONLY);
2682         if (fd < 0) {
2683                 evdns_resolv_set_defaults(flags);
2684                 return 1;
2685         }
2686
2687         if (fstat(fd, &st)) { err = 2; goto out1; }
2688         if (!st.st_size) {
2689                 evdns_resolv_set_defaults(flags);
2690                 err = (flags & DNS_OPTION_NAMESERVERS) ? 6 : 0;
2691                 goto out1;
2692         }
2693         if (st.st_size > 65535) { err = 3; goto out1; }  /* no resolv.conf should be any bigger */
2694
2695         resolv = (u8 *) malloc((size_t)st.st_size + 1);
2696         if (!resolv) { err = 4; goto out1; }
2697
2698         n = 0;
2699         while ((r = read(fd, resolv+n, (size_t)st.st_size-n)) > 0) {
2700                 n += r;
2701                 if (n == st.st_size)
2702                         break;
2703                 assert(n < st.st_size);
2704         }
2705         if (r < 0) { err = 5; goto out2; }
2706         resolv[n] = 0;   /* we malloced an extra byte; this should be fine. */
2707
2708         start = (char *) resolv;
2709         for (;;) {
2710                 char *const newline = strchr(start, '\n');
2711                 if (!newline) {
2712                         resolv_conf_parse_line(start, flags);
2713                         break;
2714                 } else {
2715                         *newline = 0;
2716                         resolv_conf_parse_line(start, flags);
2717                         start = newline + 1;
2718                 }
2719         }
2720
2721         if (!server_head && (flags & DNS_OPTION_NAMESERVERS)) {
2722                 /* no nameservers were configured. */
2723                 evdns_nameserver_ip_add("127.0.0.1");
2724                 err = 6;
2725         }
2726         if (flags & DNS_OPTION_SEARCH && (!global_search_state || global_search_state->num_domains == 0)) {
2727                 search_set_from_hostname();
2728         }
2729
2730 out2:
2731         free(resolv);
2732 out1:
2733         close(fd);
2734         return err;
2735 }
2736
2737 #ifdef WIN32
2738 /* Add multiple nameservers from a space-or-comma-separated list. */
2739 static int
2740 evdns_nameserver_ip_add_line(const char *ips) {
2741         const char *addr;
2742         char *buf;
2743         int r;
2744         while (*ips) {
2745                 while (ISSPACE(*ips) || *ips == ',' || *ips == '\t')
2746                         ++ips;
2747                 addr = ips;
2748                 while (ISDIGIT(*ips) || *ips == '.' || *ips == ':')
2749                         ++ips;
2750                 buf = malloc(ips-addr+1);
2751                 if (!buf) return 4;
2752                 memcpy(buf, addr, ips-addr);
2753                 buf[ips-addr] = '\0';
2754                 r = evdns_nameserver_ip_add(buf);
2755                 free(buf);
2756                 if (r) return r;
2757         }
2758         return 0;
2759 }
2760
2761 typedef DWORD(WINAPI *GetNetworkParams_fn_t)(FIXED_INFO *, DWORD*);
2762
2763 /* Use the windows GetNetworkParams interface in iphlpapi.dll to */
2764 /* figure out what our nameservers are. */
2765 static int
2766 load_nameservers_with_getnetworkparams(void)
2767 {
2768         /* Based on MSDN examples and inspection of  c-ares code. */
2769         FIXED_INFO *fixed;
2770         HMODULE handle = 0;
2771         ULONG size = sizeof(FIXED_INFO);
2772         void *buf = NULL;
2773         int status = 0, r, added_any;
2774         IP_ADDR_STRING *ns;
2775         GetNetworkParams_fn_t fn;
2776
2777         if (!(handle = LoadLibrary("iphlpapi.dll"))) {
2778                 log(EVDNS_LOG_WARN, "Could not open iphlpapi.dll");
2779                 status = -1;
2780                 goto done;
2781         }
2782         if (!(fn = (GetNetworkParams_fn_t) GetProcAddress(handle, "GetNetworkParams"))) {
2783                 log(EVDNS_LOG_WARN, "Could not get address of function.");
2784                 status = -1;
2785                 goto done;
2786         }
2787
2788         buf = malloc(size);
2789         if (!buf) { status = 4; goto done; }
2790         fixed = buf;
2791         r = fn(fixed, &size);
2792         if (r != ERROR_SUCCESS && r != ERROR_BUFFER_OVERFLOW) {
2793                 status = -1;
2794                 goto done;
2795         }
2796         if (r != ERROR_SUCCESS) {
2797                 free(buf);
