contrib/bind9/lib/dns/resolver.c

   1 /*
   2  * Copyright (C) 2004-2014  Internet Systems Consortium, Inc. ("ISC")
   3  * Copyright (C) 1999-2003  Internet Software Consortium.
   4  *
   5  * Permission to use, copy, modify, and/or distribute this software for any
   6  * purpose with or without fee is hereby granted, provided that the above
   7  * copyright notice and this permission notice appear in all copies.
   8  *
   9  * THE SOFTWARE IS PROVIDED "AS IS" AND ISC DISCLAIMS ALL WARRANTIES WITH
  10  * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
  11  * AND FITNESS.  IN NO EVENT SHALL ISC BE LIABLE FOR ANY SPECIAL, DIRECT,
  12  * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
  13  * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE
  14  * OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
  15  * PERFORMANCE OF THIS SOFTWARE.
  16  */
  17
  18 /* $Id$ */
  19
  20 /*! \file */
  21
  22 #include <config.h>
  23
  24 #include <isc/log.h>
  25 #include <isc/platform.h>
  26 #include <isc/print.h>
  27 #include <isc/string.h>
  28 #include <isc/random.h>
  29 #include <isc/socket.h>
  30 #include <isc/stats.h>
  31 #include <isc/task.h>
  32 #include <isc/timer.h>
  33 #include <isc/util.h>
  34
  35 #include <dns/acl.h>
  36 #include <dns/adb.h>
  37 #include <dns/cache.h>
  38 #include <dns/db.h>
  39 #include <dns/dispatch.h>
  40 #include <dns/ds.h>
  41 #include <dns/events.h>
  42 #include <dns/forward.h>
  43 #include <dns/keytable.h>
  44 #include <dns/log.h>
  45 #include <dns/message.h>
  46 #include <dns/ncache.h>
  47 #include <dns/nsec.h>
  48 #include <dns/nsec3.h>
  49 #include <dns/opcode.h>
  50 #include <dns/peer.h>
  51 #include <dns/rbt.h>
  52 #include <dns/rcode.h>
  53 #include <dns/rdata.h>
  54 #include <dns/rdataclass.h>
  55 #include <dns/rdatalist.h>
  56 #include <dns/rdataset.h>
  57 #include <dns/rdatastruct.h>
  58 #include <dns/rdatatype.h>
  59 #include <dns/resolver.h>
  60 #include <dns/result.h>
  61 #include <dns/rootns.h>
  62 #include <dns/stats.h>
  63 #include <dns/tsig.h>
  64 #include <dns/validator.h>
  65
  66 #define DNS_RESOLVER_TRACE
  67 #ifdef DNS_RESOLVER_TRACE
  68 #define RTRACE(m)       isc_log_write(dns_lctx, \
  69                                       DNS_LOGCATEGORY_RESOLVER, \
  70                                       DNS_LOGMODULE_RESOLVER, \
  71                                       ISC_LOG_DEBUG(3), \
  72                                       "res %p: %s", res, (m))
  73 #define RRTRACE(r, m)   isc_log_write(dns_lctx, \
  74                                       DNS_LOGCATEGORY_RESOLVER, \
  75                                       DNS_LOGMODULE_RESOLVER, \
  76                                       ISC_LOG_DEBUG(3), \
  77                                       "res %p: %s", (r), (m))
  78 #define FCTXTRACE(m)    isc_log_write(dns_lctx, \
  79                                       DNS_LOGCATEGORY_RESOLVER, \
  80                                       DNS_LOGMODULE_RESOLVER, \
  81                                       ISC_LOG_DEBUG(3), \
  82                                       "fctx %p(%s): %s", fctx, fctx->info, (m))
  83 #define FCTXTRACE2(m1, m2) \
  84                         isc_log_write(dns_lctx, \
  85                                       DNS_LOGCATEGORY_RESOLVER, \
  86                                       DNS_LOGMODULE_RESOLVER, \
  87                                       ISC_LOG_DEBUG(3), \
  88                                       "fctx %p(%s): %s %s", \
  89                                       fctx, fctx->info, (m1), (m2))
  90 #define FTRACE(m)       isc_log_write(dns_lctx, \
  91                                       DNS_LOGCATEGORY_RESOLVER, \
  92                                       DNS_LOGMODULE_RESOLVER, \
  93                                       ISC_LOG_DEBUG(3), \
  94                                       "fetch %p (fctx %p(%s)): %s", \
  95                                       fetch, fetch->private, \
  96                                       fetch->private->info, (m))
  97 #define QTRACE(m)       isc_log_write(dns_lctx, \
  98                                       DNS_LOGCATEGORY_RESOLVER, \
  99                                       DNS_LOGMODULE_RESOLVER, \
 100                                       ISC_LOG_DEBUG(3), \
 101                                       "resquery %p (fctx %p(%s)): %s", \
 102                                       query, query->fctx, \
 103                                       query->fctx->info, (m))
 104 #else
 105 #define RTRACE(m)
 106 #define RRTRACE(r, m)
 107 #define FCTXTRACE(m)
 108 #define FCTXTRACE2(m1, m2)
 109 #define FTRACE(m)
 110 #define QTRACE(m)
 111 #endif
 112
 113 #define US_PER_SEC 1000000U
 114 /*
 115  * The maximum time we will wait for a single query.
 116  */
 117 #define MAX_SINGLE_QUERY_TIMEOUT 9U
 118 #define MAX_SINGLE_QUERY_TIMEOUT_US (MAX_SINGLE_QUERY_TIMEOUT*US_PER_SEC)
 119
 120 /*
 121  * We need to allow a individual query time to complete / timeout.
 122  */
 123 #define MINIMUM_QUERY_TIMEOUT (MAX_SINGLE_QUERY_TIMEOUT + 1U)
 124
 125 /* The default time in seconds for the whole query to live. */
 126 #ifndef DEFAULT_QUERY_TIMEOUT
 127 #define DEFAULT_QUERY_TIMEOUT MINIMUM_QUERY_TIMEOUT
 128 #endif
 129
 130 #ifndef MAXIMUM_QUERY_TIMEOUT
 131 #define MAXIMUM_QUERY_TIMEOUT 30 /* The maximum time in seconds for the whole query to live. */
 132 #endif
 133
 134 /* The default maximum number of recursions to follow before giving up. */
 135 #ifndef DEFAULT_RECURSION_DEPTH
 136 #define DEFAULT_RECURSION_DEPTH 7
 137 #endif
 138
 139 /* The default maximum number of iterative queries to allow before giving up. */
 140 #ifndef DEFAULT_MAX_QUERIES
 141 #define DEFAULT_MAX_QUERIES 50
 142 #endif
 143
 144 /*%
 145  * Maximum EDNS0 input packet size.
 146  */
 147 #define RECV_BUFFER_SIZE                4096            /* XXXRTH  Constant. */
 148 #define EDNSOPTS                        2
 149
 150 /*%
 151  * This defines the maximum number of timeouts we will permit before we
 152  * disable EDNS0 on the query.
 153  */
 154 #define MAX_EDNS0_TIMEOUTS      3
 155
 156 typedef struct fetchctx fetchctx_t;
 157
 158 typedef struct query {
 159         /* Locked by task event serialization. */
 160         unsigned int                    magic;
 161         fetchctx_t *                    fctx;
 162         isc_mem_t *                     mctx;
 163         dns_dispatchmgr_t *             dispatchmgr;
 164         dns_dispatch_t *                dispatch;
 165         isc_boolean_t                   exclusivesocket;
 166         dns_adbaddrinfo_t *             addrinfo;
 167         isc_socket_t *                  tcpsocket;
 168         isc_time_t                      start;
 169         dns_messageid_t                 id;
 170         dns_dispentry_t *               dispentry;
 171         ISC_LINK(struct query)          link;
 172         isc_buffer_t                    buffer;
 173         isc_buffer_t                    *tsig;
 174         dns_tsigkey_t                   *tsigkey;
 175         isc_socketevent_t               sendevent;
 176         unsigned int                    options;
 177         unsigned int                    attributes;
 178         unsigned int                    sends;
 179         unsigned int                    connects;
 180         unsigned char                   data[512];
 181 } resquery_t;
 182
 183 #define QUERY_MAGIC                     ISC_MAGIC('Q', '!', '!', '!')
 184 #define VALID_QUERY(query)              ISC_MAGIC_VALID(query, QUERY_MAGIC)
 185
 186 #define RESQUERY_ATTR_CANCELED          0x02
 187
 188 #define RESQUERY_CONNECTING(q)          ((q)->connects > 0)
 189 #define RESQUERY_CANCELED(q)            (((q)->attributes & \
 190                                           RESQUERY_ATTR_CANCELED) != 0)
 191 #define RESQUERY_SENDING(q)             ((q)->sends > 0)
 192
 193 typedef enum {
 194         fetchstate_init = 0,            /*%< Start event has not run yet. */
 195         fetchstate_active,
 196         fetchstate_done                 /*%< FETCHDONE events posted. */
 197 } fetchstate;
 198
 199 typedef enum {
 200         badns_unreachable = 0,
 201         badns_response,
 202         badns_validation
 203 } badnstype_t;
 204
 205 struct fetchctx {
 206         /*% Not locked. */
 207         unsigned int                    magic;
 208         dns_resolver_t *                res;
 209         dns_name_t                      name;
 210         dns_rdatatype_t                 type;
 211         unsigned int                    options;
 212         unsigned int                    bucketnum;
 213         char *                          info;
 214         isc_mem_t *                     mctx;
 215
 216         /*% Locked by appropriate bucket lock. */
 217         fetchstate                      state;
 218         isc_boolean_t                   want_shutdown;
 219         isc_boolean_t                   cloned;
 220         isc_boolean_t                   spilled;
 221         unsigned int                    references;
 222         isc_event_t                     control_event;
 223         ISC_LINK(struct fetchctx)       link;
 224         ISC_LIST(dns_fetchevent_t)      events;
 225         /*% Locked by task event serialization. */
 226         dns_name_t                      domain;
 227         dns_rdataset_t                  nameservers;
 228         unsigned int                    attributes;
 229         isc_timer_t *                   timer;
 230         isc_time_t                      expires;
 231         isc_interval_t                  interval;
 232         dns_message_t *                 qmessage;
 233         dns_message_t *                 rmessage;
 234         ISC_LIST(resquery_t)            queries;
 235         dns_adbfindlist_t               finds;
 236         dns_adbfind_t *                 find;
 237         dns_adbfindlist_t               altfinds;
 238         dns_adbfind_t *                 altfind;
 239         dns_adbaddrinfolist_t           forwaddrs;
 240         dns_adbaddrinfolist_t           altaddrs;
 241         isc_sockaddrlist_t              forwarders;
 242         dns_fwdpolicy_t                 fwdpolicy;
 243         isc_sockaddrlist_t              bad;
 244         isc_sockaddrlist_t              edns;
 245         isc_sockaddrlist_t              edns512;
 246         isc_sockaddrlist_t              bad_edns;
 247         dns_validator_t                 *validator;
 248         ISC_LIST(dns_validator_t)       validators;
 249         dns_db_t *                      cache;
 250         dns_adb_t *                     adb;
 251         isc_boolean_t                   ns_ttl_ok;
 252         isc_uint32_t                    ns_ttl;
 253
 254         /*%
 255          * The number of events we're waiting for.
 256          */
 257         unsigned int                    pending;
 258
 259         /*%
 260          * The number of times we've "restarted" the current
 261          * nameserver set.  This acts as a failsafe to prevent
 262          * us from pounding constantly on a particular set of
 263          * servers that, for whatever reason, are not giving
 264          * us useful responses, but are responding in such a
 265          * way that they are not marked "bad".
 266          */
 267         unsigned int                    restarts;
 268
 269         /*%
 270          * The number of timeouts that have occurred since we
 271          * last successfully received a response packet.  This
 272          * is used for EDNS0 black hole detection.
 273          */
 274         unsigned int                    timeouts;
 275
 276         /*%
 277          * Look aside state for DS lookups.
 278          */
 279         dns_name_t                      nsname;
 280         dns_fetch_t *                   nsfetch;
 281         dns_rdataset_t                  nsrrset;
 282
 283         /*%
 284          * Number of queries that reference this context.
 285          */
 286         unsigned int                    nqueries;
 287
 288         /*%
 289          * The reason to print when logging a successful
 290          * response to a query.
 291          */
 292         const char *                    reason;
 293
 294         /*%
 295          * Random numbers to use for mixing up server addresses.
 296          */
 297         isc_uint32_t                    rand_buf;
 298         isc_uint32_t                    rand_bits;
 299
 300         /*%
 301          * Fetch-local statistics for detailed logging.
 302          */
 303         isc_result_t                    result; /*%< fetch result  */
 304         isc_result_t                    vresult; /*%< validation result  */
 305         int                             exitline;
 306         isc_time_t                      start;
 307         isc_uint64_t                    duration;
 308         isc_boolean_t                   logged;
 309         unsigned int                    querysent;
 310         unsigned int                    totalqueries;
 311         unsigned int                    referrals;
 312         unsigned int                    lamecount;
 313         unsigned int                    neterr;
 314         unsigned int                    badresp;
 315         unsigned int                    adberr;
 316         unsigned int                    findfail;
 317         unsigned int                    valfail;
 318         isc_boolean_t                   timeout;
 319         dns_adbaddrinfo_t               *addrinfo;
 320         isc_sockaddr_t                  *client;
 321         unsigned int                    depth;
 322 };
 323
 324 #define FCTX_MAGIC                      ISC_MAGIC('F', '!', '!', '!')
 325 #define VALID_FCTX(fctx)                ISC_MAGIC_VALID(fctx, FCTX_MAGIC)
 326
 327 #define FCTX_ATTR_HAVEANSWER            0x0001
 328 #define FCTX_ATTR_GLUING                0x0002
 329 #define FCTX_ATTR_ADDRWAIT              0x0004
 330 #define FCTX_ATTR_SHUTTINGDOWN          0x0008
 331 #define FCTX_ATTR_WANTCACHE             0x0010
 332 #define FCTX_ATTR_WANTNCACHE            0x0020
 333 #define FCTX_ATTR_NEEDEDNS0             0x0040
 334 #define FCTX_ATTR_TRIEDFIND             0x0080
 335 #define FCTX_ATTR_TRIEDALT              0x0100
 336
 337 #define HAVE_ANSWER(f)          (((f)->attributes & FCTX_ATTR_HAVEANSWER) != \
 338                                  0)
 339 #define GLUING(f)               (((f)->attributes & FCTX_ATTR_GLUING) != \
 340                                  0)
 341 #define ADDRWAIT(f)             (((f)->attributes & FCTX_ATTR_ADDRWAIT) != \
 342                                  0)
 343 #define SHUTTINGDOWN(f)         (((f)->attributes & FCTX_ATTR_SHUTTINGDOWN) \
 344                                  != 0)
 345 #define WANTCACHE(f)            (((f)->attributes & FCTX_ATTR_WANTCACHE) != 0)
 346 #define WANTNCACHE(f)           (((f)->attributes & FCTX_ATTR_WANTNCACHE) != 0)
 347 #define NEEDEDNS0(f)            (((f)->attributes & FCTX_ATTR_NEEDEDNS0) != 0)
 348 #define TRIEDFIND(f)            (((f)->attributes & FCTX_ATTR_TRIEDFIND) != 0)
 349 #define TRIEDALT(f)             (((f)->attributes & FCTX_ATTR_TRIEDALT) != 0)
 350
 351 typedef struct {
 352         dns_adbaddrinfo_t *             addrinfo;
 353         fetchctx_t *                    fctx;
 354 } dns_valarg_t;
 355
 356 struct dns_fetch {
 357         unsigned int                    magic;
 358         fetchctx_t *                    private;
 359 };
 360
 361 #define DNS_FETCH_MAGIC                 ISC_MAGIC('F', 't', 'c', 'h')
 362 #define DNS_FETCH_VALID(fetch)          ISC_MAGIC_VALID(fetch, DNS_FETCH_MAGIC)
 363
 364 typedef struct fctxbucket {
 365         isc_task_t *                    task;
 366         isc_mutex_t                     lock;
 367         ISC_LIST(fetchctx_t)            fctxs;
 368         isc_boolean_t                   exiting;
 369         isc_mem_t *                     mctx;
 370 } fctxbucket_t;
 371
 372 typedef struct alternate {
 373         isc_boolean_t                   isaddress;
 374         union   {
 375                 isc_sockaddr_t          addr;
 376                 struct {
 377                         dns_name_t      name;
 378                         in_port_t       port;
 379                 } _n;
 380         } _u;
 381         ISC_LINK(struct alternate)      link;
 382 } alternate_t;
 383
 384 typedef struct dns_badcache dns_badcache_t;
 385 struct dns_badcache {
 386         dns_badcache_t *        next;
 387         dns_rdatatype_t         type;
 388         isc_time_t              expire;
 389         unsigned int            hashval;
 390         dns_name_t              name;
 391 };
 392 #define DNS_BADCACHE_SIZE 1021
 393 #define DNS_BADCACHE_TTL(fctx) \
 394         (((fctx)->res->lame_ttl > 30 ) ? (fctx)->res->lame_ttl : 30)
 395
 396 struct dns_resolver {
 397         /* Unlocked. */
 398         unsigned int                    magic;
 399         isc_mem_t *                     mctx;
 400         isc_mutex_t                     lock;
 401         isc_mutex_t                     nlock;
 402         isc_mutex_t                     primelock;
 403         dns_rdataclass_t                rdclass;
 404         isc_socketmgr_t *               socketmgr;
 405         isc_timermgr_t *                timermgr;
 406         isc_taskmgr_t *                 taskmgr;
 407         dns_view_t *                    view;
 408         isc_boolean_t                   frozen;
 409         unsigned int                    options;
 410         dns_dispatchmgr_t *             dispatchmgr;
 411         dns_dispatchset_t *             dispatches4;
 412         isc_boolean_t                   exclusivev4;
 413         dns_dispatchset_t *             dispatches6;
 414         isc_boolean_t                   exclusivev6;
 415         unsigned int                    nbuckets;
 416         fctxbucket_t *                  buckets;
 417         isc_uint32_t                    lame_ttl;
 418         ISC_LIST(alternate_t)           alternates;
 419         isc_uint16_t                    udpsize;
 420 #if USE_ALGLOCK
 421         isc_rwlock_t                    alglock;
 422 #endif
 423         dns_rbt_t *                     algorithms;
 424 #if USE_MBSLOCK
 425         isc_rwlock_t                    mbslock;
 426 #endif
 427         dns_rbt_t *                     mustbesecure;
 428         unsigned int                    spillatmax;
 429         unsigned int                    spillatmin;
 430         isc_timer_t *                   spillattimer;
 431         isc_boolean_t                   zero_no_soa_ttl;
 432         unsigned int                    query_timeout;
 433         unsigned int                    maxdepth;
 434
 435         /* Locked by lock. */
 436         unsigned int                    references;
 437         isc_boolean_t                   exiting;
 438         isc_eventlist_t                 whenshutdown;
 439         unsigned int                    activebuckets;
 440         isc_boolean_t                   priming;
 441         unsigned int                    spillat;        /* clients-per-query */
 442
 443         /* Bad cache. */
 444         dns_badcache_t  **              badcache;
 445         unsigned int                    badcount;
 446         unsigned int                    badhash;
 447         unsigned int                    badsweep;
 448
 449         /* Locked by primelock. */
 450         dns_fetch_t *                   primefetch;
 451         /* Locked by nlock. */
 452         unsigned int                    nfctx;
 453 };
 454
 455 #define RES_MAGIC                       ISC_MAGIC('R', 'e', 's', '!')
 456 #define VALID_RESOLVER(res)             ISC_MAGIC_VALID(res, RES_MAGIC)
 457
 458 /*%
 459  * Private addrinfo flags.  These must not conflict with DNS_FETCHOPT_NOEDNS0,
 460  * which we also use as an addrinfo flag.
 461  */
 462 #define FCTX_ADDRINFO_MARK              0x0001
 463 #define FCTX_ADDRINFO_FORWARDER         0x1000
 464 #define FCTX_ADDRINFO_TRIED             0x2000
 465 #define UNMARKED(a)                     (((a)->flags & FCTX_ADDRINFO_MARK) \
 466                                          == 0)
 467 #define ISFORWARDER(a)                  (((a)->flags & \
 468                                          FCTX_ADDRINFO_FORWARDER) != 0)
 469 #define TRIED(a)                        (((a)->flags & \
 470                                          FCTX_ADDRINFO_TRIED) != 0)
 471
 472 #define NXDOMAIN(r) (((r)->attributes & DNS_RDATASETATTR_NXDOMAIN) != 0)
 473 #define NEGATIVE(r) (((r)->attributes & DNS_RDATASETATTR_NEGATIVE) != 0)
 474
 475 static void destroy(dns_resolver_t *res);
 476 static void empty_bucket(dns_resolver_t *res);
 477 static isc_result_t resquery_send(resquery_t *query);
 478 static void resquery_response(isc_task_t *task, isc_event_t *event);
 479 static void resquery_connected(isc_task_t *task, isc_event_t *event);
 480 static void fctx_try(fetchctx_t *fctx, isc_boolean_t retrying,
 481                      isc_boolean_t badcache);
 482 static void fctx_destroy(fetchctx_t *fctx);
 483 static isc_boolean_t fctx_unlink(fetchctx_t *fctx);
 484 static isc_result_t ncache_adderesult(dns_message_t *message,
 485                                       dns_db_t *cache, dns_dbnode_t *node,
 486                                       dns_rdatatype_t covers,
 487                                       isc_stdtime_t now, dns_ttl_t maxttl,
 488                                       isc_boolean_t optout,
 489                                       isc_boolean_t secure,
 490                                       dns_rdataset_t *ardataset,
 491                                       isc_result_t *eresultp);
 492 static void validated(isc_task_t *task, isc_event_t *event);
 493 static isc_boolean_t maybe_destroy(fetchctx_t *fctx, isc_boolean_t locked);
 494 static void add_bad(fetchctx_t *fctx, dns_adbaddrinfo_t *addrinfo,
 495                     isc_result_t reason, badnstype_t badtype);
 496 static inline isc_result_t findnoqname(fetchctx_t *fctx, dns_name_t *name,
 497                                        dns_rdatatype_t type,
 498                                        dns_name_t **noqname);
 499
 500 /*%
 501  * Increment resolver-related statistics counters.
 502  */
 503 static inline void
 504 inc_stats(dns_resolver_t *res, isc_statscounter_t counter) {
 505         if (res->view->resstats != NULL)
 506                 isc_stats_increment(res->view->resstats, counter);
 507 }
 508
 509 static isc_result_t
 510 valcreate(fetchctx_t *fctx, dns_adbaddrinfo_t *addrinfo, dns_name_t *name,
 511           dns_rdatatype_t type, dns_rdataset_t *rdataset,
 512           dns_rdataset_t *sigrdataset, unsigned int valoptions,
 513           isc_task_t *task)
 514 {
 515         dns_validator_t *validator = NULL;
 516         dns_valarg_t *valarg;
 517         isc_result_t result;
 518
 519         valarg = isc_mem_get(fctx->mctx, sizeof(*valarg));
 520         if (valarg == NULL)
 521                 return (ISC_R_NOMEMORY);
 522
 523         valarg->fctx = fctx;
 524         valarg->addrinfo = addrinfo;
 525
 526         if (!ISC_LIST_EMPTY(fctx->validators))
 527                 INSIST((valoptions & DNS_VALIDATOR_DEFER) != 0);
 528
 529         result = dns_validator_create(fctx->res->view, name, type, rdataset,
 530                                       sigrdataset, fctx->rmessage,
 531                                       valoptions, task, validated, valarg,
 532                                       &validator);
 533         if (result == ISC_R_SUCCESS) {
 534                 inc_stats(fctx->res, dns_resstatscounter_val);
 535                 if ((valoptions & DNS_VALIDATOR_DEFER) == 0) {
 536                         INSIST(fctx->validator == NULL);
 537                         fctx->validator = validator;
 538                 }
 539                 ISC_LIST_APPEND(fctx->validators, validator, link);
 540         } else
 541                 isc_mem_put(fctx->mctx, valarg, sizeof(*valarg));
 542         return (result);
 543 }
 544
 545 static isc_boolean_t
 546 rrsig_fromchildzone(fetchctx_t *fctx, dns_rdataset_t *rdataset) {
 547         dns_namereln_t namereln;
 548         dns_rdata_rrsig_t rrsig;
 549         dns_rdata_t rdata = DNS_RDATA_INIT;
 550         int order;
 551         isc_result_t result;
 552         unsigned int labels;
 553
 554         for (result = dns_rdataset_first(rdataset);
 555              result == ISC_R_SUCCESS;
 556              result = dns_rdataset_next(rdataset)) {
 557                 dns_rdataset_current(rdataset, &rdata);
 558                 result = dns_rdata_tostruct(&rdata, &rrsig, NULL);
 559                 RUNTIME_CHECK(result == ISC_R_SUCCESS);
 560                 namereln = dns_name_fullcompare(&rrsig.signer, &fctx->domain,
 561                                                 &order, &labels);
 562                 if (namereln == dns_namereln_subdomain)
 563                         return (ISC_TRUE);
 564                 dns_rdata_reset(&rdata);
 565         }
 566         return (ISC_FALSE);
 567 }
 568
 569 static isc_boolean_t
 570 fix_mustbedelegationornxdomain(dns_message_t *message, fetchctx_t *fctx) {
 571         dns_name_t *name;
 572         dns_name_t *domain = &fctx->domain;
 573         dns_rdataset_t *rdataset;
 574         dns_rdatatype_t type;
 575         isc_result_t result;
 576         isc_boolean_t keep_auth = ISC_FALSE;
 577
 578         if (message->rcode == dns_rcode_nxdomain)
 579                 return (ISC_FALSE);
 580
 581         /*
 582          * A DS RRset can appear anywhere in a zone, even for a delegation-only
 583          * zone.  So a response to an explicit query for this type should be
 584          * excluded from delegation-only fixup.
 585          *
 586          * SOA, NS, and DNSKEY can only exist at a zone apex, so a postive
 587          * response to a query for these types can never violate the
 588          * delegation-only assumption: if the query name is below a
 589          * zone cut, the response should normally be a referral, which should
 590          * be accepted; if the query name is below a zone cut but the server
 591          * happens to have authority for the zone of the query name, the
 592          * response is a (non-referral) answer.  But this does not violate
 593          * delegation-only because the query name must be in a different zone
 594          * due to the "apex-only" nature of these types.  Note that if the
 595          * remote server happens to have authority for a child zone of a
 596          * delegation-only zone, we may still incorrectly "fix" the response
 597          * with NXDOMAIN for queries for other types.  Unfortunately it's
 598          * generally impossible to differentiate this case from violation of
 599          * the delegation-only assumption.  Once the resolver learns the
 600          * correct zone cut, possibly via a separate query for an "apex-only"
 601          * type, queries for other types will be resolved correctly.
 602          *
 603          * A query for type ANY will be accepted if it hits an exceptional
 604          * type above in the answer section as it should be from a child
 605          * zone.
 606          *
 607          * Also accept answers with RRSIG records from the child zone.
 608          * Direct queries for RRSIG records should not be answered from
 609          * the parent zone.
 610          */
 611
 612         if (message->counts[DNS_SECTION_ANSWER] != 0 &&
 613             (fctx->type == dns_rdatatype_ns ||
 614              fctx->type == dns_rdatatype_ds ||
 615              fctx->type == dns_rdatatype_soa ||
 616              fctx->type == dns_rdatatype_any ||
 617              fctx->type == dns_rdatatype_rrsig ||
 618              fctx->type == dns_rdatatype_dnskey)) {
 619                 result = dns_message_firstname(message, DNS_SECTION_ANSWER);
 620                 while (result == ISC_R_SUCCESS) {
 621                         name = NULL;
 622                         dns_message_currentname(message, DNS_SECTION_ANSWER,
 623                                                 &name);
 624                         for (rdataset = ISC_LIST_HEAD(name->list);
 625                              rdataset != NULL;
 626                              rdataset = ISC_LIST_NEXT(rdataset, link)) {
 627                                 if (!dns_name_equal(name, &fctx->name))
 628                                         continue;
 629                                 type = rdataset->type;
 630                                 /*
 631                                  * RRsig from child?
 632                                  */
 633                                 if (type == dns_rdatatype_rrsig &&
 634                                     rrsig_fromchildzone(fctx, rdataset))
 635                                         return (ISC_FALSE);
 636                                 /*
 637                                  * Direct query for apex records or DS.
 638                                  */
 639                                 if (fctx->type == type &&
 640                                     (type == dns_rdatatype_ds ||
 641                                      type == dns_rdatatype_ns ||
 642                                      type == dns_rdatatype_soa ||
 643                                      type == dns_rdatatype_dnskey))
 644                                         return (ISC_FALSE);
 645                                 /*
 646                                  * Indirect query for apex records or DS.
 647                                  */
 648                                 if (fctx->type == dns_rdatatype_any &&
 649                                     (type == dns_rdatatype_ns ||
 650                                      type == dns_rdatatype_ds ||
 651                                      type == dns_rdatatype_soa ||
 652                                      type == dns_rdatatype_dnskey))
 653                                         return (ISC_FALSE);
 654                         }
 655                         result = dns_message_nextname(message,
 656                                                       DNS_SECTION_ANSWER);
 657                 }
 658         }
 659
 660         /*
 661          * A NODATA response to a DS query?
 662          */
 663         if (fctx->type == dns_rdatatype_ds &&
 664             message->counts[DNS_SECTION_ANSWER] == 0)
 665                 return (ISC_FALSE);
 666
 667         /* Look for referral or indication of answer from child zone? */
 668         if (message->counts[DNS_SECTION_AUTHORITY] == 0)
 669                 goto munge;
 670
 671         result = dns_message_firstname(message, DNS_SECTION_AUTHORITY);
 672         while (result == ISC_R_SUCCESS) {
 673                 name = NULL;
 674                 dns_message_currentname(message, DNS_SECTION_AUTHORITY, &name);
 675                 for (rdataset = ISC_LIST_HEAD(name->list);
 676                      rdataset != NULL;
 677                      rdataset = ISC_LIST_NEXT(rdataset, link)) {
 678                         type = rdataset->type;
 679                         if (type == dns_rdatatype_soa &&
 680                             dns_name_equal(name, domain))
 681                                 keep_auth = ISC_TRUE;
 682
 683                         if (type != dns_rdatatype_ns &&
 684                             type != dns_rdatatype_soa &&
 685                             type != dns_rdatatype_rrsig)
 686                                 continue;
 687
 688                         if (type == dns_rdatatype_rrsig) {
 689                                 if (rrsig_fromchildzone(fctx, rdataset))
 690                                         return (ISC_FALSE);
 691                                 else
 692                                         continue;
 693                         }
 694
 695                         /* NS or SOA records. */
 696                         if (dns_name_equal(name, domain)) {
 697                                 /*
 698                                  * If a query for ANY causes a negative
 699                                  * response, we can be sure that this is
 700                                  * an empty node.  For other type of queries
 701                                  * we cannot differentiate an empty node
 702                                  * from a node that just doesn't have that
 703                                  * type of record.  We only accept the former
 704                                  * case.
 705                                  */
 706                                 if (message->counts[DNS_SECTION_ANSWER] == 0 &&
 707                                     fctx->type == dns_rdatatype_any)
 708                                         return (ISC_FALSE);
 709                         } else if (dns_name_issubdomain(name, domain)) {
 710                                 /* Referral or answer from child zone. */
 711                                 return (ISC_FALSE);
 712                         }
 713                 }
 714                 result = dns_message_nextname(message, DNS_SECTION_AUTHORITY);
 715         }
 716
 717  munge:
 718         message->rcode = dns_rcode_nxdomain;
 719         message->counts[DNS_SECTION_ANSWER] = 0;
 720         if (!keep_auth)
 721                 message->counts[DNS_SECTION_AUTHORITY] = 0;
 722         message->counts[DNS_SECTION_ADDITIONAL] = 0;
 723         return (ISC_TRUE);
 724 }
 725
 726 static inline isc_result_t
 727 fctx_starttimer(fetchctx_t *fctx) {
 728         /*
 729          * Start the lifetime timer for fctx.
 730          *
 731          * This is also used for stopping the idle timer; in that
 732          * case we must purge events already posted to ensure that
 733          * no further idle events are delivered.
 734          */
 735         return (isc_timer_reset(fctx->timer, isc_timertype_once,
 736                                 &fctx->expires, NULL, ISC_TRUE));
 737 }
 738
 739 static inline void
 740 fctx_stoptimer(fetchctx_t *fctx) {
 741         isc_result_t result;
 742
 743         /*
 744          * We don't return a result if resetting the timer to inactive fails
 745          * since there's nothing to be done about it.  Resetting to inactive
 746          * should never fail anyway, since the code as currently written
 747          * cannot fail in that case.
 748          */
 749         result = isc_timer_reset(fctx->timer, isc_timertype_inactive,
 750                                   NULL, NULL, ISC_TRUE);
 751         if (result != ISC_R_SUCCESS) {
 752                 UNEXPECTED_ERROR(__FILE__, __LINE__,
 753                                  "isc_timer_reset(): %s",
 754                                  isc_result_totext(result));
 755         }
 756 }
 757
 758
 759 static inline isc_result_t
 760 fctx_startidletimer(fetchctx_t *fctx, isc_interval_t *interval) {
 761         /*
 762          * Start the idle timer for fctx.  The lifetime timer continues
 763          * to be in effect.
 764          */
 765         return (isc_timer_reset(fctx->timer, isc_timertype_once,
 766                                 &fctx->expires, interval, ISC_FALSE));
 767 }
 768
 769 /*
 770  * Stopping the idle timer is equivalent to calling fctx_starttimer(), but
 771  * we use fctx_stopidletimer for readability in the code below.
 772  */
 773 #define fctx_stopidletimer      fctx_starttimer
 774
 775
 776 static inline void
 777 resquery_destroy(resquery_t **queryp) {
 778         resquery_t *query;
 779
 780         REQUIRE(queryp != NULL);
 781         query = *queryp;
 782         REQUIRE(!ISC_LINK_LINKED(query, link));
 783
 784         INSIST(query->tcpsocket == NULL);
 785
 786         query->fctx->nqueries--;
 787         if (SHUTTINGDOWN(query->fctx)) {
 788                 dns_resolver_t *res = query->fctx->res;
 789                 if (maybe_destroy(query->fctx, ISC_FALSE))
 790                         empty_bucket(res);
 791         }
 792         query->magic = 0;
 793         isc_mem_put(query->mctx, query, sizeof(*query));
 794         *queryp = NULL;
 795 }
 796
 797 static void
 798 fctx_cancelquery(resquery_t **queryp, dns_dispatchevent_t **deventp,
 799                  isc_time_t *finish, isc_boolean_t no_response)
 800 {
 801         fetchctx_t *fctx;
 802         resquery_t *query;
 803         unsigned int rtt, rttms;
 804         unsigned int factor;
 805         dns_adbfind_t *find;
 806         dns_adbaddrinfo_t *addrinfo;
 807         isc_socket_t *socket;
 808
 809         query = *queryp;
 810         fctx = query->fctx;
 811
 812         FCTXTRACE("cancelquery");
 813
 814         REQUIRE(!RESQUERY_CANCELED(query));
 815
 816         query->attributes |= RESQUERY_ATTR_CANCELED;
 817
 818         /*
 819          * Should we update the RTT?
 820          */
 821         if (finish != NULL || no_response) {
 822                 if (finish != NULL) {
 823                         /*
 824                          * We have both the start and finish times for this
 825                          * packet, so we can compute a real RTT.
 826                          */
 827                         rtt = (unsigned int)isc_time_microdiff(finish,
 828                                                                &query->start);
 829                         factor = DNS_ADB_RTTADJDEFAULT;
 830
 831                         rttms = rtt / 1000;
 832                         if (rttms < DNS_RESOLVER_QRYRTTCLASS0) {
 833                                 inc_stats(fctx->res,
 834                                           dns_resstatscounter_queryrtt0);
 835                         } else if (rttms < DNS_RESOLVER_QRYRTTCLASS1) {
 836                                 inc_stats(fctx->res,
 837                                           dns_resstatscounter_queryrtt1);
 838                         } else if (rttms < DNS_RESOLVER_QRYRTTCLASS2) {
 839                                 inc_stats(fctx->res,
 840                                           dns_resstatscounter_queryrtt2);
 841                         } else if (rttms < DNS_RESOLVER_QRYRTTCLASS3) {
 842                                 inc_stats(fctx->res,
 843                                           dns_resstatscounter_queryrtt3);
 844                         } else if (rttms < DNS_RESOLVER_QRYRTTCLASS4) {
 845                                 inc_stats(fctx->res,
 846                                           dns_resstatscounter_queryrtt4);
 847                         } else {
 848                                 inc_stats(fctx->res,
 849                                           dns_resstatscounter_queryrtt5);
 850                         }
 851                 } else {
 852                         /*
 853                          * We don't have an RTT for this query.  Maybe the
 854                          * packet was lost, or maybe this server is very
 855                          * slow.  We don't know.  Increase the RTT.
 856                          */
 857                         INSIST(no_response);
 858                         rtt = query->addrinfo->srtt + 200000;
 859                         if (rtt > MAX_SINGLE_QUERY_TIMEOUT_US)
 860                                 rtt = MAX_SINGLE_QUERY_TIMEOUT_US;
 861                         /*
 862                          * Replace the current RTT with our value.
 863                          */
 864                         factor = DNS_ADB_RTTADJREPLACE;
 865                 }
 866                 dns_adb_adjustsrtt(fctx->adb, query->addrinfo, rtt, factor);
 867         }
 868
 869         /* Remember that the server has been tried. */
 870         if (!TRIED(query->addrinfo)) {
 871                 dns_adb_changeflags(fctx->adb, query->addrinfo,
 872                                     FCTX_ADDRINFO_TRIED, FCTX_ADDRINFO_TRIED);
 873         }
 874
 875         /*
 876          * Age RTTs of servers not tried.
 877          */
 878         factor = DNS_ADB_RTTADJAGE;
 879         if (finish != NULL)
 880                 for (addrinfo = ISC_LIST_HEAD(fctx->forwaddrs);
 881                      addrinfo != NULL;
 882                      addrinfo = ISC_LIST_NEXT(addrinfo, publink))
 883                         if (UNMARKED(addrinfo))
 884                                 dns_adb_adjustsrtt(fctx->adb, addrinfo,
 885                                                    0, factor);
 886
 887         if (finish != NULL && TRIEDFIND(fctx))
 888                 for (find = ISC_LIST_HEAD(fctx->finds);
 889                      find != NULL;
 890                      find = ISC_LIST_NEXT(find, publink))
 891                         for (addrinfo = ISC_LIST_HEAD(find->list);
 892                              addrinfo != NULL;
 893                              addrinfo = ISC_LIST_NEXT(addrinfo, publink))
 894                                 if (UNMARKED(addrinfo))
 895                                         dns_adb_adjustsrtt(fctx->adb, addrinfo,
 896                                                            0, factor);
 897
 898         if (finish != NULL && TRIEDALT(fctx)) {
 899                 for (addrinfo = ISC_LIST_HEAD(fctx->altaddrs);
 900                      addrinfo != NULL;
 901                      addrinfo = ISC_LIST_NEXT(addrinfo, publink))
 902                         if (UNMARKED(addrinfo))
 903                                 dns_adb_adjustsrtt(fctx->adb, addrinfo,
 904                                                    0, factor);
 905                 for (find = ISC_LIST_HEAD(fctx->altfinds);
 906                      find != NULL;
 907                      find = ISC_LIST_NEXT(find, publink))
 908                         for (addrinfo = ISC_LIST_HEAD(find->list);
 909                              addrinfo != NULL;
 910                              addrinfo = ISC_LIST_NEXT(addrinfo, publink))
 911                                 if (UNMARKED(addrinfo))
 912                                         dns_adb_adjustsrtt(fctx->adb, addrinfo,
 913                                                            0, factor);
 914         }
 915
 916         /*
 917          * Check for any outstanding socket events.  If they exist, cancel
 918          * them and let the event handlers finish the cleanup.  The resolver
 919          * only needs to worry about managing the connect and send events;
 920          * the dispatcher manages the recv events.
 921          */
 922         if (RESQUERY_CONNECTING(query)) {
 923                 /*
 924                  * Cancel the connect.
 925                  */
 926                 if (query->tcpsocket != NULL) {
 927                         isc_socket_cancel(query->tcpsocket, NULL,
 928                                           ISC_SOCKCANCEL_CONNECT);
 929                 } else if (query->dispentry != NULL) {
 930                         INSIST(query->exclusivesocket);
 931                         socket = dns_dispatch_getentrysocket(query->dispentry);
 932                         if (socket != NULL)
 933                                 isc_socket_cancel(socket, NULL,
 934                                                   ISC_SOCKCANCEL_CONNECT);
 935                 }
 936         } else if (RESQUERY_SENDING(query)) {
 937                 /*
 938                  * Cancel the pending send.
 939                  */
 940                 if (query->exclusivesocket && query->dispentry != NULL)
 941                         socket = dns_dispatch_getentrysocket(query->dispentry);
 942                 else
 943                         socket = dns_dispatch_getsocket(query->dispatch);
 944                 if (socket != NULL)
 945                         isc_socket_cancel(socket, NULL, ISC_SOCKCANCEL_SEND);
 946         }
 947
 948         if (query->dispentry != NULL)
 949                 dns_dispatch_removeresponse(&query->dispentry, deventp);
 950
 951         ISC_LIST_UNLINK(fctx->queries, query, link);
 952
 953         if (query->tsig != NULL)
 954                 isc_buffer_free(&query->tsig);
 955
 956         if (query->tsigkey != NULL)
 957                 dns_tsigkey_detach(&query->tsigkey);
 958
 959         if (query->dispatch != NULL)
 960                 dns_dispatch_detach(&query->dispatch);
 961
 962         if (! (RESQUERY_CONNECTING(query) || RESQUERY_SENDING(query)))
 963                 /*
 964                  * It's safe to destroy the query now.
 965                  */
 966                 resquery_destroy(&query);
 967 }
 968
 969 static void
 970 fctx_cancelqueries(fetchctx_t *fctx, isc_boolean_t no_response) {
 971         resquery_t *query, *next_query;
 972
 973         FCTXTRACE("cancelqueries");
 974
 975         for (query = ISC_LIST_HEAD(fctx->queries);
 976              query != NULL;
 977              query = next_query) {
 978                 next_query = ISC_LIST_NEXT(query, link);
 979                 fctx_cancelquery(&query, NULL, NULL, no_response);
 980         }
 981 }
 982
 983 static void
 984 fctx_cleanupfinds(fetchctx_t *fctx) {
 985         dns_adbfind_t *find, *next_find;
 986
 987         REQUIRE(ISC_LIST_EMPTY(fctx->queries));
 988
 989         for (find = ISC_LIST_HEAD(fctx->finds);
 990              find != NULL;
 991              find = next_find) {
 992                 next_find = ISC_LIST_NEXT(find, publink);
 993                 ISC_LIST_UNLINK(fctx->finds, find, publink);
 994                 dns_adb_destroyfind(&find);
 995         }
 996         fctx->find = NULL;
 997 }
 998
 999 static void
1000 fctx_cleanupaltfinds(fetchctx_t *fctx) {
1001         dns_adbfind_t *find, *next_find;
1002
1003         REQUIRE(ISC_LIST_EMPTY(fctx->queries));
1004
1005         for (find = ISC_LIST_HEAD(fctx->altfinds);
1006              find != NULL;
1007              find = next_find) {
1008                 next_find = ISC_LIST_NEXT(find, publink);
1009                 ISC_LIST_UNLINK(fctx->altfinds, find, publink);
1010                 dns_adb_destroyfind(&find);
1011         }
1012         fctx->altfind = NULL;
1013 }
1014
1015 static void
1016 fctx_cleanupforwaddrs(fetchctx_t *fctx) {
1017         dns_adbaddrinfo_t *addr, *next_addr;
1018
1019         REQUIRE(ISC_LIST_EMPTY(fctx->queries));
1020
1021         for (addr = ISC_LIST_HEAD(fctx->forwaddrs);
1022              addr != NULL;
1023              addr = next_addr) {
1024                 next_addr = ISC_LIST_NEXT(addr, publink);
1025                 ISC_LIST_UNLINK(fctx->forwaddrs, addr, publink);
1026                 dns_adb_freeaddrinfo(fctx->adb, &addr);
1027         }
1028 }
1029
1030 static void
1031 fctx_cleanupaltaddrs(fetchctx_t *fctx) {
1032         dns_adbaddrinfo_t *addr, *next_addr;
1033
1034         REQUIRE(ISC_LIST_EMPTY(fctx->queries));
1035
1036         for (addr = ISC_LIST_HEAD(fctx->altaddrs);
1037              addr != NULL;
1038              addr = next_addr) {
1039                 next_addr = ISC_LIST_NEXT(addr, publink);
1040                 ISC_LIST_UNLINK(fctx->altaddrs, addr, publink);
1041                 dns_adb_freeaddrinfo(fctx->adb, &addr);
1042         }
1043 }
1044
1045 static inline void
1046 fctx_stopeverything(fetchctx_t *fctx, isc_boolean_t no_response) {
1047         FCTXTRACE("stopeverything");
1048         fctx_cancelqueries(fctx, no_response);
1049         fctx_cleanupfinds(fctx);
1050         fctx_cleanupaltfinds(fctx);
1051         fctx_cleanupforwaddrs(fctx);
1052         fctx_cleanupaltaddrs(fctx);
1053         fctx_stoptimer(fctx);
1054 }
1055
1056 static inline void
1057 fctx_sendevents(fetchctx_t *fctx, isc_result_t result, int line) {
1058         dns_fetchevent_t *event, *next_event;
1059         isc_task_t *task;
1060         unsigned int count = 0;
1061         isc_interval_t i;
1062         isc_boolean_t logit = ISC_FALSE;
1063         isc_time_t now;
1064         unsigned int old_spillat;
1065         unsigned int new_spillat = 0;   /* initialized to silence
1066                                            compiler warnings */
1067
1068         /*
1069          * Caller must be holding the appropriate bucket lock.
