src/third_party/libevent/evdns.c

   1 /* $Id: evdns.c 6979 2006-08-04 18:31:13Z nickm $ */
   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 DNS_USE_FTIME_FOR_ID
  43 #include <sys/timeb.h>
  44 #endif
  45
  46 #ifndef DNS_USE_CPU_CLOCK_FOR_ID
  47 #ifndef DNS_USE_GETTIMEOFDAY_FOR_ID
  48 #ifndef DNS_USE_OPENSSL_FOR_ID
  49 #ifndef DNS_USE_FTIME_FOR_ID
  50 #error Must configure at least one id generation method.
  51 #error Please see the documentation.
  52 #endif
  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 #ifndef _FORTIFY_SOURCE
  78 #define _FORTIFY_SOURCE 3
  79 #endif
  80
  81 #include <string.h>
  82 #include <fcntl.h>
  83 #ifdef HAVE_SYS_TIME_H
  84 #include <sys/time.h>
  85 #endif
  86 #ifdef HAVE_STDINT_H
  87 #include <stdint.h>
  88 #endif
  89 #include <stdlib.h>
  90 #include <string.h>
  91 #include <errno.h>
  92 #include <assert.h>
  93 #ifdef HAVE_UNISTD_H
  94 #include <unistd.h>
  95 #endif
  96 #include <limits.h>
  97 #include <sys/stat.h>
  98 #include <ctype.h>
  99 #include <stdio.h>
 100 #include <stdarg.h>
 101
 102 #include "evdns.h"
 103 #include "evutil.h"
 104 #include "log.h"
 105 #ifdef WIN32
 106 #include <winsock2.h>
 107 #include <windows.h>
 108 #include <iphlpapi.h>
 109 #include <io.h>
 110 #else
 111 #include <sys/socket.h>
 112 #include <netinet/in.h>
 113 #include <arpa/inet.h>
 114 #endif
 115
 116 #ifdef HAVE_NETINET_IN6_H
 117 #include <netinet/in6.h>
 118 #endif
 119
 120 #define EVDNS_LOG_DEBUG 0
 121 #define EVDNS_LOG_WARN 1
 122
 123 #ifndef HOST_NAME_MAX
 124 #define HOST_NAME_MAX 255
 125 #endif
 126
 127 #include <stdio.h>
 128
 129 #undef MIN
 130 #define MIN(a,b) ((a)<(b)?(a):(b))
 131
 132 #ifdef __USE_ISOC99B
 133 /* libevent doesn't work without this */
 134 typedef ev_uint8_t u_char;
 135 typedef unsigned int uint;
 136 #endif
 137 #include "event.h"
 138
 139 #define u64 ev_uint64_t
 140 #define u32 ev_uint32_t
 141 #define u16 ev_uint16_t
 142 #define u8  ev_uint8_t
 143
 144 #ifdef WIN32
 145 #define open _open
 146 #define read _read
 147 #define close _close
 148 #define strdup _strdup
 149 #endif
 150
 151 #define MAX_ADDRS 32  /* maximum number of addresses from a single packet */
 152 /* which we bother recording */
 153
 154 #define TYPE_A         EVDNS_TYPE_A
 155 #define TYPE_CNAME     5
 156 #define TYPE_PTR       EVDNS_TYPE_PTR
 157 #define TYPE_AAAA      EVDNS_TYPE_AAAA
 158
 159 #define CLASS_INET     EVDNS_CLASS_INET
 160
 161 struct request {
 162         u8 *request;  /* the dns packet data */
 163         unsigned int request_len;
 164         int reissue_count;
 165         int tx_count;  /* the number of times that this packet has been sent */
 166         unsigned int request_type; /* TYPE_PTR or TYPE_A */
 167         void *user_pointer;  /* the pointer given to us for this request */
 168         evdns_callback_type user_callback;
 169         struct nameserver *ns;  /* the server which we last sent it */
 170
 171         /* elements used by the searching code */
 172         int search_index;
 173         struct search_state *search_state;
 174         char *search_origname;  /* needs to be free()ed */
 175         int search_flags;
 176
 177         /* these objects are kept in a circular list */
 178         struct request *next, *prev;
 179
 180         struct event timeout_event;
 181
 182         u16 trans_id;  /* the transaction id */
 183         char request_appended;  /* true if the request pointer is data which follows this struct */
 184         char transmit_me;  /* needs to be transmitted */
 185 };
 186
 187 #ifndef HAVE_STRUCT_IN6_ADDR
 188 struct in6_addr {
 189         u8 s6_addr[16];
 190 };
 191 #endif
 192
 193 struct reply {
 194         unsigned int type;
 195         unsigned int have_answer;
 196         union {
 197                 struct {
 198                         u32 addrcount;
 199                         u32 addresses[MAX_ADDRS];
 200                 } a;
 201                 struct {
 202                         u32 addrcount;
 203                         struct in6_addr addresses[MAX_ADDRS];
 204                 } aaaa;
 205                 struct {
 206                         char name[HOST_NAME_MAX];
 207                 } ptr;
 208         } data;
 209 };
 210
 211 struct nameserver {
 212         int socket;  /* a connected UDP socket */
 213         u32 address;
 214         u16 port;
 215         int failed_times;  /* number of times which we have given this server a chance */
 216         int timedout;  /* number of times in a row a request has timed out */
 217         struct event event;
 218         /* these objects are kept in a circular list */
 219         struct nameserver *next, *prev;
 220         struct event timeout_event;  /* used to keep the timeout for */
 221                                      /* when we next probe this server. */
 222                                      /* Valid if state == 0 */
 223         char state;  /* zero if we think that this server is down */
 224         char choked;  /* true if we have an EAGAIN from this server's socket */
 225         char write_waiting;  /* true if we are waiting for EV_WRITE events */
 226 };
 227
 228 static struct request *req_head = NULL, *req_waiting_head = NULL;
 229 static struct nameserver *server_head = NULL;
 230
 231 /* Represents a local port where we're listening for DNS requests. Right now, */
 232 /* only UDP is supported. */
 233 struct evdns_server_port {
 234         int socket; /* socket we use to read queries and write replies. */
 235         int refcnt; /* reference count. */
 236         char choked; /* Are we currently blocked from writing? */
 237         char closing; /* Are we trying to close this port, pending writes? */
 238         evdns_request_callback_fn_type user_callback; /* Fn to handle requests */
 239         void *user_data; /* Opaque pointer passed to user_callback */
 240         struct event event; /* Read/write event */
 241         /* circular list of replies that we want to write. */
 242         struct server_request *pending_replies;
 243 };
 244
 245 /* Represents part of a reply being built.      (That is, a single RR.) */
 246 struct server_reply_item {
 247         struct server_reply_item *next; /* next item in sequence. */
 248         char *name; /* name part of the RR */
 249         u16 type : 16; /* The RR type */
 250         u16 class : 16; /* The RR class (usually CLASS_INET) */
 251         u32 ttl; /* The RR TTL */
 252         char is_name; /* True iff data is a label */
 253         u16 datalen; /* Length of data; -1 if data is a label */
 254         void *data; /* The contents of the RR */
 255 };
 256
 257 /* Represents a request that we've received as a DNS server, and holds */
 258 /* the components of the reply as we're constructing it. */
 259 struct server_request {
 260         /* Pointers to the next and previous entries on the list of replies */
 261         /* that we're waiting to write.  Only set if we have tried to respond */
 262         /* and gotten EAGAIN. */
 263         struct server_request *next_pending;
 264         struct server_request *prev_pending;
 265
 266         u16 trans_id; /* Transaction id. */
 267         struct evdns_server_port *port; /* Which port received this request on? */
 268         struct sockaddr_storage addr; /* Where to send the response */
 269         socklen_t addrlen; /* length of addr */
 270
 271         int n_answer; /* how many answer RRs have been set? */
 272         int n_authority; /* how many authority RRs have been set? */
 273         int n_additional; /* how many additional RRs have been set? */
 274
 275         struct server_reply_item *answer; /* linked list of answer RRs */
 276         struct server_reply_item *authority; /* linked list of authority RRs */
 277         struct server_reply_item *additional; /* linked list of additional RRs */
 278
 279         /* Constructed response.  Only set once we're ready to send a reply. */
 280         /* Once this is set, the RR fields are cleared, and no more should be set. */
 281         char *response;
 282         size_t response_len;
 283
 284         /* Caller-visible fields: flags, questions. */
 285         struct evdns_server_request base;
 286 };
 287
 288 /* helper macro */
 289 #define OFFSET_OF(st, member) ((off_t) (((char*)&((st*)0)->member)-(char*)0))
 290
 291 /* Given a pointer to an evdns_server_request, get the corresponding */
 292 /* server_request. */
 293 #define TO_SERVER_REQUEST(base_ptr)                                                                             \
 294         ((struct server_request*)                                                                                       \
 295          (((char*)(base_ptr) - OFFSET_OF(struct server_request, base))))
 296
 297 /* The number of good nameservers that we have */
 298 static int global_good_nameservers = 0;
 299
 300 /* inflight requests are contained in the req_head list */
 301 /* and are actually going out across the network */
 302 static int global_requests_inflight = 0;
 303 /* requests which aren't inflight are in the waiting list */
 304 /* and are counted here */
 305 static int global_requests_waiting = 0;
 306
 307 static int global_max_requests_inflight = 64;
 308
 309 static struct timeval global_timeout = {5, 0};  /* 5 seconds */
 310 static int global_max_reissues = 1;  /* a reissue occurs when we get some errors from the server */
 311 static int global_max_retransmits = 3;  /* number of times we'll retransmit a request which timed out */
 312 /* number of timeouts in a row before we consider this server to be down */
 313 static int global_max_nameserver_timeout = 3;
 314
 315 /* These are the timeout values for nameservers. If we find a nameserver is down */
 316 /* we try to probe it at intervals as given below. Values are in seconds. */
 317 static const struct timeval global_nameserver_timeouts[] = {{10, 0}, {60, 0}, {300, 0}, {900, 0}, {3600, 0}};
 318 static const int global_nameserver_timeouts_length = sizeof(global_nameserver_timeouts)/sizeof(struct timeval);
 319
 320 static struct nameserver *nameserver_pick(void);
 321 static void evdns_request_insert(struct request *req, struct request **head);
 322 static void nameserver_ready_callback(int fd, short events, void *arg);
 323 static int evdns_transmit(void);
 324 static int evdns_request_transmit(struct request *req);
 325 static void nameserver_send_probe(struct nameserver *const ns);
 326 static void search_request_finished(struct request *const);
 327 static int search_try_next(struct request *const req);
 328 static int search_request_new(int type, const char *const name, int flags, evdns_callback_type user_callback, void *user_arg);
 329 static void evdns_requests_pump_waiting_queue(void);
 330 static u16 transaction_id_pick(void);
 331 static struct request *request_new(int type, const char *name, int flags, evdns_callback_type callback, void *ptr);
 332 static void request_submit(struct request *const req);
 333
 334 static int server_request_free(struct server_request *req);
 335 static void server_request_free_answers(struct server_request *req);
 336 static void server_port_free(struct evdns_server_port *port);
 337 static void server_port_ready_callback(int fd, short events, void *arg);
 338
 339 static int strtoint(const char *const str);
 340
 341 #ifdef WIN32
 342 static int
 343 last_error(int sock)
 344 {
 345         int optval, optvallen=sizeof(optval);
 346         int err = WSAGetLastError();
 347         if (err == WSAEWOULDBLOCK && sock >= 0) {
 348                 if (getsockopt(sock, SOL_SOCKET, SO_ERROR, (void*)&optval,
 349                                &optvallen))
 350                         return err;
 351                 if (optval)
 352                         return optval;
 353         }
 354         return err;
 355
 356 }
 357 static int
 358 error_is_eagain(int err)
 359 {
 360         return err == EAGAIN || err == WSAEWOULDBLOCK;
 361 }
 362 static int
 363 inet_aton(const char *c, struct in_addr *addr)
 364 {
 365         ev_uint32_t r;
 366         if (strcmp(c, "255.255.255.255") == 0) {
 367                 addr->s_addr = 0xffffffffu;
 368         } else {
 369                 r = inet_addr(c);
 370                 if (r == INADDR_NONE)
 371                         return 0;
 372                 addr->s_addr = r;
 373         }
 374         return 1;
 375 }
 376 #else
 377 #define last_error(sock) (errno)
 378 #define error_is_eagain(err) ((err) == EAGAIN)
 379 #endif
 380 #define CLOSE_SOCKET(s) EVUTIL_CLOSESOCKET(s)
 381
 382 #define ISSPACE(c) isspace((int)(unsigned char)(c))
 383 #define ISDIGIT(c) isdigit((int)(unsigned char)(c))
 384
 385 static const char *
 386 debug_ntoa(u32 address)
 387 {
 388         static char buf[32];
 389         u32 a = ntohl(address);
 390         evutil_snprintf(buf, sizeof(buf), "%d.%d.%d.%d",
 391                       (int)(u8)((a>>24)&0xff),
 392                       (int)(u8)((a>>16)&0xff),
 393                       (int)(u8)((a>>8 )&0xff),
 394                       (int)(u8)((a    )&0xff));
 395         return buf;
 396 }
 397
 398 static evdns_debug_log_fn_type evdns_log_fn = NULL;
 399
 400 void
 401 evdns_set_log_fn(evdns_debug_log_fn_type fn)
 402 {
 403   evdns_log_fn = fn;
 404 }
 405
 406 #ifdef __GNUC__
 407 #define EVDNS_LOG_CHECK  __attribute__ ((format(printf, 2, 3)))
 408 #else
 409 #define EVDNS_LOG_CHECK
 410 #endif
 411
 412 static void _evdns_log(int warn, const char *fmt, ...) EVDNS_LOG_CHECK;
 413 static void
 414 _evdns_log(int warn, const char *fmt, ...)
