contrib/unbound/iterator/iterator.c

   1 /*
   2  * iterator/iterator.c - iterative resolver DNS query response module
   3  *
   4  * Copyright (c) 2007, NLnet Labs. All rights reserved.
   5  *
   6  * This software is open source.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * Redistributions of source code must retain the above copyright notice,
  13  * this list of conditions and the following disclaimer.
  14  *
  15  * Redistributions in binary form must reproduce the above copyright notice,
  16  * this list of conditions and the following disclaimer in the documentation
  17  * and/or other materials provided with the distribution.
  18  *
  19  * Neither the name of the NLNET LABS nor the names of its contributors may
  20  * be used to endorse or promote products derived from this software without
  21  * specific prior written permission.
  22  *
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  24  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  26  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  27  * HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  28  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED
  29  * TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
  30  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  31  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  32  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  33  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  34  */
  35
  36 /**
  37  * \file
  38  *
  39  * This file contains a module that performs recusive iterative DNS query
  40  * processing.
  41  */
  42
  43 #include "config.h"
  44 #include "iterator/iterator.h"
  45 #include "iterator/iter_utils.h"
  46 #include "iterator/iter_hints.h"
  47 #include "iterator/iter_fwd.h"
  48 #include "iterator/iter_donotq.h"
  49 #include "iterator/iter_delegpt.h"
  50 #include "iterator/iter_resptype.h"
  51 #include "iterator/iter_scrub.h"
  52 #include "iterator/iter_priv.h"
  53 #include "validator/val_neg.h"
  54 #include "services/cache/dns.h"
  55 #include "services/cache/infra.h"
  56 #include "util/module.h"
  57 #include "util/netevent.h"
  58 #include "util/net_help.h"
  59 #include "util/regional.h"
  60 #include "util/data/dname.h"
  61 #include "util/data/msgencode.h"
  62 #include "util/fptr_wlist.h"
  63 #include "util/config_file.h"
  64 #include "util/random.h"
  65 #include "sldns/rrdef.h"
  66 #include "sldns/wire2str.h"
  67 #include "sldns/str2wire.h"
  68 #include "sldns/parseutil.h"
  69 #include "sldns/sbuffer.h"
  70
  71 int
  72 iter_init(struct module_env* env, int id)
  73 {
  74         struct iter_env* iter_env = (struct iter_env*)calloc(1,
  75                 sizeof(struct iter_env));
  76         if(!iter_env) {
  77                 log_err("malloc failure");
  78                 return 0;
  79         }
  80         env->modinfo[id] = (void*)iter_env;
  81         if(!iter_apply_cfg(iter_env, env->cfg)) {
  82                 log_err("iterator: could not apply configuration settings.");
  83                 return 0;
  84         }
  85
  86         return 1;
  87 }
  88
  89 /** delete caps_whitelist element */
  90 static void
  91 caps_free(struct rbnode_t* n, void* ATTR_UNUSED(d))
  92 {
  93         if(n) {
  94                 free(((struct name_tree_node*)n)->name);
  95                 free(n);
  96         }
  97 }
  98
  99 void
 100 iter_deinit(struct module_env* env, int id)
 101 {
 102         struct iter_env* iter_env;
 103         if(!env || !env->modinfo[id])
 104                 return;
 105         iter_env = (struct iter_env*)env->modinfo[id];
 106         free(iter_env->target_fetch_policy);
 107         priv_delete(iter_env->priv);
 108         donotq_delete(iter_env->donotq);
 109         if(iter_env->caps_white) {
 110                 traverse_postorder(iter_env->caps_white, caps_free, NULL);
 111                 free(iter_env->caps_white);
 112         }
 113         free(iter_env);
 114         env->modinfo[id] = NULL;
 115 }
 116
 117 /** new query for iterator */
 118 static int
 119 iter_new(struct module_qstate* qstate, int id)
 120 {
 121         struct iter_qstate* iq = (struct iter_qstate*)regional_alloc(
 122                 qstate->region, sizeof(struct iter_qstate));
 123         qstate->minfo[id] = iq;
 124         if(!iq)
 125                 return 0;
 126         memset(iq, 0, sizeof(*iq));
 127         iq->state = INIT_REQUEST_STATE;
 128         iq->final_state = FINISHED_STATE;
 129         iq->an_prepend_list = NULL;
 130         iq->an_prepend_last = NULL;
 131         iq->ns_prepend_list = NULL;
 132         iq->ns_prepend_last = NULL;
 133         iq->dp = NULL;
 134         iq->depth = 0;
 135         iq->num_target_queries = 0;
 136         iq->num_current_queries = 0;
 137         iq->query_restart_count = 0;
 138         iq->referral_count = 0;
 139         iq->sent_count = 0;
 140         iq->ratelimit_ok = 0;
 141         iq->target_count = NULL;
 142         iq->wait_priming_stub = 0;
 143         iq->refetch_glue = 0;
 144         iq->dnssec_expected = 0;
 145         iq->dnssec_lame_query = 0;
 146         iq->chase_flags = qstate->query_flags;
 147         /* Start with the (current) qname. */
 148         iq->qchase = qstate->qinfo;
 149         outbound_list_init(&iq->outlist);
 150         iq->minimise_count = 0;
 151         iq->minimise_timeout_count = 0;
 152         if (qstate->env->cfg->qname_minimisation)
 153                 iq->minimisation_state = INIT_MINIMISE_STATE;
 154         else
 155                 iq->minimisation_state = DONOT_MINIMISE_STATE;
 156
 157         memset(&iq->qinfo_out, 0, sizeof(struct query_info));
 158         return 1;
 159 }
 160
 161 /**
 162  * Transition to the next state. This can be used to advance a currently
 163  * processing event. It cannot be used to reactivate a forEvent.
 164  *
 165  * @param iq: iterator query state
 166  * @param nextstate The state to transition to.
 167  * @return true. This is so this can be called as the return value for the
 168  *         actual process*State() methods. (Transitioning to the next state
 169  *         implies further processing).
 170  */
 171 static int
 172 next_state(struct iter_qstate* iq, enum iter_state nextstate)
 173 {
 174         /* If transitioning to a "response" state, make sure that there is a
 175          * response */
 176         if(iter_state_is_responsestate(nextstate)) {
 177                 if(iq->response == NULL) {
 178                         log_err("transitioning to response state sans "
 179                                 "response.");
 180                 }
 181         }
 182         iq->state = nextstate;
 183         return 1;
 184 }
 185
 186 /**
 187  * Transition an event to its final state. Final states always either return
 188  * a result up the module chain, or reactivate a dependent event. Which
 189  * final state to transition to is set in the module state for the event when
 190  * it was created, and depends on the original purpose of the event.
 191  *
 192  * The response is stored in the qstate->buf buffer.
 193  *
 194  * @param iq: iterator query state
 195  * @return false. This is so this method can be used as the return value for
 196  *         the processState methods. (Transitioning to the final state
 197  */
 198 static int
 199 final_state(struct iter_qstate* iq)
 200 {
 201         return next_state(iq, iq->final_state);
 202 }
 203
 204 /**
 205  * Callback routine to handle errors in parent query states
 206  * @param qstate: query state that failed.
 207  * @param id: module id.
 208  * @param super: super state.
 209  */
 210 static void
 211 error_supers(struct module_qstate* qstate, int id, struct module_qstate* super)
 212 {
 213         struct iter_qstate* super_iq = (struct iter_qstate*)super->minfo[id];
 214
 215         if(qstate->qinfo.qtype == LDNS_RR_TYPE_A ||
 216                 qstate->qinfo.qtype == LDNS_RR_TYPE_AAAA) {
 217                 /* mark address as failed. */
 218                 struct delegpt_ns* dpns = NULL;
 219                 super_iq->num_target_queries--;
 220                 if(super_iq->dp)
 221                         dpns = delegpt_find_ns(super_iq->dp,
 222                                 qstate->qinfo.qname, qstate->qinfo.qname_len);
 223                 if(!dpns) {
 224                         /* not interested */
 225                         verbose(VERB_ALGO, "subq error, but not interested");
 226                         log_query_info(VERB_ALGO, "superq", &super->qinfo);
 227                         if(super_iq->dp)
 228                                 delegpt_log(VERB_ALGO, super_iq->dp);
 229                         log_assert(0);
 230                         return;
 231                 } else {
 232                         /* see if the failure did get (parent-lame) info */
 233                         if(!cache_fill_missing(super->env, super_iq->qchase.qclass,
 234                                 super->region, super_iq->dp))
 235                                 log_err("out of memory adding missing");
 236                 }
 237                 dpns->resolved = 1; /* mark as failed */
 238         }
 239         if(qstate->qinfo.qtype == LDNS_RR_TYPE_NS) {
 240                 /* prime failed to get delegation */
 241                 super_iq->dp = NULL;
 242         }
 243         /* evaluate targets again */
 244         super_iq->state = QUERYTARGETS_STATE;
 245         /* super becomes runnable, and will process this change */
 246 }
 247
 248 /**
 249  * Return an error to the client
 250  * @param qstate: our query state
 251  * @param id: module id
 252  * @param rcode: error code (DNS errcode).
 253  * @return: 0 for use by caller, to make notation easy, like:
 254  *      return error_response(..).
 255  */
 256 static int
 257 error_response(struct module_qstate* qstate, int id, int rcode)
 258 {
 259         verbose(VERB_QUERY, "return error response %s",
 260                 sldns_lookup_by_id(sldns_rcodes, rcode)?
 261                 sldns_lookup_by_id(sldns_rcodes, rcode)->name:"??");
 262         qstate->return_rcode = rcode;
 263         qstate->return_msg = NULL;
 264         qstate->ext_state[id] = module_finished;
 265         return 0;
 266 }
 267
 268 /**
 269  * Return an error to the client and cache the error code in the
 270  * message cache (so per qname, qtype, qclass).
 271  * @param qstate: our query state
 272  * @param id: module id
 273  * @param rcode: error code (DNS errcode).
 274  * @return: 0 for use by caller, to make notation easy, like:
 275  *      return error_response(..).
 276  */
 277 static int
 278 error_response_cache(struct module_qstate* qstate, int id, int rcode)
 279 {
 280         if(!qstate->no_cache_store) {
 281                 /* store in cache */
 282                 struct reply_info err;
 283                 if(qstate->prefetch_leeway > NORR_TTL) {
 284                         verbose(VERB_ALGO, "error response for prefetch in cache");
 285                         /* attempt to adjust the cache entry prefetch */
 286                         if(dns_cache_prefetch_adjust(qstate->env, &qstate->qinfo,
 287                                 NORR_TTL, qstate->query_flags))
 288                                 return error_response(qstate, id, rcode);
 289                         /* if that fails (not in cache), fall through to store err */
 290                 }
 291                 memset(&err, 0, sizeof(err));
 292                 err.flags = (uint16_t)(BIT_QR | BIT_RA);
 293                 FLAGS_SET_RCODE(err.flags, rcode);
 294                 err.qdcount = 1;
 295                 err.ttl = NORR_TTL;
 296                 err.prefetch_ttl = PREFETCH_TTL_CALC(err.ttl);
 297                 /* do not waste time trying to validate this servfail */
 298                 err.security = sec_status_indeterminate;
 299                 verbose(VERB_ALGO, "store error response in message cache");
 300                 iter_dns_store(qstate->env, &qstate->qinfo, &err, 0, 0, 0, NULL,
 301                         qstate->query_flags);
 302         }
 303         return error_response(qstate, id, rcode);
 304 }
 305
 306 /** check if prepend item is duplicate item */
 307 static int
 308 prepend_is_duplicate(struct ub_packed_rrset_key** sets, size_t to,
 309         struct ub_packed_rrset_key* dup)
 310 {
 311         size_t i;
 312         for(i=0; i<to; i++) {
 313                 if(sets[i]->rk.type == dup->rk.type &&
 314                         sets[i]->rk.rrset_class == dup->rk.rrset_class &&
 315                         sets[i]->rk.dname_len == dup->rk.dname_len &&
 316                         query_dname_compare(sets[i]->rk.dname, dup->rk.dname)
 317                         == 0)
 318                         return 1;
 319         }
 320         return 0;
 321 }
 322
 323 /** prepend the prepend list in the answer and authority section of dns_msg */
 324 static int
 325 iter_prepend(struct iter_qstate* iq, struct dns_msg* msg,
 326         struct regional* region)
 327 {
 328         struct iter_prep_list* p;
 329         struct ub_packed_rrset_key** sets;
 330         size_t num_an = 0, num_ns = 0;;
 331         for(p = iq->an_prepend_list; p; p = p->next)
 332                 num_an++;
 333         for(p = iq->ns_prepend_list; p; p = p->next)
 334                 num_ns++;
 335         if(num_an + num_ns == 0)
 336                 return 1;
 337         verbose(VERB_ALGO, "prepending %d rrsets", (int)num_an + (int)num_ns);
 338         if(num_an > RR_COUNT_MAX || num_ns > RR_COUNT_MAX ||
 339                 msg->rep->rrset_count > RR_COUNT_MAX) return 0; /* overflow */
 340         sets = regional_alloc(region, (num_an+num_ns+msg->rep->rrset_count) *
 341                 sizeof(struct ub_packed_rrset_key*));
 342         if(!sets)
 343                 return 0;
 344         /* ANSWER section */
 345         num_an = 0;
 346         for(p = iq->an_prepend_list; p; p = p->next) {
 347                 sets[num_an++] = p->rrset;
 348         }
 349         memcpy(sets+num_an, msg->rep->rrsets, msg->rep->an_numrrsets *
 350                 sizeof(struct ub_packed_rrset_key*));
 351         /* AUTH section */
 352         num_ns = 0;
 353         for(p = iq->ns_prepend_list; p; p = p->next) {
 354                 if(prepend_is_duplicate(sets+msg->rep->an_numrrsets+num_an,
 355                         num_ns, p->rrset) || prepend_is_duplicate(
 356                         msg->rep->rrsets+msg->rep->an_numrrsets,
 357                         msg->rep->ns_numrrsets, p->rrset))
 358                         continue;
 359                 sets[msg->rep->an_numrrsets + num_an + num_ns++] = p->rrset;
 360         }
 361         memcpy(sets + num_an + msg->rep->an_numrrsets + num_ns,
 362                 msg->rep->rrsets + msg->rep->an_numrrsets,
 363                 (msg->rep->ns_numrrsets + msg->rep->ar_numrrsets) *
 364                 sizeof(struct ub_packed_rrset_key*));
 365
 366         /* NXDOMAIN rcode can stay if we prepended DNAME/CNAMEs, because
 367          * this is what recursors should give. */
 368         msg->rep->rrset_count += num_an + num_ns;
 369         msg->rep->an_numrrsets += num_an;
 370         msg->rep->ns_numrrsets += num_ns;
 371         msg->rep->rrsets = sets;
 372         return 1;
 373 }
 374
 375 /**
 376  * Add rrset to ANSWER prepend list
 377  * @param qstate: query state.
 378  * @param iq: iterator query state.
 379  * @param rrset: rrset to add.
 380  * @return false on failure (malloc).
 381  */
 382 static int
 383 iter_add_prepend_answer(struct module_qstate* qstate, struct iter_qstate* iq,
 384         struct ub_packed_rrset_key* rrset)
 385 {
 386         struct iter_prep_list* p = (struct iter_prep_list*)regional_alloc(
 387                 qstate->region, sizeof(struct iter_prep_list));
 388         if(!p)
 389                 return 0;
 390         p->rrset = rrset;
 391         p->next = NULL;
 392         /* add at end */
 393         if(iq->an_prepend_last)
 394                 iq->an_prepend_last->next = p;
 395         else    iq->an_prepend_list = p;
 396         iq->an_prepend_last = p;
 397         return 1;
 398 }
 399
 400 /**
 401  * Add rrset to AUTHORITY prepend list
 402  * @param qstate: query state.
 403  * @param iq: iterator query state.
 404  * @param rrset: rrset to add.
 405  * @return false on failure (malloc).
 406  */
 407 static int
 408 iter_add_prepend_auth(struct module_qstate* qstate, struct iter_qstate* iq,
 409         struct ub_packed_rrset_key* rrset)
 410 {
 411         struct iter_prep_list* p = (struct iter_prep_list*)regional_alloc(
 412                 qstate->region, sizeof(struct iter_prep_list));
 413         if(!p)
 414                 return 0;
 415         p->rrset = rrset;
 416         p->next = NULL;
 417         /* add at end */
 418         if(iq->ns_prepend_last)
 419                 iq->ns_prepend_last->next = p;
 420         else    iq->ns_prepend_list = p;
 421         iq->ns_prepend_last = p;
 422         return 1;
 423 }
 424
 425 /**
 426  * Given a CNAME response (defined as a response containing a CNAME or DNAME
 427  * that does not answer the request), process the response, modifying the
 428  * state as necessary. This follows the CNAME/DNAME chain and returns the
 429  * final query name.
 430  *
 431  * sets the new query name, after following the CNAME/DNAME chain.
 432  * @param qstate: query state.
 433  * @param iq: iterator query state.
 434  * @param msg: the response.
 435  * @param mname: returned target new query name.
 436  * @param mname_len: length of mname.
 437  * @return false on (malloc) error.
 438  */
 439 static int
 440 handle_cname_response(struct module_qstate* qstate, struct iter_qstate* iq,
 441         struct dns_msg* msg, uint8_t** mname, size_t* mname_len)
 442 {
 443         size_t i;
 444         /* Start with the (current) qname. */
 445         *mname = iq->qchase.qname;
 446         *mname_len = iq->qchase.qname_len;
 447
 448         /* Iterate over the ANSWER rrsets in order, looking for CNAMEs and
 449          * DNAMES. */
 450         for(i=0; i<msg->rep->an_numrrsets; i++) {
 451                 struct ub_packed_rrset_key* r = msg->rep->rrsets[i];
 452                 /* If there is a (relevant) DNAME, add it to the list.