1070          */
1071         REQUIRE(fctx->state == fetchstate_done);
1072
1073         FCTXTRACE("sendevents");
1074
1075         /*
1076          * Keep some record of fetch result for logging later (if required).
1077          */
1078         fctx->result = result;
1079         fctx->exitline = line;
1080         TIME_NOW(&now);
1081         fctx->duration = isc_time_microdiff(&now, &fctx->start);
1082
1083         for (event = ISC_LIST_HEAD(fctx->events);
1084              event != NULL;
1085              event = next_event) {
1086                 next_event = ISC_LIST_NEXT(event, ev_link);
1087                 ISC_LIST_UNLINK(fctx->events, event, ev_link);
1088                 task = event->ev_sender;
1089                 event->ev_sender = fctx;
1090                 event->vresult = fctx->vresult;
1091                 if (!HAVE_ANSWER(fctx))
1092                         event->result = result;
1093
1094                 INSIST(result != ISC_R_SUCCESS ||
1095                        dns_rdataset_isassociated(event->rdataset) ||
1096                        fctx->type == dns_rdatatype_any ||
1097                        fctx->type == dns_rdatatype_rrsig ||
1098                        fctx->type == dns_rdatatype_sig);
1099
1100                 /*
1101                  * Negative results must be indicated in event->result.
1102                  */
1103                 if (dns_rdataset_isassociated(event->rdataset) &&
1104                     NEGATIVE(event->rdataset)) {
1105                         INSIST(event->result == DNS_R_NCACHENXDOMAIN ||
1106                                event->result == DNS_R_NCACHENXRRSET);
1107                 }
1108
1109                 event->qtotal = fctx->totalqueries;
1110                 isc_task_sendanddetach(&task, ISC_EVENT_PTR(&event));
1111                 count++;
1112         }
1113
1114         if ((fctx->attributes & FCTX_ATTR_HAVEANSWER) != 0 &&
1115             fctx->spilled &&
1116             (count < fctx->res->spillatmax || fctx->res->spillatmax == 0)) {
1117                 LOCK(&fctx->res->lock);
1118                 if (count == fctx->res->spillat && !fctx->res->exiting) {
1119                         old_spillat = fctx->res->spillat;
1120                         fctx->res->spillat += 5;
1121                         if (fctx->res->spillat > fctx->res->spillatmax &&
1122                             fctx->res->spillatmax != 0)
1123                                 fctx->res->spillat = fctx->res->spillatmax;
1124                         new_spillat = fctx->res->spillat;
1125                         if (new_spillat != old_spillat) {
1126                                 logit = ISC_TRUE;
1127                         }
1128                         isc_interval_set(&i, 20 * 60, 0);
1129                         result = isc_timer_reset(fctx->res->spillattimer,
1130                                                  isc_timertype_ticker, NULL,
1131                                                  &i, ISC_TRUE);
1132                         RUNTIME_CHECK(result == ISC_R_SUCCESS);
1133                 }
1134                 UNLOCK(&fctx->res->lock);
1135                 if (logit)
1136                         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
1137                                       DNS_LOGMODULE_RESOLVER, ISC_LOG_NOTICE,
1138                                       "clients-per-query increased to %u",
1139                                       new_spillat);
1140         }
1141 }
1142
1143 static inline void
1144 log_edns(fetchctx_t *fctx) {
1145         char domainbuf[DNS_NAME_FORMATSIZE];
1146
1147         if (fctx->reason == NULL)
1148                 return;
1149
1150         /*
1151          * We do not know if fctx->domain is the actual domain the record
1152          * lives in or a parent domain so we have a '?' after it.
1153          */
1154         dns_name_format(&fctx->domain, domainbuf, sizeof(domainbuf));
1155         isc_log_write(dns_lctx, DNS_LOGCATEGORY_EDNS_DISABLED,
1156                       DNS_LOGMODULE_RESOLVER, ISC_LOG_INFO,
1157                       "success resolving '%s' (in '%s'?) after %s",
1158                       fctx->info, domainbuf, fctx->reason);
1159
1160         fctx->reason = NULL;
1161 }
1162
1163 static void
1164 fctx_done(fetchctx_t *fctx, isc_result_t result, int line) {
1165         dns_resolver_t *res;
1166         isc_boolean_t no_response;
1167
1168         REQUIRE(line >= 0);
1169
1170         FCTXTRACE("done");
1171
1172         res = fctx->res;
1173
1174         if (result == ISC_R_SUCCESS) {
1175                 /*%
1176                  * Log any deferred EDNS timeout messages.
1177                  */
1178                 log_edns(fctx);
1179                 no_response = ISC_TRUE;
1180          } else
1181                 no_response = ISC_FALSE;
1182
1183         fctx->reason = NULL;
1184         fctx_stopeverything(fctx, no_response);
1185
1186         LOCK(&res->buckets[fctx->bucketnum].lock);
1187
1188         fctx->state = fetchstate_done;
1189         fctx->attributes &= ~FCTX_ATTR_ADDRWAIT;
1190         fctx_sendevents(fctx, result, line);
1191
1192         UNLOCK(&res->buckets[fctx->bucketnum].lock);
1193 }
1194
1195 static void
1196 process_sendevent(resquery_t *query, isc_event_t *event) {
1197         isc_socketevent_t *sevent = (isc_socketevent_t *)event;
1198         isc_boolean_t retry = ISC_FALSE;
1199         isc_result_t result;
1200         fetchctx_t *fctx;
1201
1202         fctx = query->fctx;
1203
1204         if (RESQUERY_CANCELED(query)) {
1205                 if (query->sends == 0 && query->connects == 0) {
1206                         /*
1207                          * This query was canceled while the
1208                          * isc_socket_sendto/connect() was in progress.
1209                          */
1210                         if (query->tcpsocket != NULL)
1211                                 isc_socket_detach(&query->tcpsocket);
1212                         resquery_destroy(&query);
1213                 }
1214         } else {
1215                 switch (sevent->result) {
1216                 case ISC_R_SUCCESS:
1217                         break;
1218
1219                 case ISC_R_HOSTUNREACH:
1220                 case ISC_R_NETUNREACH:
1221                 case ISC_R_NOPERM:
1222                 case ISC_R_ADDRNOTAVAIL:
1223                 case ISC_R_CONNREFUSED:
1224
1225                         /*
1226                          * No route to remote.
1227                          */
1228                         add_bad(fctx, query->addrinfo, sevent->result,
1229                                 badns_unreachable);
1230                         fctx_cancelquery(&query, NULL, NULL, ISC_TRUE);
1231                         retry = ISC_TRUE;
1232                         break;
1233
1234                 default:
1235                         fctx_cancelquery(&query, NULL, NULL, ISC_FALSE);
1236                         break;
1237                 }
1238         }
1239
1240         if (event->ev_type == ISC_SOCKEVENT_CONNECT)
1241                 isc_event_free(&event);
1242
1243         if (retry) {
1244                 /*
1245                  * Behave as if the idle timer has expired.  For TCP
1246                  * this may not actually reflect the latest timer.
1247                  */
1248                 fctx->attributes &= ~FCTX_ATTR_ADDRWAIT;
1249                 result = fctx_stopidletimer(fctx);
1250                 if (result != ISC_R_SUCCESS)
1251                         fctx_done(fctx, result, __LINE__);
1252                 else
1253                         fctx_try(fctx, ISC_TRUE, ISC_FALSE);
1254         }
1255 }
1256
1257 static void
1258 resquery_udpconnected(isc_task_t *task, isc_event_t *event) {
1259         resquery_t *query = event->ev_arg;
1260
1261         REQUIRE(event->ev_type == ISC_SOCKEVENT_CONNECT);
1262
1263         QTRACE("udpconnected");
1264
1265         UNUSED(task);
1266
1267         INSIST(RESQUERY_CONNECTING(query));
1268
1269         query->connects--;
1270
1271         process_sendevent(query, event);
1272 }
1273
1274 static void
1275 resquery_senddone(isc_task_t *task, isc_event_t *event) {
1276         resquery_t *query = event->ev_arg;
1277
1278         REQUIRE(event->ev_type == ISC_SOCKEVENT_SENDDONE);
1279
1280         QTRACE("senddone");
1281
1282         /*
1283          * XXXRTH
1284          *
1285          * Currently we don't wait for the senddone event before retrying
1286          * a query.  This means that if we get really behind, we may end
1287          * up doing extra work!
1288          */
1289
1290         UNUSED(task);
1291
1292         INSIST(RESQUERY_SENDING(query));
1293
1294         query->sends--;
1295
1296         process_sendevent(query, event);
1297 }
1298
1299 static inline isc_result_t
1300 fctx_addopt(dns_message_t *message, unsigned int version,
1301             isc_uint16_t udpsize, dns_ednsopt_t *ednsopts, size_t count)
1302 {
1303         dns_rdataset_t *rdataset = NULL;
1304         isc_result_t result;
1305
1306         result = dns_message_buildopt(message, &rdataset, version, udpsize,
1307                                       DNS_MESSAGEEXTFLAG_DO, ednsopts, count);
1308         if (result != ISC_R_SUCCESS)
1309                 return (result);
1310         return (dns_message_setopt(message, rdataset));
1311 }
1312
1313 static inline void
1314 fctx_setretryinterval(fetchctx_t *fctx, unsigned int rtt) {
1315         unsigned int seconds;
1316         unsigned int us;
1317
1318         /*
1319          * We retry every .8 seconds the first two times through the address
1320          * list, and then we do exponential back-off.
1321          */
1322         if (fctx->restarts < 3)
1323                 us = 800000;
1324         else
1325                 us = (800000 << (fctx->restarts - 2));
1326
1327         /*
1328          * Add a fudge factor to the expected rtt based on the current
1329          * estimate.
1330          */
1331         if (rtt < 50000)
1332                 rtt += 50000;
1333         else if (rtt < 100000)
1334                 rtt += 100000;
1335         else
1336                 rtt += 200000;
1337
1338         /*
1339          * Always wait for at least the expected rtt.
1340          */
1341         if (us < rtt)
1342                 us = rtt;
1343
1344         /*
1345          * But don't ever wait for more than 10 seconds.
1346          */
1347         if (us > MAX_SINGLE_QUERY_TIMEOUT_US)
1348                 us = MAX_SINGLE_QUERY_TIMEOUT_US;
1349
1350         seconds = us / US_PER_SEC;
1351         us -= seconds * US_PER_SEC;
1352         isc_interval_set(&fctx->interval, seconds, us * 1000);
1353 }
1354
1355 static isc_result_t
1356 fctx_query(fetchctx_t *fctx, dns_adbaddrinfo_t *addrinfo,
1357            unsigned int options)
1358 {
1359         dns_resolver_t *res;
1360         isc_task_t *task;
1361         isc_result_t result;
1362         resquery_t *query;
1363         isc_sockaddr_t addr;
1364         isc_boolean_t have_addr = ISC_FALSE;
1365         unsigned int srtt;
1366
1367         FCTXTRACE("query");
1368
1369         res = fctx->res;
1370         task = res->buckets[fctx->bucketnum].task;
1371
1372         srtt = addrinfo->srtt;
1373
1374         /*
1375          * A forwarder needs to make multiple queries. Give it at least
1376          * a second to do these in.
1377          */
1378         if (ISFORWARDER(addrinfo) && srtt < 1000000)
1379                 srtt = 1000000;
1380
1381         fctx_setretryinterval(fctx, srtt);
1382         result = fctx_startidletimer(fctx, &fctx->interval);
1383         if (result != ISC_R_SUCCESS)
1384                 return (result);
1385
1386         INSIST(ISC_LIST_EMPTY(fctx->validators));
1387
1388         dns_message_reset(fctx->rmessage, DNS_MESSAGE_INTENTPARSE);
1389
1390         query = isc_mem_get(fctx->mctx, sizeof(*query));
1391         if (query == NULL) {
1392                 result = ISC_R_NOMEMORY;
1393                 goto stop_idle_timer;
1394         }
1395         query->mctx = fctx->mctx;
1396         query->options = options;
1397         query->attributes = 0;
1398         query->sends = 0;
1399         query->connects = 0;
1400         /*
1401          * Note that the caller MUST guarantee that 'addrinfo' will remain
1402          * valid until this query is canceled.
1403          */
1404         query->addrinfo = addrinfo;
1405         TIME_NOW(&query->start);
1406
1407         /*
1408          * If this is a TCP query, then we need to make a socket and
1409          * a dispatch for it here.  Otherwise we use the resolver's
1410          * shared dispatch.
1411          */
1412         query->dispatchmgr = res->dispatchmgr;
1413         query->dispatch = NULL;
1414         query->exclusivesocket = ISC_FALSE;
1415         query->tcpsocket = NULL;
1416         if (res->view->peers != NULL) {
1417                 dns_peer_t *peer = NULL;
1418                 isc_netaddr_t dstip;
1419                 isc_netaddr_fromsockaddr(&dstip, &addrinfo->sockaddr);
1420                 result = dns_peerlist_peerbyaddr(res->view->peers,
1421                                                  &dstip, &peer);
1422                 if (result == ISC_R_SUCCESS) {
1423                         result = dns_peer_getquerysource(peer, &addr);
1424                         if (result == ISC_R_SUCCESS)
1425                                 have_addr = ISC_TRUE;
1426                 }
1427         }
1428
1429         if ((query->options & DNS_FETCHOPT_TCP) != 0) {
1430                 int pf;
1431
1432                 pf = isc_sockaddr_pf(&addrinfo->sockaddr);
1433                 if (!have_addr) {
1434                         switch (pf) {
1435                         case PF_INET:
1436                                 result = dns_dispatch_getlocaladdress(
1437                                               res->dispatches4->dispatches[0],
1438                                               &addr);
1439                                 break;
1440                         case PF_INET6:
1441                                 result = dns_dispatch_getlocaladdress(
1442                                               res->dispatches6->dispatches[0],
1443                                               &addr);
1444                                 break;
1445                         default:
1446                                 result = ISC_R_NOTIMPLEMENTED;
1447                                 break;
1448                         }
1449                         if (result != ISC_R_SUCCESS)
1450                                 goto cleanup_query;
1451                 }
1452                 isc_sockaddr_setport(&addr, 0);
1453
1454                 result = isc_socket_create(res->socketmgr, pf,
1455                                            isc_sockettype_tcp,
1456                                            &query->tcpsocket);
1457                 if (result != ISC_R_SUCCESS)
1458                         goto cleanup_query;
1459
1460 #ifndef BROKEN_TCP_BIND_BEFORE_CONNECT
1461                 result = isc_socket_bind(query->tcpsocket, &addr, 0);
1462                 if (result != ISC_R_SUCCESS)
1463                         goto cleanup_socket;
1464 #endif
1465
1466                 /*
1467                  * A dispatch will be created once the connect succeeds.
1468                  */
1469         } else {
1470                 if (have_addr) {
1471                         unsigned int attrs, attrmask;
1472                         attrs = DNS_DISPATCHATTR_UDP;
1473                         switch (isc_sockaddr_pf(&addr)) {
1474                         case AF_INET:
1475                                 attrs |= DNS_DISPATCHATTR_IPV4;
1476                                 break;
1477                         case AF_INET6:
1478                                 attrs |= DNS_DISPATCHATTR_IPV6;
1479                                 break;
1480                         default:
1481                                 result = ISC_R_NOTIMPLEMENTED;
1482                                 goto cleanup_query;
1483                         }
1484                         attrmask = DNS_DISPATCHATTR_UDP;
1485                         attrmask |= DNS_DISPATCHATTR_TCP;
1486                         attrmask |= DNS_DISPATCHATTR_IPV4;
1487                         attrmask |= DNS_DISPATCHATTR_IPV6;
1488                         result = dns_dispatch_getudp(res->dispatchmgr,
1489                                                      res->socketmgr,
1490                                                      res->taskmgr, &addr,
1491                                                      4096, 1000, 32768, 16411,
1492                                                      16433, attrs, attrmask,
1493                                                      &query->dispatch);
1494                         if (result != ISC_R_SUCCESS)
1495                                 goto cleanup_query;
1496                 } else {
1497                         switch (isc_sockaddr_pf(&addrinfo->sockaddr)) {
1498                         case PF_INET:
1499                                 dns_dispatch_attach(
1500                                     dns_resolver_dispatchv4(res),
1501                                     &query->dispatch);
1502                                 query->exclusivesocket = res->exclusivev4;
1503                                 break;
1504                         case PF_INET6:
1505                                 dns_dispatch_attach(
1506                                     dns_resolver_dispatchv6(res),
1507                                     &query->dispatch);
1508                                 query->exclusivesocket = res->exclusivev6;
1509                                 break;
1510                         default:
1511                                 result = ISC_R_NOTIMPLEMENTED;
1512                                 goto cleanup_query;
1513                         }
1514                 }
1515                 /*
1516                  * We should always have a valid dispatcher here.  If we
1517                  * don't support a protocol family, then its dispatcher
1518                  * will be NULL, but we shouldn't be finding addresses for
1519                  * protocol types we don't support, so the dispatcher
1520                  * we found should never be NULL.
1521                  */
1522                 INSIST(query->dispatch != NULL);
1523         }
1524
1525         query->dispentry = NULL;
1526         query->fctx = fctx;
1527         query->tsig = NULL;
1528         query->tsigkey = NULL;
1529         ISC_LINK_INIT(query, link);
1530         query->magic = QUERY_MAGIC;
1531
1532         if ((query->options & DNS_FETCHOPT_TCP) != 0) {
1533                 /*
1534                  * Connect to the remote server.
1535                  *
1536                  * XXXRTH  Should we attach to the socket?
1537                  */
1538                 result = isc_socket_connect(query->tcpsocket,
1539                                             &addrinfo->sockaddr, task,
1540                                             resquery_connected, query);
1541                 if (result != ISC_R_SUCCESS)
1542                         goto cleanup_socket;
1543                 query->connects++;
1544                 QTRACE("connecting via TCP");
1545         } else {
1546                 result = resquery_send(query);
1547                 if (result != ISC_R_SUCCESS)
1548                         goto cleanup_dispatch;
1549         }
1550
1551         fctx->querysent++;
1552         fctx->totalqueries++;
1553
1554         ISC_LIST_APPEND(fctx->queries, query, link);
1555         query->fctx->nqueries++;
1556         if (isc_sockaddr_pf(&addrinfo->sockaddr) == PF_INET)
1557                 inc_stats(res, dns_resstatscounter_queryv4);
1558         else
1559                 inc_stats(res, dns_resstatscounter_queryv6);
1560         if (res->view->resquerystats != NULL)
1561                 dns_rdatatypestats_increment(res->view->resquerystats,
1562                                              fctx->type);
1563
1564         return (ISC_R_SUCCESS);
1565
1566  cleanup_socket:
1567         isc_socket_detach(&query->tcpsocket);
1568
1569  cleanup_dispatch:
1570         if (query->dispatch != NULL)
1571                 dns_dispatch_detach(&query->dispatch);
1572
1573  cleanup_query:
1574         if (query->connects == 0) {
1575                 query->magic = 0;
1576                 isc_mem_put(fctx->mctx, query, sizeof(*query));
1577         }
1578
1579  stop_idle_timer:
1580         RUNTIME_CHECK(fctx_stopidletimer(fctx) == ISC_R_SUCCESS);
1581
1582         return (result);
1583 }
1584
1585 static isc_boolean_t
1586 bad_edns(fetchctx_t *fctx, isc_sockaddr_t *address) {
1587         isc_sockaddr_t *sa;
1588
1589         for (sa = ISC_LIST_HEAD(fctx->bad_edns);
1590              sa != NULL;
1591              sa = ISC_LIST_NEXT(sa, link)) {
1592                 if (isc_sockaddr_equal(sa, address))
1593                         return (ISC_TRUE);
1594         }
1595
1596         return (ISC_FALSE);
1597 }
1598
1599 static void
1600 add_bad_edns(fetchctx_t *fctx, isc_sockaddr_t *address) {
1601         isc_sockaddr_t *sa;
1602
1603         if (bad_edns(fctx, address))
1604                 return;
1605
1606         sa = isc_mem_get(fctx->mctx, sizeof(*sa));
1607         if (sa == NULL)
1608                 return;
1609
1610         *sa = *address;
1611         ISC_LIST_INITANDAPPEND(fctx->bad_edns, sa, link);
1612 }
1613
1614 static isc_boolean_t
1615 triededns(fetchctx_t *fctx, isc_sockaddr_t *address) {
1616         isc_sockaddr_t *sa;
1617
1618         for (sa = ISC_LIST_HEAD(fctx->edns);
1619              sa != NULL;
1620              sa = ISC_LIST_NEXT(sa, link)) {
1621                 if (isc_sockaddr_equal(sa, address))
1622                         return (ISC_TRUE);
1623         }
1624
1625         return (ISC_FALSE);
1626 }
1627
1628 static void
1629 add_triededns(fetchctx_t *fctx, isc_sockaddr_t *address) {
1630         isc_sockaddr_t *sa;
1631
1632         if (triededns(fctx, address))
1633                 return;
1634
1635         sa = isc_mem_get(fctx->mctx, sizeof(*sa));
1636         if (sa == NULL)
1637                 return;
1638
1639         *sa = *address;
1640         ISC_LIST_INITANDAPPEND(fctx->edns, sa, link);
1641 }
1642
1643 static isc_boolean_t
1644 triededns512(fetchctx_t *fctx, isc_sockaddr_t *address) {
1645         isc_sockaddr_t *sa;
1646
1647         for (sa = ISC_LIST_HEAD(fctx->edns512);
1648              sa != NULL;
1649              sa = ISC_LIST_NEXT(sa, link)) {
1650                 if (isc_sockaddr_equal(sa, address))
1651                         return (ISC_TRUE);
1652         }
1653
1654         return (ISC_FALSE);
1655 }
1656
1657 static void
1658 add_triededns512(fetchctx_t *fctx, isc_sockaddr_t *address) {
1659         isc_sockaddr_t *sa;
1660
1661         if (triededns512(fctx, address))
1662                 return;
1663
1664         sa = isc_mem_get(fctx->mctx, sizeof(*sa));
1665         if (sa == NULL)
1666                 return;
1667
1668         *sa = *address;
1669         ISC_LIST_INITANDAPPEND(fctx->edns512, sa, link);
1670 }
1671
1672 static isc_result_t
1673 resquery_send(resquery_t *query) {
1674         fetchctx_t *fctx;
1675         isc_result_t result;
1676         dns_name_t *qname = NULL;
1677         dns_rdataset_t *qrdataset = NULL;
1678         isc_region_t r;
1679         dns_resolver_t *res;
1680         isc_task_t *task;
1681         isc_socket_t *socket;
1682         isc_buffer_t tcpbuffer;
1683         isc_sockaddr_t *address;
1684         isc_buffer_t *buffer;
1685         isc_netaddr_t ipaddr;
1686         dns_tsigkey_t *tsigkey = NULL;
1687         dns_peer_t *peer = NULL;
1688         isc_boolean_t useedns;
1689         dns_compress_t cctx;
1690         isc_boolean_t cleanup_cctx = ISC_FALSE;
1691         isc_boolean_t secure_domain;
1692         isc_boolean_t connecting = ISC_FALSE;
1693         dns_ednsopt_t ednsopts[EDNSOPTS];
1694         unsigned ednsopt = 0;
1695
1696         fctx = query->fctx;
1697         QTRACE("send");
1698
1699         res = fctx->res;
1700         task = res->buckets[fctx->bucketnum].task;
1701         address = NULL;
1702
1703         if ((query->options & DNS_FETCHOPT_TCP) != 0) {
1704                 /*
1705                  * Reserve space for the TCP message length.
1706                  */
1707                 isc_buffer_init(&tcpbuffer, query->data, sizeof(query->data));
1708                 isc_buffer_init(&query->buffer, query->data + 2,
1709                                 sizeof(query->data) - 2);
1710                 buffer = &tcpbuffer;
1711         } else {
1712                 isc_buffer_init(&query->buffer, query->data,
1713                                 sizeof(query->data));
1714                 buffer = &query->buffer;
1715         }
1716
1717         result = dns_message_gettempname(fctx->qmessage, &qname);
1718         if (result != ISC_R_SUCCESS)
1719                 goto cleanup_temps;
1720         result = dns_message_gettemprdataset(fctx->qmessage, &qrdataset);
1721         if (result != ISC_R_SUCCESS)
1722                 goto cleanup_temps;
1723
1724         /*
1725          * Get a query id from the dispatch.
1726          */
1727         result = dns_dispatch_addresponse2(query->dispatch,
1728                                            &query->addrinfo->sockaddr,
1729                                            task,
1730                                            resquery_response,
1731                                            query,
1732                                            &query->id,
1733                                            &query->dispentry,
1734                                            res->socketmgr);
1735         if (result != ISC_R_SUCCESS)
1736                 goto cleanup_temps;
1737
1738         fctx->qmessage->opcode = dns_opcode_query;
1739
1740         /*
1741          * Set up question.
1742          */
1743         dns_name_init(qname, NULL);
1744         dns_name_clone(&fctx->name, qname);
1745         dns_rdataset_init(qrdataset);
1746         dns_rdataset_makequestion(qrdataset, res->rdclass, fctx->type);
1747         ISC_LIST_APPEND(qname->list, qrdataset, link);
1748         dns_message_addname(fctx->qmessage, qname, DNS_SECTION_QUESTION);
1749         qname = NULL;
1750         qrdataset = NULL;
1751
1752         /*
1753          * Set RD if the client has requested that we do a recursive query,
1754          * or if we're sending to a forwarder.
1755          */
1756         if ((query->options & DNS_FETCHOPT_RECURSIVE) != 0 ||
1757             ISFORWARDER(query->addrinfo))
1758                 fctx->qmessage->flags |= DNS_MESSAGEFLAG_RD;
1759
1760         /*
1761          * Set CD if the client says don't validate or the question is
1762          * under a secure entry point.
1763          */
1764         if ((query->options & DNS_FETCHOPT_NOVALIDATE) != 0) {
1765                 fctx->qmessage->flags |= DNS_MESSAGEFLAG_CD;
1766         } else if (res->view->enablevalidation) {
1767                 result = dns_view_issecuredomain(res->view, &fctx->name,
1768                                                  &secure_domain);
1769                 if (result != ISC_R_SUCCESS)
1770                         secure_domain = ISC_FALSE;
1771                 if (res->view->dlv != NULL)
1772                         secure_domain = ISC_TRUE;
1773                 if (secure_domain)
1774                         fctx->qmessage->flags |= DNS_MESSAGEFLAG_CD;
1775         }
1776
1777         /*
1778          * We don't have to set opcode because it defaults to query.
1779          */
1780         fctx->qmessage->id = query->id;
1781
1782         /*
1783          * Convert the question to wire format.
1784          */
1785         result = dns_compress_init(&cctx, -1, fctx->res->mctx);
1786         if (result != ISC_R_SUCCESS)
1787                 goto cleanup_message;
1788         cleanup_cctx = ISC_TRUE;
1789
1790         result = dns_message_renderbegin(fctx->qmessage, &cctx,
1791                                          &query->buffer);
1792         if (result != ISC_R_SUCCESS)
1793                 goto cleanup_message;
1794
1795         result = dns_message_rendersection(fctx->qmessage,
1796                                            DNS_SECTION_QUESTION, 0);
1797         if (result != ISC_R_SUCCESS)
1798                 goto cleanup_message;
1799
1800         peer = NULL;
1801         isc_netaddr_fromsockaddr(&ipaddr, &query->addrinfo->sockaddr);
1802         (void) dns_peerlist_peerbyaddr(fctx->res->view->peers, &ipaddr, &peer);
1803
1804         /*
1805          * The ADB does not know about servers with "edns no".  Check this,
1806          * and then inform the ADB for future use.
1807          */
1808         if ((query->addrinfo->flags & DNS_FETCHOPT_NOEDNS0) == 0 &&
1809             peer != NULL &&
1810             dns_peer_getsupportedns(peer, &useedns) == ISC_R_SUCCESS &&
1811             !useedns)
1812         {
1813                 query->options |= DNS_FETCHOPT_NOEDNS0;
1814                 dns_adb_changeflags(fctx->adb, query->addrinfo,
1815                                     DNS_FETCHOPT_NOEDNS0,
1816                                     DNS_FETCHOPT_NOEDNS0);
1817         }
1818
1819         /* Sync NOEDNS0 flag in addrinfo->flags and options now. */
1820         if ((query->addrinfo->flags & DNS_FETCHOPT_NOEDNS0) != 0)
1821                 query->options |= DNS_FETCHOPT_NOEDNS0;
1822
1823         /*
1824          * Handle timeouts by reducing the UDP response size to 512 bytes
1825          * then if that doesn't work disabling EDNS (includes DO) and CD.
1826          *
1827          * These timeout can be due to:
1828          *      * broken nameservers that don't respond to EDNS queries.
1829          *      * broken/misconfigured firewalls and NAT implementations
1830          *        that don't handle IP fragmentation.
1831          *      * broken/misconfigured firewalls that don't handle responses
1832          *        greater than 512 bytes.
1833          *      * broken/misconfigured firewalls that don't handle EDNS, DO
1834          *        or CD.
1835          *      * packet loss / link outage.
1836          */
1837         if (fctx->timeout) {
1838                 if ((triededns512(fctx, &query->addrinfo->sockaddr) ||
1839                      fctx->timeouts >= (MAX_EDNS0_TIMEOUTS * 2)) &&
1840                     (query->options & DNS_FETCHOPT_NOEDNS0) == 0) {
1841                         query->options |= DNS_FETCHOPT_NOEDNS0;
1842                         fctx->reason = "disabling EDNS";
1843                 } else if ((triededns(fctx, &query->addrinfo->sockaddr) ||
1844                             fctx->timeouts >= MAX_EDNS0_TIMEOUTS) &&
1845                            (query->options & DNS_FETCHOPT_NOEDNS0) == 0) {
1846                         query->options |= DNS_FETCHOPT_EDNS512;
1847                         fctx->reason = "reducing the advertised EDNS UDP "
1848                                        "packet size to 512 octets";
1849                 }
1850                 fctx->timeout = ISC_FALSE;
1851         }
1852
1853         /*
1854          * Use EDNS0, unless the caller doesn't want it, or we know that
1855          * the remote server doesn't like it.
1856          */
1857         if ((query->options & DNS_FETCHOPT_NOEDNS0) == 0) {
1858                 if ((query->addrinfo->flags & DNS_FETCHOPT_NOEDNS0) == 0) {
1859                         unsigned int version = 0;       /* Default version. */
1860                         unsigned int flags;
1861                         isc_uint16_t udpsize = res->udpsize;
1862                         isc_boolean_t reqnsid = res->view->requestnsid;
1863
1864                         flags = query->addrinfo->flags;
1865                         if ((flags & DNS_FETCHOPT_EDNSVERSIONSET) != 0) {
1866                                 version = flags & DNS_FETCHOPT_EDNSVERSIONMASK;
1867                                 version >>= DNS_FETCHOPT_EDNSVERSIONSHIFT;
1868                         }
1869                         if ((query->options & DNS_FETCHOPT_EDNS512) != 0)
1870                                 udpsize = 512;
1871                         else if (peer != NULL)
1872                                 (void)dns_peer_getudpsize(peer, &udpsize);
1873
1874                         /* request NSID for current view or peer? */
1875                         if (peer != NULL)
1876                                 (void) dns_peer_getrequestnsid(peer, &reqnsid);
1877                         if (reqnsid) {
1878                                 INSIST(ednsopt < EDNSOPTS);
1879                                 ednsopts[ednsopt].code = DNS_OPT_NSID;
1880                                 ednsopts[ednsopt].length = 0;
1881                                 ednsopts[ednsopt].value = NULL;
1882                                 ednsopt++;
1883                         }
1884                         result = fctx_addopt(fctx->qmessage, version,
1885                                              udpsize, ednsopts, ednsopt);
1886                         if (reqnsid && result == ISC_R_SUCCESS) {
1887                                 query->options |= DNS_FETCHOPT_WANTNSID;
1888                         } else if (result != ISC_R_SUCCESS) {
1889                                 /*
1890                                  * We couldn't add the OPT, but we'll press on.
1891                                  * We're not using EDNS0, so set the NOEDNS0
1892                                  * bit.
1893                                  */
1894                                 query->options |= DNS_FETCHOPT_NOEDNS0;
1895                         }
1896                 } else {
1897                         /*
1898                          * We know this server doesn't like EDNS0, so we
1899                          * won't use it.  Set the NOEDNS0 bit since we're
1900                          * not using EDNS0.
1901                          */
1902                         query->options |= DNS_FETCHOPT_NOEDNS0;
1903                 }
1904         }
1905
1906         /*
1907          * If we need EDNS0 to do this query and aren't using it, we lose.
1908          */
1909         if (NEEDEDNS0(fctx) && (query->options & DNS_FETCHOPT_NOEDNS0) != 0) {
1910                 result = DNS_R_SERVFAIL;
1911                 goto cleanup_message;
1912         }
1913
1914         if ((query->options & DNS_FETCHOPT_NOEDNS0) == 0)
1915                 add_triededns(fctx, &query->addrinfo->sockaddr);
1916
1917         if ((query->options & DNS_FETCHOPT_EDNS512) != 0)
1918                 add_triededns512(fctx, &query->addrinfo->sockaddr);
1919
1920         /*
1921          * Clear CD if EDNS is not in use.
1922          */
1923         if ((query->options & DNS_FETCHOPT_NOEDNS0) != 0)
1924                 fctx->qmessage->flags &= ~DNS_MESSAGEFLAG_CD;
1925
1926         /*
1927          * Add TSIG record tailored to the current recipient.
1928          */
1929         result = dns_view_getpeertsig(fctx->res->view, &ipaddr, &tsigkey);
1930         if (result != ISC_R_SUCCESS && result != ISC_R_NOTFOUND)
1931                 goto cleanup_message;
1932
1933         if (tsigkey != NULL) {
1934                 result = dns_message_settsigkey(fctx->qmessage, tsigkey);
1935                 dns_tsigkey_detach(&tsigkey);
1936                 if (result != ISC_R_SUCCESS)
1937                         goto cleanup_message;
1938         }
1939
1940         result = dns_message_rendersection(fctx->qmessage,
1941                                            DNS_SECTION_ADDITIONAL, 0);
1942         if (result != ISC_R_SUCCESS)
1943                 goto cleanup_message;
1944
1945         result = dns_message_renderend(fctx->qmessage);
1946         if (result != ISC_R_SUCCESS)
1947                 goto cleanup_message;
1948
1949         dns_compress_invalidate(&cctx);
1950         cleanup_cctx = ISC_FALSE;
1951
1952         if (dns_message_gettsigkey(fctx->qmessage) != NULL) {
1953                 dns_tsigkey_attach(dns_message_gettsigkey(fctx->qmessage),
1954                                    &query->tsigkey);
1955                 result = dns_message_getquerytsig(fctx->qmessage,
1956                                                   fctx->res->mctx,
1957                                                   &query->tsig);
1958                 if (result != ISC_R_SUCCESS)
1959                         goto cleanup_message;
1960         }
1961
1962         /*
1963          * If using TCP, write the length of the message at the beginning
1964          * of the buffer.
1965          */
1966         if ((query->options & DNS_FETCHOPT_TCP) != 0) {
1967                 isc_buffer_usedregion(&query->buffer, &r);
1968                 isc_buffer_putuint16(&tcpbuffer, (isc_uint16_t)r.length);
1969                 isc_buffer_add(&tcpbuffer, r.length);
1970         }
1971
1972         /*
1973          * We're now done with the query message.
1974          */
1975         dns_message_reset(fctx->qmessage, DNS_MESSAGE_INTENTRENDER);
1976
1977         if (query->exclusivesocket)
1978                 socket = dns_dispatch_getentrysocket(query->dispentry);
1979         else
1980                 socket = dns_dispatch_getsocket(query->dispatch);
1981         /*
1982          * Send the query!
1983          */
1984         if ((query->options & DNS_FETCHOPT_TCP) == 0) {
1985                 address = &query->addrinfo->sockaddr;
1986                 if (query->exclusivesocket) {
1987                         result = isc_socket_connect(socket, address, task,
1988                                                     resquery_udpconnected,
1989                                                     query);
1990                         if (result != ISC_R_SUCCESS)
1991                                 goto cleanup_message;
1992                         connecting = ISC_TRUE;
1993                         query->connects++;
1994                 }
1995         }
1996         isc_buffer_usedregion(buffer, &r);
1997
1998         /*
1999          * XXXRTH  Make sure we don't send to ourselves!  We should probably
2000          *              prune out these addresses when we get them from the ADB.
2001          */
2002         ISC_EVENT_INIT(&query->sendevent, sizeof(query->sendevent), 0, NULL,
2003                        ISC_SOCKEVENT_SENDDONE, resquery_senddone, query,
2004                        NULL, NULL, NULL);
2005         result = isc_socket_sendto2(socket, &r, task, address, NULL,
2006                                     &query->sendevent, 0);
2007         if (result != ISC_R_SUCCESS) {
2008                 if (connecting) {
2009                         /*
2010                          * This query is still connecting.
2011                          * Mark it as canceled so that it will just be
2012                          * cleaned up when the connected event is received.
2013                          * Keep fctx around until the event is processed.
2014                          */
2015                         query->fctx->nqueries++;
2016                         query->attributes |= RESQUERY_ATTR_CANCELED;
2017                 }
2018                 goto cleanup_message;
2019         }
2020
2021         query->sends++;
2022
2023         QTRACE("sent");
2024
2025         return (ISC_R_SUCCESS);
2026
2027  cleanup_message:
2028         if (cleanup_cctx)
2029                 dns_compress_invalidate(&cctx);
2030
2031         dns_message_reset(fctx->qmessage, DNS_MESSAGE_INTENTRENDER);
2032
2033         /*
2034          * Stop the dispatcher from listening.
2035          */
2036         dns_dispatch_removeresponse(&query->dispentry, NULL);
2037
2038  cleanup_temps:
2039         if (qname != NULL)
2040                 dns_message_puttempname(fctx->qmessage, &qname);
2041         if (qrdataset != NULL)
2042                 dns_message_puttemprdataset(fctx->qmessage, &qrdataset);
2043
2044         return (result);
2045 }
2046
2047 static void
2048 resquery_connected(isc_task_t *task, isc_event_t *event) {
2049         isc_socketevent_t *sevent = (isc_socketevent_t *)event;
2050         resquery_t *query = event->ev_arg;
2051         isc_boolean_t retry = ISC_FALSE;
2052         isc_interval_t interval;
2053         isc_result_t result;
2054         unsigned int attrs;
2055         fetchctx_t *fctx;
2056
2057         REQUIRE(event->ev_type == ISC_SOCKEVENT_CONNECT);
2058         REQUIRE(VALID_QUERY(query));
2059
2060         QTRACE("connected");
2061
2062         UNUSED(task);
2063
2064         /*
2065          * XXXRTH
2066          *
2067          * Currently we don't wait for the connect event before retrying
2068          * a query.  This means that if we get really behind, we may end
2069          * up doing extra work!
2070          */
2071
2072         query->connects--;
2073         fctx = query->fctx;
2074
2075         if (RESQUERY_CANCELED(query)) {
2076                 /*
2077                  * This query was canceled while the connect() was in
2078                  * progress.
2079                  */
2080                 isc_socket_detach(&query->tcpsocket);
2081                 resquery_destroy(&query);
2082         } else {
2083                 switch (sevent->result) {
2084                 case ISC_R_SUCCESS:
2085
2086                         /*
2087                          * Extend the idle timer for TCP.  20 seconds
2088                          * should be long enough for a TCP connection to be
2089                          * established, a single DNS request to be sent,
2090                          * and the response received.
2091                          */
2092                         isc_interval_set(&interval, 20, 0);
2093                         result = fctx_startidletimer(query->fctx, &interval);
2094                         if (result != ISC_R_SUCCESS) {
2095                                 fctx_cancelquery(&query, NULL, NULL, ISC_FALSE);
2096                                 fctx_done(fctx, result, __LINE__);
2097                                 break;
2098                         }
2099                         /*
2100                          * We are connected.  Create a dispatcher and
2101                          * send the query.
2102                          */
2103                         attrs = 0;
2104                         attrs |= DNS_DISPATCHATTR_TCP;
2105                         attrs |= DNS_DISPATCHATTR_PRIVATE;
2106                         attrs |= DNS_DISPATCHATTR_CONNECTED;
2107                         if (isc_sockaddr_pf(&query->addrinfo->sockaddr) ==
2108                             AF_INET)
2109                                 attrs |= DNS_DISPATCHATTR_IPV4;
2110                         else
2111                                 attrs |= DNS_DISPATCHATTR_IPV6;
2112                         attrs |= DNS_DISPATCHATTR_MAKEQUERY;
2113
2114                         result = dns_dispatch_createtcp(query->dispatchmgr,
2115                                                      query->tcpsocket,
2116                                                      query->fctx->res->taskmgr,
2117                                                      4096, 2, 1, 1, 3, attrs,
2118                                                      &query->dispatch);
2119
2120                         /*
2121                          * Regardless of whether dns_dispatch_create()
2122                          * succeeded or not, we don't need our reference
2123                          * to the socket anymore.
2124                          */
2125                         isc_socket_detach(&query->tcpsocket);
2126
2127                         if (result == ISC_R_SUCCESS)
2128                                 result = resquery_send(query);
2129
2130                         if (result != ISC_R_SUCCESS) {
2131                                 fctx_cancelquery(&query, NULL, NULL, ISC_FALSE);
2132                                 fctx_done(fctx, result, __LINE__);
2133                         }
2134                         break;
2135
2136                 case ISC_R_NETUNREACH:
2137                 case ISC_R_HOSTUNREACH:
2138                 case ISC_R_CONNREFUSED:
2139                 case ISC_R_NOPERM:
2140                 case ISC_R_ADDRNOTAVAIL:
2141                 case ISC_R_CONNECTIONRESET:
2142                         /*
2143                          * No route to remote.
2144                          */
2145                         isc_socket_detach(&query->tcpsocket);
2146                         fctx_cancelquery(&query, NULL, NULL, ISC_TRUE);
2147                         retry = ISC_TRUE;
2148                         break;
2149
2150                 default:
2151                         isc_socket_detach(&query->tcpsocket);
2152                         fctx_cancelquery(&query, NULL, NULL, ISC_FALSE);
2153                         break;
2154                 }
2155         }
2156
2157         isc_event_free(&event);
2158
2159         if (retry) {
2160                 /*
2161                  * Behave as if the idle timer has expired.  For TCP
2162                  * connections this may not actually reflect the latest timer.
2163                  */
2164                 fctx->attributes &= ~FCTX_ATTR_ADDRWAIT;
2165                 result = fctx_stopidletimer(fctx);
2166                 if (result != ISC_R_SUCCESS)
2167                         fctx_done(fctx, result, __LINE__);
2168                 else
2169                         fctx_try(fctx, ISC_TRUE, ISC_FALSE);
2170         }
2171 }
2172
2173 static void
2174 fctx_finddone(isc_task_t *task, isc_event_t *event) {
2175         fetchctx_t *fctx;
2176         dns_adbfind_t *find;
2177         dns_resolver_t *res;
2178         isc_boolean_t want_try = ISC_FALSE;
2179         isc_boolean_t want_done = ISC_FALSE;
2180         isc_boolean_t bucket_empty = ISC_FALSE;
2181         unsigned int bucketnum;
2182         isc_boolean_t destroy = ISC_FALSE;
2183
2184         find = event->ev_sender;
2185         fctx = event->ev_arg;
2186         REQUIRE(VALID_FCTX(fctx));
2187         res = fctx->res;
2188
2189         UNUSED(task);
2190
2191         FCTXTRACE("finddone");
2192
2193         bucketnum = fctx->bucketnum;
2194         LOCK(&res->buckets[bucketnum].lock);
2195
2196         INSIST(fctx->pending > 0);
2197         fctx->pending--;
2198
2199         if (ADDRWAIT(fctx)) {
2200                 /*
2201                  * The fetch is waiting for a name to be found.
2202                  */
2203                 INSIST(!SHUTTINGDOWN(fctx));
2204                 fctx->attributes &= ~FCTX_ATTR_ADDRWAIT;
2205                 if (event->ev_type == DNS_EVENT_ADBMOREADDRESSES) {
2206                         want_try = ISC_TRUE;
2207                         fctx->totalqueries += find->qtotal;
2208                 } else {
2209                         fctx->findfail++;
2210                         if (fctx->pending == 0) {
2211                                 /*
2212                                  * We've got nothing else to wait for and don't
2213                                  * know the answer.  There's nothing to do but
2214                                  * fail the fctx.
2215                                  */
2216                                 want_done = ISC_TRUE;
2217                         }
2218                 }
2219         } else if (SHUTTINGDOWN(fctx) && fctx->pending == 0 &&
2220                    fctx->nqueries == 0 && ISC_LIST_EMPTY(fctx->validators)) {
2221
2222                 if (fctx->references == 0) {
2223                         bucket_empty = fctx_unlink(fctx);
2224                         destroy = ISC_TRUE;
2225                 }
2226         }
2227         UNLOCK(&res->buckets[bucketnum].lock);
2228
2229         isc_event_free(&event);
2230         dns_adb_destroyfind(&find);
2231
2232         if (want_try)
2233                 fctx_try(fctx, ISC_TRUE, ISC_FALSE);
2234         else if (want_done)
2235                 fctx_done(fctx, ISC_R_FAILURE, __LINE__);
2236         else if (destroy) {
2237                         fctx_destroy(fctx);
2238                 if (bucket_empty)
2239                         empty_bucket(res);
2240         }
2241 }
2242
2243
2244 static inline isc_boolean_t
2245 bad_server(fetchctx_t *fctx, isc_sockaddr_t *address) {
2246         isc_sockaddr_t *sa;
2247
2248         for (sa = ISC_LIST_HEAD(fctx->bad);
2249              sa != NULL;
2250              sa = ISC_LIST_NEXT(sa, link)) {
2251                 if (isc_sockaddr_equal(sa, address))
2252                         return (ISC_TRUE);
2253         }
2254
2255         return (ISC_FALSE);
2256 }
2257
2258 static inline isc_boolean_t
2259 mark_bad(fetchctx_t *fctx) {
2260         dns_adbfind_t *curr;
2261         dns_adbaddrinfo_t *addrinfo;
2262         isc_boolean_t all_bad = ISC_TRUE;
2263
2264         /*
2265          * Mark all known bad servers, so we don't try to talk to them
2266          * again.