 415 {
 416   va_list args;
 417   static char buf[512];
 418   if (!evdns_log_fn)
 419     return;
 420   va_start(args,fmt);
 421   evutil_vsnprintf(buf, sizeof(buf), fmt, args);
 422   buf[sizeof(buf)-1] = '\0';
 423   evdns_log_fn(warn, buf);
 424   va_end(args);
 425 }
 426
 427 #define log _evdns_log
 428
 429 /* This walks the list of inflight requests to find the */
 430 /* one with a matching transaction id. Returns NULL on */
 431 /* failure */
 432 static struct request *
 433 request_find_from_trans_id(u16 trans_id) {
 434         struct request *req = req_head, *const started_at = req_head;
 435
 436         if (req) {
 437                 do {
 438                         if (req->trans_id == trans_id) return req;
 439                         req = req->next;
 440                 } while (req != started_at);
 441         }
 442
 443         return NULL;
 444 }
 445
 446 /* a libevent callback function which is called when a nameserver */
 447 /* has gone down and we want to test if it has came back to life yet */
 448 static void
 449 nameserver_prod_callback(int fd, short events, void *arg) {
 450         struct nameserver *const ns = (struct nameserver *) arg;
 451         (void)fd;
 452         (void)events;
 453
 454         nameserver_send_probe(ns);
 455 }
 456
 457 /* a libevent callback which is called when a nameserver probe (to see if */
 458 /* it has come back to life) times out. We increment the count of failed_times */
 459 /* and wait longer to send the next probe packet. */
 460 static void
 461 nameserver_probe_failed(struct nameserver *const ns) {
 462         const struct timeval * timeout;
 463         (void) evtimer_del(&ns->timeout_event);
 464         if (ns->state == 1) {
 465                 /* This can happen if the nameserver acts in a way which makes us mark */
 466                 /* it as bad and then starts sending good replies. */
 467                 return;
 468         }
 469
 470         timeout =
 471           &global_nameserver_timeouts[MIN(ns->failed_times,
 472                                           global_nameserver_timeouts_length - 1)];
 473         ns->failed_times++;
 474
 475         if (evtimer_add(&ns->timeout_event, (struct timeval *) timeout) < 0) {
 476           log(EVDNS_LOG_WARN,
 477               "Error from libevent when adding timer event for %s",
 478               debug_ntoa(ns->address));
 479           /* ???? Do more? */
 480         }
 481 }
 482
 483 /* called when a nameserver has been deemed to have failed. For example, too */
 484 /* many packets have timed out etc */
 485 static void
 486 nameserver_failed(struct nameserver *const ns, const char *msg) {
 487         struct request *req, *started_at;
 488         /* if this nameserver has already been marked as failed */
 489         /* then don't do anything */
 490         if (!ns->state) return;
 491
 492         log(EVDNS_LOG_WARN, "Nameserver %s has failed: %s",
 493             debug_ntoa(ns->address), msg);
 494         global_good_nameservers--;
 495         assert(global_good_nameservers >= 0);
 496         if (global_good_nameservers == 0) {
 497                 log(EVDNS_LOG_WARN, "All nameservers have failed");
 498         }
 499
 500         ns->state = 0;
 501         ns->failed_times = 1;
 502
 503         if (evtimer_add(&ns->timeout_event, (struct timeval *) &global_nameserver_timeouts[0]) < 0) {
 504                 log(EVDNS_LOG_WARN,
 505                     "Error from libevent when adding timer event for %s",
 506                     debug_ntoa(ns->address));
 507                 /* ???? Do more? */
 508         }
 509
 510         /* walk the list of inflight requests to see if any can be reassigned to */
 511         /* a different server. Requests in the waiting queue don't have a */
 512         /* nameserver assigned yet */
 513
 514         /* if we don't have *any* good nameservers then there's no point */
 515         /* trying to reassign requests to one */
 516         if (!global_good_nameservers) return;
 517
 518         req = req_head;
 519         started_at = req_head;
 520         if (req) {
 521                 do {
 522                         if (req->tx_count == 0 && req->ns == ns) {
 523                                 /* still waiting to go out, can be moved */
 524                                 /* to another server */
 525                                 req->ns = nameserver_pick();
 526                         }
 527                         req = req->next;
 528                 } while (req != started_at);
 529         }
 530 }
 531
 532 static void
 533 nameserver_up(struct nameserver *const ns) {
 534         if (ns->state) return;
 535         log(EVDNS_LOG_WARN, "Nameserver %s is back up",
 536             debug_ntoa(ns->address));
 537         evtimer_del(&ns->timeout_event);
 538         ns->state = 1;
 539         ns->failed_times = 0;
 540         ns->timedout = 0;
 541         global_good_nameservers++;
 542 }
 543
 544 static void
 545 request_trans_id_set(struct request *const req, const u16 trans_id) {
 546         req->trans_id = trans_id;
 547         *((u16 *) req->request) = htons(trans_id);
 548 }
 549
 550 /* Called to remove a request from a list and dealloc it. */
 551 /* head is a pointer to the head of the list it should be */
 552 /* removed from or NULL if the request isn't in a list. */
 553 static void
 554 request_finished(struct request *const req, struct request **head) {
 555         if (head) {
 556                 if (req->next == req) {
 557                         /* only item in the list */
 558                         *head = NULL;
 559                 } else {
 560                         req->next->prev = req->prev;
 561                         req->prev->next = req->next;
 562                         if (*head == req) *head = req->next;
 563                 }
 564         }
 565
 566         log(EVDNS_LOG_DEBUG, "Removing timeout for request %lx",
 567             (unsigned long) req);
 568         evtimer_del(&req->timeout_event);
 569
 570         search_request_finished(req);
 571         global_requests_inflight--;
 572
 573         if (!req->request_appended) {
 574                 /* need to free the request data on it's own */
 575                 free(req->request);
 576         } else {
 577                 /* the request data is appended onto the header */
 578                 /* so everything gets free()ed when we: */
 579         }
 580
 581         free(req);
 582
 583         evdns_requests_pump_waiting_queue();
 584 }
 585
 586 /* This is called when a server returns a funny error code. */
 587 /* We try the request again with another server. */
 588 /* */
 589 /* return: */
 590 /*   0 ok */
 591 /*   1 failed/reissue is pointless */
 592 static int
 593 request_reissue(struct request *req) {
 594         const struct nameserver *const last_ns = req->ns;
 595         /* the last nameserver should have been marked as failing */
 596         /* by the caller of this function, therefore pick will try */
 597         /* not to return it */
 598         req->ns = nameserver_pick();
 599         if (req->ns == last_ns) {
 600                 /* ... but pick did return it */
 601                 /* not a lot of point in trying again with the */
 602                 /* same server */
 603                 return 1;
 604         }
 605
 606         req->reissue_count++;
 607         req->tx_count = 0;
 608         req->transmit_me = 1;
 609
 610         return 0;
 611 }
 612
 613 /* this function looks for space on the inflight queue and promotes */
 614 /* requests from the waiting queue if it can. */
 615 static void
 616 evdns_requests_pump_waiting_queue(void) {
 617         while (global_requests_inflight < global_max_requests_inflight &&
 618             global_requests_waiting) {
 619                 struct request *req;
 620                 /* move a request from the waiting queue to the inflight queue */
 621                 assert(req_waiting_head);
 622                 if (req_waiting_head->next == req_waiting_head) {
 623                         /* only one item in the queue */
 624                         req = req_waiting_head;
 625                         req_waiting_head = NULL;
 626                 } else {
 627                         req = req_waiting_head;
 628                         req->next->prev = req->prev;
 629                         req->prev->next = req->next;
 630                         req_waiting_head = req->next;
 631                 }
 632
 633                 global_requests_waiting--;
 634                 global_requests_inflight++;
 635
 636                 req->ns = nameserver_pick();
 637                 request_trans_id_set(req, transaction_id_pick());
 638
 639                 evdns_request_insert(req, &req_head);
 640                 evdns_request_transmit(req);
 641                 evdns_transmit();
 642         }
 643 }
 644
 645 static void
 646 reply_callback(struct request *const req, u32 ttl, u32 err, struct reply *reply) {
 647         switch (req->request_type) {
 648         case TYPE_A:
 649                 if (reply)
 650                         req->user_callback(DNS_ERR_NONE, DNS_IPv4_A,
 651                                                            reply->data.a.addrcount, ttl,
 652                                                  reply->data.a.addresses,
 653                                                            req->user_pointer);
 654                 else
 655                         req->user_callback(err, 0, 0, 0, NULL, req->user_pointer);
 656                 return;
 657         case TYPE_PTR:
 658                 if (reply) {
 659                         char *name = reply->data.ptr.name;
 660                         req->user_callback(DNS_ERR_NONE, DNS_PTR, 1, ttl,
 661                                                            &name, req->user_pointer);
 662                 } else {
 663                         req->user_callback(err, 0, 0, 0, NULL,
 664                                                            req->user_pointer);
 665                 }
 666                 return;
 667         case TYPE_AAAA:
 668                 if (reply)
 669                         req->user_callback(DNS_ERR_NONE, DNS_IPv6_AAAA,
 670                                                            reply->data.aaaa.addrcount, ttl,
 671                                                            reply->data.aaaa.addresses,
 672                                                            req->user_pointer);
 673                 else
 674                         req->user_callback(err, 0, 0, 0, NULL, req->user_pointer);
 675                 return;
 676         }
 677         assert(0);
 678 }
 679
 680 /* this processes a parsed reply packet */
 681 static void
 682 reply_handle(struct request *const req, u16 flags, u32 ttl, struct reply *reply) {
 683         int error;
 684         static const int error_codes[] = {
 685                 DNS_ERR_FORMAT, DNS_ERR_SERVERFAILED, DNS_ERR_NOTEXIST,
 686                 DNS_ERR_NOTIMPL, DNS_ERR_REFUSED
 687         };
 688
 689         if (flags & 0x020f || !reply || !reply->have_answer) {
 690                 /* there was an error */
 691                 if (flags & 0x0200) {
 692                         error = DNS_ERR_TRUNCATED;
 693                 } else {
 694                         u16 error_code = (flags & 0x000f) - 1;
 695                         if (error_code > 4) {
 696                                 error = DNS_ERR_UNKNOWN;
 697                         } else {
 698                                 error = error_codes[error_code];
 699                         }
 700                 }
 701
 702                 switch(error) {
 703                 case DNS_ERR_NOTIMPL:
 704                 case DNS_ERR_REFUSED:
 705                         /* we regard these errors as marking a bad nameserver */
 706                         if (req->reissue_count < global_max_reissues) {
 707                                 char msg[64];
 708                                 evutil_snprintf(msg, sizeof(msg),
 709                                     "Bad response %d (%s)",
 710                                          error, evdns_err_to_string(error));
 711                                 nameserver_failed(req->ns, msg);
 712                                 if (!request_reissue(req)) return;
 713                         }
 714                         break;
 715                 case DNS_ERR_SERVERFAILED:
 716                         /* rcode 2 (servfailed) sometimes means "we
 717                          * are broken" and sometimes (with some binds)
 718                          * means "that request was very confusing."
 719                          * Treat this as a timeout, not a failure.
