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 recursive 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 "services/authzone.h"
  57 #include "util/module.h"
  58 #include "util/netevent.h"
  59 #include "util/net_help.h"
  60 #include "util/regional.h"
  61 #include "util/data/dname.h"
  62 #include "util/data/msgencode.h"
  63 #include "util/fptr_wlist.h"
  64 #include "util/config_file.h"
  65 #include "util/random.h"
  66 #include "sldns/rrdef.h"
  67 #include "sldns/wire2str.h"
  68 #include "sldns/str2wire.h"
  69 #include "sldns/parseutil.h"
  70 #include "sldns/sbuffer.h"
  71
  72 /* in msec */
  73 int UNKNOWN_SERVER_NICENESS = 376;
  74
  75 int
  76 iter_init(struct module_env* env, int id)
  77 {
  78         struct iter_env* iter_env = (struct iter_env*)calloc(1,
  79                 sizeof(struct iter_env));
  80         if(!iter_env) {
  81                 log_err("malloc failure");
  82                 return 0;
  83         }
  84         env->modinfo[id] = (void*)iter_env;
  85
  86         lock_basic_init(&iter_env->queries_ratelimit_lock);
  87         lock_protect(&iter_env->queries_ratelimit_lock,
  88                         &iter_env->num_queries_ratelimited,
  89                 sizeof(iter_env->num_queries_ratelimited));
  90
  91         if(!iter_apply_cfg(iter_env, env->cfg)) {
  92                 log_err("iterator: could not apply configuration settings.");
  93                 return 0;
  94         }
  95
  96         return 1;
  97 }
  98
  99 /** delete caps_whitelist element */
 100 static void
 101 caps_free(struct rbnode_type* n, void* ATTR_UNUSED(d))
 102 {
 103         if(n) {
 104                 free(((struct name_tree_node*)n)->name);
 105                 free(n);
 106         }
 107 }
 108
 109 void
 110 iter_deinit(struct module_env* env, int id)
 111 {
 112         struct iter_env* iter_env;
 113         if(!env || !env->modinfo[id])
 114                 return;
 115         iter_env = (struct iter_env*)env->modinfo[id];
 116         lock_basic_destroy(&iter_env->queries_ratelimit_lock);
 117         free(iter_env->target_fetch_policy);
 118         priv_delete(iter_env->priv);
 119         donotq_delete(iter_env->donotq);
 120         if(iter_env->caps_white) {
 121                 traverse_postorder(iter_env->caps_white, caps_free, NULL);
 122                 free(iter_env->caps_white);
 123         }
 124         free(iter_env);
 125         env->modinfo[id] = NULL;
 126 }
 127
 128 /** new query for iterator */
 129 static int
 130 iter_new(struct module_qstate* qstate, int id)
 131 {
 132         struct iter_qstate* iq = (struct iter_qstate*)regional_alloc(
 133                 qstate->region, sizeof(struct iter_qstate));
 134         qstate->minfo[id] = iq;
 135         if(!iq)
 136                 return 0;
 137         memset(iq, 0, sizeof(*iq));
 138         iq->state = INIT_REQUEST_STATE;
 139         iq->final_state = FINISHED_STATE;
 140         iq->an_prepend_list = NULL;
 141         iq->an_prepend_last = NULL;
 142         iq->ns_prepend_list = NULL;
 143         iq->ns_prepend_last = NULL;
 144         iq->dp = NULL;
 145         iq->depth = 0;
 146         iq->num_target_queries = 0;
 147         iq->num_current_queries = 0;
 148         iq->query_restart_count = 0;
 149         iq->referral_count = 0;
 150         iq->sent_count = 0;
 151         iq->ratelimit_ok = 0;
 152         iq->target_count = NULL;
 153         iq->wait_priming_stub = 0;
 154         iq->refetch_glue = 0;
 155         iq->dnssec_expected = 0;
 156         iq->dnssec_lame_query = 0;
 157         iq->chase_flags = qstate->query_flags;
 158         /* Start with the (current) qname. */
 159         iq->qchase = qstate->qinfo;
 160         outbound_list_init(&iq->outlist);
 161         iq->minimise_count = 0;
 162         iq->minimise_timeout_count = 0;
 163         if (qstate->env->cfg->qname_minimisation)
 164                 iq->minimisation_state = INIT_MINIMISE_STATE;
 165         else
 166                 iq->minimisation_state = DONOT_MINIMISE_STATE;
 167
 168         memset(&iq->qinfo_out, 0, sizeof(struct query_info));
 169         return 1;
 170 }
 171
 172 /**
 173  * Transition to the next state. This can be used to advance a currently
 174  * processing event. It cannot be used to reactivate a forEvent.
 175  *
 176  * @param iq: iterator query state
 177  * @param nextstate The state to transition to.
 178  * @return true. This is so this can be called as the return value for the
 179  *         actual process*State() methods. (Transitioning to the next state
 180  *         implies further processing).
 181  */
 182 static int
 183 next_state(struct iter_qstate* iq, enum iter_state nextstate)
 184 {
 185         /* If transitioning to a "response" state, make sure that there is a
 186          * response */
 187         if(iter_state_is_responsestate(nextstate)) {
 188                 if(iq->response == NULL) {
 189                         log_err("transitioning to response state sans "
 190                                 "response.");
 191                 }
 192         }
 193         iq->state = nextstate;
 194         return 1;
 195 }
 196
 197 /**
 198  * Transition an event to its final state. Final states always either return
 199  * a result up the module chain, or reactivate a dependent event. Which
 200  * final state to transition to is set in the module state for the event when
 201  * it was created, and depends on the original purpose of the event.
 202  *
 203  * The response is stored in the qstate->buf buffer.
 204  *
 205  * @param iq: iterator query state
 206  * @return false. This is so this method can be used as the return value for
 207  *         the processState methods. (Transitioning to the final state
 208  */
 209 static int
 210 final_state(struct iter_qstate* iq)
 211 {
 212         return next_state(iq, iq->final_state);
 213 }
 214
 215 /**
 216  * Callback routine to handle errors in parent query states
 217  * @param qstate: query state that failed.
 218  * @param id: module id.
 219  * @param super: super state.
 220  */
 221 static void
 222 error_supers(struct module_qstate* qstate, int id, struct module_qstate* super)
 223 {
 224         struct iter_qstate* super_iq = (struct iter_qstate*)super->minfo[id];
 225
 226         if(qstate->qinfo.qtype == LDNS_RR_TYPE_A ||
 227                 qstate->qinfo.qtype == LDNS_RR_TYPE_AAAA) {
 228                 /* mark address as failed. */
 229                 struct delegpt_ns* dpns = NULL;
 230                 super_iq->num_target_queries--;
 231                 if(super_iq->dp)
 232                         dpns = delegpt_find_ns(super_iq->dp,
 233                                 qstate->qinfo.qname, qstate->qinfo.qname_len);
 234                 if(!dpns) {
 235                         /* not interested */
 236                         /* this can happen, for eg. qname minimisation asked
 237                          * for an NXDOMAIN to be validated, and used qtype
 238                          * A for that, and the error of that, the name, is
 239                          * not listed in super_iq->dp */
 240                         verbose(VERB_ALGO, "subq error, but not interested");
 241                         log_query_info(VERB_ALGO, "superq", &super->qinfo);
 242                         return;
 243                 } else {
 244                         /* see if the failure did get (parent-lame) info */
 245                         if(!cache_fill_missing(super->env, super_iq->qchase.qclass,
 246                                 super->region, super_iq->dp))
 247                                 log_err("out of memory adding missing");
 248                 }
 249                 dpns->resolved = 1; /* mark as failed */
 250         }
 251         if(qstate->qinfo.qtype == LDNS_RR_TYPE_NS) {
 252                 /* prime failed to get delegation */
 253                 super_iq->dp = NULL;
 254         }
 255         /* evaluate targets again */
 256         super_iq->state = QUERYTARGETS_STATE;
 257         /* super becomes runnable, and will process this change */
 258 }
 259
 260 /**
 261  * Return an error to the client
 262  * @param qstate: our query state
 263  * @param id: module id
 264  * @param rcode: error code (DNS errcode).
 265  * @return: 0 for use by caller, to make notation easy, like:
 266  *      return error_response(..).
 267  */
 268 static int
 269 error_response(struct module_qstate* qstate, int id, int rcode)
 270 {
 271         verbose(VERB_QUERY, "return error response %s",
 272                 sldns_lookup_by_id(sldns_rcodes, rcode)?
 273                 sldns_lookup_by_id(sldns_rcodes, rcode)->name:"??");
 274         qstate->return_rcode = rcode;
 275         qstate->return_msg = NULL;
 276         qstate->ext_state[id] = module_finished;
 277         return 0;
 278 }
 279
 280 /**
 281  * Return an error to the client and cache the error code in the
 282  * message cache (so per qname, qtype, qclass).
 283  * @param qstate: our query state
 284  * @param id: module id
 285  * @param rcode: error code (DNS errcode).
 286  * @return: 0 for use by caller, to make notation easy, like:
 287  *      return error_response(..).
 288  */
 289 static int
 290 error_response_cache(struct module_qstate* qstate, int id, int rcode)
 291 {
 292         if(!qstate->no_cache_store) {
 293                 /* store in cache */
 294                 struct reply_info err;
 295                 if(qstate->prefetch_leeway > NORR_TTL) {
 296                         verbose(VERB_ALGO, "error response for prefetch in cache");
 297                         /* attempt to adjust the cache entry prefetch */
 298                         if(dns_cache_prefetch_adjust(qstate->env, &qstate->qinfo,
 299                                 NORR_TTL, qstate->query_flags))
 300                                 return error_response(qstate, id, rcode);
 301                         /* if that fails (not in cache), fall through to store err */
 302                 }
 303                 if(qstate->env->cfg->serve_expired) {
 304                         /* if serving expired contents, and such content is
 305                          * already available, don't overwrite this servfail */
 306                         struct msgreply_entry* msg;
 307                         if((msg=msg_cache_lookup(qstate->env,
 308                                 qstate->qinfo.qname, qstate->qinfo.qname_len,
 309                                 qstate->qinfo.qtype, qstate->qinfo.qclass,
 310                                 qstate->query_flags, 0,
 311                                 qstate->env->cfg->serve_expired_ttl_reset))
 312                                 != NULL) {
 313                                 if(qstate->env->cfg->serve_expired_ttl_reset) {
 314                                         struct reply_info* rep =
 315                                                 (struct reply_info*)msg->entry.data;
 316                                         if(rep && *qstate->env->now +
 317                                                 qstate->env->cfg->serve_expired_ttl  >
 318                                                 rep->serve_expired_ttl) {
 319                                                 rep->serve_expired_ttl =
 320                                                         *qstate->env->now +
 321                                                         qstate->env->cfg->serve_expired_ttl;
 322                                         }
 323                                 }
 324                                 lock_rw_unlock(&msg->entry.lock);
 325                                 return error_response(qstate, id, rcode);
 326                         }
 327                         /* serving expired contents, but nothing is cached
 328                          * at all, so the servfail cache entry is useful
 329                          * (stops waste of time on this servfail NORR_TTL) */
 330                 } else {
 331                         /* don't overwrite existing (non-expired) data in
 332                          * cache with a servfail */
 333                         struct msgreply_entry* msg;
 334                         if((msg=msg_cache_lookup(qstate->env,
 335                                 qstate->qinfo.qname, qstate->qinfo.qname_len,
 336                                 qstate->qinfo.qtype, qstate->qinfo.qclass,
 337                                 qstate->query_flags, *qstate->env->now, 0))
 338                                 != NULL) {
 339                                 struct reply_info* rep = (struct reply_info*)
 340                                         msg->entry.data;
 341                                 if(FLAGS_GET_RCODE(rep->flags) ==
 342                                         LDNS_RCODE_NOERROR ||
 343                                         FLAGS_GET_RCODE(rep->flags) ==
 344                                         LDNS_RCODE_NXDOMAIN) {
 345                                         /* we have a good entry,
 346                                          * don't overwrite */
 347                                         lock_rw_unlock(&msg->entry.lock);
 348                                         return error_response(qstate, id, rcode);
 349                                 }
 350                                 lock_rw_unlock(&msg->entry.lock);
 351                         }
 352
 353                 }
 354                 memset(&err, 0, sizeof(err));
 355                 err.flags = (uint16_t)(BIT_QR | BIT_RA);
 356                 FLAGS_SET_RCODE(err.flags, rcode);
 357                 err.qdcount = 1;
 358                 err.ttl = NORR_TTL;
 359                 err.prefetch_ttl = PREFETCH_TTL_CALC(err.ttl);
 360                 err.serve_expired_ttl = NORR_TTL;
 361                 /* do not waste time trying to validate this servfail */
 362                 err.security = sec_status_indeterminate;
 363                 verbose(VERB_ALGO, "store error response in message cache");
 364                 iter_dns_store(qstate->env, &qstate->qinfo, &err, 0, 0, 0, NULL,
 365                         qstate->query_flags);
 366         }
 367         return error_response(qstate, id, rcode);
 368 }
 369
 370 /** check if prepend item is duplicate item */
 371 static int
 372 prepend_is_duplicate(struct ub_packed_rrset_key** sets, size_t to,
 373         struct ub_packed_rrset_key* dup)
 374 {
 375         size_t i;
 376         for(i=0; i<to; i++) {
 377                 if(sets[i]->rk.type == dup->rk.type &&
 378                         sets[i]->rk.rrset_class == dup->rk.rrset_class &&
 379                         sets[i]->rk.dname_len == dup->rk.dname_len &&
 380                         query_dname_compare(sets[i]->rk.dname, dup->rk.dname)
 381                         == 0)
 382                         return 1;
 383         }
 384         return 0;
 385 }
 386
 387 /** prepend the prepend list in the answer and authority section of dns_msg */
 388 static int
 389 iter_prepend(struct iter_qstate* iq, struct dns_msg* msg,
 390         struct regional* region)
 391 {
 392         struct iter_prep_list* p;
 393         struct ub_packed_rrset_key** sets;
 394         size_t num_an = 0, num_ns = 0;;
 395         for(p = iq->an_prepend_list; p; p = p->next)
 396                 num_an++;
 397         for(p = iq->ns_prepend_list; p; p = p->next)
 398                 num_ns++;
 399         if(num_an + num_ns == 0)
 400                 return 1;
 401         verbose(VERB_ALGO, "prepending %d rrsets", (int)num_an + (int)num_ns);
 402         if(num_an > RR_COUNT_MAX || num_ns > RR_COUNT_MAX ||
 403                 msg->rep->rrset_count > RR_COUNT_MAX) return 0; /* overflow */
 404         sets = regional_alloc(region, (num_an+num_ns+msg->rep->rrset_count) *
 405                 sizeof(struct ub_packed_rrset_key*));
 406         if(!sets)
 407                 return 0;
 408         /* ANSWER section */
 409         num_an = 0;
 410         for(p = iq->an_prepend_list; p; p = p->next) {
 411                 sets[num_an++] = p->rrset;
 412                 if(ub_packed_rrset_ttl(p->rrset) < msg->rep->ttl)
 413                         msg->rep->ttl = ub_packed_rrset_ttl(p->rrset);
 414         }
 415         memcpy(sets+num_an, msg->rep->rrsets, msg->rep->an_numrrsets *
 416                 sizeof(struct ub_packed_rrset_key*));
 417         /* AUTH section */
 418         num_ns = 0;
 419         for(p = iq->ns_prepend_list; p; p = p->next) {
 420                 if(prepend_is_duplicate(sets+msg->rep->an_numrrsets+num_an,
 421                         num_ns, p->rrset) || prepend_is_duplicate(
 422                         msg->rep->rrsets+msg->rep->an_numrrsets,
 423                         msg->rep->ns_numrrsets, p->rrset))
 424                         continue;
 425                 sets[msg->rep->an_numrrsets + num_an + num_ns++] = p->rrset;
 426                 if(ub_packed_rrset_ttl(p->rrset) < msg->rep->ttl)
 427                         msg->rep->ttl = ub_packed_rrset_ttl(p->rrset);
 428         }
 429         memcpy(sets + num_an + msg->rep->an_numrrsets + num_ns,
 430                 msg->rep->rrsets + msg->rep->an_numrrsets,
 431                 (msg->rep->ns_numrrsets + msg->rep->ar_numrrsets) *
 432                 sizeof(struct ub_packed_rrset_key*));
 433
 434         /* NXDOMAIN rcode can stay if we prepended DNAME/CNAMEs, because
 435          * this is what recursors should give. */
 436         msg->rep->rrset_count += num_an + num_ns;
 437         msg->rep->an_numrrsets += num_an;
 438         msg->rep->ns_numrrsets += num_ns;
 439         msg->rep->rrsets = sets;
 440         return 1;
 441 }
 442
 443 /**
 444  * Find rrset in ANSWER prepend list.