2267          */
2268
2269         /*
2270          * Mark any bad nameservers.
2271          */
2272         for (curr = ISC_LIST_HEAD(fctx->finds);
2273              curr != NULL;
2274              curr = ISC_LIST_NEXT(curr, publink)) {
2275                 for (addrinfo = ISC_LIST_HEAD(curr->list);
2276                      addrinfo != NULL;
2277                      addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
2278                         if (bad_server(fctx, &addrinfo->sockaddr))
2279                                 addrinfo->flags |= FCTX_ADDRINFO_MARK;
2280                         else
2281                                 all_bad = ISC_FALSE;
2282                 }
2283         }
2284
2285         /*
2286          * Mark any bad forwarders.
2287          */
2288         for (addrinfo = ISC_LIST_HEAD(fctx->forwaddrs);
2289              addrinfo != NULL;
2290              addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
2291                 if (bad_server(fctx, &addrinfo->sockaddr))
2292                         addrinfo->flags |= FCTX_ADDRINFO_MARK;
2293                 else
2294                         all_bad = ISC_FALSE;
2295         }
2296
2297         /*
2298          * Mark any bad alternates.
2299          */
2300         for (curr = ISC_LIST_HEAD(fctx->altfinds);
2301              curr != NULL;
2302              curr = ISC_LIST_NEXT(curr, publink)) {
2303                 for (addrinfo = ISC_LIST_HEAD(curr->list);
2304                      addrinfo != NULL;
2305                      addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
2306                         if (bad_server(fctx, &addrinfo->sockaddr))
2307                                 addrinfo->flags |= FCTX_ADDRINFO_MARK;
2308                         else
2309                                 all_bad = ISC_FALSE;
2310                 }
2311         }
2312
2313         for (addrinfo = ISC_LIST_HEAD(fctx->altaddrs);
2314              addrinfo != NULL;
2315              addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
2316                 if (bad_server(fctx, &addrinfo->sockaddr))
2317                         addrinfo->flags |= FCTX_ADDRINFO_MARK;
2318                 else
2319                         all_bad = ISC_FALSE;
2320         }
2321
2322         return (all_bad);
2323 }
2324
2325 static void
2326 add_bad(fetchctx_t *fctx, dns_adbaddrinfo_t *addrinfo, isc_result_t reason,
2327         badnstype_t badtype)
2328 {
2329         char namebuf[DNS_NAME_FORMATSIZE];
2330         char addrbuf[ISC_SOCKADDR_FORMATSIZE];
2331         char classbuf[64];
2332         char typebuf[64];
2333         char code[64];
2334         isc_buffer_t b;
2335         isc_sockaddr_t *sa;
2336         const char *spc = "";
2337         isc_sockaddr_t *address = &addrinfo->sockaddr;
2338
2339         if (reason == DNS_R_LAME)
2340                 fctx->lamecount++;
2341         else {
2342                 switch (badtype) {
2343                 case badns_unreachable:
2344                         fctx->neterr++;
2345                         break;
2346                 case badns_response:
2347                         fctx->badresp++;
2348                         break;
2349                 case badns_validation:
2350                         break;  /* counted as 'valfail' */
2351                 }
2352         }
2353
2354         if (bad_server(fctx, address)) {
2355                 /*
2356                  * We already know this server is bad.
2357                  */
2358                 return;
2359         }
2360
2361         FCTXTRACE("add_bad");
2362
2363         sa = isc_mem_get(fctx->mctx, sizeof(*sa));
2364         if (sa == NULL)
2365                 return;
2366         *sa = *address;
2367         ISC_LIST_INITANDAPPEND(fctx->bad, sa, link);
2368
2369         if (reason == DNS_R_LAME)       /* already logged */
2370                 return;
2371
2372         if (reason == DNS_R_UNEXPECTEDRCODE &&
2373             fctx->rmessage->rcode == dns_rcode_servfail &&
2374             ISFORWARDER(addrinfo))
2375                 return;
2376
2377         if (reason == DNS_R_UNEXPECTEDRCODE) {
2378                 isc_buffer_init(&b, code, sizeof(code) - 1);
2379                 dns_rcode_totext(fctx->rmessage->rcode, &b);
2380                 code[isc_buffer_usedlength(&b)] = '\0';
2381                 spc = " ";
2382         } else if (reason == DNS_R_UNEXPECTEDOPCODE) {
2383                 isc_buffer_init(&b, code, sizeof(code) - 1);
2384                 dns_opcode_totext((dns_opcode_t)fctx->rmessage->opcode, &b);
2385                 code[isc_buffer_usedlength(&b)] = '\0';
2386                 spc = " ";
2387         } else {
2388                 code[0] = '\0';
2389         }
2390         dns_name_format(&fctx->name, namebuf, sizeof(namebuf));
2391         dns_rdatatype_format(fctx->type, typebuf, sizeof(typebuf));
2392         dns_rdataclass_format(fctx->res->rdclass, classbuf, sizeof(classbuf));
2393         isc_sockaddr_format(address, addrbuf, sizeof(addrbuf));
2394         isc_log_write(dns_lctx, DNS_LOGCATEGORY_LAME_SERVERS,
2395                       DNS_LOGMODULE_RESOLVER, ISC_LOG_INFO,
2396                       "error (%s%s%s) resolving '%s/%s/%s': %s",
2397                       dns_result_totext(reason), spc, code,
2398                       namebuf, typebuf, classbuf, addrbuf);
2399 }
2400
2401 /*
2402  * Sort addrinfo list by RTT.
2403  */
2404 static void
2405 sort_adbfind(dns_adbfind_t *find) {
2406         dns_adbaddrinfo_t *best, *curr;
2407         dns_adbaddrinfolist_t sorted;
2408
2409         /* Lame N^2 bubble sort. */
2410         ISC_LIST_INIT(sorted);
2411         while (!ISC_LIST_EMPTY(find->list)) {
2412                 best = ISC_LIST_HEAD(find->list);
2413                 curr = ISC_LIST_NEXT(best, publink);
2414                 while (curr != NULL) {
2415                         if (curr->srtt < best->srtt)
2416                                 best = curr;
2417                         curr = ISC_LIST_NEXT(curr, publink);
2418                 }
2419                 ISC_LIST_UNLINK(find->list, best, publink);
2420                 ISC_LIST_APPEND(sorted, best, publink);
2421         }
2422         find->list = sorted;
2423 }
2424
2425 /*
2426  * Sort a list of finds by server RTT.
2427  */
2428 static void
2429 sort_finds(dns_adbfindlist_t *findlist) {
2430         dns_adbfind_t *best, *curr;
2431         dns_adbfindlist_t sorted;
2432         dns_adbaddrinfo_t *addrinfo, *bestaddrinfo;
2433
2434         /* Sort each find's addrinfo list by SRTT. */
2435         for (curr = ISC_LIST_HEAD(*findlist);
2436              curr != NULL;
2437              curr = ISC_LIST_NEXT(curr, publink))
2438                 sort_adbfind(curr);
2439
2440         /* Lame N^2 bubble sort. */
2441         ISC_LIST_INIT(sorted);
2442         while (!ISC_LIST_EMPTY(*findlist)) {
2443                 best = ISC_LIST_HEAD(*findlist);
2444                 bestaddrinfo = ISC_LIST_HEAD(best->list);
2445                 INSIST(bestaddrinfo != NULL);
2446                 curr = ISC_LIST_NEXT(best, publink);
2447                 while (curr != NULL) {
2448                         addrinfo = ISC_LIST_HEAD(curr->list);
2449                         INSIST(addrinfo != NULL);
2450                         if (addrinfo->srtt < bestaddrinfo->srtt) {
2451                                 best = curr;
2452                                 bestaddrinfo = addrinfo;
2453                         }
2454                         curr = ISC_LIST_NEXT(curr, publink);
2455                 }
2456                 ISC_LIST_UNLINK(*findlist, best, publink);
2457                 ISC_LIST_APPEND(sorted, best, publink);
2458         }
2459         *findlist = sorted;
2460 }
2461
2462 static void
2463 findname(fetchctx_t *fctx, dns_name_t *name, in_port_t port,
2464          unsigned int options, unsigned int flags, isc_stdtime_t now,
2465          isc_boolean_t *need_alternate)
2466 {
2467         dns_adbaddrinfo_t *ai;
2468         dns_adbfind_t *find;
2469         dns_resolver_t *res;
2470         isc_boolean_t unshared;
2471         isc_result_t result;
2472
2473         res = fctx->res;
2474         unshared = ISC_TF((fctx->options & DNS_FETCHOPT_UNSHARED) != 0);
2475         /*
2476          * If this name is a subdomain of the query domain, tell
2477          * the ADB to start looking using zone/hint data. This keeps us
2478          * from getting stuck if the nameserver is beneath the zone cut
2479          * and we don't know its address (e.g. because the A record has
2480          * expired).
2481          */
2482         if (dns_name_issubdomain(name, &fctx->domain))
2483                 options |= DNS_ADBFIND_STARTATZONE;
2484         options |= DNS_ADBFIND_GLUEOK;
2485         options |= DNS_ADBFIND_HINTOK;
2486
2487         /*
2488          * See what we know about this address.
2489          */
2490         find = NULL;
2491         result = dns_adb_createfind2(fctx->adb,
2492                                      res->buckets[fctx->bucketnum].task,
2493                                      fctx_finddone, fctx, name,
2494                                      &fctx->name, fctx->type,
2495                                      options, now, NULL,
2496                                      res->view->dstport,
2497                                      fctx->depth + 1, &find);
2498         if (result != ISC_R_SUCCESS) {
2499                 if (result == DNS_R_ALIAS) {
2500                         /*
2501                          * XXXRTH  Follow the CNAME/DNAME chain?
2502                          */
2503                         dns_adb_destroyfind(&find);
2504                         fctx->adberr++;
2505                 }
2506         } else if (!ISC_LIST_EMPTY(find->list)) {
2507                 /*
2508                  * We have at least some of the addresses for the
2509                  * name.
2510                  */
2511                 INSIST((find->options & DNS_ADBFIND_WANTEVENT) == 0);
2512                 if (flags != 0 || port != 0) {
2513                         for (ai = ISC_LIST_HEAD(find->list);
2514                              ai != NULL;
2515                              ai = ISC_LIST_NEXT(ai, publink)) {
2516                                 ai->flags |= flags;
2517                                 if (port != 0)
2518                                         isc_sockaddr_setport(&ai->sockaddr,
2519                                                              port);
2520                         }
2521                 }
2522                 if ((flags & FCTX_ADDRINFO_FORWARDER) != 0)
2523                         ISC_LIST_APPEND(fctx->altfinds, find, publink);
2524                 else
2525                         ISC_LIST_APPEND(fctx->finds, find, publink);
2526         } else {
2527                 /*
2528                  * We don't know any of the addresses for this
2529                  * name.
2530                  */
2531                 if ((find->options & DNS_ADBFIND_WANTEVENT) != 0) {
2532                         /*
2533                          * We're looking for them and will get an
2534                          * event about it later.
2535                          */
2536                         fctx->pending++;
2537                         /*
2538                          * Bootstrap.
2539                          */
2540                         if (need_alternate != NULL &&
2541                             !*need_alternate && unshared &&
2542                             ((res->dispatches4 == NULL &&
2543                               find->result_v6 != DNS_R_NXDOMAIN) ||
2544                              (res->dispatches6 == NULL &&
2545                               find->result_v4 != DNS_R_NXDOMAIN)))
2546                                 *need_alternate = ISC_TRUE;
2547                 } else {
2548                         if ((find->options & DNS_ADBFIND_LAMEPRUNED) != 0)
2549                                 fctx->lamecount++; /* cached lame server */
2550                         else
2551                                 fctx->adberr++; /* unreachable server, etc. */
2552
2553                         /*
2554                          * If we know there are no addresses for
2555                          * the family we are using then try to add
2556                          * an alternative server.
2557                          */
2558                         if (need_alternate != NULL && !*need_alternate &&
2559                             ((res->dispatches4 == NULL &&
2560                               find->result_v6 == DNS_R_NXRRSET) ||
2561                              (res->dispatches6 == NULL &&
2562                               find->result_v4 == DNS_R_NXRRSET)))
2563                                 *need_alternate = ISC_TRUE;
2564                         dns_adb_destroyfind(&find);
2565                 }
2566         }
2567 }
2568
2569 static isc_boolean_t
2570 isstrictsubdomain(dns_name_t *name1, dns_name_t *name2) {
2571         int order;
2572         unsigned int nlabels;
2573         dns_namereln_t namereln;
2574
2575         namereln = dns_name_fullcompare(name1, name2, &order, &nlabels);
2576         return (ISC_TF(namereln == dns_namereln_subdomain));
2577 }
2578
2579 static isc_result_t
2580 fctx_getaddresses(fetchctx_t *fctx, isc_boolean_t badcache) {
2581         dns_rdata_t rdata = DNS_RDATA_INIT;
2582         isc_result_t result;
2583         dns_resolver_t *res;
2584         isc_stdtime_t now;
2585         unsigned int stdoptions = 0;
2586         isc_sockaddr_t *sa;
2587         dns_adbaddrinfo_t *ai;
2588         isc_boolean_t all_bad;
2589         dns_rdata_ns_t ns;
2590         isc_boolean_t need_alternate = ISC_FALSE;
2591
2592         FCTXTRACE("getaddresses");
2593
2594         /*
2595          * Don't pound on remote servers.  (Failsafe!)
2596          */
2597         fctx->restarts++;
2598         if (fctx->restarts > 10) {
2599                 FCTXTRACE("too many restarts");
2600                 return (DNS_R_SERVFAIL);
2601         }
2602
2603         res = fctx->res;
2604
2605         if (fctx->depth > res->maxdepth) {
2606                 FCTXTRACE("too much NS indirection");
2607                 return (DNS_R_SERVFAIL);
2608         }
2609
2610         /*
2611          * Forwarders.
2612          */
2613
2614         INSIST(ISC_LIST_EMPTY(fctx->forwaddrs));
2615         INSIST(ISC_LIST_EMPTY(fctx->altaddrs));
2616
2617         /*
2618          * If this fctx has forwarders, use them; otherwise use any
2619          * selective forwarders specified in the view; otherwise use the
2620          * resolver's forwarders (if any).
2621          */
2622         sa = ISC_LIST_HEAD(fctx->forwarders);
2623         if (sa == NULL) {
2624                 dns_forwarders_t *forwarders = NULL;
2625                 dns_name_t *name = &fctx->name;
2626                 dns_name_t suffix;
2627                 unsigned int labels;
2628                 dns_fixedname_t fixed;
2629                 dns_name_t *domain;
2630
2631                 /*
2632                  * DS records are found in the parent server.
2633                  * Strip label to get the correct forwarder (if any).
2634                  */
2635                 if (dns_rdatatype_atparent(fctx->type) &&
2636                     dns_name_countlabels(name) > 1) {
2637                         dns_name_init(&suffix, NULL);
2638                         labels = dns_name_countlabels(name);
2639                         dns_name_getlabelsequence(name, 1, labels - 1, &suffix);
2640                         name = &suffix;
2641                 }
2642
2643                 dns_fixedname_init(&fixed);
2644                 domain = dns_fixedname_name(&fixed);
2645                 result = dns_fwdtable_find2(fctx->res->view->fwdtable, name,
2646                                             domain, &forwarders);
2647                 if (result == ISC_R_SUCCESS) {
2648                         sa = ISC_LIST_HEAD(forwarders->addrs);
2649                         fctx->fwdpolicy = forwarders->fwdpolicy;
2650                         if (fctx->fwdpolicy == dns_fwdpolicy_only &&
2651                             isstrictsubdomain(domain, &fctx->domain)) {
2652                                 dns_name_free(&fctx->domain, fctx->mctx);
2653                                 dns_name_init(&fctx->domain, NULL);
2654                                 result = dns_name_dup(domain, fctx->mctx,
2655                                                       &fctx->domain);
2656                                 if (result != ISC_R_SUCCESS)
2657                                         return (result);
2658                         }
2659                 }
2660         }
2661
2662         while (sa != NULL) {
2663                 if ((isc_sockaddr_pf(sa) == AF_INET &&
2664                          fctx->res->dispatches4 == NULL) ||
2665                     (isc_sockaddr_pf(sa) == AF_INET6 &&
2666                         fctx->res->dispatches6 == NULL)) {
2667                                 sa = ISC_LIST_NEXT(sa, link);
2668                                 continue;
2669                 }
2670                 ai = NULL;
2671                 result = dns_adb_findaddrinfo(fctx->adb,
2672                                               sa, &ai, 0);  /* XXXMLG */
2673                 if (result == ISC_R_SUCCESS) {
2674                         dns_adbaddrinfo_t *cur;
2675                         ai->flags |= FCTX_ADDRINFO_FORWARDER;
2676                         cur = ISC_LIST_HEAD(fctx->forwaddrs);
2677                         while (cur != NULL && cur->srtt < ai->srtt)
2678                                 cur = ISC_LIST_NEXT(cur, publink);
2679                         if (cur != NULL)
2680                                 ISC_LIST_INSERTBEFORE(fctx->forwaddrs, cur,
2681                                                       ai, publink);
2682                         else
2683                                 ISC_LIST_APPEND(fctx->forwaddrs, ai, publink);
2684                 }
2685                 sa = ISC_LIST_NEXT(sa, link);
2686         }
2687
2688         /*
2689          * If the forwarding policy is "only", we don't need the addresses
2690          * of the nameservers.
2691          */
2692         if (fctx->fwdpolicy == dns_fwdpolicy_only)
2693                 goto out;
2694
2695         /*
2696          * Normal nameservers.
2697          */
2698
2699         stdoptions = DNS_ADBFIND_WANTEVENT | DNS_ADBFIND_EMPTYEVENT;
2700         if (fctx->restarts == 1) {
2701                 /*
2702                  * To avoid sending out a flood of queries likely to
2703                  * result in NXRRSET, we suppress fetches for address
2704                  * families we don't have the first time through,
2705                  * provided that we have addresses in some family we
2706                  * can use.
2707                  *
2708                  * We don't want to set this option all the time, since
2709                  * if fctx->restarts > 1, we've clearly been having trouble
2710                  * with the addresses we had, so getting more could help.
2711                  */
2712                 stdoptions |= DNS_ADBFIND_AVOIDFETCHES;
2713         }
2714         if (res->dispatches4 != NULL)
2715                 stdoptions |= DNS_ADBFIND_INET;
2716         if (res->dispatches6 != NULL)
2717                 stdoptions |= DNS_ADBFIND_INET6;
2718         isc_stdtime_get(&now);
2719
2720         INSIST(ISC_LIST_EMPTY(fctx->finds));
2721         INSIST(ISC_LIST_EMPTY(fctx->altfinds));
2722
2723         for (result = dns_rdataset_first(&fctx->nameservers);
2724              result == ISC_R_SUCCESS;
2725              result = dns_rdataset_next(&fctx->nameservers))
2726         {
2727                 dns_rdataset_current(&fctx->nameservers, &rdata);
2728                 /*
2729                  * Extract the name from the NS record.
2730                  */
2731                 result = dns_rdata_tostruct(&rdata, &ns, NULL);
2732                 if (result != ISC_R_SUCCESS)
2733                         continue;
2734
2735                 findname(fctx, &ns.name, 0, stdoptions, 0, now,
2736                          &need_alternate);
2737                 dns_rdata_reset(&rdata);
2738                 dns_rdata_freestruct(&ns);
2739         }
2740         if (result != ISC_R_NOMORE)
2741                 return (result);
2742
2743         /*
2744          * Do we need to use 6 to 4?
2745          */
2746         if (need_alternate) {
2747                 int family;
2748                 alternate_t *a;
2749                 family = (res->dispatches6 != NULL) ? AF_INET6 : AF_INET;
2750                 for (a = ISC_LIST_HEAD(fctx->res->alternates);
2751                      a != NULL;
2752                      a = ISC_LIST_NEXT(a, link)) {
2753                         if (!a->isaddress) {
2754                                 findname(fctx, &a->_u._n.name, a->_u._n.port,
2755                                          stdoptions, FCTX_ADDRINFO_FORWARDER,
2756                                          now, NULL);
2757                                 continue;
2758                         }
2759                         if (isc_sockaddr_pf(&a->_u.addr) != family)
2760                                 continue;
2761                         ai = NULL;
2762                         result = dns_adb_findaddrinfo(fctx->adb, &a->_u.addr,
2763                                                       &ai, 0);
2764                         if (result == ISC_R_SUCCESS) {
2765                                 dns_adbaddrinfo_t *cur;
2766                                 ai->flags |= FCTX_ADDRINFO_FORWARDER;
2767                                 cur = ISC_LIST_HEAD(fctx->altaddrs);
2768                                 while (cur != NULL && cur->srtt < ai->srtt)
2769                                         cur = ISC_LIST_NEXT(cur, publink);
2770                                 if (cur != NULL)
2771                                         ISC_LIST_INSERTBEFORE(fctx->altaddrs,
2772                                                               cur, ai, publink);
2773                                 else
2774                                         ISC_LIST_APPEND(fctx->altaddrs, ai,
2775                                                         publink);
2776                         }
2777                 }
2778         }
2779
2780  out:
2781         /*
2782          * Mark all known bad servers.
2783          */
2784         all_bad = mark_bad(fctx);
2785
2786         /*
2787          * How are we doing?
2788          */
2789         if (all_bad) {
2790                 /*
2791                  * We've got no addresses.
2792                  */
2793                 if (fctx->pending > 0) {
2794                         /*
2795                          * We're fetching the addresses, but don't have any
2796                          * yet.   Tell the caller to wait for an answer.
2797                          */
2798                         result = DNS_R_WAIT;
2799                 } else {
2800                         isc_time_t expire;
2801                         isc_interval_t i;
2802                         /*
2803                          * We've lost completely.  We don't know any
2804                          * addresses, and the ADB has told us it can't get
2805                          * them.
2806                          */
2807                         FCTXTRACE("no addresses");
2808                         isc_interval_set(&i, DNS_BADCACHE_TTL(fctx), 0);
2809                         result = isc_time_nowplusinterval(&expire, &i);
2810                         if (badcache &&
2811                             (fctx->type == dns_rdatatype_dnskey ||
2812                              fctx->type == dns_rdatatype_dlv ||
2813                              fctx->type == dns_rdatatype_ds) &&
2814                              result == ISC_R_SUCCESS)
2815                                 dns_resolver_addbadcache(fctx->res,
2816                                                          &fctx->name,
2817                                                          fctx->type, &expire);
2818                         result = ISC_R_FAILURE;
2819                 }
2820         } else {
2821                 /*
2822                  * We've found some addresses.  We might still be looking
2823                  * for more addresses.
2824                  */
2825                 sort_finds(&fctx->finds);
2826                 sort_finds(&fctx->altfinds);
2827                 result = ISC_R_SUCCESS;
2828         }
2829
2830         return (result);
2831 }
2832
2833 static inline void
2834 possibly_mark(fetchctx_t *fctx, dns_adbaddrinfo_t *addr)
2835 {
2836         isc_netaddr_t na;
2837         char buf[ISC_NETADDR_FORMATSIZE];
2838         isc_sockaddr_t *sa;
2839         isc_boolean_t aborted = ISC_FALSE;
2840         isc_boolean_t bogus;
2841         dns_acl_t *blackhole;
2842         isc_netaddr_t ipaddr;
2843         dns_peer_t *peer = NULL;
2844         dns_resolver_t *res;
2845         const char *msg = NULL;
2846
2847         sa = &addr->sockaddr;
2848
2849         res = fctx->res;
2850         isc_netaddr_fromsockaddr(&ipaddr, sa);
2851         blackhole = dns_dispatchmgr_getblackhole(res->dispatchmgr);
2852         (void) dns_peerlist_peerbyaddr(res->view->peers, &ipaddr, &peer);
2853
2854         if (blackhole != NULL) {
2855                 int match;
2856
2857                 if (dns_acl_match(&ipaddr, NULL, blackhole,
2858                                   &res->view->aclenv,
2859                                   &match, NULL) == ISC_R_SUCCESS &&
2860                     match > 0)
2861                         aborted = ISC_TRUE;
2862         }
2863
2864         if (peer != NULL &&
2865             dns_peer_getbogus(peer, &bogus) == ISC_R_SUCCESS &&
2866             bogus)
2867                 aborted = ISC_TRUE;
2868
2869         if (aborted) {
2870                 addr->flags |= FCTX_ADDRINFO_MARK;
2871                 msg = "ignoring blackholed / bogus server: ";
2872         } else if (isc_sockaddr_ismulticast(sa)) {
2873                 addr->flags |= FCTX_ADDRINFO_MARK;
2874                 msg = "ignoring multicast address: ";
2875         } else if (isc_sockaddr_isexperimental(sa)) {
2876                 addr->flags |= FCTX_ADDRINFO_MARK;
2877                 msg = "ignoring experimental address: ";
2878         } else if (sa->type.sa.sa_family != AF_INET6) {
2879                 return;
2880         } else if (IN6_IS_ADDR_V4MAPPED(&sa->type.sin6.sin6_addr)) {
2881                 addr->flags |= FCTX_ADDRINFO_MARK;
2882                 msg = "ignoring IPv6 mapped IPV4 address: ";
2883         } else if (IN6_IS_ADDR_V4COMPAT(&sa->type.sin6.sin6_addr)) {
2884                 addr->flags |= FCTX_ADDRINFO_MARK;
2885                 msg = "ignoring IPv6 compatibility IPV4 address: ";
2886         } else
2887                 return;
2888
2889         if (!isc_log_wouldlog(dns_lctx, ISC_LOG_DEBUG(3)))
2890                 return;
2891
2892         isc_netaddr_fromsockaddr(&na, sa);
2893         isc_netaddr_format(&na, buf, sizeof(buf));
2894         FCTXTRACE2(msg, buf);
2895 }
2896
2897 static inline dns_adbaddrinfo_t *
2898 fctx_nextaddress(fetchctx_t *fctx) {
2899         dns_adbfind_t *find, *start;
2900         dns_adbaddrinfo_t *addrinfo;
2901         dns_adbaddrinfo_t *faddrinfo;
2902
2903         /*
2904          * Return the next untried address, if any.
2905          */
2906
2907         /*
2908          * Find the first unmarked forwarder (if any).
2909          */
2910         for (addrinfo = ISC_LIST_HEAD(fctx->forwaddrs);
2911              addrinfo != NULL;
2912              addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
2913                 if (!UNMARKED(addrinfo))
2914                         continue;
2915                 possibly_mark(fctx, addrinfo);
2916                 if (UNMARKED(addrinfo)) {
2917                         addrinfo->flags |= FCTX_ADDRINFO_MARK;
2918                         fctx->find = NULL;
2919                         return (addrinfo);
2920                 }
2921         }
2922
2923         /*
2924          * No forwarders.  Move to the next find.
2925          */
2926
2927         fctx->attributes |= FCTX_ATTR_TRIEDFIND;
2928
2929         find = fctx->find;
2930         if (find == NULL)
2931                 find = ISC_LIST_HEAD(fctx->finds);
2932         else {
2933                 find = ISC_LIST_NEXT(find, publink);
2934                 if (find == NULL)
2935                         find = ISC_LIST_HEAD(fctx->finds);
2936         }
2937
2938         /*
2939          * Find the first unmarked addrinfo.
2940          */
2941         addrinfo = NULL;
2942         if (find != NULL) {
2943                 start = find;
2944                 do {
2945                         for (addrinfo = ISC_LIST_HEAD(find->list);
2946                              addrinfo != NULL;
2947                              addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
2948                                 if (!UNMARKED(addrinfo))
2949                                         continue;
2950                                 possibly_mark(fctx, addrinfo);
2951                                 if (UNMARKED(addrinfo)) {
2952                                         addrinfo->flags |= FCTX_ADDRINFO_MARK;
2953                                         break;
2954                                 }
2955                         }
2956                         if (addrinfo != NULL)
2957                                 break;
2958                         find = ISC_LIST_NEXT(find, publink);
2959                         if (find == NULL)
2960                                 find = ISC_LIST_HEAD(fctx->finds);
2961                 } while (find != start);
2962         }
2963
2964         fctx->find = find;
2965         if (addrinfo != NULL)
2966                 return (addrinfo);
2967
2968         /*
2969          * No nameservers left.  Try alternates.
2970          */
2971
2972         fctx->attributes |= FCTX_ATTR_TRIEDALT;
2973
2974         find = fctx->altfind;
2975         if (find == NULL)
2976                 find = ISC_LIST_HEAD(fctx->altfinds);
2977         else {
2978                 find = ISC_LIST_NEXT(find, publink);
2979                 if (find == NULL)
2980                         find = ISC_LIST_HEAD(fctx->altfinds);
2981         }
2982
2983         /*
2984          * Find the first unmarked addrinfo.
2985          */
2986         addrinfo = NULL;
2987         if (find != NULL) {
2988                 start = find;
2989                 do {
2990                         for (addrinfo = ISC_LIST_HEAD(find->list);
2991                              addrinfo != NULL;
2992                              addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
2993                                 if (!UNMARKED(addrinfo))
2994                                         continue;
2995                                 possibly_mark(fctx, addrinfo);
2996                                 if (UNMARKED(addrinfo)) {
2997                                         addrinfo->flags |= FCTX_ADDRINFO_MARK;
2998                                         break;
2999                                 }
3000                         }
3001                         if (addrinfo != NULL)
3002                                 break;
3003                         find = ISC_LIST_NEXT(find, publink);
3004                         if (find == NULL)
3005                                 find = ISC_LIST_HEAD(fctx->altfinds);
3006                 } while (find != start);
3007         }
3008
3009         faddrinfo = addrinfo;
3010
3011         /*
3012          * See if we have a better alternate server by address.
3013          */
3014
3015         for (addrinfo = ISC_LIST_HEAD(fctx->altaddrs);
3016              addrinfo != NULL;
3017              addrinfo = ISC_LIST_NEXT(addrinfo, publink)) {
3018                 if (!UNMARKED(addrinfo))
3019                         continue;
3020                 possibly_mark(fctx, addrinfo);
3021                 if (UNMARKED(addrinfo) &&
3022                     (faddrinfo == NULL ||
3023                      addrinfo->srtt < faddrinfo->srtt)) {
3024                         if (faddrinfo != NULL)
3025                                 faddrinfo->flags &= ~FCTX_ADDRINFO_MARK;
3026                         addrinfo->flags |= FCTX_ADDRINFO_MARK;
3027                         break;
3028                 }
3029         }
3030
3031         if (addrinfo == NULL) {
3032                 addrinfo = faddrinfo;
3033                 fctx->altfind = find;
3034         }
3035
3036         return (addrinfo);
3037 }
3038
3039 static void
3040 fctx_try(fetchctx_t *fctx, isc_boolean_t retrying, isc_boolean_t badcache) {
3041         isc_result_t result;
3042         dns_adbaddrinfo_t *addrinfo;
3043
3044         FCTXTRACE("try");
3045
3046         REQUIRE(!ADDRWAIT(fctx));
3047
3048         if (fctx->totalqueries > DEFAULT_MAX_QUERIES)
3049                 fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
3050
3051         addrinfo = fctx_nextaddress(fctx);
3052         if (addrinfo == NULL) {
3053                 /*
3054                  * We have no more addresses.  Start over.
3055                  */
3056                 fctx_cancelqueries(fctx, ISC_TRUE);
3057                 fctx_cleanupfinds(fctx);
3058                 fctx_cleanupaltfinds(fctx);
3059                 fctx_cleanupforwaddrs(fctx);
3060                 fctx_cleanupaltaddrs(fctx);
3061                 result = fctx_getaddresses(fctx, badcache);
3062                 if (result == DNS_R_WAIT) {
3063                         /*
3064                          * Sleep waiting for addresses.
3065                          */
3066                         FCTXTRACE("addrwait");
3067                         fctx->attributes |= FCTX_ATTR_ADDRWAIT;
3068                         return;
3069                 } else if (result != ISC_R_SUCCESS) {
3070                         /*
3071                          * Something bad happened.
3072                          */
3073                         fctx_done(fctx, result, __LINE__);
3074                         return;
3075                 }
3076
3077                 addrinfo = fctx_nextaddress(fctx);
3078                 /*
3079                  * While we may have addresses from the ADB, they
3080                  * might be bad ones.  In this case, return SERVFAIL.
3081                  */
3082                 if (addrinfo == NULL) {
3083                         fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
3084                         return;
3085                 }
3086         }
3087
3088         result = fctx_query(fctx, addrinfo, fctx->options);
3089         if (result != ISC_R_SUCCESS)
3090                 fctx_done(fctx, result, __LINE__);
3091         else if (retrying)
3092                 inc_stats(fctx->res, dns_resstatscounter_retry);
3093 }
3094
3095 static isc_boolean_t
3096 fctx_unlink(fetchctx_t *fctx) {
3097         dns_resolver_t *res;
3098         unsigned int bucketnum;
3099
3100         /*
3101          * Caller must be holding the bucket lock.
3102          */
3103
3104         REQUIRE(VALID_FCTX(fctx));
3105         REQUIRE(fctx->state == fetchstate_done ||
3106                 fctx->state == fetchstate_init);
3107         REQUIRE(ISC_LIST_EMPTY(fctx->events));
3108         REQUIRE(ISC_LIST_EMPTY(fctx->queries));
3109         REQUIRE(ISC_LIST_EMPTY(fctx->finds));
3110         REQUIRE(ISC_LIST_EMPTY(fctx->altfinds));
3111         REQUIRE(fctx->pending == 0);
3112         REQUIRE(fctx->references == 0);
3113         REQUIRE(ISC_LIST_EMPTY(fctx->validators));
3114
3115         FCTXTRACE("unlink");
3116
3117         res = fctx->res;
3118         bucketnum = fctx->bucketnum;
3119
3120         ISC_LIST_UNLINK(res->buckets[bucketnum].fctxs, fctx, link);
3121
3122         LOCK(&res->nlock);
3123         res->nfctx--;
3124         UNLOCK(&res->nlock);
3125
3126         if (res->buckets[bucketnum].exiting &&
3127             ISC_LIST_EMPTY(res->buckets[bucketnum].fctxs))
3128                 return (ISC_TRUE);
3129
3130         return (ISC_FALSE);
3131 }
3132
3133 static void
3134 fctx_destroy(fetchctx_t *fctx) {
3135         isc_sockaddr_t *sa, *next_sa;
3136
3137         REQUIRE(VALID_FCTX(fctx));
3138         REQUIRE(fctx->state == fetchstate_done ||
3139                 fctx->state == fetchstate_init);
3140         REQUIRE(ISC_LIST_EMPTY(fctx->events));
3141         REQUIRE(ISC_LIST_EMPTY(fctx->queries));
3142         REQUIRE(ISC_LIST_EMPTY(fctx->finds));
3143         REQUIRE(ISC_LIST_EMPTY(fctx->altfinds));
3144         REQUIRE(fctx->pending == 0);
3145         REQUIRE(fctx->references == 0);
3146         REQUIRE(ISC_LIST_EMPTY(fctx->validators));
3147         REQUIRE(!ISC_LINK_LINKED(fctx, link));
3148
3149         FCTXTRACE("destroy");
3150
3151         /*
3152          * Free bad.
3153          */
3154         for (sa = ISC_LIST_HEAD(fctx->bad);
3155              sa != NULL;
3156              sa = next_sa) {
3157                 next_sa = ISC_LIST_NEXT(sa, link);
3158                 ISC_LIST_UNLINK(fctx->bad, sa, link);
3159                 isc_mem_put(fctx->mctx, sa, sizeof(*sa));
3160         }
3161
3162         for (sa = ISC_LIST_HEAD(fctx->edns);
3163              sa != NULL;
3164              sa = next_sa) {
3165                 next_sa = ISC_LIST_NEXT(sa, link);
3166                 ISC_LIST_UNLINK(fctx->edns, sa, link);
3167                 isc_mem_put(fctx->mctx, sa, sizeof(*sa));
3168         }
3169
3170         for (sa = ISC_LIST_HEAD(fctx->edns512);
3171              sa != NULL;
3172              sa = next_sa) {
3173                 next_sa = ISC_LIST_NEXT(sa, link);
3174                 ISC_LIST_UNLINK(fctx->edns512, sa, link);
3175                 isc_mem_put(fctx->mctx, sa, sizeof(*sa));
3176         }
3177
3178         for (sa = ISC_LIST_HEAD(fctx->bad_edns);
3179              sa != NULL;
3180              sa = next_sa) {
3181                 next_sa = ISC_LIST_NEXT(sa, link);
3182                 ISC_LIST_UNLINK(fctx->bad_edns, sa, link);
3183                 isc_mem_put(fctx->mctx, sa, sizeof(*sa));
3184         }
3185
3186         isc_timer_detach(&fctx->timer);
3187         dns_message_destroy(&fctx->rmessage);
3188         dns_message_destroy(&fctx->qmessage);
3189         if (dns_name_countlabels(&fctx->domain) > 0)
3190                 dns_name_free(&fctx->domain, fctx->mctx);
3191         if (dns_rdataset_isassociated(&fctx->nameservers))
3192                 dns_rdataset_disassociate(&fctx->nameservers);
3193         dns_name_free(&fctx->name, fctx->mctx);
3194         dns_db_detach(&fctx->cache);
3195         dns_adb_detach(&fctx->adb);
3196         isc_mem_free(fctx->mctx, fctx->info);
3197         isc_mem_putanddetach(&fctx->mctx, fctx, sizeof(*fctx));
3198 }
3199
3200 /*
3201  * Fetch event handlers.
3202  */
3203
3204 static void
3205 fctx_timeout(isc_task_t *task, isc_event_t *event) {
3206         fetchctx_t *fctx = event->ev_arg;
3207         isc_timerevent_t *tevent = (isc_timerevent_t *)event;
3208         resquery_t *query;
3209
3210         REQUIRE(VALID_FCTX(fctx));
3211
3212         UNUSED(task);
3213
3214         FCTXTRACE("timeout");
3215
3216         inc_stats(fctx->res, dns_resstatscounter_querytimeout);
3217
3218         if (event->ev_type == ISC_TIMEREVENT_LIFE) {
3219                 fctx->reason = NULL;
3220                 fctx_done(fctx, ISC_R_TIMEDOUT, __LINE__);
3221         } else {
3222                 isc_result_t result;
3223
3224                 fctx->timeouts++;
3225                 fctx->timeout = ISC_TRUE;
3226                 /*
3227                  * We could cancel the running queries here, or we could let
3228                  * them keep going.  Since we normally use separate sockets for
3229                  * different queries, we adopt the former approach to reduce
3230                  * the number of open sockets: cancel the oldest query if it
3231                  * expired after the query had started (this is usually the
3232                  * case but is not always so, depending on the task schedule
3233                  * timing).
3234                  */
3235                 query = ISC_LIST_HEAD(fctx->queries);
3236                 if (query != NULL &&
3237                     isc_time_compare(&tevent->due, &query->start) >= 0) {
3238                         fctx_cancelquery(&query, NULL, NULL, ISC_TRUE);
3239                 }
3240                 fctx->attributes &= ~FCTX_ATTR_ADDRWAIT;
3241                 /*
3242                  * Our timer has triggered.  Reestablish the fctx lifetime
3243                  * timer.
3244                  */
3245                 result = fctx_starttimer(fctx);
3246                 if (result != ISC_R_SUCCESS)
3247                         fctx_done(fctx, result, __LINE__);
3248                 else
3249                         /*
3250                          * Keep trying.
3251                          */
3252                         fctx_try(fctx, ISC_TRUE, ISC_FALSE);
3253         }
3254
3255         isc_event_free(&event);
3256 }
3257
3258 static void
3259 fctx_shutdown(fetchctx_t *fctx) {
3260         isc_event_t *cevent;
3261
3262         /*
3263          * Start the shutdown process for fctx, if it isn't already underway.
3264          */
3265
3266         FCTXTRACE("shutdown");
3267
3268         /*
3269          * The caller must be holding the appropriate bucket lock.
3270          */
3271
3272         if (fctx->want_shutdown)
3273                 return;
3274
3275         fctx->want_shutdown = ISC_TRUE;
3276
3277         /*
3278          * Unless we're still initializing (in which case the
3279          * control event is still outstanding), we need to post
3280          * the control event to tell the fetch we want it to
3281          * exit.
3282          */
3283         if (fctx->state != fetchstate_init) {
3284                 cevent = &fctx->control_event;
3285                 isc_task_send(fctx->res->buckets[fctx->bucketnum].task,
3286                               &cevent);
3287         }
3288 }
3289
3290 static void
3291 fctx_doshutdown(isc_task_t *task, isc_event_t *event) {
3292         fetchctx_t *fctx = event->ev_arg;
3293         isc_boolean_t bucket_empty = ISC_FALSE;
3294         dns_resolver_t *res;
3295         unsigned int bucketnum;
3296         dns_validator_t *validator;
3297         isc_boolean_t destroy = ISC_FALSE;
3298
3299         REQUIRE(VALID_FCTX(fctx));
3300
3301         UNUSED(task);
3302
3303         res = fctx->res;
3304         bucketnum = fctx->bucketnum;
3305
3306         FCTXTRACE("doshutdown");
3307
3308         /*
3309          * An fctx that is shutting down is no longer in ADDRWAIT mode.
3310          */
3311         fctx->attributes &= ~FCTX_ATTR_ADDRWAIT;
3312
3313         /*
3314          * Cancel all pending validators.  Note that this must be done
3315          * without the bucket lock held, since that could cause deadlock.
3316          */
3317         validator = ISC_LIST_HEAD(fctx->validators);
3318         while (validator != NULL) {
3319                 dns_validator_cancel(validator);
3320                 validator = ISC_LIST_NEXT(validator, link);
3321         }
3322
3323         if (fctx->nsfetch != NULL)
3324                 dns_resolver_cancelfetch(fctx->nsfetch);
3325
3326         /*
3327          * Shut down anything that is still running on behalf of this
3328          * fetch.  To avoid deadlock with the ADB, we must do this
3329          * before we lock the bucket lock.
3330          */
3331         fctx_stopeverything(fctx, ISC_FALSE);
3332
3333         LOCK(&res->buckets[bucketnum].lock);
3334
3335         fctx->attributes |= FCTX_ATTR_SHUTTINGDOWN;
3336
3337         INSIST(fctx->state == fetchstate_active ||
3338                fctx->state == fetchstate_done);
3339         INSIST(fctx->want_shutdown);
3340
3341         if (fctx->state != fetchstate_done) {
3342                 fctx->state = fetchstate_done;
3343                 fctx_sendevents(fctx, ISC_R_CANCELED, __LINE__);
3344         }
3345
3346         if (fctx->references == 0 && fctx->pending == 0 &&
3347             fctx->nqueries == 0 && ISC_LIST_EMPTY(fctx->validators)) {
3348                 bucket_empty = fctx_unlink(fctx);
3349                 destroy = ISC_TRUE;
3350         }
3351
3352         UNLOCK(&res->buckets[bucketnum].lock);
3353
3354         if (destroy) {
3355                 fctx_destroy(fctx);
3356                 if (bucket_empty)
3357                         empty_bucket(res);
3358         }
3359 }
3360
3361 static void
3362 fctx_start(isc_task_t *task, isc_event_t *event) {
3363         fetchctx_t *fctx = event->ev_arg;
3364         isc_boolean_t done = ISC_FALSE, bucket_empty = ISC_FALSE;
3365         dns_resolver_t *res;
3366         unsigned int bucketnum;
3367         isc_boolean_t destroy = ISC_FALSE;
3368
3369         REQUIRE(VALID_FCTX(fctx));
3370
3371         UNUSED(task);
3372
3373         res = fctx->res;
3374         bucketnum = fctx->bucketnum;
3375
3376         FCTXTRACE("start");
3377
3378         LOCK(&res->buckets[bucketnum].lock);
3379
3380         INSIST(fctx->state == fetchstate_init);
3381         if (fctx->want_shutdown) {
3382                 /*
3383                  * We haven't started this fctx yet, and we've been requested
3384                  * to shut it down.
3385                  */
3386                 fctx->attributes |= FCTX_ATTR_SHUTTINGDOWN;
3387                 fctx->state = fetchstate_done;
3388                 fctx_sendevents(fctx, ISC_R_CANCELED, __LINE__);
3389                 /*
3390                  * Since we haven't started, we INSIST that we have no
3391                  * pending ADB finds and no pending validations.
3392                  */
3393                 INSIST(fctx->pending == 0);
3394                 INSIST(fctx->nqueries == 0);
3395                 INSIST(ISC_LIST_EMPTY(fctx->validators));
3396                 if (fctx->references == 0) {
3397                         /*
3398                          * It's now safe to destroy this fctx.
3399                          */
3400                         bucket_empty = fctx_unlink(fctx);
3401                         destroy = ISC_TRUE;
3402                 }
3403                 done = ISC_TRUE;
3404         } else {
3405                 /*
3406                  * Normal fctx startup.
3407                  */
3408                 fctx->state = fetchstate_active;
3409                 fctx->totalqueries = 0;
3410                 /*
3411                  * Reset the control event for later use in shutting down
3412                  * the fctx.