 720                          */
 721                         log(EVDNS_LOG_DEBUG, "Got a SERVERFAILED from nameserver %s; "
 722                                 "will allow the request to time out.",
 723                                 debug_ntoa(req->ns->address));
 724                         break;
 725                 default:
 726                         /* we got a good reply from the nameserver */
 727                         nameserver_up(req->ns);
 728                 }
 729
 730                 if (req->search_state && req->request_type != TYPE_PTR) {
 731                         /* if we have a list of domains to search in,
 732                          * try the next one */
 733                         if (!search_try_next(req)) {
 734                                 /* a new request was issued so this
 735                                  * request is finished and */
 736                                 /* the user callback will be made when
 737                                  * that request (or a */
 738                                 /* child of it) finishes. */
 739                                 request_finished(req, &req_head);
 740                                 return;
 741                         }
 742                 }
 743
 744                 /* all else failed. Pass the failure up */
 745                 reply_callback(req, 0, error, NULL);
 746                 request_finished(req, &req_head);
 747         } else {
 748                 /* all ok, tell the user */
 749                 reply_callback(req, ttl, 0, reply);
 750                 nameserver_up(req->ns);
 751                 request_finished(req, &req_head);
 752         }
 753 }
 754
 755 static int
 756 name_parse(u8 *packet, int length, int *idx, char *name_out, int name_out_len) {
 757         int name_end = -1;
 758         int j = *idx;
 759         int ptr_count = 0;
 760 #define GET32(x) do { if (j + 4 > length) goto err; memcpy(&_t32, packet + j, 4); j += 4; x = ntohl(_t32); } while(0)
 761 #define GET16(x) do { if (j + 2 > length) goto err; memcpy(&_t, packet + j, 2); j += 2; x = ntohs(_t); } while(0)
 762 #define GET8(x) do { if (j >= length) goto err; x = packet[j++]; } while(0)
 763
 764         char *cp = name_out;
 765         const char *const end = name_out + name_out_len;
 766
 767         /* Normally, names are a series of length prefixed strings terminated */
 768         /* with a length of 0 (the lengths are u8's < 63). */
 769         /* However, the length can start with a pair of 1 bits and that */
 770         /* means that the next 14 bits are a pointer within the current */
 771         /* packet. */
 772
 773         for(;;) {
 774                 u8 label_len;
 775                 if (j >= length) return -1;
 776                 GET8(label_len);
 777                 if (!label_len) break;
 778                 if (label_len & 0xc0) {
 779                         u8 ptr_low;
 780                         GET8(ptr_low);
 781                         if (name_end < 0) name_end = j;
 782                         j = (((int)label_len & 0x3f) << 8) + ptr_low;
 783                         /* Make sure that the target offset is in-bounds. */
 784                         if (j < 0 || j >= length) return -1;
 785                         /* If we've jumped more times than there are characters in the
 786                          * message, we must have a loop. */
 787                         if (++ptr_count > length) return -1;
 788                         continue;
 789                 }
 790                 if (label_len > 63) return -1;
 791                 if (cp != name_out) {
 792                         if (cp + 1 >= end) return -1;
 793                         *cp++ = '.';
 794                 }
 795                 if (cp + label_len >= end) return -1;
 796                 memcpy(cp, packet + j, label_len);
 797                 cp += label_len;
 798                 j += label_len;
 799         }
 800         if (cp >= end) return -1;
 801         *cp = '\0';
 802         if (name_end < 0)
 803                 *idx = j;
 804         else
 805                 *idx = name_end;
 806         return 0;
 807  err:
 808         return -1;
 809 }
 810
 811 /* parses a raw request from a nameserver */
 812 static int
 813 reply_parse(u8 *packet, int length) {
 814         int j = 0, k = 0;  /* index into packet */
 815         u16 _t;  /* used by the macros */
 816         u32 _t32;  /* used by the macros */
 817         char tmp_name[256], cmp_name[256]; /* used by the macros */
 818
 819         u16 trans_id, questions, answers, authority, additional, datalength;
 820         u16 flags = 0;
 821         u32 ttl, ttl_r = 0xffffffff;
 822         struct reply reply;
 823         struct request *req = NULL;
 824         unsigned int i;
 825
 826         GET16(trans_id);
 827         GET16(flags);
 828         GET16(questions);
 829         GET16(answers);
 830         GET16(authority);
 831         GET16(additional);
 832         (void) authority; /* suppress "unused variable" warnings. */
 833         (void) additional; /* suppress "unused variable" warnings. */
 834
 835         req = request_find_from_trans_id(trans_id);
 836         if (!req) return -1;
 837
 838         memset(&reply, 0, sizeof(reply));
 839
 840         /* If it's not an answer, it doesn't correspond to any request. */
 841         if (!(flags & 0x8000)) return -1;  /* must be an answer */
 842         if (flags & 0x020f) {
 843                 /* there was an error */
 844                 goto err;
 845         }
 846         /* if (!answers) return; */  /* must have an answer of some form */
 847
 848         /* This macro skips a name in the DNS reply. */
 849 #define SKIP_NAME \
 850         do { tmp_name[0] = '\0';                                \
 851                 if (name_parse(packet, length, &j, tmp_name, sizeof(tmp_name))<0)\
 852                         goto err;                               \
 853         } while(0)
 854 #define TEST_NAME \
 855         do { tmp_name[0] = '\0';                                \
 856                 cmp_name[0] = '\0';                             \
 857                 k = j;                                          \
 858                 if (name_parse(packet, length, &j, tmp_name, sizeof(tmp_name))<0)\
 859                         goto err;                                       \
 860                 if (name_parse(req->request, req->request_len, &k, cmp_name, sizeof(cmp_name))<0)       \
 861                         goto err;                               \
 862                 if (memcmp(tmp_name, cmp_name, strlen (tmp_name)) != 0) \
 863                         return (-1); /* we ignore mismatching names */  \
 864         } while(0)
 865
 866         reply.type = req->request_type;
 867
 868         /* skip over each question in the reply */
 869         for (i = 0; i < questions; ++i) {
 870                 /* the question looks like
 871                  *   <label:name><u16:type><u16:class>
 872                  */
 873                 TEST_NAME;
 874                 j += 4;
 875                 if (j > length) goto err;
 876         }
 877
 878         /* now we have the answer section which looks like
 879          * <label:name><u16:type><u16:class><u32:ttl><u16:len><data...>
 880          */
 881
 882         for (i = 0; i < answers; ++i) {
 883                 u16 type, class;
 884
 885                 SKIP_NAME;
 886                 GET16(type);
 887                 GET16(class);
 888                 GET32(ttl);
 889                 GET16(datalength);
 890
 891                 if (type == TYPE_A && class == CLASS_INET) {
 892                         int addrcount, addrtocopy;
 893                         if (req->request_type != TYPE_A) {
 894                                 j += datalength; continue;
 895                         }
 896                         if ((datalength & 3) != 0) /* not an even number of As. */
 897                             goto err;
 898                         addrcount = datalength >> 2;
 899                         addrtocopy = MIN(MAX_ADDRS - reply.data.a.addrcount, (unsigned)addrcount);
 900
 901                         ttl_r = MIN(ttl_r, ttl);
 902                         /* we only bother with the first four addresses. */
 903                         if (j + 4*addrtocopy > length) goto err;
 904                         memcpy(&reply.data.a.addresses[reply.data.a.addrcount],
 905                                    packet + j, 4*addrtocopy);
 906                         j += 4*addrtocopy;
 907                         reply.data.a.addrcount += addrtocopy;
 908                         reply.have_answer = 1;
 909                         if (reply.data.a.addrcount == MAX_ADDRS) break;
 910                 } else if (type == TYPE_PTR && class == CLASS_INET) {
 911                         if (req->request_type != TYPE_PTR) {
 912                                 j += datalength; continue;
 913                         }
 914                         if (name_parse(packet, length, &j, reply.data.ptr.name,
 915                                                    sizeof(reply.data.ptr.name))<0)
 916                                 goto err;
 917                         ttl_r = MIN(ttl_r, ttl);
 918                         reply.have_answer = 1;
 919                         break;
 920                 } else if (type == TYPE_AAAA && class == CLASS_INET) {
 921                         int addrcount, addrtocopy;
 922                         if (req->request_type != TYPE_AAAA) {
 923                                 j += datalength; continue;
 924                         }
 925                         if ((datalength & 15) != 0) /* not an even number of AAAAs. */
 926                                 goto err;
 927                         addrcount = datalength >> 4;  /* each address is 16 bytes long */
 928                         addrtocopy = MIN(MAX_ADDRS - reply.data.aaaa.addrcount, (unsigned)addrcount);
 929                         ttl_r = MIN(ttl_r, ttl);
 930
 931                         /* we only bother with the first four addresses. */
 932                         if (j + 16*addrtocopy > length) goto err;
 933                         memcpy(&reply.data.aaaa.addresses[reply.data.aaaa.addrcount],
 934                                    packet + j, 16*addrtocopy);
 935                         reply.data.aaaa.addrcount += addrtocopy;
 936                         j += 16*addrtocopy;
 937                         reply.have_answer = 1;
 938                         if (reply.data.aaaa.addrcount == MAX_ADDRS) break;
 939                 } else {
 940                         /* skip over any other type of resource */
 941                         j += datalength;
 942                 }
 943         }
 944
 945         reply_handle(req, flags, ttl_r, &reply);
 946         return 0;
 947  err:
 948         if (req)
 949                 reply_handle(req, flags, 0, NULL);
 950         return -1;
 951 }
 952
 953 /* Parse a raw request (packet,length) sent to a nameserver port (port) from */
 954 /* a DNS client (addr,addrlen), and if it's well-formed, call the corresponding */
 955 /* callback. */
 956 static int
 957 request_parse(u8 *packet, int length, struct evdns_server_port *port, struct sockaddr *addr, socklen_t addrlen)
 958 {
 959         int j = 0;      /* index into packet */
 960         u16 _t;  /* used by the macros */
 961         char tmp_name[256]; /* used by the macros */
 962
 963         int i;
 964         u16 trans_id, flags, questions, answers, authority, additional;
 965         struct server_request *server_req = NULL;
 966
 967         /* Get the header fields */
 968         GET16(trans_id);
 969         GET16(flags);
 970         GET16(questions);
 971         GET16(answers);
 972         GET16(authority);
 973         GET16(additional);
 974
 975         if (flags & 0x8000) return -1; /* Must not be an answer. */
 976         flags &= 0x0110; /* Only RD and CD get preserved. */
 977
 978         server_req = malloc(sizeof(struct server_request));
 979         if (server_req == NULL) return -1;
 980         memset(server_req, 0, sizeof(struct server_request));
 981
 982         server_req->trans_id = trans_id;
 983         memcpy(&server_req->addr, addr, addrlen);
 984         server_req->addrlen = addrlen;
 985
 986         server_req->base.flags = flags;
 987         server_req->base.nquestions = 0;
 988         server_req->base.questions = malloc(sizeof(struct evdns_server_question *) * questions);
 989         if (server_req->base.questions == NULL)
 990                 goto err;
 991
 992         for (i = 0; i < questions; ++i) {
 993                 u16 type, class;
 994                 struct evdns_server_question *q;
 995                 int namelen;
 996                 if (name_parse(packet, length, &j, tmp_name, sizeof(tmp_name))<0)
 997                         goto err;
 998                 GET16(type);
 999                 GET16(class);
1000                 namelen = strlen(tmp_name);
1001                 q = malloc(sizeof(struct evdns_server_question) + namelen);
1002                 if (!q)
1003                         goto err;
1004                 q->type = type;
1005                 q->dns_question_class = class;
1006                 memcpy(q->name, tmp_name, namelen+1);
1007                 server_req->base.questions[server_req->base.nquestions++] = q;
1008         }
1009
1010         /* Ignore answers, authority, and additional. */
1011
1012         server_req->port = port;
1013         port->refcnt++;
1014
1015         /* Only standard queries are supported. */
1016         if (flags & 0x7800) {
1017                 evdns_server_request_respond(&(server_req->base), DNS_ERR_NOTIMPL);
1018                 return -1;
1019         }
1020
1021         port->user_callback(&(server_req->base), port->user_data);
1022
1023         return 0;
1024 err:
1025         if (server_req) {
1026                 if (server_req->base.questions) {
1027                         for (i = 0; i < server_req->base.nquestions; ++i)
1028                                 free(server_req->base.questions[i]);
1029                         free(server_req->base.questions);
1030                 }
1031                 free(server_req);
1032         }
1033         return -1;
1034
1035 #undef SKIP_NAME
1036 #undef GET32
1037 #undef GET16
1038 #undef GET8
1039 }
1040
1041 static u16
1042 default_transaction_id_fn(void)
1043 {
1044         u16 trans_id;
1045 #ifdef DNS_USE_CPU_CLOCK_FOR_ID
1046         struct timespec ts;
1047         static int clkid = -1;
1048         if (clkid == -1) {
1049                 clkid = CLOCK_REALTIME;
1050 #ifdef CLOCK_MONOTONIC
1051                 if (clock_gettime(CLOCK_MONOTONIC, &ts) != -1)
1052                         clkid = CLOCK_MONOTONIC;
1053 #endif
1054         }
1055         if (clock_gettime(clkid, &ts) == -1)
1056                 event_err(1, "clock_gettime");
1057         trans_id = ts.tv_nsec & 0xffff;
1058 #endif
1059
1060 #ifdef DNS_USE_FTIME_FOR_ID
1061         struct _timeb tb;
1062         _ftime(&tb);
1063         trans_id = tb.millitm & 0xffff;
1064 #endif
1065
1066 #ifdef DNS_USE_GETTIMEOFDAY_FOR_ID
1067         struct timeval tv;
1068         evutil_gettimeofday(&tv, NULL);
1069         trans_id = tv.tv_usec & 0xffff;
1070 #endif
1071
1072 #ifdef DNS_USE_OPENSSL_FOR_ID
1073         if (RAND_pseudo_bytes((u8 *) &trans_id, 2) == -1) {
1074                 /* in the case that the RAND call fails we back */
1075                 /* down to using gettimeofday. */
1076                 /*
1077                   struct timeval tv;
1078                   evutil_gettimeofday(&tv, NULL);
1079                   trans_id = tv.tv_usec & 0xffff;
1080                 */
1081                 abort();
1082         }
1083 #endif
1084         return trans_id;
1085 }
1086
1087 static ev_uint16_t (*trans_id_function)(void) = default_transaction_id_fn;
1088
1089 void
1090 evdns_set_transaction_id_fn(ev_uint16_t (*fn)(void))
1091 {
1092         if (fn)
1093                 trans_id_function = fn;
1094         else
1095                 trans_id_function = default_transaction_id_fn;
1096 }
1097
1098 /* Try to choose a strong transaction id which isn't already in flight */
1099 static u16
1100 transaction_id_pick(void) {
1101         for (;;) {
1102                 const struct request *req = req_head, *started_at;
1103                 u16 trans_id = trans_id_function();
1104