2798                 buf = malloc(size);
2799                 if (!buf) { status = 4; goto done; }
2800                 fixed = buf;
2801                 r = fn(fixed, &size);
2802                 if (r != ERROR_SUCCESS) {
2803                         log(EVDNS_LOG_DEBUG, "fn() failed.");
2804                         status = -1;
2805                         goto done;
2806                 }
2807         }
2808
2809         assert(fixed);
2810         added_any = 0;
2811         ns = &(fixed->DnsServerList);
2812         while (ns) {
2813                 r = evdns_nameserver_ip_add_line(ns->IpAddress.String);
2814                 if (r) {
2815                         log(EVDNS_LOG_DEBUG,"Could not add nameserver %s to list,error: %d",
2816                                 (ns->IpAddress.String),(int)GetLastError());
2817                         status = r;
2818                         goto done;
2819                 } else {
2820                         log(EVDNS_LOG_DEBUG,"Succesfully added %s as nameserver",ns->IpAddress.String);
2821                 }
2822
2823                 added_any++;
2824                 ns = ns->Next;
2825         }
2826
2827         if (!added_any) {
2828                 log(EVDNS_LOG_DEBUG, "No nameservers added.");
2829                 status = -1;
2830         }
2831
2832  done:
2833         if (buf)
2834                 free(buf);
2835         if (handle)
2836                 FreeLibrary(handle);
2837         return status;
2838 }
2839
2840 static int
2841 config_nameserver_from_reg_key(HKEY key, const char *subkey)
2842 {
2843         char *buf;
2844         DWORD bufsz = 0, type = 0;
2845         int status = 0;
2846
2847         if (RegQueryValueEx(key, subkey, 0, &type, NULL, &bufsz)
2848             != ERROR_MORE_DATA)
2849                 return -1;
2850         if (!(buf = malloc(bufsz)))
2851                 return -1;
2852
2853         if (RegQueryValueEx(key, subkey, 0, &type, (LPBYTE)buf, &bufsz)
2854             == ERROR_SUCCESS && bufsz > 1) {
2855                 status = evdns_nameserver_ip_add_line(buf);
2856         }
2857
2858         free(buf);
2859         return status;
2860 }
2861
2862 #define SERVICES_KEY "System\\CurrentControlSet\\Services\\"
2863 #define WIN_NS_9X_KEY  SERVICES_KEY "VxD\\MSTCP"
2864 #define WIN_NS_NT_KEY  SERVICES_KEY "Tcpip\\Parameters"
2865
2866 static int
2867 load_nameservers_from_registry(void)
2868 {
2869         int found = 0;
2870         int r;
2871 #define TRY(k, name) \
2872         if (!found && config_nameserver_from_reg_key(k,name) == 0) {    \
2873                 log(EVDNS_LOG_DEBUG,"Found nameservers in %s/%s",#k,name); \
2874                 found = 1;                                              \
2875         } else if (!found) {                                            \
2876                 log(EVDNS_LOG_DEBUG,"Didn't find nameservers in %s/%s", \
2877                     #k,#name);                                          \
2878         }
2879
2880         if (((int)GetVersion()) > 0) { /* NT */
2881                 HKEY nt_key = 0, interfaces_key = 0;
2882
2883                 if (RegOpenKeyEx(HKEY_LOCAL_MACHINE, WIN_NS_NT_KEY, 0,
2884                                  KEY_READ, &nt_key) != ERROR_SUCCESS) {
2885                         log(EVDNS_LOG_DEBUG,"Couldn't open nt key, %d",(int)GetLastError());
2886                         return -1;
2887                 }
2888                 r = RegOpenKeyEx(nt_key, "Interfaces", 0,
2889                              KEY_QUERY_VALUE|KEY_ENUMERATE_SUB_KEYS,
2890                              &interfaces_key);
2891                 if (r != ERROR_SUCCESS) {
2892                         log(EVDNS_LOG_DEBUG,"Couldn't open interfaces key, %d",(int)GetLastError());
2893                         return -1;
2894                 }
2895                 TRY(nt_key, "NameServer");
2896                 TRY(nt_key, "DhcpNameServer");
2897                 TRY(interfaces_key, "NameServer");
2898                 TRY(interfaces_key, "DhcpNameServer");
2899                 RegCloseKey(interfaces_key);
2900                 RegCloseKey(nt_key);
2901         } else {
2902                 HKEY win_key = 0;
2903                 if (RegOpenKeyEx(HKEY_LOCAL_MACHINE, WIN_NS_9X_KEY, 0,
2904                                  KEY_READ, &win_key) != ERROR_SUCCESS) {