3413                  */
3414                 ISC_EVENT_INIT(event, sizeof(*event), 0, NULL,
3415                                DNS_EVENT_FETCHCONTROL, fctx_doshutdown, fctx,
3416                                NULL, NULL, NULL);
3417         }
3418
3419         UNLOCK(&res->buckets[bucketnum].lock);
3420
3421         if (!done) {
3422                 isc_result_t result;
3423
3424                 INSIST(!destroy);
3425
3426                 /*
3427                  * All is well.  Start working on the fetch.
3428                  */
3429                 result = fctx_starttimer(fctx);
3430                 if (result != ISC_R_SUCCESS)
3431                         fctx_done(fctx, result, __LINE__);
3432                 else
3433                         fctx_try(fctx, ISC_FALSE, ISC_FALSE);
3434         } else if (destroy) {
3435                         fctx_destroy(fctx);
3436                 if (bucket_empty)
3437                         empty_bucket(res);
3438         }
3439 }
3440
3441 /*
3442  * Fetch Creation, Joining, and Cancelation.
3443  */
3444
3445 static inline isc_result_t
3446 fctx_join(fetchctx_t *fctx, isc_task_t *task, isc_sockaddr_t *client,
3447           dns_messageid_t id, isc_taskaction_t action, void *arg,
3448           dns_rdataset_t *rdataset, dns_rdataset_t *sigrdataset,
3449           dns_fetch_t *fetch)
3450 {
3451         isc_task_t *clone;
3452         dns_fetchevent_t *event;
3453
3454         FCTXTRACE("join");
3455
3456         /*
3457          * We store the task we're going to send this event to in the
3458          * sender field.  We'll make the fetch the sender when we actually
3459          * send the event.
3460          */
3461         clone = NULL;
3462         isc_task_attach(task, &clone);
3463         event = (dns_fetchevent_t *)
3464                 isc_event_allocate(fctx->res->mctx, clone, DNS_EVENT_FETCHDONE,
3465                                    action, arg, sizeof(*event));
3466         if (event == NULL) {
3467                 isc_task_detach(&clone);
3468                 return (ISC_R_NOMEMORY);
3469         }
3470         event->result = DNS_R_SERVFAIL;
3471         event->qtype = fctx->type;
3472         event->db = NULL;
3473         event->node = NULL;
3474         event->rdataset = rdataset;
3475         event->sigrdataset = sigrdataset;
3476         event->fetch = fetch;
3477         event->client = client;
3478         event->id = id;
3479         event->qtotal = 0;
3480         dns_fixedname_init(&event->foundname);
3481
3482         /*
3483          * Make sure that we can store the sigrdataset in the
3484          * first event if it is needed by any of the events.
3485          */
3486         if (event->sigrdataset != NULL)
3487                 ISC_LIST_PREPEND(fctx->events, event, ev_link);
3488         else
3489                 ISC_LIST_APPEND(fctx->events, event, ev_link);
3490         fctx->references++;
3491         fctx->client = client;
3492
3493         fetch->magic = DNS_FETCH_MAGIC;
3494         fetch->private = fctx;
3495
3496         return (ISC_R_SUCCESS);
3497 }
3498
3499 static inline void
3500 log_ns_ttl(fetchctx_t *fctx, const char *where) {
3501         char namebuf[DNS_NAME_FORMATSIZE];
3502         char domainbuf[DNS_NAME_FORMATSIZE];
3503
3504         dns_name_format(&fctx->name, namebuf, sizeof(namebuf));
3505         dns_name_format(&fctx->domain, domainbuf, sizeof(domainbuf));
3506         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
3507                       DNS_LOGMODULE_RESOLVER, ISC_LOG_DEBUG(10),
3508                       "log_ns_ttl: fctx %p: %s: %s (in '%s'?): %u %u",
3509                       fctx, where, namebuf, domainbuf,
3510                       fctx->ns_ttl_ok, fctx->ns_ttl);
3511 }
3512
3513 static isc_result_t
3514 fctx_create(dns_resolver_t *res, dns_name_t *name, dns_rdatatype_t type,
3515             dns_name_t *domain, dns_rdataset_t *nameservers,
3516             unsigned int options, unsigned int bucketnum, unsigned int depth,
3517             fetchctx_t **fctxp)
3518 {
3519         fetchctx_t *fctx;
3520         isc_result_t result;
3521         isc_result_t iresult;
3522         isc_interval_t interval;
3523         dns_fixedname_t fixed;
3524         unsigned int findoptions = 0;
3525         char buf[DNS_NAME_FORMATSIZE + DNS_RDATATYPE_FORMATSIZE];
3526         char typebuf[DNS_RDATATYPE_FORMATSIZE];
3527         dns_name_t suffix;
3528         isc_mem_t *mctx;
3529
3530         /*
3531          * Caller must be holding the lock for bucket number 'bucketnum'.
3532          */
3533         REQUIRE(fctxp != NULL && *fctxp == NULL);
3534
3535         mctx = res->buckets[bucketnum].mctx;
3536         fctx = isc_mem_get(mctx, sizeof(*fctx));
3537         if (fctx == NULL)
3538                 return (ISC_R_NOMEMORY);
3539         dns_name_format(name, buf, sizeof(buf));
3540         dns_rdatatype_format(type, typebuf, sizeof(typebuf));
3541         strcat(buf, "/");       /* checked */
3542         strcat(buf, typebuf);   /* checked */
3543         fctx->info = isc_mem_strdup(mctx, buf);
3544         if (fctx->info == NULL) {
3545                 result = ISC_R_NOMEMORY;
3546                 goto cleanup_fetch;
3547         }
3548         FCTXTRACE("create");
3549         dns_name_init(&fctx->name, NULL);
3550         result = dns_name_dup(name, mctx, &fctx->name);
3551         if (result != ISC_R_SUCCESS)
3552                 goto cleanup_info;
3553         dns_name_init(&fctx->domain, NULL);
3554         dns_rdataset_init(&fctx->nameservers);
3555
3556         fctx->type = type;
3557         fctx->options = options;
3558         /*
3559          * Note!  We do not attach to the task.  We are relying on the
3560          * resolver to ensure that this task doesn't go away while we are
3561          * using it.
3562          */
3563         fctx->res = res;
3564         fctx->references = 0;
3565         fctx->bucketnum = bucketnum;
3566         fctx->state = fetchstate_init;
3567         fctx->want_shutdown = ISC_FALSE;
3568         fctx->cloned = ISC_FALSE;
3569         fctx->depth = depth;
3570         ISC_LIST_INIT(fctx->queries);
3571         ISC_LIST_INIT(fctx->finds);
3572         ISC_LIST_INIT(fctx->altfinds);
3573         ISC_LIST_INIT(fctx->forwaddrs);
3574         ISC_LIST_INIT(fctx->altaddrs);
3575         ISC_LIST_INIT(fctx->forwarders);
3576         fctx->fwdpolicy = dns_fwdpolicy_none;
3577         ISC_LIST_INIT(fctx->bad);
3578         ISC_LIST_INIT(fctx->edns);
3579         ISC_LIST_INIT(fctx->edns512);
3580         ISC_LIST_INIT(fctx->bad_edns);
3581         ISC_LIST_INIT(fctx->validators);
3582         fctx->validator = NULL;
3583         fctx->find = NULL;
3584         fctx->altfind = NULL;
3585         fctx->pending = 0;
3586         fctx->restarts = 0;
3587         fctx->querysent = 0;
3588         fctx->totalqueries = 0;
3589         fctx->referrals = 0;
3590         TIME_NOW(&fctx->start);
3591         fctx->timeouts = 0;
3592         fctx->lamecount = 0;
3593         fctx->adberr = 0;
3594         fctx->neterr = 0;
3595         fctx->badresp = 0;
3596         fctx->findfail = 0;
3597         fctx->valfail = 0;
3598         fctx->result = ISC_R_FAILURE;
3599         fctx->vresult = ISC_R_SUCCESS;
3600         fctx->exitline = -1;    /* sentinel */
3601         fctx->logged = ISC_FALSE;
3602         fctx->attributes = 0;
3603         fctx->spilled = ISC_FALSE;
3604         fctx->nqueries = 0;
3605         fctx->reason = NULL;
3606         fctx->rand_buf = 0;
3607         fctx->rand_bits = 0;
3608         fctx->timeout = ISC_FALSE;
3609         fctx->addrinfo = NULL;
3610         fctx->client = NULL;
3611         fctx->ns_ttl = 0;
3612         fctx->ns_ttl_ok = ISC_FALSE;
3613
3614         dns_name_init(&fctx->nsname, NULL);
3615         fctx->nsfetch = NULL;
3616         dns_rdataset_init(&fctx->nsrrset);
3617
3618         if (domain == NULL) {
3619                 dns_forwarders_t *forwarders = NULL;
3620                 unsigned int labels;
3621                 dns_name_t *fwdname = name;
3622
3623                 /*
3624                  * DS records are found in the parent server.
3625                  * Strip label to get the correct forwarder (if any).
3626                  */
3627                 if (dns_rdatatype_atparent(fctx->type) &&
3628                     dns_name_countlabels(name) > 1) {
3629                         dns_name_init(&suffix, NULL);
3630                         labels = dns_name_countlabels(name);
3631                         dns_name_getlabelsequence(name, 1, labels - 1, &suffix);
3632                         fwdname = &suffix;
3633                 }
3634                 dns_fixedname_init(&fixed);
3635                 domain = dns_fixedname_name(&fixed);
3636                 result = dns_fwdtable_find2(fctx->res->view->fwdtable, fwdname,
3637                                             domain, &forwarders);
3638                 if (result == ISC_R_SUCCESS)
3639                         fctx->fwdpolicy = forwarders->fwdpolicy;
3640
3641                 if (fctx->fwdpolicy != dns_fwdpolicy_only) {
3642                         /*
3643                          * The caller didn't supply a query domain and
3644                          * nameservers, and we're not in forward-only mode,
3645                          * so find the best nameservers to use.
3646                          */
3647                         if (dns_rdatatype_atparent(fctx->type))
3648                                 findoptions |= DNS_DBFIND_NOEXACT;
3649                         result = dns_view_findzonecut(res->view, fwdname,
3650                                                       domain, 0, findoptions,
3651                                                       ISC_TRUE,
3652                                                       &fctx->nameservers,
3653                                                       NULL);
3654                         if (result != ISC_R_SUCCESS)
3655                                 goto cleanup_name;
3656
3657                         result = dns_name_dup(domain, mctx, &fctx->domain);
3658                         if (result != ISC_R_SUCCESS) {
3659                                 dns_rdataset_disassociate(&fctx->nameservers);
3660                                 goto cleanup_name;
3661                         }
3662                         fctx->ns_ttl = fctx->nameservers.ttl;
3663                         fctx->ns_ttl_ok = ISC_TRUE;
3664                 } else {
3665                         /*
3666                          * We're in forward-only mode.  Set the query domain.
3667                          */
3668                         result = dns_name_dup(domain, mctx, &fctx->domain);
3669                         if (result != ISC_R_SUCCESS)
3670                                 goto cleanup_name;
3671                 }
3672         } else {
3673                 result = dns_name_dup(domain, mctx, &fctx->domain);
3674                 if (result != ISC_R_SUCCESS)
3675                         goto cleanup_name;
3676                 dns_rdataset_clone(nameservers, &fctx->nameservers);
3677                 fctx->ns_ttl = fctx->nameservers.ttl;
3678                 fctx->ns_ttl_ok = ISC_TRUE;
3679         }
3680
3681         log_ns_ttl(fctx, "fctx_create");
3682
3683         INSIST(dns_name_issubdomain(&fctx->name, &fctx->domain));
3684
3685         fctx->qmessage = NULL;
3686         result = dns_message_create(mctx, DNS_MESSAGE_INTENTRENDER,
3687                                     &fctx->qmessage);
3688
3689         if (result != ISC_R_SUCCESS)
3690                 goto cleanup_domain;
3691
3692         fctx->rmessage = NULL;
3693         result = dns_message_create(mctx, DNS_MESSAGE_INTENTPARSE,
3694                                     &fctx->rmessage);
3695
3696         if (result != ISC_R_SUCCESS)
3697                 goto cleanup_qmessage;
3698
3699         /*
3700          * Compute an expiration time for the entire fetch.
3701          */
3702         isc_interval_set(&interval, res->query_timeout, 0);
3703         iresult = isc_time_nowplusinterval(&fctx->expires, &interval);
3704         if (iresult != ISC_R_SUCCESS) {
3705                 UNEXPECTED_ERROR(__FILE__, __LINE__,
3706                                  "isc_time_nowplusinterval: %s",
3707                                  isc_result_totext(iresult));
3708                 result = ISC_R_UNEXPECTED;
3709                 goto cleanup_rmessage;
3710         }
3711
3712         /*
3713          * Default retry interval initialization.  We set the interval now
3714          * mostly so it won't be uninitialized.  It will be set to the
3715          * correct value before a query is issued.
3716          */
3717         isc_interval_set(&fctx->interval, 2, 0);
3718
3719         /*
3720          * Create an inactive timer.  It will be made active when the fetch
3721          * is actually started.
3722          */
3723         fctx->timer = NULL;
3724         iresult = isc_timer_create(res->timermgr, isc_timertype_inactive,
3725                                    NULL, NULL,
3726                                    res->buckets[bucketnum].task, fctx_timeout,
3727                                    fctx, &fctx->timer);
3728         if (iresult != ISC_R_SUCCESS) {
3729                 UNEXPECTED_ERROR(__FILE__, __LINE__,
3730                                  "isc_timer_create: %s",
3731                                  isc_result_totext(iresult));
3732                 result = ISC_R_UNEXPECTED;
3733                 goto cleanup_rmessage;
3734         }
3735
3736         /*
3737          * Attach to the view's cache and adb.
3738          */
3739         fctx->cache = NULL;
3740         dns_db_attach(res->view->cachedb, &fctx->cache);
3741         fctx->adb = NULL;
3742         dns_adb_attach(res->view->adb, &fctx->adb);
3743         fctx->mctx = NULL;
3744         isc_mem_attach(mctx, &fctx->mctx);
3745
3746         ISC_LIST_INIT(fctx->events);
3747         ISC_LINK_INIT(fctx, link);
3748         fctx->magic = FCTX_MAGIC;
3749
3750         ISC_LIST_APPEND(res->buckets[bucketnum].fctxs, fctx, link);
3751
3752         LOCK(&res->nlock);
3753         res->nfctx++;
3754         UNLOCK(&res->nlock);
3755
3756         *fctxp = fctx;
3757
3758         return (ISC_R_SUCCESS);
3759
3760  cleanup_rmessage:
3761         dns_message_destroy(&fctx->rmessage);
3762
3763  cleanup_qmessage:
3764         dns_message_destroy(&fctx->qmessage);
3765
3766  cleanup_domain:
3767         if (dns_name_countlabels(&fctx->domain) > 0)
3768                 dns_name_free(&fctx->domain, mctx);
3769         if (dns_rdataset_isassociated(&fctx->nameservers))
3770                 dns_rdataset_disassociate(&fctx->nameservers);
3771
3772  cleanup_name:
3773         dns_name_free(&fctx->name, mctx);
3774
3775  cleanup_info:
3776         isc_mem_free(mctx, fctx->info);
3777
3778  cleanup_fetch:
3779         isc_mem_put(mctx, fctx, sizeof(*fctx));
3780
3781         return (result);
3782 }
3783
3784 /*
3785  * Handle Responses
3786  */
3787 static inline isc_boolean_t
3788 is_lame(fetchctx_t *fctx) {
3789         dns_message_t *message = fctx->rmessage;
3790         dns_name_t *name;
3791         dns_rdataset_t *rdataset;
3792         isc_result_t result;
3793
3794         if (message->rcode != dns_rcode_noerror &&
3795             message->rcode != dns_rcode_nxdomain)
3796                 return (ISC_FALSE);
3797
3798         if (message->counts[DNS_SECTION_ANSWER] != 0)
3799                 return (ISC_FALSE);
3800
3801         if (message->counts[DNS_SECTION_AUTHORITY] == 0)
3802                 return (ISC_FALSE);
3803
3804         result = dns_message_firstname(message, DNS_SECTION_AUTHORITY);
3805         while (result == ISC_R_SUCCESS) {
3806                 name = NULL;
3807                 dns_message_currentname(message, DNS_SECTION_AUTHORITY, &name);
3808                 for (rdataset = ISC_LIST_HEAD(name->list);
3809                      rdataset != NULL;
3810                      rdataset = ISC_LIST_NEXT(rdataset, link)) {
3811                         dns_namereln_t namereln;
3812                         int order;
3813                         unsigned int labels;
3814                         if (rdataset->type != dns_rdatatype_ns)
3815                                 continue;
3816                         namereln = dns_name_fullcompare(name, &fctx->domain,
3817                                                         &order, &labels);
3818                         if (namereln == dns_namereln_equal &&
3819                             (message->flags & DNS_MESSAGEFLAG_AA) != 0)
3820                                 return (ISC_FALSE);
3821                         if (namereln == dns_namereln_subdomain)
3822                                 return (ISC_FALSE);
3823                         return (ISC_TRUE);
3824                 }
3825                 result = dns_message_nextname(message, DNS_SECTION_AUTHORITY);
3826         }
3827
3828         return (ISC_FALSE);
3829 }
3830
3831 static inline void
3832 log_lame(fetchctx_t *fctx, dns_adbaddrinfo_t *addrinfo) {
3833         char namebuf[DNS_NAME_FORMATSIZE];
3834         char domainbuf[DNS_NAME_FORMATSIZE];
3835         char addrbuf[ISC_SOCKADDR_FORMATSIZE];
3836
3837         dns_name_format(&fctx->name, namebuf, sizeof(namebuf));
3838         dns_name_format(&fctx->domain, domainbuf, sizeof(domainbuf));
3839         isc_sockaddr_format(&addrinfo->sockaddr, addrbuf, sizeof(addrbuf));
3840         isc_log_write(dns_lctx, DNS_LOGCATEGORY_LAME_SERVERS,
3841                       DNS_LOGMODULE_RESOLVER, ISC_LOG_INFO,
3842                       "lame server resolving '%s' (in '%s'?): %s",
3843                       namebuf, domainbuf, addrbuf);
3844 }
3845
3846 static inline void
3847 log_formerr(fetchctx_t *fctx, const char *format, ...) {
3848         char nsbuf[ISC_SOCKADDR_FORMATSIZE];
3849         char clbuf[ISC_SOCKADDR_FORMATSIZE];
3850         const char *clmsg = "";
3851         char msgbuf[2048];
3852         va_list args;
3853
3854         va_start(args, format);
3855         vsnprintf(msgbuf, sizeof(msgbuf), format, args);
3856         va_end(args);
3857
3858         isc_sockaddr_format(&fctx->addrinfo->sockaddr, nsbuf, sizeof(nsbuf));
3859
3860         if (fctx->client != NULL) {
3861                 clmsg = " for client ";
3862                 isc_sockaddr_format(fctx->client, clbuf, sizeof(clbuf));
3863         } else {
3864                 clbuf[0] = '\0';
3865         }
3866
3867         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
3868                       DNS_LOGMODULE_RESOLVER, ISC_LOG_NOTICE,
3869                       "DNS format error from %s resolving %s%s%s: %s",
3870                       nsbuf, fctx->info, clmsg, clbuf, msgbuf);
3871 }
3872
3873 static inline isc_result_t
3874 same_question(fetchctx_t *fctx) {
3875         isc_result_t result;
3876         dns_message_t *message = fctx->rmessage;
3877         dns_name_t *name;
3878         dns_rdataset_t *rdataset;
3879
3880         /*
3881          * Caller must be holding the fctx lock.
3882          */
3883
3884         /*
3885          * XXXRTH  Currently we support only one question.
3886          */
3887         if (message->counts[DNS_SECTION_QUESTION] != 1) {
3888                 log_formerr(fctx, "too many questions");
3889                 return (DNS_R_FORMERR);
3890         }
3891
3892         result = dns_message_firstname(message, DNS_SECTION_QUESTION);
3893         if (result != ISC_R_SUCCESS)
3894                 return (result);
3895         name = NULL;
3896         dns_message_currentname(message, DNS_SECTION_QUESTION, &name);
3897         rdataset = ISC_LIST_HEAD(name->list);
3898         INSIST(rdataset != NULL);
3899         INSIST(ISC_LIST_NEXT(rdataset, link) == NULL);
3900
3901         if (fctx->type != rdataset->type ||
3902             fctx->res->rdclass != rdataset->rdclass ||
3903             !dns_name_equal(&fctx->name, name)) {
3904                 char namebuf[DNS_NAME_FORMATSIZE];
3905                 char class[DNS_RDATACLASS_FORMATSIZE];
3906                 char type[DNS_RDATATYPE_FORMATSIZE];
3907
3908                 dns_name_format(name, namebuf, sizeof(namebuf));
3909                 dns_rdataclass_format(rdataset->rdclass, class, sizeof(class));
3910                 dns_rdatatype_format(rdataset->type, type, sizeof(type));
3911                 log_formerr(fctx, "question section mismatch: got %s/%s/%s",
3912                             namebuf, class, type);
3913                 return (DNS_R_FORMERR);
3914         }
3915
3916         return (ISC_R_SUCCESS);
3917 }
3918
3919 static void
3920 clone_results(fetchctx_t *fctx) {
3921         dns_fetchevent_t *event, *hevent;
3922         isc_result_t result;
3923         dns_name_t *name, *hname;
3924
3925         FCTXTRACE("clone_results");
3926
3927         /*
3928          * Set up any other events to have the same data as the first
3929          * event.
3930          *
3931          * Caller must be holding the appropriate lock.
3932          */
3933
3934         fctx->cloned = ISC_TRUE;
3935         hevent = ISC_LIST_HEAD(fctx->events);
3936         if (hevent == NULL)
3937                 return;
3938         hname = dns_fixedname_name(&hevent->foundname);
3939         for (event = ISC_LIST_NEXT(hevent, ev_link);
3940              event != NULL;
3941              event = ISC_LIST_NEXT(event, ev_link)) {
3942                 name = dns_fixedname_name(&event->foundname);
3943                 result = dns_name_copy(hname, name, NULL);
3944                 if (result != ISC_R_SUCCESS)
3945                         event->result = result;
3946                 else
3947                         event->result = hevent->result;
3948                 dns_db_attach(hevent->db, &event->db);
3949                 dns_db_attachnode(hevent->db, hevent->node, &event->node);
3950                 INSIST(hevent->rdataset != NULL);
3951                 INSIST(event->rdataset != NULL);
3952                 if (dns_rdataset_isassociated(hevent->rdataset))
3953                         dns_rdataset_clone(hevent->rdataset, event->rdataset);
3954                 INSIST(! (hevent->sigrdataset == NULL &&
3955                           event->sigrdataset != NULL));
3956                 if (hevent->sigrdataset != NULL &&
3957                     dns_rdataset_isassociated(hevent->sigrdataset) &&
3958                     event->sigrdataset != NULL)
3959                         dns_rdataset_clone(hevent->sigrdataset,
3960                                            event->sigrdataset);
3961         }
3962 }
3963
3964 #define CACHE(r)        (((r)->attributes & DNS_RDATASETATTR_CACHE) != 0)
3965 #define ANSWER(r)       (((r)->attributes & DNS_RDATASETATTR_ANSWER) != 0)
3966 #define ANSWERSIG(r)    (((r)->attributes & DNS_RDATASETATTR_ANSWERSIG) != 0)
3967 #define EXTERNAL(r)     (((r)->attributes & DNS_RDATASETATTR_EXTERNAL) != 0)
3968 #define CHAINING(r)     (((r)->attributes & DNS_RDATASETATTR_CHAINING) != 0)
3969 #define CHASE(r)        (((r)->attributes & DNS_RDATASETATTR_CHASE) != 0)
3970 #define CHECKNAMES(r)   (((r)->attributes & DNS_RDATASETATTR_CHECKNAMES) != 0)
3971
3972
3973 /*
3974  * Destroy '*fctx' if it is ready to be destroyed (i.e., if it has
3975  * no references and is no longer waiting for any events).
3976  *
3977  * Requires:
3978  *      '*fctx' is shutting down.
3979  *
3980  * Returns:
3981  *      true if the resolver is exiting and this is the last fctx in the bucket.
3982  */
3983 static isc_boolean_t
3984 maybe_destroy(fetchctx_t *fctx, isc_boolean_t locked) {
3985         unsigned int bucketnum;
3986         isc_boolean_t bucket_empty = ISC_FALSE;
3987         dns_resolver_t *res = fctx->res;
3988         dns_validator_t *validator, *next_validator;
3989         isc_boolean_t destroy = ISC_FALSE;
3990
3991         REQUIRE(SHUTTINGDOWN(fctx));
3992
3993         bucketnum = fctx->bucketnum;
3994         if (!locked)
3995                 LOCK(&res->buckets[bucketnum].lock);
3996         if (fctx->pending != 0 || fctx->nqueries != 0)
3997                 goto unlock;
3998
3999         for (validator = ISC_LIST_HEAD(fctx->validators);
4000              validator != NULL; validator = next_validator) {
4001                 next_validator = ISC_LIST_NEXT(validator, link);
4002                 dns_validator_cancel(validator);
4003         }
4004
4005         if (fctx->references == 0 && ISC_LIST_EMPTY(fctx->validators)) {
4006                 bucket_empty = fctx_unlink(fctx);
4007                 destroy = ISC_TRUE;
4008         }
4009  unlock:
4010         if (!locked)
4011                 UNLOCK(&res->buckets[bucketnum].lock);
4012         if (destroy)
4013                 fctx_destroy(fctx);
4014         return (bucket_empty);
4015 }
4016
4017 /*
4018  * The validator has finished.
4019  */
4020 static void
4021 validated(isc_task_t *task, isc_event_t *event) {
4022         dns_adbaddrinfo_t *addrinfo;
4023         dns_dbnode_t *node = NULL;
4024         dns_dbnode_t *nsnode = NULL;
4025         dns_fetchevent_t *hevent;
4026         dns_name_t *name;
4027         dns_rdataset_t *ardataset = NULL;
4028         dns_rdataset_t *asigrdataset = NULL;
4029         dns_rdataset_t *rdataset;
4030         dns_rdataset_t *sigrdataset;
4031         dns_resolver_t *res;
4032         dns_valarg_t *valarg;
4033         dns_validatorevent_t *vevent;
4034         fetchctx_t *fctx;
4035         isc_boolean_t chaining;
4036         isc_boolean_t negative;
4037         isc_boolean_t sentresponse;
4038         isc_result_t eresult = ISC_R_SUCCESS;
4039         isc_result_t result = ISC_R_SUCCESS;
4040         isc_stdtime_t now;
4041         isc_uint32_t ttl;
4042
4043         UNUSED(task); /* for now */
4044
4045         REQUIRE(event->ev_type == DNS_EVENT_VALIDATORDONE);
4046         valarg = event->ev_arg;
4047         fctx = valarg->fctx;
4048         res = fctx->res;
4049         addrinfo = valarg->addrinfo;
4050         REQUIRE(VALID_FCTX(fctx));
4051         REQUIRE(!ISC_LIST_EMPTY(fctx->validators));
4052
4053         vevent = (dns_validatorevent_t *)event;
4054         fctx->vresult = vevent->result;
4055
4056         FCTXTRACE("received validation completion event");
4057
4058         LOCK(&res->buckets[fctx->bucketnum].lock);
4059
4060         ISC_LIST_UNLINK(fctx->validators, vevent->validator, link);
4061         fctx->validator = NULL;
4062
4063         /*
4064          * Destroy the validator early so that we can
4065          * destroy the fctx if necessary.
4066          */
4067         dns_validator_destroy(&vevent->validator);
4068         isc_mem_put(fctx->mctx, valarg, sizeof(*valarg));
4069
4070         negative = ISC_TF(vevent->rdataset == NULL);
4071
4072         sentresponse = ISC_TF((fctx->options & DNS_FETCHOPT_NOVALIDATE) != 0);
4073
4074         /*
4075          * If shutting down, ignore the results.  Check to see if we're
4076          * done waiting for validator completions and ADB pending events; if
4077          * so, destroy the fctx.
4078          */
4079         if (SHUTTINGDOWN(fctx) && !sentresponse) {
4080                 isc_uint32_t bucketnum = fctx->bucketnum;
4081                 isc_boolean_t bucket_empty;
4082                 bucket_empty = maybe_destroy(fctx, ISC_TRUE);
4083                 UNLOCK(&res->buckets[bucketnum].lock);
4084                 if (bucket_empty)
4085                         empty_bucket(res);
4086                 goto cleanup_event;
4087         }
4088
4089         isc_stdtime_get(&now);
4090
4091         /*
4092          * If chaining, we need to make sure that the right result code is
4093          * returned, and that the rdatasets are bound.
4094          */
4095         if (vevent->result == ISC_R_SUCCESS &&
4096             !negative &&
4097             vevent->rdataset != NULL &&
4098             CHAINING(vevent->rdataset))
4099         {
4100                 if (vevent->rdataset->type == dns_rdatatype_cname)
4101                         eresult = DNS_R_CNAME;
4102                 else {
4103                         INSIST(vevent->rdataset->type == dns_rdatatype_dname);
4104                         eresult = DNS_R_DNAME;
4105                 }
4106                 chaining = ISC_TRUE;
4107         } else
4108                 chaining = ISC_FALSE;
4109
4110         /*
4111          * Either we're not shutting down, or we are shutting down but want
4112          * to cache the result anyway (if this was a validation started by
4113          * a query with cd set)
4114          */
4115
4116         hevent = ISC_LIST_HEAD(fctx->events);
4117         if (hevent != NULL) {
4118                 if (!negative && !chaining &&
4119                     (fctx->type == dns_rdatatype_any ||
4120                      fctx->type == dns_rdatatype_rrsig ||
4121                      fctx->type == dns_rdatatype_sig)) {
4122                         /*
4123                          * Don't bind rdatasets; the caller
4124                          * will iterate the node.
4125                          */
4126                 } else {
4127                         ardataset = hevent->rdataset;
4128                         asigrdataset = hevent->sigrdataset;
4129                 }
4130         }
4131
4132         if (vevent->result != ISC_R_SUCCESS) {
4133                 FCTXTRACE("validation failed");
4134                 inc_stats(res, dns_resstatscounter_valfail);
4135                 fctx->valfail++;
4136                 fctx->vresult = vevent->result;
4137                 if (fctx->vresult != DNS_R_BROKENCHAIN) {
4138                         result = ISC_R_NOTFOUND;
4139                         if (vevent->rdataset != NULL)
4140                                 result = dns_db_findnode(fctx->cache,
4141                                                          vevent->name,
4142                                                          ISC_TRUE, &node);
4143                         if (result == ISC_R_SUCCESS)
4144                                 (void)dns_db_deleterdataset(fctx->cache, node,
4145                                                              NULL,
4146                                                             vevent->type, 0);
4147                         if (result == ISC_R_SUCCESS &&
4148                              vevent->sigrdataset != NULL)
4149                                 (void)dns_db_deleterdataset(fctx->cache, node,
4150                                                             NULL,
4151                                                             dns_rdatatype_rrsig,
4152                                                             vevent->type);
4153                         if (result == ISC_R_SUCCESS)
4154                                 dns_db_detachnode(fctx->cache, &node);
4155                 }
4156                 if (fctx->vresult == DNS_R_BROKENCHAIN && !negative) {
4157                         /*
4158                          * Cache the data as pending for later validation.
4159                          */
4160                         result = ISC_R_NOTFOUND;
4161                         if (vevent->rdataset != NULL)
4162                                 result = dns_db_findnode(fctx->cache,
4163                                                          vevent->name,
4164                                                          ISC_TRUE, &node);
4165                         if (result == ISC_R_SUCCESS) {
4166                                 (void)dns_db_addrdataset(fctx->cache, node,
4167                                                          NULL, now,
4168                                                          vevent->rdataset, 0,
4169                                                          NULL);
4170                         }
4171                         if (result == ISC_R_SUCCESS &&
4172                             vevent->sigrdataset != NULL)
4173                                 (void)dns_db_addrdataset(fctx->cache, node,
4174                                                          NULL, now,
4175                                                          vevent->sigrdataset,
4176                                                          0, NULL);
4177                         if (result == ISC_R_SUCCESS)
4178                                 dns_db_detachnode(fctx->cache, &node);
4179                 }
4180                 result = fctx->vresult;
4181                 add_bad(fctx, addrinfo, result, badns_validation);
4182                 isc_event_free(&event);
4183                 UNLOCK(&res->buckets[fctx->bucketnum].lock);
4184                 INSIST(fctx->validator == NULL);
4185                 fctx->validator = ISC_LIST_HEAD(fctx->validators);
4186                 if (fctx->validator != NULL)
4187                         dns_validator_send(fctx->validator);
4188                 else if (sentresponse)
4189                         fctx_done(fctx, result, __LINE__); /* Locks bucket. */
4190                 else if (result == DNS_R_BROKENCHAIN) {
4191                         isc_result_t tresult;
4192                         isc_time_t expire;
4193                         isc_interval_t i;
4194
4195                         isc_interval_set(&i, DNS_BADCACHE_TTL(fctx), 0);
4196                         tresult = isc_time_nowplusinterval(&expire, &i);
4197                         if (negative &&
4198                             (fctx->type == dns_rdatatype_dnskey ||
4199                              fctx->type == dns_rdatatype_dlv ||
4200                              fctx->type == dns_rdatatype_ds) &&
4201                              tresult == ISC_R_SUCCESS)
4202                                 dns_resolver_addbadcache(res, &fctx->name,
4203                                                          fctx->type, &expire);
4204                         fctx_done(fctx, result, __LINE__); /* Locks bucket. */
4205                 } else
4206                         fctx_try(fctx, ISC_TRUE, ISC_TRUE); /* Locks bucket. */
4207                 return;
4208         }
4209
4210
4211         if (negative) {
4212                 dns_rdatatype_t covers;
4213                 FCTXTRACE("nonexistence validation OK");
4214
4215                 inc_stats(res, dns_resstatscounter_valnegsuccess);
4216
4217                 if (fctx->rmessage->rcode == dns_rcode_nxdomain)
4218                         covers = dns_rdatatype_any;
4219                 else
4220                         covers = fctx->type;
4221
4222                 result = dns_db_findnode(fctx->cache, vevent->name, ISC_TRUE,
4223                                          &node);
4224                 if (result != ISC_R_SUCCESS)
4225                         goto noanswer_response;
4226
4227                 /*
4228                  * If we are asking for a SOA record set the cache time
4229                  * to zero to facilitate locating the containing zone of
4230                  * a arbitrary zone.
4231                  */
4232                 ttl = res->view->maxncachettl;
4233                 if (fctx->type == dns_rdatatype_soa &&
4234                     covers == dns_rdatatype_any && res->zero_no_soa_ttl)
4235                         ttl = 0;
4236
4237                 result = ncache_adderesult(fctx->rmessage, fctx->cache, node,
4238                                            covers, now, ttl, vevent->optout,
4239                                            vevent->secure, ardataset, &eresult);
4240                 if (result != ISC_R_SUCCESS)
4241                         goto noanswer_response;
4242                 goto answer_response;
4243         } else
4244                 inc_stats(res, dns_resstatscounter_valsuccess);
4245
4246         FCTXTRACE("validation OK");
4247
4248         if (vevent->proofs[DNS_VALIDATOR_NOQNAMEPROOF] != NULL) {
4249                 result = dns_rdataset_addnoqname(vevent->rdataset,
4250                                    vevent->proofs[DNS_VALIDATOR_NOQNAMEPROOF]);
4251                 RUNTIME_CHECK(result == ISC_R_SUCCESS);
4252                 INSIST(vevent->sigrdataset != NULL);
4253                 vevent->sigrdataset->ttl = vevent->rdataset->ttl;
4254                 if (vevent->proofs[DNS_VALIDATOR_CLOSESTENCLOSER] != NULL) {
4255                         result = dns_rdataset_addclosest(vevent->rdataset,
4256                                  vevent->proofs[DNS_VALIDATOR_CLOSESTENCLOSER]);
4257                         RUNTIME_CHECK(result == ISC_R_SUCCESS);
4258                 }
4259         } else if (vevent->rdataset->trust == dns_trust_answer &&
4260                    vevent->rdataset->type != dns_rdatatype_rrsig)
4261         {
4262                 isc_result_t tresult;
4263                 dns_name_t *noqname = NULL;
4264                 tresult = findnoqname(fctx, vevent->name,
4265                                       vevent->rdataset->type, &noqname);
4266                 if (tresult == ISC_R_SUCCESS && noqname != NULL) {
4267                         tresult = dns_rdataset_addnoqname(vevent->rdataset,
4268                                                           noqname);
4269                         RUNTIME_CHECK(tresult == ISC_R_SUCCESS);
4270                 }
4271         }
4272
4273         /*
4274          * The data was already cached as pending data.
4275          * Re-cache it as secure and bind the cached
4276          * rdatasets to the first event on the fetch
4277          * event list.
4278          */
4279         result = dns_db_findnode(fctx->cache, vevent->name, ISC_TRUE, &node);
4280         if (result != ISC_R_SUCCESS)
4281                 goto noanswer_response;
4282
4283         result = dns_db_addrdataset(fctx->cache, node, NULL, now,
4284                                     vevent->rdataset, 0, ardataset);
4285         if (result != ISC_R_SUCCESS &&
4286             result != DNS_R_UNCHANGED)
4287                 goto noanswer_response;
4288         if (ardataset != NULL && NEGATIVE(ardataset)) {
4289                 if (NXDOMAIN(ardataset))
4290                         eresult = DNS_R_NCACHENXDOMAIN;
4291                 else
4292                         eresult = DNS_R_NCACHENXRRSET;
4293         } else if (vevent->sigrdataset != NULL) {
4294                 result = dns_db_addrdataset(fctx->cache, node, NULL, now,
4295                                             vevent->sigrdataset, 0,
4296                                             asigrdataset);
4297                 if (result != ISC_R_SUCCESS &&
4298                     result != DNS_R_UNCHANGED)
4299                         goto noanswer_response;
4300         }
4301
4302         if (sentresponse) {
4303                 isc_boolean_t bucket_empty = ISC_FALSE;
4304                 /*
4305                  * If we only deferred the destroy because we wanted to cache
4306                  * the data, destroy now.
4307                  */
4308                 dns_db_detachnode(fctx->cache, &node);
4309                 if (SHUTTINGDOWN(fctx))
4310                         bucket_empty = maybe_destroy(fctx, ISC_TRUE);
4311                 UNLOCK(&res->buckets[fctx->bucketnum].lock);
4312                 if (bucket_empty)
4313                         empty_bucket(res);
4314                 goto cleanup_event;
4315         }
4316
4317         if (!ISC_LIST_EMPTY(fctx->validators)) {
4318                 INSIST(!negative);
4319                 INSIST(fctx->type == dns_rdatatype_any ||
4320                        fctx->type == dns_rdatatype_rrsig ||
4321                        fctx->type == dns_rdatatype_sig);
4322                 /*
4323                  * Don't send a response yet - we have
4324                  * more rdatasets that still need to
4325                  * be validated.
4326                  */
4327                 dns_db_detachnode(fctx->cache, &node);
4328                 UNLOCK(&res->buckets[fctx->bucketnum].lock);
4329                 dns_validator_send(ISC_LIST_HEAD(fctx->validators));
4330                 goto cleanup_event;
4331         }
4332
4333  answer_response:
4334         /*
4335          * Cache any NS/NSEC records that happened to be validated.
4336          */
4337         result = dns_message_firstname(fctx->rmessage, DNS_SECTION_AUTHORITY);
4338         while (result == ISC_R_SUCCESS) {
4339                 name = NULL;
4340                 dns_message_currentname(fctx->rmessage, DNS_SECTION_AUTHORITY,
4341                                         &name);
4342                 for (rdataset = ISC_LIST_HEAD(name->list);
4343                      rdataset != NULL;
4344                      rdataset = ISC_LIST_NEXT(rdataset, link)) {
4345                         if ((rdataset->type != dns_rdatatype_ns &&
4346                              rdataset->type != dns_rdatatype_nsec) ||
4347                             rdataset->trust != dns_trust_secure)
4348                                 continue;
4349                         for (sigrdataset = ISC_LIST_HEAD(name->list);
4350                              sigrdataset != NULL;
4351                              sigrdataset = ISC_LIST_NEXT(sigrdataset, link)) {
4352                                 if (sigrdataset->type != dns_rdatatype_rrsig ||
4353                                     sigrdataset->covers != rdataset->type)
4354                                         continue;
4355                                 break;
4356                         }
4357                         if (sigrdataset == NULL ||
4358                             sigrdataset->trust != dns_trust_secure)
4359                                 continue;
4360                         result = dns_db_findnode(fctx->cache, name, ISC_TRUE,
4361                                                  &nsnode);
4362                         if (result != ISC_R_SUCCESS)
4363                                 continue;
4364
4365                         result = dns_db_addrdataset(fctx->cache, nsnode, NULL,
4366                                                     now, rdataset, 0, NULL);
4367                         if (result == ISC_R_SUCCESS)
4368                                 result = dns_db_addrdataset(fctx->cache, nsnode,
4369                                                             NULL, now,
4370                                                             sigrdataset, 0,
4371                                                             NULL);
4372                         dns_db_detachnode(fctx->cache, &nsnode);
4373                         if (result != ISC_R_SUCCESS)
4374                                 continue;
4375                 }
4376                 result = dns_message_nextname(fctx->rmessage,
4377                                               DNS_SECTION_AUTHORITY);
4378         }
4379
4380         result = ISC_R_SUCCESS;
4381
4382         /*
4383          * Respond with an answer, positive or negative,
4384          * as opposed to an error.  'node' must be non-NULL.
4385          */
4386
4387         fctx->attributes |= FCTX_ATTR_HAVEANSWER;
4388
4389         if (hevent != NULL) {
4390                 /*
4391                  * Negative results must be indicated in event->result.
4392                  */
4393                 if (dns_rdataset_isassociated(hevent->rdataset) &&
4394                     NEGATIVE(hevent->rdataset)) {
4395                         INSIST(eresult == DNS_R_NCACHENXDOMAIN ||
4396                                eresult == DNS_R_NCACHENXRRSET);
4397                 }
4398                 hevent->result = eresult;
4399                 RUNTIME_CHECK(dns_name_copy(vevent->name,
4400                               dns_fixedname_name(&hevent->foundname), NULL)
4401                               == ISC_R_SUCCESS);
4402                 dns_db_attach(fctx->cache, &hevent->db);
4403                 dns_db_transfernode(fctx->cache, &node, &hevent->node);
4404                 clone_results(fctx);
4405         }
4406
4407  noanswer_response:
4408         if (node != NULL)
4409                 dns_db_detachnode(fctx->cache, &node);
4410
4411         UNLOCK(&res->buckets[fctx->bucketnum].lock);
4412         fctx_done(fctx, result, __LINE__); /* Locks bucket. */
4413
4414  cleanup_event:
4415         INSIST(node == NULL);
4416         isc_event_free(&event);
4417 }
4418
4419 static void
4420 fctx_log(void *arg, int level, const char *fmt, ...) {
4421         char msgbuf[2048];
4422         va_list args;
4423         fetchctx_t *fctx = arg;
4424
4425         va_start(args, fmt);
4426         vsnprintf(msgbuf, sizeof(msgbuf), fmt, args);
4427         va_end(args);
4428
4429         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
4430                       DNS_LOGMODULE_RESOLVER, level,
4431                       "fctx %p(%s): %s", fctx, fctx->info, msgbuf);
4432 }
4433
4434 static inline isc_result_t
4435 findnoqname(fetchctx_t *fctx, dns_name_t *name, dns_rdatatype_t type,
4436             dns_name_t **noqnamep)
4437 {
4438         dns_rdataset_t *nrdataset, *next, *sigrdataset;
4439         dns_rdata_rrsig_t rrsig;
4440         isc_result_t result;
4441         unsigned int labels;
4442         dns_section_t section;
4443         dns_name_t *zonename;
4444         dns_fixedname_t fzonename;
4445         dns_name_t *closest;
4446         dns_fixedname_t fclosest;
4447         dns_name_t *nearest;
4448         dns_fixedname_t fnearest;
4449         dns_rdatatype_t found = dns_rdatatype_none;
4450         dns_name_t *noqname = NULL;
4451
4452         FCTXTRACE("findnoqname");
4453
4454         REQUIRE(noqnamep != NULL && *noqnamep == NULL);
4455
4456         /*
4457          * Find the SIG for this rdataset, if we have it.