1105                 if (trans_id == 0xffff) continue;
1106                 /* now check to see if that id is already inflight */
1107                 req = started_at = req_head;
1108                 if (req) {
1109                         do {
1110                                 if (req->trans_id == trans_id) break;
1111                                 req = req->next;
1112                         } while (req != started_at);
1113                 }
1114                 /* we didn't find it, so this is a good id */
1115                 if (req == started_at) return trans_id;
1116         }
1117 }
1118
1119 /* choose a namesever to use. This function will try to ignore */
1120 /* nameservers which we think are down and load balance across the rest */
1121 /* by updating the server_head global each time. */
1122 static struct nameserver *
1123 nameserver_pick(void) {
1124         struct nameserver *started_at = server_head, *picked;
1125         if (!server_head) return NULL;
1126
1127         /* if we don't have any good nameservers then there's no */
1128         /* point in trying to find one. */
1129         if (!global_good_nameservers) {
1130                 server_head = server_head->next;
1131                 return server_head;
1132         }
1133
1134         /* remember that nameservers are in a circular list */
1135         for (;;) {
1136                 if (server_head->state) {
1137                         /* we think this server is currently good */
1138                         picked = server_head;
1139                         server_head = server_head->next;
1140                         return picked;
1141                 }
1142
1143                 server_head = server_head->next;
1144                 if (server_head == started_at) {
1145                         /* all the nameservers seem to be down */
1146                         /* so we just return this one and hope for the */
1147                         /* best */
1148                         assert(global_good_nameservers == 0);
1149                         picked = server_head;
1150                         server_head = server_head->next;
1151                         return picked;
1152                 }
1153         }
1154 }
1155
1156 static int
1157 address_is_correct(struct nameserver *ns, struct sockaddr *sa, socklen_t slen)
1158 {
1159         struct sockaddr_in *sin = (struct sockaddr_in*) sa;
1160         if (sa->sa_family != AF_INET || slen != sizeof(struct sockaddr_in))
1161                 return 0;
1162         if (sin->sin_addr.s_addr != ns->address)
1163                 return 0;
1164         return 1;
1165 }
1166
1167 /* this is called when a namesever socket is ready for reading */
1168 static void
1169 nameserver_read(struct nameserver *ns) {
1170         u8 packet[1500];
1171         struct sockaddr_storage ss;
1172         socklen_t addrlen = sizeof(ss);
1173
1174         for (;;) {
1175                 const int r = recvfrom(ns->socket, packet, sizeof(packet), 0,
1176                     (struct sockaddr*)&ss, &addrlen);
1177                 if (r < 0) {
1178                         int err = last_error(ns->socket);
1179                         if (error_is_eagain(err)) return;
1180                         nameserver_failed(ns, strerror(err));
1181                         return;
1182                 }
1183                 if (!address_is_correct(ns, (struct sockaddr*)&ss, addrlen)) {
1184                         log(EVDNS_LOG_WARN, "Address mismatch on received "
1185                             "DNS packet.");
1186                         return;
1187                 }
1188                 ns->timedout = 0;
1189                 reply_parse(packet, r);
1190         }
1191 }
1192
1193 /* Read a packet from a DNS client on a server port s, parse it, and */
1194 /* act accordingly. */
1195 static void
1196 server_port_read(struct evdns_server_port *s) {
1197         u8 packet[1500];
1198         struct sockaddr_storage addr;
1199         socklen_t addrlen;
1200         int r;
1201
1202         for (;;) {
1203                 addrlen = sizeof(struct sockaddr_storage);
1204                 r = recvfrom(s->socket, packet, sizeof(packet), 0,
1205                                          (struct sockaddr*) &addr, &addrlen);
1206                 if (r < 0) {
1207                         int err = last_error(s->socket);
1208                         if (error_is_eagain(err)) return;
1209                         log(EVDNS_LOG_WARN, "Error %s (%d) while reading request.",
1210                                 strerror(err), err);
1211                         return;
1212                 }
1213                 request_parse(packet, r, s, (struct sockaddr*) &addr, addrlen);
1214         }
1215 }
1216
1217 /* Try to write all pending replies on a given DNS server port. */
1218 static void
1219 server_port_flush(struct evdns_server_port *port)
1220 {
1221         while (port->pending_replies) {
1222                 struct server_request *req = port->pending_replies;
1223                 int r = sendto(port->socket, req->response, req->response_len, 0,
1224                            (struct sockaddr*) &req->addr, req->addrlen);
1225                 if (r < 0) {
1226                         int err = last_error(port->socket);
1227                         if (error_is_eagain(err))
1228                                 return;
1229                         log(EVDNS_LOG_WARN, "Error %s (%d) while writing response to port; dropping", strerror(err), err);
1230                 }
1231                 if (server_request_free(req)) {
1232                         /* we released the last reference to req->port. */
1233                         return;
1234                 }
1235         }
1236
1237         /* We have no more pending requests; stop listening for 'writeable' events. */
1238         (void) event_del(&port->event);
1239         event_set(&port->event, port->socket, EV_READ | EV_PERSIST,
1240                           server_port_ready_callback, port);
1241         if (event_add(&port->event, NULL) < 0) {
1242                 log(EVDNS_LOG_WARN, "Error from libevent when adding event for DNS server.");
1243                 /* ???? Do more? */
1244         }
1245 }
1246
1247 /* set if we are waiting for the ability to write to this server. */
1248 /* if waiting is true then we ask libevent for EV_WRITE events, otherwise */
1249 /* we stop these events. */
1250 static void
1251 nameserver_write_waiting(struct nameserver *ns, char waiting) {
1252         if (ns->write_waiting == waiting) return;
1253
1254         ns->write_waiting = waiting;
1255         (void) event_del(&ns->event);
1256         event_set(&ns->event, ns->socket, EV_READ | (waiting ? EV_WRITE : 0) | EV_PERSIST,
1257                         nameserver_ready_callback, ns);
1258         if (event_add(&ns->event, NULL) < 0) {
1259           log(EVDNS_LOG_WARN, "Error from libevent when adding event for %s",
1260               debug_ntoa(ns->address));
1261           /* ???? Do more? */
1262         }
1263 }
1264
1265 /* a callback function. Called by libevent when the kernel says that */
1266 /* a nameserver socket is ready for writing or reading */
1267 static void
1268 nameserver_ready_callback(int fd, short events, void *arg) {
1269         struct nameserver *ns = (struct nameserver *) arg;
1270         (void)fd;
1271
1272         if (events & EV_WRITE) {
1273                 ns->choked = 0;
1274                 if (!evdns_transmit()) {
1275                         nameserver_write_waiting(ns, 0);
1276                 }
1277         }
1278         if (events & EV_READ) {
1279                 nameserver_read(ns);
1280         }
1281 }
1282
1283 /* a callback function. Called by libevent when the kernel says that */
1284 /* a server socket is ready for writing or reading. */
1285 static void
1286 server_port_ready_callback(int fd, short events, void *arg) {
1287         struct evdns_server_port *port = (struct evdns_server_port *) arg;
1288         (void) fd;
1289
1290         if (events & EV_WRITE) {
1291                 port->choked = 0;
1292                 server_port_flush(port);
1293         }
1294         if (events & EV_READ) {
1295                 server_port_read(port);
1296         }
1297 }
1298
1299 /* This is an inefficient representation; only use it via the dnslabel_table_*
1300  * functions, so that is can be safely replaced with something smarter later. */
1301 #define MAX_LABELS 128
1302 /* Structures used to implement name compression */
1303 struct dnslabel_entry { char *v; off_t pos; };
1304 struct dnslabel_table {
1305         int n_labels; /* number of current entries */
1306         /* map from name to position in message */
1307         struct dnslabel_entry labels[MAX_LABELS];
1308 };
1309
1310 /* Initialize dnslabel_table. */
1311 static void
1312 dnslabel_table_init(struct dnslabel_table *table)
1313 {
1314         table->n_labels = 0;
1315 }
1316
1317 /* Free all storage held by table, but not the table itself. */
1318 static void
1319 dnslabel_clear(struct dnslabel_table *table)
1320 {
1321         int i;
1322         for (i = 0; i < table->n_labels; ++i)
1323                 free(table->labels[i].v);
1324         table->n_labels = 0;
1325 }
1326
1327 /* return the position of the label in the current message, or -1 if the label */
1328 /* hasn't been used yet. */
1329 static int
1330 dnslabel_table_get_pos(const struct dnslabel_table *table, const char *label)
1331 {
1332         int i;
1333         for (i = 0; i < table->n_labels; ++i) {
1334                 if (!strcmp(label, table->labels[i].v))
1335                         return table->labels[i].pos;
1336         }
1337         return -1;
1338 }
1339
1340 /* remember that we've used the label at position pos */
1341 static int
1342 dnslabel_table_add(struct dnslabel_table *table, const char *label, off_t pos)
1343 {
1344         char *v;
1345         int p;
1346         if (table->n_labels == MAX_LABELS)
1347                 return (-1);
1348         v = strdup(label);
1349         if (v == NULL)
1350                 return (-1);
1351         p = table->n_labels++;
1352         table->labels[p].v = v;
1353         table->labels[p].pos = pos;
1354
1355         return (0);
1356 }
1357
1358 /* Converts a string to a length-prefixed set of DNS labels, starting */
1359 /* at buf[j]. name and buf must not overlap. name_len should be the length */
1360 /* of name.      table is optional, and is used for compression. */
1361 /* */
1362 /* Input: abc.def */
1363 /* Output: <3>abc<3>def<0> */
1364 /* */
1365 /* Returns the first index after the encoded name, or negative on error. */
1366 /*       -1      label was > 63 bytes */
1367 /*       -2      name too long to fit in buffer. */
1368 /* */
1369 static off_t
1370 dnsname_to_labels(u8 *const buf, size_t buf_len, off_t j,
1371                                   const char *name, const int name_len,
1372                                   struct dnslabel_table *table) {
1373         const char *end = name + name_len;
1374         int ref = 0;
1375         u16 _t;
1376
1377 #define APPEND16(x) do {                                                   \
1378                 if (j + 2 > (off_t)buf_len)                                \
1379                         goto overflow;                                             \
1380                 _t = htons(x);                                                     \
1381                 memcpy(buf + j, &_t, 2);                                   \
1382                 j += 2;                                                                    \
1383         } while (0)
1384 #define APPEND32(x) do {                                                   \
1385                 if (j + 4 > (off_t)buf_len)                                \
1386                         goto overflow;                                             \
1387                 _t32 = htonl(x);                                                   \
1388                 memcpy(buf + j, &_t32, 4);                                 \
1389                 j += 4;                                                                    \
1390         } while (0)
1391
1392         if (name_len > 255) return -2;
1393
1394         for (;;) {
1395                 const char *const start = name;
1396                 if (table && (ref = dnslabel_table_get_pos(table, name)) >= 0) {
1397                         APPEND16(ref | 0xc000);
1398                         return j;
1399                 }
1400                 name = strchr(name, '.');
1401                 if (!name) {
1402                         const unsigned int label_len = end - start;
1403                         if (label_len > 63) return -1;
1404                         if ((size_t)(j+label_len+1) > buf_len) return -2;
1405                         if (table) dnslabel_table_add(table, start, j);
1406                         buf[j++] = label_len;
1407
1408                         memcpy(buf + j, start, end - start);
1409                         j += end - start;
1410                         break;
1411                 } else {
1412                         /* append length of the label. */
1413                         const unsigned int label_len = name - start;
1414                         if (label_len > 63) return -1;
1415                         if ((size_t)(j+label_len+1) > buf_len) return -2;
1416                         if (table) dnslabel_table_add(table, start, j);
1417                         buf[j++] = label_len;
1418
1419                         memcpy(buf + j, start, name - start);
1420                         j += name - start;
1421                         /* hop over the '.' */
1422                         name++;
1423                 }
1424         }
1425
1426         /* the labels must be terminated by a 0. */
1427         /* It's possible that the name ended in a . */
1428         /* in which case the zero is already there */
1429         if (!j || buf[j-1]) buf[j++] = 0;
1430         return j;
1431  overflow:
1432         return (-2);
1433 }
1434
1435 /* Finds the length of a dns request for a DNS name of the given */
1436 /* length. The actual request may be smaller than the value returned */
1437 /* here */
1438 static int
1439 evdns_request_len(const int name_len) {
1440         return 96 + /* length of the DNS standard header */
1441                 name_len + 2 +
1442                 4;  /* space for the resource type */
1443 }
1444
1445 /* build a dns request packet into buf. buf should be at least as long */
1446 /* as evdns_request_len told you it should be. */
1447 /* */
1448 /* Returns the amount of space used. Negative on error. */
1449 static int
1450 evdns_request_data_build(const char *const name, const int name_len,
1451     const u16 trans_id, const u16 type, const u16 class,
1452     u8 *const buf, size_t buf_len) {
1453         off_t j = 0;  /* current offset into buf */
1454         u16 _t;  /* used by the macros */
1455
1456         APPEND16(trans_id);
1457         APPEND16(0x0100);  /* standard query, recusion needed */
1458         APPEND16(1);  /* one question */
1459         APPEND16(0);  /* no answers */
1460         APPEND16(0);  /* no authority */
1461         APPEND16(0);  /* no additional */
1462
1463         j = dnsname_to_labels(buf, buf_len, j, name, name_len, NULL);
1464         if (j < 0) {
1465                 return (int)j;
1466         }
1467
1468         APPEND16(type);
1469         APPEND16(class);
1470
1471         return (int)j;
1472  overflow:
1473         return (-1);
1474 }
1475
1476 /* exported function */
1477 struct evdns_server_port *
1478 evdns_add_server_port(int socket, int is_tcp, evdns_request_callback_fn_type cb, void *user_data)
1479 {
1480         struct evdns_server_port *port;
1481         if (!(port = malloc(sizeof(struct evdns_server_port))))
1482                 return NULL;
1483         memset(port, 0, sizeof(struct evdns_server_port));
1484
1485         assert(!is_tcp); /* TCP sockets not yet implemented */
1486         port->socket = socket;
1487         port->refcnt = 1;
1488         port->choked = 0;
1489         port->closing = 0;
1490         port->user_callback = cb;
1491         port->user_data = user_data;
1492         port->pending_replies = NULL;
1493
1494         event_set(&port->event, port->socket, EV_READ | EV_PERSIST,
1495                           server_port_ready_callback, port);
1496         event_add(&port->event, NULL); /* check return. */
1497         return port;
1498 }
1499
1500 /* exported function */