 453                  * We always expect there to be CNAME that was generated
 454                  * by this DNAME following, so we don't process the DNAME
 455                  * directly.  */
 456                 if(ntohs(r->rk.type) == LDNS_RR_TYPE_DNAME &&
 457                         dname_strict_subdomain_c(*mname, r->rk.dname)) {
 458                         if(!iter_add_prepend_answer(qstate, iq, r))
 459                                 return 0;
 460                         continue;
 461                 }
 462
 463                 if(ntohs(r->rk.type) == LDNS_RR_TYPE_CNAME &&
 464                         query_dname_compare(*mname, r->rk.dname) == 0) {
 465                         /* Add this relevant CNAME rrset to the prepend list.*/
 466                         if(!iter_add_prepend_answer(qstate, iq, r))
 467                                 return 0;
 468                         get_cname_target(r, mname, mname_len);
 469                 }
 470
 471                 /* Other rrsets in the section are ignored. */
 472         }
 473         /* add authority rrsets to authority prepend, for wildcarded CNAMEs */
 474         for(i=msg->rep->an_numrrsets; i<msg->rep->an_numrrsets +
 475                 msg->rep->ns_numrrsets; i++) {
 476                 struct ub_packed_rrset_key* r = msg->rep->rrsets[i];
 477                 /* only add NSEC/NSEC3, as they may be needed for validation */
 478                 if(ntohs(r->rk.type) == LDNS_RR_TYPE_NSEC ||
 479                         ntohs(r->rk.type) == LDNS_RR_TYPE_NSEC3) {
 480                         if(!iter_add_prepend_auth(qstate, iq, r))
 481                                 return 0;
 482                 }
 483         }
 484         return 1;
 485 }
 486
 487 /** see if target name is caps-for-id whitelisted */
 488 static int
 489 is_caps_whitelisted(struct iter_env* ie, struct iter_qstate* iq)
 490 {
 491         if(!ie->caps_white) return 0; /* no whitelist, or no capsforid */
 492         return name_tree_lookup(ie->caps_white, iq->qchase.qname,
 493                 iq->qchase.qname_len, dname_count_labels(iq->qchase.qname),
 494                 iq->qchase.qclass) != NULL;
 495 }
 496
 497 /** create target count structure for this query */
 498 static void
 499 target_count_create(struct iter_qstate* iq)
 500 {
 501         if(!iq->target_count) {
 502                 iq->target_count = (int*)calloc(2, sizeof(int));
 503                 /* if calloc fails we simply do not track this number */
 504                 if(iq->target_count)
 505                         iq->target_count[0] = 1;
 506         }
 507 }
 508
 509 static void
 510 target_count_increase(struct iter_qstate* iq, int num)
 511 {
 512         target_count_create(iq);
 513         if(iq->target_count)
 514                 iq->target_count[1] += num;
 515 }
 516
 517 /**
 518  * Generate a subrequest.
 519  * Generate a local request event. Local events are tied to this module, and
 520  * have a corresponding (first tier) event that is waiting for this event to
 521  * resolve to continue.
 522  *
 523  * @param qname The query name for this request.
 524  * @param qnamelen length of qname
 525  * @param qtype The query type for this request.
 526  * @param qclass The query class for this request.
 527  * @param qstate The event that is generating this event.
 528  * @param id: module id.
 529  * @param iq: The iterator state that is generating this event.
 530  * @param initial_state The initial response state (normally this
 531  *          is QUERY_RESP_STATE, unless it is known that the request won't
 532  *          need iterative processing
 533  * @param finalstate The final state for the response to this request.
 534  * @param subq_ret: if newly allocated, the subquerystate, or NULL if it does
 535  *      not need initialisation.
 536  * @param v: if true, validation is done on the subquery.
 537  * @return false on error (malloc).
 538  */
 539 static int
 540 generate_sub_request(uint8_t* qname, size_t qnamelen, uint16_t qtype,
 541         uint16_t qclass, struct module_qstate* qstate, int id,
 542         struct iter_qstate* iq, enum iter_state initial_state,
 543         enum iter_state finalstate, struct module_qstate** subq_ret, int v)
 544 {
 545         struct module_qstate* subq = NULL;
 546         struct iter_qstate* subiq = NULL;
 547         uint16_t qflags = 0; /* OPCODE QUERY, no flags */
 548         struct query_info qinf;
 549         int prime = (finalstate == PRIME_RESP_STATE)?1:0;
 550         int valrec = 0;
 551         qinf.qname = qname;
 552         qinf.qname_len = qnamelen;
 553         qinf.qtype = qtype;
 554         qinf.qclass = qclass;
 555         qinf.local_alias = NULL;
 556
 557         /* RD should be set only when sending the query back through the INIT
 558          * state. */
 559         if(initial_state == INIT_REQUEST_STATE)
 560                 qflags |= BIT_RD;
 561         /* We set the CD flag so we can send this through the "head" of
 562          * the resolution chain, which might have a validator. We are
 563          * uninterested in validating things not on the direct resolution
 564          * path.  */
 565         if(!v) {
 566                 qflags |= BIT_CD;
 567                 valrec = 1;
 568         }
 569
 570         /* attach subquery, lookup existing or make a new one */
 571         fptr_ok(fptr_whitelist_modenv_attach_sub(qstate->env->attach_sub));
 572         if(!(*qstate->env->attach_sub)(qstate, &qinf, qflags, prime, valrec,
 573                 &subq)) {
 574                 return 0;
 575         }
 576         *subq_ret = subq;
 577         if(subq) {
 578                 /* initialise the new subquery */
 579                 subq->curmod = id;
 580                 subq->ext_state[id] = module_state_initial;
 581                 subq->minfo[id] = regional_alloc(subq->region,
 582                         sizeof(struct iter_qstate));
 583                 if(!subq->minfo[id]) {
 584                         log_err("init subq: out of memory");
 585                         fptr_ok(fptr_whitelist_modenv_kill_sub(
 586                                 qstate->env->kill_sub));
 587                         (*qstate->env->kill_sub)(subq);
 588                         return 0;
 589                 }
 590                 subiq = (struct iter_qstate*)subq->minfo[id];
 591                 memset(subiq, 0, sizeof(*subiq));
 592                 subiq->num_target_queries = 0;
 593                 target_count_create(iq);
 594                 subiq->target_count = iq->target_count;
 595                 if(iq->target_count)
 596                         iq->target_count[0] ++; /* extra reference */
 597                 subiq->num_current_queries = 0;
 598                 subiq->depth = iq->depth+1;
 599                 outbound_list_init(&subiq->outlist);
 600                 subiq->state = initial_state;
 601                 subiq->final_state = finalstate;
 602                 subiq->qchase = subq->qinfo;
 603                 subiq->chase_flags = subq->query_flags;
 604                 subiq->refetch_glue = 0;
 605                 if(qstate->env->cfg->qname_minimisation)
 606                         subiq->minimisation_state = INIT_MINIMISE_STATE;
 607                 else
 608                         subiq->minimisation_state = DONOT_MINIMISE_STATE;
 609                 memset(&subiq->qinfo_out, 0, sizeof(struct query_info));
 610         }
 611         return 1;
 612 }
 613
 614 /**
 615  * Generate and send a root priming request.
 616  * @param qstate: the qtstate that triggered the need to prime.
 617  * @param iq: iterator query state.
 618  * @param id: module id.
 619  * @param qclass: the class to prime.
 620  * @return 0 on failure
 621  */
 622 static int
 623 prime_root(struct module_qstate* qstate, struct iter_qstate* iq, int id,
 624         uint16_t qclass)
 625 {
 626         struct delegpt* dp;
 627         struct module_qstate* subq;
 628         verbose(VERB_DETAIL, "priming . %s NS",
 629                 sldns_lookup_by_id(sldns_rr_classes, (int)qclass)?
 630                 sldns_lookup_by_id(sldns_rr_classes, (int)qclass)->name:"??");
 631         dp = hints_lookup_root(qstate->env->hints, qclass);
 632         if(!dp) {
 633                 verbose(VERB_ALGO, "Cannot prime due to lack of hints");
 634                 return 0;
 635         }
 636         /* Priming requests start at the QUERYTARGETS state, skipping
 637          * the normal INIT state logic (which would cause an infloop). */
 638         if(!generate_sub_request((uint8_t*)"\000", 1, LDNS_RR_TYPE_NS,
 639                 qclass, qstate, id, iq, QUERYTARGETS_STATE, PRIME_RESP_STATE,
 640                 &subq, 0)) {
 641                 verbose(VERB_ALGO, "could not prime root");
 642                 return 0;
 643         }
 644         if(subq) {
 645                 struct iter_qstate* subiq =
 646                         (struct iter_qstate*)subq->minfo[id];
 647                 /* Set the initial delegation point to the hint.
 648                  * copy dp, it is now part of the root prime query.
 649                  * dp was part of in the fixed hints structure. */
 650                 subiq->dp = delegpt_copy(dp, subq->region);
 651                 if(!subiq->dp) {
 652                         log_err("out of memory priming root, copydp");
 653                         fptr_ok(fptr_whitelist_modenv_kill_sub(
 654                                 qstate->env->kill_sub));
 655                         (*qstate->env->kill_sub)(subq);
 656                         return 0;
 657                 }
 658                 /* there should not be any target queries. */
 659                 subiq->num_target_queries = 0;
 660                 subiq->dnssec_expected = iter_indicates_dnssec(
 661                         qstate->env, subiq->dp, NULL, subq->qinfo.qclass);
 662         }
 663
 664         /* this module stops, our submodule starts, and does the query. */
 665         qstate->ext_state[id] = module_wait_subquery;
 666         return 1;
 667 }
 668
 669 /**
 670  * Generate and process a stub priming request. This method tests for the
 671  * need to prime a stub zone, so it is safe to call for every request.
 672  *
 673  * @param qstate: the qtstate that triggered the need to prime.
 674  * @param iq: iterator query state.
 675  * @param id: module id.
 676  * @param qname: request name.
 677  * @param qclass: request class.
 678  * @return true if a priming subrequest was made, false if not. The will only
 679  *         issue a priming request if it detects an unprimed stub.
 680  *         Uses value of 2 to signal during stub-prime in root-prime situation
 681  *         that a noprime-stub is available and resolution can continue.
 682  */
 683 static int
 684 prime_stub(struct module_qstate* qstate, struct iter_qstate* iq, int id,
 685         uint8_t* qname, uint16_t qclass)
 686 {
 687         /* Lookup the stub hint. This will return null if the stub doesn't
 688          * need to be re-primed. */
 689         struct iter_hints_stub* stub;
 690         struct delegpt* stub_dp;
 691         struct module_qstate* subq;
 692
 693         if(!qname) return 0;
 694         stub = hints_lookup_stub(qstate->env->hints, qname, qclass, iq->dp);
 695         /* The stub (if there is one) does not need priming. */
 696         if(!stub)
 697                 return 0;
 698         stub_dp = stub->dp;
 699
 700         /* is it a noprime stub (always use) */
 701         if(stub->noprime) {
 702                 int r = 0;
 703                 if(iq->dp == NULL) r = 2;
 704                 /* copy the dp out of the fixed hints structure, so that
 705                  * it can be changed when servicing this query */
 706                 iq->dp = delegpt_copy(stub_dp, qstate->region);
 707                 if(!iq->dp) {
 708                         log_err("out of memory priming stub");
 709                         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 710                         return 1; /* return 1 to make module stop, with error */
 711                 }
 712                 log_nametypeclass(VERB_DETAIL, "use stub", stub_dp->name,
 713                         LDNS_RR_TYPE_NS, qclass);
 714                 return r;
 715         }
 716
 717         /* Otherwise, we need to (re)prime the stub. */
 718         log_nametypeclass(VERB_DETAIL, "priming stub", stub_dp->name,
 719                 LDNS_RR_TYPE_NS, qclass);
 720
 721         /* Stub priming events start at the QUERYTARGETS state to avoid the
 722          * redundant INIT state processing. */
 723         if(!generate_sub_request(stub_dp->name, stub_dp->namelen,
 724                 LDNS_RR_TYPE_NS, qclass, qstate, id, iq,
 725                 QUERYTARGETS_STATE, PRIME_RESP_STATE, &subq, 0)) {
 726                 verbose(VERB_ALGO, "could not prime stub");
 727                 (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 728                 return 1; /* return 1 to make module stop, with error */
 729         }
 730         if(subq) {
 731                 struct iter_qstate* subiq =
 732                         (struct iter_qstate*)subq->minfo[id];
 733
 734                 /* Set the initial delegation point to the hint. */
 735                 /* make copy to avoid use of stub dp by different qs/threads */
 736                 subiq->dp = delegpt_copy(stub_dp, subq->region);
 737                 if(!subiq->dp) {
 738                         log_err("out of memory priming stub, copydp");
 739                         fptr_ok(fptr_whitelist_modenv_kill_sub(
 740                                 qstate->env->kill_sub));
 741                         (*qstate->env->kill_sub)(subq);
 742                         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 743                         return 1; /* return 1 to make module stop, with error */
 744                 }
 745                 /* there should not be any target queries -- although there
 746                  * wouldn't be anyway, since stub hints never have
 747                  * missing targets. */
 748                 subiq->num_target_queries = 0;
 749                 subiq->wait_priming_stub = 1;
 750                 subiq->dnssec_expected = iter_indicates_dnssec(
 751                         qstate->env, subiq->dp, NULL, subq->qinfo.qclass);
 752         }
 753
 754         /* this module stops, our submodule starts, and does the query. */
 755         qstate->ext_state[id] = module_wait_subquery;
 756         return 1;
 757 }
 758
 759 /**
 760  * Generate A and AAAA checks for glue that is in-zone for the referral
 761  * we just got to obtain authoritative information on the adresses.
 762  *
 763  * @param qstate: the qtstate that triggered the need to prime.
 764  * @param iq: iterator query state.
 765  * @param id: module id.
 766  */
 767 static void
 768 generate_a_aaaa_check(struct module_qstate* qstate, struct iter_qstate* iq,
 769         int id)
 770 {
 771         struct iter_env* ie = (struct iter_env*)qstate->env->modinfo[id];
 772         struct module_qstate* subq;
 773         size_t i;
 774         struct reply_info* rep = iq->response->rep;
 775         struct ub_packed_rrset_key* s;
 776         log_assert(iq->dp);
 777
 778         if(iq->depth == ie->max_dependency_depth)
 779                 return;
 780         /* walk through additional, and check if in-zone,
 781          * only relevant A, AAAA are left after scrub anyway */
 782         for(i=rep->an_numrrsets+rep->ns_numrrsets; i<rep->rrset_count; i++) {
 783                 s = rep->rrsets[i];
 784                 /* check *ALL* addresses that are transmitted in additional*/
 785                 /* is it an address ? */
 786                 if( !(ntohs(s->rk.type)==LDNS_RR_TYPE_A ||
 787                         ntohs(s->rk.type)==LDNS_RR_TYPE_AAAA)) {
 788                         continue;
 789                 }
 790                 /* is this query the same as the A/AAAA check for it */
 791                 if(qstate->qinfo.qtype == ntohs(s->rk.type) &&
 792                         qstate->qinfo.qclass == ntohs(s->rk.rrset_class) &&
 793                         query_dname_compare(qstate->qinfo.qname,
 794                                 s->rk.dname)==0 &&
 795                         (qstate->query_flags&BIT_RD) &&
 796                         !(qstate->query_flags&BIT_CD))
 797                         continue;
 798
 799                 /* generate subrequest for it */
 800                 log_nametypeclass(VERB_ALGO, "schedule addr fetch",
 801                         s->rk.dname, ntohs(s->rk.type),
 802                         ntohs(s->rk.rrset_class));
 803                 if(!generate_sub_request(s->rk.dname, s->rk.dname_len,
 804                         ntohs(s->rk.type), ntohs(s->rk.rrset_class),
 805                         qstate, id, iq,
 806                         INIT_REQUEST_STATE, FINISHED_STATE, &subq, 1)) {
 807                         verbose(VERB_ALGO, "could not generate addr check");
 808                         return;
 809                 }
 810                 /* ignore subq - not need for more init */
 811         }
 812 }
 813
 814 /**
 815  * Generate a NS check request to obtain authoritative information
 816  * on an NS rrset.
 817  *
 818  * @param qstate: the qtstate that triggered the need to prime.
 819  * @param iq: iterator query state.
 820  * @param id: module id.
 821  */
 822 static void
 823 generate_ns_check(struct module_qstate* qstate, struct iter_qstate* iq, int id)
 824 {
 825         struct iter_env* ie = (struct iter_env*)qstate->env->modinfo[id];
 826         struct module_qstate* subq;
 827         log_assert(iq->dp);
 828
 829         if(iq->depth == ie->max_dependency_depth)
 830                 return;
 831         /* is this query the same as the nscheck? */
 832         if(qstate->qinfo.qtype == LDNS_RR_TYPE_NS &&
 833                 query_dname_compare(iq->dp->name, qstate->qinfo.qname)==0 &&
 834                 (qstate->query_flags&BIT_RD) && !(qstate->query_flags&BIT_CD)){
 835                 /* spawn off A, AAAA queries for in-zone glue to check */
 836                 generate_a_aaaa_check(qstate, iq, id);
 837                 return;
 838         }
 839
 840         log_nametypeclass(VERB_ALGO, "schedule ns fetch",
 841                 iq->dp->name, LDNS_RR_TYPE_NS, iq->qchase.qclass);
 842         if(!generate_sub_request(iq->dp->name, iq->dp->namelen,
 843                 LDNS_RR_TYPE_NS, iq->qchase.qclass, qstate, id, iq,
 844                 INIT_REQUEST_STATE, FINISHED_STATE, &subq, 1)) {
 845                 verbose(VERB_ALGO, "could not generate ns check");
 846                 return;
 847         }
 848         if(subq) {
 849                 struct iter_qstate* subiq =
 850                         (struct iter_qstate*)subq->minfo[id];
 851
 852                 /* make copy to avoid use of stub dp by different qs/threads */
 853                 /* refetch glue to start higher up the tree */
 854                 subiq->refetch_glue = 1;
 855                 subiq->dp = delegpt_copy(iq->dp, subq->region);
 856                 if(!subiq->dp) {
 857                         log_err("out of memory generating ns check, copydp");
 858                         fptr_ok(fptr_whitelist_modenv_kill_sub(
 859                                 qstate->env->kill_sub));
 860                         (*qstate->env->kill_sub)(subq);
 861                         return;
 862                 }
 863         }
 864 }
 865
 866 /**
 867  * Generate a DNSKEY prefetch query to get the DNSKEY for the DS record we
 868  * just got in a referral (where we have dnssec_expected, thus have trust
 869  * anchors above it).  Note that right after calling this routine the
 870  * iterator detached subqueries (because of following the referral), and thus
 871  * the DNSKEY query becomes detached, its return stored in the cache for
 872  * later lookup by the validator.  This cache lookup by the validator avoids
 873  * the roundtrip incurred by the DNSKEY query.  The DNSKEY query is now
 874  * performed at about the same time the original query is sent to the domain,
 875  * thus the two answers are likely to be returned at about the same time,
 876  * saving a roundtrip from the validated lookup.
 877  *
 878  * @param qstate: the qtstate that triggered the need to prime.
 879  * @param iq: iterator query state.
 880  * @param id: module id.