2905                         log(EVDNS_LOG_DEBUG, "Couldn't open registry key, %d", (int)GetLastError());
2906                         return -1;
2907                 }
2908                 TRY(win_key, "NameServer");
2909                 RegCloseKey(win_key);
2910         }
2911
2912         if (found == 0) {
2913                 log(EVDNS_LOG_WARN,"Didn't find any nameservers.");
2914         }
2915
2916         return found ? 0 : -1;
2917 #undef TRY
2918 }
2919
2920 int
2921 evdns_config_windows_nameservers(void)
2922 {
2923         if (load_nameservers_with_getnetworkparams() == 0)
2924                 return 0;
2925         return load_nameservers_from_registry();
2926 }
2927 #endif
2928
2929 int
2930 evdns_init(void)
2931 {
2932         int res = 0;
2933 #ifdef WIN32
2934         evdns_config_windows_nameservers();
2935 #else
2936         res = evdns_resolv_conf_parse(DNS_OPTIONS_ALL, "/etc/resolv.conf");
2937 #endif
2938
2939         return (res);
2940 }
2941
2942 const char *
2943 evdns_err_to_string(int err)
2944 {
2945     switch (err) {
2946         case DNS_ERR_NONE: return "no error";
2947         case DNS_ERR_FORMAT: return "misformatted query";
2948         case DNS_ERR_SERVERFAILED: return "server failed";
2949         case DNS_ERR_NOTEXIST: return "name does not exist";
2950         case DNS_ERR_NOTIMPL: return "query not implemented";
2951         case DNS_ERR_REFUSED: return "refused";
2952
2953         case DNS_ERR_TRUNCATED: return "reply truncated or ill-formed";
2954         case DNS_ERR_UNKNOWN: return "unknown";
2955         case DNS_ERR_TIMEOUT: return "request timed out";
2956         case DNS_ERR_SHUTDOWN: return "dns subsystem shut down";
2957         default: return "[Unknown error code]";
2958     }
2959 }
2960
2961 void
2962 evdns_shutdown(int fail_requests)
2963 {
2964         struct nameserver *server, *server_next;
2965         struct search_domain *dom, *dom_next;
2966
2967         while (req_head) {
2968                 if (fail_requests)
2969                         reply_callback(req_head, 0, DNS_ERR_SHUTDOWN, NULL);
2970                 request_finished(req_head, &req_head);
2971         }
2972         while (req_waiting_head) {
2973                 if (fail_requests)
2974                         reply_callback(req_waiting_head, 0, DNS_ERR_SHUTDOWN, NULL);
2975                 request_finished(req_waiting_head, &req_waiting_head);
2976         }
2977         global_requests_inflight = global_requests_waiting = 0;
2978
2979         for (server = server_head; server; server = server_next) {
2980                 server_next = server->next;
2981                 if (server->socket >= 0)
2982                         CLOSE_SOCKET(server->socket);
2983                 (void) event_del(&server->event);
2984                 if (server->state == 0)
2985                         (void) event_del(&server->timeout_event);
2986                 free(server);
2987                 if (server_next == server_head)
2988                         break;
2989         }
2990         server_head = NULL;
2991         global_good_nameservers = 0;
2992
2993         if (global_search_state) {
2994                 for (dom = global_search_state->head; dom; dom = dom_next) {
2995                         dom_next = dom->next;
2996                         free(dom);
2997                 }
2998                 free(global_search_state);
2999                 global_search_state = NULL;
3000         }
3001         evdns_log_fn = NULL;
3002 }
3003
3004 #ifdef EVDNS_MAIN
3005 void
3006 main_callback(int result, char type, int count, int ttl,
3007                           void *addrs, void *orig) {
3008         char *n = (char*)orig;
3009         int i;
3010         for (i = 0; i < count; ++i) {
3011                 if (type == DNS_IPv4_A) {
3012                         printf("%s: %s\n", n, debug_ntoa(((u32*)addrs)[i]));
3013                 } else if (type == DNS_PTR) {
3014                         printf("%s: %s\n", n, ((char**)addrs)[i]);
3015                 }
3016         }
3017         if (!count) {
3018                 printf("%s: No answer (%d)\n", n, result);
3019         }
3020         fflush(stdout);
3021 }
3022 void
3023 evdns_server_callback(struct evdns_server_request *req, void *data)
3024 {
3025         int i, r;
3026         (void)data;