 445  * to avoid duplicate DNAMEs when a DNAME is traversed twice.
 446  * @param iq: iterator query state.
 447  * @param rrset: rrset to add.
 448  * @return false if not found
 449  */
 450 static int
 451 iter_find_rrset_in_prepend_answer(struct iter_qstate* iq,
 452         struct ub_packed_rrset_key* rrset)
 453 {
 454         struct iter_prep_list* p = iq->an_prepend_list;
 455         while(p) {
 456                 if(ub_rrset_compare(p->rrset, rrset) == 0 &&
 457                         rrsetdata_equal((struct packed_rrset_data*)p->rrset
 458                         ->entry.data, (struct packed_rrset_data*)rrset
 459                         ->entry.data))
 460                         return 1;
 461                 p = p->next;
 462         }
 463         return 0;
 464 }
 465
 466 /**
 467  * Add rrset to ANSWER prepend list
 468  * @param qstate: query state.
 469  * @param iq: iterator query state.
 470  * @param rrset: rrset to add.
 471  * @return false on failure (malloc).
 472  */
 473 static int
 474 iter_add_prepend_answer(struct module_qstate* qstate, struct iter_qstate* iq,
 475         struct ub_packed_rrset_key* rrset)
 476 {
 477         struct iter_prep_list* p = (struct iter_prep_list*)regional_alloc(
 478                 qstate->region, sizeof(struct iter_prep_list));
 479         if(!p)
 480                 return 0;
 481         p->rrset = rrset;
 482         p->next = NULL;
 483         /* add at end */
 484         if(iq->an_prepend_last)
 485                 iq->an_prepend_last->next = p;
 486         else    iq->an_prepend_list = p;
 487         iq->an_prepend_last = p;
 488         return 1;
 489 }
 490
 491 /**
 492  * Add rrset to AUTHORITY prepend list
 493  * @param qstate: query state.
 494  * @param iq: iterator query state.
 495  * @param rrset: rrset to add.
 496  * @return false on failure (malloc).
 497  */
 498 static int
 499 iter_add_prepend_auth(struct module_qstate* qstate, struct iter_qstate* iq,
 500         struct ub_packed_rrset_key* rrset)
 501 {
 502         struct iter_prep_list* p = (struct iter_prep_list*)regional_alloc(
 503                 qstate->region, sizeof(struct iter_prep_list));
 504         if(!p)
 505                 return 0;
 506         p->rrset = rrset;
 507         p->next = NULL;
 508         /* add at end */
 509         if(iq->ns_prepend_last)
 510                 iq->ns_prepend_last->next = p;
 511         else    iq->ns_prepend_list = p;
 512         iq->ns_prepend_last = p;
 513         return 1;
 514 }
 515
 516 /**
 517  * Given a CNAME response (defined as a response containing a CNAME or DNAME
 518  * that does not answer the request), process the response, modifying the
 519  * state as necessary. This follows the CNAME/DNAME chain and returns the
 520  * final query name.
 521  *
 522  * sets the new query name, after following the CNAME/DNAME chain.
 523  * @param qstate: query state.
 524  * @param iq: iterator query state.
 525  * @param msg: the response.
 526  * @param mname: returned target new query name.
 527  * @param mname_len: length of mname.
 528  * @return false on (malloc) error.
 529  */
 530 static int
 531 handle_cname_response(struct module_qstate* qstate, struct iter_qstate* iq,
 532         struct dns_msg* msg, uint8_t** mname, size_t* mname_len)
 533 {
 534         size_t i;
 535         /* Start with the (current) qname. */
 536         *mname = iq->qchase.qname;
 537         *mname_len = iq->qchase.qname_len;
 538
 539         /* Iterate over the ANSWER rrsets in order, looking for CNAMEs and
 540          * DNAMES. */
 541         for(i=0; i<msg->rep->an_numrrsets; i++) {
 542                 struct ub_packed_rrset_key* r = msg->rep->rrsets[i];
 543                 /* If there is a (relevant) DNAME, add it to the list.
 544                  * We always expect there to be CNAME that was generated
 545                  * by this DNAME following, so we don't process the DNAME
 546                  * directly.  */
 547                 if(ntohs(r->rk.type) == LDNS_RR_TYPE_DNAME &&
 548                         dname_strict_subdomain_c(*mname, r->rk.dname) &&
 549                         !iter_find_rrset_in_prepend_answer(iq, r)) {
 550                         if(!iter_add_prepend_answer(qstate, iq, r))
 551                                 return 0;
 552                         continue;
 553                 }
 554
 555                 if(ntohs(r->rk.type) == LDNS_RR_TYPE_CNAME &&
 556                         query_dname_compare(*mname, r->rk.dname) == 0 &&
 557                         !iter_find_rrset_in_prepend_answer(iq, r)) {
 558                         /* Add this relevant CNAME rrset to the prepend list.*/
 559                         if(!iter_add_prepend_answer(qstate, iq, r))
 560                                 return 0;
 561                         get_cname_target(r, mname, mname_len);
 562                 }
 563
 564                 /* Other rrsets in the section are ignored. */
 565         }
 566         /* add authority rrsets to authority prepend, for wildcarded CNAMEs */
 567         for(i=msg->rep->an_numrrsets; i<msg->rep->an_numrrsets +
 568                 msg->rep->ns_numrrsets; i++) {
 569                 struct ub_packed_rrset_key* r = msg->rep->rrsets[i];
 570                 /* only add NSEC/NSEC3, as they may be needed for validation */
 571                 if(ntohs(r->rk.type) == LDNS_RR_TYPE_NSEC ||
 572                         ntohs(r->rk.type) == LDNS_RR_TYPE_NSEC3) {
 573                         if(!iter_add_prepend_auth(qstate, iq, r))
 574                                 return 0;
 575                 }
 576         }
 577         return 1;
 578 }
 579
 580 /** see if last resort is possible - does config allow queries to parent */
 581 static int
 582 can_have_last_resort(struct module_env* env, uint8_t* nm, size_t nmlen,
 583         uint16_t qclass, struct delegpt** retdp)
 584 {
 585         struct delegpt* fwddp;
 586         struct iter_hints_stub* stub;
 587         int labs = dname_count_labels(nm);
 588         /* do not process a last resort (the parent side) if a stub
 589          * or forward is configured, because we do not want to go 'above'
 590          * the configured servers */
 591         if(!dname_is_root(nm) && (stub = (struct iter_hints_stub*)
 592                 name_tree_find(&env->hints->tree, nm, nmlen, labs, qclass)) &&
 593                 /* has_parent side is turned off for stub_first, where we
 594                  * are allowed to go to the parent */
 595                 stub->dp->has_parent_side_NS) {
 596                 if(retdp) *retdp = stub->dp;
 597                 return 0;
 598         }
 599         if((fwddp = forwards_find(env->fwds, nm, qclass)) &&
 600                 /* has_parent_side is turned off for forward_first, where
 601                  * we are allowed to go to the parent */
 602                 fwddp->has_parent_side_NS) {
 603                 if(retdp) *retdp = fwddp;
 604                 return 0;
 605         }
 606         return 1;
 607 }
 608
 609 /** see if target name is caps-for-id whitelisted */
 610 static int
 611 is_caps_whitelisted(struct iter_env* ie, struct iter_qstate* iq)
 612 {
 613         if(!ie->caps_white) return 0; /* no whitelist, or no capsforid */
 614         return name_tree_lookup(ie->caps_white, iq->qchase.qname,
 615                 iq->qchase.qname_len, dname_count_labels(iq->qchase.qname),
 616                 iq->qchase.qclass) != NULL;
 617 }
 618
 619 /** create target count structure for this query */
 620 static void
 621 target_count_create(struct iter_qstate* iq)
 622 {
 623         if(!iq->target_count) {
 624                 iq->target_count = (int*)calloc(2, sizeof(int));
 625                 /* if calloc fails we simply do not track this number */
 626                 if(iq->target_count)
 627                         iq->target_count[0] = 1;
 628         }
 629 }
 630
 631 static void
 632 target_count_increase(struct iter_qstate* iq, int num)
 633 {
 634         target_count_create(iq);
 635         if(iq->target_count)
 636                 iq->target_count[1] += num;
 637 }
 638
 639 /**
 640  * Generate a subrequest.
 641  * Generate a local request event. Local events are tied to this module, and
 642  * have a corresponding (first tier) event that is waiting for this event to
 643  * resolve to continue.
 644  *
 645  * @param qname The query name for this request.
 646  * @param qnamelen length of qname
 647  * @param qtype The query type for this request.
 648  * @param qclass The query class for this request.
 649  * @param qstate The event that is generating this event.
 650  * @param id: module id.
 651  * @param iq: The iterator state that is generating this event.
 652  * @param initial_state The initial response state (normally this
 653  *          is QUERY_RESP_STATE, unless it is known that the request won't
 654  *          need iterative processing
 655  * @param finalstate The final state for the response to this request.
 656  * @param subq_ret: if newly allocated, the subquerystate, or NULL if it does
 657  *      not need initialisation.
 658  * @param v: if true, validation is done on the subquery.
 659  * @return false on error (malloc).
 660  */
 661 static int
 662 generate_sub_request(uint8_t* qname, size_t qnamelen, uint16_t qtype,
 663         uint16_t qclass, struct module_qstate* qstate, int id,
 664         struct iter_qstate* iq, enum iter_state initial_state,
 665         enum iter_state finalstate, struct module_qstate** subq_ret, int v)
 666 {
 667         struct module_qstate* subq = NULL;
 668         struct iter_qstate* subiq = NULL;
 669         uint16_t qflags = 0; /* OPCODE QUERY, no flags */
 670         struct query_info qinf;
 671         int prime = (finalstate == PRIME_RESP_STATE)?1:0;
 672         int valrec = 0;
 673         qinf.qname = qname;
 674         qinf.qname_len = qnamelen;
 675         qinf.qtype = qtype;
 676         qinf.qclass = qclass;
 677         qinf.local_alias = NULL;
 678
 679         /* RD should be set only when sending the query back through the INIT
 680          * state. */
 681         if(initial_state == INIT_REQUEST_STATE)
 682                 qflags |= BIT_RD;
 683         /* We set the CD flag so we can send this through the "head" of
 684          * the resolution chain, which might have a validator. We are
 685          * uninterested in validating things not on the direct resolution
 686          * path.  */
 687         if(!v) {
 688                 qflags |= BIT_CD;
 689                 valrec = 1;
 690         }
 691
 692         /* attach subquery, lookup existing or make a new one */
 693         fptr_ok(fptr_whitelist_modenv_attach_sub(qstate->env->attach_sub));
 694         if(!(*qstate->env->attach_sub)(qstate, &qinf, qflags, prime, valrec,
 695                 &subq)) {
 696                 return 0;
 697         }
 698         *subq_ret = subq;
 699         if(subq) {
 700                 /* initialise the new subquery */
 701                 subq->curmod = id;
 702                 subq->ext_state[id] = module_state_initial;
 703                 subq->minfo[id] = regional_alloc(subq->region,
 704                         sizeof(struct iter_qstate));
 705                 if(!subq->minfo[id]) {
 706                         log_err("init subq: out of memory");
 707                         fptr_ok(fptr_whitelist_modenv_kill_sub(
 708                                 qstate->env->kill_sub));
 709                         (*qstate->env->kill_sub)(subq);
 710                         return 0;
 711                 }
 712                 subiq = (struct iter_qstate*)subq->minfo[id];
 713                 memset(subiq, 0, sizeof(*subiq));
 714                 subiq->num_target_queries = 0;
 715                 target_count_create(iq);
 716                 subiq->target_count = iq->target_count;
 717                 if(iq->target_count)
 718                         iq->target_count[0] ++; /* extra reference */
 719                 subiq->num_current_queries = 0;
 720                 subiq->depth = iq->depth+1;
 721                 outbound_list_init(&subiq->outlist);
 722                 subiq->state = initial_state;
 723                 subiq->final_state = finalstate;
 724                 subiq->qchase = subq->qinfo;
 725                 subiq->chase_flags = subq->query_flags;
 726                 subiq->refetch_glue = 0;
 727                 if(qstate->env->cfg->qname_minimisation)
 728                         subiq->minimisation_state = INIT_MINIMISE_STATE;
 729                 else
 730                         subiq->minimisation_state = DONOT_MINIMISE_STATE;
 731                 memset(&subiq->qinfo_out, 0, sizeof(struct query_info));
 732         }
 733         return 1;
 734 }
 735
 736 /**
 737  * Generate and send a root priming request.
 738  * @param qstate: the qtstate that triggered the need to prime.
 739  * @param iq: iterator query state.
 740  * @param id: module id.
 741  * @param qclass: the class to prime.
 742  * @return 0 on failure
 743  */
 744 static int
 745 prime_root(struct module_qstate* qstate, struct iter_qstate* iq, int id,
 746         uint16_t qclass)
 747 {
 748         struct delegpt* dp;
 749         struct module_qstate* subq;
 750         verbose(VERB_DETAIL, "priming . %s NS",
 751                 sldns_lookup_by_id(sldns_rr_classes, (int)qclass)?
 752                 sldns_lookup_by_id(sldns_rr_classes, (int)qclass)->name:"??");
 753         dp = hints_lookup_root(qstate->env->hints, qclass);
 754         if(!dp) {
 755                 verbose(VERB_ALGO, "Cannot prime due to lack of hints");
 756                 return 0;
 757         }
 758         /* Priming requests start at the QUERYTARGETS state, skipping
 759          * the normal INIT state logic (which would cause an infloop). */
 760         if(!generate_sub_request((uint8_t*)"\000", 1, LDNS_RR_TYPE_NS,
 761                 qclass, qstate, id, iq, QUERYTARGETS_STATE, PRIME_RESP_STATE,
 762                 &subq, 0)) {
 763                 verbose(VERB_ALGO, "could not prime root");
 764                 return 0;
 765         }
 766         if(subq) {
 767                 struct iter_qstate* subiq =
 768                         (struct iter_qstate*)subq->minfo[id];
 769                 /* Set the initial delegation point to the hint.