 881  */
 882 static void
 883 generate_dnskey_prefetch(struct module_qstate* qstate,
 884         struct iter_qstate* iq, int id)
 885 {
 886         struct module_qstate* subq;
 887         log_assert(iq->dp);
 888
 889         /* is this query the same as the prefetch? */
 890         if(qstate->qinfo.qtype == LDNS_RR_TYPE_DNSKEY &&
 891                 query_dname_compare(iq->dp->name, qstate->qinfo.qname)==0 &&
 892                 (qstate->query_flags&BIT_RD) && !(qstate->query_flags&BIT_CD)){
 893                 return;
 894         }
 895
 896         /* if the DNSKEY is in the cache this lookup will stop quickly */
 897         log_nametypeclass(VERB_ALGO, "schedule dnskey prefetch",
 898                 iq->dp->name, LDNS_RR_TYPE_DNSKEY, iq->qchase.qclass);
 899         if(!generate_sub_request(iq->dp->name, iq->dp->namelen,
 900                 LDNS_RR_TYPE_DNSKEY, iq->qchase.qclass, qstate, id, iq,
 901                 INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0)) {
 902                 /* we'll be slower, but it'll work */
 903                 verbose(VERB_ALGO, "could not generate dnskey prefetch");
 904                 return;
 905         }
 906         if(subq) {
 907                 struct iter_qstate* subiq =
 908                         (struct iter_qstate*)subq->minfo[id];
 909                 /* this qstate has the right delegation for the dnskey lookup*/
 910                 /* make copy to avoid use of stub dp by different qs/threads */
 911                 subiq->dp = delegpt_copy(iq->dp, subq->region);
 912                 /* if !subiq->dp, it'll start from the cache, no problem */
 913         }
 914 }
 915
 916 /**
 917  * See if the query needs forwarding.
 918  *
 919  * @param qstate: query state.
 920  * @param iq: iterator query state.
 921  * @return true if the request is forwarded, false if not.
 922  *      If returns true but, iq->dp is NULL then a malloc failure occurred.
 923  */
 924 static int
 925 forward_request(struct module_qstate* qstate, struct iter_qstate* iq)
 926 {
 927         struct delegpt* dp;
 928         uint8_t* delname = iq->qchase.qname;
 929         size_t delnamelen = iq->qchase.qname_len;
 930         if(iq->refetch_glue) {
 931                 delname = iq->dp->name;
 932                 delnamelen = iq->dp->namelen;
 933         }
 934         /* strip one label off of DS query to lookup higher for it */
 935         if( (iq->qchase.qtype == LDNS_RR_TYPE_DS || iq->refetch_glue)
 936                 && !dname_is_root(iq->qchase.qname))
 937                 dname_remove_label(&delname, &delnamelen);
 938         dp = forwards_lookup(qstate->env->fwds, delname, iq->qchase.qclass);
 939         if(!dp)
 940                 return 0;
 941         /* send recursion desired to forward addr */
 942         iq->chase_flags |= BIT_RD;
 943         iq->dp = delegpt_copy(dp, qstate->region);
 944         /* iq->dp checked by caller */
 945         verbose(VERB_ALGO, "forwarding request");
 946         return 1;
 947 }
 948
 949 /**
 950  * Process the initial part of the request handling. This state roughly
 951  * corresponds to resolver algorithms steps 1 (find answer in cache) and 2
 952  * (find the best servers to ask).
 953  *
 954  * Note that all requests start here, and query restarts revisit this state.
 955  *
 956  * This state either generates: 1) a response, from cache or error, 2) a
 957  * priming event, or 3) forwards the request to the next state (init2,
 958  * generally).
 959  *
 960  * @param qstate: query state.
 961  * @param iq: iterator query state.
 962  * @param ie: iterator shared global environment.
 963  * @param id: module id.
 964  * @return true if the event needs more request processing immediately,
 965  *         false if not.
 966  */
 967 static int
 968 processInitRequest(struct module_qstate* qstate, struct iter_qstate* iq,
 969         struct iter_env* ie, int id)
 970 {
 971         uint8_t* delname;
 972         size_t delnamelen;
 973         struct dns_msg* msg = NULL;
 974
 975         log_query_info(VERB_DETAIL, "resolving", &qstate->qinfo);
 976         /* check effort */
 977
 978         /* We enforce a maximum number of query restarts. This is primarily a
 979          * cheap way to prevent CNAME loops. */
 980         if(iq->query_restart_count > MAX_RESTART_COUNT) {
 981                 verbose(VERB_QUERY, "request has exceeded the maximum number"
 982                         " of query restarts with %d", iq->query_restart_count);
 983                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 984         }
 985
 986         /* We enforce a maximum recursion/dependency depth -- in general,
 987          * this is unnecessary for dependency loops (although it will
 988          * catch those), but it provides a sensible limit to the amount
 989          * of work required to answer a given query. */
 990         verbose(VERB_ALGO, "request has dependency depth of %d", iq->depth);
 991         if(iq->depth > ie->max_dependency_depth) {
 992                 verbose(VERB_QUERY, "request has exceeded the maximum "
 993                         "dependency depth with depth of %d", iq->depth);
 994                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 995         }
 996
 997         /* If the request is qclass=ANY, setup to generate each class */
 998         if(qstate->qinfo.qclass == LDNS_RR_CLASS_ANY) {
 999                 iq->qchase.qclass = 0;
1000                 return next_state(iq, COLLECT_CLASS_STATE);
1001         }
1002
1003         /* Resolver Algorithm Step 1 -- Look for the answer in local data. */
1004
1005         /* This either results in a query restart (CNAME cache response), a
1006          * terminating response (ANSWER), or a cache miss (null). */
1007
1008         if(qstate->blacklist) {
1009                 /* if cache, or anything else, was blacklisted then
1010                  * getting older results from cache is a bad idea, no cache */
1011                 verbose(VERB_ALGO, "cache blacklisted, going to the network");
1012                 msg = NULL;
1013         } else if(!qstate->no_cache_lookup) {
1014                 msg = dns_cache_lookup(qstate->env, iq->qchase.qname,
1015                         iq->qchase.qname_len, iq->qchase.qtype,
1016                         iq->qchase.qclass, qstate->query_flags,
1017                         qstate->region, qstate->env->scratch);
1018                 if(!msg && qstate->env->neg_cache) {
1019                         /* lookup in negative cache; may result in
1020                          * NOERROR/NODATA or NXDOMAIN answers that need validation */
1021                         msg = val_neg_getmsg(qstate->env->neg_cache, &iq->qchase,
1022                                 qstate->region, qstate->env->rrset_cache,
1023                                 qstate->env->scratch_buffer,
1024                                 *qstate->env->now, 1/*add SOA*/, NULL);
1025                 }
1026                 /* item taken from cache does not match our query name, thus
1027                  * security needs to be re-examined later */
1028                 if(msg && query_dname_compare(qstate->qinfo.qname,
1029                         iq->qchase.qname) != 0)
1030                         msg->rep->security = sec_status_unchecked;
1031         }
1032         if(msg) {
1033                 /* handle positive cache response */
1034                 enum response_type type = response_type_from_cache(msg,
1035                         &iq->qchase);
1036                 if(verbosity >= VERB_ALGO) {
1037                         log_dns_msg("msg from cache lookup", &msg->qinfo,
1038                                 msg->rep);
1039                         verbose(VERB_ALGO, "msg ttl is %d, prefetch ttl %d",
1040                                 (int)msg->rep->ttl,
1041                                 (int)msg->rep->prefetch_ttl);
1042                 }
1043
1044                 if(type == RESPONSE_TYPE_CNAME) {
1045                         uint8_t* sname = 0;
1046                         size_t slen = 0;
1047                         verbose(VERB_ALGO, "returning CNAME response from "
1048                                 "cache");
1049                         if(!handle_cname_response(qstate, iq, msg,
1050                                 &sname, &slen))
1051                                 return error_response(qstate, id,
1052                                         LDNS_RCODE_SERVFAIL);
1053                         iq->qchase.qname = sname;
1054                         iq->qchase.qname_len = slen;
1055                         /* This *is* a query restart, even if it is a cheap
1056                          * one. */
1057                         iq->dp = NULL;
1058                         iq->refetch_glue = 0;
1059                         iq->query_restart_count++;
1060                         iq->sent_count = 0;
1061                         sock_list_insert(&qstate->reply_origin, NULL, 0, qstate->region);
1062                         if(qstate->env->cfg->qname_minimisation)
1063                                 iq->minimisation_state = INIT_MINIMISE_STATE;
1064                         return next_state(iq, INIT_REQUEST_STATE);
1065                 }
1066
1067                 /* if from cache, NULL, else insert 'cache IP' len=0 */
1068                 if(qstate->reply_origin)
1069                         sock_list_insert(&qstate->reply_origin, NULL, 0, qstate->region);
1070                 /* it is an answer, response, to final state */
1071                 verbose(VERB_ALGO, "returning answer from cache.");
1072                 iq->response = msg;
1073                 return final_state(iq);
1074         }
1075
1076         /* attempt to forward the request */
1077         if(forward_request(qstate, iq))
1078         {
1079                 if(!iq->dp) {
1080                         log_err("alloc failure for forward dp");
1081                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1082                 }
1083                 iq->refetch_glue = 0;
1084                 iq->minimisation_state = DONOT_MINIMISE_STATE;
1085                 /* the request has been forwarded.
1086                  * forwarded requests need to be immediately sent to the
1087                  * next state, QUERYTARGETS. */
1088                 return next_state(iq, QUERYTARGETS_STATE);
1089         }
1090
1091         /* Resolver Algorithm Step 2 -- find the "best" servers. */
1092
1093         /* first, adjust for DS queries. To avoid the grandparent problem,
1094          * we just look for the closest set of server to the parent of qname.
1095          * When re-fetching glue we also need to ask the parent.
1096          */
1097         if(iq->refetch_glue) {
1098                 if(!iq->dp) {
1099                         log_err("internal or malloc fail: no dp for refetch");
1100                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1101                 }
1102                 delname = iq->dp->name;
1103                 delnamelen = iq->dp->namelen;
1104         } else {
1105                 delname = iq->qchase.qname;
1106                 delnamelen = iq->qchase.qname_len;
1107         }
1108         if(iq->qchase.qtype == LDNS_RR_TYPE_DS || iq->refetch_glue ||
1109            (iq->qchase.qtype == LDNS_RR_TYPE_NS && qstate->prefetch_leeway)) {
1110                 /* remove first label from delname, root goes to hints,
1111                  * but only to fetch glue, not for qtype=DS. */
1112                 /* also when prefetching an NS record, fetch it again from
1113                  * its parent, just as if it expired, so that you do not
1114                  * get stuck on an older nameserver that gives old NSrecords */
1115                 if(dname_is_root(delname) && (iq->refetch_glue ||
1116                         (iq->qchase.qtype == LDNS_RR_TYPE_NS &&
1117                         qstate->prefetch_leeway)))
1118                         delname = NULL; /* go to root priming */
1119                 else    dname_remove_label(&delname, &delnamelen);
1120         }
1121         /* delname is the name to lookup a delegation for. If NULL rootprime */
1122         while(1) {
1123
1124                 /* Lookup the delegation in the cache. If null, then the
1125                  * cache needs to be primed for the qclass. */
1126                 if(delname)
1127                      iq->dp = dns_cache_find_delegation(qstate->env, delname,
1128                         delnamelen, iq->qchase.qtype, iq->qchase.qclass,
1129                         qstate->region, &iq->deleg_msg,
1130                         *qstate->env->now+qstate->prefetch_leeway);
1131                 else iq->dp = NULL;
1132
1133                 /* If the cache has returned nothing, then we have a
1134                  * root priming situation. */
1135                 if(iq->dp == NULL) {
1136                         /* if there is a stub, then no root prime needed */
1137                         int r = prime_stub(qstate, iq, id, delname,
1138                                 iq->qchase.qclass);
1139                         if(r == 2)
1140                                 break; /* got noprime-stub-zone, continue */
1141                         else if(r)
1142                                 return 0; /* stub prime request made */
1143                         if(forwards_lookup_root(qstate->env->fwds,
1144                                 iq->qchase.qclass)) {
1145                                 /* forward zone root, no root prime needed */
1146                                 /* fill in some dp - safety belt */
1147                                 iq->dp = hints_lookup_root(qstate->env->hints,
1148                                         iq->qchase.qclass);
1149                                 if(!iq->dp) {
1150                                         log_err("internal error: no hints dp");
1151                                         return error_response(qstate, id,
1152                                                 LDNS_RCODE_SERVFAIL);
1153                                 }
1154                                 iq->dp = delegpt_copy(iq->dp, qstate->region);
1155                                 if(!iq->dp) {
1156                                         log_err("out of memory in safety belt");
1157                                         return error_response(qstate, id,
1158                                                 LDNS_RCODE_SERVFAIL);
1159                                 }
1160                                 return next_state(iq, INIT_REQUEST_2_STATE);
1161                         }
1162                         /* Note that the result of this will set a new
1163                          * DelegationPoint based on the result of priming. */
1164                         if(!prime_root(qstate, iq, id, iq->qchase.qclass))
1165                                 return error_response(qstate, id,
1166                                         LDNS_RCODE_REFUSED);
1167
1168                         /* priming creates and sends a subordinate query, with
1169                          * this query as the parent. So further processing for
1170                          * this event will stop until reactivated by the
1171                          * results of priming. */
1172                         return 0;
1173                 }
1174                 if(!iq->ratelimit_ok && qstate->prefetch_leeway)
1175                         iq->ratelimit_ok = 1; /* allow prefetches, this keeps
1176                         otherwise valid data in the cache */
1177                 if(!iq->ratelimit_ok && infra_ratelimit_exceeded(
1178                         qstate->env->infra_cache, iq->dp->name,
1179                         iq->dp->namelen, *qstate->env->now)) {
1180                         /* and increment the rate, so that the rate for time
1181                          * now will also exceed the rate, keeping cache fresh */
1182                         (void)infra_ratelimit_inc(qstate->env->infra_cache,
1183                                 iq->dp->name, iq->dp->namelen,
1184                                 *qstate->env->now);
1185                         /* see if we are passed through with slip factor */
1186                         if(qstate->env->cfg->ratelimit_factor != 0 &&
1187                                 ub_random_max(qstate->env->rnd,
1188                                     qstate->env->cfg->ratelimit_factor) == 1) {
1189                                 iq->ratelimit_ok = 1;
1190                                 log_nametypeclass(VERB_ALGO, "ratelimit allowed through for "
1191                                         "delegation point", iq->dp->name,
1192                                         LDNS_RR_TYPE_NS, LDNS_RR_CLASS_IN);
1193                         } else {
1194                                 log_nametypeclass(VERB_ALGO, "ratelimit exceeded with "
1195                                         "delegation point", iq->dp->name,
1196                                         LDNS_RR_TYPE_NS, LDNS_RR_CLASS_IN);
1197                                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1198                         }
1199                 }
1200
1201                 /* see if this dp not useless.
1202                  * It is useless if:
1203                  *      o all NS items are required glue.
1204                  *        or the query is for NS item that is required glue.
1205                  *      o no addresses are provided.
1206                  *      o RD qflag is on.
1207                  * Instead, go up one level, and try to get even further
1208                  * If the root was useless, use safety belt information.
1209                  * Only check cache returns, because replies for servers
1210                  * could be useless but lead to loops (bumping into the
1211                  * same server reply) if useless-checked.
1212                  */
1213                 if(iter_dp_is_useless(&qstate->qinfo, qstate->query_flags,
1214                         iq->dp)) {
1215                         if(dname_is_root(iq->dp->name)) {
1216                                 /* use safety belt */
1217                                 verbose(VERB_QUERY, "Cache has root NS but "
1218                                 "no addresses. Fallback to the safety belt.");
1219                                 iq->dp = hints_lookup_root(qstate->env->hints,
1220                                         iq->qchase.qclass);
1221                                 /* note deleg_msg is from previous lookup,
1222                                  * but RD is on, so it is not used */
1223                                 if(!iq->dp) {
1224                                         log_err("internal error: no hints dp");
1225                                         return error_response(qstate, id,
1226                                                 LDNS_RCODE_REFUSED);
1227                                 }
1228                                 iq->dp = delegpt_copy(iq->dp, qstate->region);
1229                                 if(!iq->dp) {
1230                                         log_err("out of memory in safety belt");
1231                                         return error_response(qstate, id,
1232                                                 LDNS_RCODE_SERVFAIL);
1233                                 }
1234                                 break;
1235                         } else {
1236                                 verbose(VERB_ALGO,
1237                                         "cache delegation was useless:");
1238                                 delegpt_log(VERB_ALGO, iq->dp);
1239                                 /* go up */
1240                                 delname = iq->dp->name;
1241                                 delnamelen = iq->dp->namelen;
1242                                 dname_remove_label(&delname, &delnamelen);
1243                         }
1244                 } else break;
1245         }
1246
1247         verbose(VERB_ALGO, "cache delegation returns delegpt");
1248         delegpt_log(VERB_ALGO, iq->dp);
1249
1250         /* Otherwise, set the current delegation point and move on to the
1251          * next state. */
1252         return next_state(iq, INIT_REQUEST_2_STATE);
1253 }
1254
1255 /**
1256  * Process the second part of the initial request handling. This state
1257  * basically exists so that queries that generate root priming events have
1258  * the same init processing as ones that do not. Request events that reach
1259  * this state must have a valid currentDelegationPoint set.
1260  *
1261  * This part is primarly handling stub zone priming. Events that reach this
1262  * state must have a current delegation point.
1263  *
1264  * @param qstate: query state.
1265  * @param iq: iterator query state.
1266  * @param id: module id.
1267  * @return true if the event needs more request processing immediately,
1268  *         false if not.
1269  */
1270 static int
1271 processInitRequest2(struct module_qstate* qstate, struct iter_qstate* iq,
1272         int id)
1273 {
1274         uint8_t* delname;
1275         size_t delnamelen;
1276         log_query_info(VERB_QUERY, "resolving (init part 2): ",
1277                 &qstate->qinfo);
1278
1279         if(iq->refetch_glue) {
1280                 if(!iq->dp) {
1281                         log_err("internal or malloc fail: no dp for refetch");
1282                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1283                 }
1284                 delname = iq->dp->name;
1285                 delnamelen = iq->dp->namelen;
1286         } else {
1287                 delname = iq->qchase.qname;
1288                 delnamelen = iq->qchase.qname_len;
1289         }
1290         if(iq->qchase.qtype == LDNS_RR_TYPE_DS || iq->refetch_glue) {
1291                 if(!dname_is_root(delname))
1292                         dname_remove_label(&delname, &delnamelen);
1293                 iq->refetch_glue = 0; /* if CNAME causes restart, no refetch */
1294         }
1295         /* Check to see if we need to prime a stub zone. */
1296         if(prime_stub(qstate, iq, id, delname, iq->qchase.qclass)) {
1297                 /* A priming sub request was made */
1298                 return 0;
1299         }
1300
1301         /* most events just get forwarded to the next state. */
1302         return next_state(iq, INIT_REQUEST_3_STATE);
1303 }
1304
1305 /**
1306  * Process the third part of the initial request handling. This state exists
1307  * as a separate state so that queries that generate stub priming events
1308  * will get the tail end of the init process but not repeat the stub priming
1309  * check.
1310  *
1311  * @param qstate: query state.
1312  * @param iq: iterator query state.
1313  * @param id: module id.
1314  * @return true, advancing the event to the QUERYTARGETS_STATE.