4458          */
4459         for (sigrdataset = ISC_LIST_HEAD(name->list);
4460              sigrdataset != NULL;
4461              sigrdataset = ISC_LIST_NEXT(sigrdataset, link)) {
4462                 if (sigrdataset->type == dns_rdatatype_rrsig &&
4463                     sigrdataset->covers == type)
4464                         break;
4465         }
4466
4467         if (sigrdataset == NULL)
4468                 return (ISC_R_NOTFOUND);
4469
4470         labels = dns_name_countlabels(name);
4471
4472         for (result = dns_rdataset_first(sigrdataset);
4473              result == ISC_R_SUCCESS;
4474              result = dns_rdataset_next(sigrdataset)) {
4475                 dns_rdata_t rdata = DNS_RDATA_INIT;
4476                 dns_rdataset_current(sigrdataset, &rdata);
4477                 result = dns_rdata_tostruct(&rdata, &rrsig, NULL);
4478                 RUNTIME_CHECK(result == ISC_R_SUCCESS);
4479                 /* Wildcard has rrsig.labels < labels - 1. */
4480                 if (rrsig.labels + 1U >= labels)
4481                         continue;
4482                 break;
4483         }
4484
4485         if (result == ISC_R_NOMORE)
4486                 return (ISC_R_NOTFOUND);
4487         if (result != ISC_R_SUCCESS)
4488                 return (result);
4489
4490         dns_fixedname_init(&fzonename);
4491         zonename = dns_fixedname_name(&fzonename);
4492         dns_fixedname_init(&fclosest);
4493         closest = dns_fixedname_name(&fclosest);
4494         dns_fixedname_init(&fnearest);
4495         nearest = dns_fixedname_name(&fnearest);
4496
4497 #define NXND(x) ((x) == ISC_R_SUCCESS)
4498
4499         section = DNS_SECTION_AUTHORITY;
4500         for (result = dns_message_firstname(fctx->rmessage, section);
4501              result == ISC_R_SUCCESS;
4502              result = dns_message_nextname(fctx->rmessage, section)) {
4503                 dns_name_t *nsec = NULL;
4504                 dns_message_currentname(fctx->rmessage, section, &nsec);
4505                 for (nrdataset = ISC_LIST_HEAD(nsec->list);
4506                       nrdataset != NULL; nrdataset = next) {
4507                         isc_boolean_t data = ISC_FALSE, exists = ISC_FALSE;
4508                         isc_boolean_t optout = ISC_FALSE, unknown = ISC_FALSE;
4509                         isc_boolean_t setclosest = ISC_FALSE;
4510                         isc_boolean_t setnearest = ISC_FALSE;
4511
4512                         next = ISC_LIST_NEXT(nrdataset, link);
4513                         if (nrdataset->type != dns_rdatatype_nsec &&
4514                             nrdataset->type != dns_rdatatype_nsec3)
4515                                 continue;
4516
4517                         if (nrdataset->type == dns_rdatatype_nsec &&
4518                             NXND(dns_nsec_noexistnodata(type, name, nsec,
4519                                                         nrdataset, &exists,
4520                                                         &data, NULL, fctx_log,
4521                                                         fctx)))
4522                         {
4523                                 if (!exists) {
4524                                         noqname = nsec;
4525                                         found = dns_rdatatype_nsec;
4526                                 }
4527                         }
4528
4529                         if (nrdataset->type == dns_rdatatype_nsec3 &&
4530                             NXND(dns_nsec3_noexistnodata(type, name, nsec,
4531                                                          nrdataset, zonename,
4532                                                          &exists, &data,
4533                                                          &optout, &unknown,
4534                                                          &setclosest,
4535                                                          &setnearest,
4536                                                          closest, nearest,
4537                                                          fctx_log, fctx)))
4538                         {
4539                                 if (!exists && setnearest) {
4540                                         noqname = nsec;
4541                                         found = dns_rdatatype_nsec3;
4542                                 }
4543                         }
4544                 }
4545         }
4546         if (result == ISC_R_NOMORE)
4547                 result = ISC_R_SUCCESS;
4548         if (noqname != NULL) {
4549                 for (sigrdataset = ISC_LIST_HEAD(noqname->list);
4550                      sigrdataset != NULL;
4551                      sigrdataset = ISC_LIST_NEXT(sigrdataset, link)) {
4552                         if (sigrdataset->type == dns_rdatatype_rrsig &&
4553                             sigrdataset->covers == found)
4554                                 break;
4555                 }
4556                 if (sigrdataset != NULL)
4557                         *noqnamep = noqname;
4558         }
4559         return (result);
4560 }
4561
4562 static inline isc_result_t
4563 cache_name(fetchctx_t *fctx, dns_name_t *name, dns_adbaddrinfo_t *addrinfo,
4564            isc_stdtime_t now)
4565 {
4566         dns_rdataset_t *rdataset, *sigrdataset;
4567         dns_rdataset_t *addedrdataset, *ardataset, *asigrdataset;
4568         dns_rdataset_t *valrdataset = NULL, *valsigrdataset = NULL;
4569         dns_dbnode_t *node, **anodep;
4570         dns_db_t **adbp;
4571         dns_name_t *aname;
4572         dns_resolver_t *res;
4573         isc_boolean_t need_validation, secure_domain, have_answer;
4574         isc_result_t result, eresult;
4575         dns_fetchevent_t *event;
4576         unsigned int options;
4577         isc_task_t *task;
4578         isc_boolean_t fail;
4579         unsigned int valoptions = 0;
4580
4581         /*
4582          * The appropriate bucket lock must be held.
4583          */
4584
4585         res = fctx->res;
4586         need_validation = ISC_FALSE;
4587         POST(need_validation);
4588         secure_domain = ISC_FALSE;
4589         have_answer = ISC_FALSE;
4590         eresult = ISC_R_SUCCESS;
4591         task = res->buckets[fctx->bucketnum].task;
4592
4593         /*
4594          * Is DNSSEC validation required for this name?
4595          */
4596         if (res->view->enablevalidation) {
4597                 result = dns_view_issecuredomain(res->view, name,
4598                                                  &secure_domain);
4599                 if (result != ISC_R_SUCCESS)
4600                         return (result);
4601
4602                 if (!secure_domain && res->view->dlv != NULL) {
4603                         valoptions = DNS_VALIDATOR_DLV;
4604                         secure_domain = ISC_TRUE;
4605                 }
4606         }
4607
4608         if ((fctx->options & DNS_FETCHOPT_NOVALIDATE) != 0)
4609                 need_validation = ISC_FALSE;
4610         else
4611                 need_validation = secure_domain;
4612
4613         adbp = NULL;
4614         aname = NULL;
4615         anodep = NULL;
4616         ardataset = NULL;
4617         asigrdataset = NULL;
4618         event = NULL;
4619         if ((name->attributes & DNS_NAMEATTR_ANSWER) != 0 &&
4620             !need_validation) {
4621                 have_answer = ISC_TRUE;
4622                 event = ISC_LIST_HEAD(fctx->events);
4623                 if (event != NULL) {
4624                         adbp = &event->db;
4625                         aname = dns_fixedname_name(&event->foundname);
4626                         result = dns_name_copy(name, aname, NULL);
4627                         if (result != ISC_R_SUCCESS)
4628                                 return (result);
4629                         anodep = &event->node;
4630                         /*
4631                          * If this is an ANY, SIG or RRSIG query, we're not
4632                          * going to return any rdatasets, unless we encountered
4633                          * a CNAME or DNAME as "the answer".  In this case,
4634                          * we're going to return DNS_R_CNAME or DNS_R_DNAME
4635                          * and we must set up the rdatasets.
4636                          */
4637                         if ((fctx->type != dns_rdatatype_any &&
4638                              fctx->type != dns_rdatatype_rrsig &&
4639                              fctx->type != dns_rdatatype_sig) ||
4640                             (name->attributes & DNS_NAMEATTR_CHAINING) != 0) {
4641                                 ardataset = event->rdataset;
4642                                 asigrdataset = event->sigrdataset;
4643                         }
4644                 }
4645         }
4646
4647         /*
4648          * Find or create the cache node.
4649          */
4650         node = NULL;
4651         result = dns_db_findnode(fctx->cache, name, ISC_TRUE, &node);
4652         if (result != ISC_R_SUCCESS)
4653                 return (result);
4654
4655         /*
4656          * Cache or validate each cacheable rdataset.
4657          */
4658         fail = ISC_TF((fctx->res->options & DNS_RESOLVER_CHECKNAMESFAIL) != 0);
4659         for (rdataset = ISC_LIST_HEAD(name->list);
4660              rdataset != NULL;
4661              rdataset = ISC_LIST_NEXT(rdataset, link)) {
4662                 if (!CACHE(rdataset))
4663                         continue;
4664                 if (CHECKNAMES(rdataset)) {
4665                         char namebuf[DNS_NAME_FORMATSIZE];
4666                         char typebuf[DNS_RDATATYPE_FORMATSIZE];
4667                         char classbuf[DNS_RDATATYPE_FORMATSIZE];
4668
4669                         dns_name_format(name, namebuf, sizeof(namebuf));
4670                         dns_rdatatype_format(rdataset->type, typebuf,
4671                                              sizeof(typebuf));
4672                         dns_rdataclass_format(rdataset->rdclass, classbuf,
4673                                               sizeof(classbuf));
4674                         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
4675                                       DNS_LOGMODULE_RESOLVER, ISC_LOG_NOTICE,
4676                                       "check-names %s %s/%s/%s",
4677                                       fail ? "failure" : "warning",
4678                                       namebuf, typebuf, classbuf);
4679                         if (fail) {
4680                                 if (ANSWER(rdataset)) {
4681                                         dns_db_detachnode(fctx->cache, &node);
4682                                         return (DNS_R_BADNAME);
4683                                 }
4684                                 continue;
4685                         }
4686                 }
4687
4688                 /*
4689                  * Enforce the configure maximum cache TTL.
4690                  */
4691                 if (rdataset->ttl > res->view->maxcachettl)
4692                         rdataset->ttl = res->view->maxcachettl;
4693
4694                 /*
4695                  * Find the SIG for this rdataset, if we have it.
4696                  */
4697                 for (sigrdataset = ISC_LIST_HEAD(name->list);
4698                      sigrdataset != NULL;
4699                      sigrdataset = ISC_LIST_NEXT(sigrdataset, link)) {
4700                         if (sigrdataset->type == dns_rdatatype_rrsig &&
4701                             sigrdataset->covers == rdataset->type)
4702                                 break;
4703                 }
4704
4705                 /*
4706                  * If this RRset is in a secure domain, is in bailiwick,
4707                  * and is not glue, attempt DNSSEC validation.  (We do not
4708                  * attempt to validate glue or out-of-bailiwick data--even
4709                  * though there might be some performance benefit to doing
4710                  * so--because it makes it simpler and safer to ensure that
4711                  * records from a secure domain are only cached if validated
4712                  * within the context of a query to the domain that owns
4713                  * them.)
4714                  */
4715                 if (secure_domain && rdataset->trust != dns_trust_glue &&
4716                     !EXTERNAL(rdataset)) {
4717                         dns_trust_t trust;
4718
4719                         /*
4720                          * RRSIGs are validated as part of validating the
4721                          * type they cover.
4722                          */
4723                         if (rdataset->type == dns_rdatatype_rrsig)
4724                                 continue;
4725
4726                         if (sigrdataset == NULL) {
4727                                 if (!ANSWER(rdataset) && need_validation) {
4728                                         /*
4729                                          * Ignore non-answer rdatasets that
4730                                          * are missing signatures.
4731                                          */
4732                                         continue;
4733                                 }
4734                         }
4735
4736                         /*
4737                          * Normalize the rdataset and sigrdataset TTLs.
4738                          */
4739                         if (sigrdataset != NULL) {
4740                                 rdataset->ttl = ISC_MIN(rdataset->ttl,
4741                                                         sigrdataset->ttl);
4742                                 sigrdataset->ttl = rdataset->ttl;
4743                         }
4744
4745                         /*
4746                          * Cache this rdataset/sigrdataset pair as
4747                          * pending data.  Track whether it was additional
4748                          * or not.
4749                          */
4750                         if (rdataset->trust == dns_trust_additional)
4751                                 trust = dns_trust_pending_additional;
4752                         else
4753                                 trust = dns_trust_pending_answer;
4754
4755                         rdataset->trust = trust;
4756                         if (sigrdataset != NULL)
4757                                 sigrdataset->trust = trust;
4758                         if (!need_validation || !ANSWER(rdataset)) {
4759                                 if (ANSWER(rdataset) &&
4760                                    rdataset->type != dns_rdatatype_rrsig) {
4761                                         isc_result_t tresult;
4762                                         dns_name_t *noqname = NULL;
4763                                         tresult = findnoqname(fctx, name,
4764                                                               rdataset->type,
4765                                                               &noqname);
4766                                         if (tresult == ISC_R_SUCCESS &&
4767                                             noqname != NULL) {
4768                                                 tresult =
4769                                                      dns_rdataset_addnoqname(
4770                                                             rdataset, noqname);
4771                                                 RUNTIME_CHECK(tresult ==
4772                                                               ISC_R_SUCCESS);
4773                                         }
4774                                 }
4775                                 addedrdataset = ardataset;
4776                                 result = dns_db_addrdataset(fctx->cache, node,
4777                                                             NULL, now, rdataset,
4778                                                             0, addedrdataset);
4779                                 if (result == DNS_R_UNCHANGED) {
4780                                         result = ISC_R_SUCCESS;
4781                                         if (!need_validation &&
4782                                             ardataset != NULL &&
4783                                             NEGATIVE(ardataset)) {
4784                                                 /*
4785                                                  * The answer in the cache is
4786                                                  * better than the answer we
4787                                                  * found, and is a negative
4788                                                  * cache entry, so we must set
4789                                                  * eresult appropriately.
4790                                                  */
4791                                                 if (NXDOMAIN(ardataset))
4792                                                         eresult =
4793                                                            DNS_R_NCACHENXDOMAIN;
4794                                                 else
4795                                                         eresult =
4796                                                            DNS_R_NCACHENXRRSET;
4797                                                 /*
4798                                                  * We have a negative response
4799                                                  * from the cache so don't
4800                                                  * attempt to add the RRSIG
4801                                                  * rrset.
4802                                                  */
4803                                                 continue;
4804                                         }
4805                                 }
4806                                 if (result != ISC_R_SUCCESS)
4807                                         break;
4808                                 if (sigrdataset != NULL) {
4809                                         addedrdataset = asigrdataset;
4810                                         result = dns_db_addrdataset(fctx->cache,
4811                                                                 node, NULL, now,
4812                                                                 sigrdataset, 0,
4813                                                                 addedrdataset);
4814                                         if (result == DNS_R_UNCHANGED)
4815                                                 result = ISC_R_SUCCESS;
4816                                         if (result != ISC_R_SUCCESS)
4817                                                 break;
4818                                 } else if (!ANSWER(rdataset))
4819                                         continue;
4820                         }
4821
4822                         if (ANSWER(rdataset) && need_validation) {
4823                                 if (fctx->type != dns_rdatatype_any &&
4824                                     fctx->type != dns_rdatatype_rrsig &&
4825                                     fctx->type != dns_rdatatype_sig) {
4826                                         /*
4827                                          * This is The Answer.  We will
4828                                          * validate it, but first we cache
4829                                          * the rest of the response - it may
4830                                          * contain useful keys.
4831                                          */
4832                                         INSIST(valrdataset == NULL &&
4833                                                valsigrdataset == NULL);
4834                                         valrdataset = rdataset;
4835                                         valsigrdataset = sigrdataset;
4836                                 } else {
4837                                         /*
4838                                          * This is one of (potentially)
4839                                          * multiple answers to an ANY
4840                                          * or SIG query.  To keep things
4841                                          * simple, we just start the
4842                                          * validator right away rather
4843                                          * than caching first and
4844                                          * having to remember which
4845                                          * rdatasets needed validation.
4846                                          */
4847                                         result = valcreate(fctx, addrinfo,
4848                                                            name, rdataset->type,
4849                                                            rdataset,
4850                                                            sigrdataset,
4851                                                            valoptions, task);
4852                                         /*
4853                                          * Defer any further validations.
4854                                          * This prevents multiple validators
4855                                          * from manipulating fctx->rmessage
4856                                          * simultaneously.
4857                                          */
4858                                         valoptions |= DNS_VALIDATOR_DEFER;
4859                                 }
4860                         } else if (CHAINING(rdataset)) {
4861                                 if (rdataset->type == dns_rdatatype_cname)
4862                                         eresult = DNS_R_CNAME;
4863                                 else {
4864                                         INSIST(rdataset->type ==
4865                                                dns_rdatatype_dname);
4866                                         eresult = DNS_R_DNAME;
4867                                 }
4868                         }
4869                 } else if (!EXTERNAL(rdataset)) {
4870                         /*
4871                          * It's OK to cache this rdataset now.
4872                          */
4873                         if (ANSWER(rdataset))
4874                                 addedrdataset = ardataset;
4875                         else if (ANSWERSIG(rdataset))
4876                                 addedrdataset = asigrdataset;
4877                         else
4878                                 addedrdataset = NULL;
4879                         if (CHAINING(rdataset)) {
4880                                 if (rdataset->type == dns_rdatatype_cname)
4881                                         eresult = DNS_R_CNAME;
4882                                 else {
4883                                         INSIST(rdataset->type ==
4884                                                dns_rdatatype_dname);
4885                                         eresult = DNS_R_DNAME;
4886                                 }
4887                         }
4888                         if (rdataset->trust == dns_trust_glue &&
4889                             (rdataset->type == dns_rdatatype_ns ||
4890                              (rdataset->type == dns_rdatatype_rrsig &&
4891                               rdataset->covers == dns_rdatatype_ns))) {
4892                                 /*
4893                                  * If the trust level is 'dns_trust_glue'
4894                                  * then we are adding data from a referral
4895                                  * we got while executing the search algorithm.
4896                                  * New referral data always takes precedence
4897                                  * over the existing cache contents.
4898                                  */
4899                                 options = DNS_DBADD_FORCE;
4900                         } else
4901                                 options = 0;
4902
4903                         if (ANSWER(rdataset) &&
4904                            rdataset->type != dns_rdatatype_rrsig) {
4905                                 isc_result_t tresult;
4906                                 dns_name_t *noqname = NULL;
4907                                 tresult = findnoqname(fctx, name,
4908                                                       rdataset->type, &noqname);
4909                                 if (tresult == ISC_R_SUCCESS &&
4910                                     noqname != NULL) {
4911                                         tresult = dns_rdataset_addnoqname(
4912                                                             rdataset, noqname);
4913                                         RUNTIME_CHECK(tresult == ISC_R_SUCCESS);
4914                                 }
4915                         }
4916
4917                         /*
4918                          * Now we can add the rdataset.
4919                          */
4920                         result = dns_db_addrdataset(fctx->cache,
4921                                                     node, NULL, now,
4922                                                     rdataset,
4923                                                     options,
4924                                                     addedrdataset);
4925
4926                         if (result == DNS_R_UNCHANGED) {
4927                                 if (ANSWER(rdataset) &&
4928                                     ardataset != NULL &&
4929                                     NEGATIVE(ardataset)) {
4930                                         /*
4931                                          * The answer in the cache is better
4932                                          * than the answer we found, and is
4933                                          * a negative cache entry, so we
4934                                          * must set eresult appropriately.
4935                                          */
4936                                         if (NXDOMAIN(ardataset))
4937                                                 eresult = DNS_R_NCACHENXDOMAIN;
4938                                         else
4939                                                 eresult = DNS_R_NCACHENXRRSET;
4940                                 }
4941                                 result = ISC_R_SUCCESS;
4942                         } else if (result != ISC_R_SUCCESS)
4943                                 break;
4944                 }
4945         }
4946
4947         if (valrdataset != NULL)
4948                 result = valcreate(fctx, addrinfo, name, fctx->type,
4949                                    valrdataset, valsigrdataset, valoptions,
4950                                    task);
4951
4952         if (result == ISC_R_SUCCESS && have_answer) {
4953                 fctx->attributes |= FCTX_ATTR_HAVEANSWER;
4954                 if (event != NULL) {
4955                         /*
4956                          * Negative results must be indicated in event->result.
4957                          */
4958                         if (dns_rdataset_isassociated(event->rdataset) &&
4959                             NEGATIVE(event->rdataset)) {
4960                                 INSIST(eresult == DNS_R_NCACHENXDOMAIN ||
4961                                        eresult == DNS_R_NCACHENXRRSET);
4962                         }
4963                         event->result = eresult;
4964                         dns_db_attach(fctx->cache, adbp);
4965                         dns_db_transfernode(fctx->cache, &node, anodep);
4966                         clone_results(fctx);
4967                 }
4968         }
4969
4970         if (node != NULL)
4971                 dns_db_detachnode(fctx->cache, &node);
4972
4973         return (result);
4974 }
4975
4976 static inline isc_result_t
4977 cache_message(fetchctx_t *fctx, dns_adbaddrinfo_t *addrinfo, isc_stdtime_t now)
4978 {
4979         isc_result_t result;
4980         dns_section_t section;
4981         dns_name_t *name;
4982
4983         FCTXTRACE("cache_message");
4984
4985         fctx->attributes &= ~FCTX_ATTR_WANTCACHE;
4986
4987         LOCK(&fctx->res->buckets[fctx->bucketnum].lock);
4988
4989         for (section = DNS_SECTION_ANSWER;
4990              section <= DNS_SECTION_ADDITIONAL;
4991              section++) {
4992                 result = dns_message_firstname(fctx->rmessage, section);
4993                 while (result == ISC_R_SUCCESS) {
4994                         name = NULL;
4995                         dns_message_currentname(fctx->rmessage, section,
4996                                                 &name);
4997                         if ((name->attributes & DNS_NAMEATTR_CACHE) != 0) {
4998                                 result = cache_name(fctx, name, addrinfo, now);
4999                                 if (result != ISC_R_SUCCESS)
5000                                         break;
5001                         }
5002                         result = dns_message_nextname(fctx->rmessage, section);
5003                 }
5004                 if (result != ISC_R_NOMORE)
5005                         break;
5006         }
5007         if (result == ISC_R_NOMORE)
5008                 result = ISC_R_SUCCESS;
5009
5010         UNLOCK(&fctx->res->buckets[fctx->bucketnum].lock);
5011
5012         return (result);
5013 }
5014
5015 /*
5016  * Do what dns_ncache_addoptout() does, and then compute an appropriate eresult.
5017  */
5018 static isc_result_t
5019 ncache_adderesult(dns_message_t *message, dns_db_t *cache, dns_dbnode_t *node,
5020                   dns_rdatatype_t covers, isc_stdtime_t now, dns_ttl_t maxttl,
5021                   isc_boolean_t optout, isc_boolean_t secure,
5022                   dns_rdataset_t *ardataset, isc_result_t *eresultp)
5023 {
5024         isc_result_t result;
5025         dns_rdataset_t rdataset;
5026
5027         if (ardataset == NULL) {
5028                 dns_rdataset_init(&rdataset);
5029                 ardataset = &rdataset;
5030         }
5031         if (secure)
5032                 result = dns_ncache_addoptout(message, cache, node, covers,
5033                                               now, maxttl, optout, ardataset);
5034         else
5035                 result = dns_ncache_add(message, cache, node, covers, now,
5036                                         maxttl, ardataset);
5037         if (result == DNS_R_UNCHANGED || result == ISC_R_SUCCESS) {
5038                 /*
5039                  * If the cache now contains a negative entry and we
5040                  * care about whether it is DNS_R_NCACHENXDOMAIN or
5041                  * DNS_R_NCACHENXRRSET then extract it.
5042                  */
5043                 if (NEGATIVE(ardataset)) {
5044                         /*
5045                          * The cache data is a negative cache entry.
5046                          */
5047                         if (NXDOMAIN(ardataset))
5048                                 *eresultp = DNS_R_NCACHENXDOMAIN;
5049                         else
5050                                 *eresultp = DNS_R_NCACHENXRRSET;
5051                 } else {
5052                         /*
5053                          * Either we don't care about the nature of the
5054                          * cache rdataset (because no fetch is interested
5055                          * in the outcome), or the cache rdataset is not
5056                          * a negative cache entry.  Whichever case it is,
5057                          * we can return success.
5058                          *
5059                          * XXXRTH  There's a CNAME/DNAME problem here.
5060                          */
5061                         *eresultp = ISC_R_SUCCESS;
5062                 }
5063                 result = ISC_R_SUCCESS;
5064         }
5065         if (ardataset == &rdataset && dns_rdataset_isassociated(ardataset))
5066                 dns_rdataset_disassociate(ardataset);
5067
5068         return (result);
5069 }
5070
5071 static inline isc_result_t
5072 ncache_message(fetchctx_t *fctx, dns_adbaddrinfo_t *addrinfo,
5073                dns_rdatatype_t covers, isc_stdtime_t now)
5074 {
5075         isc_result_t result, eresult;
5076         dns_name_t *name;
5077         dns_resolver_t *res;
5078         dns_db_t **adbp;
5079         dns_dbnode_t *node, **anodep;
5080         dns_rdataset_t *ardataset;
5081         isc_boolean_t need_validation, secure_domain;
5082         dns_name_t *aname;
5083         dns_fetchevent_t *event;
5084         isc_uint32_t ttl;
5085         unsigned int valoptions = 0;
5086
5087         FCTXTRACE("ncache_message");
5088
5089         fctx->attributes &= ~FCTX_ATTR_WANTNCACHE;
5090
5091         res = fctx->res;
5092         need_validation = ISC_FALSE;
5093         POST(need_validation);
5094         secure_domain = ISC_FALSE;
5095         eresult = ISC_R_SUCCESS;
5096         name = &fctx->name;
5097         node = NULL;
5098
5099         /*
5100          * XXXMPA remove when we follow cnames and adjust the setting
5101          * of FCTX_ATTR_WANTNCACHE in noanswer_response().
5102          */
5103         INSIST(fctx->rmessage->counts[DNS_SECTION_ANSWER] == 0);
5104
5105         /*
5106          * Is DNSSEC validation required for this name?
5107          */
5108         if (fctx->res->view->enablevalidation) {
5109                 result = dns_view_issecuredomain(res->view, name,
5110                                                  &secure_domain);
5111                 if (result != ISC_R_SUCCESS)
5112                         return (result);
5113
5114                 if (!secure_domain && res->view->dlv != NULL) {
5115                         valoptions = DNS_VALIDATOR_DLV;
5116                         secure_domain = ISC_TRUE;
5117                 }
5118         }
5119
5120         if ((fctx->options & DNS_FETCHOPT_NOVALIDATE) != 0)
5121                 need_validation = ISC_FALSE;
5122         else
5123                 need_validation = secure_domain;
5124
5125         if (secure_domain) {
5126                 /*
5127                  * Mark all rdatasets as pending.
5128                  */
5129                 dns_rdataset_t *trdataset;
5130                 dns_name_t *tname;
5131
5132                 result = dns_message_firstname(fctx->rmessage,
5133                                                DNS_SECTION_AUTHORITY);
5134                 while (result == ISC_R_SUCCESS) {
5135                         tname = NULL;
5136                         dns_message_currentname(fctx->rmessage,
5137                                                 DNS_SECTION_AUTHORITY,
5138                                                 &tname);
5139                         for (trdataset = ISC_LIST_HEAD(tname->list);
5140                              trdataset != NULL;
5141                              trdataset = ISC_LIST_NEXT(trdataset, link))
5142                                 trdataset->trust = dns_trust_pending_answer;
5143                         result = dns_message_nextname(fctx->rmessage,
5144                                                       DNS_SECTION_AUTHORITY);
5145                 }
5146                 if (result != ISC_R_NOMORE)
5147                         return (result);
5148
5149         }
5150
5151         if (need_validation) {
5152                 /*
5153                  * Do negative response validation.
5154                  */
5155                 result = valcreate(fctx, addrinfo, name, fctx->type,
5156                                    NULL, NULL, valoptions,
5157                                    res->buckets[fctx->bucketnum].task);
5158                 /*
5159                  * If validation is necessary, return now.  Otherwise continue
5160                  * to process the message, letting the validation complete
5161                  * in its own good time.
5162                  */
5163                 return (result);
5164         }
5165
5166         LOCK(&res->buckets[fctx->bucketnum].lock);
5167
5168         adbp = NULL;
5169         aname = NULL;
5170         anodep = NULL;
5171         ardataset = NULL;
5172         if (!HAVE_ANSWER(fctx)) {
5173                 event = ISC_LIST_HEAD(fctx->events);
5174                 if (event != NULL) {
5175                         adbp = &event->db;
5176                         aname = dns_fixedname_name(&event->foundname);
5177                         result = dns_name_copy(name, aname, NULL);
5178                         if (result != ISC_R_SUCCESS)
5179                                 goto unlock;
5180                         anodep = &event->node;
5181                         ardataset = event->rdataset;
5182                 }
5183         } else
5184                 event = NULL;
5185
5186         result = dns_db_findnode(fctx->cache, name, ISC_TRUE, &node);
5187         if (result != ISC_R_SUCCESS)
5188                 goto unlock;
5189
5190         /*
5191          * If we are asking for a SOA record set the cache time
5192          * to zero to facilitate locating the containing zone of
5193          * a arbitrary zone.
5194          */
5195         ttl = fctx->res->view->maxncachettl;
5196         if (fctx->type == dns_rdatatype_soa &&
5197             covers == dns_rdatatype_any &&
5198             fctx->res->zero_no_soa_ttl)
5199                 ttl = 0;
5200
5201         result = ncache_adderesult(fctx->rmessage, fctx->cache, node,
5202                                    covers, now, ttl, ISC_FALSE,
5203                                    ISC_FALSE, ardataset, &eresult);
5204         if (result != ISC_R_SUCCESS)
5205                 goto unlock;
5206
5207         if (!HAVE_ANSWER(fctx)) {
5208                 fctx->attributes |= FCTX_ATTR_HAVEANSWER;
5209                 if (event != NULL) {
5210                         event->result = eresult;
5211                         dns_db_attach(fctx->cache, adbp);
5212                         dns_db_transfernode(fctx->cache, &node, anodep);
5213                         clone_results(fctx);
5214                 }
5215         }
5216
5217  unlock:
5218         UNLOCK(&res->buckets[fctx->bucketnum].lock);
5219
5220         if (node != NULL)
5221                 dns_db_detachnode(fctx->cache, &node);
5222
5223         return (result);
5224 }
5225
5226 static inline void
5227 mark_related(dns_name_t *name, dns_rdataset_t *rdataset,
5228              isc_boolean_t external, isc_boolean_t gluing)
5229 {
5230         name->attributes |= DNS_NAMEATTR_CACHE;
5231         if (gluing) {
5232                 rdataset->trust = dns_trust_glue;
5233                 /*
5234                  * Glue with 0 TTL causes problems.  We force the TTL to
5235                  * 1 second to prevent this.
5236                  */
5237                 if (rdataset->ttl == 0)
5238                         rdataset->ttl = 1;
5239         } else
5240                 rdataset->trust = dns_trust_additional;
5241         /*
5242          * Avoid infinite loops by only marking new rdatasets.
5243          */
5244         if (!CACHE(rdataset)) {
5245                 name->attributes |= DNS_NAMEATTR_CHASE;
5246                 rdataset->attributes |= DNS_RDATASETATTR_CHASE;
5247         }
5248         rdataset->attributes |= DNS_RDATASETATTR_CACHE;
5249         if (external)
5250                 rdataset->attributes |= DNS_RDATASETATTR_EXTERNAL;
5251 }
5252
5253 static isc_result_t
5254 check_section(void *arg, dns_name_t *addname, dns_rdatatype_t type,
5255               dns_section_t section)
5256 {
5257         fetchctx_t *fctx = arg;
5258         isc_result_t result;
5259         dns_name_t *name;
5260         dns_rdataset_t *rdataset;
5261         isc_boolean_t external;
5262         dns_rdatatype_t rtype;
5263         isc_boolean_t gluing;
5264
5265         REQUIRE(VALID_FCTX(fctx));
5266
5267 #if CHECK_FOR_GLUE_IN_ANSWER
5268         if (section == DNS_SECTION_ANSWER && type != dns_rdatatype_a)
5269                 return (ISC_R_SUCCESS);
5270 #endif
5271
5272         if (GLUING(fctx))
5273                 gluing = ISC_TRUE;
5274         else
5275                 gluing = ISC_FALSE;
5276         name = NULL;
5277         rdataset = NULL;
5278         result = dns_message_findname(fctx->rmessage, section, addname,
5279                                       dns_rdatatype_any, 0, &name, NULL);
5280         if (result == ISC_R_SUCCESS) {
5281                 external = ISC_TF(!dns_name_issubdomain(name, &fctx->domain));
5282                 if (type == dns_rdatatype_a) {
5283                         for (rdataset = ISC_LIST_HEAD(name->list);
5284                              rdataset != NULL;
5285                              rdataset = ISC_LIST_NEXT(rdataset, link)) {
5286                                 if (rdataset->type == dns_rdatatype_rrsig)
5287                                         rtype = rdataset->covers;
5288                                 else
5289                                         rtype = rdataset->type;
5290                                 if (rtype == dns_rdatatype_a ||
5291                                     rtype == dns_rdatatype_aaaa)
5292                                         mark_related(name, rdataset, external,
5293                                                      gluing);
5294                         }
5295                 } else {
5296                         result = dns_message_findtype(name, type, 0,
5297                                                       &rdataset);
5298                         if (result == ISC_R_SUCCESS) {
5299                                 mark_related(name, rdataset, external, gluing);
5300                                 /*
5301                                  * Do we have its SIG too?
5302                                  */
5303                                 rdataset = NULL;
5304                                 result = dns_message_findtype(name,
5305                                                       dns_rdatatype_rrsig,
5306                                                       type, &rdataset);
5307                                 if (result == ISC_R_SUCCESS)
5308                                         mark_related(name, rdataset, external,
5309                                                      gluing);
5310                         }
5311                 }
5312         }
5313
5314         return (ISC_R_SUCCESS);
5315 }
5316
5317 static isc_result_t
5318 check_related(void *arg, dns_name_t *addname, dns_rdatatype_t type) {
5319         return (check_section(arg, addname, type, DNS_SECTION_ADDITIONAL));
5320 }
5321
5322 #ifndef CHECK_FOR_GLUE_IN_ANSWER
5323 #define CHECK_FOR_GLUE_IN_ANSWER 0
5324 #endif
5325 #if CHECK_FOR_GLUE_IN_ANSWER
5326 static isc_result_t
5327 check_answer(void *arg, dns_name_t *addname, dns_rdatatype_t type) {
5328         return (check_section(arg, addname, type, DNS_SECTION_ANSWER));
5329 }
5330 #endif
5331
5332 static void
5333 chase_additional(fetchctx_t *fctx) {
5334         isc_boolean_t rescan;
5335         dns_section_t section = DNS_SECTION_ADDITIONAL;
5336         isc_result_t result;
5337
5338  again:
5339         rescan = ISC_FALSE;
5340
5341         for (result = dns_message_firstname(fctx->rmessage, section);
5342              result == ISC_R_SUCCESS;
5343              result = dns_message_nextname(fctx->rmessage, section)) {
5344                 dns_name_t *name = NULL;
5345                 dns_rdataset_t *rdataset;
5346                 dns_message_currentname(fctx->rmessage, DNS_SECTION_ADDITIONAL,
5347                                         &name);
5348                 if ((name->attributes & DNS_NAMEATTR_CHASE) == 0)
5349                         continue;
5350                 name->attributes &= ~DNS_NAMEATTR_CHASE;
5351                 for (rdataset = ISC_LIST_HEAD(name->list);
5352                      rdataset != NULL;
5353                      rdataset = ISC_LIST_NEXT(rdataset, link)) {
5354                         if (CHASE(rdataset)) {
5355                                 rdataset->attributes &= ~DNS_RDATASETATTR_CHASE;
5356                                 (void)dns_rdataset_additionaldata(rdataset,
5357                                                                   check_related,
5358                                                                   fctx);
5359                                 rescan = ISC_TRUE;
5360                         }
5361                 }
5362         }
5363         if (rescan)
5364                 goto again;
5365 }
5366
5367 static inline isc_result_t
5368 cname_target(dns_rdataset_t *rdataset, dns_name_t *tname) {
5369         isc_result_t result;
5370         dns_rdata_t rdata = DNS_RDATA_INIT;
5371         dns_rdata_cname_t cname;
5372
5373         result = dns_rdataset_first(rdataset);
5374         if (result != ISC_R_SUCCESS)
5375                 return (result);
5376         dns_rdataset_current(rdataset, &rdata);
5377         result = dns_rdata_tostruct(&rdata, &cname, NULL);
5378         if (result != ISC_R_SUCCESS)
5379                 return (result);
5380         dns_name_init(tname, NULL);
5381         dns_name_clone(&cname.cname, tname);
5382         dns_rdata_freestruct(&cname);
5383
5384         return (ISC_R_SUCCESS);
5385 }
5386
5387 static inline isc_result_t
5388 dname_target(fetchctx_t *fctx, dns_rdataset_t *rdataset, dns_name_t *qname,
5389              dns_name_t *oname, dns_fixedname_t *fixeddname)
5390 {
5391         isc_result_t result;
5392         dns_rdata_t rdata = DNS_RDATA_INIT;
5393         unsigned int nlabels;
5394         int order;
5395         dns_namereln_t namereln;
5396         dns_rdata_dname_t dname;
5397         dns_fixedname_t prefix;
5398
5399         /*
5400          * Get the target name of the DNAME.
5401          */
5402         result = dns_rdataset_first(rdataset);
5403         if (result != ISC_R_SUCCESS)
5404                 return (result);
5405         dns_rdataset_current(rdataset, &rdata);
5406         result = dns_rdata_tostruct(&rdata, &dname, NULL);
5407         if (result != ISC_R_SUCCESS)
5408                 return (result);
5409
5410         /*
5411          * Get the prefix of qname.
5412          */
5413         namereln = dns_name_fullcompare(qname, oname, &order, &nlabels);
5414         if (namereln != dns_namereln_subdomain) {
5415                 char qbuf[DNS_NAME_FORMATSIZE];
5416                 char obuf[DNS_NAME_FORMATSIZE];
5417
5418                 dns_rdata_freestruct(&dname);
5419                 dns_name_format(qname, qbuf, sizeof(qbuf));
5420                 dns_name_format(oname, obuf, sizeof(obuf));
5421                 log_formerr(fctx, "unrelated DNAME in answer: "
5422                                    "%s is not in %s", qbuf, obuf);
5423                 return (DNS_R_FORMERR);
5424         }
5425         dns_fixedname_init(&prefix);
5426         dns_name_split(qname, nlabels, dns_fixedname_name(&prefix), NULL);
5427         dns_fixedname_init(fixeddname);
5428         result = dns_name_concatenate(dns_fixedname_name(&prefix),
5429                                       &dname.dname,
5430                                       dns_fixedname_name(fixeddname), NULL);
5431         dns_rdata_freestruct(&dname);
5432         return (result);
5433 }
5434
5435 static isc_boolean_t
5436 is_answeraddress_allowed(dns_view_t *view, dns_name_t *name,
5437                          dns_rdataset_t *rdataset)
5438 {
5439         isc_result_t result;
5440         dns_rdata_t rdata = DNS_RDATA_INIT;
5441         struct in_addr ina;
5442         struct in6_addr in6a;
5443         isc_netaddr_t netaddr;
5444         char addrbuf[ISC_NETADDR_FORMATSIZE];
5445         char namebuf[DNS_NAME_FORMATSIZE];
5446         char classbuf[64];
5447         char typebuf[64];
5448         int match;
5449
5450         /* By default, we allow any addresses. */
5451         if (view->denyansweracl == NULL)
5452                 return (ISC_TRUE);
5453
5454         /*
5455          * If the owner name matches one in the exclusion list, either exactly
5456          * or partially, allow it.
5457          */
5458         if (view->answeracl_exclude != NULL) {
5459                 dns_rbtnode_t *node = NULL;
5460
5461                 result = dns_rbt_findnode(view->answeracl_exclude, name, NULL,
5462                                           &node, NULL, 0, NULL, NULL);
5463
5464                 if (result == ISC_R_SUCCESS || result == DNS_R_PARTIALMATCH)
5465                         return (ISC_TRUE);
5466         }
5467
5468         /*
5469          * Otherwise, search the filter list for a match for each address
5470          * record.  If a match is found, the address should be filtered,
5471          * so should the entire answer.
5472          */
5473         for (result = dns_rdataset_first(rdataset);
5474              result == ISC_R_SUCCESS;
5475              result = dns_rdataset_next(rdataset)) {
5476                 dns_rdata_reset(&rdata);
5477                 dns_rdataset_current(rdataset, &rdata);
5478                 if (rdataset->type == dns_rdatatype_a) {
5479                         INSIST(rdata.length == sizeof(ina.s_addr));
5480                         memmove(&ina.s_addr, rdata.data, sizeof(ina.s_addr));
5481                         isc_netaddr_fromin(&netaddr, &ina);
5482                 } else {
5483                         INSIST(rdata.length == sizeof(in6a.s6_addr));
5484                         memmove(in6a.s6_addr, rdata.data, sizeof(in6a.s6_addr));
5485                         isc_netaddr_fromin6(&netaddr, &in6a);
5486                 }
5487
5488                 result = dns_acl_match(&netaddr, NULL, view->denyansweracl,
5489                                        &view->aclenv, &match, NULL);
5490
5491                 if (result == ISC_R_SUCCESS && match > 0) {
5492                         isc_netaddr_format(&netaddr, addrbuf, sizeof(addrbuf));
5493                         dns_name_format(name, namebuf, sizeof(namebuf));
5494                         dns_rdatatype_format(rdataset->type, typebuf,
5495                                              sizeof(typebuf));
5496                         dns_rdataclass_format(rdataset->rdclass, classbuf,
5497                                               sizeof(classbuf));
5498                         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
5499                                       DNS_LOGMODULE_RESOLVER, ISC_LOG_NOTICE,
5500                                       "answer address %s denied for %s/%s/%s",
5501                                       addrbuf, namebuf, typebuf, classbuf);
5502                         return (ISC_FALSE);
5503                 }
5504         }
5505
5506         return (ISC_TRUE);
5507 }
5508
5509 static isc_boolean_t
5510 is_answertarget_allowed(dns_view_t *view, dns_name_t *name,
5511                         dns_rdatatype_t type, dns_name_t *tname,
5512                         dns_name_t *domain)
5513 {
5514         isc_result_t result;
5515         dns_rbtnode_t *node = NULL;
5516         char qnamebuf[DNS_NAME_FORMATSIZE];
5517         char tnamebuf[DNS_NAME_FORMATSIZE];
5518         char classbuf[64];
5519         char typebuf[64];
5520
5521         /* By default, we allow any target name. */
5522         if (view->denyanswernames == NULL)
5523                 return (ISC_TRUE);
5524
5525         /*
5526          * If the owner name matches one in the exclusion list, either exactly
5527          * or partially, allow it.
5528          */
5529         if (view->answernames_exclude != NULL) {
5530                 result = dns_rbt_findnode(view->answernames_exclude, name, NULL,
5531                                           &node, NULL, 0, NULL, NULL);
5532                 if (result == ISC_R_SUCCESS || result == DNS_R_PARTIALMATCH)
5533                         return (ISC_TRUE);
5534         }
5535
5536         /*
5537          * If the target name is a subdomain of the search domain, allow it.
5538          */
5539         if (dns_name_issubdomain(tname, domain))
5540                 return (ISC_TRUE);
5541
5542         /*
5543          * Otherwise, apply filters.
5544          */
5545         result = dns_rbt_findnode(view->denyanswernames, tname, NULL, &node,
5546                                   NULL, 0, NULL, NULL);
5547         if (result == ISC_R_SUCCESS || result == DNS_R_PARTIALMATCH) {
5548                 dns_name_format(name, qnamebuf, sizeof(qnamebuf));
5549                 dns_name_format(tname, tnamebuf, sizeof(tnamebuf));
5550                 dns_rdatatype_format(type, typebuf, sizeof(typebuf));
5551                 dns_rdataclass_format(view->rdclass, classbuf,
5552                                       sizeof(classbuf));
5553                 isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
5554                               DNS_LOGMODULE_RESOLVER, ISC_LOG_NOTICE,
5555                               "%s target %s denied for %s/%s",
5556                               typebuf, tnamebuf, qnamebuf, classbuf);
5557                 return (ISC_FALSE);
5558         }
5559
5560         return (ISC_TRUE);
5561 }
5562
5563 static void
5564 trim_ns_ttl(fetchctx_t *fctx, dns_name_t *name, dns_rdataset_t *rdataset) {
5565         char ns_namebuf[DNS_NAME_FORMATSIZE];
5566         char namebuf[DNS_NAME_FORMATSIZE];
5567         char tbuf[DNS_RDATATYPE_FORMATSIZE];
5568
5569         if (fctx->ns_ttl_ok && rdataset->ttl > fctx->ns_ttl) {
5570                 dns_name_format(name, ns_namebuf, sizeof(ns_namebuf));
5571                 dns_name_format(&fctx->name, namebuf, sizeof(namebuf));
5572                 dns_rdatatype_format(fctx->type, tbuf, sizeof(tbuf));
5573
5574                 isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
5575                               DNS_LOGMODULE_RESOLVER, ISC_LOG_DEBUG(10),
5576                               "fctx %p: trimming ttl of %s/NS for %s/%s: "
5577                               "%u -> %u", fctx, ns_namebuf, namebuf, tbuf,
5578                               rdataset->ttl, fctx->ns_ttl);
5579                 rdataset->ttl = fctx->ns_ttl;
5580         }
5581 }
5582
5583 /*
5584  * Handle a no-answer response (NXDOMAIN, NXRRSET, or referral).
5585  * If look_in_options has LOOK_FOR_NS_IN_ANSWER then we look in the answer
5586  * section for the NS RRset if the query type is NS; if it has
5587  * LOOK_FOR_GLUE_IN_ANSWER we look for glue incorrectly returned in the answer
5588  * section for A and AAAA queries.
5589  */
5590 #define LOOK_FOR_NS_IN_ANSWER 0x1
5591 #define LOOK_FOR_GLUE_IN_ANSWER 0x2
5592
5593 static isc_result_t
5594 noanswer_response(fetchctx_t *fctx, dns_name_t *oqname,
5595                   unsigned int look_in_options)
5596 {
5597         isc_result_t result;
5598         dns_message_t *message;
5599         dns_name_t *name, *qname, *ns_name, *soa_name, *ds_name, *save_name;
5600         dns_rdataset_t *rdataset, *ns_rdataset;
5601         isc_boolean_t aa, negative_response;
5602         dns_rdatatype_t type, save_type;
5603         dns_section_t section;
5604
5605         FCTXTRACE("noanswer_response");
5606
5607         if ((look_in_options & LOOK_FOR_NS_IN_ANSWER) != 0) {
5608                 INSIST(fctx->type == dns_rdatatype_ns);
5609                 section = DNS_SECTION_ANSWER;
5610         } else
5611                 section = DNS_SECTION_AUTHORITY;
5612
5613         message = fctx->rmessage;
5614
5615         /*
5616          * Setup qname.