1501 void
1502 evdns_close_server_port(struct evdns_server_port *port)
1503 {
1504         if (--port->refcnt == 0)
1505                 server_port_free(port);
1506         port->closing = 1;
1507 }
1508
1509 /* exported function */
1510 int
1511 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)
1512 {
1513         struct server_request *req = TO_SERVER_REQUEST(_req);
1514         struct server_reply_item **itemp, *item;
1515         int *countp;
1516
1517         if (req->response) /* have we already answered? */
1518                 return (-1);
1519
1520         switch (section) {
1521         case EVDNS_ANSWER_SECTION:
1522                 itemp = &req->answer;
1523                 countp = &req->n_answer;
1524                 break;
1525         case EVDNS_AUTHORITY_SECTION:
1526                 itemp = &req->authority;
1527                 countp = &req->n_authority;
1528                 break;
1529         case EVDNS_ADDITIONAL_SECTION:
1530                 itemp = &req->additional;
1531                 countp = &req->n_additional;
1532                 break;
1533         default:
1534                 return (-1);
1535         }
1536         while (*itemp) {
1537                 itemp = &((*itemp)->next);
1538         }
1539         item = malloc(sizeof(struct server_reply_item));
1540         if (!item)
1541                 return -1;
1542         item->next = NULL;
1543         if (!(item->name = strdup(name))) {
1544                 free(item);
1545                 return -1;
1546         }
1547         item->type = type;
1548         item->dns_question_class = class;
1549         item->ttl = ttl;
1550         item->is_name = is_name != 0;
1551         item->datalen = 0;
1552         item->data = NULL;
1553         if (data) {
1554                 if (item->is_name) {
1555                         if (!(item->data = strdup(data))) {
1556                                 free(item->name);
1557                                 free(item);
1558                                 return -1;
1559                         }
1560                         item->datalen = (u16)-1;
1561                 } else {
1562                         if (!(item->data = malloc(datalen))) {
1563                                 free(item->name);
1564                                 free(item);
1565                                 return -1;
1566                         }
1567                         item->datalen = datalen;
1568                         memcpy(item->data, data, datalen);
1569                 }
1570         }
1571
1572         *itemp = item;
1573         ++(*countp);
1574         return 0;
1575 }
1576
1577 /* exported function */
1578 int
1579 evdns_server_request_add_a_reply(struct evdns_server_request *req, const char *name, int n, void *addrs, int ttl)
1580 {
1581         return evdns_server_request_add_reply(
1582                   req, EVDNS_ANSWER_SECTION, name, TYPE_A, CLASS_INET,
1583                   ttl, n*4, 0, addrs);
1584 }
1585
1586 /* exported function */
1587 int
1588 evdns_server_request_add_aaaa_reply(struct evdns_server_request *req, const char *name, int n, void *addrs, int ttl)
1589 {
1590         return evdns_server_request_add_reply(
1591                   req, EVDNS_ANSWER_SECTION, name, TYPE_AAAA, CLASS_INET,
1592                   ttl, n*16, 0, addrs);
1593 }
1594
1595 /* exported function */
1596 int
1597 evdns_server_request_add_ptr_reply(struct evdns_server_request *req, struct in_addr *in, const char *inaddr_name, const char *hostname, int ttl)
1598 {
1599         u32 a;
1600         char buf[32];
1601         assert(in || inaddr_name);
1602         assert(!(in && inaddr_name));
1603         if (in) {
1604                 a = ntohl(in->s_addr);
1605                 evutil_snprintf(buf, sizeof(buf), "%d.%d.%d.%d.in-addr.arpa",
1606                                 (int)(u8)((a    )&0xff),
1607                                 (int)(u8)((a>>8 )&0xff),
1608                                 (int)(u8)((a>>16)&0xff),
1609                                 (int)(u8)((a>>24)&0xff));
1610                 inaddr_name = buf;
1611         }
1612         return evdns_server_request_add_reply(
1613                   req, EVDNS_ANSWER_SECTION, inaddr_name, TYPE_PTR, CLASS_INET,
1614                   ttl, -1, 1, hostname);
1615 }
1616
1617 /* exported function */
1618 int
1619 evdns_server_request_add_cname_reply(struct evdns_server_request *req, const char *name, const char *cname, int ttl)
1620 {
1621         return evdns_server_request_add_reply(
1622                   req, EVDNS_ANSWER_SECTION, name, TYPE_CNAME, CLASS_INET,
1623                   ttl, -1, 1, cname);
1624 }
1625
1626
1627 static int
1628 evdns_server_request_format_response(struct server_request *req, int err)
1629 {
1630         unsigned char buf[1500];
1631         size_t buf_len = sizeof(buf);
1632         off_t j = 0, r;
1633         u16 _t;
1634         u32 _t32;
1635         int i;
1636         u16 flags;
1637         struct dnslabel_table table;
1638
1639         if (err < 0 || err > 15) return -1;
1640
1641         /* Set response bit and error code; copy OPCODE and RD fields from
1642          * question; copy RA and AA if set by caller. */
1643         flags = req->base.flags;
1644         flags |= (0x8000 | err);
1645
1646         dnslabel_table_init(&table);
1647         APPEND16(req->trans_id);
1648         APPEND16(flags);
1649         APPEND16(req->base.nquestions);
1650         APPEND16(req->n_answer);
1651         APPEND16(req->n_authority);
1652         APPEND16(req->n_additional);
1653
1654         /* Add questions. */
1655         for (i=0; i < req->base.nquestions; ++i) {
1656                 const char *s = req->base.questions[i]->name;
1657                 j = dnsname_to_labels(buf, buf_len, j, s, strlen(s), &table);
1658                 if (j < 0) {
1659                         dnslabel_clear(&table);
1660                         return (int) j;
1661                 }
1662                 APPEND16(req->base.questions[i]->type);
1663                 APPEND16(req->base.questions[i]->dns_question_class);
1664         }
1665
1666         /* Add answer, authority, and additional sections. */
1667         for (i=0; i<3; ++i) {
1668                 struct server_reply_item *item;
1669                 if (i==0)
1670                         item = req->answer;
1671                 else if (i==1)
1672                         item = req->authority;
1673                 else
1674                         item = req->additional;
1675                 while (item) {
1676                         r = dnsname_to_labels(buf, buf_len, j, item->name, strlen(item->name), &table);
1677                         if (r < 0)
1678                                 goto overflow;
1679                         j = r;
1680
1681                         APPEND16(item->type);
1682                         APPEND16(item->dns_question_class);
1683                         APPEND32(item->ttl);
1684                         if (item->is_name) {
1685                                 off_t len_idx = j, name_start;
1686                                 j += 2;
1687                                 name_start = j;
1688                                 r = dnsname_to_labels(buf, buf_len, j, item->data, strlen(item->data), &table);
1689                                 if (r < 0)
1690                                         goto overflow;
1691                                 j = r;
1692                                 _t = htons( (short) (j-name_start) );
1693                                 memcpy(buf+len_idx, &_t, 2);
1694                         } else {
1695                                 APPEND16(item->datalen);
1696                                 if (j+item->datalen > (off_t)buf_len)
1697                                         goto overflow;
1698                                 memcpy(buf+j, item->data, item->datalen);
1699                                 j += item->datalen;
1700                         }
1701                         item = item->next;
1702                 }
1703         }
1704
1705         if (j > 512) {
1706 overflow:
1707                 j = 512;
1708                 buf[2] |= 0x02; /* set the truncated bit. */
1709         }
1710
1711         req->response_len = j;
1712
1713         if (!(req->response = malloc(req->response_len))) {
1714                 server_request_free_answers(req);
1715                 dnslabel_clear(&table);
1716                 return (-1);
1717         }
1718         memcpy(req->response, buf, req->response_len);
1719         server_request_free_answers(req);
1720         dnslabel_clear(&table);
1721         return (0);
1722 }
1723
1724 /* exported function */
1725 int
1726 evdns_server_request_respond(struct evdns_server_request *_req, int err)
1727 {
1728         struct server_request *req = TO_SERVER_REQUEST(_req);
1729         struct evdns_server_port *port = req->port;
1730         int r;
1731         if (!req->response) {
1732                 if ((r = evdns_server_request_format_response(req, err))<0)
1733                         return r;
1734         }
1735
1736         r = sendto(port->socket, req->response, req->response_len, 0,
1737                            (struct sockaddr*) &req->addr, req->addrlen);
1738         if (r<0) {
1739                 int sock_err = last_error(port->socket);
1740                 if (! error_is_eagain(sock_err))
1741                         return -1;
1742
1743                 if (port->pending_replies) {
1744                         req->prev_pending = port->pending_replies->prev_pending;
1745                         req->next_pending = port->pending_replies;
1746                         req->prev_pending->next_pending =
1747                                 req->next_pending->prev_pending = req;
1748                 } else {
1749                         req->prev_pending = req->next_pending = req;
1750                         port->pending_replies = req;
1751                         port->choked = 1;
1752
1753                         (void) event_del(&port->event);
1754                         event_set(&port->event, port->socket, (port->closing?0:EV_READ) | EV_WRITE | EV_PERSIST, server_port_ready_callback, port);
1755
1756                         if (event_add(&port->event, NULL) < 0) {
1757                                 log(EVDNS_LOG_WARN, "Error from libevent when adding event for DNS server");
1758                         }
1759
1760                 }
1761
1762                 return 1;
1763         }
1764         if (server_request_free(req))
1765                 return 0;
1766
1767         if (port->pending_replies)
1768                 server_port_flush(port);
1769
1770         return 0;
1771 }
1772
1773 /* Free all storage held by RRs in req. */
1774 static void
1775 server_request_free_answers(struct server_request *req)
1776 {
1777         struct server_reply_item *victim, *next, **list;
1778         int i;
1779         for (i = 0; i < 3; ++i) {
1780                 if (i==0)
1781                         list = &req->answer;
1782                 else if (i==1)
1783                         list = &req->authority;
1784                 else
1785                         list = &req->additional;
1786
1787                 victim = *list;
1788                 while (victim) {
1789                         next = victim->next;
1790                         free(victim->name);
1791                         if (victim->data)
1792                                 free(victim->data);
1793                         free(victim);
1794                         victim = next;
1795                 }
1796                 *list = NULL;
1797         }
1798 }
1799
1800 /* Free all storage held by req, and remove links to it. */
1801 /* return true iff we just wound up freeing the server_port. */
1802 static int
1803 server_request_free(struct server_request *req)
1804 {
1805         int i, rc=1;
1806         if (req->base.questions) {
1807                 for (i = 0; i < req->base.nquestions; ++i)
1808                         free(req->base.questions[i]);
1809                 free(req->base.questions);
1810         }
1811
1812         if (req->port) {
1813                 if (req->port->pending_replies == req) {
1814                         if (req->next_pending)
1815                                 req->port->pending_replies = req->next_pending;
1816                         else
1817                                 req->port->pending_replies = NULL;
1818                 }
1819                 rc = --req->port->refcnt;
1820         }
1821
1822         if (req->response) {
1823                 free(req->response);
1824         }
1825
1826         server_request_free_answers(req);
1827
1828         if (req->next_pending && req->next_pending != req) {
1829                 req->next_pending->prev_pending = req->prev_pending;
1830                 req->prev_pending->next_pending = req->next_pending;
1831         }
1832
1833         if (rc == 0) {
1834                 server_port_free(req->port);
1835                 free(req);
1836                 return (1);
1837         }
1838         free(req);
1839         return (0);
1840 }
1841
1842 /* Free all storage held by an evdns_server_port.  Only called when  */
1843 static void
1844 server_port_free(struct evdns_server_port *port)
1845 {
1846         assert(port);
1847         assert(!port->refcnt);
1848         assert(!port->pending_replies);
1849         if (port->socket > 0) {
1850                 CLOSE_SOCKET(port->socket);
1851                 port->socket = -1;
1852         }
1853         (void) event_del(&port->event);
1854         /* XXXX actually free the port? -NM */
1855 }
1856
1857 /* exported function */
1858 int
1859 evdns_server_request_drop(struct evdns_server_request *_req)
1860 {
1861         struct server_request *req = TO_SERVER_REQUEST(_req);
1862         server_request_free(req);
1863         return 0;
1864 }
1865
1866 /* exported function */
1867 int
1868 evdns_server_request_get_requesting_addr(struct evdns_server_request *_req, struct sockaddr *sa, int addr_len)
1869 {
1870         struct server_request *req = TO_SERVER_REQUEST(_req);
1871         if (addr_len < (int)req->addrlen)
1872                 return -1;
1873         memcpy(sa, &(req->addr), req->addrlen);
1874         return req->addrlen;
1875 }
1876
1877 #undef APPEND16
1878 #undef APPEND32
1879
1880 /* this is a libevent callback function which is called when a request */
1881 /* has timed out. */
1882 static void
1883 evdns_request_timeout_callback(int fd, short events, void *arg) {
1884         struct request *const req = (struct request *) arg;
1885         (void) fd;
1886         (void) events;
1887
1888         log(EVDNS_LOG_DEBUG, "Request %lx timed out", (unsigned long) arg);
1889
1890         req->ns->timedout++;
1891         if (req->ns->timedout > global_max_nameserver_timeout) {
1892                 req->ns->timedout = 0;
1893                 nameserver_failed(req->ns, "request timed out.");
1894         }
1895
1896         (void) evtimer_del(&req->timeout_event);
1897         if (req->tx_count >= global_max_retransmits) {
1898                 /* this request has failed */
1899                 reply_callback(req, 0, DNS_ERR_TIMEOUT, NULL);
1900                 request_finished(req, &req_head);
1901         } else {
1902                 /* retransmit it */
1903                 evdns_request_transmit(req);
1904         }
1905 }
1906
1907 /* try to send a request to a given server. */
1908 /* */
1909 /* return: */
1910 /*   0 ok */
1911 /*   1 temporary failure */
1912 /*   2 other failure */
1913 static int
1914 evdns_request_transmit_to(struct request *req, struct nameserver *server) {
1915         struct sockaddr_in sin;
1916         int r;
1917         memset(&sin, 0, sizeof(sin));
1918         sin.sin_addr.s_addr = req->ns->address;
1919         sin.sin_port = req->ns->port;
1920         sin.sin_family = AF_INET;
1921
1922         r = sendto(server->socket, req->request, req->request_len, 0,
1923             (struct sockaddr*)&sin, sizeof(sin));
1924         if (r < 0) {
1925                 int err = last_error(server->socket);
1926                 if (error_is_eagain(err)) return 1;
1927                 nameserver_failed(req->ns, strerror(err));
1928                 return 2;
1929         } else if (r != (int)req->request_len) {
1930                 return 1;  /* short write */
1931         } else {
1932                 return 0;
1933         }
1934 }
1935
1936 /* try to send a request, updating the fields of the request */