1315  */
1316 static int
1317 processInitRequest3(struct module_qstate* qstate, struct iter_qstate* iq,
1318         int id)
1319 {
1320         log_query_info(VERB_QUERY, "resolving (init part 3): ",
1321                 &qstate->qinfo);
1322         /* if the cache reply dp equals a validation anchor or msg has DS,
1323          * then DNSSEC RRSIGs are expected in the reply */
1324         iq->dnssec_expected = iter_indicates_dnssec(qstate->env, iq->dp,
1325                 iq->deleg_msg, iq->qchase.qclass);
1326
1327         /* If the RD flag wasn't set, then we just finish with the
1328          * cached referral as the response. */
1329         if(!(qstate->query_flags & BIT_RD)) {
1330                 iq->response = iq->deleg_msg;
1331                 if(verbosity >= VERB_ALGO && iq->response)
1332                         log_dns_msg("no RD requested, using delegation msg",
1333                                 &iq->response->qinfo, iq->response->rep);
1334                 if(qstate->reply_origin)
1335                         sock_list_insert(&qstate->reply_origin, NULL, 0, qstate->region);
1336                 return final_state(iq);
1337         }
1338         /* After this point, unset the RD flag -- this query is going to
1339          * be sent to an auth. server. */
1340         iq->chase_flags &= ~BIT_RD;
1341
1342         /* if dnssec expected, fetch key for the trust-anchor or cached-DS */
1343         if(iq->dnssec_expected && qstate->env->cfg->prefetch_key &&
1344                 !(qstate->query_flags&BIT_CD)) {
1345                 generate_dnskey_prefetch(qstate, iq, id);
1346                 fptr_ok(fptr_whitelist_modenv_detach_subs(
1347                         qstate->env->detach_subs));
1348                 (*qstate->env->detach_subs)(qstate);
1349         }
1350
1351         /* Jump to the next state. */
1352         return next_state(iq, QUERYTARGETS_STATE);
1353 }
1354
1355 /**
1356  * Given a basic query, generate a parent-side "target" query.
1357  * These are subordinate queries for missing delegation point target addresses,
1358  * for which only the parent of the delegation provides correct IP addresses.
1359  *
1360  * @param qstate: query state.
1361  * @param iq: iterator query state.
1362  * @param id: module id.
1363  * @param name: target qname.
1364  * @param namelen: target qname length.
1365  * @param qtype: target qtype (either A or AAAA).
1366  * @param qclass: target qclass.
1367  * @return true on success, false on failure.
1368  */
1369 static int
1370 generate_parentside_target_query(struct module_qstate* qstate,
1371         struct iter_qstate* iq, int id, uint8_t* name, size_t namelen,
1372         uint16_t qtype, uint16_t qclass)
1373 {
1374         struct module_qstate* subq;
1375         if(!generate_sub_request(name, namelen, qtype, qclass, qstate,
1376                 id, iq, INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0))
1377                 return 0;
1378         if(subq) {
1379                 struct iter_qstate* subiq =
1380                         (struct iter_qstate*)subq->minfo[id];
1381                 /* blacklist the cache - we want to fetch parent stuff */
1382                 sock_list_insert(&subq->blacklist, NULL, 0, subq->region);
1383                 subiq->query_for_pside_glue = 1;
1384                 if(dname_subdomain_c(name, iq->dp->name)) {
1385                         subiq->dp = delegpt_copy(iq->dp, subq->region);
1386                         subiq->dnssec_expected = iter_indicates_dnssec(
1387                                 qstate->env, subiq->dp, NULL,
1388                                 subq->qinfo.qclass);
1389                         subiq->refetch_glue = 1;
1390                 } else {
1391                         subiq->dp = dns_cache_find_delegation(qstate->env,
1392                                 name, namelen, qtype, qclass, subq->region,
1393                                 &subiq->deleg_msg,
1394                                 *qstate->env->now+subq->prefetch_leeway);
1395                         /* if no dp, then it's from root, refetch unneeded */
1396                         if(subiq->dp) {
1397                                 subiq->dnssec_expected = iter_indicates_dnssec(
1398                                         qstate->env, subiq->dp, NULL,
1399                                         subq->qinfo.qclass);
1400                                 subiq->refetch_glue = 1;
1401                         }
1402                 }
1403         }
1404         log_nametypeclass(VERB_QUERY, "new pside target", name, qtype, qclass);
1405         return 1;
1406 }
1407
1408 /**
1409  * Given a basic query, generate a "target" query. These are subordinate
1410  * queries for missing delegation point target addresses.
1411  *
1412  * @param qstate: query state.
1413  * @param iq: iterator query state.
1414  * @param id: module id.
1415  * @param name: target qname.
1416  * @param namelen: target qname length.
1417  * @param qtype: target qtype (either A or AAAA).
1418  * @param qclass: target qclass.
1419  * @return true on success, false on failure.
1420  */
1421 static int
1422 generate_target_query(struct module_qstate* qstate, struct iter_qstate* iq,
1423         int id, uint8_t* name, size_t namelen, uint16_t qtype, uint16_t qclass)
1424 {
1425         struct module_qstate* subq;
1426         if(!generate_sub_request(name, namelen, qtype, qclass, qstate,
1427                 id, iq, INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0))
1428                 return 0;
1429         log_nametypeclass(VERB_QUERY, "new target", name, qtype, qclass);
1430         return 1;
1431 }
1432
1433 /**
1434  * Given an event at a certain state, generate zero or more target queries
1435  * for it's current delegation point.
1436  *
1437  * @param qstate: query state.
1438  * @param iq: iterator query state.
1439  * @param ie: iterator shared global environment.
1440  * @param id: module id.
1441  * @param maxtargets: The maximum number of targets to query for.
1442  *      if it is negative, there is no maximum number of targets.
1443  * @param num: returns the number of queries generated and processed,
1444  *      which may be zero if there were no missing targets.
1445  * @return false on error.
1446  */
1447 static int
1448 query_for_targets(struct module_qstate* qstate, struct iter_qstate* iq,
1449         struct iter_env* ie, int id, int maxtargets, int* num)
1450 {
1451         int query_count = 0;
1452         struct delegpt_ns* ns;
1453         int missing;
1454         int toget = 0;
1455
1456         if(iq->depth == ie->max_dependency_depth)
1457                 return 0;
1458         if(iq->depth > 0 && iq->target_count &&
1459                 iq->target_count[1] > MAX_TARGET_COUNT) {
1460                 char s[LDNS_MAX_DOMAINLEN+1];
1461                 dname_str(qstate->qinfo.qname, s);
1462                 verbose(VERB_QUERY, "request %s has exceeded the maximum "
1463                         "number of glue fetches %d", s, iq->target_count[1]);
1464                 return 0;
1465         }
1466
1467         iter_mark_cycle_targets(qstate, iq->dp);
1468         missing = (int)delegpt_count_missing_targets(iq->dp);
1469         log_assert(maxtargets != 0); /* that would not be useful */
1470
1471         /* Generate target requests. Basically, any missing targets
1472          * are queried for here, regardless if it is necessary to do
1473          * so to continue processing. */
1474         if(maxtargets < 0 || maxtargets > missing)
1475                 toget = missing;
1476         else    toget = maxtargets;
1477         if(toget == 0) {
1478                 *num = 0;
1479                 return 1;
1480         }
1481         /* select 'toget' items from the total of 'missing' items */
1482         log_assert(toget <= missing);
1483
1484         /* loop over missing targets */
1485         for(ns = iq->dp->nslist; ns; ns = ns->next) {
1486                 if(ns->resolved)
1487                         continue;
1488
1489                 /* randomly select this item with probability toget/missing */
1490                 if(!iter_ns_probability(qstate->env->rnd, toget, missing)) {
1491                         /* do not select this one, next; select toget number
1492                          * of items from a list one less in size */
1493                         missing --;
1494                         continue;
1495                 }
1496
1497                 if(ie->supports_ipv6 && !ns->got6) {
1498                         /* Send the AAAA request. */
1499                         if(!generate_target_query(qstate, iq, id,
1500                                 ns->name, ns->namelen,
1501                                 LDNS_RR_TYPE_AAAA, iq->qchase.qclass)) {
1502                                 *num = query_count;
1503                                 if(query_count > 0)
1504                                         qstate->ext_state[id] = module_wait_subquery;
1505                                 return 0;
1506                         }
1507                         query_count++;
1508                 }
1509                 /* Send the A request. */
1510                 if(ie->supports_ipv4 && !ns->got4) {
1511                         if(!generate_target_query(qstate, iq, id,
1512                                 ns->name, ns->namelen,
1513                                 LDNS_RR_TYPE_A, iq->qchase.qclass)) {
1514                                 *num = query_count;
1515                                 if(query_count > 0)
1516                                         qstate->ext_state[id] = module_wait_subquery;
1517                                 return 0;
1518                         }
1519                         query_count++;
1520                 }
1521
1522                 /* mark this target as in progress. */
1523                 ns->resolved = 1;
1524                 missing--;
1525                 toget--;
1526                 if(toget == 0)
1527                         break;
1528         }
1529         *num = query_count;
1530         if(query_count > 0)
1531                 qstate->ext_state[id] = module_wait_subquery;
1532
1533         return 1;
1534 }
1535
1536 /** see if last resort is possible - does config allow queries to parent */
1537 static int
1538 can_have_last_resort(struct module_env* env, struct delegpt* dp,
1539         struct iter_qstate* iq)
1540 {
1541         struct delegpt* fwddp;
1542         struct iter_hints_stub* stub;
1543         /* do not process a last resort (the parent side) if a stub
1544          * or forward is configured, because we do not want to go 'above'
1545          * the configured servers */
1546         if(!dname_is_root(dp->name) && (stub = (struct iter_hints_stub*)
1547                 name_tree_find(&env->hints->tree, dp->name, dp->namelen,
1548                 dp->namelabs, iq->qchase.qclass)) &&
1549                 /* has_parent side is turned off for stub_first, where we
1550                  * are allowed to go to the parent */
1551                 stub->dp->has_parent_side_NS) {
1552                 verbose(VERB_QUERY, "configured stub servers failed -- returning SERVFAIL");
1553                 return 0;
1554         }
1555         if((fwddp = forwards_find(env->fwds, dp->name, iq->qchase.qclass)) &&
1556                 /* has_parent_side is turned off for forward_first, where
1557                  * we are allowed to go to the parent */
1558                 fwddp->has_parent_side_NS) {
1559                 verbose(VERB_QUERY, "configured forward servers failed -- returning SERVFAIL");
1560                 return 0;
1561         }
1562         return 1;
1563 }
1564
1565 /**
1566  * Called by processQueryTargets when it would like extra targets to query
1567  * but it seems to be out of options.  At last resort some less appealing
1568  * options are explored.  If there are no more options, the result is SERVFAIL
1569  *
1570  * @param qstate: query state.
1571  * @param iq: iterator query state.
1572  * @param ie: iterator shared global environment.
1573  * @param id: module id.
1574  * @return true if the event requires more request processing immediately,
1575  *         false if not.
1576  */
1577 static int
1578 processLastResort(struct module_qstate* qstate, struct iter_qstate* iq,
1579         struct iter_env* ie, int id)
1580 {
1581         struct delegpt_ns* ns;
1582         int query_count = 0;
1583         verbose(VERB_ALGO, "No more query targets, attempting last resort");
1584         log_assert(iq->dp);
1585
1586         if(!can_have_last_resort(qstate->env, iq->dp, iq)) {
1587                 /* fail -- no more targets, no more hope of targets, no hope
1588                  * of a response. */
1589                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1590         }
1591         if(!iq->dp->has_parent_side_NS && dname_is_root(iq->dp->name)) {
1592                 struct delegpt* p = hints_lookup_root(qstate->env->hints,
1593                         iq->qchase.qclass);
1594                 if(p) {
1595                         struct delegpt_ns* ns;
1596                         struct delegpt_addr* a;
1597                         iq->chase_flags &= ~BIT_RD; /* go to authorities */
1598                         for(ns = p->nslist; ns; ns=ns->next) {
1599                                 (void)delegpt_add_ns(iq->dp, qstate->region,
1600                                         ns->name, ns->lame);
1601                         }
1602                         for(a = p->target_list; a; a=a->next_target) {
1603                                 (void)delegpt_add_addr(iq->dp, qstate->region,
1604                                         &a->addr, a->addrlen, a->bogus,
1605                                         a->lame);
1606                         }
1607                 }
1608                 iq->dp->has_parent_side_NS = 1;
1609         } else if(!iq->dp->has_parent_side_NS) {
1610                 if(!iter_lookup_parent_NS_from_cache(qstate->env, iq->dp,
1611                         qstate->region, &qstate->qinfo)
1612                         || !iq->dp->has_parent_side_NS) {
1613                         /* if: malloc failure in lookup go up to try */
1614                         /* if: no parent NS in cache - go up one level */
1615                         verbose(VERB_ALGO, "try to grab parent NS");
1616                         iq->store_parent_NS = iq->dp;
1617                         iq->chase_flags &= ~BIT_RD; /* go to authorities */
1618                         iq->deleg_msg = NULL;
1619                         iq->refetch_glue = 1;
1620                         iq->query_restart_count++;
1621                         iq->sent_count = 0;
1622                         if(qstate->env->cfg->qname_minimisation)
1623                                 iq->minimisation_state = INIT_MINIMISE_STATE;
1624                         return next_state(iq, INIT_REQUEST_STATE);
1625                 }
1626         }
1627         /* see if that makes new names available */
1628         if(!cache_fill_missing(qstate->env, iq->qchase.qclass,
1629                 qstate->region, iq->dp))
1630                 log_err("out of memory in cache_fill_missing");
1631         if(iq->dp->usable_list) {
1632                 verbose(VERB_ALGO, "try parent-side-name, w. glue from cache");
1633                 return next_state(iq, QUERYTARGETS_STATE);
1634         }
1635         /* try to fill out parent glue from cache */
1636         if(iter_lookup_parent_glue_from_cache(qstate->env, iq->dp,
1637                 qstate->region, &qstate->qinfo)) {
1638                 /* got parent stuff from cache, see if we can continue */
1639                 verbose(VERB_ALGO, "try parent-side glue from cache");
1640                 return next_state(iq, QUERYTARGETS_STATE);
1641         }
1642         /* query for an extra name added by the parent-NS record */
1643         if(delegpt_count_missing_targets(iq->dp) > 0) {
1644                 int qs = 0;
1645                 verbose(VERB_ALGO, "try parent-side target name");
1646                 if(!query_for_targets(qstate, iq, ie, id, 1, &qs)) {
1647                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1648                 }
1649                 iq->num_target_queries += qs;
1650                 target_count_increase(iq, qs);
1651                 if(qs != 0) {
1652                         qstate->ext_state[id] = module_wait_subquery;
1653                         return 0; /* and wait for them */
1654                 }
1655         }
1656         if(iq->depth == ie->max_dependency_depth) {
1657                 verbose(VERB_QUERY, "maxdepth and need more nameservers, fail");
1658                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1659         }
1660         if(iq->depth > 0 && iq->target_count &&
1661                 iq->target_count[1] > MAX_TARGET_COUNT) {
1662                 char s[LDNS_MAX_DOMAINLEN+1];
1663                 dname_str(qstate->qinfo.qname, s);
1664                 verbose(VERB_QUERY, "request %s has exceeded the maximum "
1665                         "number of glue fetches %d", s, iq->target_count[1]);
1666                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1667         }
1668         /* mark cycle targets for parent-side lookups */
1669         iter_mark_pside_cycle_targets(qstate, iq->dp);
1670         /* see if we can issue queries to get nameserver addresses */
1671         /* this lookup is not randomized, but sequential. */
1672         for(ns = iq->dp->nslist; ns; ns = ns->next) {
1673                 /* query for parent-side A and AAAA for nameservers */
1674                 if(ie->supports_ipv6 && !ns->done_pside6) {
1675                         /* Send the AAAA request. */
1676                         if(!generate_parentside_target_query(qstate, iq, id,
1677                                 ns->name, ns->namelen,
1678                                 LDNS_RR_TYPE_AAAA, iq->qchase.qclass))
1679                                 return error_response(qstate, id,
1680                                         LDNS_RCODE_SERVFAIL);
1681                         ns->done_pside6 = 1;
1682                         query_count++;
1683                 }
1684                 if(ie->supports_ipv4 && !ns->done_pside4) {
1685                         /* Send the A request. */
1686                         if(!generate_parentside_target_query(qstate, iq, id,
1687                                 ns->name, ns->namelen,
1688                                 LDNS_RR_TYPE_A, iq->qchase.qclass))
1689                                 return error_response(qstate, id,
1690                                         LDNS_RCODE_SERVFAIL);
1691                         ns->done_pside4 = 1;
1692                         query_count++;
1693                 }
1694                 if(query_count != 0) { /* suspend to await results */
1695                         verbose(VERB_ALGO, "try parent-side glue lookup");
1696                         iq->num_target_queries += query_count;
1697                         target_count_increase(iq, query_count);
1698                         qstate->ext_state[id] = module_wait_subquery;
1699                         return 0;
1700                 }
1701         }
1702
1703         /* if this was a parent-side glue query itself, then store that
1704          * failure in cache. */
1705         if(!qstate->no_cache_store && iq->query_for_pside_glue
1706                 && !iq->pside_glue)
1707                         iter_store_parentside_neg(qstate->env, &qstate->qinfo,
1708                                 iq->deleg_msg?iq->deleg_msg->rep:
1709                                 (iq->response?iq->response->rep:NULL));
1710
1711         verbose(VERB_QUERY, "out of query targets -- returning SERVFAIL");
1712         /* fail -- no more targets, no more hope of targets, no hope
1713          * of a response. */
1714         return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1715 }
1716
1717 /**
1718  * Try to find the NS record set that will resolve a qtype DS query. Due
1719  * to grandparent/grandchild reasons we did not get a proper lookup right
1720  * away.  We need to create type NS queries until we get the right parent
1721  * for this lookup.  We remove labels from the query to find the right point.
1722  * If we end up at the old dp name, then there is no solution.
1723  *
1724  * @param qstate: query state.
1725  * @param iq: iterator query state.
1726  * @param id: module id.
1727  * @return true if the event requires more immediate processing, false if
1728  *         not. This is generally only true when forwarding the request to
1729  *         the final state (i.e., on answer).