5617          */
5618         if (oqname == NULL) {
5619                 /*
5620                  * We have a normal, non-chained negative response or
5621                  * referral.
5622                  */
5623                 if ((message->flags & DNS_MESSAGEFLAG_AA) != 0)
5624                         aa = ISC_TRUE;
5625                 else
5626                         aa = ISC_FALSE;
5627                 qname = &fctx->name;
5628         } else {
5629                 /*
5630                  * We're being invoked by answer_response() after it has
5631                  * followed a CNAME/DNAME chain.
5632                  */
5633                 qname = oqname;
5634                 aa = ISC_FALSE;
5635                 /*
5636                  * If the current qname is not a subdomain of the query
5637                  * domain, there's no point in looking at the authority
5638                  * section without doing DNSSEC validation.
5639                  *
5640                  * Until we do that validation, we'll just return success
5641                  * in this case.
5642                  */
5643                 if (!dns_name_issubdomain(qname, &fctx->domain))
5644                         return (ISC_R_SUCCESS);
5645         }
5646
5647         /*
5648          * We have to figure out if this is a negative response, or a
5649          * referral.
5650          */
5651
5652         /*
5653          * Sometimes we can tell if its a negative response by looking at
5654          * the message header.
5655          */
5656         negative_response = ISC_FALSE;
5657         if (message->rcode == dns_rcode_nxdomain ||
5658             (message->counts[DNS_SECTION_ANSWER] == 0 &&
5659              message->counts[DNS_SECTION_AUTHORITY] == 0))
5660                 negative_response = ISC_TRUE;
5661
5662         /*
5663          * Process the authority section.
5664          */
5665         ns_name = NULL;
5666         ns_rdataset = NULL;
5667         soa_name = NULL;
5668         ds_name = NULL;
5669         save_name = NULL;
5670         save_type = dns_rdatatype_none;
5671         result = dns_message_firstname(message, section);
5672         while (result == ISC_R_SUCCESS) {
5673                 name = NULL;
5674                 dns_message_currentname(message, section, &name);
5675                 if (dns_name_issubdomain(name, &fctx->domain)) {
5676                         /*
5677                          * Look for NS/SOA RRsets first.
5678                          */
5679                         for (rdataset = ISC_LIST_HEAD(name->list);
5680                              rdataset != NULL;
5681                              rdataset = ISC_LIST_NEXT(rdataset, link)) {
5682                                 type = rdataset->type;
5683                                 if (type == dns_rdatatype_rrsig)
5684                                         type = rdataset->covers;
5685                                 if (((type == dns_rdatatype_ns ||
5686                                       type == dns_rdatatype_soa) &&
5687                                      !dns_name_issubdomain(qname, name))) {
5688                                         char qbuf[DNS_NAME_FORMATSIZE];
5689                                         char nbuf[DNS_NAME_FORMATSIZE];
5690                                         char tbuf[DNS_RDATATYPE_FORMATSIZE];
5691                                         dns_rdatatype_format(fctx->type, tbuf,
5692                                                              sizeof(tbuf));
5693                                         dns_name_format(name, nbuf,
5694                                                              sizeof(nbuf));
5695                                         dns_name_format(qname, qbuf,
5696                                                              sizeof(qbuf));
5697                                         log_formerr(fctx,
5698                                                     "unrelated %s %s in "
5699                                                     "%s authority section",
5700                                                     tbuf, qbuf, nbuf);
5701                                         return (DNS_R_FORMERR);
5702                                 }
5703                                 if (type == dns_rdatatype_ns) {
5704                                         /*
5705                                          * NS or RRSIG NS.
5706                                          *
5707                                          * Only one set of NS RRs is allowed.
5708                                          */
5709                                         if (rdataset->type ==
5710                                             dns_rdatatype_ns) {
5711                                                 if (ns_name != NULL &&
5712                                                     name != ns_name) {
5713                                                         log_formerr(fctx,
5714                                                                 "multiple NS "
5715                                                                 "RRsets in "
5716                                                                 "authority "
5717                                                                 "section");
5718                                                         return (DNS_R_FORMERR);
5719                                                 }
5720                                                 ns_name = name;
5721                                                 ns_rdataset = rdataset;
5722                                         }
5723                                         name->attributes |=
5724                                                 DNS_NAMEATTR_CACHE;
5725                                         rdataset->attributes |=
5726                                                 DNS_RDATASETATTR_CACHE;
5727                                         rdataset->trust = dns_trust_glue;
5728                                 }
5729                                 if (type == dns_rdatatype_soa) {
5730                                         /*
5731                                          * SOA, or RRSIG SOA.
5732                                          *
5733                                          * Only one SOA is allowed.
5734                                          */
5735                                         if (rdataset->type ==
5736                                             dns_rdatatype_soa) {
5737                                                 if (soa_name != NULL &&
5738                                                     name != soa_name) {
5739                                                         log_formerr(fctx,
5740                                                                 "multiple SOA "
5741                                                                 "RRs in "
5742                                                                 "authority "
5743                                                                 "section");
5744                                                         return (DNS_R_FORMERR);
5745                                                 }
5746                                                 soa_name = name;
5747                                         }
5748                                         name->attributes |=
5749                                                 DNS_NAMEATTR_NCACHE;
5750                                         rdataset->attributes |=
5751                                                 DNS_RDATASETATTR_NCACHE;
5752                                         if (aa)
5753                                                 rdataset->trust =
5754                                                     dns_trust_authauthority;
5755                                         else if (ISFORWARDER(fctx->addrinfo))
5756                                                 rdataset->trust =
5757                                                         dns_trust_answer;
5758                                         else
5759                                                 rdataset->trust =
5760                                                         dns_trust_additional;
5761                                 }
5762                         }
5763                 }
5764                 result = dns_message_nextname(message, section);
5765                 if (result == ISC_R_NOMORE)
5766                         break;
5767                 else if (result != ISC_R_SUCCESS)
5768                         return (result);
5769         }
5770
5771         log_ns_ttl(fctx, "noanswer_response");
5772
5773         if (ns_rdataset != NULL && dns_name_equal(&fctx->domain, ns_name) &&
5774             !dns_name_equal(ns_name, dns_rootname))
5775                 trim_ns_ttl(fctx, ns_name, ns_rdataset);
5776
5777         /*
5778          * A negative response has a SOA record (Type 2)
5779          * and a optional NS RRset (Type 1) or it has neither
5780          * a SOA or a NS RRset (Type 3, handled above) or
5781          * rcode is NXDOMAIN (handled above) in which case
5782          * the NS RRset is allowed (Type 4).
5783          */
5784         if (soa_name != NULL)
5785                 negative_response = ISC_TRUE;
5786
5787         result = dns_message_firstname(message, section);
5788         while (result == ISC_R_SUCCESS) {
5789                 name = NULL;
5790                 dns_message_currentname(message, section, &name);
5791                 if (dns_name_issubdomain(name, &fctx->domain)) {
5792                         for (rdataset = ISC_LIST_HEAD(name->list);
5793                              rdataset != NULL;
5794                              rdataset = ISC_LIST_NEXT(rdataset, link)) {
5795                                 type = rdataset->type;
5796                                 if (type == dns_rdatatype_rrsig)
5797                                         type = rdataset->covers;
5798                                 if (type == dns_rdatatype_nsec ||
5799                                     type == dns_rdatatype_nsec3) {
5800                                         /*
5801                                          * NSEC or RRSIG NSEC.
5802                                          */
5803                                         if (negative_response) {
5804                                                 name->attributes |=
5805                                                         DNS_NAMEATTR_NCACHE;
5806                                                 rdataset->attributes |=
5807                                                         DNS_RDATASETATTR_NCACHE;
5808                                         } else if (type == dns_rdatatype_nsec) {
5809                                                 name->attributes |=
5810                                                         DNS_NAMEATTR_CACHE;
5811                                                 rdataset->attributes |=
5812                                                         DNS_RDATASETATTR_CACHE;
5813                                         }
5814                                         if (aa)
5815                                                 rdataset->trust =
5816                                                     dns_trust_authauthority;
5817                                         else if (ISFORWARDER(fctx->addrinfo))
5818                                                 rdataset->trust =
5819                                                         dns_trust_answer;
5820                                         else
5821                                                 rdataset->trust =
5822                                                         dns_trust_additional;
5823                                         /*
5824                                          * No additional data needs to be
5825                                          * marked.
5826                                          */
5827                                 } else if (type == dns_rdatatype_ds) {
5828                                         /*
5829                                          * DS or SIG DS.
5830                                          *
5831                                          * These should only be here if
5832                                          * this is a referral, and there
5833                                          * should only be one DS RRset.
5834                                          */
5835                                         if (ns_name == NULL) {
5836                                                 log_formerr(fctx,
5837                                                             "DS with no "
5838                                                             "referral");
5839                                                 return (DNS_R_FORMERR);
5840                                         }
5841                                         if (rdataset->type ==
5842                                             dns_rdatatype_ds) {
5843                                                 if (ds_name != NULL &&
5844                                                     name != ds_name) {
5845                                                         log_formerr(fctx,
5846                                                                 "DS doesn't "
5847                                                                 "match "
5848                                                                 "referral "
5849                                                                 "(NS)");
5850                                                         return (DNS_R_FORMERR);
5851                                                 }
5852                                                 ds_name = name;
5853                                         }
5854                                         name->attributes |=
5855                                                 DNS_NAMEATTR_CACHE;
5856                                         rdataset->attributes |=
5857                                                 DNS_RDATASETATTR_CACHE;
5858                                         if (aa)
5859                                                 rdataset->trust =
5860                                                     dns_trust_authauthority;
5861                                         else if (ISFORWARDER(fctx->addrinfo))
5862                                                 rdataset->trust =
5863                                                         dns_trust_answer;
5864                                         else
5865                                                 rdataset->trust =
5866                                                         dns_trust_additional;
5867                                 }
5868                         }
5869                 } else {
5870                         save_name = name;
5871                         save_type = ISC_LIST_HEAD(name->list)->type;
5872                 }
5873                 result = dns_message_nextname(message, section);
5874                 if (result == ISC_R_NOMORE)
5875                         break;
5876                 else if (result != ISC_R_SUCCESS)
5877                         return (result);
5878         }
5879
5880         /*
5881          * Trigger lookups for DNS nameservers.
5882          */
5883         if (negative_response && message->rcode == dns_rcode_noerror &&
5884             fctx->type == dns_rdatatype_ds && soa_name != NULL &&
5885             dns_name_equal(soa_name, qname) &&
5886             !dns_name_equal(qname, dns_rootname))
5887                 return (DNS_R_CHASEDSSERVERS);
5888
5889         /*
5890          * Did we find anything?
5891          */
5892         if (!negative_response && ns_name == NULL) {
5893                 /*
5894                  * Nope.
5895                  */
5896                 if (oqname != NULL) {
5897                         /*
5898                          * We've already got a partial CNAME/DNAME chain,
5899                          * and haven't found else anything useful here, but
5900                          * no error has occurred since we have an answer.
5901                          */
5902                         return (ISC_R_SUCCESS);
5903                 } else {
5904                         /*
5905                          * The responder is insane.
5906                          */
5907                         if (save_name == NULL) {
5908                                 log_formerr(fctx, "invalid response");
5909                                 return (DNS_R_FORMERR);
5910                         }
5911                         if (!dns_name_issubdomain(save_name, &fctx->domain)) {
5912                                 char nbuf[DNS_NAME_FORMATSIZE];
5913                                 char dbuf[DNS_NAME_FORMATSIZE];
5914                                 char tbuf[DNS_RDATATYPE_FORMATSIZE];
5915
5916                                 dns_rdatatype_format(save_type, tbuf,
5917                                         sizeof(tbuf));
5918                                 dns_name_format(save_name, nbuf, sizeof(nbuf));
5919                                 dns_name_format(&fctx->domain, dbuf,
5920                                         sizeof(dbuf));
5921
5922                                 log_formerr(fctx, "Name %s (%s) not subdomain"
5923                                         " of zone %s -- invalid response",
5924                                         nbuf, tbuf, dbuf);
5925                         } else {
5926                                 log_formerr(fctx, "invalid response");
5927                         }
5928                         return (DNS_R_FORMERR);
5929                 }
5930         }
5931
5932         /*
5933          * If we found both NS and SOA, they should be the same name.
5934          */
5935         if (ns_name != NULL && soa_name != NULL && ns_name != soa_name) {
5936                 log_formerr(fctx, "NS/SOA mismatch");
5937                 return (DNS_R_FORMERR);
5938         }
5939
5940         /*
5941          * Do we have a referral?  (We only want to follow a referral if
5942          * we're not following a chain.)
5943          */
5944         if (!negative_response && ns_name != NULL && oqname == NULL) {
5945                 /*
5946                  * We already know ns_name is a subdomain of fctx->domain.
5947                  * If ns_name is equal to fctx->domain, we're not making
5948                  * progress.  We return DNS_R_FORMERR so that we'll keep
5949                  * trying other servers.
5950                  */
5951                 if (dns_name_equal(ns_name, &fctx->domain)) {
5952                         log_formerr(fctx, "non-improving referral");
5953                         return (DNS_R_FORMERR);
5954                 }
5955
5956                 /*
5957                  * If the referral name is not a parent of the query
5958                  * name, consider the responder insane.
5959                  */
5960                 if (! dns_name_issubdomain(&fctx->name, ns_name)) {
5961                         /* Logged twice */
5962                         log_formerr(fctx, "referral to non-parent");
5963                         FCTXTRACE("referral to non-parent");
5964                         return (DNS_R_FORMERR);
5965                 }
5966
5967                 /*
5968                  * Mark any additional data related to this rdataset.
5969                  * It's important that we do this before we change the
5970                  * query domain.
5971                  */
5972                 INSIST(ns_rdataset != NULL);
5973                 fctx->attributes |= FCTX_ATTR_GLUING;
5974                 (void)dns_rdataset_additionaldata(ns_rdataset, check_related,
5975                                                   fctx);
5976 #if CHECK_FOR_GLUE_IN_ANSWER
5977                 /*
5978                  * Look in the answer section for "glue" that is incorrectly
5979                  * returned as a answer.  This is needed if the server also
5980                  * minimizes the response size by not adding records to the
5981                  * additional section that are in the answer section or if
5982                  * the record gets dropped due to message size constraints.
5983                  */
5984                 if ((look_in_options & LOOK_FOR_GLUE_IN_ANSWER) != 0 &&
5985                     (fctx->type == dns_rdatatype_aaaa ||
5986                      fctx->type == dns_rdatatype_a))
5987                         (void)dns_rdataset_additionaldata(ns_rdataset,
5988                                                           check_answer, fctx);
5989 #endif
5990                 fctx->attributes &= ~FCTX_ATTR_GLUING;
5991                 /*
5992                  * NS rdatasets with 0 TTL cause problems.
5993                  * dns_view_findzonecut() will not find them when we
5994                  * try to follow the referral, and we'll SERVFAIL
5995                  * because the best nameservers are now above QDOMAIN.
5996                  * We force the TTL to 1 second to prevent this.
5997                  */
5998                 if (ns_rdataset->ttl == 0)
5999                         ns_rdataset->ttl = 1;
6000                 /*
6001                  * Set the current query domain to the referral name.
6002                  *
6003                  * XXXRTH  We should check if we're in forward-only mode, and
6004                  *              if so we should bail out.
6005                  */
6006                 INSIST(dns_name_countlabels(&fctx->domain) > 0);
6007                 dns_name_free(&fctx->domain, fctx->mctx);
6008                 if (dns_rdataset_isassociated(&fctx->nameservers))
6009                         dns_rdataset_disassociate(&fctx->nameservers);
6010                 dns_name_init(&fctx->domain, NULL);
6011                 result = dns_name_dup(ns_name, fctx->mctx, &fctx->domain);
6012                 if (result != ISC_R_SUCCESS)
6013                         return (result);
6014                 fctx->attributes |= FCTX_ATTR_WANTCACHE;
6015                 fctx->ns_ttl_ok = ISC_FALSE;
6016                 log_ns_ttl(fctx, "DELEGATION");
6017                 return (DNS_R_DELEGATION);
6018         }
6019
6020         /*
6021          * Since we're not doing a referral, we don't want to cache any
6022          * NS RRs we may have found.
6023          */
6024         if (ns_name != NULL)
6025                 ns_name->attributes &= ~DNS_NAMEATTR_CACHE;
6026
6027         if (negative_response && oqname == NULL)
6028                 fctx->attributes |= FCTX_ATTR_WANTNCACHE;
6029
6030         return (ISC_R_SUCCESS);
6031 }
6032
6033 static isc_result_t
6034 answer_response(fetchctx_t *fctx) {
6035         isc_result_t result;
6036         dns_message_t *message;
6037         dns_name_t *name, *qname, tname, *ns_name;
6038         dns_rdataset_t *rdataset, *ns_rdataset;
6039         isc_boolean_t done, external, chaining, aa, found, want_chaining;
6040         isc_boolean_t have_answer, found_cname, found_type, wanted_chaining;
6041         unsigned int aflag;
6042         dns_rdatatype_t type;
6043         dns_fixedname_t dname, fqname;
6044         dns_view_t *view;
6045
6046         FCTXTRACE("answer_response");
6047
6048         message = fctx->rmessage;
6049
6050         /*
6051          * Examine the answer section, marking those rdatasets which are
6052          * part of the answer and should be cached.
6053          */
6054
6055         done = ISC_FALSE;
6056         found_cname = ISC_FALSE;
6057         found_type = ISC_FALSE;
6058         chaining = ISC_FALSE;
6059         have_answer = ISC_FALSE;
6060         want_chaining = ISC_FALSE;
6061         POST(want_chaining);
6062         if ((message->flags & DNS_MESSAGEFLAG_AA) != 0)
6063                 aa = ISC_TRUE;
6064         else
6065                 aa = ISC_FALSE;
6066         qname = &fctx->name;
6067         type = fctx->type;
6068         view = fctx->res->view;
6069         result = dns_message_firstname(message, DNS_SECTION_ANSWER);
6070         while (!done && result == ISC_R_SUCCESS) {
6071                 name = NULL;
6072                 dns_message_currentname(message, DNS_SECTION_ANSWER, &name);
6073                 external = ISC_TF(!dns_name_issubdomain(name, &fctx->domain));
6074                 if (dns_name_equal(name, qname)) {
6075                         wanted_chaining = ISC_FALSE;
6076                         for (rdataset = ISC_LIST_HEAD(name->list);
6077                              rdataset != NULL;
6078                              rdataset = ISC_LIST_NEXT(rdataset, link)) {
6079                                 found = ISC_FALSE;
6080                                 want_chaining = ISC_FALSE;
6081                                 aflag = 0;
6082                                 if (rdataset->type == dns_rdatatype_nsec3) {
6083                                         /*
6084                                          * NSEC3 records are not allowed to
6085                                          * appear in the answer section.
6086                                          */
6087                                         log_formerr(fctx, "NSEC3 in answer");
6088                                         return (DNS_R_FORMERR);
6089                                 }
6090
6091                                 /*
6092                                  * Apply filters, if given, on answers to reject
6093                                  * a malicious attempt of rebinding.
6094                                  */
6095                                 if ((rdataset->type == dns_rdatatype_a ||
6096                                      rdataset->type == dns_rdatatype_aaaa) &&
6097                                     !is_answeraddress_allowed(view, name,
6098                                                               rdataset)) {
6099                                         return (DNS_R_SERVFAIL);
6100                                 }
6101
6102                                 if (rdataset->type == type && !found_cname) {
6103                                         /*
6104                                          * We've found an ordinary answer.
6105                                          */
6106                                         found = ISC_TRUE;
6107                                         found_type = ISC_TRUE;
6108                                         done = ISC_TRUE;
6109                                         aflag = DNS_RDATASETATTR_ANSWER;
6110                                 } else if (type == dns_rdatatype_any) {
6111                                         /*
6112                                          * We've found an answer matching
6113                                          * an ANY query.  There may be
6114                                          * more.
6115                                          */
6116                                         found = ISC_TRUE;
6117                                         aflag = DNS_RDATASETATTR_ANSWER;
6118                                 } else if (rdataset->type == dns_rdatatype_rrsig
6119                                            && rdataset->covers == type
6120                                            && !found_cname) {
6121                                         /*
6122                                          * We've found a signature that
6123                                          * covers the type we're looking for.
6124                                          */
6125                                         found = ISC_TRUE;
6126                                         found_type = ISC_TRUE;
6127                                         aflag = DNS_RDATASETATTR_ANSWERSIG;
6128                                 } else if (rdataset->type ==
6129                                            dns_rdatatype_cname
6130                                            && !found_type) {
6131                                         /*
6132                                          * We're looking for something else,
6133                                          * but we found a CNAME.
6134                                          *
6135                                          * Getting a CNAME response for some
6136                                          * query types is an error, see
6137                                          * RFC 4035, Section 2.5.
6138                                          */
6139                                         if (type == dns_rdatatype_rrsig ||
6140                                             type == dns_rdatatype_key ||
6141                                             type == dns_rdatatype_nsec) {
6142                                                 char buf[DNS_RDATATYPE_FORMATSIZE];
6143                                                 dns_rdatatype_format(fctx->type,
6144                                                               buf, sizeof(buf));
6145                                                 log_formerr(fctx,
6146                                                             "CNAME response "
6147                                                             "for %s RR", buf);
6148                                                 return (DNS_R_FORMERR);
6149                                         }
6150                                         found = ISC_TRUE;
6151                                         found_cname = ISC_TRUE;
6152                                         want_chaining = ISC_TRUE;
6153                                         aflag = DNS_RDATASETATTR_ANSWER;
6154                                         result = cname_target(rdataset,
6155                                                               &tname);
6156                                         if (result != ISC_R_SUCCESS)
6157                                                 return (result);
6158                                         /* Apply filters on the target name. */
6159                                         if (!is_answertarget_allowed(view,
6160                                                         name,
6161                                                         rdataset->type,
6162                                                         &tname,
6163                                                         &fctx->domain)) {
6164                                                 return (DNS_R_SERVFAIL);
6165                                         }
6166                                 } else if (rdataset->type == dns_rdatatype_rrsig
6167                                            && rdataset->covers ==
6168                                            dns_rdatatype_cname
6169                                            && !found_type) {
6170                                         /*
6171                                          * We're looking for something else,
6172                                          * but we found a SIG CNAME.
6173                                          */
6174                                         found = ISC_TRUE;
6175                                         found_cname = ISC_TRUE;
6176                                         aflag = DNS_RDATASETATTR_ANSWERSIG;
6177                                 }
6178
6179                                 if (found) {
6180                                         /*
6181                                          * We've found an answer to our
6182                                          * question.
6183                                          */
6184                                         name->attributes |=
6185                                                 DNS_NAMEATTR_CACHE;
6186                                         rdataset->attributes |=
6187                                                 DNS_RDATASETATTR_CACHE;
6188                                         rdataset->trust = dns_trust_answer;
6189                                         if (!chaining) {
6190                                                 /*
6191                                                  * This data is "the" answer
6192                                                  * to our question only if
6193                                                  * we're not chaining (i.e.
6194                                                  * if we haven't followed
6195                                                  * a CNAME or DNAME).
6196                                                  */
6197                                                 INSIST(!external);
6198                                                 if (aflag ==
6199                                                     DNS_RDATASETATTR_ANSWER)
6200                                                         have_answer = ISC_TRUE;
6201                                                 name->attributes |=
6202                                                         DNS_NAMEATTR_ANSWER;
6203                                                 rdataset->attributes |= aflag;
6204                                                 if (aa)
6205                                                         rdataset->trust =
6206                                                           dns_trust_authanswer;
6207                                         } else if (external) {
6208                                                 /*
6209                                                  * This data is outside of
6210                                                  * our query domain, and
6211                                                  * may not be cached.
6212                                                  */
6213                                                 rdataset->attributes |=
6214                                                     DNS_RDATASETATTR_EXTERNAL;
6215                                         }
6216
6217                                         /*
6218                                          * Mark any additional data related
6219                                          * to this rdataset.
6220                                          */
6221                                         (void)dns_rdataset_additionaldata(
6222                                                         rdataset,
6223                                                         check_related,
6224                                                         fctx);
6225
6226                                         /*
6227                                          * CNAME chaining.
6228                                          */
6229                                         if (want_chaining) {
6230                                                 wanted_chaining = ISC_TRUE;
6231                                                 name->attributes |=
6232                                                         DNS_NAMEATTR_CHAINING;
6233                                                 rdataset->attributes |=
6234                                                     DNS_RDATASETATTR_CHAINING;
6235                                                 qname = &tname;
6236                                         }
6237                                 }
6238                                 /*
6239                                  * We could add an "else" clause here and
6240                                  * log that we're ignoring this rdataset.
6241                                  */
6242                         }
6243                         /*
6244                          * If wanted_chaining is true, we've done
6245                          * some chaining as the result of processing
6246                          * this node, and thus we need to set
6247                          * chaining to true.
6248                          *
6249                          * We don't set chaining inside of the
6250                          * rdataset loop because doing that would
6251                          * cause us to ignore the signatures of
6252                          * CNAMEs.
6253                          */
6254                         if (wanted_chaining)
6255                                 chaining = ISC_TRUE;
6256                 } else {
6257                         /*
6258                          * Look for a DNAME (or its SIG).  Anything else is
6259                          * ignored.
6260                          */
6261                         wanted_chaining = ISC_FALSE;
6262                         for (rdataset = ISC_LIST_HEAD(name->list);
6263                              rdataset != NULL;
6264                              rdataset = ISC_LIST_NEXT(rdataset, link)) {
6265                                 isc_boolean_t found_dname = ISC_FALSE;
6266                                 dns_name_t *dname_name;
6267
6268                                 found = ISC_FALSE;
6269                                 aflag = 0;
6270                                 if (rdataset->type == dns_rdatatype_dname) {
6271                                         /*
6272                                          * We're looking for something else,
6273                                          * but we found a DNAME.
6274                                          *
6275                                          * If we're not chaining, then the
6276                                          * DNAME should not be external.
6277                                          */
6278                                         if (!chaining && external) {
6279                                                 log_formerr(fctx,
6280                                                             "external DNAME");
6281                                                 return (DNS_R_FORMERR);
6282                                         }
6283                                         found = ISC_TRUE;
6284                                         want_chaining = ISC_TRUE;
6285                                         POST(want_chaining);
6286                                         aflag = DNS_RDATASETATTR_ANSWER;
6287                                         result = dname_target(fctx, rdataset,
6288                                                               qname, name,
6289                                                               &dname);
6290                                         if (result == ISC_R_NOSPACE) {
6291                                                 /*
6292                                                  * We can't construct the
6293                                                  * DNAME target.  Do not
6294                                                  * try to continue.
6295                                                  */
6296                                                 want_chaining = ISC_FALSE;
6297                                                 POST(want_chaining);
6298                                         } else if (result != ISC_R_SUCCESS)
6299                                                 return (result);
6300                                         else
6301                                                 found_dname = ISC_TRUE;
6302
6303                                         dname_name = dns_fixedname_name(&dname);
6304                                         if (!is_answertarget_allowed(view,
6305                                                         qname,
6306                                                         rdataset->type,
6307                                                         dname_name,
6308                                                         &fctx->domain)) {
6309                                                 return (DNS_R_SERVFAIL);
6310                                         }
6311                                 } else if (rdataset->type == dns_rdatatype_rrsig
6312                                            && rdataset->covers ==
6313                                            dns_rdatatype_dname) {
6314                                         /*
6315                                          * We've found a signature that
6316                                          * covers the DNAME.
6317                                          */
6318                                         found = ISC_TRUE;
6319                                         aflag = DNS_RDATASETATTR_ANSWERSIG;
6320                                 }
6321
6322                                 if (found) {
6323                                         /*
6324                                          * We've found an answer to our
6325                                          * question.
6326                                          */
6327                                         name->attributes |=
6328                                                 DNS_NAMEATTR_CACHE;
6329                                         rdataset->attributes |=
6330                                                 DNS_RDATASETATTR_CACHE;
6331                                         rdataset->trust = dns_trust_answer;
6332                                         if (!chaining) {
6333                                                 /*
6334                                                  * This data is "the" answer
6335                                                  * to our question only if
6336                                                  * we're not chaining.
6337                                                  */
6338                                                 INSIST(!external);
6339                                                 if (aflag ==
6340                                                     DNS_RDATASETATTR_ANSWER)
6341                                                         have_answer = ISC_TRUE;
6342                                                 name->attributes |=
6343                                                         DNS_NAMEATTR_ANSWER;
6344                                                 rdataset->attributes |= aflag;
6345                                                 if (aa)
6346                                                         rdataset->trust =
6347                                                           dns_trust_authanswer;
6348                                         } else if (external) {
6349                                                 rdataset->attributes |=
6350                                                     DNS_RDATASETATTR_EXTERNAL;
6351                                         }
6352
6353                                         /*
6354                                          * DNAME chaining.
6355                                          */
6356                                         if (found_dname) {
6357                                                 /*
6358                                                  * Copy the dname into the
6359                                                  * qname fixed name.
6360                                                  *
6361                                                  * Although we check for
6362                                                  * failure of the copy
6363                                                  * operation, in practice it
6364                                                  * should never fail since
6365                                                  * we already know that the
6366                                                  * result fits in a fixedname.
6367                                                  */
6368                                                 dns_fixedname_init(&fqname);
6369                                                 result = dns_name_copy(
6370                                                   dns_fixedname_name(&dname),
6371                                                   dns_fixedname_name(&fqname),
6372                                                   NULL);
6373                                                 if (result != ISC_R_SUCCESS)
6374                                                         return (result);
6375                                                 wanted_chaining = ISC_TRUE;
6376                                                 name->attributes |=
6377                                                         DNS_NAMEATTR_CHAINING;
6378                                                 rdataset->attributes |=
6379                                                     DNS_RDATASETATTR_CHAINING;
6380                                                 qname = dns_fixedname_name(
6381                                                                    &fqname);
6382                                         }
6383                                 }
6384                         }
6385                         if (wanted_chaining)
6386                                 chaining = ISC_TRUE;
6387                 }
6388                 result = dns_message_nextname(message, DNS_SECTION_ANSWER);
6389         }
6390         if (result == ISC_R_NOMORE)
6391                 result = ISC_R_SUCCESS;
6392         if (result != ISC_R_SUCCESS)
6393                 return (result);
6394
6395         /*
6396          * We should have found an answer.
6397          */
6398         if (!have_answer) {
6399                 log_formerr(fctx, "reply has no answer");
6400                 return (DNS_R_FORMERR);
6401         }
6402
6403         /*
6404          * This response is now potentially cacheable.
6405          */
6406         fctx->attributes |= FCTX_ATTR_WANTCACHE;
6407
6408         /*
6409          * Did chaining end before we got the final answer?
6410          */
6411         if (chaining) {
6412                 /*
6413                  * Yes.  This may be a negative reply, so hand off
6414                  * authority section processing to the noanswer code.
6415                  * If it isn't a noanswer response, no harm will be
6416                  * done.
6417                  */
6418                 return (noanswer_response(fctx, qname, 0));
6419         }
6420
6421         /*
6422          * We didn't end with an incomplete chain, so the rcode should be
6423          * "no error".
6424          */
6425         if (message->rcode != dns_rcode_noerror) {
6426                 log_formerr(fctx, "CNAME/DNAME chain complete, but RCODE "
6427                                   "indicates error");
6428                 return (DNS_R_FORMERR);
6429         }
6430
6431         /*
6432          * Examine the authority section (if there is one).
6433          *
6434          * We expect there to be only one owner name for all the rdatasets
6435          * in this section, and we expect that it is not external.
6436          */
6437         done = ISC_FALSE;
6438         ns_name = NULL;
6439         ns_rdataset = NULL;
6440         result = dns_message_firstname(message, DNS_SECTION_AUTHORITY);
6441         while (!done && result == ISC_R_SUCCESS) {
6442                 name = NULL;
6443                 dns_message_currentname(message, DNS_SECTION_AUTHORITY, &name);
6444                 external = ISC_TF(!dns_name_issubdomain(name, &fctx->domain));
6445                 if (!external) {
6446                         /*
6447                          * We expect to find NS or SIG NS rdatasets, and
6448                          * nothing else.
6449                          */
6450                         for (rdataset = ISC_LIST_HEAD(name->list);
6451                              rdataset != NULL;
6452                              rdataset = ISC_LIST_NEXT(rdataset, link)) {
6453                                 if (rdataset->type == dns_rdatatype_ns ||
6454                                     (rdataset->type == dns_rdatatype_rrsig &&
6455                                      rdataset->covers == dns_rdatatype_ns)) {
6456                                         name->attributes |=
6457                                                 DNS_NAMEATTR_CACHE;
6458                                         rdataset->attributes |=
6459                                                 DNS_RDATASETATTR_CACHE;
6460                                         if (aa && !chaining)
6461                                                 rdataset->trust =
6462                                                     dns_trust_authauthority;
6463                                         else
6464                                                 rdataset->trust =
6465                                                     dns_trust_additional;
6466
6467                                         if (rdataset->type == dns_rdatatype_ns) {
6468                                                 ns_name = name;
6469                                                 ns_rdataset = rdataset;
6470                                         }
6471                                         /*
6472                                          * Mark any additional data related
6473                                          * to this rdataset.
6474                                          */
6475                                         (void)dns_rdataset_additionaldata(
6476                                                         rdataset,
6477                                                         check_related,
6478                                                         fctx);
6479                                         done = ISC_TRUE;
6480                                 }
6481                         }
6482                 }
6483                 result = dns_message_nextname(message, DNS_SECTION_AUTHORITY);
6484         }
6485         if (result == ISC_R_NOMORE)
6486                 result = ISC_R_SUCCESS;
6487
6488         log_ns_ttl(fctx, "answer_response");
6489
6490         if (ns_rdataset != NULL && dns_name_equal(&fctx->domain, ns_name) &&
6491             !dns_name_equal(ns_name, dns_rootname))
6492                 trim_ns_ttl(fctx, ns_name, ns_rdataset);
6493
6494         return (result);
6495 }
6496
6497 static isc_boolean_t
6498 fctx_decreference(fetchctx_t *fctx) {
6499         isc_boolean_t bucket_empty = ISC_FALSE;
6500
6501         INSIST(fctx->references > 0);
6502         fctx->references--;
6503         if (fctx->references == 0) {
6504                 /*
6505                  * No one cares about the result of this fetch anymore.
6506                  */
6507                 if (fctx->pending == 0 && fctx->nqueries == 0 &&
6508                     ISC_LIST_EMPTY(fctx->validators) && SHUTTINGDOWN(fctx)) {
6509                         /*
6510                          * This fctx is already shutdown; we were just
6511                          * waiting for the last reference to go away.
6512                          */
6513                         bucket_empty = fctx_unlink(fctx);
6514                         fctx_destroy(fctx);
6515                 } else {
6516                         /*
6517                          * Initiate shutdown.
6518                          */
6519                         fctx_shutdown(fctx);
6520                 }
6521         }
6522         return (bucket_empty);
6523 }
6524
6525 static void
6526 resume_dslookup(isc_task_t *task, isc_event_t *event) {
6527         dns_fetchevent_t *fevent;
6528         dns_resolver_t *res;
6529         fetchctx_t *fctx;
6530         isc_result_t result;
6531         isc_boolean_t bucket_empty;
6532         isc_boolean_t locked = ISC_FALSE;
6533         unsigned int bucketnum;
6534         dns_rdataset_t nameservers;
6535         dns_fixedname_t fixed;
6536         dns_name_t *domain;
6537
6538         REQUIRE(event->ev_type == DNS_EVENT_FETCHDONE);
6539         fevent = (dns_fetchevent_t *)event;
6540         fctx = event->ev_arg;
6541         REQUIRE(VALID_FCTX(fctx));
6542         res = fctx->res;
6543
6544         UNUSED(task);
6545         FCTXTRACE("resume_dslookup");
6546
6547         if (fevent->node != NULL)
6548                 dns_db_detachnode(fevent->db, &fevent->node);
6549         if (fevent->db != NULL)
6550                 dns_db_detach(&fevent->db);
6551
6552         dns_rdataset_init(&nameservers);
6553
6554         bucketnum = fctx->bucketnum;
6555         if (fevent->result == ISC_R_CANCELED) {
6556                 dns_resolver_destroyfetch(&fctx->nsfetch);
6557                 fctx_done(fctx, ISC_R_CANCELED, __LINE__);
6558         } else if (fevent->result == ISC_R_SUCCESS) {
6559
6560                 FCTXTRACE("resuming DS lookup");
6561
6562                 dns_resolver_destroyfetch(&fctx->nsfetch);
6563                 if (dns_rdataset_isassociated(&fctx->nameservers))
6564                         dns_rdataset_disassociate(&fctx->nameservers);
6565                 dns_rdataset_clone(fevent->rdataset, &fctx->nameservers);
6566                 fctx->ns_ttl = fctx->nameservers.ttl;
6567                 fctx->ns_ttl_ok = ISC_TRUE;
6568                 log_ns_ttl(fctx, "resume_dslookup");
6569                 dns_name_free(&fctx->domain, fctx->mctx);
6570                 dns_name_init(&fctx->domain, NULL);
6571                 result = dns_name_dup(&fctx->nsname, fctx->mctx, &fctx->domain);
6572                 if (result != ISC_R_SUCCESS) {
6573                         fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
6574                         goto cleanup;
6575                 }
6576                 /*
6577                  * Try again.
6578                  */
6579                 fctx_try(fctx, ISC_TRUE, ISC_FALSE);
6580         } else {
6581                 unsigned int n;
6582                 dns_rdataset_t *nsrdataset = NULL;
6583
6584                 /*
6585                  * Retrieve state from fctx->nsfetch before we destroy it.
6586                  */
6587                 dns_fixedname_init(&fixed);
6588                 domain = dns_fixedname_name(&fixed);
6589                 dns_name_copy(&fctx->nsfetch->private->domain, domain, NULL);
6590                 if (dns_name_equal(&fctx->nsname, domain)) {
6591                         fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
6592                         dns_resolver_destroyfetch(&fctx->nsfetch);
6593                         goto cleanup;
6594                 }
6595                 if (dns_rdataset_isassociated(
6596                     &fctx->nsfetch->private->nameservers)) {
6597                         dns_rdataset_clone(
6598                             &fctx->nsfetch->private->nameservers,
6599                             &nameservers);
6600                         nsrdataset = &nameservers;
6601                 } else
6602                         domain = NULL;
6603                 dns_resolver_destroyfetch(&fctx->nsfetch);
6604                 n = dns_name_countlabels(&fctx->nsname);
6605                 dns_name_getlabelsequence(&fctx->nsname, 1, n - 1,
6606                                           &fctx->nsname);
6607
6608                 if (dns_rdataset_isassociated(fevent->rdataset))
6609                         dns_rdataset_disassociate(fevent->rdataset);
6610                 FCTXTRACE("continuing to look for parent's NS records");
6611                 result = dns_resolver_createfetch(fctx->res, &fctx->nsname,
6612                                                   dns_rdatatype_ns, domain,
6613                                                   nsrdataset, NULL, 0, task,
6614                                                   resume_dslookup, fctx,
6615                                                   &fctx->nsrrset, NULL,
6616                                                   &fctx->nsfetch);
6617                 if (result != ISC_R_SUCCESS)
6618                         fctx_done(fctx, result, __LINE__);
6619                 else {
6620                         LOCK(&res->buckets[bucketnum].lock);
6621                         locked = ISC_TRUE;
6622                         fctx->references++;
6623                 }
6624         }
6625
6626  cleanup:
6627         if (dns_rdataset_isassociated(&nameservers))
6628                 dns_rdataset_disassociate(&nameservers);
6629         if (dns_rdataset_isassociated(fevent->rdataset))
6630                 dns_rdataset_disassociate(fevent->rdataset);
6631         INSIST(fevent->sigrdataset == NULL);
6632         isc_event_free(&event);
6633         if (!locked)
6634                 LOCK(&res->buckets[bucketnum].lock);
6635         bucket_empty = fctx_decreference(fctx);
6636         UNLOCK(&res->buckets[bucketnum].lock);
6637         if (bucket_empty)
6638                 empty_bucket(res);
6639 }
6640
6641 static inline void
6642 checknamessection(dns_message_t *message, dns_section_t section) {
6643         isc_result_t result;
6644         dns_name_t *name;
6645         dns_rdata_t rdata = DNS_RDATA_INIT;
6646         dns_rdataset_t *rdataset;
6647
6648         for (result = dns_message_firstname(message, section);
6649              result == ISC_R_SUCCESS;
6650              result = dns_message_nextname(message, section))
6651         {
6652                 name = NULL;
6653                 dns_message_currentname(message, section, &name);
6654                 for (rdataset = ISC_LIST_HEAD(name->list);
6655                      rdataset != NULL;
6656                      rdataset = ISC_LIST_NEXT(rdataset, link)) {
6657                         for (result = dns_rdataset_first(rdataset);
6658                              result == ISC_R_SUCCESS;
6659                              result = dns_rdataset_next(rdataset)) {
6660                                 dns_rdataset_current(rdataset, &rdata);
6661                                 if (!dns_rdata_checkowner(name, rdata.rdclass,
6662                                                           rdata.type,
6663                                                           ISC_FALSE) ||
6664                                     !dns_rdata_checknames(&rdata, name, NULL))
6665                                 {
6666                                         rdataset->attributes |=
6667                                                 DNS_RDATASETATTR_CHECKNAMES;
6668                                 }
6669                                 dns_rdata_reset(&rdata);
6670                         }
6671                 }
6672         }
6673 }
6674
6675 static void
6676 checknames(dns_message_t *message) {
6677
6678         checknamessection(message, DNS_SECTION_ANSWER);
6679         checknamessection(message, DNS_SECTION_AUTHORITY);
6680         checknamessection(message, DNS_SECTION_ADDITIONAL);
6681 }
6682
6683 /*
6684  * Log server NSID at log level 'level'
6685  */
6686 static void
6687 log_nsid(isc_buffer_t *opt, size_t nsid_len, resquery_t *query,
6688          int level, isc_mem_t *mctx)
6689 {
6690         static const char hex[17] = "0123456789abcdef";
6691         char addrbuf[ISC_SOCKADDR_FORMATSIZE];
6692         isc_uint16_t buflen, i;
6693         unsigned char *p, *buf, *nsid;
6694
6695         /* Allocate buffer for storing hex version of the NSID */
6696         buflen = (isc_uint16_t)nsid_len * 2 + 1;
6697         buf = isc_mem_get(mctx, buflen);
6698         if (buf == NULL)
6699                 return;
6700
6701         /* Convert to hex */
6702         p = buf;
6703         nsid = isc_buffer_current(opt);
6704         for (i = 0; i < nsid_len; i++) {
6705                 *p++ = hex[(nsid[0] >> 4) & 0xf];
6706                 *p++ = hex[nsid[0] & 0xf];
6707                 nsid++;
6708         }
6709         *p = '\0';
6710
6711         isc_sockaddr_format(&query->addrinfo->sockaddr, addrbuf,
6712                             sizeof(addrbuf));
6713         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
6714                       DNS_LOGMODULE_RESOLVER, level,
6715                       "received NSID '%s' from %s", buf, addrbuf);
6716
6717         /* Clean up */
6718         isc_mem_put(mctx, buf, buflen);
6719         return;
6720 }
6721
6722 static void
6723 log_packet(dns_message_t *message, int level, isc_mem_t *mctx) {
6724         isc_buffer_t buffer;
6725         char *buf = NULL;
6726         int len = 1024;
6727         isc_result_t result;
6728
6729         if (! isc_log_wouldlog(dns_lctx, level))
6730                 return;
6731
6732         /*
6733          * Note that these are multiline debug messages.  We want a newline
6734          * to appear in the log after each message.