3027         /* dummy; give 192.168.11.11 as an answer for all A questions,
3028          *      give foo.bar.example.com as an answer for all PTR questions. */
3029         for (i = 0; i < req->nquestions; ++i) {
3030                 u32 ans = htonl(0xc0a80b0bUL);
3031                 if (req->questions[i]->type == EVDNS_TYPE_A &&
3032                         req->questions[i]->class == EVDNS_CLASS_INET) {
3033                         printf(" -- replying for %s (A)\n", req->questions[i]->name);
3034                         r = evdns_server_request_add_a_reply(req, req->questions[i]->name,
3035                                                                                   1, &ans, 10);
3036                         if (r<0)
3037                                 printf("eeep, didn't work.\n");
3038                 } else if (req->questions[i]->type == EVDNS_TYPE_PTR &&
3039                                    req->questions[i]->class == EVDNS_CLASS_INET) {
3040                         printf(" -- replying for %s (PTR)\n", req->questions[i]->name);
3041                         r = evdns_server_request_add_ptr_reply(req, NULL, req->questions[i]->name,
3042                                                                                         "foo.bar.example.com", 10);
3043                 } else {
3044                         printf(" -- skipping %s [%d %d]\n", req->questions[i]->name,
3045                                    req->questions[i]->type, req->questions[i]->class);
3046                 }
3047         }
3048
3049         r = evdns_request_respond(req, 0);
3050         if (r<0)
3051                 printf("eeek, couldn't send reply.\n");
3052 }
3053
3054 void
3055 logfn(int is_warn, const char *msg) {
3056   (void) is_warn;
3057   fprintf(stderr, "%s\n", msg);
3058 }
3059 int
3060 main(int c, char **v) {
3061         int idx;
3062         int reverse = 0, verbose = 1, servertest = 0;
3063         if (c<2) {
3064                 fprintf(stderr, "syntax: %s [-x] [-v] hostname\n", v[0]);
3065                 fprintf(stderr, "syntax: %s [-servertest]\n", v[0]);
3066                 return 1;
3067         }
3068         idx = 1;
3069         while (idx < c && v[idx][0] == '-') {
3070                 if (!strcmp(v[idx], "-x"))
3071                         reverse = 1;
3072                 else if (!strcmp(v[idx], "-v"))
3073                         verbose = 1;
3074                 else if (!strcmp(v[idx], "-servertest"))
3075                         servertest = 1;
3076                 else
3077                         fprintf(stderr, "Unknown option %s\n", v[idx]);
3078                 ++idx;
3079         }
3080         event_init();
3081         if (verbose)
3082                 evdns_set_log_fn(logfn);
3083         evdns_resolv_conf_parse(DNS_OPTION_NAMESERVERS, "/etc/resolv.conf");
3084         if (servertest) {
3085                 int sock;
3086                 struct sockaddr_in my_addr;
3087                 sock = socket(PF_INET, SOCK_DGRAM, 0);
3088                 fcntl(sock, F_SETFL, O_NONBLOCK);
3089                 my_addr.sin_family = AF_INET;
3090                 my_addr.sin_port = htons(10053);
3091                 my_addr.sin_addr.s_addr = INADDR_ANY;
3092                 if (bind(sock, (struct sockaddr*)&my_addr, sizeof(my_addr))<0) {
3093                         perror("bind");
3094                         exit(1);
3095                 }
3096                 evdns_add_server_port(sock, 0, evdns_server_callback, NULL);
3097         }
3098         for (; idx < c; ++idx) {
3099                 if (reverse) {
3100                         struct in_addr addr;
3101                         if (!inet_aton(v[idx], &addr)) {
3102                                 fprintf(stderr, "Skipping non-IP %s\n", v[idx]);
3103                                 continue;
3104                         }
3105                         fprintf(stderr, "resolving %s...\n",v[idx]);
3106                         evdns_resolve_reverse(&addr, 0, main_callback, v[idx]);
3107                 } else {
3108                         fprintf(stderr, "resolving (fwd) %s...\n",v[idx]);
3109                         evdns_resolve_ipv4(v[idx], 0, main_callback, v[idx]);
3110                 }
3111         }
3112         fflush(stdout);
3113         event_dispatch();
3114         return 0;
3115 }
3116 #endif