1730  */
1731 static int
1732 processDSNSFind(struct module_qstate* qstate, struct iter_qstate* iq, int id)
1733 {
1734         struct module_qstate* subq = NULL;
1735         verbose(VERB_ALGO, "processDSNSFind");
1736
1737         if(!iq->dsns_point) {
1738                 /* initialize */
1739                 iq->dsns_point = iq->qchase.qname;
1740                 iq->dsns_point_len = iq->qchase.qname_len;
1741         }
1742         /* robustcheck for internal error: we are not underneath the dp */
1743         if(!dname_subdomain_c(iq->dsns_point, iq->dp->name)) {
1744                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1745         }
1746
1747         /* go up one (more) step, until we hit the dp, if so, end */
1748         dname_remove_label(&iq->dsns_point, &iq->dsns_point_len);
1749         if(query_dname_compare(iq->dsns_point, iq->dp->name) == 0) {
1750                 /* there was no inbetween nameserver, use the old delegation
1751                  * point again.  And this time, because dsns_point is nonNULL
1752                  * we are going to accept the (bad) result */
1753                 iq->state = QUERYTARGETS_STATE;
1754                 return 1;
1755         }
1756         iq->state = DSNS_FIND_STATE;
1757
1758         /* spawn NS lookup (validation not needed, this is for DS lookup) */
1759         log_nametypeclass(VERB_ALGO, "fetch nameservers",
1760                 iq->dsns_point, LDNS_RR_TYPE_NS, iq->qchase.qclass);
1761         if(!generate_sub_request(iq->dsns_point, iq->dsns_point_len,
1762                 LDNS_RR_TYPE_NS, iq->qchase.qclass, qstate, id, iq,
1763                 INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0)) {
1764                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1765         }
1766
1767         return 0;
1768 }
1769
1770 /**
1771  * This is the request event state where the request will be sent to one of
1772  * its current query targets. This state also handles issuing target lookup
1773  * queries for missing target IP addresses. Queries typically iterate on
1774  * this state, both when they are just trying different targets for a given
1775  * delegation point, and when they change delegation points. This state
1776  * roughly corresponds to RFC 1034 algorithm steps 3 and 4.
1777  *
1778  * @param qstate: query state.
1779  * @param iq: iterator query state.
1780  * @param ie: iterator shared global environment.
1781  * @param id: module id.
1782  * @return true if the event requires more request processing immediately,
1783  *         false if not. This state only returns true when it is generating
1784  *         a SERVFAIL response because the query has hit a dead end.
1785  */
1786 static int
1787 processQueryTargets(struct module_qstate* qstate, struct iter_qstate* iq,
1788         struct iter_env* ie, int id)
1789 {
1790         int tf_policy;
1791         struct delegpt_addr* target;
1792         struct outbound_entry* outq;
1793
1794         /* NOTE: a request will encounter this state for each target it
1795          * needs to send a query to. That is, at least one per referral,
1796          * more if some targets timeout or return throwaway answers. */
1797
1798         log_query_info(VERB_QUERY, "processQueryTargets:", &qstate->qinfo);
1799         verbose(VERB_ALGO, "processQueryTargets: targetqueries %d, "
1800                 "currentqueries %d sentcount %d", iq->num_target_queries,
1801                 iq->num_current_queries, iq->sent_count);
1802
1803         /* Make sure that we haven't run away */
1804         /* FIXME: is this check even necessary? */
1805         if(iq->referral_count > MAX_REFERRAL_COUNT) {
1806                 verbose(VERB_QUERY, "request has exceeded the maximum "
1807                         "number of referrrals with %d", iq->referral_count);
1808                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1809         }
1810         if(iq->sent_count > MAX_SENT_COUNT) {
1811                 verbose(VERB_QUERY, "request has exceeded the maximum "
1812                         "number of sends with %d", iq->sent_count);
1813                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1814         }
1815
1816         /* Make sure we have a delegation point, otherwise priming failed
1817          * or another failure occurred */
1818         if(!iq->dp) {
1819                 verbose(VERB_QUERY, "Failed to get a delegation, giving up");
1820                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1821         }
1822         if(!ie->supports_ipv6)
1823                 delegpt_no_ipv6(iq->dp);
1824         if(!ie->supports_ipv4)
1825                 delegpt_no_ipv4(iq->dp);
1826         delegpt_log(VERB_ALGO, iq->dp);
1827
1828         if(iq->num_current_queries>0) {
1829                 /* already busy answering a query, this restart is because
1830                  * more delegpt addrs became available, wait for existing
1831                  * query. */
1832                 verbose(VERB_ALGO, "woke up, but wait for outstanding query");
1833                 qstate->ext_state[id] = module_wait_reply;
1834                 return 0;
1835         }
1836
1837         tf_policy = 0;
1838         /* < not <=, because although the array is large enough for <=, the
1839          * generated query will immediately be discarded due to depth and
1840          * that servfail is cached, which is not good as opportunism goes. */
1841         if(iq->depth < ie->max_dependency_depth
1842                 && iq->sent_count < TARGET_FETCH_STOP) {
1843                 tf_policy = ie->target_fetch_policy[iq->depth];
1844         }
1845
1846         /* if in 0x20 fallback get as many targets as possible */
1847         if(iq->caps_fallback) {
1848                 int extra = 0;
1849                 size_t naddr, nres, navail;
1850                 if(!query_for_targets(qstate, iq, ie, id, -1, &extra)) {
1851                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1852                 }
1853                 iq->num_target_queries += extra;
1854                 target_count_increase(iq, extra);
1855                 if(iq->num_target_queries > 0) {
1856                         /* wait to get all targets, we want to try em */
1857                         verbose(VERB_ALGO, "wait for all targets for fallback");
1858                         qstate->ext_state[id] = module_wait_reply;
1859                         return 0;
1860                 }
1861                 /* did we do enough fallback queries already? */
1862                 delegpt_count_addr(iq->dp, &naddr, &nres, &navail);
1863                 /* the current caps_server is the number of fallbacks sent.
1864                  * the original query is one that matched too, so we have
1865                  * caps_server+1 number of matching queries now */
1866                 if(iq->caps_server+1 >= naddr*3 ||
1867                         iq->caps_server*2+2 >= MAX_SENT_COUNT) {
1868                         /* *2 on sentcount check because ipv6 may fail */
1869                         /* we're done, process the response */
1870                         verbose(VERB_ALGO, "0x20 fallback had %d responses "
1871                                 "match for %d wanted, done.",
1872                                 (int)iq->caps_server+1, (int)naddr*3);
1873                         iq->response = iq->caps_response;
1874                         iq->caps_fallback = 0;
1875                         iter_dec_attempts(iq->dp, 3); /* space for fallback */
1876                         iq->num_current_queries++; /* RespState decrements it*/
1877                         iq->referral_count++; /* make sure we don't loop */
1878                         iq->sent_count = 0;
1879                         iq->state = QUERY_RESP_STATE;
1880                         return 1;
1881                 }
1882                 verbose(VERB_ALGO, "0x20 fallback number %d",
1883                         (int)iq->caps_server);
1884
1885         /* if there is a policy to fetch missing targets
1886          * opportunistically, do it. we rely on the fact that once a
1887          * query (or queries) for a missing name have been issued,
1888          * they will not show up again. */
1889         } else if(tf_policy != 0) {
1890                 int extra = 0;
1891                 verbose(VERB_ALGO, "attempt to get extra %d targets",
1892                         tf_policy);
1893                 (void)query_for_targets(qstate, iq, ie, id, tf_policy, &extra);
1894                 /* errors ignored, these targets are not strictly necessary for
1895                  * this result, we do not have to reply with SERVFAIL */
1896                 iq->num_target_queries += extra;
1897                 target_count_increase(iq, extra);
1898         }
1899
1900         /* Add the current set of unused targets to our queue. */
1901         delegpt_add_unused_targets(iq->dp);
1902
1903         /* Select the next usable target, filtering out unsuitable targets. */
1904         target = iter_server_selection(ie, qstate->env, iq->dp,
1905                 iq->dp->name, iq->dp->namelen, iq->qchase.qtype,
1906                 &iq->dnssec_lame_query, &iq->chase_to_rd,
1907                 iq->num_target_queries, qstate->blacklist);
1908
1909         /* If no usable target was selected... */
1910         if(!target) {
1911                 /* Here we distinguish between three states: generate a new
1912                  * target query, just wait, or quit (with a SERVFAIL).
1913                  * We have the following information: number of active
1914                  * target queries, number of active current queries,
1915                  * the presence of missing targets at this delegation
1916                  * point, and the given query target policy. */
1917
1918                 /* Check for the wait condition. If this is true, then
1919                  * an action must be taken. */
1920                 if(iq->num_target_queries==0 && iq->num_current_queries==0) {
1921                         /* If there is nothing to wait for, then we need
1922                          * to distinguish between generating (a) new target
1923                          * query, or failing. */
1924                         if(delegpt_count_missing_targets(iq->dp) > 0) {
1925                                 int qs = 0;
1926                                 verbose(VERB_ALGO, "querying for next "
1927                                         "missing target");
1928                                 if(!query_for_targets(qstate, iq, ie, id,
1929                                         1, &qs)) {
1930                                         return error_response(qstate, id,
1931                                                 LDNS_RCODE_SERVFAIL);
1932                                 }
1933                                 if(qs == 0 &&
1934                                    delegpt_count_missing_targets(iq->dp) == 0){
1935                                         /* it looked like there were missing
1936                                          * targets, but they did not turn up.
1937                                          * Try the bad choices again (if any),
1938                                          * when we get back here missing==0,
1939                                          * so this is not a loop. */
1940                                         return 1;
1941                                 }
1942                                 iq->num_target_queries += qs;
1943                                 target_count_increase(iq, qs);
1944                         }
1945                         /* Since a target query might have been made, we
1946                          * need to check again. */
1947                         if(iq->num_target_queries == 0) {
1948                                 /* if in capsforid fallback, instead of last
1949                                  * resort, we agree with the current reply
1950                                  * we have (if any) (our count of addrs bad)*/
1951                                 if(iq->caps_fallback && iq->caps_reply) {
1952                                         /* we're done, process the response */
1953                                         verbose(VERB_ALGO, "0x20 fallback had %d responses, "
1954                                                 "but no more servers except "
1955                                                 "last resort, done.",
1956                                                 (int)iq->caps_server+1);
1957                                         iq->response = iq->caps_response;
1958                                         iq->caps_fallback = 0;
1959                                         iter_dec_attempts(iq->dp, 3); /* space for fallback */
1960                                         iq->num_current_queries++; /* RespState decrements it*/
1961                                         iq->referral_count++; /* make sure we don't loop */
1962                                         iq->sent_count = 0;
1963                                         iq->state = QUERY_RESP_STATE;
1964                                         return 1;
1965                                 }
1966                                 return processLastResort(qstate, iq, ie, id);
1967                         }
1968                 }
1969
1970                 /* otherwise, we have no current targets, so submerge
1971                  * until one of the target or direct queries return. */
1972                 if(iq->num_target_queries>0 && iq->num_current_queries>0) {
1973                         verbose(VERB_ALGO, "no current targets -- waiting "
1974                                 "for %d targets to resolve or %d outstanding"
1975                                 " queries to respond", iq->num_target_queries,
1976                                 iq->num_current_queries);
1977                         qstate->ext_state[id] = module_wait_reply;
1978                 } else if(iq->num_target_queries>0) {
1979                         verbose(VERB_ALGO, "no current targets -- waiting "
1980                                 "for %d targets to resolve.",
1981                                 iq->num_target_queries);
1982                         qstate->ext_state[id] = module_wait_subquery;
1983                 } else {
1984                         verbose(VERB_ALGO, "no current targets -- waiting "
1985                                 "for %d outstanding queries to respond.",
1986                                 iq->num_current_queries);
1987                         qstate->ext_state[id] = module_wait_reply;
1988                 }
1989                 return 0;
1990         }
1991
1992         /* if not forwarding, check ratelimits per delegationpoint name */
1993         if(!(iq->chase_flags & BIT_RD) && !iq->ratelimit_ok) {
1994                 if(!infra_ratelimit_inc(qstate->env->infra_cache, iq->dp->name,
1995                         iq->dp->namelen, *qstate->env->now)) {
1996                         verbose(VERB_ALGO, "query exceeded ratelimits");
1997                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1998                 }
1999         }
2000
2001         if(iq->minimisation_state == INIT_MINIMISE_STATE) {
2002                 /* (Re)set qinfo_out to (new) delegation point, except when
2003                  * qinfo_out is already a subdomain of dp. This happens when
2004                  * increasing by more than one label at once (QNAMEs with more
2005                  * than MAX_MINIMISE_COUNT labels). */
2006                 if(!(iq->qinfo_out.qname_len
2007                         && dname_subdomain_c(iq->qchase.qname,
2008                                 iq->qinfo_out.qname)
2009                         && dname_subdomain_c(iq->qinfo_out.qname,
2010                                 iq->dp->name))) {
2011                         iq->qinfo_out.qname = iq->dp->name;
2012                         iq->qinfo_out.qname_len = iq->dp->namelen;
2013                         iq->qinfo_out.qtype = LDNS_RR_TYPE_A;
2014                         iq->qinfo_out.qclass = iq->qchase.qclass;
2015                         iq->qinfo_out.local_alias = NULL;
2016                         iq->minimise_count = 0;
2017                 }
2018
2019                 iq->minimisation_state = MINIMISE_STATE;
2020         }
2021         if(iq->minimisation_state == MINIMISE_STATE) {
2022                 int qchaselabs = dname_count_labels(iq->qchase.qname);
2023                 int labdiff = qchaselabs -
2024                         dname_count_labels(iq->qinfo_out.qname);
2025
2026                 iq->qinfo_out.qname = iq->qchase.qname;
2027                 iq->qinfo_out.qname_len = iq->qchase.qname_len;
2028                 iq->minimise_count++;
2029                 iq->minimise_timeout_count = 0;
2030
2031                 iter_dec_attempts(iq->dp, 1);
2032
2033                 /* Limit number of iterations for QNAMEs with more
2034                  * than MAX_MINIMISE_COUNT labels. Send first MINIMISE_ONE_LAB
2035                  * labels of QNAME always individually.
2036                  */
2037                 if(qchaselabs > MAX_MINIMISE_COUNT && labdiff > 1 &&
2038                         iq->minimise_count > MINIMISE_ONE_LAB) {
2039                         if(iq->minimise_count < MAX_MINIMISE_COUNT) {
2040                                 int multilabs = qchaselabs - 1 -
2041                                         MINIMISE_ONE_LAB;
2042                                 int extralabs = multilabs /
2043                                         MINIMISE_MULTIPLE_LABS;
2044
2045                                 if (MAX_MINIMISE_COUNT - iq->minimise_count >=
2046                                         multilabs % MINIMISE_MULTIPLE_LABS)
2047                                         /* Default behaviour is to add 1 label
2048                                          * every iteration. Therefore, decrement
2049                                          * the extralabs by 1 */
2050                                         extralabs--;
2051                                 if (extralabs < labdiff)
2052                                         labdiff -= extralabs;
2053                                 else
2054                                         labdiff = 1;
2055                         }
2056                         /* Last minimised iteration, send all labels with
2057                          * QTYPE=NS */
2058                         else
2059                                 labdiff = 1;
2060                 }
2061
2062                 if(labdiff > 1) {
2063                         verbose(VERB_QUERY, "removing %d labels", labdiff-1);
2064                         dname_remove_labels(&iq->qinfo_out.qname,
2065                                 &iq->qinfo_out.qname_len,
2066                                 labdiff-1);
2067                 }
2068                 if(labdiff < 1 || (labdiff < 2
2069                         && (iq->qchase.qtype == LDNS_RR_TYPE_DS
2070                         || iq->qchase.qtype == LDNS_RR_TYPE_A)))
2071                         /* Stop minimising this query, resolve "as usual" */
2072                         iq->minimisation_state = DONOT_MINIMISE_STATE;
2073                 else if(!qstate->no_cache_lookup) {
2074                         struct dns_msg* msg = dns_cache_lookup(qstate->env,
2075                                 iq->qinfo_out.qname, iq->qinfo_out.qname_len,
2076                                 iq->qinfo_out.qtype, iq->qinfo_out.qclass,
2077                                 qstate->query_flags, qstate->region,
2078                                 qstate->env->scratch);
2079                         if(msg && msg->rep->an_numrrsets == 0
2080                                 && FLAGS_GET_RCODE(msg->rep->flags) ==
2081                                 LDNS_RCODE_NOERROR)
2082                                 /* no need to send query if it is already
2083                                  * cached as NOERROR/NODATA */
2084                                 return 1;
2085                 }
2086         }
2087         if(iq->minimisation_state == SKIP_MINIMISE_STATE) {
2088                 iq->minimise_timeout_count++;
2089                 if(iq->minimise_timeout_count < MAX_MINIMISE_TIMEOUT_COUNT)
2090                         /* Do not increment qname, continue incrementing next
2091                          * iteration */
2092                         iq->minimisation_state = MINIMISE_STATE;
2093                 else if(!qstate->env->cfg->qname_minimisation_strict)
2094                         /* Too many time-outs detected for this QNAME and QTYPE.
2095                          * We give up, disable QNAME minimisation. */
2096                         iq->minimisation_state = DONOT_MINIMISE_STATE;
2097         }
2098         if(iq->minimisation_state == DONOT_MINIMISE_STATE)
2099                 iq->qinfo_out = iq->qchase;
2100
2101         /* We have a valid target. */
2102         if(verbosity >= VERB_QUERY) {
2103                 log_query_info(VERB_QUERY, "sending query:", &iq->qinfo_out);
2104                 log_name_addr(VERB_QUERY, "sending to target:", iq->dp->name,
2105                         &target->addr, target->addrlen);
2106                 verbose(VERB_ALGO, "dnssec status: %s%s",
2107                         iq->dnssec_expected?"expected": "not expected",
2108                         iq->dnssec_lame_query?" but lame_query anyway": "");
2109         }
2110         fptr_ok(fptr_whitelist_modenv_send_query(qstate->env->send_query));
2111         outq = (*qstate->env->send_query)(&iq->qinfo_out,
2112                 iq->chase_flags | (iq->chase_to_rd?BIT_RD:0),
2113                 /* unset CD if to forwarder(RD set) and not dnssec retry
2114                  * (blacklist nonempty) and no trust-anchors are configured
2115                  * above the qname or on the first attempt when dnssec is on */
2116                 EDNS_DO| ((iq->chase_to_rd||(iq->chase_flags&BIT_RD)!=0)&&
2117                 !qstate->blacklist&&(!iter_indicates_dnssec_fwd(qstate->env,
2118                 &iq->qinfo_out)||target->attempts==1)?0:BIT_CD),
2119                 iq->dnssec_expected, iq->caps_fallback || is_caps_whitelisted(
2120                 ie, iq), &target->addr, target->addrlen,
2121                 iq->dp->name, iq->dp->namelen,
2122                 (iq->dp->ssl_upstream || qstate->env->cfg->ssl_upstream), qstate);
2123         if(!outq) {
2124                 log_addr(VERB_DETAIL, "error sending query to auth server",
2125                         &target->addr, target->addrlen);
2126                 if(!(iq->chase_flags & BIT_RD) && !iq->ratelimit_ok)
2127                     infra_ratelimit_dec(qstate->env->infra_cache, iq->dp->name,
2128                         iq->dp->namelen, *qstate->env->now);
2129                 return next_state(iq, QUERYTARGETS_STATE);
2130         }
2131         outbound_list_insert(&iq->outlist, outq);
2132         iq->num_current_queries++;
2133         iq->sent_count++;
2134         qstate->ext_state[id] = module_wait_reply;
2135
2136         return 0;
2137 }
2138
2139 /** find NS rrset in given list */
2140 static struct ub_packed_rrset_key*
2141 find_NS(struct reply_info* rep, size_t from, size_t to)
2142 {
2143         size_t i;
2144         for(i=from; i<to; i++) {
2145                 if(ntohs(rep->rrsets[i]->rk.type) == LDNS_RR_TYPE_NS)
2146                         return rep->rrsets[i];
2147         }
2148         return NULL;
2149 }
2150
2151
2152 /**
2153  * Process the query response. All queries end up at this state first. This
2154  * process generally consists of analyzing the response and routing the
2155  * event to the next state (either bouncing it back to a request state, or
2156  * terminating the processing for this event).