6735          */
6736
6737         do {
6738                 buf = isc_mem_get(mctx, len);
6739                 if (buf == NULL)
6740                         break;
6741                 isc_buffer_init(&buffer, buf, len);
6742                 result = dns_message_totext(message, &dns_master_style_debug,
6743                                             0, &buffer);
6744                 if (result == ISC_R_NOSPACE) {
6745                         isc_mem_put(mctx, buf, len);
6746                         len += 1024;
6747                 } else if (result == ISC_R_SUCCESS)
6748                         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
6749                                       DNS_LOGMODULE_RESOLVER, level,
6750                                       "received packet:\n%.*s",
6751                                       (int)isc_buffer_usedlength(&buffer),
6752                                       buf);
6753         } while (result == ISC_R_NOSPACE);
6754
6755         if (buf != NULL)
6756                 isc_mem_put(mctx, buf, len);
6757 }
6758
6759 static isc_boolean_t
6760 iscname(fetchctx_t *fctx) {
6761         isc_result_t result;
6762
6763         result = dns_message_findname(fctx->rmessage, DNS_SECTION_ANSWER,
6764                                       &fctx->name, dns_rdatatype_cname, 0,
6765                                       NULL, NULL);
6766         return (result == ISC_R_SUCCESS ? ISC_TRUE : ISC_FALSE);
6767 }
6768
6769 static isc_boolean_t
6770 betterreferral(fetchctx_t *fctx) {
6771         isc_result_t result;
6772         dns_name_t *name;
6773         dns_rdataset_t *rdataset;
6774         dns_message_t *message = fctx->rmessage;
6775
6776         for (result = dns_message_firstname(message, DNS_SECTION_AUTHORITY);
6777              result == ISC_R_SUCCESS;
6778              result = dns_message_nextname(message, DNS_SECTION_AUTHORITY)) {
6779                 name = NULL;
6780                 dns_message_currentname(message, DNS_SECTION_AUTHORITY, &name);
6781                 if (!isstrictsubdomain(name, &fctx->domain))
6782                         continue;
6783                 for (rdataset = ISC_LIST_HEAD(name->list);
6784                      rdataset != NULL;
6785                      rdataset = ISC_LIST_NEXT(rdataset, link))
6786                         if (rdataset->type == dns_rdatatype_ns)
6787                                 return (ISC_TRUE);
6788         }
6789         return (ISC_FALSE);
6790 }
6791
6792 static void
6793 process_opt(resquery_t *query, dns_rdataset_t *opt) {
6794         dns_rdata_t rdata;
6795         isc_buffer_t optbuf;
6796         isc_result_t result;
6797         isc_uint16_t optcode;
6798         isc_uint16_t optlen;
6799
6800         result = dns_rdataset_first(opt);
6801         if (result == ISC_R_SUCCESS) {
6802                 dns_rdata_init(&rdata);
6803                 dns_rdataset_current(opt, &rdata);
6804                 isc_buffer_init(&optbuf, rdata.data, rdata.length);
6805                 isc_buffer_add(&optbuf, rdata.length);
6806                 while (isc_buffer_remaininglength(&optbuf) >= 4) {
6807                         optcode = isc_buffer_getuint16(&optbuf);
6808                         optlen = isc_buffer_getuint16(&optbuf);
6809                         INSIST(optlen <= isc_buffer_remaininglength(&optbuf));
6810                         switch (optcode) {
6811                         case DNS_OPT_NSID:
6812                                 if (query->options & DNS_FETCHOPT_WANTNSID)
6813                                         log_nsid(&optbuf, optlen, query,
6814                                                  ISC_LOG_INFO,
6815                                                  query->fctx->res->mctx);
6816                                 isc_buffer_forward(&optbuf, optlen);
6817                                 break;
6818                         default:
6819                                 isc_buffer_forward(&optbuf, optlen);
6820                                 break;
6821                         }
6822                 }
6823                 INSIST(isc_buffer_remaininglength(&optbuf) == 0U);
6824         }
6825 }
6826
6827 static void
6828 resquery_response(isc_task_t *task, isc_event_t *event) {
6829         isc_result_t result = ISC_R_SUCCESS;
6830         resquery_t *query = event->ev_arg;
6831         dns_dispatchevent_t *devent = (dns_dispatchevent_t *)event;
6832         isc_boolean_t keep_trying, get_nameservers, resend;
6833         isc_boolean_t truncated;
6834         dns_message_t *message;
6835         dns_rdataset_t *opt;
6836         fetchctx_t *fctx;
6837         dns_name_t *fname;
6838         dns_fixedname_t foundname;
6839         isc_stdtime_t now;
6840         isc_time_t tnow, *finish;
6841         dns_adbaddrinfo_t *addrinfo;
6842         unsigned int options;
6843         unsigned int findoptions;
6844         isc_result_t broken_server;
6845         badnstype_t broken_type = badns_response;
6846         isc_boolean_t no_response;
6847
6848         REQUIRE(VALID_QUERY(query));
6849         fctx = query->fctx;
6850         options = query->options;
6851         REQUIRE(VALID_FCTX(fctx));
6852         REQUIRE(event->ev_type == DNS_EVENT_DISPATCH);
6853
6854         QTRACE("response");
6855
6856         if (isc_sockaddr_pf(&query->addrinfo->sockaddr) == PF_INET)
6857                 inc_stats(fctx->res, dns_resstatscounter_responsev4);
6858         else
6859                 inc_stats(fctx->res, dns_resstatscounter_responsev6);
6860
6861         (void)isc_timer_touch(fctx->timer);
6862
6863         keep_trying = ISC_FALSE;
6864         broken_server = ISC_R_SUCCESS;
6865         get_nameservers = ISC_FALSE;
6866         resend = ISC_FALSE;
6867         truncated = ISC_FALSE;
6868         finish = NULL;
6869         no_response = ISC_FALSE;
6870
6871         if (fctx->res->exiting) {
6872                 result = ISC_R_SHUTTINGDOWN;
6873                 goto done;
6874         }
6875
6876         fctx->timeouts = 0;
6877         fctx->timeout = ISC_FALSE;
6878         fctx->addrinfo = query->addrinfo;
6879
6880         /*
6881          * XXXRTH  We should really get the current time just once.  We
6882          *              need a routine to convert from an isc_time_t to an
6883          *              isc_stdtime_t.
6884          */
6885         TIME_NOW(&tnow);
6886         finish = &tnow;
6887         isc_stdtime_get(&now);
6888
6889         /*
6890          * Did the dispatcher have a problem?
6891          */
6892         if (devent->result != ISC_R_SUCCESS) {
6893                 if (devent->result == ISC_R_EOF &&
6894                     (query->options & DNS_FETCHOPT_NOEDNS0) == 0) {
6895                         /*
6896                          * The problem might be that they
6897                          * don't understand EDNS0.  Turn it
6898                          * off and try again.
6899                          */
6900                         options |= DNS_FETCHOPT_NOEDNS0;
6901                         resend = ISC_TRUE;
6902                         add_bad_edns(fctx, &query->addrinfo->sockaddr);
6903                 } else {
6904                         /*
6905                          * There's no hope for this query.
6906                          */
6907                         keep_trying = ISC_TRUE;
6908
6909                         /*
6910                          * If this is a network error on an exclusive query
6911                          * socket, mark the server as bad so that we won't try
6912                          * it for this fetch again.  Also adjust finish and
6913                          * no_response so that we penalize this address in SRTT
6914                          * adjustment later.
6915                          */
6916                         if (query->exclusivesocket &&
6917                             (devent->result == ISC_R_HOSTUNREACH ||
6918                              devent->result == ISC_R_NETUNREACH ||
6919                              devent->result == ISC_R_CONNREFUSED ||
6920                              devent->result == ISC_R_CANCELED)) {
6921                                     broken_server = devent->result;
6922                                     broken_type = badns_unreachable;
6923                                     finish = NULL;
6924                                     no_response = ISC_TRUE;
6925                         }
6926                 }
6927                 goto done;
6928         }
6929
6930         message = fctx->rmessage;
6931
6932         if (query->tsig != NULL) {
6933                 result = dns_message_setquerytsig(message, query->tsig);
6934                 if (result != ISC_R_SUCCESS)
6935                         goto done;
6936         }
6937
6938         if (query->tsigkey) {
6939                 result = dns_message_settsigkey(message, query->tsigkey);
6940                 if (result != ISC_R_SUCCESS)
6941                         goto done;
6942         }
6943
6944         dns_message_setclass(message, fctx->res->rdclass);
6945
6946         result = dns_message_parse(message, &devent->buffer, 0);
6947         if (result != ISC_R_SUCCESS) {
6948                 switch (result) {
6949                 case ISC_R_UNEXPECTEDEND:
6950                         if (!message->question_ok ||
6951                             (message->flags & DNS_MESSAGEFLAG_TC) == 0 ||
6952                             (options & DNS_FETCHOPT_TCP) != 0) {
6953                                 /*
6954                                  * Either the message ended prematurely,
6955                                  * and/or wasn't marked as being truncated,
6956                                  * and/or this is a response to a query we
6957                                  * sent over TCP.  In all of these cases,
6958                                  * something is wrong with the remote
6959                                  * server and we don't want to retry using
6960                                  * TCP.
6961                                  */
6962                                 if ((query->options & DNS_FETCHOPT_NOEDNS0)
6963                                     == 0) {
6964                                         /*
6965                                          * The problem might be that they
6966                                          * don't understand EDNS0.  Turn it
6967                                          * off and try again.
6968                                          */
6969                                         options |= DNS_FETCHOPT_NOEDNS0;
6970                                         resend = ISC_TRUE;
6971                                         add_bad_edns(fctx,
6972                                                     &query->addrinfo->sockaddr);
6973                                         inc_stats(fctx->res,
6974                                                  dns_resstatscounter_edns0fail);
6975                                 } else {
6976                                         broken_server = result;
6977                                         keep_trying = ISC_TRUE;
6978                                 }
6979                                 goto done;
6980                         }
6981                         /*
6982                          * We defer retrying via TCP for a bit so we can
6983                          * check out this message further.
6984                          */
6985                         truncated = ISC_TRUE;
6986                         break;
6987                 case DNS_R_FORMERR:
6988                         if ((query->options & DNS_FETCHOPT_NOEDNS0) == 0) {
6989                                 /*
6990                                  * The problem might be that they
6991                                  * don't understand EDNS0.  Turn it
6992                                  * off and try again.
6993                                  */
6994                                 options |= DNS_FETCHOPT_NOEDNS0;
6995                                 resend = ISC_TRUE;
6996                                 add_bad_edns(fctx, &query->addrinfo->sockaddr);
6997                                 inc_stats(fctx->res,
6998                                                  dns_resstatscounter_edns0fail);
6999                         } else {
7000                                 broken_server = DNS_R_UNEXPECTEDRCODE;
7001                                 keep_trying = ISC_TRUE;
7002                         }
7003                         goto done;
7004                 default:
7005                         /*
7006                          * Something bad has happened.
7007                          */
7008                         goto done;
7009                 }
7010         }
7011
7012
7013         /*
7014          * Log the incoming packet.
7015          */
7016         log_packet(message, ISC_LOG_DEBUG(10), fctx->res->mctx);
7017
7018         if (message->rdclass != fctx->res->rdclass) {
7019                 resend = ISC_TRUE;
7020                 FCTXTRACE("bad class");
7021                 goto done;
7022         }
7023
7024         /*
7025          * Process receive opt record.
7026          */
7027         opt = dns_message_getopt(message);
7028         if (opt != NULL)
7029                 process_opt(query, opt);
7030
7031         /*
7032          * If the message is signed, check the signature.  If not, this
7033          * returns success anyway.
7034          */
7035         result = dns_message_checksig(message, fctx->res->view);
7036         if (result != ISC_R_SUCCESS)
7037                 goto done;
7038
7039         /*
7040          * The dispatcher should ensure we only get responses with QR set.
7041          */
7042         INSIST((message->flags & DNS_MESSAGEFLAG_QR) != 0);
7043         /*
7044          * INSIST() that the message comes from the place we sent it to,
7045          * since the dispatch code should ensure this.
7046          *
7047          * INSIST() that the message id is correct (this should also be
7048          * ensured by the dispatch code).
7049          */
7050
7051         /*
7052          * We have an affirmative response to the query and we have
7053          * previously got a response from this server which indicated
7054          * EDNS may not be supported so we can now cache the lack of
7055          * EDNS support.
7056          */
7057         if (opt == NULL &&
7058             (message->rcode == dns_rcode_noerror ||
7059              message->rcode == dns_rcode_nxdomain ||
7060              message->rcode == dns_rcode_refused ||
7061              message->rcode == dns_rcode_yxdomain) &&
7062              bad_edns(fctx, &query->addrinfo->sockaddr)) {
7063                 char addrbuf[ISC_SOCKADDR_FORMATSIZE];
7064                 isc_sockaddr_format(&query->addrinfo->sockaddr, addrbuf,
7065                                     sizeof(addrbuf));
7066                 dns_adb_changeflags(fctx->adb, query->addrinfo,
7067                                     DNS_FETCHOPT_NOEDNS0,
7068                                     DNS_FETCHOPT_NOEDNS0);
7069         }
7070
7071         /*
7072          * Deal with truncated responses by retrying using TCP.
7073          */
7074         if ((message->flags & DNS_MESSAGEFLAG_TC) != 0)
7075                 truncated = ISC_TRUE;
7076
7077         if (truncated) {
7078                 inc_stats(fctx->res, dns_resstatscounter_truncated);
7079                 if ((options & DNS_FETCHOPT_TCP) != 0) {
7080                         broken_server = DNS_R_TRUNCATEDTCP;
7081                         keep_trying = ISC_TRUE;
7082                 } else {
7083                         options |= DNS_FETCHOPT_TCP;
7084                         resend = ISC_TRUE;
7085                 }
7086                 goto done;
7087         }
7088
7089         /*
7090          * Is it a query response?
7091          */
7092         if (message->opcode != dns_opcode_query) {
7093                 /* XXXRTH Log */
7094                 broken_server = DNS_R_UNEXPECTEDOPCODE;
7095                 keep_trying = ISC_TRUE;
7096                 goto done;
7097         }
7098
7099         /*
7100          * Update statistics about erroneous responses.
7101          */
7102         if (message->rcode != dns_rcode_noerror) {
7103                 switch (message->rcode) {
7104                 case dns_rcode_nxdomain:
7105                         inc_stats(fctx->res, dns_resstatscounter_nxdomain);
7106                         break;
7107                 case dns_rcode_servfail:
7108                         inc_stats(fctx->res, dns_resstatscounter_servfail);
7109                         break;
7110                 case dns_rcode_formerr:
7111                         inc_stats(fctx->res, dns_resstatscounter_formerr);
7112                         break;
7113                 default:
7114                         inc_stats(fctx->res, dns_resstatscounter_othererror);
7115                         break;
7116                 }
7117         }
7118
7119         /*
7120          * Is the remote server broken, or does it dislike us?
7121          */
7122         if (message->rcode != dns_rcode_noerror &&
7123             message->rcode != dns_rcode_nxdomain) {
7124                 if (((message->rcode == dns_rcode_formerr ||
7125                       message->rcode == dns_rcode_notimp) ||
7126                      (message->rcode == dns_rcode_servfail &&
7127                       dns_message_getopt(message) == NULL)) &&
7128                     (query->options & DNS_FETCHOPT_NOEDNS0) == 0) {
7129                         /*
7130                          * It's very likely they don't like EDNS0.
7131                          * If the response code is SERVFAIL, also check if the
7132                          * response contains an OPT RR and don't cache the
7133                          * failure since it can be returned for various other
7134                          * reasons.
7135                          *
7136                          * XXXRTH  We should check if the question
7137                          *              we're asking requires EDNS0, and
7138                          *              if so, we should bail out.
7139                          */
7140                         options |= DNS_FETCHOPT_NOEDNS0;
7141                         resend = ISC_TRUE;
7142                         /*
7143                          * Remember that they may not like EDNS0.
7144                          */
7145                         add_bad_edns(fctx, &query->addrinfo->sockaddr);
7146                         inc_stats(fctx->res, dns_resstatscounter_edns0fail);
7147                 } else if (message->rcode == dns_rcode_formerr) {
7148                         if (ISFORWARDER(query->addrinfo)) {
7149                                 /*
7150                                  * This forwarder doesn't understand us,
7151                                  * but other forwarders might.  Keep trying.
7152                                  */
7153                                 broken_server = DNS_R_REMOTEFORMERR;
7154                                 keep_trying = ISC_TRUE;
7155                         } else {
7156                                 /*
7157                                  * The server doesn't understand us.  Since
7158                                  * all servers for a zone need similar
7159                                  * capabilities, we assume that we will get
7160                                  * FORMERR from all servers, and thus we
7161                                  * cannot make any more progress with this
7162                                  * fetch.
7163                                  */
7164                                 log_formerr(fctx, "server sent FORMERR");
7165                                 result = DNS_R_FORMERR;
7166                         }
7167                 } else if (message->rcode == dns_rcode_yxdomain) {
7168                         /*
7169                          * DNAME mapping failed because the new name
7170                          * was too long.  There's no chance of success
7171                          * for this fetch.
7172                          */
7173                         result = DNS_R_YXDOMAIN;
7174                 } else if (message->rcode == dns_rcode_badvers) {
7175                         unsigned int flags, mask;
7176                         unsigned int version;
7177
7178                         resend = ISC_TRUE;
7179                         INSIST(opt != NULL);
7180                         version = (opt->ttl >> 16) & 0xff;
7181                         flags = (version << DNS_FETCHOPT_EDNSVERSIONSHIFT) |
7182                                 DNS_FETCHOPT_EDNSVERSIONSET;
7183                         mask = DNS_FETCHOPT_EDNSVERSIONMASK |
7184                                DNS_FETCHOPT_EDNSVERSIONSET;
7185                         switch (version) {
7186                         case 0:
7187                                 dns_adb_changeflags(fctx->adb, query->addrinfo,
7188                                                     flags, mask);
7189                                 break;
7190                         default:
7191                                 broken_server = DNS_R_BADVERS;
7192                                 keep_trying = ISC_TRUE;
7193                                 break;
7194                         }
7195                 } else {
7196                         /*
7197                          * XXXRTH log.
7198                          */
7199                         broken_server = DNS_R_UNEXPECTEDRCODE;
7200                         INSIST(broken_server != ISC_R_SUCCESS);
7201                         keep_trying = ISC_TRUE;
7202                 }
7203                 goto done;
7204         }
7205
7206         /*
7207          * Is the question the same as the one we asked?
7208          */
7209         result = same_question(fctx);
7210         if (result != ISC_R_SUCCESS) {
7211                 /* XXXRTH Log */
7212                 if (result == DNS_R_FORMERR)
7213                         keep_trying = ISC_TRUE;
7214                 goto done;
7215         }
7216
7217         /*
7218          * Is the server lame?
7219          */
7220         if (fctx->res->lame_ttl != 0 && !ISFORWARDER(query->addrinfo) &&
7221             is_lame(fctx)) {
7222                 inc_stats(fctx->res, dns_resstatscounter_lame);
7223                 log_lame(fctx, query->addrinfo);
7224                 result = dns_adb_marklame(fctx->adb, query->addrinfo,
7225                                           &fctx->name, fctx->type,
7226                                           now + fctx->res->lame_ttl);
7227                 if (result != ISC_R_SUCCESS)
7228                         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
7229                                       DNS_LOGMODULE_RESOLVER, ISC_LOG_ERROR,
7230                                       "could not mark server as lame: %s",
7231                                       isc_result_totext(result));
7232                 broken_server = DNS_R_LAME;
7233                 keep_trying = ISC_TRUE;
7234                 goto done;
7235         }
7236
7237         /*
7238          * Enforce delegations only zones like NET and COM.
7239          */
7240         if (!ISFORWARDER(query->addrinfo) &&
7241             dns_view_isdelegationonly(fctx->res->view, &fctx->domain) &&
7242             !dns_name_equal(&fctx->domain, &fctx->name) &&
7243             fix_mustbedelegationornxdomain(message, fctx)) {
7244                 char namebuf[DNS_NAME_FORMATSIZE];
7245                 char domainbuf[DNS_NAME_FORMATSIZE];
7246                 char addrbuf[ISC_SOCKADDR_FORMATSIZE];
7247                 char classbuf[64];
7248                 char typebuf[64];
7249
7250                 dns_name_format(&fctx->name, namebuf, sizeof(namebuf));
7251                 dns_name_format(&fctx->domain, domainbuf, sizeof(domainbuf));
7252                 dns_rdatatype_format(fctx->type, typebuf, sizeof(typebuf));
7253                 dns_rdataclass_format(fctx->res->rdclass, classbuf,
7254                                       sizeof(classbuf));
7255                 isc_sockaddr_format(&query->addrinfo->sockaddr, addrbuf,
7256                                     sizeof(addrbuf));
7257
7258                 isc_log_write(dns_lctx, DNS_LOGCATEGORY_DELEGATION_ONLY,
7259                              DNS_LOGMODULE_RESOLVER, ISC_LOG_NOTICE,
7260                              "enforced delegation-only for '%s' (%s/%s/%s) "
7261                              "from %s",
7262                              domainbuf, namebuf, typebuf, classbuf, addrbuf);
7263         }
7264
7265         if ((fctx->res->options & DNS_RESOLVER_CHECKNAMES) != 0)
7266                 checknames(message);
7267
7268         /*
7269          * Clear cache bits.
7270          */
7271         fctx->attributes &= ~(FCTX_ATTR_WANTNCACHE | FCTX_ATTR_WANTCACHE);
7272
7273         /*
7274          * Did we get any answers?
7275          */
7276         if (message->counts[DNS_SECTION_ANSWER] > 0 &&
7277             (message->rcode == dns_rcode_noerror ||
7278              message->rcode == dns_rcode_nxdomain)) {
7279                 /*
7280                  * [normal case]
7281                  * We've got answers.  If it has an authoritative answer or an
7282                  * answer from a forwarder, we're done.
7283                  */
7284                 if ((message->flags & DNS_MESSAGEFLAG_AA) != 0 ||
7285                     ISFORWARDER(query->addrinfo))
7286                         result = answer_response(fctx);
7287                 else if (iscname(fctx) &&
7288                          fctx->type != dns_rdatatype_any &&
7289                          fctx->type != dns_rdatatype_cname) {
7290                         /*
7291                          * A BIND8 server could return a non-authoritative
7292                          * answer when a CNAME is followed.  We should treat
7293                          * it as a valid answer.
7294                          */
7295                         result = answer_response(fctx);
7296                 } else if (fctx->type != dns_rdatatype_ns &&
7297                            !betterreferral(fctx)) {
7298                         /*
7299                          * Lame response !!!.
7300                          */
7301                         result = answer_response(fctx);
7302                 } else {
7303                         if (fctx->type == dns_rdatatype_ns) {
7304                                 /*
7305                                  * A BIND 8 server could incorrectly return a
7306                                  * non-authoritative answer to an NS query
7307                                  * instead of a referral. Since this answer
7308                                  * lacks the SIGs necessary to do DNSSEC
7309                                  * validation, we must invoke the following
7310                                  * special kludge to treat it as a referral.
7311                                  */
7312                                 result = noanswer_response(fctx, NULL,
7313                                                    LOOK_FOR_NS_IN_ANSWER);
7314                         } else {
7315                                 /*
7316                                  * Some other servers may still somehow include
7317                                  * an answer when it should return a referral
7318                                  * with an empty answer.  Check to see if we can
7319                                  * treat this as a referral by ignoring the
7320                                  * answer.  Further more, there may be an
7321                                  * implementation that moves A/AAAA glue records
7322                                  * to the answer section for that type of
7323                                  * delegation when the query is for that glue
7324                                  * record.  LOOK_FOR_GLUE_IN_ANSWER will handle
7325                                  * such a corner case.
7326                                  */
7327                                 result = noanswer_response(fctx, NULL,
7328                                                    LOOK_FOR_GLUE_IN_ANSWER);
7329                         }
7330                         if (result != DNS_R_DELEGATION) {
7331                                 /*
7332                                  * At this point, AA is not set, the response
7333                                  * is not a referral, and the server is not a
7334                                  * forwarder.  It is technically lame and it's
7335                                  * easier to treat it as such than to figure out
7336                                  * some more elaborate course of action.
7337                                  */
7338                                 broken_server = DNS_R_LAME;
7339                                 keep_trying = ISC_TRUE;
7340                                 goto done;
7341                         }
7342                         goto force_referral;
7343                 }
7344                 if (result != ISC_R_SUCCESS) {
7345                         if (result == DNS_R_FORMERR)
7346                                 keep_trying = ISC_TRUE;
7347                         goto done;
7348                 }
7349         } else if (message->counts[DNS_SECTION_AUTHORITY] > 0 ||
7350                    message->rcode == dns_rcode_noerror ||
7351                    message->rcode == dns_rcode_nxdomain) {
7352                 /*
7353                  * NXDOMAIN, NXRDATASET, or referral.
7354                  */
7355                 result = noanswer_response(fctx, NULL, 0);
7356                 switch (result) {
7357                 case ISC_R_SUCCESS:
7358                 case DNS_R_CHASEDSSERVERS:
7359                         break;
7360                 case DNS_R_DELEGATION:
7361  force_referral:
7362                         /*
7363                          * We don't have the answer, but we know a better
7364                          * place to look.
7365                          */
7366                         get_nameservers = ISC_TRUE;
7367                         keep_trying = ISC_TRUE;
7368                         /*
7369                          * We have a new set of name servers, and it
7370                          * has not experienced any restarts yet.
7371                          */
7372                         fctx->restarts = 0;
7373
7374                         /*
7375                          * Update local statistics counters collected for each
7376                          * new zone.
7377                          */
7378                         fctx->referrals++;
7379                         fctx->querysent = 0;
7380                         fctx->lamecount = 0;
7381                         fctx->neterr = 0;
7382                         fctx->badresp = 0;
7383                         fctx->adberr = 0;
7384
7385                         result = ISC_R_SUCCESS;
7386                         break;
7387                 default:
7388                         /*
7389                          * Something has gone wrong.
7390                          */
7391                         if (result == DNS_R_FORMERR)
7392                                 keep_trying = ISC_TRUE;
7393                         goto done;
7394                 }
7395         } else {
7396                 /*
7397                  * The server is insane.
7398                  */
7399                 /* XXXRTH Log */
7400                 broken_server = DNS_R_UNEXPECTEDRCODE;
7401                 keep_trying = ISC_TRUE;
7402                 goto done;
7403         }
7404
7405         /*
7406          * Follow additional section data chains.
7407          */
7408         chase_additional(fctx);
7409
7410         /*
7411          * Cache the cacheable parts of the message.  This may also cause
7412          * work to be queued to the DNSSEC validator.
7413          */
7414         if (WANTCACHE(fctx)) {
7415                 result = cache_message(fctx, query->addrinfo, now);
7416                 if (result != ISC_R_SUCCESS)
7417                         goto done;
7418         }
7419
7420         /*
7421          * Ncache the negatively cacheable parts of the message.  This may
7422          * also cause work to be queued to the DNSSEC validator.
7423          */
7424         if (WANTNCACHE(fctx)) {
7425                 dns_rdatatype_t covers;
7426                 if (message->rcode == dns_rcode_nxdomain)
7427                         covers = dns_rdatatype_any;
7428                 else
7429                         covers = fctx->type;
7430
7431                 /*
7432                  * Cache any negative cache entries in the message.
7433                  */
7434                 result = ncache_message(fctx, query->addrinfo, covers, now);
7435         }
7436
7437  done:
7438         /*
7439          * Remember the query's addrinfo, in case we need to mark the
7440          * server as broken.
7441          */
7442         addrinfo = query->addrinfo;
7443
7444         /*
7445          * Cancel the query.
7446          *
7447          * XXXRTH  Don't cancel the query if waiting for validation?
7448          */
7449         fctx_cancelquery(&query, &devent, finish, no_response);
7450
7451         if (keep_trying) {
7452                 if (result == DNS_R_FORMERR)
7453                         broken_server = DNS_R_FORMERR;
7454                 if (broken_server != ISC_R_SUCCESS) {
7455                         /*
7456                          * Add this server to the list of bad servers for
7457                          * this fctx.
7458                          */
7459                         add_bad(fctx, addrinfo, broken_server, broken_type);
7460                 }
7461
7462                 if (get_nameservers) {
7463                         dns_name_t *name;
7464                         dns_fixedname_init(&foundname);
7465                         fname = dns_fixedname_name(&foundname);
7466                         if (result != ISC_R_SUCCESS) {
7467                                 fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
7468                                 return;
7469                         }
7470                         findoptions = 0;
7471                         if (dns_rdatatype_atparent(fctx->type))
7472                                 findoptions |= DNS_DBFIND_NOEXACT;
7473                         if ((options & DNS_FETCHOPT_UNSHARED) == 0)
7474                                 name = &fctx->name;
7475                         else
7476                                 name = &fctx->domain;
7477                         result = dns_view_findzonecut(fctx->res->view,
7478                                                       name, fname,
7479                                                       now, findoptions,
7480                                                       ISC_TRUE,
7481                                                       &fctx->nameservers,
7482                                                       NULL);
7483                         if (result != ISC_R_SUCCESS) {
7484                                 FCTXTRACE("couldn't find a zonecut");
7485                                 fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
7486                                 return;
7487                         }
7488                         if (!dns_name_issubdomain(fname, &fctx->domain)) {
7489                                 /*
7490                                  * The best nameservers are now above our
7491                                  * QDOMAIN.
7492                                  */
7493                                 FCTXTRACE("nameservers now above QDOMAIN");
7494                                 fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
7495                                 return;
7496                         }
7497                         dns_name_free(&fctx->domain, fctx->mctx);
7498                         dns_name_init(&fctx->domain, NULL);
7499                         result = dns_name_dup(fname, fctx->mctx, &fctx->domain);
7500                         if (result != ISC_R_SUCCESS) {
7501                                 fctx_done(fctx, DNS_R_SERVFAIL, __LINE__);
7502                                 return;
7503                         }
7504                         fctx->ns_ttl = fctx->nameservers.ttl;
7505                         fctx->ns_ttl_ok = ISC_TRUE;
7506                         fctx_cancelqueries(fctx, ISC_TRUE);
7507                         fctx_cleanupfinds(fctx);
7508                         fctx_cleanupaltfinds(fctx);
7509                         fctx_cleanupforwaddrs(fctx);
7510                         fctx_cleanupaltaddrs(fctx);
7511                 }
7512                 /*
7513                  * Try again.
7514                  */
7515                 fctx_try(fctx, !get_nameservers, ISC_FALSE);
7516         } else if (resend) {
7517                 /*
7518                  * Resend (probably with changed options).
7519                  */
7520                 FCTXTRACE("resend");
7521                 inc_stats(fctx->res, dns_resstatscounter_retry);
7522                 result = fctx_query(fctx, addrinfo, options);
7523                 if (result != ISC_R_SUCCESS)
7524                         fctx_done(fctx, result, __LINE__);
7525         } else if (result == ISC_R_SUCCESS && !HAVE_ANSWER(fctx)) {
7526                 /*
7527                  * All has gone well so far, but we are waiting for the
7528                  * DNSSEC validator to validate the answer.
7529                  */
7530                 FCTXTRACE("wait for validator");
7531                 fctx_cancelqueries(fctx, ISC_TRUE);
7532                 /*
7533                  * We must not retransmit while the validator is working;
7534                  * it has references to the current rmessage.
7535                  */
7536                 result = fctx_stopidletimer(fctx);
7537                 if (result != ISC_R_SUCCESS)
7538                         fctx_done(fctx, result, __LINE__);
7539         } else if (result == DNS_R_CHASEDSSERVERS) {
7540                 unsigned int n;
7541                 add_bad(fctx, addrinfo, result, broken_type);
7542                 fctx_cancelqueries(fctx, ISC_TRUE);
7543                 fctx_cleanupfinds(fctx);
7544                 fctx_cleanupforwaddrs(fctx);
7545
7546                 n = dns_name_countlabels(&fctx->name);
7547                 dns_name_getlabelsequence(&fctx->name, 1, n - 1, &fctx->nsname);
7548
7549                 FCTXTRACE("suspending DS lookup to find parent's NS records");
7550
7551                 result = dns_resolver_createfetch(fctx->res, &fctx->nsname,
7552                                                   dns_rdatatype_ns,
7553                                                   NULL, NULL, NULL, 0, task,
7554                                                   resume_dslookup, fctx,
7555                                                   &fctx->nsrrset, NULL,
7556                                                   &fctx->nsfetch);
7557                 if (result != ISC_R_SUCCESS)
7558                         fctx_done(fctx, result, __LINE__);
7559                 else {
7560                         LOCK(&fctx->res->buckets[fctx->bucketnum].lock);
7561                         fctx->references++;
7562                         UNLOCK(&fctx->res->buckets[fctx->bucketnum].lock);
7563                         result = fctx_stopidletimer(fctx);
7564                         if (result != ISC_R_SUCCESS)
7565                                 fctx_done(fctx, result, __LINE__);
7566                 }
7567         } else {
7568                 /*
7569                  * We're done.
7570                  */
7571                 fctx_done(fctx, result, __LINE__);
7572         }
7573 }
7574
7575
7576 /***
7577  *** Resolver Methods
7578  ***/
7579 static void
7580 destroy_badcache(dns_resolver_t *res) {
7581         dns_badcache_t *bad, *next;
7582         unsigned int i;
7583
7584         if (res->badcache != NULL) {
7585                 for (i = 0; i < res->badhash; i++)
7586                         for (bad = res->badcache[i]; bad != NULL;
7587                              bad = next) {
7588                                 next = bad->next;
7589                                 isc_mem_put(res->mctx, bad, sizeof(*bad) +
7590                                             bad->name.length);
7591                                 res->badcount--;
7592                         }
7593                 isc_mem_put(res->mctx, res->badcache,
7594                             sizeof(*res->badcache) * res->badhash);
7595                 res->badcache = NULL;
7596                 res->badhash = 0;
7597                 INSIST(res->badcount == 0);
7598         }
7599 }
7600
7601 static void
7602 destroy(dns_resolver_t *res) {
7603         unsigned int i;
7604         alternate_t *a;
7605
7606         REQUIRE(res->references == 0);
7607         REQUIRE(!res->priming);
7608         REQUIRE(res->primefetch == NULL);
7609
7610         RTRACE("destroy");
7611
7612         INSIST(res->nfctx == 0);
7613
7614         DESTROYLOCK(&res->primelock);
7615         DESTROYLOCK(&res->nlock);
7616         DESTROYLOCK(&res->lock);
7617         for (i = 0; i < res->nbuckets; i++) {
7618                 INSIST(ISC_LIST_EMPTY(res->buckets[i].fctxs));
7619                 isc_task_shutdown(res->buckets[i].task);
7620                 isc_task_detach(&res->buckets[i].task);
7621                 DESTROYLOCK(&res->buckets[i].lock);
7622                 isc_mem_detach(&res->buckets[i].mctx);
7623         }
7624         isc_mem_put(res->mctx, res->buckets,
7625                     res->nbuckets * sizeof(fctxbucket_t));
7626         if (res->dispatches4 != NULL)
7627                 dns_dispatchset_destroy(&res->dispatches4);
7628         if (res->dispatches6 != NULL)
7629                 dns_dispatchset_destroy(&res->dispatches6);
7630         while ((a = ISC_LIST_HEAD(res->alternates)) != NULL) {
7631                 ISC_LIST_UNLINK(res->alternates, a, link);
7632                 if (!a->isaddress)
7633                         dns_name_free(&a->_u._n.name, res->mctx);
7634                 isc_mem_put(res->mctx, a, sizeof(*a));
7635         }
7636         dns_resolver_reset_algorithms(res);
7637         destroy_badcache(res);
7638         dns_resolver_resetmustbesecure(res);
7639 #if USE_ALGLOCK
7640         isc_rwlock_destroy(&res->alglock);
7641 #endif
7642 #if USE_MBSLOCK
7643         isc_rwlock_destroy(&res->mbslock);
7644 #endif
7645         isc_timer_detach(&res->spillattimer);
7646         res->magic = 0;
7647         isc_mem_put(res->mctx, res, sizeof(*res));
7648 }
7649
7650 static void
7651 send_shutdown_events(dns_resolver_t *res) {
7652         isc_event_t *event, *next_event;
7653         isc_task_t *etask;
7654
7655         /*
7656          * Caller must be holding the resolver lock.
7657          */
7658
7659         for (event = ISC_LIST_HEAD(res->whenshutdown);
7660              event != NULL;
7661              event = next_event) {
7662                 next_event = ISC_LIST_NEXT(event, ev_link);
7663                 ISC_LIST_UNLINK(res->whenshutdown, event, ev_link);
7664                 etask = event->ev_sender;
7665                 event->ev_sender = res;
7666                 isc_task_sendanddetach(&etask, &event);
7667         }
7668 }
7669
7670 static void
7671 empty_bucket(dns_resolver_t *res) {
7672         RTRACE("empty_bucket");
7673
7674         LOCK(&res->lock);
7675
7676         INSIST(res->activebuckets > 0);
7677         res->activebuckets--;
7678         if (res->activebuckets == 0)
7679                 send_shutdown_events(res);
7680
7681         UNLOCK(&res->lock);
7682 }
7683
7684 static void
7685 spillattimer_countdown(isc_task_t *task, isc_event_t *event) {
7686         dns_resolver_t *res = event->ev_arg;
7687         isc_result_t result;
7688         unsigned int count;
7689         isc_boolean_t logit = ISC_FALSE;
7690
7691         REQUIRE(VALID_RESOLVER(res));
7692
7693         UNUSED(task);
7694
7695         LOCK(&res->lock);
7696         INSIST(!res->exiting);
7697         if (res->spillat > res->spillatmin) {
7698                 res->spillat--;
7699                 logit = ISC_TRUE;
7700         }
7701         if (res->spillat <= res->spillatmin) {
7702                 result = isc_timer_reset(res->spillattimer,
7703                                          isc_timertype_inactive, NULL,
7704                                          NULL, ISC_TRUE);
7705                 RUNTIME_CHECK(result == ISC_R_SUCCESS);
7706         }
7707         count = res->spillat;
7708         UNLOCK(&res->lock);
7709         if (logit)
7710                 isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
7711                               DNS_LOGMODULE_RESOLVER, ISC_LOG_NOTICE,
7712                               "clients-per-query decreased to %u", count);
7713
7714         isc_event_free(&event);
7715 }
7716
7717 isc_result_t
7718 dns_resolver_create(dns_view_t *view,
7719                     isc_taskmgr_t *taskmgr,
7720                     unsigned int ntasks, unsigned int ndisp,
7721                     isc_socketmgr_t *socketmgr,
7722                     isc_timermgr_t *timermgr,
7723                     unsigned int options,
7724                     dns_dispatchmgr_t *dispatchmgr,
7725                     dns_dispatch_t *dispatchv4,
7726                     dns_dispatch_t *dispatchv6,
7727                     dns_resolver_t **resp)
7728 {
7729         dns_resolver_t *res;
7730         isc_result_t result = ISC_R_SUCCESS;
7731         unsigned int i, buckets_created = 0;
7732         isc_task_t *task = NULL;
7733         char name[16];
7734         unsigned dispattr;
7735
7736         /*
7737          * Create a resolver.
7738          */
7739
7740         REQUIRE(DNS_VIEW_VALID(view));
7741         REQUIRE(ntasks > 0);
7742         REQUIRE(ndisp > 0);
7743         REQUIRE(resp != NULL && *resp == NULL);
7744         REQUIRE(dispatchmgr != NULL);
7745         REQUIRE(dispatchv4 != NULL || dispatchv6 != NULL);
7746
7747         res = isc_mem_get(view->mctx, sizeof(*res));
7748         if (res == NULL)
7749                 return (ISC_R_NOMEMORY);
7750         RTRACE("create");
7751         res->mctx = view->mctx;
7752         res->rdclass = view->rdclass;
7753         res->socketmgr = socketmgr;
7754         res->timermgr = timermgr;
7755         res->taskmgr = taskmgr;
7756         res->dispatchmgr = dispatchmgr;
7757         res->view = view;
7758         res->options = options;
7759         res->lame_ttl = 0;
7760         ISC_LIST_INIT(res->alternates);
7761         res->udpsize = RECV_BUFFER_SIZE;
7762         res->algorithms = NULL;
7763         res->badcache = NULL;
7764         res->badcount = 0;
7765         res->badhash = 0;
7766         res->badsweep = 0;
7767         res->mustbesecure = NULL;
7768         res->spillatmin = res->spillat = 10;
7769         res->spillatmax = 100;
7770         res->spillattimer = NULL;
7771         res->zero_no_soa_ttl = ISC_FALSE;
7772         res->query_timeout = DEFAULT_QUERY_TIMEOUT;
7773         res->maxdepth = DEFAULT_RECURSION_DEPTH;
7774         res->nbuckets = ntasks;
7775         res->activebuckets = ntasks;
7776         res->buckets = isc_mem_get(view->mctx,
7777                                    ntasks * sizeof(fctxbucket_t));
7778         if (res->buckets == NULL) {
7779                 result = ISC_R_NOMEMORY;
7780                 goto cleanup_res;
7781         }
7782         for (i = 0; i < ntasks; i++) {
7783                 result = isc_mutex_init(&res->buckets[i].lock);
7784                 if (result != ISC_R_SUCCESS)
7785                         goto cleanup_buckets;
7786                 res->buckets[i].task = NULL;
7787                 result = isc_task_create(taskmgr, 0, &res->buckets[i].task);
7788                 if (result != ISC_R_SUCCESS) {
7789                         DESTROYLOCK(&res->buckets[i].lock);
7790                         goto cleanup_buckets;
7791                 }
7792                 res->buckets[i].mctx = NULL;
7793                 snprintf(name, sizeof(name), "res%u", i);
7794 #ifdef ISC_PLATFORM_USETHREADS
7795                 /*
7796                  * Use a separate memory context for each bucket to reduce
7797                  * contention among multiple threads.  Do this only when
7798                  * enabling threads because it will be require more memory.
7799                  */
7800                 result = isc_mem_create(0, 0, &res->buckets[i].mctx);
7801                 if (result != ISC_R_SUCCESS) {
7802                         isc_task_detach(&res->buckets[i].task);
7803                         DESTROYLOCK(&res->buckets[i].lock);
7804                         goto cleanup_buckets;
7805                 }
7806                 isc_mem_setname(res->buckets[i].mctx, name, NULL);
7807 #else
7808                 isc_mem_attach(view->mctx, &res->buckets[i].mctx);
7809 #endif
7810                 isc_task_setname(res->buckets[i].task, name, res);
7811                 ISC_LIST_INIT(res->buckets[i].fctxs);
7812                 res->buckets[i].exiting = ISC_FALSE;
7813                 buckets_created++;
7814         }
7815
7816         res->dispatches4 = NULL;
7817         if (dispatchv4 != NULL) {
7818                 dns_dispatchset_create(view->mctx, socketmgr, taskmgr,
7819                                        dispatchv4, &res->dispatches4, ndisp);
7820                 dispattr = dns_dispatch_getattributes(dispatchv4);
7821                 res->exclusivev4 =
7822                         ISC_TF((dispattr & DNS_DISPATCHATTR_EXCLUSIVE) != 0);
7823         }
7824
7825         res->dispatches6 = NULL;
7826         if (dispatchv6 != NULL) {
7827                 dns_dispatchset_create(view->mctx, socketmgr, taskmgr,
7828                                        dispatchv6, &res->dispatches6, ndisp);
7829                 dispattr = dns_dispatch_getattributes(dispatchv6);
7830                 res->exclusivev6 =
7831                         ISC_TF((dispattr & DNS_DISPATCHATTR_EXCLUSIVE) != 0);
7832         }
7833
7834         res->references = 1;
7835         res->exiting = ISC_FALSE;
7836         res->frozen = ISC_FALSE;
7837         ISC_LIST_INIT(res->whenshutdown);
7838         res->priming = ISC_FALSE;
7839         res->primefetch = NULL;
7840         res->nfctx = 0;
7841
7842         result = isc_mutex_init(&res->lock);
7843         if (result != ISC_R_SUCCESS)
7844                 goto cleanup_dispatches;
7845
7846         result = isc_mutex_init(&res->nlock);
7847         if (result != ISC_R_SUCCESS)
7848                 goto cleanup_lock;
7849
7850         result = isc_mutex_init(&res->primelock);
7851         if (result != ISC_R_SUCCESS)
7852                 goto cleanup_nlock;
7853
7854         task = NULL;
7855         result = isc_task_create(taskmgr, 0, &task);
7856         if (result != ISC_R_SUCCESS)
7857                 goto cleanup_primelock;
7858
7859         result = isc_timer_create(timermgr, isc_timertype_inactive, NULL, NULL,
7860                                   task, spillattimer_countdown, res,
7861                                   &res->spillattimer);
7862         isc_task_detach(&task);
7863         if (result != ISC_R_SUCCESS)
7864                 goto cleanup_primelock;
7865
7866 #if USE_ALGLOCK
7867         result = isc_rwlock_init(&res->alglock, 0, 0);
7868         if (result != ISC_R_SUCCESS)
7869                 goto cleanup_spillattimer;
7870 #endif
7871 #if USE_MBSLOCK
7872         result = isc_rwlock_init(&res->mbslock, 0, 0);
7873         if (result != ISC_R_SUCCESS)
7874                 goto cleanup_alglock;
7875 #endif
7876
7877         res->magic = RES_MAGIC;
7878
7879         *resp = res;
7880
7881         return (ISC_R_SUCCESS);
7882
7883 #if USE_MBSLOCK
7884  cleanup_alglock:
7885 #if USE_ALGLOCK
7886         isc_rwlock_destroy(&res->alglock);
7887 #endif
7888 #endif
7889 #if USE_ALGLOCK || USE_MBSLOCK
7890  cleanup_spillattimer:
7891         isc_timer_detach(&res->spillattimer);
7892 #endif
7893
7894  cleanup_primelock:
7895         DESTROYLOCK(&res->primelock);
7896
7897  cleanup_nlock:
7898         DESTROYLOCK(&res->nlock);
7899
7900  cleanup_lock:
7901         DESTROYLOCK(&res->lock);
7902
7903  cleanup_dispatches:
7904         if (res->dispatches6 != NULL)
7905                 dns_dispatchset_destroy(&res->dispatches6);
7906         if (res->dispatches4 != NULL)
7907                 dns_dispatchset_destroy(&res->dispatches4);
7908
7909  cleanup_buckets:
7910         for (i = 0; i < buckets_created; i++) {
7911                 isc_mem_detach(&res->buckets[i].mctx);
7912                 DESTROYLOCK(&res->buckets[i].lock);
7913                 isc_task_shutdown(res->buckets[i].task);
7914                 isc_task_detach(&res->buckets[i].task);
7915         }
7916         isc_mem_put(view->mctx, res->buckets,
7917                     res->nbuckets * sizeof(fctxbucket_t));
7918
7919  cleanup_res:
7920         isc_mem_put(view->mctx, res, sizeof(*res));
7921
7922         return (result);
7923 }
7924
7925 #ifdef BIND9
7926 static void
7927 prime_done(isc_task_t *task, isc_event_t *event) {
7928         dns_resolver_t *res;
7929         dns_fetchevent_t *fevent;
7930         dns_fetch_t *fetch;
7931         dns_db_t *db = NULL;
7932
7933         REQUIRE(event->ev_type == DNS_EVENT_FETCHDONE);
7934         fevent = (dns_fetchevent_t *)event;
7935         res = event->ev_arg;
7936         REQUIRE(VALID_RESOLVER(res));
7937
7938         UNUSED(task);
7939
7940         LOCK(&res->lock);
7941
7942         INSIST(res->priming);
7943         res->priming = ISC_FALSE;
7944         LOCK(&res->primelock);
7945         fetch = res->primefetch;
7946         res->primefetch = NULL;
7947         UNLOCK(&res->primelock);
7948
7949         UNLOCK(&res->lock);
7950
7951         if (fevent->result == ISC_R_SUCCESS &&
7952             res->view->cache != NULL && res->view->hints != NULL) {
7953                 dns_cache_attachdb(res->view->cache, &db);
7954                 dns_root_checkhints(res->view, res->view->hints, db);
7955                 dns_db_detach(&db);
7956         }
7957
7958         if (fevent->node != NULL)
7959                 dns_db_detachnode(fevent->db, &fevent->node);
7960         if (fevent->db != NULL)
7961                 dns_db_detach(&fevent->db);
7962         if (dns_rdataset_isassociated(fevent->rdataset))
7963                 dns_rdataset_disassociate(fevent->rdataset);
7964         INSIST(fevent->sigrdataset == NULL);
7965
7966         isc_mem_put(res->mctx, fevent->rdataset, sizeof(*fevent->rdataset));
7967
7968         isc_event_free(&event);
7969         dns_resolver_destroyfetch(&fetch);
7970 }
7971
7972 void
7973 dns_resolver_prime(dns_resolver_t *res) {
7974         isc_boolean_t want_priming = ISC_FALSE;
7975         dns_rdataset_t *rdataset;
7976         isc_result_t result;
7977
7978         REQUIRE(VALID_RESOLVER(res));
7979         REQUIRE(res->frozen);
7980
7981         RTRACE("dns_resolver_prime");
7982
7983         LOCK(&res->lock);
7984
7985         if (!res->exiting && !res->priming) {
7986                 INSIST(res->primefetch == NULL);
7987                 res->priming = ISC_TRUE;
7988                 want_priming = ISC_TRUE;
7989         }
7990
7991         UNLOCK(&res->lock);
7992
7993         if (want_priming) {
7994                 /*
7995                  * To avoid any possible recursive locking problems, we
7996                  * start the priming fetch like any other fetch, and holding
7997                  * no resolver locks.  No one else will try to start it
7998                  * because we're the ones who set res->priming to true.