 770                  * copy dp, it is now part of the root prime query.
 771                  * dp was part of in the fixed hints structure. */
 772                 subiq->dp = delegpt_copy(dp, subq->region);
 773                 if(!subiq->dp) {
 774                         log_err("out of memory priming root, copydp");
 775                         fptr_ok(fptr_whitelist_modenv_kill_sub(
 776                                 qstate->env->kill_sub));
 777                         (*qstate->env->kill_sub)(subq);
 778                         return 0;
 779                 }
 780                 /* there should not be any target queries. */
 781                 subiq->num_target_queries = 0;
 782                 subiq->dnssec_expected = iter_indicates_dnssec(
 783                         qstate->env, subiq->dp, NULL, subq->qinfo.qclass);
 784         }
 785
 786         /* this module stops, our submodule starts, and does the query. */
 787         qstate->ext_state[id] = module_wait_subquery;
 788         return 1;
 789 }
 790
 791 /**
 792  * Generate and process a stub priming request. This method tests for the
 793  * need to prime a stub zone, so it is safe to call for every request.
 794  *
 795  * @param qstate: the qtstate that triggered the need to prime.
 796  * @param iq: iterator query state.
 797  * @param id: module id.
 798  * @param qname: request name.
 799  * @param qclass: request class.
 800  * @return true if a priming subrequest was made, false if not. The will only
 801  *         issue a priming request if it detects an unprimed stub.
 802  *         Uses value of 2 to signal during stub-prime in root-prime situation
 803  *         that a noprime-stub is available and resolution can continue.
 804  */
 805 static int
 806 prime_stub(struct module_qstate* qstate, struct iter_qstate* iq, int id,
 807         uint8_t* qname, uint16_t qclass)
 808 {
 809         /* Lookup the stub hint. This will return null if the stub doesn't
 810          * need to be re-primed. */
 811         struct iter_hints_stub* stub;
 812         struct delegpt* stub_dp;
 813         struct module_qstate* subq;
 814
 815         if(!qname) return 0;
 816         stub = hints_lookup_stub(qstate->env->hints, qname, qclass, iq->dp);
 817         /* The stub (if there is one) does not need priming. */
 818         if(!stub)
 819                 return 0;
 820         stub_dp = stub->dp;
 821         /* if we have an auth_zone dp, and stub is equal, don't prime stub
 822          * yet, unless we want to fallback and avoid the auth_zone */
 823         if(!iq->auth_zone_avoid && iq->dp && iq->dp->auth_dp &&
 824                 query_dname_compare(iq->dp->name, stub_dp->name) == 0)
 825                 return 0;
 826
 827         /* is it a noprime stub (always use) */
 828         if(stub->noprime) {
 829                 int r = 0;
 830                 if(iq->dp == NULL) r = 2;
 831                 /* copy the dp out of the fixed hints structure, so that
 832                  * it can be changed when servicing this query */
 833                 iq->dp = delegpt_copy(stub_dp, qstate->region);
 834                 if(!iq->dp) {
 835                         log_err("out of memory priming stub");
 836                         errinf(qstate, "malloc failure, priming stub");
 837                         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 838                         return 1; /* return 1 to make module stop, with error */
 839                 }
 840                 log_nametypeclass(VERB_DETAIL, "use stub", stub_dp->name,
 841                         LDNS_RR_TYPE_NS, qclass);
 842                 return r;
 843         }
 844
 845         /* Otherwise, we need to (re)prime the stub. */
 846         log_nametypeclass(VERB_DETAIL, "priming stub", stub_dp->name,
 847                 LDNS_RR_TYPE_NS, qclass);
 848
 849         /* Stub priming events start at the QUERYTARGETS state to avoid the
 850          * redundant INIT state processing. */
 851         if(!generate_sub_request(stub_dp->name, stub_dp->namelen,
 852                 LDNS_RR_TYPE_NS, qclass, qstate, id, iq,
 853                 QUERYTARGETS_STATE, PRIME_RESP_STATE, &subq, 0)) {
 854                 verbose(VERB_ALGO, "could not prime stub");
 855                 errinf(qstate, "could not generate lookup for stub prime");
 856                 (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 857                 return 1; /* return 1 to make module stop, with error */
 858         }
 859         if(subq) {
 860                 struct iter_qstate* subiq =
 861                         (struct iter_qstate*)subq->minfo[id];
 862
 863                 /* Set the initial delegation point to the hint. */
 864                 /* make copy to avoid use of stub dp by different qs/threads */
 865                 subiq->dp = delegpt_copy(stub_dp, subq->region);
 866                 if(!subiq->dp) {
 867                         log_err("out of memory priming stub, copydp");
 868                         fptr_ok(fptr_whitelist_modenv_kill_sub(
 869                                 qstate->env->kill_sub));
 870                         (*qstate->env->kill_sub)(subq);
 871                         errinf(qstate, "malloc failure, in stub prime");
 872                         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
 873                         return 1; /* return 1 to make module stop, with error */
 874                 }
 875                 /* there should not be any target queries -- although there
 876                  * wouldn't be anyway, since stub hints never have
 877                  * missing targets. */
 878                 subiq->num_target_queries = 0;
 879                 subiq->wait_priming_stub = 1;
 880                 subiq->dnssec_expected = iter_indicates_dnssec(
 881                         qstate->env, subiq->dp, NULL, subq->qinfo.qclass);
 882         }
 883
 884         /* this module stops, our submodule starts, and does the query. */
 885         qstate->ext_state[id] = module_wait_subquery;
 886         return 1;
 887 }
 888
 889 /**
 890  * Generate a delegation point for an auth zone (unless cached dp is better)
 891  * false on alloc failure.
 892  */
 893 static int
 894 auth_zone_delegpt(struct module_qstate* qstate, struct iter_qstate* iq,
 895         uint8_t* delname, size_t delnamelen)
 896 {
 897         struct auth_zone* z;
 898         if(iq->auth_zone_avoid)
 899                 return 1;
 900         if(!delname) {
 901                 delname = iq->qchase.qname;
 902                 delnamelen = iq->qchase.qname_len;
 903         }
 904         lock_rw_rdlock(&qstate->env->auth_zones->lock);
 905         z = auth_zones_find_zone(qstate->env->auth_zones, delname, delnamelen,
 906                 qstate->qinfo.qclass);
 907         if(!z) {
 908                 lock_rw_unlock(&qstate->env->auth_zones->lock);
 909                 return 1;
 910         }
 911         lock_rw_rdlock(&z->lock);
 912         lock_rw_unlock(&qstate->env->auth_zones->lock);
 913         if(z->for_upstream) {
 914                 if(iq->dp && query_dname_compare(z->name, iq->dp->name) == 0
 915                         && iq->dp->auth_dp && qstate->blacklist &&
 916                         z->fallback_enabled) {
 917                         /* cache is blacklisted and fallback, and we
 918                          * already have an auth_zone dp */
 919                         if(verbosity>=VERB_ALGO) {
 920                                 char buf[255+1];
 921                                 dname_str(z->name, buf);
 922                                 verbose(VERB_ALGO, "auth_zone %s "
 923                                   "fallback because cache blacklisted",
 924                                   buf);
 925                         }
 926                         lock_rw_unlock(&z->lock);
 927                         iq->dp = NULL;
 928                         return 1;
 929                 }
 930                 if(iq->dp==NULL || dname_subdomain_c(z->name, iq->dp->name)) {
 931                         struct delegpt* dp;
 932                         if(qstate->blacklist && z->fallback_enabled) {
 933                                 /* cache is blacklisted because of a DNSSEC
 934                                  * validation failure, and the zone allows
 935                                  * fallback to the internet, query there. */
 936                                 if(verbosity>=VERB_ALGO) {
 937                                         char buf[255+1];
 938                                         dname_str(z->name, buf);
 939                                         verbose(VERB_ALGO, "auth_zone %s "
 940                                           "fallback because cache blacklisted",
 941                                           buf);
 942                                 }
 943                                 lock_rw_unlock(&z->lock);
 944                                 return 1;
 945                         }
 946                         dp = (struct delegpt*)regional_alloc_zero(
 947                                 qstate->region, sizeof(*dp));
 948                         if(!dp) {
 949                                 log_err("alloc failure");
 950                                 if(z->fallback_enabled) {
 951                                         lock_rw_unlock(&z->lock);
 952                                         return 1; /* just fallback */
 953                                 }
 954                                 lock_rw_unlock(&z->lock);
 955                                 errinf(qstate, "malloc failure");
 956                                 return 0;
 957                         }
 958                         dp->name = regional_alloc_init(qstate->region,
 959                                 z->name, z->namelen);
 960                         if(!dp->name) {
 961                                 log_err("alloc failure");
 962                                 if(z->fallback_enabled) {
 963                                         lock_rw_unlock(&z->lock);
 964                                         return 1; /* just fallback */
 965                                 }
 966                                 lock_rw_unlock(&z->lock);
 967                                 errinf(qstate, "malloc failure");
 968                                 return 0;
 969                         }
 970                         dp->namelen = z->namelen;
 971                         dp->namelabs = z->namelabs;
 972                         dp->auth_dp = 1;
 973                         iq->dp = dp;
 974                 }
 975         }
 976
 977         lock_rw_unlock(&z->lock);
 978         return 1;
 979 }
 980
 981 /**
 982  * Generate A and AAAA checks for glue that is in-zone for the referral
 983  * we just got to obtain authoritative information on the addresses.
 984  *
 985  * @param qstate: the qtstate that triggered the need to prime.
 986  * @param iq: iterator query state.
 987  * @param id: module id.
 988  */
 989 static void
 990 generate_a_aaaa_check(struct module_qstate* qstate, struct iter_qstate* iq,
 991         int id)
 992 {
 993         struct iter_env* ie = (struct iter_env*)qstate->env->modinfo[id];
 994         struct module_qstate* subq;
 995         size_t i;
 996         struct reply_info* rep = iq->response->rep;
 997         struct ub_packed_rrset_key* s;
 998         log_assert(iq->dp);
 999
1000         if(iq->depth == ie->max_dependency_depth)
1001                 return;
1002         /* walk through additional, and check if in-zone,
1003          * only relevant A, AAAA are left after scrub anyway */
1004         for(i=rep->an_numrrsets+rep->ns_numrrsets; i<rep->rrset_count; i++) {
1005                 s = rep->rrsets[i];
1006                 /* check *ALL* addresses that are transmitted in additional*/
1007                 /* is it an address ? */
1008                 if( !(ntohs(s->rk.type)==LDNS_RR_TYPE_A ||
1009                         ntohs(s->rk.type)==LDNS_RR_TYPE_AAAA)) {
1010                         continue;
1011                 }
1012                 /* is this query the same as the A/AAAA check for it */
1013                 if(qstate->qinfo.qtype == ntohs(s->rk.type) &&
1014                         qstate->qinfo.qclass == ntohs(s->rk.rrset_class) &&
1015                         query_dname_compare(qstate->qinfo.qname,
1016                                 s->rk.dname)==0 &&
1017                         (qstate->query_flags&BIT_RD) &&
1018                         !(qstate->query_flags&BIT_CD))
1019                         continue;
1020
1021                 /* generate subrequest for it */
1022                 log_nametypeclass(VERB_ALGO, "schedule addr fetch",
1023                         s->rk.dname, ntohs(s->rk.type),
1024                         ntohs(s->rk.rrset_class));
1025                 if(!generate_sub_request(s->rk.dname, s->rk.dname_len,
1026                         ntohs(s->rk.type), ntohs(s->rk.rrset_class),
1027                         qstate, id, iq,
1028                         INIT_REQUEST_STATE, FINISHED_STATE, &subq, 1)) {
1029                         verbose(VERB_ALGO, "could not generate addr check");
1030                         return;
1031                 }
1032                 /* ignore subq - not need for more init */
1033         }
1034 }
1035
1036 /**
1037  * Generate a NS check request to obtain authoritative information
1038  * on an NS rrset.
1039  *
1040  * @param qstate: the qtstate that triggered the need to prime.
1041  * @param iq: iterator query state.
1042  * @param id: module id.
1043  */
1044 static void
1045 generate_ns_check(struct module_qstate* qstate, struct iter_qstate* iq, int id)
1046 {
1047         struct iter_env* ie = (struct iter_env*)qstate->env->modinfo[id];
1048         struct module_qstate* subq;
1049         log_assert(iq->dp);
1050
1051         if(iq->depth == ie->max_dependency_depth)
1052                 return;
1053         if(!can_have_last_resort(qstate->env, iq->dp->name, iq->dp->namelen,
1054                 iq->qchase.qclass, NULL))
1055                 return;
1056         /* is this query the same as the nscheck? */
1057         if(qstate->qinfo.qtype == LDNS_RR_TYPE_NS &&
1058                 query_dname_compare(iq->dp->name, qstate->qinfo.qname)==0 &&
1059                 (qstate->query_flags&BIT_RD) && !(qstate->query_flags&BIT_CD)){
1060                 /* spawn off A, AAAA queries for in-zone glue to check */
1061                 generate_a_aaaa_check(qstate, iq, id);
1062                 return;
1063         }
1064         /* no need to get the NS record for DS, it is above the zonecut */
1065         if(qstate->qinfo.qtype == LDNS_RR_TYPE_DS)
1066                 return;
1067
1068         log_nametypeclass(VERB_ALGO, "schedule ns fetch",
1069                 iq->dp->name, LDNS_RR_TYPE_NS, iq->qchase.qclass);
1070         if(!generate_sub_request(iq->dp->name, iq->dp->namelen,
1071                 LDNS_RR_TYPE_NS, iq->qchase.qclass, qstate, id, iq,
1072                 INIT_REQUEST_STATE, FINISHED_STATE, &subq, 1)) {
1073                 verbose(VERB_ALGO, "could not generate ns check");
1074                 return;
1075         }
1076         if(subq) {
1077                 struct iter_qstate* subiq =
1078                         (struct iter_qstate*)subq->minfo[id];
1079
1080                 /* make copy to avoid use of stub dp by different qs/threads */
1081                 /* refetch glue to start higher up the tree */
1082                 subiq->refetch_glue = 1;
1083                 subiq->dp = delegpt_copy(iq->dp, subq->region);
1084                 if(!subiq->dp) {
1085                         log_err("out of memory generating ns check, copydp");
1086                         fptr_ok(fptr_whitelist_modenv_kill_sub(
1087                                 qstate->env->kill_sub));
1088                         (*qstate->env->kill_sub)(subq);
1089                         return;
1090                 }
1091         }
1092 }
1093
1094 /**
1095  * Generate a DNSKEY prefetch query to get the DNSKEY for the DS record we
1096  * just got in a referral (where we have dnssec_expected, thus have trust
1097  * anchors above it).  Note that right after calling this routine the
1098  * iterator detached subqueries (because of following the referral), and thus
1099  * the DNSKEY query becomes detached, its return stored in the cache for
1100  * later lookup by the validator.  This cache lookup by the validator avoids
1101  * the roundtrip incurred by the DNSKEY query.  The DNSKEY query is now
1102  * performed at about the same time the original query is sent to the domain,
1103  * thus the two answers are likely to be returned at about the same time,
1104  * saving a roundtrip from the validated lookup.