2157  *
2158  * @param qstate: query state.
2159  * @param iq: iterator query state.
2160  * @param id: module id.
2161  * @return true if the event requires more immediate processing, false if
2162  *         not. This is generally only true when forwarding the request to
2163  *         the final state (i.e., on answer).
2164  */
2165 static int
2166 processQueryResponse(struct module_qstate* qstate, struct iter_qstate* iq,
2167         int id)
2168 {
2169         int dnsseclame = 0;
2170         enum response_type type;
2171         iq->num_current_queries--;
2172         if(iq->response == NULL) {
2173                 /* Don't increment qname when QNAME minimisation is enabled */
2174                 if(qstate->env->cfg->qname_minimisation)
2175                         iq->minimisation_state = SKIP_MINIMISE_STATE;
2176                 iq->chase_to_rd = 0;
2177                 iq->dnssec_lame_query = 0;
2178                 verbose(VERB_ALGO, "query response was timeout");
2179                 return next_state(iq, QUERYTARGETS_STATE);
2180         }
2181         type = response_type_from_server(
2182                 (int)((iq->chase_flags&BIT_RD) || iq->chase_to_rd),
2183                 iq->response, &iq->qchase, iq->dp);
2184         iq->chase_to_rd = 0;
2185         if(type == RESPONSE_TYPE_REFERRAL && (iq->chase_flags&BIT_RD)) {
2186                 /* When forwarding (RD bit is set), we handle referrals
2187                  * differently. No queries should be sent elsewhere */
2188                 type = RESPONSE_TYPE_ANSWER;
2189         }
2190         if(!qstate->env->cfg->disable_dnssec_lame_check && iq->dnssec_expected
2191                 && !iq->dnssec_lame_query &&
2192                 !(iq->chase_flags&BIT_RD)
2193                 && iq->sent_count < DNSSEC_LAME_DETECT_COUNT
2194                 && type != RESPONSE_TYPE_LAME
2195                 && type != RESPONSE_TYPE_REC_LAME
2196                 && type != RESPONSE_TYPE_THROWAWAY
2197                 && type != RESPONSE_TYPE_UNTYPED) {
2198                 /* a possible answer, see if it is missing DNSSEC */
2199                 /* but not when forwarding, so we dont mark fwder lame */
2200                 if(!iter_msg_has_dnssec(iq->response)) {
2201                         /* Mark this address as dnsseclame in this dp,
2202                          * because that will make serverselection disprefer
2203                          * it, but also, once it is the only final option,
2204                          * use dnssec-lame-bypass if it needs to query there.*/
2205                         if(qstate->reply) {
2206                                 struct delegpt_addr* a = delegpt_find_addr(
2207                                         iq->dp, &qstate->reply->addr,
2208                                         qstate->reply->addrlen);
2209                                 if(a) a->dnsseclame = 1;
2210                         }
2211                         /* test the answer is from the zone we expected,
2212                          * otherwise, (due to parent,child on same server), we
2213                          * might mark the server,zone lame inappropriately */
2214                         if(!iter_msg_from_zone(iq->response, iq->dp, type,
2215                                 iq->qchase.qclass))
2216                                 qstate->reply = NULL;
2217                         type = RESPONSE_TYPE_LAME;
2218                         dnsseclame = 1;
2219                 }
2220         } else iq->dnssec_lame_query = 0;
2221         /* see if referral brings us close to the target */
2222         if(type == RESPONSE_TYPE_REFERRAL) {
2223                 struct ub_packed_rrset_key* ns = find_NS(
2224                         iq->response->rep, iq->response->rep->an_numrrsets,
2225                         iq->response->rep->an_numrrsets
2226                         + iq->response->rep->ns_numrrsets);
2227                 if(!ns) ns = find_NS(iq->response->rep, 0,
2228                                 iq->response->rep->an_numrrsets);
2229                 if(!ns || !dname_strict_subdomain_c(ns->rk.dname, iq->dp->name)
2230                         || !dname_subdomain_c(iq->qchase.qname, ns->rk.dname)){
2231                         verbose(VERB_ALGO, "bad referral, throwaway");
2232                         type = RESPONSE_TYPE_THROWAWAY;
2233                 } else
2234                         iter_scrub_ds(iq->response, ns, iq->dp->name);
2235         } else iter_scrub_ds(iq->response, NULL, NULL);
2236
2237         /* handle each of the type cases */
2238         if(type == RESPONSE_TYPE_ANSWER) {
2239                 /* ANSWER type responses terminate the query algorithm,
2240                  * so they sent on their */
2241                 if(verbosity >= VERB_DETAIL) {
2242                         verbose(VERB_DETAIL, "query response was %s",
2243                                 FLAGS_GET_RCODE(iq->response->rep->flags)
2244                                 ==LDNS_RCODE_NXDOMAIN?"NXDOMAIN ANSWER":
2245                                 (iq->response->rep->an_numrrsets?"ANSWER":
2246                                 "nodata ANSWER"));
2247                 }
2248                 /* if qtype is DS, check we have the right level of answer,
2249                  * like grandchild answer but we need the middle, reject it */
2250                 if(iq->qchase.qtype == LDNS_RR_TYPE_DS && !iq->dsns_point
2251                         && !(iq->chase_flags&BIT_RD)
2252                         && iter_ds_toolow(iq->response, iq->dp)
2253                         && iter_dp_cangodown(&iq->qchase, iq->dp)) {
2254                         /* close down outstanding requests to be discarded */
2255                         outbound_list_clear(&iq->outlist);
2256                         iq->num_current_queries = 0;
2257                         fptr_ok(fptr_whitelist_modenv_detach_subs(
2258                                 qstate->env->detach_subs));
2259                         (*qstate->env->detach_subs)(qstate);
2260                         iq->num_target_queries = 0;
2261                         return processDSNSFind(qstate, iq, id);
2262                 }
2263                 if(!qstate->no_cache_store)
2264                         iter_dns_store(qstate->env, &iq->response->qinfo,
2265                                 iq->response->rep, 0, qstate->prefetch_leeway,
2266                                 iq->dp&&iq->dp->has_parent_side_NS,
2267                                 qstate->region, qstate->query_flags);
2268                 /* close down outstanding requests to be discarded */
2269                 outbound_list_clear(&iq->outlist);
2270                 iq->num_current_queries = 0;
2271                 fptr_ok(fptr_whitelist_modenv_detach_subs(
2272                         qstate->env->detach_subs));
2273                 (*qstate->env->detach_subs)(qstate);
2274                 iq->num_target_queries = 0;
2275                 if(qstate->reply)
2276                         sock_list_insert(&qstate->reply_origin,
2277                                 &qstate->reply->addr, qstate->reply->addrlen,
2278                                 qstate->region);
2279                 if(iq->minimisation_state != DONOT_MINIMISE_STATE) {
2280                         if(FLAGS_GET_RCODE(iq->response->rep->flags) !=
2281                                 LDNS_RCODE_NOERROR) {
2282                                 if(qstate->env->cfg->qname_minimisation_strict)
2283                                         return final_state(iq);
2284                                 /* Best effort qname-minimisation.
2285                                  * Stop minimising and send full query when
2286                                  * RCODE is not NOERROR. */
2287                                 iq->minimisation_state = DONOT_MINIMISE_STATE;
2288                         }
2289                         if(FLAGS_GET_RCODE(iq->response->rep->flags) ==
2290                                 LDNS_RCODE_NXDOMAIN) {
2291                                 /* Stop resolving when NXDOMAIN is DNSSEC
2292                                  * signed. Based on assumption that namservers
2293                                  * serving signed zones do not return NXDOMAIN
2294                                  * for empty-non-terminals. */
2295                                 if(iq->dnssec_expected)
2296                                         return final_state(iq);
2297                                 /* Make subrequest to validate intermediate
2298                                  * NXDOMAIN if harden-below-nxdomain is
2299                                  * enabled. */
2300                                 if(qstate->env->cfg->harden_below_nxdomain) {
2301                                         struct module_qstate* subq = NULL;
2302                                         log_query_info(VERB_QUERY,
2303                                                 "schedule NXDOMAIN validation:",
2304                                                 &iq->response->qinfo);
2305                                         if(!generate_sub_request(
2306                                                 iq->response->qinfo.qname,
2307                                                 iq->response->qinfo.qname_len,
2308                                                 iq->response->qinfo.qtype,
2309                                                 iq->response->qinfo.qclass,
2310                                                 qstate, id, iq,
2311                                                 INIT_REQUEST_STATE,
2312                                                 FINISHED_STATE, &subq, 1))
2313                                                 verbose(VERB_ALGO,
2314                                                 "could not validate NXDOMAIN "
2315                                                 "response");
2316                                 }
2317                         }
2318                         return next_state(iq, QUERYTARGETS_STATE);
2319                 }
2320                 return final_state(iq);
2321         } else if(type == RESPONSE_TYPE_REFERRAL) {
2322                 /* REFERRAL type responses get a reset of the
2323                  * delegation point, and back to the QUERYTARGETS_STATE. */
2324                 verbose(VERB_DETAIL, "query response was REFERRAL");
2325
2326                 if(!(iq->chase_flags & BIT_RD) && !iq->ratelimit_ok) {
2327                         /* we have a referral, no ratelimit, we can send
2328                          * our queries to the given name */
2329                         infra_ratelimit_dec(qstate->env->infra_cache,
2330                                 iq->dp->name, iq->dp->namelen,
2331                                 *qstate->env->now);
2332                 }
2333
2334                 /* if hardened, only store referral if we asked for it */
2335                 if(!qstate->no_cache_store &&
2336                 (!qstate->env->cfg->harden_referral_path ||
2337                     (  qstate->qinfo.qtype == LDNS_RR_TYPE_NS
2338                         && (qstate->query_flags&BIT_RD)
2339                         && !(qstate->query_flags&BIT_CD)
2340                            /* we know that all other NS rrsets are scrubbed
2341                             * away, thus on referral only one is left.
2342                             * see if that equals the query name... */
2343                         && ( /* auth section, but sometimes in answer section*/
2344                           reply_find_rrset_section_ns(iq->response->rep,
2345                                 iq->qchase.qname, iq->qchase.qname_len,
2346                                 LDNS_RR_TYPE_NS, iq->qchase.qclass)
2347                           || reply_find_rrset_section_an(iq->response->rep,
2348                                 iq->qchase.qname, iq->qchase.qname_len,
2349                                 LDNS_RR_TYPE_NS, iq->qchase.qclass)
2350                           )
2351                     ))) {
2352                         /* Store the referral under the current query */
2353                         /* no prefetch-leeway, since its not the answer */
2354                         iter_dns_store(qstate->env, &iq->response->qinfo,
2355                                 iq->response->rep, 1, 0, 0, NULL, 0);
2356                         if(iq->store_parent_NS)
2357                                 iter_store_parentside_NS(qstate->env,
2358                                         iq->response->rep);
2359                         if(qstate->env->neg_cache)
2360                                 val_neg_addreferral(qstate->env->neg_cache,
2361                                         iq->response->rep, iq->dp->name);
2362                 }
2363                 /* store parent-side-in-zone-glue, if directly queried for */
2364                 if(!qstate->no_cache_store && iq->query_for_pside_glue
2365                         && !iq->pside_glue) {
2366                                 iq->pside_glue = reply_find_rrset(iq->response->rep,
2367                                         iq->qchase.qname, iq->qchase.qname_len,
2368                                         iq->qchase.qtype, iq->qchase.qclass);
2369                                 if(iq->pside_glue) {
2370                                         log_rrset_key(VERB_ALGO, "found parent-side "
2371                                                 "glue", iq->pside_glue);
2372                                         iter_store_parentside_rrset(qstate->env,
2373                                                 iq->pside_glue);
2374                                 }
2375                 }
2376
2377                 /* Reset the event state, setting the current delegation
2378                  * point to the referral. */
2379                 iq->deleg_msg = iq->response;
2380                 iq->dp = delegpt_from_message(iq->response, qstate->region);
2381                 if (qstate->env->cfg->qname_minimisation)
2382                         iq->minimisation_state = INIT_MINIMISE_STATE;
2383                 if(!iq->dp)
2384                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2385                 if(!cache_fill_missing(qstate->env, iq->qchase.qclass,
2386                         qstate->region, iq->dp))
2387                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2388                 if(iq->store_parent_NS && query_dname_compare(iq->dp->name,
2389                         iq->store_parent_NS->name) == 0)
2390                         iter_merge_retry_counts(iq->dp, iq->store_parent_NS);
2391                 delegpt_log(VERB_ALGO, iq->dp);
2392                 /* Count this as a referral. */
2393                 iq->referral_count++;
2394                 iq->sent_count = 0;
2395                 /* see if the next dp is a trust anchor, or a DS was sent
2396                  * along, indicating dnssec is expected for next zone */
2397                 iq->dnssec_expected = iter_indicates_dnssec(qstate->env,
2398                         iq->dp, iq->response, iq->qchase.qclass);
2399                 /* if dnssec, validating then also fetch the key for the DS */
2400                 if(iq->dnssec_expected && qstate->env->cfg->prefetch_key &&
2401                         !(qstate->query_flags&BIT_CD))
2402                         generate_dnskey_prefetch(qstate, iq, id);
2403
2404                 /* spawn off NS and addr to auth servers for the NS we just
2405                  * got in the referral. This gets authoritative answer
2406                  * (answer section trust level) rrset.
2407                  * right after, we detach the subs, answer goes to cache. */
2408                 if(qstate->env->cfg->harden_referral_path)
2409                         generate_ns_check(qstate, iq, id);
2410
2411                 /* stop current outstanding queries.
2412                  * FIXME: should the outstanding queries be waited for and
2413                  * handled? Say by a subquery that inherits the outbound_entry.
2414                  */
2415                 outbound_list_clear(&iq->outlist);
2416                 iq->num_current_queries = 0;
2417                 fptr_ok(fptr_whitelist_modenv_detach_subs(
2418                         qstate->env->detach_subs));
2419                 (*qstate->env->detach_subs)(qstate);
2420                 iq->num_target_queries = 0;
2421                 verbose(VERB_ALGO, "cleared outbound list for next round");
2422                 return next_state(iq, QUERYTARGETS_STATE);
2423         } else if(type == RESPONSE_TYPE_CNAME) {
2424                 uint8_t* sname = NULL;
2425                 size_t snamelen = 0;
2426                 /* CNAME type responses get a query restart (i.e., get a
2427                  * reset of the query state and go back to INIT_REQUEST_STATE).
2428                  */
2429                 verbose(VERB_DETAIL, "query response was CNAME");
2430                 if(verbosity >= VERB_ALGO)
2431                         log_dns_msg("cname msg", &iq->response->qinfo,
2432                                 iq->response->rep);
2433                 /* if qtype is DS, check we have the right level of answer,
2434                  * like grandchild answer but we need the middle, reject it */
2435                 if(iq->qchase.qtype == LDNS_RR_TYPE_DS && !iq->dsns_point
2436                         && !(iq->chase_flags&BIT_RD)
2437                         && iter_ds_toolow(iq->response, iq->dp)
2438                         && iter_dp_cangodown(&iq->qchase, iq->dp)) {
2439                         outbound_list_clear(&iq->outlist);
2440                         iq->num_current_queries = 0;
2441                         fptr_ok(fptr_whitelist_modenv_detach_subs(
2442                                 qstate->env->detach_subs));
2443                         (*qstate->env->detach_subs)(qstate);
2444                         iq->num_target_queries = 0;
2445                         return processDSNSFind(qstate, iq, id);
2446                 }
2447                 /* Process the CNAME response. */
2448                 if(!handle_cname_response(qstate, iq, iq->response,
2449                         &sname, &snamelen))
2450                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2451                 /* cache the CNAME response under the current query */
2452                 /* NOTE : set referral=1, so that rrsets get stored but not
2453                  * the partial query answer (CNAME only). */
2454                 /* prefetchleeway applied because this updates answer parts */
2455                 if(!qstate->no_cache_store)
2456                         iter_dns_store(qstate->env, &iq->response->qinfo,
2457                                 iq->response->rep, 1, qstate->prefetch_leeway,
2458                                 iq->dp&&iq->dp->has_parent_side_NS, NULL,
2459                                 qstate->query_flags);
2460                 /* set the current request's qname to the new value. */
2461                 iq->qchase.qname = sname;
2462                 iq->qchase.qname_len = snamelen;
2463                 if (qstate->env->cfg->qname_minimisation)
2464                         iq->minimisation_state = INIT_MINIMISE_STATE;
2465                 /* Clear the query state, since this is a query restart. */
2466                 iq->deleg_msg = NULL;
2467                 iq->dp = NULL;
2468                 iq->dsns_point = NULL;
2469                 /* Note the query restart. */
2470                 iq->query_restart_count++;
2471                 iq->sent_count = 0;
2472
2473                 /* stop current outstanding queries.
2474                  * FIXME: should the outstanding queries be waited for and
2475                  * handled? Say by a subquery that inherits the outbound_entry.
2476                  */
2477                 outbound_list_clear(&iq->outlist);
2478                 iq->num_current_queries = 0;
2479                 fptr_ok(fptr_whitelist_modenv_detach_subs(
2480                         qstate->env->detach_subs));
2481                 (*qstate->env->detach_subs)(qstate);
2482                 iq->num_target_queries = 0;
2483                 if(qstate->reply)
2484                         sock_list_insert(&qstate->reply_origin,
2485                                 &qstate->reply->addr, qstate->reply->addrlen,
2486                                 qstate->region);
2487                 verbose(VERB_ALGO, "cleared outbound list for query restart");
2488                 /* go to INIT_REQUEST_STATE for new qname. */
2489                 return next_state(iq, INIT_REQUEST_STATE);
2490         } else if(type == RESPONSE_TYPE_LAME) {
2491                 /* Cache the LAMEness. */
2492                 verbose(VERB_DETAIL, "query response was %sLAME",
2493                         dnsseclame?"DNSSEC ":"");
2494                 if(!dname_subdomain_c(iq->qchase.qname, iq->dp->name)) {
2495                         log_err("mark lame: mismatch in qname and dpname");
2496                         /* throwaway this reply below */
2497                 } else if(qstate->reply) {
2498                         /* need addr for lameness cache, but we may have
2499                          * gotten this from cache, so test to be sure */
2500                         if(!infra_set_lame(qstate->env->infra_cache,
2501                                 &qstate->reply->addr, qstate->reply->addrlen,
2502                                 iq->dp->name, iq->dp->namelen,
2503                                 *qstate->env->now, dnsseclame, 0,
2504                                 iq->qchase.qtype))
2505                                 log_err("mark host lame: out of memory");
2506                 }
2507         } else if(type == RESPONSE_TYPE_REC_LAME) {
2508                 /* Cache the LAMEness. */
2509                 verbose(VERB_DETAIL, "query response REC_LAME: "
2510                         "recursive but not authoritative server");
2511                 if(!dname_subdomain_c(iq->qchase.qname, iq->dp->name)) {
2512                         log_err("mark rec_lame: mismatch in qname and dpname");
2513                         /* throwaway this reply below */
2514                 } else if(qstate->reply) {
2515                         /* need addr for lameness cache, but we may have
2516                          * gotten this from cache, so test to be sure */
2517                         verbose(VERB_DETAIL, "mark as REC_LAME");
2518                         if(!infra_set_lame(qstate->env->infra_cache,
2519                                 &qstate->reply->addr, qstate->reply->addrlen,
2520                                 iq->dp->name, iq->dp->namelen,
2521                                 *qstate->env->now, 0, 1, iq->qchase.qtype))
2522                                 log_err("mark host lame: out of memory");
2523                 }
2524         } else if(type == RESPONSE_TYPE_THROWAWAY) {
2525                 /* LAME and THROWAWAY responses are handled the same way.