1937 /* as needed */
1938 /* */
1939 /* return: */
1940 /*   0 ok */
1941 /*   1 failed */
1942 static int
1943 evdns_request_transmit(struct request *req) {
1944         int retcode = 0, r;
1945
1946         /* if we fail to send this packet then this flag marks it */
1947         /* for evdns_transmit */
1948         req->transmit_me = 1;
1949         if (req->trans_id == 0xffff) abort();
1950
1951         if (req->ns->choked) {
1952                 /* don't bother trying to write to a socket */
1953                 /* which we have had EAGAIN from */
1954                 return 1;
1955         }
1956
1957         r = evdns_request_transmit_to(req, req->ns);
1958         switch (r) {
1959         case 1:
1960                 /* temp failure */
1961                 req->ns->choked = 1;
1962                 nameserver_write_waiting(req->ns, 1);
1963                 return 1;
1964         case 2:
1965                 /* failed in some other way */
1966                 retcode = 1;
1967                 /* fall through */
1968         default:
1969                 /* all ok */
1970                 log(EVDNS_LOG_DEBUG,
1971                     "Setting timeout for request %lx", (unsigned long) req);
1972                 if (evtimer_add(&req->timeout_event, &global_timeout) < 0) {
1973                   log(EVDNS_LOG_WARN,
1974                       "Error from libevent when adding timer for request %lx",
1975                       (unsigned long) req);
1976                   /* ???? Do more? */
1977                 }
1978                 req->tx_count++;
1979                 req->transmit_me = 0;
1980                 return retcode;
1981         }
1982 }
1983
1984 static void
1985 nameserver_probe_callback(int result, char type, int count, int ttl, void *addresses, void *arg) {
1986         struct nameserver *const ns = (struct nameserver *) arg;
1987         (void) type;
1988         (void) count;
1989         (void) ttl;
1990         (void) addresses;
1991
1992         if (result == DNS_ERR_NONE || result == DNS_ERR_NOTEXIST) {
1993                 /* this is a good reply */
1994                 nameserver_up(ns);
1995         } else nameserver_probe_failed(ns);
1996 }
1997
1998 static void
1999 nameserver_send_probe(struct nameserver *const ns) {
2000         struct request *req;
2001         /* here we need to send a probe to a given nameserver */
2002         /* in the hope that it is up now. */
2003
2004         log(EVDNS_LOG_DEBUG, "Sending probe to %s", debug_ntoa(ns->address));
2005
2006         req = request_new(TYPE_A, "www.google.com", DNS_QUERY_NO_SEARCH, nameserver_probe_callback, ns);
2007         if (!req) return;
2008         /* we force this into the inflight queue no matter what */
2009         request_trans_id_set(req, transaction_id_pick());
2010         req->ns = ns;
2011         request_submit(req);
2012 }
2013
2014 /* returns: */
2015 /*   0 didn't try to transmit anything */
2016 /*   1 tried to transmit something */
2017 static int
2018 evdns_transmit(void) {
2019         char did_try_to_transmit = 0;
2020
2021         if (req_head) {
2022                 struct request *const started_at = req_head, *req = req_head;
2023                 /* first transmit all the requests which are currently waiting */
2024                 do {
2025                         if (req->transmit_me) {
2026                                 did_try_to_transmit = 1;
2027                                 evdns_request_transmit(req);
2028                         }
2029
2030                         req = req->next;
2031                 } while (req != started_at);
2032         }
2033
2034         return did_try_to_transmit;
2035 }
2036
2037 /* exported function */
2038 int
2039 evdns_count_nameservers(void)
2040 {
2041         const struct nameserver *server = server_head;
2042         int n = 0;
2043         if (!server)
2044                 return 0;
2045         do {
2046                 ++n;
2047                 server = server->next;
2048         } while (server != server_head);
2049         return n;
2050 }
2051
2052 /* exported function */
2053 int
2054 evdns_clear_nameservers_and_suspend(void)
2055 {
2056         struct nameserver *server = server_head, *started_at = server_head;
2057         struct request *req = req_head, *req_started_at = req_head;
2058
2059         if (!server)
2060                 return 0;
2061         while (1) {
2062                 struct nameserver *next = server->next;
2063                 (void) event_del(&server->event);
2064                 if (evtimer_initialized(&server->timeout_event))
2065                         (void) evtimer_del(&server->timeout_event);
2066                 if (server->socket >= 0)
2067                         CLOSE_SOCKET(server->socket);
2068                 free(server);
2069                 if (next == started_at)
2070                         break;
2071                 server = next;
2072         }
2073         server_head = NULL;
2074         global_good_nameservers = 0;
2075
2076         while (req) {
2077                 struct request *next = req->next;
2078                 req->tx_count = req->reissue_count = 0;
2079                 req->ns = NULL;
2080                 /* ???? What to do about searches? */
2081                 (void) evtimer_del(&req->timeout_event);
2082                 req->trans_id = 0;
2083                 req->transmit_me = 0;
2084
2085                 global_requests_waiting++;
2086                 evdns_request_insert(req, &req_waiting_head);
2087                 /* We want to insert these suspended elements at the front of
2088                  * the waiting queue, since they were pending before any of
2089                  * the waiting entries were added.  This is a circular list,
2090                  * so we can just shift the start back by one.*/
2091                 req_waiting_head = req_waiting_head->prev;
2092
2093                 if (next == req_started_at)
2094                         break;
2095                 req = next;
2096         }
2097         req_head = NULL;
2098         global_requests_inflight = 0;
2099
2100         return 0;
2101 }
2102
2103
2104 /* exported function */
2105 int
2106 evdns_resume(void)
2107 {
2108         evdns_requests_pump_waiting_queue();
2109         return 0;
2110 }
2111
2112 static int
2113 _evdns_nameserver_add_impl(unsigned long int address, int port) {
2114         /* first check to see if we already have this nameserver */
2115
2116         const struct nameserver *server = server_head, *const started_at = server_head;
2117         struct nameserver *ns;
2118         int err = 0;
2119         if (server) {
2120                 do {
2121                         if (server->address == address) return 3;
2122                         server = server->next;
2123                 } while (server != started_at);
2124         }
2125
2126         ns = (struct nameserver *) malloc(sizeof(struct nameserver));
2127         if (!ns) return -1;
2128
2129         memset(ns, 0, sizeof(struct nameserver));
2130
2131         evtimer_set(&ns->timeout_event, nameserver_prod_callback, ns);
2132
2133         ns->socket = socket(PF_INET, SOCK_DGRAM, 0);
2134         if (ns->socket < 0) { err = 1; goto out1; }
2135         evutil_make_socket_nonblocking(ns->socket);
2136
2137         ns->address = address;
2138         ns->port = htons(port);
2139         ns->state = 1;
2140         event_set(&ns->event, ns->socket, EV_READ | EV_PERSIST, nameserver_ready_callback, ns);
2141         if (event_add(&ns->event, NULL) < 0) {
2142           err = 2;
2143           goto out2;
2144         }
2145
2146         log(EVDNS_LOG_DEBUG, "Added nameserver %s", debug_ntoa(address));
2147
2148         /* insert this nameserver into the list of them */
2149         if (!server_head) {
2150                 ns->next = ns->prev = ns;
2151                 server_head = ns;
2152         } else {
2153                 ns->next = server_head->next;
2154                 ns->prev = server_head;
2155                 server_head->next = ns;
2156                 if (server_head->prev == server_head) {
2157                         server_head->prev = ns;
2158                 }
2159         }
2160
2161         global_good_nameservers++;
2162
2163         return 0;
2164
2165 out2:
2166         CLOSE_SOCKET(ns->socket);
2167 out1:
2168         free(ns);
2169         log(EVDNS_LOG_WARN, "Unable to add nameserver %s: error %d", debug_ntoa(address), err);
2170         return err;
2171 }
2172
2173 /* exported function */
2174 int
2175 evdns_nameserver_add(unsigned long int address) {
2176         return _evdns_nameserver_add_impl(address, 53);
2177 }
2178
2179 /* exported function */
2180 int
2181 evdns_nameserver_ip_add(const char *ip_as_string) {
2182         struct in_addr ina;
2183         int port;
2184         char buf[20];
2185         const char *cp;
2186         cp = strchr(ip_as_string, ':');
2187         if (! cp) {
2188                 cp = ip_as_string;
2189                 port = 53;
2190         } else {
2191                 port = strtoint(cp+1);
2192                 if (port < 0 || port > 65535) {
2193                         return 4;
2194                 }
2195                 if ((cp-ip_as_string) >= (int)sizeof(buf)) {
2196                         return 4;
2197                 }
2198                 memcpy(buf, ip_as_string, cp-ip_as_string);
2199                 buf[cp-ip_as_string] = '\0';
2200                 cp = buf;
2201         }
2202         if (!inet_aton(cp, &ina)) {
2203                 return 4;
2204         }
2205         return _evdns_nameserver_add_impl(ina.s_addr, port);
2206 }
2207
2208 /* insert into the tail of the queue */
2209 static void
2210 evdns_request_insert(struct request *req, struct request **head) {
2211         if (!*head) {
2212                 *head = req;
2213                 req->next = req->prev = req;
2214                 return;
2215         }
2216
2217         req->prev = (*head)->prev;
2218         req->prev->next = req;
2219         req->next = *head;
2220         (*head)->prev = req;
2221 }
2222
2223 static int
2224 string_num_dots(const char *s) {
2225         int count = 0;
2226         while ((s = strchr(s, '.'))) {
2227                 s++;
2228                 count++;
2229         }
2230         return count;
2231 }
2232
2233 static struct request *
2234 request_new(int type, const char *name, int flags,
2235     evdns_callback_type callback, void *user_ptr) {
2236         const char issuing_now =
2237             (global_requests_inflight < global_max_requests_inflight) ? 1 : 0;
2238
2239         const int name_len = strlen(name);
2240         const int request_max_len = evdns_request_len(name_len);
2241         const u16 trans_id = issuing_now ? transaction_id_pick() : 0xffff;
2242         /* the request data is alloced in a single block with the header */
2243         struct request *const req =
2244             (struct request *) malloc(sizeof(struct request) + request_max_len);
2245         int rlen;
2246         (void) flags;
2247
2248         if (!req) return NULL;
2249         memset(req, 0, sizeof(struct request));
2250
2251         evtimer_set(&req->timeout_event, evdns_request_timeout_callback, req);
2252
2253         /* request data lives just after the header */
2254         req->request = ((u8 *) req) + sizeof(struct request);
2255         /* denotes that the request data shouldn't be free()ed */
2256         req->request_appended = 1;
2257         rlen = evdns_request_data_build(name, name_len, trans_id,
2258             type, CLASS_INET, req->request, request_max_len);
2259         if (rlen < 0)
2260                 goto err1;
2261         req->request_len = rlen;
2262         req->trans_id = trans_id;
2263         req->tx_count = 0;
2264         req->request_type = type;
2265         req->user_pointer = user_ptr;
2266         req->user_callback = callback;
2267         req->ns = issuing_now ? nameserver_pick() : NULL;
2268         req->next = req->prev = NULL;
2269
2270         return req;
2271 err1:
2272         free(req);
2273         return NULL;
2274 }
2275
2276 static void
2277 request_submit(struct request *const req) {
2278         if (req->ns) {
2279                 /* if it has a nameserver assigned then this is going */
2280                 /* straight into the inflight queue */
2281                 evdns_request_insert(req, &req_head);
2282                 global_requests_inflight++;
2283                 evdns_request_transmit(req);
2284         } else {
2285                 evdns_request_insert(req, &req_waiting_head);
2286                 global_requests_waiting++;
2287         }
2288 }
2289
2290 /* exported function */
2291 int evdns_resolve_ipv4(const char *name, int flags,
2292     evdns_callback_type callback, void *ptr) {
2293         log(EVDNS_LOG_DEBUG, "Resolve requested for %s", name);
2294         if (flags & DNS_QUERY_NO_SEARCH) {
2295                 struct request *const req =
2296                         request_new(TYPE_A, name, flags, callback, ptr);
2297                 if (req == NULL)
2298                         return (1);
2299                 request_submit(req);
2300                 return (0);
2301         } else {
2302                 return (search_request_new(TYPE_A, name, flags, callback, ptr));
2303         }
2304 }
2305
2306 /* exported function */
2307 int evdns_resolve_ipv6(const char *name, int flags,
2308                                            evdns_callback_type callback, void *ptr) {
2309         log(EVDNS_LOG_DEBUG, "Resolve requested for %s", name);
2310         if (flags & DNS_QUERY_NO_SEARCH) {
2311                 struct request *const req =
2312                         request_new(TYPE_AAAA, name, flags, callback, ptr);
2313                 if (req == NULL)
2314                         return (1);
2315                 request_submit(req);
2316                 return (0);
2317         } else {
2318                 return (search_request_new(TYPE_AAAA, name, flags, callback, ptr));
2319         }
2320 }
2321
2322 int evdns_resolve_reverse(const struct in_addr *in, int flags, evdns_callback_type callback, void *ptr) {
2323         char buf[32];
2324         struct request *req;
2325         u32 a;
2326         assert(in);
2327         a = ntohl(in->s_addr);
2328         evutil_snprintf(buf, sizeof(buf), "%d.%d.%d.%d.in-addr.arpa",
2329                         (int)(u8)((a    )&0xff),
2330                         (int)(u8)((a>>8 )&0xff),
2331                         (int)(u8)((a>>16)&0xff),
2332                         (int)(u8)((a>>24)&0xff));
2333         log(EVDNS_LOG_DEBUG, "Resolve requested for %s (reverse)", buf);
2334         req = request_new(TYPE_PTR, buf, flags, callback, ptr);
2335         if (!req) return 1;
2336         request_submit(req);
2337         return 0;
2338 }
2339
2340 int evdns_resolve_reverse_ipv6(const struct in6_addr *in, int flags, evdns_callback_type callback, void *ptr) {
2341         /* 32 nybbles, 32 periods, "ip6.arpa", NUL. */
2342         char buf[73];
2343         char *cp;
2344         struct request *req;
2345         int i;
2346         assert(in);
2347         cp = buf;
2348         for (i=15; i >= 0; --i) {
2349                 u8 byte = in->s6_addr[i];
2350                 *cp++ = "0123456789abcdef"[byte & 0x0f];
2351                 *cp++ = '.';
2352                 *cp++ = "0123456789abcdef"[byte >> 4];
2353                 *cp++ = '.';
2354         }
2355         assert(cp + strlen("ip6.arpa") < buf+sizeof(buf));
2356         memcpy(cp, "ip6.arpa", strlen("ip6.arpa")+1);
2357         log(EVDNS_LOG_DEBUG, "Resolve requested for %s (reverse)", buf);
2358         req = request_new(TYPE_PTR, buf, flags, callback, ptr);
2359         if (!req) return 1;
2360         request_submit(req);
2361         return 0;
2362 }
2363
2364 /*/////////////////////////////////////////////////////////////////// */
2365 /* Search support */
2366 /* */
2367 /* the libc resolver has support for searching a number of domains */
2368 /* to find a name. If nothing else then it takes the single domain */
2369 /* from the gethostname() call. */
2370 /* */
2371 /* It can also be configured via the domain and search options in a */
2372 /* resolv.conf. */
2373 /* */
2374 /* The ndots option controls how many dots it takes for the resolver */
2375 /* to decide that a name is non-local and so try a raw lookup first. */
2376
2377 struct search_domain {
2378         int len;
2379         struct search_domain *next;
2380         /* the text string is appended to this structure */