1105  *
1106  * @param qstate: the qtstate that triggered the need to prime.
1107  * @param iq: iterator query state.
1108  * @param id: module id.
1109  */
1110 static void
1111 generate_dnskey_prefetch(struct module_qstate* qstate,
1112         struct iter_qstate* iq, int id)
1113 {
1114         struct module_qstate* subq;
1115         log_assert(iq->dp);
1116
1117         /* is this query the same as the prefetch? */
1118         if(qstate->qinfo.qtype == LDNS_RR_TYPE_DNSKEY &&
1119                 query_dname_compare(iq->dp->name, qstate->qinfo.qname)==0 &&
1120                 (qstate->query_flags&BIT_RD) && !(qstate->query_flags&BIT_CD)){
1121                 return;
1122         }
1123
1124         /* if the DNSKEY is in the cache this lookup will stop quickly */
1125         log_nametypeclass(VERB_ALGO, "schedule dnskey prefetch",
1126                 iq->dp->name, LDNS_RR_TYPE_DNSKEY, iq->qchase.qclass);
1127         if(!generate_sub_request(iq->dp->name, iq->dp->namelen,
1128                 LDNS_RR_TYPE_DNSKEY, iq->qchase.qclass, qstate, id, iq,
1129                 INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0)) {
1130                 /* we'll be slower, but it'll work */
1131                 verbose(VERB_ALGO, "could not generate dnskey prefetch");
1132                 return;
1133         }
1134         if(subq) {
1135                 struct iter_qstate* subiq =
1136                         (struct iter_qstate*)subq->minfo[id];
1137                 /* this qstate has the right delegation for the dnskey lookup*/
1138                 /* make copy to avoid use of stub dp by different qs/threads */
1139                 subiq->dp = delegpt_copy(iq->dp, subq->region);
1140                 /* if !subiq->dp, it'll start from the cache, no problem */
1141         }
1142 }
1143
1144 /**
1145  * See if the query needs forwarding.
1146  *
1147  * @param qstate: query state.
1148  * @param iq: iterator query state.
1149  * @return true if the request is forwarded, false if not.
1150  *      If returns true but, iq->dp is NULL then a malloc failure occurred.
1151  */
1152 static int
1153 forward_request(struct module_qstate* qstate, struct iter_qstate* iq)
1154 {
1155         struct delegpt* dp;
1156         uint8_t* delname = iq->qchase.qname;
1157         size_t delnamelen = iq->qchase.qname_len;
1158         if(iq->refetch_glue && iq->dp) {
1159                 delname = iq->dp->name;
1160                 delnamelen = iq->dp->namelen;
1161         }
1162         /* strip one label off of DS query to lookup higher for it */
1163         if( (iq->qchase.qtype == LDNS_RR_TYPE_DS || iq->refetch_glue)
1164                 && !dname_is_root(iq->qchase.qname))
1165                 dname_remove_label(&delname, &delnamelen);
1166         dp = forwards_lookup(qstate->env->fwds, delname, iq->qchase.qclass);
1167         if(!dp)
1168                 return 0;
1169         /* send recursion desired to forward addr */
1170         iq->chase_flags |= BIT_RD;
1171         iq->dp = delegpt_copy(dp, qstate->region);
1172         /* iq->dp checked by caller */
1173         verbose(VERB_ALGO, "forwarding request");
1174         return 1;
1175 }
1176
1177 /**
1178  * Process the initial part of the request handling. This state roughly
1179  * corresponds to resolver algorithms steps 1 (find answer in cache) and 2
1180  * (find the best servers to ask).
1181  *
1182  * Note that all requests start here, and query restarts revisit this state.
1183  *
1184  * This state either generates: 1) a response, from cache or error, 2) a
1185  * priming event, or 3) forwards the request to the next state (init2,
1186  * generally).
1187  *
1188  * @param qstate: query state.
1189  * @param iq: iterator query state.
1190  * @param ie: iterator shared global environment.
1191  * @param id: module id.
1192  * @return true if the event needs more request processing immediately,
1193  *         false if not.
1194  */
1195 static int
1196 processInitRequest(struct module_qstate* qstate, struct iter_qstate* iq,
1197         struct iter_env* ie, int id)
1198 {
1199         uint8_t* delname;
1200         size_t delnamelen;
1201         struct dns_msg* msg = NULL;
1202
1203         log_query_info(VERB_DETAIL, "resolving", &qstate->qinfo);
1204         /* check effort */
1205
1206         /* We enforce a maximum number of query restarts. This is primarily a
1207          * cheap way to prevent CNAME loops. */
1208         if(iq->query_restart_count > MAX_RESTART_COUNT) {
1209                 verbose(VERB_QUERY, "request has exceeded the maximum number"
1210                         " of query restarts with %d", iq->query_restart_count);
1211                 errinf(qstate, "request has exceeded the maximum number "
1212                         "restarts (eg. indirections)");
1213                 if(iq->qchase.qname)
1214                         errinf_dname(qstate, "stop at", iq->qchase.qname);
1215                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1216         }
1217
1218         /* We enforce a maximum recursion/dependency depth -- in general,
1219          * this is unnecessary for dependency loops (although it will
1220          * catch those), but it provides a sensible limit to the amount
1221          * of work required to answer a given query. */
1222         verbose(VERB_ALGO, "request has dependency depth of %d", iq->depth);
1223         if(iq->depth > ie->max_dependency_depth) {
1224                 verbose(VERB_QUERY, "request has exceeded the maximum "
1225                         "dependency depth with depth of %d", iq->depth);
1226                 errinf(qstate, "request has exceeded the maximum dependency "
1227                         "depth (eg. nameserver lookup recursion)");
1228                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1229         }
1230
1231         /* If the request is qclass=ANY, setup to generate each class */
1232         if(qstate->qinfo.qclass == LDNS_RR_CLASS_ANY) {
1233                 iq->qchase.qclass = 0;
1234                 return next_state(iq, COLLECT_CLASS_STATE);
1235         }
1236
1237         /*
1238          * If we are restricted by a forward-zone or a stub-zone, we
1239          * can't re-fetch glue for this delegation point.
1240          * we won’t try to re-fetch glue if the iq->dp is null.
1241          */
1242         if (iq->refetch_glue &&
1243                 iq->dp &&
1244                 !can_have_last_resort(qstate->env, iq->dp->name,
1245                      iq->dp->namelen, iq->qchase.qclass, NULL)) {
1246             iq->refetch_glue = 0;
1247         }
1248
1249         /* Resolver Algorithm Step 1 -- Look for the answer in local data. */
1250
1251         /* This either results in a query restart (CNAME cache response), a
1252          * terminating response (ANSWER), or a cache miss (null). */
1253
1254         if (iter_stub_fwd_no_cache(qstate, &iq->qchase)) {
1255                 /* Asked to not query cache. */
1256                 verbose(VERB_ALGO, "no-cache set, going to the network");
1257                 qstate->no_cache_lookup = 1;
1258                 qstate->no_cache_store = 1;
1259                 msg = NULL;
1260         } else if(qstate->blacklist) {
1261                 /* if cache, or anything else, was blacklisted then
1262                  * getting older results from cache is a bad idea, no cache */
1263                 verbose(VERB_ALGO, "cache blacklisted, going to the network");
1264                 msg = NULL;
1265         } else if(!qstate->no_cache_lookup) {
1266                 msg = dns_cache_lookup(qstate->env, iq->qchase.qname,
1267                         iq->qchase.qname_len, iq->qchase.qtype,
1268                         iq->qchase.qclass, qstate->query_flags,
1269                         qstate->region, qstate->env->scratch, 0);
1270                 if(!msg && qstate->env->neg_cache &&
1271                         iter_qname_indicates_dnssec(qstate->env, &iq->qchase)) {
1272                         /* lookup in negative cache; may result in
1273                          * NOERROR/NODATA or NXDOMAIN answers that need validation */
1274                         msg = val_neg_getmsg(qstate->env->neg_cache, &iq->qchase,
1275                                 qstate->region, qstate->env->rrset_cache,
1276                                 qstate->env->scratch_buffer,
1277                                 *qstate->env->now, 1/*add SOA*/, NULL,
1278                                 qstate->env->cfg);
1279                 }
1280                 /* item taken from cache does not match our query name, thus
1281                  * security needs to be re-examined later */
1282                 if(msg && query_dname_compare(qstate->qinfo.qname,
1283                         iq->qchase.qname) != 0)
1284                         msg->rep->security = sec_status_unchecked;
1285         }
1286         if(msg) {
1287                 /* handle positive cache response */
1288                 enum response_type type = response_type_from_cache(msg,
1289                         &iq->qchase);
1290                 if(verbosity >= VERB_ALGO) {
1291                         log_dns_msg("msg from cache lookup", &msg->qinfo,
1292                                 msg->rep);
1293                         verbose(VERB_ALGO, "msg ttl is %d, prefetch ttl %d",
1294                                 (int)msg->rep->ttl,
1295                                 (int)msg->rep->prefetch_ttl);
1296                 }
1297
1298                 if(type == RESPONSE_TYPE_CNAME) {
1299                         uint8_t* sname = 0;
1300                         size_t slen = 0;
1301                         verbose(VERB_ALGO, "returning CNAME response from "
1302                                 "cache");
1303                         if(!handle_cname_response(qstate, iq, msg,
1304                                 &sname, &slen)) {
1305                                 errinf(qstate, "failed to prepend CNAME "
1306                                         "components, malloc failure");
1307                                 return error_response(qstate, id,
1308                                         LDNS_RCODE_SERVFAIL);
1309                         }
1310                         iq->qchase.qname = sname;
1311                         iq->qchase.qname_len = slen;
1312                         /* This *is* a query restart, even if it is a cheap
1313                          * one. */
1314                         iq->dp = NULL;
1315                         iq->refetch_glue = 0;
1316                         iq->query_restart_count++;
1317                         iq->sent_count = 0;
1318                         sock_list_insert(&qstate->reply_origin, NULL, 0, qstate->region);
1319                         if(qstate->env->cfg->qname_minimisation)
1320                                 iq->minimisation_state = INIT_MINIMISE_STATE;
1321                         return next_state(iq, INIT_REQUEST_STATE);
1322                 }
1323
1324                 /* if from cache, NULL, else insert 'cache IP' len=0 */
1325                 if(qstate->reply_origin)
1326                         sock_list_insert(&qstate->reply_origin, NULL, 0, qstate->region);
1327                 if(FLAGS_GET_RCODE(msg->rep->flags) == LDNS_RCODE_SERVFAIL)
1328                         errinf(qstate, "SERVFAIL in cache");
1329                 /* it is an answer, response, to final state */
1330                 verbose(VERB_ALGO, "returning answer from cache.");
1331                 iq->response = msg;
1332                 return final_state(iq);
1333         }
1334
1335         /* attempt to forward the request */
1336         if(forward_request(qstate, iq))
1337         {
1338                 if(!iq->dp) {
1339                         log_err("alloc failure for forward dp");
1340                         errinf(qstate, "malloc failure for forward zone");
1341                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1342                 }
1343                 iq->refetch_glue = 0;
1344                 iq->minimisation_state = DONOT_MINIMISE_STATE;
1345                 /* the request has been forwarded.
1346                  * forwarded requests need to be immediately sent to the
1347                  * next state, QUERYTARGETS. */
1348                 return next_state(iq, QUERYTARGETS_STATE);
1349         }
1350
1351         /* Resolver Algorithm Step 2 -- find the "best" servers. */
1352
1353         /* first, adjust for DS queries. To avoid the grandparent problem,
1354          * we just look for the closest set of server to the parent of qname.
1355          * When re-fetching glue we also need to ask the parent.
1356          */
1357         if(iq->refetch_glue) {
1358                 if(!iq->dp) {
1359                         log_err("internal or malloc fail: no dp for refetch");
1360                         errinf(qstate, "malloc failure, for delegation info");
1361                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1362                 }
1363                 delname = iq->dp->name;
1364                 delnamelen = iq->dp->namelen;
1365         } else {
1366                 delname = iq->qchase.qname;
1367                 delnamelen = iq->qchase.qname_len;
1368         }
1369         if(iq->qchase.qtype == LDNS_RR_TYPE_DS || iq->refetch_glue ||
1370            (iq->qchase.qtype == LDNS_RR_TYPE_NS && qstate->prefetch_leeway
1371            && can_have_last_resort(qstate->env, delname, delnamelen, iq->qchase.qclass, NULL))) {
1372                 /* remove first label from delname, root goes to hints,
1373                  * but only to fetch glue, not for qtype=DS. */
1374                 /* also when prefetching an NS record, fetch it again from
1375                  * its parent, just as if it expired, so that you do not
1376                  * get stuck on an older nameserver that gives old NSrecords */
1377                 if(dname_is_root(delname) && (iq->refetch_glue ||
1378                         (iq->qchase.qtype == LDNS_RR_TYPE_NS &&
1379                         qstate->prefetch_leeway)))
1380                         delname = NULL; /* go to root priming */
1381                 else    dname_remove_label(&delname, &delnamelen);
1382         }
1383         /* delname is the name to lookup a delegation for. If NULL rootprime */
1384         while(1) {
1385
1386                 /* Lookup the delegation in the cache. If null, then the
1387                  * cache needs to be primed for the qclass. */
1388                 if(delname)
1389                      iq->dp = dns_cache_find_delegation(qstate->env, delname,
1390                         delnamelen, iq->qchase.qtype, iq->qchase.qclass,
1391                         qstate->region, &iq->deleg_msg,
1392                         *qstate->env->now+qstate->prefetch_leeway);
1393                 else iq->dp = NULL;
1394
1395                 /* If the cache has returned nothing, then we have a
1396                  * root priming situation. */
1397                 if(iq->dp == NULL) {
1398                         int r;
1399                         /* if under auth zone, no prime needed */
1400                         if(!auth_zone_delegpt(qstate, iq, delname, delnamelen))
1401                                 return error_response(qstate, id,
1402                                         LDNS_RCODE_SERVFAIL);
1403                         if(iq->dp) /* use auth zone dp */
1404                                 return next_state(iq, INIT_REQUEST_2_STATE);
1405                         /* if there is a stub, then no root prime needed */
1406                         r = prime_stub(qstate, iq, id, delname,
1407                                 iq->qchase.qclass);
1408                         if(r == 2)
1409                                 break; /* got noprime-stub-zone, continue */
1410                         else if(r)
1411                                 return 0; /* stub prime request made */
1412                         if(forwards_lookup_root(qstate->env->fwds,
1413                                 iq->qchase.qclass)) {
1414                                 /* forward zone root, no root prime needed */
1415                                 /* fill in some dp - safety belt */
1416                                 iq->dp = hints_lookup_root(qstate->env->hints,
1417                                         iq->qchase.qclass);
1418                                 if(!iq->dp) {
1419                                         log_err("internal error: no hints dp");
1420                                         errinf(qstate, "no hints for this class");
1421                                         return error_response(qstate, id,
1422                                                 LDNS_RCODE_SERVFAIL);
1423                                 }
1424                                 iq->dp = delegpt_copy(iq->dp, qstate->region);
1425                                 if(!iq->dp) {
1426                                         log_err("out of memory in safety belt");
1427                                         errinf(qstate, "malloc failure, in safety belt");
1428                                         return error_response(qstate, id,
1429                                                 LDNS_RCODE_SERVFAIL);
1430                                 }
1431                                 return next_state(iq, INIT_REQUEST_2_STATE);
1432                         }
1433                         /* Note that the result of this will set a new
1434                          * DelegationPoint based on the result of priming. */
1435                         if(!prime_root(qstate, iq, id, iq->qchase.qclass))
1436                                 return error_response(qstate, id,
1437                                         LDNS_RCODE_REFUSED);
1438
1439                         /* priming creates and sends a subordinate query, with
1440                          * this query as the parent. So further processing for
1441                          * this event will stop until reactivated by the
1442                          * results of priming. */
1443                         return 0;
1444                 }
1445                 if(!iq->ratelimit_ok && qstate->prefetch_leeway)
1446                         iq->ratelimit_ok = 1; /* allow prefetches, this keeps
1447                         otherwise valid data in the cache */
1448                 if(!iq->ratelimit_ok && infra_ratelimit_exceeded(
1449                         qstate->env->infra_cache, iq->dp->name,
1450                         iq->dp->namelen, *qstate->env->now)) {
1451                         /* and increment the rate, so that the rate for time
1452                          * now will also exceed the rate, keeping cache fresh */
1453                         (void)infra_ratelimit_inc(qstate->env->infra_cache,
1454                                 iq->dp->name, iq->dp->namelen,
1455                                 *qstate->env->now, &qstate->qinfo,
1456                                 qstate->reply);
1457                         /* see if we are passed through with slip factor */
1458                         if(qstate->env->cfg->ratelimit_factor != 0 &&
1459                                 ub_random_max(qstate->env->rnd,
1460                                     qstate->env->cfg->ratelimit_factor) == 1) {
1461                                 iq->ratelimit_ok = 1;
1462                                 log_nametypeclass(VERB_ALGO, "ratelimit allowed through for "
1463                                         "delegation point", iq->dp->name,
1464                                         LDNS_RR_TYPE_NS, LDNS_RR_CLASS_IN);
1465                         } else {
1466                                 lock_basic_lock(&ie->queries_ratelimit_lock);
1467                                 ie->num_queries_ratelimited++;
1468                                 lock_basic_unlock(&ie->queries_ratelimit_lock);
1469                                 log_nametypeclass(VERB_ALGO, "ratelimit exceeded with "
1470                                         "delegation point", iq->dp->name,
1471                                         LDNS_RR_TYPE_NS, LDNS_RR_CLASS_IN);
1472                                 qstate->was_ratelimited = 1;
1473                                 errinf(qstate, "query was ratelimited");
1474                                 errinf_dname(qstate, "for zone", iq->dp->name);
1475                                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1476                         }
1477                 }
1478
1479                 /* see if this dp not useless.