2526                  * In this case, the event is just sent directly back to
2527                  * the QUERYTARGETS_STATE without resetting anything,
2528                  * because, clearly, the next target must be tried. */
2529                 verbose(VERB_DETAIL, "query response was THROWAWAY");
2530         } else {
2531                 log_warn("A query response came back with an unknown type: %d",
2532                         (int)type);
2533         }
2534
2535         /* LAME, THROWAWAY and "unknown" all end up here.
2536          * Recycle to the QUERYTARGETS state to hopefully try a
2537          * different target. */
2538         if (qstate->env->cfg->qname_minimisation &&
2539                 !qstate->env->cfg->qname_minimisation_strict)
2540                 iq->minimisation_state = DONOT_MINIMISE_STATE;
2541         return next_state(iq, QUERYTARGETS_STATE);
2542 }
2543
2544 /**
2545  * Return priming query results to interested super querystates.
2546  *
2547  * Sets the delegation point and delegation message (not nonRD queries).
2548  * This is a callback from walk_supers.
2549  *
2550  * @param qstate: priming query state that finished.
2551  * @param id: module id.
2552  * @param forq: the qstate for which priming has been done.
2553  */
2554 static void
2555 prime_supers(struct module_qstate* qstate, int id, struct module_qstate* forq)
2556 {
2557         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
2558         struct delegpt* dp = NULL;
2559
2560         log_assert(qstate->is_priming || foriq->wait_priming_stub);
2561         log_assert(qstate->return_rcode == LDNS_RCODE_NOERROR);
2562         /* Convert our response to a delegation point */
2563         dp = delegpt_from_message(qstate->return_msg, forq->region);
2564         if(!dp) {
2565                 /* if there is no convertable delegation point, then
2566                  * the ANSWER type was (presumably) a negative answer. */
2567                 verbose(VERB_ALGO, "prime response was not a positive "
2568                         "ANSWER; failing");
2569                 foriq->dp = NULL;
2570                 foriq->state = QUERYTARGETS_STATE;
2571                 return;
2572         }
2573
2574         log_query_info(VERB_DETAIL, "priming successful for", &qstate->qinfo);
2575         delegpt_log(VERB_ALGO, dp);
2576         foriq->dp = dp;
2577         foriq->deleg_msg = dns_copy_msg(qstate->return_msg, forq->region);
2578         if(!foriq->deleg_msg) {
2579                 log_err("copy prime response: out of memory");
2580                 foriq->dp = NULL;
2581                 foriq->state = QUERYTARGETS_STATE;
2582                 return;
2583         }
2584
2585         /* root priming responses go to init stage 2, priming stub
2586          * responses to to stage 3. */
2587         if(foriq->wait_priming_stub) {
2588                 foriq->state = INIT_REQUEST_3_STATE;
2589                 foriq->wait_priming_stub = 0;
2590         } else  foriq->state = INIT_REQUEST_2_STATE;
2591         /* because we are finished, the parent will be reactivated */
2592 }
2593
2594 /**
2595  * This handles the response to a priming query. This is used to handle both
2596  * root and stub priming responses. This is basically the equivalent of the
2597  * QUERY_RESP_STATE, but will not handle CNAME responses and will treat
2598  * REFERRALs as ANSWERS. It will also update and reactivate the originating
2599  * event.
2600  *
2601  * @param qstate: query state.
2602  * @param id: module id.
2603  * @return true if the event needs more immediate processing, false if not.
2604  *         This state always returns false.
2605  */
2606 static int
2607 processPrimeResponse(struct module_qstate* qstate, int id)
2608 {
2609         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
2610         enum response_type type;
2611         iq->response->rep->flags &= ~(BIT_RD|BIT_RA); /* ignore rec-lame */
2612         type = response_type_from_server(
2613                 (int)((iq->chase_flags&BIT_RD) || iq->chase_to_rd),
2614                 iq->response, &iq->qchase, iq->dp);
2615         if(type == RESPONSE_TYPE_ANSWER) {
2616                 qstate->return_rcode = LDNS_RCODE_NOERROR;
2617                 qstate->return_msg = iq->response;
2618         } else {
2619                 qstate->return_rcode = LDNS_RCODE_SERVFAIL;
2620                 qstate->return_msg = NULL;
2621         }
2622
2623         /* validate the root or stub after priming (if enabled).
2624          * This is the same query as the prime query, but with validation.
2625          * Now that we are primed, the additional queries that validation
2626          * may need can be resolved, such as DLV. */
2627         if(qstate->env->cfg->harden_referral_path) {
2628                 struct module_qstate* subq = NULL;
2629                 log_nametypeclass(VERB_ALGO, "schedule prime validation",
2630                         qstate->qinfo.qname, qstate->qinfo.qtype,
2631                         qstate->qinfo.qclass);
2632                 if(!generate_sub_request(qstate->qinfo.qname,
2633                         qstate->qinfo.qname_len, qstate->qinfo.qtype,
2634                         qstate->qinfo.qclass, qstate, id, iq,
2635                         INIT_REQUEST_STATE, FINISHED_STATE, &subq, 1)) {
2636                         verbose(VERB_ALGO, "could not generate prime check");
2637                 }
2638                 generate_a_aaaa_check(qstate, iq, id);
2639         }
2640
2641         /* This event is finished. */
2642         qstate->ext_state[id] = module_finished;
2643         return 0;
2644 }
2645
2646 /**
2647  * Do final processing on responses to target queries. Events reach this
2648  * state after the iterative resolution algorithm terminates. This state is
2649  * responsible for reactiving the original event, and housekeeping related
2650  * to received target responses (caching, updating the current delegation
2651  * point, etc).
2652  * Callback from walk_supers for every super state that is interested in
2653  * the results from this query.
2654  *
2655  * @param qstate: query state.
2656  * @param id: module id.
2657  * @param forq: super query state.
2658  */
2659 static void
2660 processTargetResponse(struct module_qstate* qstate, int id,
2661         struct module_qstate* forq)
2662 {
2663         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
2664         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
2665         struct ub_packed_rrset_key* rrset;
2666         struct delegpt_ns* dpns;
2667         log_assert(qstate->return_rcode == LDNS_RCODE_NOERROR);
2668
2669         foriq->state = QUERYTARGETS_STATE;
2670         log_query_info(VERB_ALGO, "processTargetResponse", &qstate->qinfo);
2671         log_query_info(VERB_ALGO, "processTargetResponse super", &forq->qinfo);
2672
2673         /* Tell the originating event that this target query has finished
2674          * (regardless if it succeeded or not). */
2675         foriq->num_target_queries--;
2676
2677         /* check to see if parent event is still interested (in orig name).  */
2678         if(!foriq->dp) {
2679                 verbose(VERB_ALGO, "subq: parent not interested, was reset");
2680                 return; /* not interested anymore */
2681         }
2682         dpns = delegpt_find_ns(foriq->dp, qstate->qinfo.qname,
2683                         qstate->qinfo.qname_len);
2684         if(!dpns) {
2685                 /* If not interested, just stop processing this event */
2686                 verbose(VERB_ALGO, "subq: parent not interested anymore");
2687                 /* could be because parent was jostled out of the cache,
2688                    and a new identical query arrived, that does not want it*/
2689                 return;
2690         }
2691
2692         /* if iq->query_for_pside_glue then add the pside_glue (marked lame) */
2693         if(iq->pside_glue) {
2694                 /* if the pside_glue is NULL, then it could not be found,
2695                  * the done_pside is already set when created and a cache
2696                  * entry created in processFinished so nothing to do here */
2697                 log_rrset_key(VERB_ALGO, "add parentside glue to dp",
2698                         iq->pside_glue);
2699                 if(!delegpt_add_rrset(foriq->dp, forq->region,
2700                         iq->pside_glue, 1))
2701                         log_err("out of memory adding pside glue");
2702         }
2703
2704         /* This response is relevant to the current query, so we
2705          * add (attempt to add, anyway) this target(s) and reactivate
2706          * the original event.
2707          * NOTE: we could only look for the AnswerRRset if the
2708          * response type was ANSWER. */
2709         rrset = reply_find_answer_rrset(&iq->qchase, qstate->return_msg->rep);
2710         if(rrset) {
2711                 /* if CNAMEs have been followed - add new NS to delegpt. */
2712                 /* BTW. RFC 1918 says NS should not have got CNAMEs. Robust. */
2713                 if(!delegpt_find_ns(foriq->dp, rrset->rk.dname,
2714                         rrset->rk.dname_len)) {
2715                         /* if dpns->lame then set newcname ns lame too */
2716                         if(!delegpt_add_ns(foriq->dp, forq->region,
2717                                 rrset->rk.dname, dpns->lame))
2718                                 log_err("out of memory adding cnamed-ns");
2719                 }
2720                 /* if dpns->lame then set the address(es) lame too */
2721                 if(!delegpt_add_rrset(foriq->dp, forq->region, rrset,
2722                         dpns->lame))
2723                         log_err("out of memory adding targets");
2724                 verbose(VERB_ALGO, "added target response");
2725                 delegpt_log(VERB_ALGO, foriq->dp);
2726         } else {
2727                 verbose(VERB_ALGO, "iterator TargetResponse failed");
2728                 dpns->resolved = 1; /* fail the target */
2729         }
2730 }
2731
2732 /**
2733  * Process response for DS NS Find queries, that attempt to find the delegation
2734  * point where we ask the DS query from.
2735  *
2736  * @param qstate: query state.
2737  * @param id: module id.
2738  * @param forq: super query state.
2739  */
2740 static void
2741 processDSNSResponse(struct module_qstate* qstate, int id,
2742         struct module_qstate* forq)
2743 {
2744         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
2745
2746         /* if the finished (iq->response) query has no NS set: continue
2747          * up to look for the right dp; nothing to change, do DPNSstate */
2748         if(qstate->return_rcode != LDNS_RCODE_NOERROR)
2749                 return; /* seek further */
2750         /* find the NS RRset (without allowing CNAMEs) */
2751         if(!reply_find_rrset(qstate->return_msg->rep, qstate->qinfo.qname,
2752                 qstate->qinfo.qname_len, LDNS_RR_TYPE_NS,
2753                 qstate->qinfo.qclass)){
2754                 return; /* seek further */
2755         }
2756
2757         /* else, store as DP and continue at querytargets */
2758         foriq->state = QUERYTARGETS_STATE;
2759         foriq->dp = delegpt_from_message(qstate->return_msg, forq->region);
2760         if(!foriq->dp) {
2761                 log_err("out of memory in dsns dp alloc");
2762                 return; /* dp==NULL in QUERYTARGETS makes SERVFAIL */
2763         }
2764         /* success, go query the querytargets in the new dp (and go down) */
2765 }
2766
2767 /**
2768  * Process response for qclass=ANY queries for a particular class.
2769  * Append to result or error-exit.
2770  *
2771  * @param qstate: query state.
2772  * @param id: module id.
2773  * @param forq: super query state.
2774  */
2775 static void
2776 processClassResponse(struct module_qstate* qstate, int id,
2777         struct module_qstate* forq)
2778 {
2779         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
2780         struct dns_msg* from = qstate->return_msg;
2781         log_query_info(VERB_ALGO, "processClassResponse", &qstate->qinfo);
2782         log_query_info(VERB_ALGO, "processClassResponse super", &forq->qinfo);
2783         if(qstate->return_rcode != LDNS_RCODE_NOERROR) {
2784                 /* cause servfail for qclass ANY query */
2785                 foriq->response = NULL;
2786                 foriq->state = FINISHED_STATE;
2787                 return;
2788         }
2789         /* append result */
2790         if(!foriq->response) {
2791                 /* allocate the response: copy RCODE, sec_state */
2792                 foriq->response = dns_copy_msg(from, forq->region);
2793                 if(!foriq->response) {
2794                         log_err("malloc failed for qclass ANY response");
2795                         foriq->state = FINISHED_STATE;
2796                         return;
2797                 }
2798                 foriq->response->qinfo.qclass = forq->qinfo.qclass;
2799                 /* qclass ANY does not receive the AA flag on replies */
2800                 foriq->response->rep->authoritative = 0;
2801         } else {
2802                 struct dns_msg* to = foriq->response;
2803                 /* add _from_ this response _to_ existing collection */
2804                 /* if there are records, copy RCODE */
2805                 /* lower sec_state if this message is lower */
2806                 if(from->rep->rrset_count != 0) {
2807                         size_t n = from->rep->rrset_count+to->rep->rrset_count;
2808                         struct ub_packed_rrset_key** dest, **d;
2809                         /* copy appropriate rcode */
2810                         to->rep->flags = from->rep->flags;
2811                         /* copy rrsets */
2812                         if(from->rep->rrset_count > RR_COUNT_MAX ||
2813                                 to->rep->rrset_count > RR_COUNT_MAX) {
2814                                 log_err("malloc failed (too many rrsets) in collect ANY");
2815                                 foriq->state = FINISHED_STATE;
2816                                 return; /* integer overflow protection */
2817                         }
2818                         dest = regional_alloc(forq->region, sizeof(dest[0])*n);
2819                         if(!dest) {
2820                                 log_err("malloc failed in collect ANY");
2821                                 foriq->state = FINISHED_STATE;
2822                                 return;
2823                         }
2824                         d = dest;
2825                         /* copy AN */
2826                         memcpy(dest, to->rep->rrsets, to->rep->an_numrrsets
2827                                 * sizeof(dest[0]));
2828                         dest += to->rep->an_numrrsets;
2829                         memcpy(dest, from->rep->rrsets, from->rep->an_numrrsets
2830                                 * sizeof(dest[0]));
2831                         dest += from->rep->an_numrrsets;
2832                         /* copy NS */
2833                         memcpy(dest, to->rep->rrsets+to->rep->an_numrrsets,
2834                                 to->rep->ns_numrrsets * sizeof(dest[0]));
2835                         dest += to->rep->ns_numrrsets;
2836                         memcpy(dest, from->rep->rrsets+from->rep->an_numrrsets,
2837                                 from->rep->ns_numrrsets * sizeof(dest[0]));
2838                         dest += from->rep->ns_numrrsets;
2839                         /* copy AR */
2840                         memcpy(dest, to->rep->rrsets+to->rep->an_numrrsets+
2841                                 to->rep->ns_numrrsets,
2842                                 to->rep->ar_numrrsets * sizeof(dest[0]));
2843                         dest += to->rep->ar_numrrsets;
2844                         memcpy(dest, from->rep->rrsets+from->rep->an_numrrsets+
2845                                 from->rep->ns_numrrsets,
2846                                 from->rep->ar_numrrsets * sizeof(dest[0]));
2847                         /* update counts */
2848                         to->rep->rrsets = d;
2849                         to->rep->an_numrrsets += from->rep->an_numrrsets;
2850                         to->rep->ns_numrrsets += from->rep->ns_numrrsets;
2851                         to->rep->ar_numrrsets += from->rep->ar_numrrsets;
2852                         to->rep->rrset_count = n;
2853                 }
2854                 if(from->rep->security < to->rep->security) /* lowest sec */
2855                         to->rep->security = from->rep->security;
2856                 if(from->rep->qdcount != 0) /* insert qd if appropriate */
2857                         to->rep->qdcount = from->rep->qdcount;
2858                 if(from->rep->ttl < to->rep->ttl) /* use smallest TTL */
2859                         to->rep->ttl = from->rep->ttl;
2860                 if(from->rep->prefetch_ttl < to->rep->prefetch_ttl)
2861                         to->rep->prefetch_ttl = from->rep->prefetch_ttl;
2862         }
2863         /* are we done? */
2864         foriq->num_current_queries --;
2865         if(foriq->num_current_queries == 0)
2866                 foriq->state = FINISHED_STATE;
2867 }
2868
2869 /**
2870  * Collect class ANY responses and make them into one response.  This
2871  * state is started and it creates queries for all classes (that have
2872  * root hints).  The answers are then collected.
2873  *
2874  * @param qstate: query state.
2875  * @param id: module id.
2876  * @return true if the event needs more immediate processing, false if not.
2877  */
2878 static int
2879 processCollectClass(struct module_qstate* qstate, int id)
2880 {
2881         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
2882         struct module_qstate* subq;
2883         /* If qchase.qclass == 0 then send out queries for all classes.
2884          * Otherwise, do nothing (wait for all answers to arrive and the
2885          * processClassResponse to put them together, and that moves us
2886          * towards the Finished state when done. */
2887         if(iq->qchase.qclass == 0) {
2888                 uint16_t c = 0;
2889                 iq->qchase.qclass = LDNS_RR_CLASS_ANY;
2890                 while(iter_get_next_root(qstate->env->hints,
2891                         qstate->env->fwds, &c)) {
2892                         /* generate query for this class */
2893                         log_nametypeclass(VERB_ALGO, "spawn collect query",
2894                                 qstate->qinfo.qname, qstate->qinfo.qtype, c);
2895                         if(!generate_sub_request(qstate->qinfo.qname,
2896                                 qstate->qinfo.qname_len, qstate->qinfo.qtype,
2897                                 c, qstate, id, iq, INIT_REQUEST_STATE,
2898                                 FINISHED_STATE, &subq,
2899                                 (int)!(qstate->query_flags&BIT_CD))) {
2900                                 return error_response(qstate, id,
2901                                         LDNS_RCODE_SERVFAIL);
2902                         }
2903                         /* ignore subq, no special init required */
2904                         iq->num_current_queries ++;
2905                         if(c == 0xffff)
2906                                 break;
2907                         else c++;
2908                 }
2909                 /* if no roots are configured at all, return */
2910                 if(iq->num_current_queries == 0) {
2911                         verbose(VERB_ALGO, "No root hints or fwds, giving up "
2912                                 "on qclass ANY");
2913                         return error_response(qstate, id, LDNS_RCODE_REFUSED);
2914                 }
2915                 /* return false, wait for queries to return */
2916         }
2917         /* if woke up here because of an answer, wait for more answers */
2918         return 0;
2919 }
2920
2921 /**
2922  * This handles the final state for first-tier responses (i.e., responses to
2923  * externally generated queries).