2381 };
2382
2383 struct search_state {
2384         int refcount;
2385         int ndots;
2386         int num_domains;
2387         struct search_domain *head;
2388 };
2389
2390 static struct search_state *global_search_state = NULL;
2391
2392 static void
2393 search_state_decref(struct search_state *const state) {
2394         if (!state) return;
2395         state->refcount--;
2396         if (!state->refcount) {
2397                 struct search_domain *next, *dom;
2398                 for (dom = state->head; dom; dom = next) {
2399                         next = dom->next;
2400                         free(dom);
2401                 }
2402                 free(state);
2403         }
2404 }
2405
2406 static struct search_state *
2407 search_state_new(void) {
2408         struct search_state *state = (struct search_state *) malloc(sizeof(struct search_state));
2409         if (!state) return NULL;
2410         memset(state, 0, sizeof(struct search_state));
2411         state->refcount = 1;
2412         state->ndots = 1;
2413
2414         return state;
2415 }
2416
2417 static void
2418 search_postfix_clear(void) {
2419         search_state_decref(global_search_state);
2420
2421         global_search_state = search_state_new();
2422 }
2423
2424 /* exported function */
2425 void
2426 evdns_search_clear(void) {
2427         search_postfix_clear();
2428 }
2429
2430 static void
2431 search_postfix_add(const char *domain) {
2432         int domain_len;
2433         struct search_domain *sdomain;
2434         while (domain[0] == '.') domain++;
2435         domain_len = strlen(domain);
2436
2437         if (!global_search_state) global_search_state = search_state_new();
2438         if (!global_search_state) return;
2439         global_search_state->num_domains++;
2440
2441         sdomain = (struct search_domain *) malloc(sizeof(struct search_domain) + domain_len);
2442         if (!sdomain) return;
2443         memcpy( ((u8 *) sdomain) + sizeof(struct search_domain), domain, domain_len);
2444         sdomain->next = global_search_state->head;
2445         sdomain->len = domain_len;
2446
2447         global_search_state->head = sdomain;
2448 }
2449
2450 /* reverse the order of members in the postfix list. This is needed because, */
2451 /* when parsing resolv.conf we push elements in the wrong order */
2452 static void
2453 search_reverse(void) {
2454         struct search_domain *cur, *prev = NULL, *next;
2455         cur = global_search_state->head;
2456         while (cur) {
2457                 next = cur->next;
2458                 cur->next = prev;
2459                 prev = cur;
2460                 cur = next;
2461         }
2462
2463         global_search_state->head = prev;
2464 }
2465
2466 /* exported function */
2467 void
2468 evdns_search_add(const char *domain) {
2469         search_postfix_add(domain);
2470 }
2471
2472 /* exported function */
2473 void
2474 evdns_search_ndots_set(const int ndots) {
2475         if (!global_search_state) global_search_state = search_state_new();
2476         if (!global_search_state) return;
2477         global_search_state->ndots = ndots;
2478 }
2479
2480 static void
2481 search_set_from_hostname(void) {
2482         char hostname[HOST_NAME_MAX + 1], *domainname;
2483
2484         search_postfix_clear();
2485         if (gethostname(hostname, sizeof(hostname))) return;
2486         domainname = strchr(hostname, '.');
2487         if (!domainname) return;
2488         search_postfix_add(domainname);
2489 }
2490
2491 /* warning: returns malloced string */
2492 static char *
2493 search_make_new(const struct search_state *const state, int n, const char *const base_name) {
2494         const int base_len = strlen(base_name);
2495         const char need_to_append_dot = base_name[base_len - 1] == '.' ? 0 : 1;
2496         struct search_domain *dom;
2497
2498         for (dom = state->head; dom; dom = dom->next) {
2499                 if (!n--) {
2500                         /* this is the postfix we want */
2501                         /* the actual postfix string is kept at the end of the structure */
2502                         const u8 *const postfix = ((u8 *) dom) + sizeof(struct search_domain);
2503                         const int postfix_len = dom->len;
2504                         char *const newname = (char *) malloc(base_len + need_to_append_dot + postfix_len + 1);
2505                         if (!newname) return NULL;
2506                         memcpy(newname, base_name, base_len);
2507                         if (need_to_append_dot) newname[base_len] = '.';
2508                         memcpy(newname + base_len + need_to_append_dot, postfix, postfix_len);
2509                         newname[base_len + need_to_append_dot + postfix_len] = 0;
2510                         return newname;
2511                 }
2512         }
2513
2514         /* we ran off the end of the list and still didn't find the requested string */
2515         abort();
2516         return NULL; /* unreachable; stops warnings in some compilers. */
2517 }
2518
2519 static int
2520 search_request_new(int type, const char *const name, int flags, evdns_callback_type user_callback, void *user_arg) {
2521         assert(type == TYPE_A || type == TYPE_AAAA);
2522         if ( ((flags & DNS_QUERY_NO_SEARCH) == 0) &&
2523              global_search_state &&
2524                  global_search_state->num_domains) {
2525                 /* we have some domains to search */
2526                 struct request *req;
2527                 if (string_num_dots(name) >= global_search_state->ndots) {
2528                         req = request_new(type, name, flags, user_callback, user_arg);
2529                         if (!req) return 1;
2530                         req->search_index = -1;
2531                 } else {
2532                         char *const new_name = search_make_new(global_search_state, 0, name);
2533                         if (!new_name) return 1;
2534                         req = request_new(type, new_name, flags, user_callback, user_arg);
2535                         free(new_name);
2536                         if (!req) return 1;
2537                         req->search_index = 0;
2538                 }
2539                 req->search_origname = strdup(name);
2540                 req->search_state = global_search_state;
2541                 req->search_flags = flags;
2542                 global_search_state->refcount++;
2543                 request_submit(req);
2544                 return 0;
2545         } else {
2546                 struct request *const req = request_new(type, name, flags, user_callback, user_arg);
2547                 if (!req) return 1;
2548                 request_submit(req);
2549                 return 0;
2550         }
2551 }
2552
2553 /* this is called when a request has failed to find a name. We need to check */
2554 /* if it is part of a search and, if so, try the next name in the list */
2555 /* returns: */
2556 /*   0 another request has been submitted */
2557 /*   1 no more requests needed */
2558 static int
2559 search_try_next(struct request *const req) {
2560         if (req->search_state) {
2561                 /* it is part of a search */
2562                 char *new_name;
2563                 struct request *newreq;
2564                 req->search_index++;
2565                 if (req->search_index >= req->search_state->num_domains) {
2566                         /* no more postfixes to try, however we may need to try */
2567                         /* this name without a postfix */
2568                         if (string_num_dots(req->search_origname) < req->search_state->ndots) {
2569                                 /* yep, we need to try it raw */
2570                                 newreq = request_new(req->request_type, req->search_origname, req->search_flags, req->user_callback, req->user_pointer);
2571                                 log(EVDNS_LOG_DEBUG, "Search: trying raw query %s", req->search_origname);
2572                                 if (newreq) {
2573                                         request_submit(newreq);
2574                                         return 0;
2575                                 }
2576                         }
2577                         return 1;
2578                 }
2579
2580                 new_name = search_make_new(req->search_state, req->search_index, req->search_origname);
2581                 if (!new_name) return 1;
2582                 log(EVDNS_LOG_DEBUG, "Search: now trying %s (%d)", new_name, req->search_index);
2583                 newreq = request_new(req->request_type, new_name, req->search_flags, req->user_callback, req->user_pointer);
2584                 free(new_name);
2585                 if (!newreq) return 1;
2586                 newreq->search_origname = req->search_origname;
2587                 req->search_origname = NULL;
2588                 newreq->search_state = req->search_state;
2589                 newreq->search_flags = req->search_flags;
2590                 newreq->search_index = req->search_index;
2591                 newreq->search_state->refcount++;
2592                 request_submit(newreq);
2593                 return 0;
2594         }
2595         return 1;
2596 }
2597
2598 static void
2599 search_request_finished(struct request *const req) {
2600         if (req->search_state) {
2601                 search_state_decref(req->search_state);
2602                 req->search_state = NULL;
2603         }
2604         if (req->search_origname) {
2605                 free(req->search_origname);
2606                 req->search_origname = NULL;
2607         }
2608 }
2609
2610 /*/////////////////////////////////////////////////////////////////// */
2611 /* Parsing resolv.conf files */
2612
2613 static void
2614 evdns_resolv_set_defaults(int flags) {
2615         /* if the file isn't found then we assume a local resolver */
2616         if (flags & DNS_OPTION_SEARCH) search_set_from_hostname();
2617         if (flags & DNS_OPTION_NAMESERVERS) evdns_nameserver_ip_add("127.0.0.1");
2618 }
2619
2620 #ifndef HAVE_STRTOK_R
2621 static char *
2622 strtok_r(char *s, const char *delim, char **state) {
2623         return strtok(s, delim);
2624 }
2625 #endif
2626
2627 /* helper version of atoi which returns -1 on error */
2628 static int
2629 strtoint(const char *const str) {
2630         char *endptr;
2631         const int r = strtol(str, &endptr, 10);
2632         if (*endptr) return -1;
2633         return r;
2634 }
2635
2636 /* helper version of atoi that returns -1 on error and clips to bounds. */
2637 static int
2638 strtoint_clipped(const char *const str, int min, int max)
2639 {
2640         int r = strtoint(str);
2641         if (r == -1)
2642                 return r;
2643         else if (r<min)
2644                 return min;
2645         else if (r>max)
2646                 return max;
2647         else
2648                 return r;
2649 }
2650
2651 /* exported function */
2652 int
2653 evdns_set_option(const char *option, const char *val, int flags)
2654 {
2655         if (!strncmp(option, "ndots:", 6)) {
2656                 const int ndots = strtoint(val);
2657                 if (ndots == -1) return -1;
2658                 if (!(flags & DNS_OPTION_SEARCH)) return 0;
2659                 log(EVDNS_LOG_DEBUG, "Setting ndots to %d", ndots);
2660                 if (!global_search_state) global_search_state = search_state_new();
2661                 if (!global_search_state) return -1;
2662                 global_search_state->ndots = ndots;
2663         } else if (!strncmp(option, "timeout:", 8)) {
2664                 const int timeout = strtoint(val);
2665                 if (timeout == -1) return -1;
2666                 if (!(flags & DNS_OPTION_MISC)) return 0;
2667                 log(EVDNS_LOG_DEBUG, "Setting timeout to %d", timeout);
2668                 global_timeout.tv_sec = timeout;
2669         } else if (!strncmp(option, "max-timeouts:", 12)) {
2670                 const int maxtimeout = strtoint_clipped(val, 1, 255);
2671                 if (maxtimeout == -1) return -1;
2672                 if (!(flags & DNS_OPTION_MISC)) return 0;
2673                 log(EVDNS_LOG_DEBUG, "Setting maximum allowed timeouts to %d",
2674                         maxtimeout);
2675                 global_max_nameserver_timeout = maxtimeout;
2676         } else if (!strncmp(option, "max-inflight:", 13)) {
2677                 const int maxinflight = strtoint_clipped(val, 1, 65000);
2678                 if (maxinflight == -1) return -1;
2679                 if (!(flags & DNS_OPTION_MISC)) return 0;
2680                 log(EVDNS_LOG_DEBUG, "Setting maximum inflight requests to %d",
2681                         maxinflight);
2682                 global_max_requests_inflight = maxinflight;
2683         } else if (!strncmp(option, "attempts:", 9)) {
2684                 int retries = strtoint(val);
2685                 if (retries == -1) return -1;
2686                 if (retries > 255) retries = 255;
2687                 if (!(flags & DNS_OPTION_MISC)) return 0;
2688                 log(EVDNS_LOG_DEBUG, "Setting retries to %d", retries);
2689                 global_max_retransmits = retries;
2690         }
2691         return 0;
2692 }
2693
2694 static void
2695 resolv_conf_parse_line(char *const start, int flags) {
2696         char *strtok_state;
2697         static const char *const delims = " \t";
2698 #define NEXT_TOKEN strtok_r(NULL, delims, &strtok_state)
2699
2700         char *const first_token = strtok_r(start, delims, &strtok_state);
2701         if (!first_token) return;
2702
2703         if (!strcmp(first_token, "nameserver") && (flags & DNS_OPTION_NAMESERVERS)) {
2704                 const char *const nameserver = NEXT_TOKEN;
2705                 struct in_addr ina;
2706
2707                 if (inet_aton(nameserver, &ina)) {
2708                         /* address is valid */
2709                         evdns_nameserver_add(ina.s_addr);
2710                 }
2711         } else if (!strcmp(first_token, "domain") && (flags & DNS_OPTION_SEARCH)) {
2712                 const char *const domain = NEXT_TOKEN;
2713                 if (domain) {
2714                         search_postfix_clear();
2715                         search_postfix_add(domain);
2716                 }
2717         } else if (!strcmp(first_token, "search") && (flags & DNS_OPTION_SEARCH)) {
2718                 const char *domain;
2719                 search_postfix_clear();
2720
2721                 while ((domain = NEXT_TOKEN)) {
2722                         search_postfix_add(domain);
2723                 }
2724                 search_reverse();
2725         } else if (!strcmp(first_token, "options")) {
2726                 const char *option;
2727                 while ((option = NEXT_TOKEN)) {
2728                         const char *val = strchr(option, ':');
2729                         evdns_set_option(option, val ? val+1 : "", flags);
2730                 }
2731         }
2732 #undef NEXT_TOKEN
2733 }
2734
2735 /* exported function */
2736 /* returns: */
2737 /*   0 no errors */
2738 /*   1 failed to open file */
2739 /*   2 failed to stat file */
2740 /*   3 file too large */
2741 /*   4 out of memory */
2742 /*   5 short read from file */
2743 int
2744 evdns_resolv_conf_parse(int flags, const char *const filename) {
2745         struct stat st;
2746         int fd, n, r;
2747         u8 *resolv;
2748         char *start;
2749         int err = 0;
2750
2751         log(EVDNS_LOG_DEBUG, "Parsing resolv.conf file %s", filename);
2752
2753         fd = open(filename, O_RDONLY);
2754         if (fd < 0) {
2755                 evdns_resolv_set_defaults(flags);
2756                 return 1;
2757         }
2758
2759         if (fstat(fd, &st)) { err = 2; goto out1; }
2760         if (!st.st_size) {
2761                 evdns_resolv_set_defaults(flags);
2762                 err = (flags & DNS_OPTION_NAMESERVERS) ? 6 : 0;
2763                 goto out1;
2764         }
2765         if (st.st_size > 65535) { err = 3; goto out1; }  /* no resolv.conf should be any bigger */
2766
2767         resolv = (u8 *) malloc((size_t)st.st_size + 1);
2768         if (!resolv) { err = 4; goto out1; }
2769
2770         n = 0;
2771         while ((r = read(fd, resolv+n, (size_t)st.st_size-n)) > 0) {
2772                 n += r;
2773                 if (n == st.st_size)
2774                         break;
2775                 assert(n < st.st_size);