1480                  * It is useless if:
1481                  *      o all NS items are required glue.
1482                  *        or the query is for NS item that is required glue.
1483                  *      o no addresses are provided.
1484                  *      o RD qflag is on.
1485                  * Instead, go up one level, and try to get even further
1486                  * If the root was useless, use safety belt information.
1487                  * Only check cache returns, because replies for servers
1488                  * could be useless but lead to loops (bumping into the
1489                  * same server reply) if useless-checked.
1490                  */
1491                 if(iter_dp_is_useless(&qstate->qinfo, qstate->query_flags,
1492                         iq->dp)) {
1493                         struct delegpt* retdp = NULL;
1494                         if(!can_have_last_resort(qstate->env, iq->dp->name, iq->dp->namelen, iq->qchase.qclass, &retdp)) {
1495                                 if(retdp) {
1496                                         verbose(VERB_QUERY, "cache has stub "
1497                                                 "or fwd but no addresses, "
1498                                                 "fallback to config");
1499                                         iq->dp = delegpt_copy(retdp,
1500                                                 qstate->region);
1501                                         if(!iq->dp) {
1502                                                 log_err("out of memory in "
1503                                                         "stub/fwd fallback");
1504                                                 errinf(qstate, "malloc failure, for fallback to config");
1505                                                 return error_response(qstate,
1506                                                     id, LDNS_RCODE_SERVFAIL);
1507                                         }
1508                                         break;
1509                                 }
1510                                 verbose(VERB_ALGO, "useless dp "
1511                                         "but cannot go up, servfail");
1512                                 delegpt_log(VERB_ALGO, iq->dp);
1513                                 errinf(qstate, "no useful nameservers, "
1514                                         "and cannot go up");
1515                                 errinf_dname(qstate, "for zone", iq->dp->name);
1516                                 return error_response(qstate, id,
1517                                         LDNS_RCODE_SERVFAIL);
1518                         }
1519                         if(dname_is_root(iq->dp->name)) {
1520                                 /* use safety belt */
1521                                 verbose(VERB_QUERY, "Cache has root NS but "
1522                                 "no addresses. Fallback to the safety belt.");
1523                                 iq->dp = hints_lookup_root(qstate->env->hints,
1524                                         iq->qchase.qclass);
1525                                 /* note deleg_msg is from previous lookup,
1526                                  * but RD is on, so it is not used */
1527                                 if(!iq->dp) {
1528                                         log_err("internal error: no hints dp");
1529                                         return error_response(qstate, id,
1530                                                 LDNS_RCODE_REFUSED);
1531                                 }
1532                                 iq->dp = delegpt_copy(iq->dp, qstate->region);
1533                                 if(!iq->dp) {
1534                                         log_err("out of memory in safety belt");
1535                                         errinf(qstate, "malloc failure, in safety belt, for root");
1536                                         return error_response(qstate, id,
1537                                                 LDNS_RCODE_SERVFAIL);
1538                                 }
1539                                 break;
1540                         } else {
1541                                 verbose(VERB_ALGO,
1542                                         "cache delegation was useless:");
1543                                 delegpt_log(VERB_ALGO, iq->dp);
1544                                 /* go up */
1545                                 delname = iq->dp->name;
1546                                 delnamelen = iq->dp->namelen;
1547                                 dname_remove_label(&delname, &delnamelen);
1548                         }
1549                 } else break;
1550         }
1551
1552         verbose(VERB_ALGO, "cache delegation returns delegpt");
1553         delegpt_log(VERB_ALGO, iq->dp);
1554
1555         /* Otherwise, set the current delegation point and move on to the
1556          * next state. */
1557         return next_state(iq, INIT_REQUEST_2_STATE);
1558 }
1559
1560 /**
1561  * Process the second part of the initial request handling. This state
1562  * basically exists so that queries that generate root priming events have
1563  * the same init processing as ones that do not. Request events that reach
1564  * this state must have a valid currentDelegationPoint set.
1565  *
1566  * This part is primarily handling stub zone priming. Events that reach this
1567  * state must have a current delegation point.
1568  *
1569  * @param qstate: query state.
1570  * @param iq: iterator query state.
1571  * @param id: module id.
1572  * @return true if the event needs more request processing immediately,
1573  *         false if not.
1574  */
1575 static int
1576 processInitRequest2(struct module_qstate* qstate, struct iter_qstate* iq,
1577         int id)
1578 {
1579         uint8_t* delname;
1580         size_t delnamelen;
1581         log_query_info(VERB_QUERY, "resolving (init part 2): ",
1582                 &qstate->qinfo);
1583
1584         delname = iq->qchase.qname;
1585         delnamelen = iq->qchase.qname_len;
1586         if(iq->refetch_glue) {
1587                 struct iter_hints_stub* stub;
1588                 if(!iq->dp) {
1589                         log_err("internal or malloc fail: no dp for refetch");
1590                         errinf(qstate, "malloc failure, no delegation info");
1591                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1592                 }
1593                 /* Do not send queries above stub, do not set delname to dp if
1594                  * this is above stub without stub-first. */
1595                 stub = hints_lookup_stub(
1596                         qstate->env->hints, iq->qchase.qname, iq->qchase.qclass,
1597                         iq->dp);
1598                 if(!stub || !stub->dp->has_parent_side_NS ||
1599                         dname_subdomain_c(iq->dp->name, stub->dp->name)) {
1600                         delname = iq->dp->name;
1601                         delnamelen = iq->dp->namelen;
1602                 }
1603         }
1604         if(iq->qchase.qtype == LDNS_RR_TYPE_DS || iq->refetch_glue) {
1605                 if(!dname_is_root(delname))
1606                         dname_remove_label(&delname, &delnamelen);
1607                 iq->refetch_glue = 0; /* if CNAME causes restart, no refetch */
1608         }
1609
1610         /* see if we have an auth zone to answer from, improves dp from cache
1611          * (if any dp from cache) with auth zone dp, if that is lower */
1612         if(!auth_zone_delegpt(qstate, iq, delname, delnamelen))
1613                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1614
1615         /* Check to see if we need to prime a stub zone. */
1616         if(prime_stub(qstate, iq, id, delname, iq->qchase.qclass)) {
1617                 /* A priming sub request was made */
1618                 return 0;
1619         }
1620
1621         /* most events just get forwarded to the next state. */
1622         return next_state(iq, INIT_REQUEST_3_STATE);
1623 }
1624
1625 /**
1626  * Process the third part of the initial request handling. This state exists
1627  * as a separate state so that queries that generate stub priming events
1628  * will get the tail end of the init process but not repeat the stub priming
1629  * check.
1630  *
1631  * @param qstate: query state.
1632  * @param iq: iterator query state.
1633  * @param id: module id.
1634  * @return true, advancing the event to the QUERYTARGETS_STATE.
1635  */
1636 static int
1637 processInitRequest3(struct module_qstate* qstate, struct iter_qstate* iq,
1638         int id)
1639 {
1640         log_query_info(VERB_QUERY, "resolving (init part 3): ",
1641                 &qstate->qinfo);
1642         /* if the cache reply dp equals a validation anchor or msg has DS,
1643          * then DNSSEC RRSIGs are expected in the reply */
1644         iq->dnssec_expected = iter_indicates_dnssec(qstate->env, iq->dp,
1645                 iq->deleg_msg, iq->qchase.qclass);
1646
1647         /* If the RD flag wasn't set, then we just finish with the
1648          * cached referral as the response. */
1649         if(!(qstate->query_flags & BIT_RD) && iq->deleg_msg) {
1650                 iq->response = iq->deleg_msg;
1651                 if(verbosity >= VERB_ALGO && iq->response)
1652                         log_dns_msg("no RD requested, using delegation msg",
1653                                 &iq->response->qinfo, iq->response->rep);
1654                 if(qstate->reply_origin)
1655                         sock_list_insert(&qstate->reply_origin, NULL, 0, qstate->region);
1656                 return final_state(iq);
1657         }
1658         /* After this point, unset the RD flag -- this query is going to
1659          * be sent to an auth. server. */
1660         iq->chase_flags &= ~BIT_RD;
1661
1662         /* if dnssec expected, fetch key for the trust-anchor or cached-DS */
1663         if(iq->dnssec_expected && qstate->env->cfg->prefetch_key &&
1664                 !(qstate->query_flags&BIT_CD)) {
1665                 generate_dnskey_prefetch(qstate, iq, id);
1666                 fptr_ok(fptr_whitelist_modenv_detach_subs(
1667                         qstate->env->detach_subs));
1668                 (*qstate->env->detach_subs)(qstate);
1669         }
1670
1671         /* Jump to the next state. */
1672         return next_state(iq, QUERYTARGETS_STATE);
1673 }
1674
1675 /**
1676  * Given a basic query, generate a parent-side "target" query.
1677  * These are subordinate queries for missing delegation point target addresses,
1678  * for which only the parent of the delegation provides correct IP addresses.
1679  *
1680  * @param qstate: query state.
1681  * @param iq: iterator query state.
1682  * @param id: module id.
1683  * @param name: target qname.
1684  * @param namelen: target qname length.
1685  * @param qtype: target qtype (either A or AAAA).
1686  * @param qclass: target qclass.
1687  * @return true on success, false on failure.
1688  */
1689 static int
1690 generate_parentside_target_query(struct module_qstate* qstate,
1691         struct iter_qstate* iq, int id, uint8_t* name, size_t namelen,
1692         uint16_t qtype, uint16_t qclass)
1693 {
1694         struct module_qstate* subq;
1695         if(!generate_sub_request(name, namelen, qtype, qclass, qstate,
1696                 id, iq, INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0))
1697                 return 0;
1698         if(subq) {
1699                 struct iter_qstate* subiq =
1700                         (struct iter_qstate*)subq->minfo[id];
1701                 /* blacklist the cache - we want to fetch parent stuff */
1702                 sock_list_insert(&subq->blacklist, NULL, 0, subq->region);
1703                 subiq->query_for_pside_glue = 1;
1704                 if(dname_subdomain_c(name, iq->dp->name)) {
1705                         subiq->dp = delegpt_copy(iq->dp, subq->region);
1706                         subiq->dnssec_expected = iter_indicates_dnssec(
1707                                 qstate->env, subiq->dp, NULL,
1708                                 subq->qinfo.qclass);
1709                         subiq->refetch_glue = 1;
1710                 } else {
1711                         subiq->dp = dns_cache_find_delegation(qstate->env,
1712                                 name, namelen, qtype, qclass, subq->region,
1713                                 &subiq->deleg_msg,
1714                                 *qstate->env->now+subq->prefetch_leeway);
1715                         /* if no dp, then it's from root, refetch unneeded */
1716                         if(subiq->dp) {
1717                                 subiq->dnssec_expected = iter_indicates_dnssec(
1718                                         qstate->env, subiq->dp, NULL,
1719                                         subq->qinfo.qclass);
1720                                 subiq->refetch_glue = 1;
1721                         }
1722                 }
1723         }
1724         log_nametypeclass(VERB_QUERY, "new pside target", name, qtype, qclass);
1725         return 1;
1726 }
1727
1728 /**
1729  * Given a basic query, generate a "target" query. These are subordinate
1730  * queries for missing delegation point target addresses.
1731  *
1732  * @param qstate: query state.
1733  * @param iq: iterator query state.
1734  * @param id: module id.
1735  * @param name: target qname.
1736  * @param namelen: target qname length.
1737  * @param qtype: target qtype (either A or AAAA).
1738  * @param qclass: target qclass.
1739  * @return true on success, false on failure.
1740  */
1741 static int
1742 generate_target_query(struct module_qstate* qstate, struct iter_qstate* iq,
1743         int id, uint8_t* name, size_t namelen, uint16_t qtype, uint16_t qclass)
1744 {
1745         struct module_qstate* subq;
1746         if(!generate_sub_request(name, namelen, qtype, qclass, qstate,
1747                 id, iq, INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0))
1748                 return 0;
1749         log_nametypeclass(VERB_QUERY, "new target", name, qtype, qclass);
1750         return 1;
1751 }
1752
1753 /**
1754  * Given an event at a certain state, generate zero or more target queries
1755  * for it's current delegation point.
1756  *
1757  * @param qstate: query state.
1758  * @param iq: iterator query state.
1759  * @param ie: iterator shared global environment.
1760  * @param id: module id.
1761  * @param maxtargets: The maximum number of targets to query for.
1762  *      if it is negative, there is no maximum number of targets.
1763  * @param num: returns the number of queries generated and processed,
1764  *      which may be zero if there were no missing targets.
1765  * @return false on error.