2924  *
2925  * @param qstate: query state.
2926  * @param iq: iterator query state.
2927  * @param id: module id.
2928  * @return true if the event needs more processing, false if not. Since this
2929  *         is the final state for an event, it always returns false.
2930  */
2931 static int
2932 processFinished(struct module_qstate* qstate, struct iter_qstate* iq,
2933         int id)
2934 {
2935         log_query_info(VERB_QUERY, "finishing processing for",
2936                 &qstate->qinfo);
2937
2938         /* store negative cache element for parent side glue. */
2939         if(!qstate->no_cache_store && iq->query_for_pside_glue
2940                 && !iq->pside_glue)
2941                         iter_store_parentside_neg(qstate->env, &qstate->qinfo,
2942                                 iq->deleg_msg?iq->deleg_msg->rep:
2943                                 (iq->response?iq->response->rep:NULL));
2944         if(!iq->response) {
2945                 verbose(VERB_ALGO, "No response is set, servfail");
2946                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2947         }
2948
2949         /* Make sure that the RA flag is set (since the presence of
2950          * this module means that recursion is available) */
2951         iq->response->rep->flags |= BIT_RA;
2952
2953         /* Clear the AA flag */
2954         /* FIXME: does this action go here or in some other module? */
2955         iq->response->rep->flags &= ~BIT_AA;
2956
2957         /* make sure QR flag is on */
2958         iq->response->rep->flags |= BIT_QR;
2959
2960         /* we have finished processing this query */
2961         qstate->ext_state[id] = module_finished;
2962
2963         /* TODO:  we are using a private TTL, trim the response. */
2964         /* if (mPrivateTTL > 0){IterUtils.setPrivateTTL(resp, mPrivateTTL); } */
2965
2966         /* prepend any items we have accumulated */
2967         if(iq->an_prepend_list || iq->ns_prepend_list) {
2968                 if(!iter_prepend(iq, iq->response, qstate->region)) {
2969                         log_err("prepend rrsets: out of memory");
2970                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2971                 }
2972                 /* reset the query name back */
2973                 iq->response->qinfo = qstate->qinfo;
2974                 /* the security state depends on the combination */
2975                 iq->response->rep->security = sec_status_unchecked;
2976                 /* store message with the finished prepended items,
2977                  * but only if we did recursion. The nonrecursion referral
2978                  * from cache does not need to be stored in the msg cache. */
2979                 if(!qstate->no_cache_store && qstate->query_flags&BIT_RD) {
2980                         iter_dns_store(qstate->env, &qstate->qinfo,
2981                                 iq->response->rep, 0, qstate->prefetch_leeway,
2982                                 iq->dp&&iq->dp->has_parent_side_NS,
2983                                 qstate->region, qstate->query_flags);
2984                 }
2985         }
2986         qstate->return_rcode = LDNS_RCODE_NOERROR;
2987         qstate->return_msg = iq->response;
2988         return 0;
2989 }
2990
2991 /*
2992  * Return priming query results to interestes super querystates.
2993  *
2994  * Sets the delegation point and delegation message (not nonRD queries).
2995  * This is a callback from walk_supers.
2996  *
2997  * @param qstate: query state that finished.
2998  * @param id: module id.
2999  * @param super: the qstate to inform.
3000  */
3001 void
3002 iter_inform_super(struct module_qstate* qstate, int id,
3003         struct module_qstate* super)
3004 {
3005         if(!qstate->is_priming && super->qinfo.qclass == LDNS_RR_CLASS_ANY)
3006                 processClassResponse(qstate, id, super);
3007         else if(super->qinfo.qtype == LDNS_RR_TYPE_DS && ((struct iter_qstate*)
3008                 super->minfo[id])->state == DSNS_FIND_STATE)
3009                 processDSNSResponse(qstate, id, super);
3010         else if(qstate->return_rcode != LDNS_RCODE_NOERROR)
3011                 error_supers(qstate, id, super);
3012         else if(qstate->is_priming)
3013                 prime_supers(qstate, id, super);
3014         else    processTargetResponse(qstate, id, super);
3015 }
3016
3017 /**
3018  * Handle iterator state.
3019  * Handle events. This is the real processing loop for events, responsible
3020  * for moving events through the various states. If a processing method
3021  * returns true, then it will be advanced to the next state. If false, then
3022  * processing will stop.
3023  *
3024  * @param qstate: query state.
3025  * @param ie: iterator shared global environment.
3026  * @param iq: iterator query state.
3027  * @param id: module id.
3028  */
3029 static void
3030 iter_handle(struct module_qstate* qstate, struct iter_qstate* iq,
3031         struct iter_env* ie, int id)
3032 {
3033         int cont = 1;
3034         while(cont) {
3035                 verbose(VERB_ALGO, "iter_handle processing q with state %s",
3036                         iter_state_to_string(iq->state));
3037                 switch(iq->state) {
3038                         case INIT_REQUEST_STATE:
3039                                 cont = processInitRequest(qstate, iq, ie, id);
3040                                 break;
3041                         case INIT_REQUEST_2_STATE:
3042                                 cont = processInitRequest2(qstate, iq, id);
3043                                 break;
3044                         case INIT_REQUEST_3_STATE:
3045                                 cont = processInitRequest3(qstate, iq, id);
3046                                 break;
3047                         case QUERYTARGETS_STATE:
3048                                 cont = processQueryTargets(qstate, iq, ie, id);
3049                                 break;
3050                         case QUERY_RESP_STATE:
3051                                 cont = processQueryResponse(qstate, iq, id);
3052                                 break;
3053                         case PRIME_RESP_STATE:
3054                                 cont = processPrimeResponse(qstate, id);
3055                                 break;
3056                         case COLLECT_CLASS_STATE:
3057                                 cont = processCollectClass(qstate, id);
3058                                 break;
3059                         case DSNS_FIND_STATE:
3060                                 cont = processDSNSFind(qstate, iq, id);
3061                                 break;
3062                         case FINISHED_STATE:
3063                                 cont = processFinished(qstate, iq, id);
3064                                 break;
3065                         default:
3066                                 log_warn("iterator: invalid state: %d",
3067                                         iq->state);
3068                                 cont = 0;
3069                                 break;
3070                 }
3071         }
3072 }
3073
3074 /**
3075  * This is the primary entry point for processing request events. Note that
3076  * this method should only be used by external modules.
3077  * @param qstate: query state.
3078  * @param ie: iterator shared global environment.
3079  * @param iq: iterator query state.
3080  * @param id: module id.
3081  */
3082 static void
3083 process_request(struct module_qstate* qstate, struct iter_qstate* iq,
3084         struct iter_env* ie, int id)
3085 {
3086         /* external requests start in the INIT state, and finish using the
3087          * FINISHED state. */
3088         iq->state = INIT_REQUEST_STATE;
3089         iq->final_state = FINISHED_STATE;
3090         verbose(VERB_ALGO, "process_request: new external request event");
3091         iter_handle(qstate, iq, ie, id);
3092 }
3093
3094 /** process authoritative server reply */
3095 static void
3096 process_response(struct module_qstate* qstate, struct iter_qstate* iq,
3097         struct iter_env* ie, int id, struct outbound_entry* outbound,
3098         enum module_ev event)
3099 {
3100         struct msg_parse* prs;
3101         struct edns_data edns;
3102         sldns_buffer* pkt;
3103
3104         verbose(VERB_ALGO, "process_response: new external response event");
3105         iq->response = NULL;
3106         iq->state = QUERY_RESP_STATE;
3107         if(event == module_event_noreply || event == module_event_error) {
3108                 if(event == module_event_noreply && iq->sent_count >= 3 &&
3109                         qstate->env->cfg->use_caps_bits_for_id &&
3110                         !iq->caps_fallback) {
3111                         /* start fallback */
3112                         iq->caps_fallback = 1;
3113                         iq->caps_server = 0;
3114                         iq->caps_reply = NULL;
3115                         iq->caps_response = NULL;
3116                         iq->state = QUERYTARGETS_STATE;
3117                         iq->num_current_queries--;
3118                         /* need fresh attempts for the 0x20 fallback, if
3119                          * that was the cause for the failure */
3120                         iter_dec_attempts(iq->dp, 3);
3121                         verbose(VERB_DETAIL, "Capsforid: timeouts, starting fallback");
3122                         goto handle_it;
3123                 }
3124                 goto handle_it;
3125         }
3126         if( (event != module_event_reply && event != module_event_capsfail)
3127                 || !qstate->reply) {
3128                 log_err("Bad event combined with response");
3129                 outbound_list_remove(&iq->outlist, outbound);
3130                 (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3131                 return;
3132         }
3133
3134         /* parse message */
3135         prs = (struct msg_parse*)regional_alloc(qstate->env->scratch,
3136                 sizeof(struct msg_parse));
3137         if(!prs) {
3138                 log_err("out of memory on incoming message");
3139                 /* like packet got dropped */
3140                 goto handle_it;
3141         }
3142         memset(prs, 0, sizeof(*prs));
3143         memset(&edns, 0, sizeof(edns));
3144         pkt = qstate->reply->c->buffer;
3145         sldns_buffer_set_position(pkt, 0);
3146         if(parse_packet(pkt, prs, qstate->env->scratch) != LDNS_RCODE_NOERROR) {
3147                 verbose(VERB_ALGO, "parse error on reply packet");
3148                 goto handle_it;
3149         }
3150         /* edns is not examined, but removed from message to help cache */
3151         if(parse_extract_edns(prs, &edns, qstate->env->scratch) !=
3152                 LDNS_RCODE_NOERROR)
3153                 goto handle_it;
3154
3155         /* Copy the edns options we may got from the back end */
3156         if(edns.opt_list) {
3157                 qstate->edns_opts_back_in = edns_opt_copy_region(edns.opt_list,
3158                         qstate->region);
3159                 if(!qstate->edns_opts_back_in) {
3160                         log_err("out of memory on incoming message");
3161                         /* like packet got dropped */
3162                         goto handle_it;
3163                 }
3164         }
3165
3166         /* remove CD-bit, we asked for in case we handle validation ourself */
3167         prs->flags &= ~BIT_CD;
3168
3169         /* normalize and sanitize: easy to delete items from linked lists */
3170         if(!scrub_message(pkt, prs, &iq->qinfo_out, iq->dp->name,
3171                 qstate->env->scratch, qstate->env, ie)) {
3172                 /* if 0x20 enabled, start fallback, but we have no message */
3173                 if(event == module_event_capsfail && !iq->caps_fallback) {
3174                         iq->caps_fallback = 1;
3175                         iq->caps_server = 0;
3176                         iq->caps_reply = NULL;
3177                         iq->caps_response = NULL;
3178                         iq->state = QUERYTARGETS_STATE;
3179                         iq->num_current_queries--;
3180                         verbose(VERB_DETAIL, "Capsforid: scrub failed, starting fallback with no response");
3181                 }
3182                 goto handle_it;
3183         }
3184
3185         /* allocate response dns_msg in region */
3186         iq->response = dns_alloc_msg(pkt, prs, qstate->region);
3187         if(!iq->response)
3188                 goto handle_it;
3189         log_query_info(VERB_DETAIL, "response for", &qstate->qinfo);
3190         log_name_addr(VERB_DETAIL, "reply from", iq->dp->name,
3191                 &qstate->reply->addr, qstate->reply->addrlen);
3192         if(verbosity >= VERB_ALGO)
3193                 log_dns_msg("incoming scrubbed packet:", &iq->response->qinfo,
3194                         iq->response->rep);
3195
3196         if(event == module_event_capsfail || iq->caps_fallback) {
3197                 /* for fallback we care about main answer, not additionals */
3198                 /* removing that makes comparison more likely to succeed */
3199                 caps_strip_reply(iq->response->rep);
3200                 if(!iq->caps_fallback) {
3201                         /* start fallback */
3202                         iq->caps_fallback = 1;
3203                         iq->caps_server = 0;
3204                         iq->caps_reply = iq->response->rep;
3205                         iq->caps_response = iq->response;
3206                         iq->state = QUERYTARGETS_STATE;
3207                         iq->num_current_queries--;
3208                         verbose(VERB_DETAIL, "Capsforid: starting fallback");
3209                         goto handle_it;
3210                 } else {
3211                         /* check if reply is the same, otherwise, fail */
3212                         if(!iq->caps_reply) {
3213                                 iq->caps_reply = iq->response->rep;
3214                                 iq->caps_response = iq->response;
3215                                 iq->caps_server = -1; /*become zero at ++,
3216                                 so that we start the full set of trials */
3217                         } else if(caps_failed_rcode(iq->caps_reply) &&
3218                                 !caps_failed_rcode(iq->response->rep)) {
3219                                 /* prefer to upgrade to non-SERVFAIL */
3220                                 iq->caps_reply = iq->response->rep;
3221                                 iq->caps_response = iq->response;
3222                         } else if(!caps_failed_rcode(iq->caps_reply) &&
3223                                 caps_failed_rcode(iq->response->rep)) {
3224                                 /* if we have non-SERVFAIL as answer then
3225                                  * we can ignore SERVFAILs for the equality
3226                                  * comparison */
3227                                 /* no instructions here, skip other else */
3228                         } else if(caps_failed_rcode(iq->caps_reply) &&
3229                                 caps_failed_rcode(iq->response->rep)) {
3230                                 /* failure is same as other failure in fallbk*/
3231                                 /* no instructions here, skip other else */
3232                         } else if(!reply_equal(iq->response->rep, iq->caps_reply,
3233                                 qstate->env->scratch)) {
3234                                 verbose(VERB_DETAIL, "Capsforid fallback: "
3235                                         "getting different replies, failed");
3236                                 outbound_list_remove(&iq->outlist, outbound);
3237                                 (void)error_response(qstate, id,
3238                                         LDNS_RCODE_SERVFAIL);
3239                                 return;
3240                         }
3241                         /* continue the fallback procedure at next server */
3242                         iq->caps_server++;
3243                         iq->state = QUERYTARGETS_STATE;
3244                         iq->num_current_queries--;
3245                         verbose(VERB_DETAIL, "Capsforid: reply is equal. "
3246                                 "go to next fallback");
3247                         goto handle_it;
3248                 }
3249         }
3250         iq->caps_fallback = 0; /* if we were in fallback, 0x20 is OK now */
3251
3252 handle_it:
3253         outbound_list_remove(&iq->outlist, outbound);
3254         iter_handle(qstate, iq, ie, id);
3255 }
3256
3257 void
3258 iter_operate(struct module_qstate* qstate, enum module_ev event, int id,
3259         struct outbound_entry* outbound)
3260 {
3261         struct iter_env* ie = (struct iter_env*)qstate->env->modinfo[id];
3262         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
3263         verbose(VERB_QUERY, "iterator[module %d] operate: extstate:%s event:%s",
3264                 id, strextstate(qstate->ext_state[id]), strmodulevent(event));
3265         if(iq) log_query_info(VERB_QUERY, "iterator operate: query",
3266                 &qstate->qinfo);
3267         if(iq && qstate->qinfo.qname != iq->qchase.qname)
3268                 log_query_info(VERB_QUERY, "iterator operate: chased to",
3269                         &iq->qchase);
3270
3271         /* perform iterator state machine */
3272         if((event == module_event_new || event == module_event_pass) &&
3273                 iq == NULL) {
3274                 if(!iter_new(qstate, id)) {
3275                         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3276                         return;
3277                 }
3278                 iq = (struct iter_qstate*)qstate->minfo[id];
3279                 process_request(qstate, iq, ie, id);
3280                 return;
3281         }
3282         if(iq && event == module_event_pass) {
3283                 iter_handle(qstate, iq, ie, id);
3284                 return;
3285         }
3286         if(iq && outbound) {
3287                 process_response(qstate, iq, ie, id, outbound, event);
3288                 return;
3289         }
3290         if(event == module_event_error) {
3291                 verbose(VERB_ALGO, "got called with event error, giving up");
3292                 (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3293                 return;
3294         }
3295
3296         log_err("bad event for iterator");
3297         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3298 }
3299
3300 void
3301 iter_clear(struct module_qstate* qstate, int id)
3302 {
3303         struct iter_qstate* iq;
3304         if(!qstate)
3305                 return;
3306         iq = (struct iter_qstate*)qstate->minfo[id];
3307         if(iq) {
3308                 outbound_list_clear(&iq->outlist);
3309                 if(iq->target_count && --iq->target_count[0] == 0)
3310                         free(iq->target_count);
3311                 iq->num_current_queries = 0;
3312         }
3313         qstate->minfo[id] = NULL;
3314 }
3315
3316 size_t
3317 iter_get_mem(struct module_env* env, int id)
3318 {
3319         struct iter_env* ie = (struct iter_env*)env->modinfo[id];
3320         if(!ie)
3321                 return 0;
3322         return sizeof(*ie) + sizeof(int)*((size_t)ie->max_dependency_depth+1)
3323                 + donotq_get_mem(ie->donotq) + priv_get_mem(ie->priv);
3324 }
3325
3326 /**
3327  * The iterator function block
3328  */
3329 static struct module_func_block iter_block = {
3330         "iterator",
3331         &iter_init, &iter_deinit, &iter_operate, &iter_inform_super,
3332         &iter_clear, &iter_get_mem
3333 };
3334
3335 struct module_func_block*
3336 iter_get_funcblock(void)
3337 {
3338         return &iter_block;
3339 }
3340
3341 const char*
3342 iter_state_to_string(enum iter_state state)
3343 {
3344         switch (state)
3345         {
3346         case INIT_REQUEST_STATE :
3347                 return "INIT REQUEST STATE";
3348         case INIT_REQUEST_2_STATE :
3349                 return "INIT REQUEST STATE (stage 2)";
3350         case INIT_REQUEST_3_STATE:
3351                 return "INIT REQUEST STATE (stage 3)";
3352         case QUERYTARGETS_STATE :
3353                 return "QUERY TARGETS STATE";
3354         case PRIME_RESP_STATE :
3355                 return "PRIME RESPONSE STATE";
3356         case COLLECT_CLASS_STATE :
3357                 return "COLLECT CLASS STATE";
3358         case DSNS_FIND_STATE :
3359                 return "DSNS FIND STATE";
3360         case QUERY_RESP_STATE :
3361                 return "QUERY RESPONSE STATE";
3362         case FINISHED_STATE :
3363                 return "FINISHED RESPONSE STATE";
3364         default :
3365                 return "UNKNOWN ITER STATE";
3366         }
3367 }
3368
3369 int
3370 iter_state_is_responsestate(enum iter_state s)
3371 {
3372         switch(s) {
3373                 case INIT_REQUEST_STATE :
3374                 case INIT_REQUEST_2_STATE :
3375                 case INIT_REQUEST_3_STATE :
3376                 case QUERYTARGETS_STATE :
3377                 case COLLECT_CLASS_STATE :
3378                         return 0;
3379                 default:
3380                         break;
3381         }
3382         return 1;
3383 }