2776         }
2777         if (r < 0) { err = 5; goto out2; }
2778         resolv[n] = 0;   /* we malloced an extra byte; this should be fine. */
2779
2780         start = (char *) resolv;
2781         for (;;) {
2782                 char *const newline = strchr(start, '\n');
2783                 if (!newline) {
2784                         resolv_conf_parse_line(start, flags);
2785                         break;
2786                 } else {
2787                         *newline = 0;
2788                         resolv_conf_parse_line(start, flags);
2789                         start = newline + 1;
2790                 }
2791         }
2792
2793         if (!server_head && (flags & DNS_OPTION_NAMESERVERS)) {
2794                 /* no nameservers were configured. */
2795                 evdns_nameserver_ip_add("127.0.0.1");
2796                 err = 6;
2797         }
2798         if (flags & DNS_OPTION_SEARCH && (!global_search_state || global_search_state->num_domains == 0)) {
2799                 search_set_from_hostname();
2800         }
2801
2802 out2:
2803         free(resolv);
2804 out1:
2805         close(fd);
2806         return err;
2807 }
2808
2809 #ifdef WIN32
2810 /* Add multiple nameservers from a space-or-comma-separated list. */
2811 static int
2812 evdns_nameserver_ip_add_line(const char *ips) {
2813         const char *addr;
2814         char *buf;
2815         int r;
2816         while (*ips) {
2817                 while (ISSPACE(*ips) || *ips == ',' || *ips == '\t')
2818                         ++ips;
2819                 addr = ips;
2820                 while (ISDIGIT(*ips) || *ips == '.' || *ips == ':')
2821                         ++ips;
2822                 buf = malloc(ips-addr+1);
2823                 if (!buf) return 4;
2824                 memcpy(buf, addr, ips-addr);
2825                 buf[ips-addr] = '\0';
2826                 r = evdns_nameserver_ip_add(buf);
2827                 free(buf);
2828                 if (r) return r;
2829         }
2830         return 0;
2831 }
2832
2833 typedef DWORD(WINAPI *GetNetworkParams_fn_t)(FIXED_INFO *, DWORD*);
2834
2835 /* Use the windows GetNetworkParams interface in iphlpapi.dll to */
2836 /* figure out what our nameservers are. */
2837 static int
2838 load_nameservers_with_getnetworkparams(void)
2839 {
2840         /* Based on MSDN examples and inspection of  c-ares code. */
2841         FIXED_INFO *fixed;
2842         HMODULE handle = 0;
2843         ULONG size = sizeof(FIXED_INFO);
2844         void *buf = NULL;
2845         int status = 0, r, added_any;
2846         IP_ADDR_STRING *ns;
2847         GetNetworkParams_fn_t fn;
2848
2849         if (!(handle = LoadLibrary("iphlpapi.dll"))) {
2850                 log(EVDNS_LOG_WARN, "Could not open iphlpapi.dll");
2851                 status = -1;
2852                 goto done;
2853         }
2854         if (!(fn = (GetNetworkParams_fn_t) GetProcAddress(handle, "GetNetworkParams"))) {
2855                 log(EVDNS_LOG_WARN, "Could not get address of function.");
2856                 status = -1;
2857                 goto done;
2858         }
2859
2860         buf = malloc(size);
2861         if (!buf) { status = 4; goto done; }
2862         fixed = buf;
2863         r = fn(fixed, &size);
2864         if (r != ERROR_SUCCESS && r != ERROR_BUFFER_OVERFLOW) {
2865                 status = -1;
2866                 goto done;
2867         }
2868         if (r != ERROR_SUCCESS) {
2869                 free(buf);
2870                 buf = malloc(size);
2871                 if (!buf) { status = 4; goto done; }
2872                 fixed = buf;
2873                 r = fn(fixed, &size);
2874                 if (r != ERROR_SUCCESS) {
2875                         log(EVDNS_LOG_DEBUG, "fn() failed.");
2876                         status = -1;
2877                         goto done;
2878                 }
2879         }
2880
2881         assert(fixed);
2882         added_any = 0;
2883         ns = &(fixed->DnsServerList);
2884         while (ns) {
2885                 r = evdns_nameserver_ip_add_line(ns->IpAddress.String);
2886                 if (r) {
2887                         log(EVDNS_LOG_DEBUG,"Could not add nameserver %s to list,error: %d",
2888                                 (ns->IpAddress.String),(int)GetLastError());
2889                         status = r;
2890                         goto done;
2891                 } else {
2892                         log(EVDNS_LOG_DEBUG,"Succesfully added %s as nameserver",ns->IpAddress.String);
2893                 }
2894
2895                 added_any++;
2896                 ns = ns->Next;
2897         }
2898
2899         if (!added_any) {
2900                 log(EVDNS_LOG_DEBUG, "No nameservers added.");
2901                 status = -1;
2902         }
2903
2904  done:
2905         if (buf)
2906                 free(buf);
2907         if (handle)
2908                 FreeLibrary(handle);
2909         return status;
2910 }
2911
2912 static int
2913 config_nameserver_from_reg_key(HKEY key, const char *subkey)
2914 {
2915         char *buf;
2916         DWORD bufsz = 0, type = 0;
2917         int status = 0;
2918
2919         if (RegQueryValueEx(key, subkey, 0, &type, NULL, &bufsz)
2920             != ERROR_MORE_DATA)
2921                 return -1;
2922         if (!(buf = malloc(bufsz)))
2923                 return -1;
2924
2925         if (RegQueryValueEx(key, subkey, 0, &type, (LPBYTE)buf, &bufsz)
2926             == ERROR_SUCCESS && bufsz > 1) {
2927                 status = evdns_nameserver_ip_add_line(buf);
2928         }
2929
2930         free(buf);
2931         return status;
2932 }
2933
2934 #define SERVICES_KEY "System\\CurrentControlSet\\Services\\"
2935 #define WIN_NS_9X_KEY  SERVICES_KEY "VxD\\MSTCP"
2936 #define WIN_NS_NT_KEY  SERVICES_KEY "Tcpip\\Parameters"
2937
2938 static int
2939 load_nameservers_from_registry(void)
2940 {
2941         int found = 0;
2942         int r;
2943 #define TRY(k, name) \
2944         if (!found && config_nameserver_from_reg_key(k,name) == 0) {    \
2945                 log(EVDNS_LOG_DEBUG,"Found nameservers in %s/%s",#k,name); \
2946                 found = 1;                                              \
2947         } else if (!found) {                                            \
2948                 log(EVDNS_LOG_DEBUG,"Didn't find nameservers in %s/%s", \
2949                     #k,#name);                                          \
2950         }
2951
2952         if (((int)GetVersion()) > 0) { /* NT */
2953                 HKEY nt_key = 0, interfaces_key = 0;
2954
2955                 if (RegOpenKeyEx(HKEY_LOCAL_MACHINE, WIN_NS_NT_KEY, 0,
2956                                  KEY_READ, &nt_key) != ERROR_SUCCESS) {
2957                         log(EVDNS_LOG_DEBUG,"Couldn't open nt key, %d",(int)GetLastError());
2958                         return -1;
2959                 }
2960                 r = RegOpenKeyEx(nt_key, "Interfaces", 0,
2961                              KEY_QUERY_VALUE|KEY_ENUMERATE_SUB_KEYS,
2962                              &interfaces_key);
2963                 if (r != ERROR_SUCCESS) {
2964                         log(EVDNS_LOG_DEBUG,"Couldn't open interfaces key, %d",(int)GetLastError());
2965                         return -1;
2966                 }
2967                 TRY(nt_key, "NameServer");
2968                 TRY(nt_key, "DhcpNameServer");
2969                 TRY(interfaces_key, "NameServer");
2970                 TRY(interfaces_key, "DhcpNameServer");
2971                 RegCloseKey(interfaces_key);
2972                 RegCloseKey(nt_key);
2973         } else {
2974                 HKEY win_key = 0;
2975                 if (RegOpenKeyEx(HKEY_LOCAL_MACHINE, WIN_NS_9X_KEY, 0,
2976                                  KEY_READ, &win_key) != ERROR_SUCCESS) {
2977                         log(EVDNS_LOG_DEBUG, "Couldn't open registry key, %d", (int)GetLastError());
2978                         return -1;
2979                 }
2980                 TRY(win_key, "NameServer");
2981                 RegCloseKey(win_key);
2982         }
2983
2984         if (found == 0) {
2985                 log(EVDNS_LOG_WARN,"Didn't find any nameservers.");
2986         }
2987
2988         return found ? 0 : -1;
2989 #undef TRY
2990 }
2991
2992 int
2993 evdns_config_windows_nameservers(void)
2994 {
2995         if (load_nameservers_with_getnetworkparams() == 0)
2996                 return 0;
2997         return load_nameservers_from_registry();
2998 }
2999 #endif
3000
3001 int
3002 evdns_init(void)
3003 {
3004         int res = 0;
3005 #ifdef WIN32
3006         res = evdns_config_windows_nameservers();
3007 #else
3008         res = evdns_resolv_conf_parse(DNS_OPTIONS_ALL, "/etc/resolv.conf");
3009 #endif
3010
3011         return (res);
3012 }
3013
3014 const char *
3015 evdns_err_to_string(int err)
3016 {
3017     switch (err) {
3018         case DNS_ERR_NONE: return "no error";
3019         case DNS_ERR_FORMAT: return "misformatted query";
3020         case DNS_ERR_SERVERFAILED: return "server failed";
3021         case DNS_ERR_NOTEXIST: return "name does not exist";
3022         case DNS_ERR_NOTIMPL: return "query not implemented";
3023         case DNS_ERR_REFUSED: return "refused";
3024
3025         case DNS_ERR_TRUNCATED: return "reply truncated or ill-formed";
3026         case DNS_ERR_UNKNOWN: return "unknown";
3027         case DNS_ERR_TIMEOUT: return "request timed out";
3028         case DNS_ERR_SHUTDOWN: return "dns subsystem shut down";
3029         default: return "[Unknown error code]";
3030     }
3031 }
3032
3033 void
3034 evdns_shutdown(int fail_requests)
3035 {
3036         struct nameserver *server, *server_next;
3037         struct search_domain *dom, *dom_next;
3038
3039         while (req_head) {
3040                 if (fail_requests)
3041                         reply_callback(req_head, 0, DNS_ERR_SHUTDOWN, NULL);
3042                 request_finished(req_head, &req_head);
3043         }
3044         while (req_waiting_head) {
3045                 if (fail_requests)
3046                         reply_callback(req_waiting_head, 0, DNS_ERR_SHUTDOWN, NULL);
3047                 request_finished(req_waiting_head, &req_waiting_head);
3048         }
3049         global_requests_inflight = global_requests_waiting = 0;
3050
3051         for (server = server_head; server; server = server_next) {
3052                 server_next = server->next;
3053                 if (server->socket >= 0)
3054                         CLOSE_SOCKET(server->socket);
3055                 (void) event_del(&server->event);
3056                 if (server->state == 0)
3057                         (void) event_del(&server->timeout_event);
3058                 free(server);
3059                 if (server_next == server_head)
3060                         break;
3061         }
3062         server_head = NULL;
3063         global_good_nameservers = 0;
3064
3065         if (global_search_state) {
3066                 for (dom = global_search_state->head; dom; dom = dom_next) {
3067                         dom_next = dom->next;
3068                         free(dom);
3069                 }
3070                 free(global_search_state);
3071                 global_search_state = NULL;
3072         }
3073         evdns_log_fn = NULL;
3074 }
3075
3076 #ifdef EVDNS_MAIN
3077 void
3078 main_callback(int result, char type, int count, int ttl,
3079                           void *addrs, void *orig) {
3080         char *n = (char*)orig;
3081         int i;
3082         for (i = 0; i < count; ++i) {
3083                 if (type == DNS_IPv4_A) {
3084                         printf("%s: %s\n", n, debug_ntoa(((u32*)addrs)[i]));
3085                 } else if (type == DNS_PTR) {
3086                         printf("%s: %s\n", n, ((char**)addrs)[i]);
3087                 }
3088         }
3089         if (!count) {
3090                 printf("%s: No answer (%d)\n", n, result);
3091         }
3092         fflush(stdout);
3093 }
3094 void
3095 evdns_server_callback(struct evdns_server_request *req, void *data)
3096 {
3097         int i, r;
3098         (void)data;
3099         /* dummy; give 192.168.11.11 as an answer for all A questions,
3100          *      give foo.bar.example.com as an answer for all PTR questions. */
3101         for (i = 0; i < req->nquestions; ++i) {
3102                 u32 ans = htonl(0xc0a80b0bUL);
3103                 if (req->questions[i]->type == EVDNS_TYPE_A &&
3104                         req->questions[i]->dns_question_class == EVDNS_CLASS_INET) {
3105                         printf(" -- replying for %s (A)\n", req->questions[i]->name);
3106                         r = evdns_server_request_add_a_reply(req, req->questions[i]->name,
3107                                                                                   1, &ans, 10);
3108                         if (r<0)
3109                                 printf("eeep, didn't work.\n");
3110                 } else if (req->questions[i]->type == EVDNS_TYPE_PTR &&
3111                                    req->questions[i]->dns_question_class == EVDNS_CLASS_INET) {
3112                         printf(" -- replying for %s (PTR)\n", req->questions[i]->name);
3113                         r = evdns_server_request_add_ptr_reply(req, NULL, req->questions[i]->name,
3114                                                                                         "foo.bar.example.com", 10);
3115                 } else {
3116                         printf(" -- skipping %s [%d %d]\n", req->questions[i]->name,
3117                                    req->questions[i]->type, req->questions[i]->dns_question_class);
3118                 }
3119         }
3120
3121         r = evdns_request_respond(req, 0);
3122         if (r<0)
3123                 printf("eeek, couldn't send reply.\n");
3124 }
3125
3126 void
3127 logfn(int is_warn, const char *msg) {
3128   (void) is_warn;
3129   fprintf(stderr, "%s\n", msg);
3130 }
3131 int
3132 main(int c, char **v) {
3133         int idx;
3134         int reverse = 0, verbose = 1, servertest = 0;
3135         if (c<2) {
3136                 fprintf(stderr, "syntax: %s [-x] [-v] hostname\n", v[0]);
3137                 fprintf(stderr, "syntax: %s [-servertest]\n", v[0]);
3138                 return 1;
3139         }
3140         idx = 1;
3141         while (idx < c && v[idx][0] == '-') {
3142                 if (!strcmp(v[idx], "-x"))
3143                         reverse = 1;
3144                 else if (!strcmp(v[idx], "-v"))
3145                         verbose = 1;
3146                 else if (!strcmp(v[idx], "-servertest"))
3147                         servertest = 1;
3148                 else
3149                         fprintf(stderr, "Unknown option %s\n", v[idx]);
3150                 ++idx;
3151         }
3152         event_init();
3153         if (verbose)
3154                 evdns_set_log_fn(logfn);
3155         evdns_resolv_conf_parse(DNS_OPTION_NAMESERVERS, "/etc/resolv.conf");
3156         if (servertest) {
3157                 int sock;
3158                 struct sockaddr_in my_addr;
3159                 sock = socket(PF_INET, SOCK_DGRAM, 0);
3160                 evutil_make_socket_nonblocking(sock);
3161                 my_addr.sin_family = AF_INET;
3162                 my_addr.sin_port = htons(10053);
3163                 my_addr.sin_addr.s_addr = INADDR_ANY;
3164                 if (bind(sock, (struct sockaddr*)&my_addr, sizeof(my_addr))<0) {
3165                         perror("bind");
3166                         exit(1);
3167                 }
3168                 evdns_add_server_port(sock, 0, evdns_server_callback, NULL);
3169         }
3170         for (; idx < c; ++idx) {
3171                 if (reverse) {
3172                         struct in_addr addr;
3173                         if (!inet_aton(v[idx], &addr)) {
3174                                 fprintf(stderr, "Skipping non-IP %s\n", v[idx]);
3175                                 continue;
3176                         }
3177                         fprintf(stderr, "resolving %s...\n",v[idx]);
3178                         evdns_resolve_reverse(&addr, 0, main_callback, v[idx]);
3179                 } else {
3180                         fprintf(stderr, "resolving (fwd) %s...\n",v[idx]);
3181                         evdns_resolve_ipv4(v[idx], 0, main_callback, v[idx]);
3182                 }
3183         }
3184         fflush(stdout);
3185         event_dispatch();
3186         return 0;
3187 }
3188 #endif