1766  */
1767 static int
1768 query_for_targets(struct module_qstate* qstate, struct iter_qstate* iq,
1769         struct iter_env* ie, int id, int maxtargets, int* num)
1770 {
1771         int query_count = 0;
1772         struct delegpt_ns* ns;
1773         int missing;
1774         int toget = 0;
1775
1776         if(iq->depth == ie->max_dependency_depth)
1777                 return 0;
1778         if(iq->depth > 0 && iq->target_count &&
1779                 iq->target_count[1] > MAX_TARGET_COUNT) {
1780                 char s[LDNS_MAX_DOMAINLEN+1];
1781                 dname_str(qstate->qinfo.qname, s);
1782                 verbose(VERB_QUERY, "request %s has exceeded the maximum "
1783                         "number of glue fetches %d", s, iq->target_count[1]);
1784                 return 0;
1785         }
1786
1787         iter_mark_cycle_targets(qstate, iq->dp);
1788         missing = (int)delegpt_count_missing_targets(iq->dp);
1789         log_assert(maxtargets != 0); /* that would not be useful */
1790
1791         /* Generate target requests. Basically, any missing targets
1792          * are queried for here, regardless if it is necessary to do
1793          * so to continue processing. */
1794         if(maxtargets < 0 || maxtargets > missing)
1795                 toget = missing;
1796         else    toget = maxtargets;
1797         if(toget == 0) {
1798                 *num = 0;
1799                 return 1;
1800         }
1801         /* select 'toget' items from the total of 'missing' items */
1802         log_assert(toget <= missing);
1803
1804         /* loop over missing targets */
1805         for(ns = iq->dp->nslist; ns; ns = ns->next) {
1806                 if(ns->resolved)
1807                         continue;
1808
1809                 /* randomly select this item with probability toget/missing */
1810                 if(!iter_ns_probability(qstate->env->rnd, toget, missing)) {
1811                         /* do not select this one, next; select toget number
1812                          * of items from a list one less in size */
1813                         missing --;
1814                         continue;
1815                 }
1816
1817                 if(ie->supports_ipv6 && !ns->got6) {
1818                         /* Send the AAAA request. */
1819                         if(!generate_target_query(qstate, iq, id,
1820                                 ns->name, ns->namelen,
1821                                 LDNS_RR_TYPE_AAAA, iq->qchase.qclass)) {
1822                                 *num = query_count;
1823                                 if(query_count > 0)
1824                                         qstate->ext_state[id] = module_wait_subquery;
1825                                 return 0;
1826                         }
1827                         query_count++;
1828                 }
1829                 /* Send the A request. */
1830                 if(ie->supports_ipv4 && !ns->got4) {
1831                         if(!generate_target_query(qstate, iq, id,
1832                                 ns->name, ns->namelen,
1833                                 LDNS_RR_TYPE_A, iq->qchase.qclass)) {
1834                                 *num = query_count;
1835                                 if(query_count > 0)
1836                                         qstate->ext_state[id] = module_wait_subquery;
1837                                 return 0;
1838                         }
1839                         query_count++;
1840                 }
1841
1842                 /* mark this target as in progress. */
1843                 ns->resolved = 1;
1844                 missing--;
1845                 toget--;
1846                 if(toget == 0)
1847                         break;
1848         }
1849         *num = query_count;
1850         if(query_count > 0)
1851                 qstate->ext_state[id] = module_wait_subquery;
1852
1853         return 1;
1854 }
1855
1856 /**
1857  * Called by processQueryTargets when it would like extra targets to query
1858  * but it seems to be out of options.  At last resort some less appealing
1859  * options are explored.  If there are no more options, the result is SERVFAIL
1860  *
1861  * @param qstate: query state.
1862  * @param iq: iterator query state.
1863  * @param ie: iterator shared global environment.
1864  * @param id: module id.
1865  * @return true if the event requires more request processing immediately,
1866  *         false if not.
1867  */
1868 static int
1869 processLastResort(struct module_qstate* qstate, struct iter_qstate* iq,
1870         struct iter_env* ie, int id)
1871 {
1872         struct delegpt_ns* ns;
1873         int query_count = 0;
1874         verbose(VERB_ALGO, "No more query targets, attempting last resort");
1875         log_assert(iq->dp);
1876
1877         if(!can_have_last_resort(qstate->env, iq->dp->name, iq->dp->namelen,
1878                 iq->qchase.qclass, NULL)) {
1879                 /* fail -- no more targets, no more hope of targets, no hope
1880                  * of a response. */
1881                 errinf(qstate, "all the configured stub or forward servers failed,");
1882                 errinf_dname(qstate, "at zone", iq->dp->name);
1883                 verbose(VERB_QUERY, "configured stub or forward servers failed -- returning SERVFAIL");
1884                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1885         }
1886         if(!iq->dp->has_parent_side_NS && dname_is_root(iq->dp->name)) {
1887                 struct delegpt* p = hints_lookup_root(qstate->env->hints,
1888                         iq->qchase.qclass);
1889                 if(p) {
1890                         struct delegpt_addr* a;
1891                         iq->chase_flags &= ~BIT_RD; /* go to authorities */
1892                         for(ns = p->nslist; ns; ns=ns->next) {
1893                                 (void)delegpt_add_ns(iq->dp, qstate->region,
1894                                         ns->name, ns->lame);
1895                         }
1896                         for(a = p->target_list; a; a=a->next_target) {
1897                                 (void)delegpt_add_addr(iq->dp, qstate->region,
1898                                         &a->addr, a->addrlen, a->bogus,
1899                                         a->lame, a->tls_auth_name);
1900                         }
1901                 }
1902                 iq->dp->has_parent_side_NS = 1;
1903         } else if(!iq->dp->has_parent_side_NS) {
1904                 if(!iter_lookup_parent_NS_from_cache(qstate->env, iq->dp,
1905                         qstate->region, &qstate->qinfo)
1906                         || !iq->dp->has_parent_side_NS) {
1907                         /* if: malloc failure in lookup go up to try */
1908                         /* if: no parent NS in cache - go up one level */
1909                         verbose(VERB_ALGO, "try to grab parent NS");
1910                         iq->store_parent_NS = iq->dp;
1911                         iq->chase_flags &= ~BIT_RD; /* go to authorities */
1912                         iq->deleg_msg = NULL;
1913                         iq->refetch_glue = 1;
1914                         iq->query_restart_count++;
1915                         iq->sent_count = 0;
1916                         if(qstate->env->cfg->qname_minimisation)
1917                                 iq->minimisation_state = INIT_MINIMISE_STATE;
1918                         return next_state(iq, INIT_REQUEST_STATE);
1919                 }
1920         }
1921         /* see if that makes new names available */
1922         if(!cache_fill_missing(qstate->env, iq->qchase.qclass,
1923                 qstate->region, iq->dp))
1924                 log_err("out of memory in cache_fill_missing");
1925         if(iq->dp->usable_list) {
1926                 verbose(VERB_ALGO, "try parent-side-name, w. glue from cache");
1927                 return next_state(iq, QUERYTARGETS_STATE);
1928         }
1929         /* try to fill out parent glue from cache */
1930         if(iter_lookup_parent_glue_from_cache(qstate->env, iq->dp,
1931                 qstate->region, &qstate->qinfo)) {
1932                 /* got parent stuff from cache, see if we can continue */
1933                 verbose(VERB_ALGO, "try parent-side glue from cache");
1934                 return next_state(iq, QUERYTARGETS_STATE);
1935         }
1936         /* query for an extra name added by the parent-NS record */
1937         if(delegpt_count_missing_targets(iq->dp) > 0) {
1938                 int qs = 0;
1939                 verbose(VERB_ALGO, "try parent-side target name");
1940                 if(!query_for_targets(qstate, iq, ie, id, 1, &qs)) {
1941                         errinf(qstate, "could not fetch nameserver");
1942                         errinf_dname(qstate, "at zone", iq->dp->name);
1943                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
1944                 }
1945                 iq->num_target_queries += qs;
1946                 target_count_increase(iq, qs);
1947                 if(qs != 0) {
1948                         qstate->ext_state[id] = module_wait_subquery;
1949                         return 0; /* and wait for them */
1950                 }
1951         }
1952         if(iq->depth == ie->max_dependency_depth) {
1953                 verbose(VERB_QUERY, "maxdepth and need more nameservers, fail");
1954                 errinf(qstate, "cannot fetch more nameservers because at max dependency depth");
1955                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1956         }
1957         if(iq->depth > 0 && iq->target_count &&
1958                 iq->target_count[1] > MAX_TARGET_COUNT) {
1959                 char s[LDNS_MAX_DOMAINLEN+1];
1960                 dname_str(qstate->qinfo.qname, s);
1961                 verbose(VERB_QUERY, "request %s has exceeded the maximum "
1962                         "number of glue fetches %d", s, iq->target_count[1]);
1963                 errinf(qstate, "exceeded the maximum number of glue fetches");
1964                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
1965         }
1966         /* mark cycle targets for parent-side lookups */
1967         iter_mark_pside_cycle_targets(qstate, iq->dp);
1968         /* see if we can issue queries to get nameserver addresses */
1969         /* this lookup is not randomized, but sequential. */
1970         for(ns = iq->dp->nslist; ns; ns = ns->next) {
1971                 /* if this nameserver is at a delegation point, but that
1972                  * delegation point is a stub and we cannot go higher, skip*/
1973                 if( ((ie->supports_ipv6 && !ns->done_pside6) ||
1974                     (ie->supports_ipv4 && !ns->done_pside4)) &&
1975                     !can_have_last_resort(qstate->env, ns->name, ns->namelen,
1976                         iq->qchase.qclass, NULL)) {
1977                         log_nametypeclass(VERB_ALGO, "cannot pside lookup ns "
1978                                 "because it is also a stub/forward,",
1979                                 ns->name, LDNS_RR_TYPE_NS, iq->qchase.qclass);
1980                         if(ie->supports_ipv6) ns->done_pside6 = 1;
1981                         if(ie->supports_ipv4) ns->done_pside4 = 1;
1982                         continue;
1983                 }
1984                 /* query for parent-side A and AAAA for nameservers */
1985                 if(ie->supports_ipv6 && !ns->done_pside6) {
1986                         /* Send the AAAA request. */
1987                         if(!generate_parentside_target_query(qstate, iq, id,
1988                                 ns->name, ns->namelen,
1989                                 LDNS_RR_TYPE_AAAA, iq->qchase.qclass)) {
1990                                 errinf_dname(qstate, "could not generate nameserver AAAA lookup for", ns->name);
1991                                 return error_response(qstate, id,
1992                                         LDNS_RCODE_SERVFAIL);
1993                         }
1994                         ns->done_pside6 = 1;
1995                         query_count++;
1996                 }
1997                 if(ie->supports_ipv4 && !ns->done_pside4) {
1998                         /* Send the A request. */
1999                         if(!generate_parentside_target_query(qstate, iq, id,
2000                                 ns->name, ns->namelen,
2001                                 LDNS_RR_TYPE_A, iq->qchase.qclass)) {
2002                                 errinf_dname(qstate, "could not generate nameserver A lookup for", ns->name);
2003                                 return error_response(qstate, id,
2004                                         LDNS_RCODE_SERVFAIL);
2005                         }
2006                         ns->done_pside4 = 1;
2007                         query_count++;
2008                 }
2009                 if(query_count != 0) { /* suspend to await results */
2010                         verbose(VERB_ALGO, "try parent-side glue lookup");
2011                         iq->num_target_queries += query_count;
2012                         target_count_increase(iq, query_count);
2013                         qstate->ext_state[id] = module_wait_subquery;
2014                         return 0;
2015                 }
2016         }
2017
2018         /* if this was a parent-side glue query itself, then store that
2019          * failure in cache. */
2020         if(!qstate->no_cache_store && iq->query_for_pside_glue
2021                 && !iq->pside_glue)
2022                         iter_store_parentside_neg(qstate->env, &qstate->qinfo,
2023                                 iq->deleg_msg?iq->deleg_msg->rep:
2024                                 (iq->response?iq->response->rep:NULL));
2025
2026         errinf(qstate, "all servers for this domain failed,");
2027         errinf_dname(qstate, "at zone", iq->dp->name);
2028         verbose(VERB_QUERY, "out of query targets -- returning SERVFAIL");
2029         /* fail -- no more targets, no more hope of targets, no hope
2030          * of a response. */
2031         return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
2032 }
2033
2034 /**
2035  * Try to find the NS record set that will resolve a qtype DS query. Due
2036  * to grandparent/grandchild reasons we did not get a proper lookup right
2037  * away.  We need to create type NS queries until we get the right parent
2038  * for this lookup.  We remove labels from the query to find the right point.
2039  * If we end up at the old dp name, then there is no solution.
2040  *
2041  * @param qstate: query state.
2042  * @param iq: iterator query state.
2043  * @param id: module id.
2044  * @return true if the event requires more immediate processing, false if
2045  *         not. This is generally only true when forwarding the request to
2046  *         the final state (i.e., on answer).
2047  */
2048 static int
2049 processDSNSFind(struct module_qstate* qstate, struct iter_qstate* iq, int id)
2050 {
2051         struct module_qstate* subq = NULL;
2052         verbose(VERB_ALGO, "processDSNSFind");
2053
2054         if(!iq->dsns_point) {
2055                 /* initialize */
2056                 iq->dsns_point = iq->qchase.qname;
2057                 iq->dsns_point_len = iq->qchase.qname_len;
2058         }
2059         /* robustcheck for internal error: we are not underneath the dp */
2060         if(!dname_subdomain_c(iq->dsns_point, iq->dp->name)) {
2061                 errinf_dname(qstate, "for DS query parent-child nameserver search the query is not under the zone", iq->dp->name);
2062                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
2063         }
2064
2065         /* go up one (more) step, until we hit the dp, if so, end */
2066         dname_remove_label(&iq->dsns_point, &iq->dsns_point_len);
2067         if(query_dname_compare(iq->dsns_point, iq->dp->name) == 0) {
2068                 /* there was no inbetween nameserver, use the old delegation
2069                  * point again.  And this time, because dsns_point is nonNULL
2070                  * we are going to accept the (bad) result */
2071                 iq->state = QUERYTARGETS_STATE;
2072                 return 1;
2073         }
2074         iq->state = DSNS_FIND_STATE;
2075
2076         /* spawn NS lookup (validation not needed, this is for DS lookup) */
2077         log_nametypeclass(VERB_ALGO, "fetch nameservers",
2078                 iq->dsns_point, LDNS_RR_TYPE_NS, iq->qchase.qclass);
2079         if(!generate_sub_request(iq->dsns_point, iq->dsns_point_len,
2080                 LDNS_RR_TYPE_NS, iq->qchase.qclass, qstate, id, iq,
2081                 INIT_REQUEST_STATE, FINISHED_STATE, &subq, 0)) {
2082                 errinf_dname(qstate, "for DS query parent-child nameserver search, could not generate NS lookup for", iq->dsns_point);
2083                 return error_response_cache(qstate, id, LDNS_RCODE_SERVFAIL);
2084         }
2085
2086         return 0;
2087 }
2088
2089 /**
2090  * This is the request event state where the request will be sent to one of
2091  * its current query targets. This state also handles issuing target lookup
2092  * queries for missing target IP addresses. Queries typically iterate on
2093  * this state, both when they are just trying different targets for a given
2094  * delegation point, and when they change delegation points. This state
2095  * roughly corresponds to RFC 1034 algorithm steps 3 and 4.
2096  *
2097  * @param qstate: query state.
2098  * @param iq: iterator query state.
2099  * @param ie: iterator shared global environment.