7999                  * Any other callers of dns_resolver_prime() while we're
8000                  * running will see that res->priming is already true and
8001                  * do nothing.
8002                  */
8003                 RTRACE("priming");
8004                 rdataset = isc_mem_get(res->mctx, sizeof(*rdataset));
8005                 if (rdataset == NULL) {
8006                         LOCK(&res->lock);
8007                         INSIST(res->priming);
8008                         INSIST(res->primefetch == NULL);
8009                         res->priming = ISC_FALSE;
8010                         UNLOCK(&res->lock);
8011                         return;
8012                 }
8013                 dns_rdataset_init(rdataset);
8014                 LOCK(&res->primelock);
8015                 result = dns_resolver_createfetch(res, dns_rootname,
8016                                                   dns_rdatatype_ns,
8017                                                   NULL, NULL, NULL, 0,
8018                                                   res->buckets[0].task,
8019                                                   prime_done,
8020                                                   res, rdataset, NULL,
8021                                                   &res->primefetch);
8022                 UNLOCK(&res->primelock);
8023                 if (result != ISC_R_SUCCESS) {
8024                         LOCK(&res->lock);
8025                         INSIST(res->priming);
8026                         res->priming = ISC_FALSE;
8027                         UNLOCK(&res->lock);
8028                 }
8029         }
8030 }
8031 #endif /* BIND9 */
8032
8033 void
8034 dns_resolver_freeze(dns_resolver_t *res) {
8035         /*
8036          * Freeze resolver.
8037          */
8038
8039         REQUIRE(VALID_RESOLVER(res));
8040
8041         res->frozen = ISC_TRUE;
8042 }
8043
8044 void
8045 dns_resolver_attach(dns_resolver_t *source, dns_resolver_t **targetp) {
8046         REQUIRE(VALID_RESOLVER(source));
8047         REQUIRE(targetp != NULL && *targetp == NULL);
8048
8049         RRTRACE(source, "attach");
8050         LOCK(&source->lock);
8051         REQUIRE(!source->exiting);
8052
8053         INSIST(source->references > 0);
8054         source->references++;
8055         INSIST(source->references != 0);
8056         UNLOCK(&source->lock);
8057
8058         *targetp = source;
8059 }
8060
8061 void
8062 dns_resolver_whenshutdown(dns_resolver_t *res, isc_task_t *task,
8063                           isc_event_t **eventp)
8064 {
8065         isc_task_t *clone;
8066         isc_event_t *event;
8067
8068         REQUIRE(VALID_RESOLVER(res));
8069         REQUIRE(eventp != NULL);
8070
8071         event = *eventp;
8072         *eventp = NULL;
8073
8074         LOCK(&res->lock);
8075
8076         if (res->exiting && res->activebuckets == 0) {
8077                 /*
8078                  * We're already shutdown.  Send the event.
8079                  */
8080                 event->ev_sender = res;
8081                 isc_task_send(task, &event);
8082         } else {
8083                 clone = NULL;
8084                 isc_task_attach(task, &clone);
8085                 event->ev_sender = clone;
8086                 ISC_LIST_APPEND(res->whenshutdown, event, ev_link);
8087         }
8088
8089         UNLOCK(&res->lock);
8090 }
8091
8092 void
8093 dns_resolver_shutdown(dns_resolver_t *res) {
8094         unsigned int i;
8095         fetchctx_t *fctx;
8096         isc_result_t result;
8097
8098         REQUIRE(VALID_RESOLVER(res));
8099
8100         RTRACE("shutdown");
8101
8102         LOCK(&res->lock);
8103
8104         if (!res->exiting) {
8105                 RTRACE("exiting");
8106                 res->exiting = ISC_TRUE;
8107
8108                 for (i = 0; i < res->nbuckets; i++) {
8109                         LOCK(&res->buckets[i].lock);
8110                         for (fctx = ISC_LIST_HEAD(res->buckets[i].fctxs);
8111                              fctx != NULL;
8112                              fctx = ISC_LIST_NEXT(fctx, link))
8113                                 fctx_shutdown(fctx);
8114                         if (res->dispatches4 != NULL && !res->exclusivev4) {
8115                                 dns_dispatchset_cancelall(res->dispatches4,
8116                                                           res->buckets[i].task);
8117                         }
8118                         if (res->dispatches6 != NULL && !res->exclusivev6) {
8119                                 dns_dispatchset_cancelall(res->dispatches6,
8120                                                           res->buckets[i].task);
8121                         }
8122                         res->buckets[i].exiting = ISC_TRUE;
8123                         if (ISC_LIST_EMPTY(res->buckets[i].fctxs)) {
8124                                 INSIST(res->activebuckets > 0);
8125                                 res->activebuckets--;
8126                         }
8127                         UNLOCK(&res->buckets[i].lock);
8128                 }
8129                 if (res->activebuckets == 0)
8130                         send_shutdown_events(res);
8131                 result = isc_timer_reset(res->spillattimer,
8132                                          isc_timertype_inactive, NULL,
8133                                          NULL, ISC_TRUE);
8134                 RUNTIME_CHECK(result == ISC_R_SUCCESS);
8135         }
8136
8137         UNLOCK(&res->lock);
8138 }
8139
8140 void
8141 dns_resolver_detach(dns_resolver_t **resp) {
8142         dns_resolver_t *res;
8143         isc_boolean_t need_destroy = ISC_FALSE;
8144
8145         REQUIRE(resp != NULL);
8146         res = *resp;
8147         REQUIRE(VALID_RESOLVER(res));
8148
8149         RTRACE("detach");
8150
8151         LOCK(&res->lock);
8152
8153         INSIST(res->references > 0);
8154         res->references--;
8155         if (res->references == 0) {
8156                 INSIST(res->exiting && res->activebuckets == 0);
8157                 need_destroy = ISC_TRUE;
8158         }
8159
8160         UNLOCK(&res->lock);
8161
8162         if (need_destroy)
8163                 destroy(res);
8164
8165         *resp = NULL;
8166 }
8167
8168 static inline isc_boolean_t
8169 fctx_match(fetchctx_t *fctx, dns_name_t *name, dns_rdatatype_t type,
8170            unsigned int options)
8171 {
8172         /*
8173          * Don't match fetch contexts that are shutting down.
8174          */
8175         if (fctx->cloned || fctx->state == fetchstate_done ||
8176             ISC_LIST_EMPTY(fctx->events))
8177                 return (ISC_FALSE);
8178
8179         if (fctx->type != type || fctx->options != options)
8180                 return (ISC_FALSE);
8181         return (dns_name_equal(&fctx->name, name));
8182 }
8183
8184 static inline void
8185 log_fetch(dns_name_t *name, dns_rdatatype_t type) {
8186         char namebuf[DNS_NAME_FORMATSIZE];
8187         char typebuf[DNS_RDATATYPE_FORMATSIZE];
8188         int level = ISC_LOG_DEBUG(1);
8189
8190         if (! isc_log_wouldlog(dns_lctx, level))
8191                 return;
8192
8193         dns_name_format(name, namebuf, sizeof(namebuf));
8194         dns_rdatatype_format(type, typebuf, sizeof(typebuf));
8195
8196         isc_log_write(dns_lctx, DNS_LOGCATEGORY_RESOLVER,
8197                       DNS_LOGMODULE_RESOLVER, level,
8198                       "createfetch: %s %s", namebuf, typebuf);
8199 }
8200
8201 isc_result_t
8202 dns_resolver_createfetch(dns_resolver_t *res, dns_name_t *name,
8203                          dns_rdatatype_t type,
8204                          dns_name_t *domain, dns_rdataset_t *nameservers,
8205                          dns_forwarders_t *forwarders,
8206                          unsigned int options, isc_task_t *task,
8207                          isc_taskaction_t action, void *arg,
8208                          dns_rdataset_t *rdataset,
8209                          dns_rdataset_t *sigrdataset,
8210                          dns_fetch_t **fetchp)
8211 {
8212         return (dns_resolver_createfetch3(res, name, type, domain,
8213                                           nameservers, forwarders, NULL, 0,
8214                                           options, 0, task, action, arg,
8215                                           rdataset, sigrdataset, fetchp));
8216 }
8217
8218 isc_result_t
8219 dns_resolver_createfetch2(dns_resolver_t *res, dns_name_t *name,
8220                           dns_rdatatype_t type,
8221                           dns_name_t *domain, dns_rdataset_t *nameservers,
8222                           dns_forwarders_t *forwarders,
8223                           isc_sockaddr_t *client, dns_messageid_t id,
8224                           unsigned int options, isc_task_t *task,
8225                           isc_taskaction_t action, void *arg,
8226                           dns_rdataset_t *rdataset,
8227                           dns_rdataset_t *sigrdataset,
8228                           dns_fetch_t **fetchp)
8229 {
8230         return (dns_resolver_createfetch3(res, name, type, domain,
8231                                           nameservers, forwarders, client, id,
8232                                           options, 0, task, action, arg,
8233                                           rdataset, sigrdataset, fetchp));
8234 }
8235
8236 isc_result_t
8237 dns_resolver_createfetch3(dns_resolver_t *res, dns_name_t *name,
8238                           dns_rdatatype_t type,
8239                           dns_name_t *domain, dns_rdataset_t *nameservers,
8240                           dns_forwarders_t *forwarders,
8241                           isc_sockaddr_t *client, dns_messageid_t id,
8242                           unsigned int options, unsigned int depth,
8243                           isc_task_t *task,
8244                           isc_taskaction_t action, void *arg,
8245                           dns_rdataset_t *rdataset,
8246                           dns_rdataset_t *sigrdataset,
8247                           dns_fetch_t **fetchp)
8248 {
8249         dns_fetch_t *fetch;
8250         fetchctx_t *fctx = NULL;
8251         isc_result_t result = ISC_R_SUCCESS;
8252         unsigned int bucketnum;
8253         isc_boolean_t new_fctx = ISC_FALSE;
8254         isc_event_t *event;
8255         unsigned int count = 0;
8256         unsigned int spillat;
8257         unsigned int spillatmin;
8258         isc_boolean_t destroy = ISC_FALSE;
8259
8260         UNUSED(forwarders);
8261
8262         REQUIRE(VALID_RESOLVER(res));
8263         REQUIRE(res->frozen);
8264         /* XXXRTH  Check for meta type */
8265         if (domain != NULL) {
8266                 REQUIRE(DNS_RDATASET_VALID(nameservers));
8267                 REQUIRE(nameservers->type == dns_rdatatype_ns);
8268         } else
8269                 REQUIRE(nameservers == NULL);
8270         REQUIRE(forwarders == NULL);
8271         REQUIRE(!dns_rdataset_isassociated(rdataset));
8272         REQUIRE(sigrdataset == NULL ||
8273                 !dns_rdataset_isassociated(sigrdataset));
8274         REQUIRE(fetchp != NULL && *fetchp == NULL);
8275
8276         log_fetch(name, type);
8277
8278         /*
8279          * XXXRTH  use a mempool?
8280          */
8281         fetch = isc_mem_get(res->mctx, sizeof(*fetch));
8282         if (fetch == NULL)
8283                 return (ISC_R_NOMEMORY);
8284
8285         bucketnum = dns_name_fullhash(name, ISC_FALSE) % res->nbuckets;
8286
8287         LOCK(&res->lock);
8288         spillat = res->spillat;
8289         spillatmin = res->spillatmin;
8290         UNLOCK(&res->lock);
8291         LOCK(&res->buckets[bucketnum].lock);
8292
8293         if (res->buckets[bucketnum].exiting) {
8294                 result = ISC_R_SHUTTINGDOWN;
8295                 goto unlock;
8296         }
8297
8298         if ((options & DNS_FETCHOPT_UNSHARED) == 0) {
8299                 for (fctx = ISC_LIST_HEAD(res->buckets[bucketnum].fctxs);
8300                      fctx != NULL;
8301                      fctx = ISC_LIST_NEXT(fctx, link)) {
8302                         if (fctx_match(fctx, name, type, options))
8303                                 break;
8304                 }
8305         }
8306
8307         /*
8308          * Is this a duplicate?
8309          */
8310         if (fctx != NULL && client != NULL) {
8311                 dns_fetchevent_t *fevent;
8312                 for (fevent = ISC_LIST_HEAD(fctx->events);
8313                      fevent != NULL;
8314                      fevent = ISC_LIST_NEXT(fevent, ev_link)) {
8315                         if (fevent->client != NULL && fevent->id == id &&
8316                             isc_sockaddr_equal(fevent->client, client)) {
8317                                 result = DNS_R_DUPLICATE;
8318                                 goto unlock;
8319                         }
8320                         count++;
8321                 }
8322         }
8323         if (count >= spillatmin && spillatmin != 0) {
8324                 INSIST(fctx != NULL);
8325                 if (count >= spillat)
8326                         fctx->spilled = ISC_TRUE;
8327                 if (fctx->spilled) {
8328                         result = DNS_R_DROP;
8329                         goto unlock;
8330                 }
8331         }
8332
8333         if (fctx == NULL) {
8334                 result = fctx_create(res, name, type, domain, nameservers,
8335                                      options, bucketnum, depth, &fctx);
8336                 if (result != ISC_R_SUCCESS)
8337                         goto unlock;
8338                 new_fctx = ISC_TRUE;
8339         } else if (fctx->depth > depth)
8340                 fctx->depth = depth;
8341
8342         result = fctx_join(fctx, task, client, id, action, arg,
8343                            rdataset, sigrdataset, fetch);
8344         if (new_fctx) {
8345                 if (result == ISC_R_SUCCESS) {
8346                         /*
8347                          * Launch this fctx.
8348                          */
8349                         event = &fctx->control_event;
8350                         ISC_EVENT_INIT(event, sizeof(*event), 0, NULL,
8351                                        DNS_EVENT_FETCHCONTROL,
8352                                        fctx_start, fctx, NULL,
8353                                        NULL, NULL);
8354                         isc_task_send(res->buckets[bucketnum].task, &event);
8355                 } else {
8356                         /*
8357                          * We don't care about the result of fctx_unlink()
8358                          * since we know we're not exiting.
8359                          */
8360                         (void)fctx_unlink(fctx);
8361                         destroy = ISC_TRUE;
8362                 }
8363         }
8364
8365  unlock:
8366         UNLOCK(&res->buckets[bucketnum].lock);
8367
8368         if (destroy)
8369                 fctx_destroy(fctx);
8370
8371         if (result == ISC_R_SUCCESS) {
8372                 FTRACE("created");
8373                 *fetchp = fetch;
8374         } else
8375                 isc_mem_put(res->mctx, fetch, sizeof(*fetch));
8376
8377         return (result);
8378 }
8379
8380 void
8381 dns_resolver_cancelfetch(dns_fetch_t *fetch) {
8382         fetchctx_t *fctx;
8383         dns_resolver_t *res;
8384         dns_fetchevent_t *event, *next_event;
8385         isc_task_t *etask;
8386
8387         REQUIRE(DNS_FETCH_VALID(fetch));
8388         fctx = fetch->private;
8389         REQUIRE(VALID_FCTX(fctx));
8390         res = fctx->res;
8391
8392         FTRACE("cancelfetch");
8393
8394         LOCK(&res->buckets[fctx->bucketnum].lock);
8395
8396         /*
8397          * Find the completion event for this fetch (as opposed
8398          * to those for other fetches that have joined the same
8399          * fctx) and send it with result = ISC_R_CANCELED.
8400          */
8401         event = NULL;
8402         if (fctx->state != fetchstate_done) {
8403                 for (event = ISC_LIST_HEAD(fctx->events);
8404                      event != NULL;
8405                      event = next_event) {
8406                         next_event = ISC_LIST_NEXT(event, ev_link);
8407                         if (event->fetch == fetch) {
8408                                 ISC_LIST_UNLINK(fctx->events, event, ev_link);
8409                                 break;
8410                         }
8411                 }
8412         }
8413         if (event != NULL) {
8414                 etask = event->ev_sender;
8415                 event->ev_sender = fctx;
8416                 event->result = ISC_R_CANCELED;
8417                 isc_task_sendanddetach(&etask, ISC_EVENT_PTR(&event));
8418         }
8419         /*
8420          * The fctx continues running even if no fetches remain;
8421          * the answer is still cached.
8422          */
8423
8424         UNLOCK(&res->buckets[fctx->bucketnum].lock);
8425 }
8426
8427 void
8428 dns_resolver_destroyfetch(dns_fetch_t **fetchp) {
8429         dns_fetch_t *fetch;
8430         dns_resolver_t *res;
8431         dns_fetchevent_t *event, *next_event;
8432         fetchctx_t *fctx;
8433         unsigned int bucketnum;
8434         isc_boolean_t bucket_empty;
8435
8436         REQUIRE(fetchp != NULL);
8437         fetch = *fetchp;
8438         REQUIRE(DNS_FETCH_VALID(fetch));
8439         fctx = fetch->private;
8440         REQUIRE(VALID_FCTX(fctx));
8441         res = fctx->res;
8442
8443         FTRACE("destroyfetch");
8444
8445         bucketnum = fctx->bucketnum;
8446         LOCK(&res->buckets[bucketnum].lock);
8447
8448         /*
8449          * Sanity check: the caller should have gotten its event before
8450          * trying to destroy the fetch.
8451          */
8452         event = NULL;
8453         if (fctx->state != fetchstate_done) {
8454                 for (event = ISC_LIST_HEAD(fctx->events);
8455                      event != NULL;
8456                      event = next_event) {
8457                         next_event = ISC_LIST_NEXT(event, ev_link);
8458                         RUNTIME_CHECK(event->fetch != fetch);
8459                 }
8460         }
8461
8462         bucket_empty = fctx_decreference(fctx);
8463
8464         UNLOCK(&res->buckets[bucketnum].lock);
8465
8466         isc_mem_put(res->mctx, fetch, sizeof(*fetch));
8467         *fetchp = NULL;
8468
8469         if (bucket_empty)
8470                 empty_bucket(res);
8471 }
8472
8473 void
8474 dns_resolver_logfetch(dns_fetch_t *fetch, isc_log_t *lctx,
8475                       isc_logcategory_t *category, isc_logmodule_t *module,
8476                       int level, isc_boolean_t duplicateok)
8477 {
8478         fetchctx_t *fctx;
8479         dns_resolver_t *res;
8480         char domainbuf[DNS_NAME_FORMATSIZE];
8481
8482         REQUIRE(DNS_FETCH_VALID(fetch));
8483         fctx = fetch->private;
8484         REQUIRE(VALID_FCTX(fctx));
8485         res = fctx->res;
8486
8487         LOCK(&res->buckets[fctx->bucketnum].lock);
8488
8489         INSIST(fctx->exitline >= 0);
8490         if (!fctx->logged || duplicateok) {
8491                 dns_name_format(&fctx->domain, domainbuf, sizeof(domainbuf));
8492                 isc_log_write(lctx, category, module, level,
8493                               "fetch completed at %s:%d for %s in "
8494                               "%" ISC_PRINT_QUADFORMAT "u."
8495                               "%06" ISC_PRINT_QUADFORMAT "u: %s/%s "
8496                               "[domain:%s,referral:%u,restart:%u,qrysent:%u,"
8497                               "timeout:%u,lame:%u,neterr:%u,badresp:%u,"
8498                               "adberr:%u,findfail:%u,valfail:%u]",
8499                               __FILE__, fctx->exitline, fctx->info,
8500                               fctx->duration / US_PER_SEC,
8501                               fctx->duration % US_PER_SEC,
8502                               isc_result_totext(fctx->result),
8503                               isc_result_totext(fctx->vresult), domainbuf,
8504                               fctx->referrals, fctx->restarts,
8505                               fctx->querysent, fctx->timeouts, fctx->lamecount,
8506                               fctx->neterr, fctx->badresp, fctx->adberr,
8507                               fctx->findfail, fctx->valfail);
8508                 fctx->logged = ISC_TRUE;
8509         }
8510
8511         UNLOCK(&res->buckets[fctx->bucketnum].lock);
8512 }
8513
8514 dns_dispatchmgr_t *
8515 dns_resolver_dispatchmgr(dns_resolver_t *resolver) {
8516         REQUIRE(VALID_RESOLVER(resolver));
8517         return (resolver->dispatchmgr);
8518 }
8519
8520 dns_dispatch_t *
8521 dns_resolver_dispatchv4(dns_resolver_t *resolver) {
8522         REQUIRE(VALID_RESOLVER(resolver));
8523         return (dns_dispatchset_get(resolver->dispatches4));
8524 }
8525
8526 dns_dispatch_t *
8527 dns_resolver_dispatchv6(dns_resolver_t *resolver) {
8528         REQUIRE(VALID_RESOLVER(resolver));
8529         return (dns_dispatchset_get(resolver->dispatches6));
8530 }
8531
8532 isc_socketmgr_t *
8533 dns_resolver_socketmgr(dns_resolver_t *resolver) {
8534         REQUIRE(VALID_RESOLVER(resolver));
8535         return (resolver->socketmgr);
8536 }
8537
8538 isc_taskmgr_t *
8539 dns_resolver_taskmgr(dns_resolver_t *resolver) {
8540         REQUIRE(VALID_RESOLVER(resolver));
8541         return (resolver->taskmgr);
8542 }
8543
8544 isc_uint32_t
8545 dns_resolver_getlamettl(dns_resolver_t *resolver) {
8546         REQUIRE(VALID_RESOLVER(resolver));
8547         return (resolver->lame_ttl);
8548 }
8549
8550 void
8551 dns_resolver_setlamettl(dns_resolver_t *resolver, isc_uint32_t lame_ttl) {
8552         REQUIRE(VALID_RESOLVER(resolver));
8553         resolver->lame_ttl = lame_ttl;
8554 }
8555
8556 unsigned int
8557 dns_resolver_nrunning(dns_resolver_t *resolver) {
8558         unsigned int n;
8559         LOCK(&resolver->nlock);
8560         n = resolver->nfctx;
8561         UNLOCK(&resolver->nlock);
8562         return (n);
8563 }
8564
8565 isc_result_t
8566 dns_resolver_addalternate(dns_resolver_t *resolver, isc_sockaddr_t *alt,
8567                           dns_name_t *name, in_port_t port) {
8568         alternate_t *a;
8569         isc_result_t result;
8570
8571         REQUIRE(VALID_RESOLVER(resolver));
8572         REQUIRE(!resolver->frozen);
8573         REQUIRE((alt == NULL) ^ (name == NULL));
8574
8575         a = isc_mem_get(resolver->mctx, sizeof(*a));
8576         if (a == NULL)
8577                 return (ISC_R_NOMEMORY);
8578         if (alt != NULL) {
8579                 a->isaddress = ISC_TRUE;
8580                 a->_u.addr = *alt;
8581         } else {
8582                 a->isaddress = ISC_FALSE;
8583                 a->_u._n.port = port;
8584                 dns_name_init(&a->_u._n.name, NULL);
8585                 result = dns_name_dup(name, resolver->mctx, &a->_u._n.name);
8586                 if (result != ISC_R_SUCCESS) {
8587                         isc_mem_put(resolver->mctx, a, sizeof(*a));
8588                         return (result);
8589                 }
8590         }
8591         ISC_LINK_INIT(a, link);
8592         ISC_LIST_APPEND(resolver->alternates, a, link);
8593
8594         return (ISC_R_SUCCESS);
8595 }
8596
8597 void
8598 dns_resolver_setudpsize(dns_resolver_t *resolver, isc_uint16_t udpsize) {
8599         REQUIRE(VALID_RESOLVER(resolver));
8600         resolver->udpsize = udpsize;
8601 }
8602
8603 isc_uint16_t
8604 dns_resolver_getudpsize(dns_resolver_t *resolver) {
8605         REQUIRE(VALID_RESOLVER(resolver));
8606         return (resolver->udpsize);
8607 }
8608
8609 void
8610 dns_resolver_flushbadcache(dns_resolver_t *resolver, dns_name_t *name) {
8611         unsigned int i;
8612         dns_badcache_t *bad, *prev, *next;
8613
8614         REQUIRE(VALID_RESOLVER(resolver));
8615
8616         LOCK(&resolver->lock);
8617         if (resolver->badcache == NULL)
8618                 goto unlock;
8619
8620         if (name != NULL) {
8621                 isc_time_t now;
8622                 isc_result_t result;
8623                 result = isc_time_now(&now);
8624                 if (result != ISC_R_SUCCESS)
8625                         isc_time_settoepoch(&now);
8626                 i = dns_name_hash(name, ISC_FALSE) % resolver->badhash;
8627                 prev = NULL;
8628                 for (bad = resolver->badcache[i]; bad != NULL; bad = next) {
8629                         int n;
8630                         next = bad->next;
8631                         n = isc_time_compare(&bad->expire, &now);
8632                         if (n < 0 || dns_name_equal(name, &bad->name)) {
8633                                 if (prev == NULL)
8634                                         resolver->badcache[i] = bad->next;
8635                                 else
8636                                         prev->next = bad->next;
8637                                 isc_mem_put(resolver->mctx, bad, sizeof(*bad) +
8638                                             bad->name.length);
8639                                 resolver->badcount--;
8640                         } else
8641                                 prev = bad;
8642                 }
8643         } else
8644                 destroy_badcache(resolver);
8645
8646  unlock:
8647         UNLOCK(&resolver->lock);
8648
8649 }
8650
8651 static void
8652 resizehash(dns_resolver_t *resolver, isc_time_t *now, isc_boolean_t grow) {
8653         unsigned int newsize;
8654         dns_badcache_t **new, *bad, *next;
8655         unsigned int i;
8656
8657         if (grow)
8658                 newsize = resolver->badhash * 2 + 1;
8659         else
8660                 newsize = (resolver->badhash - 1) / 2;
8661
8662         new = isc_mem_get(resolver->mctx,
8663                           sizeof(*resolver->badcache) * newsize);
8664         if (new == NULL)
8665                 return;
8666         memset(new, 0, sizeof(*resolver->badcache) * newsize);
8667         for (i = 0; i < resolver->badhash; i++) {
8668                 for (bad = resolver->badcache[i]; bad != NULL; bad = next) {
8669                         next = bad->next;
8670                         if (isc_time_compare(&bad->expire, now) < 0) {
8671                                 isc_mem_put(resolver->mctx, bad, sizeof(*bad) +
8672                                             bad->name.length);
8673                                 resolver->badcount--;
8674                         } else {
8675                                 bad->next = new[bad->hashval % newsize];
8676                                 new[bad->hashval % newsize] = bad;
8677                         }
8678                 }
8679         }
8680         isc_mem_put(resolver->mctx, resolver->badcache,
8681                     sizeof(*resolver->badcache) * resolver->badhash);
8682         resolver->badhash = newsize;
8683         resolver->badcache = new;
8684 }
8685
8686 void
8687 dns_resolver_addbadcache(dns_resolver_t *resolver, dns_name_t *name,
8688                          dns_rdatatype_t type, isc_time_t *expire)
8689 {
8690         isc_time_t now;
8691         isc_result_t result = ISC_R_SUCCESS;
8692         unsigned int i, hashval;
8693         dns_badcache_t *bad, *prev, *next;
8694
8695         REQUIRE(VALID_RESOLVER(resolver));
8696
8697         LOCK(&resolver->lock);
8698         if (resolver->badcache == NULL) {
8699                 resolver->badcache = isc_mem_get(resolver->mctx,
8700                                                  sizeof(*resolver->badcache) *
8701                                                  DNS_BADCACHE_SIZE);
8702                 if (resolver->badcache == NULL)
8703                         goto cleanup;
8704                 resolver->badhash = DNS_BADCACHE_SIZE;
8705                 memset(resolver->badcache, 0, sizeof(*resolver->badcache) *
8706                        resolver->badhash);
8707         }
8708
8709         result = isc_time_now(&now);
8710         if (result != ISC_R_SUCCESS)
8711                 isc_time_settoepoch(&now);
8712         hashval = dns_name_hash(name, ISC_FALSE);
8713         i = hashval % resolver->badhash;
8714         prev = NULL;
8715         for (bad = resolver->badcache[i]; bad != NULL; bad = next) {
8716                 next = bad->next;
8717                 if (bad->type == type && dns_name_equal(name, &bad->name))
8718                         break;
8719                 if (isc_time_compare(&bad->expire, &now) < 0) {
8720                         if (prev == NULL)
8721                                 resolver->badcache[i] = bad->next;
8722                         else
8723                                 prev->next = bad->next;
8724                         isc_mem_put(resolver->mctx, bad, sizeof(*bad) +
8725                                     bad->name.length);
8726                         resolver->badcount--;
8727                 } else
8728                         prev = bad;
8729         }
8730         if (bad == NULL) {
8731                 isc_buffer_t buffer;
8732                 bad = isc_mem_get(resolver->mctx, sizeof(*bad) + name->length);
8733                 if (bad == NULL)
8734                         goto cleanup;
8735                 bad->type = type;
8736                 bad->hashval = hashval;
8737                 bad->expire = *expire;
8738                 isc_buffer_init(&buffer, bad + 1, name->length);
8739                 dns_name_init(&bad->name, NULL);
8740                 dns_name_copy(name, &bad->name, &buffer);
8741                 bad->next = resolver->badcache[i];
8742                 resolver->badcache[i] = bad;
8743                 resolver->badcount++;
8744                 if (resolver->badcount > resolver->badhash * 8)
8745                         resizehash(resolver, &now, ISC_TRUE);
8746                 if (resolver->badcount < resolver->badhash * 2 &&
8747                     resolver->badhash > DNS_BADCACHE_SIZE)
8748                         resizehash(resolver, &now, ISC_FALSE);
8749         } else
8750                 bad->expire = *expire;
8751  cleanup:
8752         UNLOCK(&resolver->lock);
8753 }
8754
8755 isc_boolean_t
8756 dns_resolver_getbadcache(dns_resolver_t *resolver, dns_name_t *name,
8757                          dns_rdatatype_t type, isc_time_t *now)
8758 {
8759         dns_badcache_t *bad, *prev, *next;
8760         isc_boolean_t answer = ISC_FALSE;
8761         unsigned int i;
8762
8763         REQUIRE(VALID_RESOLVER(resolver));
8764
8765         LOCK(&resolver->lock);
8766         if (resolver->badcache == NULL)
8767                 goto unlock;
8768
8769         i = dns_name_hash(name, ISC_FALSE) % resolver->badhash;
8770         prev = NULL;
8771         for (bad = resolver->badcache[i]; bad != NULL; bad = next) {
8772                 next = bad->next;
8773                 /*
8774                  * Search the hash list. Clean out expired records as we go.
8775                  */
8776                 if (isc_time_compare(&bad->expire, now) < 0) {
8777                         if (prev != NULL)
8778                                 prev->next = bad->next;
8779                         else
8780                                 resolver->badcache[i] = bad->next;
8781                         isc_mem_put(resolver->mctx, bad, sizeof(*bad) +
8782                                     bad->name.length);
8783                         resolver->badcount--;
8784                         continue;
8785                 }
8786                 if (bad->type == type && dns_name_equal(name, &bad->name)) {
8787                         answer = ISC_TRUE;
8788                         break;
8789                 }
8790                 prev = bad;
8791         }
8792
8793         /*
8794          * Slow sweep to clean out stale records.
8795          */
8796         i = resolver->badsweep++ % resolver->badhash;
8797         bad = resolver->badcache[i];
8798         if (bad != NULL && isc_time_compare(&bad->expire, now) < 0) {
8799                 resolver->badcache[i] = bad->next;
8800                 isc_mem_put(resolver->mctx, bad, sizeof(*bad) +
8801                             bad->name.length);
8802                 resolver->badcount--;
8803         }
8804
8805  unlock:
8806         UNLOCK(&resolver->lock);
8807         return (answer);
8808 }
8809
8810 void
8811 dns_resolver_printbadcache(dns_resolver_t *resolver, FILE *fp) {
8812         char namebuf[DNS_NAME_FORMATSIZE];
8813         char typebuf[DNS_RDATATYPE_FORMATSIZE];
8814         dns_badcache_t *bad, *next, *prev;
8815         isc_time_t now;
8816         unsigned int i;
8817         isc_uint64_t t;
8818
8819         LOCK(&resolver->lock);
8820         fprintf(fp, ";\n; Bad cache\n;\n");
8821
8822         if (resolver->badcache == NULL)
8823                 goto unlock;
8824
8825         TIME_NOW(&now);
8826         for (i = 0; i < resolver->badhash; i++) {
8827                 prev = NULL;
8828                 for (bad = resolver->badcache[i]; bad != NULL; bad = next) {
8829                         next = bad->next;
8830                         if (isc_time_compare(&bad->expire, &now) < 0) {
8831                                 if (prev != NULL)
8832                                         prev->next = bad->next;
8833                                 else
8834                                         resolver->badcache[i] = bad->next;
8835                                 isc_mem_put(resolver->mctx, bad, sizeof(*bad) +
8836                                             bad->name.length);
8837                                 resolver->badcount--;
8838                                 continue;
8839                         }
8840                         prev = bad;
8841                         dns_name_format(&bad->name, namebuf, sizeof(namebuf));
8842                         dns_rdatatype_format(bad->type, typebuf,
8843                                              sizeof(typebuf));
8844                         t = isc_time_microdiff(&bad->expire, &now);
8845                         t /= 1000;
8846                         fprintf(fp, "; %s/%s [ttl "
8847                                 "%" ISC_PLATFORM_QUADFORMAT "u]\n",
8848                                 namebuf, typebuf, t);
8849                 }
8850         }
8851
8852  unlock:
8853         UNLOCK(&resolver->lock);
8854 }
8855
8856 static void
8857 free_algorithm(void *node, void *arg) {
8858         unsigned char *algorithms = node;
8859         isc_mem_t *mctx = arg;
8860
8861         isc_mem_put(mctx, algorithms, *algorithms);
8862 }
8863
8864 void
8865 dns_resolver_reset_algorithms(dns_resolver_t *resolver) {
8866
8867         REQUIRE(VALID_RESOLVER(resolver));
8868
8869 #if USE_ALGLOCK
8870         RWLOCK(&resolver->alglock, isc_rwlocktype_write);
8871 #endif
8872         if (resolver->algorithms != NULL)
8873                 dns_rbt_destroy(&resolver->algorithms);
8874 #if USE_ALGLOCK
8875         RWUNLOCK(&resolver->alglock, isc_rwlocktype_write);
8876 #endif
8877 }
8878
8879 isc_result_t
8880 dns_resolver_disable_algorithm(dns_resolver_t *resolver, dns_name_t *name,
8881                                unsigned int alg)
8882 {
8883         unsigned int len, mask;
8884         unsigned char *new;
8885         unsigned char *algorithms;
8886         isc_result_t result;
8887         dns_rbtnode_t *node = NULL;
8888
8889         REQUIRE(VALID_RESOLVER(resolver));
8890         if (alg > 255)
8891                 return (ISC_R_RANGE);
8892
8893 #if USE_ALGLOCK
8894         RWLOCK(&resolver->alglock, isc_rwlocktype_write);
8895 #endif
8896         if (resolver->algorithms == NULL) {
8897                 result = dns_rbt_create(resolver->mctx, free_algorithm,
8898                                         resolver->mctx, &resolver->algorithms);
8899                 if (result != ISC_R_SUCCESS)
8900                         goto cleanup;
8901         }
8902
8903         len = alg/8 + 2;
8904         mask = 1 << (alg%8);
8905
8906         result = dns_rbt_addnode(resolver->algorithms, name, &node);
8907
8908         if (result == ISC_R_SUCCESS || result == ISC_R_EXISTS) {
8909                 algorithms = node->data;
8910                 if (algorithms == NULL || len > *algorithms) {
8911                         new = isc_mem_get(resolver->mctx, len);
8912                         if (new == NULL) {
8913                                 result = ISC_R_NOMEMORY;
8914                                 goto cleanup;
8915                         }
8916                         memset(new, 0, len);
8917                         if (algorithms != NULL)
8918                                 memmove(new, algorithms, *algorithms);
8919                         new[len-1] |= mask;
8920                         *new = len;
8921                         node->data = new;
8922                         if (algorithms != NULL)
8923                                 isc_mem_put(resolver->mctx, algorithms,
8924                                             *algorithms);
8925                 } else
8926                         algorithms[len-1] |= mask;
8927         }
8928         result = ISC_R_SUCCESS;
8929  cleanup:
8930 #if USE_ALGLOCK
8931         RWUNLOCK(&resolver->alglock, isc_rwlocktype_write);
8932 #endif
8933         return (result);
8934 }
8935
8936 isc_boolean_t
8937 dns_resolver_algorithm_supported(dns_resolver_t *resolver, dns_name_t *name,
8938                                  unsigned int alg)
8939 {
8940         unsigned int len, mask;
8941         unsigned char *algorithms;
8942         void *data = NULL;
8943         isc_result_t result;
8944         isc_boolean_t found = ISC_FALSE;
8945
8946         REQUIRE(VALID_RESOLVER(resolver));
8947
8948         /*
8949          * DH is unsupported for DNSKEYs, see RFC 4034 sec. A.1.
8950          */
8951         if ((alg == DST_ALG_DH) || (alg == DST_ALG_INDIRECT))
8952                 return (ISC_FALSE);
8953
8954 #if USE_ALGLOCK
8955         RWLOCK(&resolver->alglock, isc_rwlocktype_read);
8956 #endif
8957         if (resolver->algorithms == NULL)
8958                 goto unlock;
8959         result = dns_rbt_findname(resolver->algorithms, name, 0, NULL, &data);
8960         if (result == ISC_R_SUCCESS || result == DNS_R_PARTIALMATCH) {
8961                 len = alg/8 + 2;
8962                 mask = 1 << (alg%8);
8963                 algorithms = data;
8964                 if (len <= *algorithms && (algorithms[len-1] & mask) != 0)
8965                         found = ISC_TRUE;
8966         }
8967  unlock:
8968 #if USE_ALGLOCK
8969         RWUNLOCK(&resolver->alglock, isc_rwlocktype_read);
8970 #endif
8971         if (found)
8972                 return (ISC_FALSE);
8973
8974         return (dst_algorithm_supported(alg));
8975 }
8976
8977 isc_boolean_t
8978 dns_resolver_digest_supported(dns_resolver_t *resolver, unsigned int digest) {
8979
8980         UNUSED(resolver);
8981         return (dns_ds_digest_supported(digest));
8982 }
8983
8984 void
8985 dns_resolver_resetmustbesecure(dns_resolver_t *resolver) {
8986
8987         REQUIRE(VALID_RESOLVER(resolver));
8988
8989 #if USE_MBSLOCK
8990         RWLOCK(&resolver->mbslock, isc_rwlocktype_write);
8991 #endif
8992         if (resolver->mustbesecure != NULL)
8993                 dns_rbt_destroy(&resolver->mustbesecure);
8994 #if USE_MBSLOCK
8995         RWUNLOCK(&resolver->mbslock, isc_rwlocktype_write);
8996 #endif
8997 }
8998
8999 static isc_boolean_t yes = ISC_TRUE, no = ISC_FALSE;
9000
9001 isc_result_t
9002 dns_resolver_setmustbesecure(dns_resolver_t *resolver, dns_name_t *name,
9003                              isc_boolean_t value)
9004 {
9005         isc_result_t result;
9006
9007         REQUIRE(VALID_RESOLVER(resolver));
9008
9009 #if USE_MBSLOCK
9010         RWLOCK(&resolver->mbslock, isc_rwlocktype_write);
9011 #endif
9012         if (resolver->mustbesecure == NULL) {
9013                 result = dns_rbt_create(resolver->mctx, NULL, NULL,
9014                                         &resolver->mustbesecure);
9015                 if (result != ISC_R_SUCCESS)
9016                         goto cleanup;
9017         }
9018         result = dns_rbt_addname(resolver->mustbesecure, name,
9019                                  value ? &yes : &no);
9020  cleanup:
9021 #if USE_MBSLOCK
9022         RWUNLOCK(&resolver->mbslock, isc_rwlocktype_write);
9023 #endif
9024         return (result);
9025 }
9026
9027 isc_boolean_t
9028 dns_resolver_getmustbesecure(dns_resolver_t *resolver, dns_name_t *name) {
9029         void *data = NULL;
9030         isc_boolean_t value = ISC_FALSE;
9031         isc_result_t result;
9032
9033         REQUIRE(VALID_RESOLVER(resolver));
9034
9035 #if USE_MBSLOCK
9036         RWLOCK(&resolver->mbslock, isc_rwlocktype_read);
9037 #endif
9038         if (resolver->mustbesecure == NULL)
9039                 goto unlock;
9040         result = dns_rbt_findname(resolver->mustbesecure, name, 0, NULL, &data);
9041         if (result == ISC_R_SUCCESS || result == DNS_R_PARTIALMATCH)
9042                 value = *(isc_boolean_t*)data;
9043  unlock:
9044 #if USE_MBSLOCK
9045         RWUNLOCK(&resolver->mbslock, isc_rwlocktype_read);
9046 #endif
9047         return (value);
9048 }
9049
9050 void
9051 dns_resolver_getclientsperquery(dns_resolver_t *resolver, isc_uint32_t *cur,
9052                                 isc_uint32_t *min, isc_uint32_t *max)
9053 {
9054         REQUIRE(VALID_RESOLVER(resolver));
9055
9056         LOCK(&resolver->lock);
9057         if (cur != NULL)
9058                 *cur = resolver->spillat;
9059         if (min != NULL)
9060                 *min = resolver->spillatmin;
9061         if (max != NULL)
9062                 *max = resolver->spillatmax;
9063         UNLOCK(&resolver->lock);
9064 }
9065
9066 void
9067 dns_resolver_setclientsperquery(dns_resolver_t *resolver, isc_uint32_t min,
9068                                 isc_uint32_t max)
9069 {
9070         REQUIRE(VALID_RESOLVER(resolver));
9071
9072         LOCK(&resolver->lock);
9073         resolver->spillatmin = resolver->spillat = min;
9074         resolver->spillatmax = max;
9075         UNLOCK(&resolver->lock);
9076 }
9077
9078 isc_boolean_t
9079 dns_resolver_getzeronosoattl(dns_resolver_t *resolver) {
9080         REQUIRE(VALID_RESOLVER(resolver));
9081
9082         return (resolver->zero_no_soa_ttl);
9083 }
9084
9085 void
9086 dns_resolver_setzeronosoattl(dns_resolver_t *resolver, isc_boolean_t state) {
9087         REQUIRE(VALID_RESOLVER(resolver));
9088
9089         resolver->zero_no_soa_ttl = state;
9090 }
9091
9092 unsigned int
9093 dns_resolver_getoptions(dns_resolver_t *resolver) {
9094         REQUIRE(VALID_RESOLVER(resolver));
9095
9096         return (resolver->options);
9097 }
9098
9099 unsigned int
9100 dns_resolver_gettimeout(dns_resolver_t *resolver) {
9101         REQUIRE(VALID_RESOLVER(resolver));
9102
9103         return (resolver->query_timeout);
9104 }
9105
9106 void
9107 dns_resolver_settimeout(dns_resolver_t *resolver, unsigned int seconds) {
9108         REQUIRE(VALID_RESOLVER(resolver));
9109
9110         if (seconds == 0)
9111                 seconds = DEFAULT_QUERY_TIMEOUT;
9112         if (seconds > MAXIMUM_QUERY_TIMEOUT)
9113                 seconds = MAXIMUM_QUERY_TIMEOUT;
9114         if (seconds < MINIMUM_QUERY_TIMEOUT)
9115                 seconds =  MINIMUM_QUERY_TIMEOUT;
9116
9117         resolver->query_timeout = seconds;
9118 }
9119
9120 void
9121 dns_resolver_setmaxdepth(dns_resolver_t *resolver, unsigned int maxdepth) {
9122         REQUIRE(VALID_RESOLVER(resolver));
9123         resolver->maxdepth = maxdepth;
9124 }
9125
9126 unsigned int
9127 dns_resolver_getmaxdepth(dns_resolver_t *resolver) {
9128         REQUIRE(VALID_RESOLVER(resolver));
9129         return (resolver->maxdepth);
9130 }