2100  * @param id: module id.
2101  * @return true if the event requires more request processing immediately,
2102  *         false if not. This state only returns true when it is generating
2103  *         a SERVFAIL response because the query has hit a dead end.
2104  */
2105 static int
2106 processQueryTargets(struct module_qstate* qstate, struct iter_qstate* iq,
2107         struct iter_env* ie, int id)
2108 {
2109         int tf_policy;
2110         struct delegpt_addr* target;
2111         struct outbound_entry* outq;
2112         int auth_fallback = 0;
2113         uint8_t* qout_orig = NULL;
2114         size_t qout_orig_len = 0;
2115
2116         /* NOTE: a request will encounter this state for each target it
2117          * needs to send a query to. That is, at least one per referral,
2118          * more if some targets timeout or return throwaway answers. */
2119
2120         log_query_info(VERB_QUERY, "processQueryTargets:", &qstate->qinfo);
2121         verbose(VERB_ALGO, "processQueryTargets: targetqueries %d, "
2122                 "currentqueries %d sentcount %d", iq->num_target_queries,
2123                 iq->num_current_queries, iq->sent_count);
2124
2125         /* Make sure that we haven't run away */
2126         /* FIXME: is this check even necessary? */
2127         if(iq->referral_count > MAX_REFERRAL_COUNT) {
2128                 verbose(VERB_QUERY, "request has exceeded the maximum "
2129                         "number of referrrals with %d", iq->referral_count);
2130                 errinf(qstate, "exceeded the maximum of referrals");
2131                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2132         }
2133         if(iq->sent_count > MAX_SENT_COUNT) {
2134                 verbose(VERB_QUERY, "request has exceeded the maximum "
2135                         "number of sends with %d", iq->sent_count);
2136                 errinf(qstate, "exceeded the maximum number of sends");
2137                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2138         }
2139
2140         /* Make sure we have a delegation point, otherwise priming failed
2141          * or another failure occurred */
2142         if(!iq->dp) {
2143                 verbose(VERB_QUERY, "Failed to get a delegation, giving up");
2144                 errinf(qstate, "failed to get a delegation (eg. prime failure)");
2145                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2146         }
2147         if(!ie->supports_ipv6)
2148                 delegpt_no_ipv6(iq->dp);
2149         if(!ie->supports_ipv4)
2150                 delegpt_no_ipv4(iq->dp);
2151         delegpt_log(VERB_ALGO, iq->dp);
2152
2153         if(iq->num_current_queries>0) {
2154                 /* already busy answering a query, this restart is because
2155                  * more delegpt addrs became available, wait for existing
2156                  * query. */
2157                 verbose(VERB_ALGO, "woke up, but wait for outstanding query");
2158                 qstate->ext_state[id] = module_wait_reply;
2159                 return 0;
2160         }
2161
2162         if(iq->minimisation_state == INIT_MINIMISE_STATE
2163                 && !(iq->chase_flags & BIT_RD)) {
2164                 /* (Re)set qinfo_out to (new) delegation point, except when
2165                  * qinfo_out is already a subdomain of dp. This happens when
2166                  * increasing by more than one label at once (QNAMEs with more
2167                  * than MAX_MINIMISE_COUNT labels). */
2168                 if(!(iq->qinfo_out.qname_len
2169                         && dname_subdomain_c(iq->qchase.qname,
2170                                 iq->qinfo_out.qname)
2171                         && dname_subdomain_c(iq->qinfo_out.qname,
2172                                 iq->dp->name))) {
2173                         iq->qinfo_out.qname = iq->dp->name;
2174                         iq->qinfo_out.qname_len = iq->dp->namelen;
2175                         iq->qinfo_out.qtype = LDNS_RR_TYPE_A;
2176                         iq->qinfo_out.qclass = iq->qchase.qclass;
2177                         iq->qinfo_out.local_alias = NULL;
2178                         iq->minimise_count = 0;
2179                 }
2180
2181                 iq->minimisation_state = MINIMISE_STATE;
2182         }
2183         if(iq->minimisation_state == MINIMISE_STATE) {
2184                 int qchaselabs = dname_count_labels(iq->qchase.qname);
2185                 int labdiff = qchaselabs -
2186                         dname_count_labels(iq->qinfo_out.qname);
2187
2188                 qout_orig = iq->qinfo_out.qname;
2189                 qout_orig_len = iq->qinfo_out.qname_len;
2190                 iq->qinfo_out.qname = iq->qchase.qname;
2191                 iq->qinfo_out.qname_len = iq->qchase.qname_len;
2192                 iq->minimise_count++;
2193                 iq->minimise_timeout_count = 0;
2194
2195                 iter_dec_attempts(iq->dp, 1);
2196
2197                 /* Limit number of iterations for QNAMEs with more
2198                  * than MAX_MINIMISE_COUNT labels. Send first MINIMISE_ONE_LAB
2199                  * labels of QNAME always individually.
2200                  */
2201                 if(qchaselabs > MAX_MINIMISE_COUNT && labdiff > 1 &&
2202                         iq->minimise_count > MINIMISE_ONE_LAB) {
2203                         if(iq->minimise_count < MAX_MINIMISE_COUNT) {
2204                                 int multilabs = qchaselabs - 1 -
2205                                         MINIMISE_ONE_LAB;
2206                                 int extralabs = multilabs /
2207                                         MINIMISE_MULTIPLE_LABS;
2208
2209                                 if (MAX_MINIMISE_COUNT - iq->minimise_count >=
2210                                         multilabs % MINIMISE_MULTIPLE_LABS)
2211                                         /* Default behaviour is to add 1 label
2212                                          * every iteration. Therefore, decrement
2213                                          * the extralabs by 1 */
2214                                         extralabs--;
2215                                 if (extralabs < labdiff)
2216                                         labdiff -= extralabs;
2217                                 else
2218                                         labdiff = 1;
2219                         }
2220                         /* Last minimised iteration, send all labels with
2221                          * QTYPE=NS */
2222                         else
2223                                 labdiff = 1;
2224                 }
2225
2226                 if(labdiff > 1) {
2227                         verbose(VERB_QUERY, "removing %d labels", labdiff-1);
2228                         dname_remove_labels(&iq->qinfo_out.qname,
2229                                 &iq->qinfo_out.qname_len,
2230                                 labdiff-1);
2231                 }
2232                 if(labdiff < 1 || (labdiff < 2
2233                         && (iq->qchase.qtype == LDNS_RR_TYPE_DS
2234                         || iq->qchase.qtype == LDNS_RR_TYPE_A)))
2235                         /* Stop minimising this query, resolve "as usual" */
2236                         iq->minimisation_state = DONOT_MINIMISE_STATE;
2237                 else if(!qstate->no_cache_lookup) {
2238                         struct dns_msg* msg = dns_cache_lookup(qstate->env,
2239                                 iq->qinfo_out.qname, iq->qinfo_out.qname_len,
2240                                 iq->qinfo_out.qtype, iq->qinfo_out.qclass,
2241                                 qstate->query_flags, qstate->region,
2242                                 qstate->env->scratch, 0);
2243                         if(msg && msg->rep->an_numrrsets == 0
2244                                 && FLAGS_GET_RCODE(msg->rep->flags) ==
2245                                 LDNS_RCODE_NOERROR)
2246                                 /* no need to send query if it is already
2247                                  * cached as NOERROR/NODATA */
2248                                 return 1;
2249                 }
2250         }
2251         if(iq->minimisation_state == SKIP_MINIMISE_STATE) {
2252                 if(iq->minimise_timeout_count < MAX_MINIMISE_TIMEOUT_COUNT)
2253                         /* Do not increment qname, continue incrementing next
2254                          * iteration */
2255                         iq->minimisation_state = MINIMISE_STATE;
2256                 else if(!qstate->env->cfg->qname_minimisation_strict)
2257                         /* Too many time-outs detected for this QNAME and QTYPE.
2258                          * We give up, disable QNAME minimisation. */
2259                         iq->minimisation_state = DONOT_MINIMISE_STATE;
2260         }
2261         if(iq->minimisation_state == DONOT_MINIMISE_STATE)
2262                 iq->qinfo_out = iq->qchase;
2263
2264         /* now find an answer to this query */
2265         /* see if authority zones have an answer */
2266         /* now we know the dp, we can check the auth zone for locally hosted
2267          * contents */
2268         if(!iq->auth_zone_avoid && qstate->blacklist) {
2269                 if(auth_zones_can_fallback(qstate->env->auth_zones,
2270                         iq->dp->name, iq->dp->namelen, iq->qinfo_out.qclass)) {
2271                         /* if cache is blacklisted and this zone allows us
2272                          * to fallback to the internet, then do so, and
2273                          * fetch results from the internet servers */
2274                         iq->auth_zone_avoid = 1;
2275                 }
2276         }
2277         if(iq->auth_zone_avoid) {
2278                 iq->auth_zone_avoid = 0;
2279                 auth_fallback = 1;
2280         } else if(auth_zones_lookup(qstate->env->auth_zones, &iq->qinfo_out,
2281                 qstate->region, &iq->response, &auth_fallback, iq->dp->name,
2282                 iq->dp->namelen)) {
2283                 /* use this as a response to be processed by the iterator */
2284                 if(verbosity >= VERB_ALGO) {
2285                         log_dns_msg("msg from auth zone",
2286                                 &iq->response->qinfo, iq->response->rep);
2287                 }
2288                 if((iq->chase_flags&BIT_RD) && !(iq->response->rep->flags&BIT_AA)) {
2289                         verbose(VERB_ALGO, "forwarder, ignoring referral from auth zone");
2290                 } else {
2291                         lock_rw_wrlock(&qstate->env->auth_zones->lock);
2292                         qstate->env->auth_zones->num_query_up++;
2293                         lock_rw_unlock(&qstate->env->auth_zones->lock);
2294                         iq->num_current_queries++;
2295                         iq->chase_to_rd = 0;
2296                         iq->dnssec_lame_query = 0;
2297                         iq->auth_zone_response = 1;
2298                         return next_state(iq, QUERY_RESP_STATE);
2299                 }
2300         }
2301         iq->auth_zone_response = 0;
2302         if(auth_fallback == 0) {
2303                 /* like we got servfail from the auth zone lookup, and
2304                  * no internet fallback */
2305                 verbose(VERB_ALGO, "auth zone lookup failed, no fallback,"
2306                         " servfail");
2307                 errinf(qstate, "auth zone lookup failed, fallback is off");
2308                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2309         }
2310         if(iq->dp->auth_dp) {
2311                 /* we wanted to fallback, but had no delegpt, only the
2312                  * auth zone generated delegpt, create an actual one */
2313                 iq->auth_zone_avoid = 1;
2314                 return next_state(iq, INIT_REQUEST_STATE);
2315         }
2316         /* but mostly, fallback==1 (like, when no such auth zone exists)
2317          * and we continue with lookups */
2318
2319         tf_policy = 0;
2320         /* < not <=, because although the array is large enough for <=, the
2321          * generated query will immediately be discarded due to depth and
2322          * that servfail is cached, which is not good as opportunism goes. */
2323         if(iq->depth < ie->max_dependency_depth
2324                 && iq->sent_count < TARGET_FETCH_STOP) {
2325                 tf_policy = ie->target_fetch_policy[iq->depth];
2326         }
2327
2328         /* if in 0x20 fallback get as many targets as possible */
2329         if(iq->caps_fallback) {
2330                 int extra = 0;
2331                 size_t naddr, nres, navail;
2332                 if(!query_for_targets(qstate, iq, ie, id, -1, &extra)) {
2333                         errinf(qstate, "could not fetch nameservers for 0x20 fallback");
2334                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2335                 }
2336                 iq->num_target_queries += extra;
2337                 target_count_increase(iq, extra);
2338                 if(iq->num_target_queries > 0) {
2339                         /* wait to get all targets, we want to try em */
2340                         verbose(VERB_ALGO, "wait for all targets for fallback");
2341                         qstate->ext_state[id] = module_wait_reply;
2342                         /* undo qname minimise step because we'll get back here
2343                          * to do it again */
2344                         if(qout_orig && iq->minimise_count > 0) {
2345                                 iq->minimise_count--;
2346                                 iq->qinfo_out.qname = qout_orig;
2347                                 iq->qinfo_out.qname_len = qout_orig_len;
2348                         }
2349                         return 0;
2350                 }
2351                 /* did we do enough fallback queries already? */
2352                 delegpt_count_addr(iq->dp, &naddr, &nres, &navail);
2353                 /* the current caps_server is the number of fallbacks sent.
2354                  * the original query is one that matched too, so we have
2355                  * caps_server+1 number of matching queries now */
2356                 if(iq->caps_server+1 >= naddr*3 ||
2357                         iq->caps_server*2+2 >= MAX_SENT_COUNT) {
2358                         /* *2 on sentcount check because ipv6 may fail */
2359                         /* we're done, process the response */
2360                         verbose(VERB_ALGO, "0x20 fallback had %d responses "
2361                                 "match for %d wanted, done.",
2362                                 (int)iq->caps_server+1, (int)naddr*3);
2363                         iq->response = iq->caps_response;
2364                         iq->caps_fallback = 0;
2365                         iter_dec_attempts(iq->dp, 3); /* space for fallback */
2366                         iq->num_current_queries++; /* RespState decrements it*/
2367                         iq->referral_count++; /* make sure we don't loop */
2368                         iq->sent_count = 0;
2369                         iq->state = QUERY_RESP_STATE;
2370                         return 1;
2371                 }
2372                 verbose(VERB_ALGO, "0x20 fallback number %d",
2373                         (int)iq->caps_server);
2374
2375         /* if there is a policy to fetch missing targets
2376          * opportunistically, do it. we rely on the fact that once a
2377          * query (or queries) for a missing name have been issued,
2378          * they will not show up again. */
2379         } else if(tf_policy != 0) {
2380                 int extra = 0;
2381                 verbose(VERB_ALGO, "attempt to get extra %d targets",
2382                         tf_policy);
2383                 (void)query_for_targets(qstate, iq, ie, id, tf_policy, &extra);
2384                 /* errors ignored, these targets are not strictly necessary for
2385                  * this result, we do not have to reply with SERVFAIL */
2386                 iq->num_target_queries += extra;
2387                 target_count_increase(iq, extra);
2388         }
2389
2390         /* Add the current set of unused targets to our queue. */
2391         delegpt_add_unused_targets(iq->dp);
2392
2393         /* Select the next usable target, filtering out unsuitable targets. */
2394         target = iter_server_selection(ie, qstate->env, iq->dp,
2395                 iq->dp->name, iq->dp->namelen, iq->qchase.qtype,
2396                 &iq->dnssec_lame_query, &iq->chase_to_rd,
2397                 iq->num_target_queries, qstate->blacklist,
2398                 qstate->prefetch_leeway);
2399
2400         /* If no usable target was selected... */
2401         if(!target) {
2402                 /* Here we distinguish between three states: generate a new
2403                  * target query, just wait, or quit (with a SERVFAIL).
2404                  * We have the following information: number of active
2405                  * target queries, number of active current queries,
2406                  * the presence of missing targets at this delegation
2407                  * point, and the given query target policy. */
2408
2409                 /* Check for the wait condition. If this is true, then
2410                  * an action must be taken. */
2411                 if(iq->num_target_queries==0 && iq->num_current_queries==0) {
2412                         /* If there is nothing to wait for, then we need
2413                          * to distinguish between generating (a) new target
2414                          * query, or failing. */
2415                         if(delegpt_count_missing_targets(iq->dp) > 0) {
2416                                 int qs = 0;
2417                                 verbose(VERB_ALGO, "querying for next "
2418                                         "missing target");
2419                                 if(!query_for_targets(qstate, iq, ie, id,
2420                                         1, &qs)) {
2421                                         errinf(qstate, "could not fetch nameserver");
2422                                         errinf_dname(qstate, "at zone", iq->dp->name);
2423                                         return error_response(qstate, id,
2424                                                 LDNS_RCODE_SERVFAIL);
2425                                 }
2426                                 if(qs == 0 &&
2427                                    delegpt_count_missing_targets(iq->dp) == 0){
2428                                         /* it looked like there were missing
2429                                          * targets, but they did not turn up.
2430                                          * Try the bad choices again (if any),
2431                                          * when we get back here missing==0,
2432                                          * so this is not a loop. */
2433                                         return 1;
2434                                 }
2435                                 iq->num_target_queries += qs;
2436                                 target_count_increase(iq, qs);
2437                         }
2438                         /* Since a target query might have been made, we
2439                          * need to check again. */
2440                         if(iq->num_target_queries == 0) {
2441                                 /* if in capsforid fallback, instead of last
2442                                  * resort, we agree with the current reply
2443                                  * we have (if any) (our count of addrs bad)*/
2444                                 if(iq->caps_fallback && iq->caps_reply) {
2445                                         /* we're done, process the response */
2446                                         verbose(VERB_ALGO, "0x20 fallback had %d responses, "
2447                                                 "but no more servers except "
2448                                                 "last resort, done.",
2449                                                 (int)iq->caps_server+1);
2450                                         iq->response = iq->caps_response;
2451                                         iq->caps_fallback = 0;
2452                                         iter_dec_attempts(iq->dp, 3); /* space for fallback */
2453                                         iq->num_current_queries++; /* RespState decrements it*/
2454                                         iq->referral_count++; /* make sure we don't loop */
2455                                         iq->sent_count = 0;
2456                                         iq->state = QUERY_RESP_STATE;
2457                                         return 1;
2458                                 }
2459                                 return processLastResort(qstate, iq, ie, id);
2460                         }
2461                 }
2462
2463                 /* otherwise, we have no current targets, so submerge
2464                  * until one of the target or direct queries return. */
2465                 if(iq->num_target_queries>0 && iq->num_current_queries>0) {
2466                         verbose(VERB_ALGO, "no current targets -- waiting "
2467                                 "for %d targets to resolve or %d outstanding"
2468                                 " queries to respond", iq->num_target_queries,
2469                                 iq->num_current_queries);
2470                         qstate->ext_state[id] = module_wait_reply;
2471                 } else if(iq->num_target_queries>0) {
2472                         verbose(VERB_ALGO, "no current targets -- waiting "
2473                                 "for %d targets to resolve.",
2474                                 iq->num_target_queries);
2475                         qstate->ext_state[id] = module_wait_subquery;
2476                 } else {
2477                         verbose(VERB_ALGO, "no current targets -- waiting "
2478                                 "for %d outstanding queries to respond.",
2479                                 iq->num_current_queries);
2480                         qstate->ext_state[id] = module_wait_reply;
2481                 }
2482                 /* undo qname minimise step because we'll get back here
2483                  * to do it again */
2484                 if(qout_orig && iq->minimise_count > 0) {
2485                         iq->minimise_count--;
2486                         iq->qinfo_out.qname = qout_orig;
2487                         iq->qinfo_out.qname_len = qout_orig_len;
2488                 }
2489                 return 0;
2490         }
2491
2492         /* if not forwarding, check ratelimits per delegationpoint name */
2493         if(!(iq->chase_flags & BIT_RD) && !iq->ratelimit_ok) {
2494                 if(!infra_ratelimit_inc(qstate->env->infra_cache, iq->dp->name,
2495                         iq->dp->namelen, *qstate->env->now, &qstate->qinfo,
2496                         qstate->reply)) {
2497                         lock_basic_lock(&ie->queries_ratelimit_lock);
2498                         ie->num_queries_ratelimited++;
2499                         lock_basic_unlock(&ie->queries_ratelimit_lock);
2500                         verbose(VERB_ALGO, "query exceeded ratelimits");
2501                         qstate->was_ratelimited = 1;
2502                         errinf_dname(qstate, "exceeded ratelimit for zone",
2503                                 iq->dp->name);
2504                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2505                 }
2506         }
2507
2508         /* We have a valid target. */
2509         if(verbosity >= VERB_QUERY) {
2510                 log_query_info(VERB_QUERY, "sending query:", &iq->qinfo_out);
2511                 log_name_addr(VERB_QUERY, "sending to target:", iq->dp->name,
2512                         &target->addr, target->addrlen);
2513                 verbose(VERB_ALGO, "dnssec status: %s%s",
2514                         iq->dnssec_expected?"expected": "not expected",
2515                         iq->dnssec_lame_query?" but lame_query anyway": "");
2516         }
2517         fptr_ok(fptr_whitelist_modenv_send_query(qstate->env->send_query));
2518         outq = (*qstate->env->send_query)(&iq->qinfo_out,
2519                 iq->chase_flags | (iq->chase_to_rd?BIT_RD:0),
2520                 /* unset CD if to forwarder(RD set) and not dnssec retry
2521                  * (blacklist nonempty) and no trust-anchors are configured
2522                  * above the qname or on the first attempt when dnssec is on */
2523                 EDNS_DO| ((iq->chase_to_rd||(iq->chase_flags&BIT_RD)!=0)&&
2524                 !qstate->blacklist&&(!iter_qname_indicates_dnssec(qstate->env,
2525                 &iq->qinfo_out)||target->attempts==1)?0:BIT_CD),
2526                 iq->dnssec_expected, iq->caps_fallback || is_caps_whitelisted(
2527                 ie, iq), &target->addr, target->addrlen,
2528                 iq->dp->name, iq->dp->namelen,
2529                 (iq->dp->ssl_upstream || qstate->env->cfg->ssl_upstream),
2530                 target->tls_auth_name, qstate);
2531         if(!outq) {
2532                 log_addr(VERB_DETAIL, "error sending query to auth server",
2533                         &target->addr, target->addrlen);
2534                 if(!(iq->chase_flags & BIT_RD) && !iq->ratelimit_ok)
2535                     infra_ratelimit_dec(qstate->env->infra_cache, iq->dp->name,
2536                         iq->dp->namelen, *qstate->env->now);
2537                 if(qstate->env->cfg->qname_minimisation)
2538                         iq->minimisation_state = SKIP_MINIMISE_STATE;
2539                 return next_state(iq, QUERYTARGETS_STATE);
2540         }
2541         outbound_list_insert(&iq->outlist, outq);
2542         iq->num_current_queries++;
2543         iq->sent_count++;
2544         qstate->ext_state[id] = module_wait_reply;
2545
2546         return 0;
2547 }
2548
2549 /** find NS rrset in given list */
2550 static struct ub_packed_rrset_key*
2551 find_NS(struct reply_info* rep, size_t from, size_t to)
2552 {
2553         size_t i;
2554         for(i=from; i<to; i++) {
2555                 if(ntohs(rep->rrsets[i]->rk.type) == LDNS_RR_TYPE_NS)
2556                         return rep->rrsets[i];
2557         }
2558         return NULL;
2559 }
2560
2561
2562 /**
2563  * Process the query response. All queries end up at this state first. This
2564  * process generally consists of analyzing the response and routing the
2565  * event to the next state (either bouncing it back to a request state, or
2566  * terminating the processing for this event).
2567  *
2568  * @param qstate: query state.
2569  * @param iq: iterator query state.
2570  * @param id: module id.
2571  * @return true if the event requires more immediate processing, false if
2572  *         not. This is generally only true when forwarding the request to
2573  *         the final state (i.e., on answer).
2574  */
2575 static int
2576 processQueryResponse(struct module_qstate* qstate, struct iter_qstate* iq,
2577         int id)
2578 {
2579         int dnsseclame = 0;
2580         enum response_type type;
2581         iq->num_current_queries--;
2582
2583         if(!inplace_cb_query_response_call(qstate->env, qstate, iq->response))
2584                 log_err("unable to call query_response callback");
2585
2586         if(iq->response == NULL) {
2587                 /* Don't increment qname when QNAME minimisation is enabled */
2588                 if(qstate->env->cfg->qname_minimisation) {
2589                         iq->minimise_timeout_count++;
2590                         iq->minimisation_state = SKIP_MINIMISE_STATE;
2591                 }
2592                 iq->chase_to_rd = 0;
2593                 iq->dnssec_lame_query = 0;
2594                 verbose(VERB_ALGO, "query response was timeout");
2595                 return next_state(iq, QUERYTARGETS_STATE);
2596         }
2597         type = response_type_from_server(
2598                 (int)((iq->chase_flags&BIT_RD) || iq->chase_to_rd),
2599                 iq->response, &iq->qinfo_out, iq->dp);
2600         iq->chase_to_rd = 0;
2601         if(type == RESPONSE_TYPE_REFERRAL && (iq->chase_flags&BIT_RD) &&
2602                 !iq->auth_zone_response) {
2603                 /* When forwarding (RD bit is set), we handle referrals
2604                  * differently. No queries should be sent elsewhere */
2605                 type = RESPONSE_TYPE_ANSWER;
2606         }
2607         if(!qstate->env->cfg->disable_dnssec_lame_check && iq->dnssec_expected
2608                 && !iq->dnssec_lame_query &&
2609                 !(iq->chase_flags&BIT_RD)
2610                 && iq->sent_count < DNSSEC_LAME_DETECT_COUNT
2611                 && type != RESPONSE_TYPE_LAME
2612                 && type != RESPONSE_TYPE_REC_LAME
2613                 && type != RESPONSE_TYPE_THROWAWAY
2614                 && type != RESPONSE_TYPE_UNTYPED) {
2615                 /* a possible answer, see if it is missing DNSSEC */
2616                 /* but not when forwarding, so we dont mark fwder lame */
2617                 if(!iter_msg_has_dnssec(iq->response)) {
2618                         /* Mark this address as dnsseclame in this dp,
2619                          * because that will make serverselection disprefer
2620                          * it, but also, once it is the only final option,
2621                          * use dnssec-lame-bypass if it needs to query there.*/
2622                         if(qstate->reply) {
2623                                 struct delegpt_addr* a = delegpt_find_addr(
2624                                         iq->dp, &qstate->reply->addr,
2625                                         qstate->reply->addrlen);
2626                                 if(a) a->dnsseclame = 1;
2627                         }
2628                         /* test the answer is from the zone we expected,
2629                          * otherwise, (due to parent,child on same server), we
2630                          * might mark the server,zone lame inappropriately */
2631                         if(!iter_msg_from_zone(iq->response, iq->dp, type,
2632                                 iq->qchase.qclass))
2633                                 qstate->reply = NULL;
2634                         type = RESPONSE_TYPE_LAME;
2635                         dnsseclame = 1;
2636                 }
2637         } else iq->dnssec_lame_query = 0;
2638         /* see if referral brings us close to the target */
2639         if(type == RESPONSE_TYPE_REFERRAL) {
2640                 struct ub_packed_rrset_key* ns = find_NS(
2641                         iq->response->rep, iq->response->rep->an_numrrsets,
2642                         iq->response->rep->an_numrrsets
2643                         + iq->response->rep->ns_numrrsets);
2644                 if(!ns) ns = find_NS(iq->response->rep, 0,
2645                                 iq->response->rep->an_numrrsets);
2646                 if(!ns || !dname_strict_subdomain_c(ns->rk.dname, iq->dp->name)
2647                         || !dname_subdomain_c(iq->qchase.qname, ns->rk.dname)){
2648                         verbose(VERB_ALGO, "bad referral, throwaway");
2649                         type = RESPONSE_TYPE_THROWAWAY;
2650                 } else
2651                         iter_scrub_ds(iq->response, ns, iq->dp->name);
2652         } else iter_scrub_ds(iq->response, NULL, NULL);
2653         if(type == RESPONSE_TYPE_THROWAWAY &&
2654                 FLAGS_GET_RCODE(iq->response->rep->flags) == LDNS_RCODE_YXDOMAIN) {
2655                 /* YXDOMAIN is a permanent error, no need to retry */
2656                 type = RESPONSE_TYPE_ANSWER;
2657         }
2658         if(type == RESPONSE_TYPE_CNAME && iq->response->rep->an_numrrsets >= 1
2659                 && ntohs(iq->response->rep->rrsets[0]->rk.type) == LDNS_RR_TYPE_DNAME) {
2660                 uint8_t* sname = NULL;
2661                 size_t snamelen = 0;
2662                 get_cname_target(iq->response->rep->rrsets[0], &sname,
2663                         &snamelen);
2664                 if(snamelen && dname_subdomain_c(sname, iq->response->rep->rrsets[0]->rk.dname)) {
2665                         /* DNAME to a subdomain loop; do not recurse */
2666                         type = RESPONSE_TYPE_ANSWER;
2667                 }
2668         } else if(type == RESPONSE_TYPE_CNAME &&
2669                 iq->qchase.qtype == LDNS_RR_TYPE_CNAME &&
2670                 iq->minimisation_state == MINIMISE_STATE &&
2671                 query_dname_compare(iq->qchase.qname, iq->qinfo_out.qname) == 0) {
2672                 /* The minimised query for full QTYPE and hidden QTYPE can be
2673                  * classified as CNAME response type, even when the original
2674                  * QTYPE=CNAME. This should be treated as answer response type.
2675                  */
2676                 type = RESPONSE_TYPE_ANSWER;
2677         }
2678
2679         /* handle each of the type cases */
2680         if(type == RESPONSE_TYPE_ANSWER) {
2681                 /* ANSWER type responses terminate the query algorithm,
2682                  * so they sent on their */
2683                 if(verbosity >= VERB_DETAIL) {
2684                         verbose(VERB_DETAIL, "query response was %s",
2685                                 FLAGS_GET_RCODE(iq->response->rep->flags)
2686                                 ==LDNS_RCODE_NXDOMAIN?"NXDOMAIN ANSWER":
2687                                 (iq->response->rep->an_numrrsets?"ANSWER":
2688                                 "nodata ANSWER"));
2689                 }
2690                 /* if qtype is DS, check we have the right level of answer,
2691                  * like grandchild answer but we need the middle, reject it */
2692                 if(iq->qchase.qtype == LDNS_RR_TYPE_DS && !iq->dsns_point
2693                         && !(iq->chase_flags&BIT_RD)
2694                         && iter_ds_toolow(iq->response, iq->dp)
2695                         && iter_dp_cangodown(&iq->qchase, iq->dp)) {
2696                         /* close down outstanding requests to be discarded */
2697                         outbound_list_clear(&iq->outlist);
2698                         iq->num_current_queries = 0;
2699                         fptr_ok(fptr_whitelist_modenv_detach_subs(
2700                                 qstate->env->detach_subs));
2701                         (*qstate->env->detach_subs)(qstate);
2702                         iq->num_target_queries = 0;
2703                         return processDSNSFind(qstate, iq, id);
2704                 }
2705                 if(!qstate->no_cache_store)
2706                         iter_dns_store(qstate->env, &iq->response->qinfo,
2707                                 iq->response->rep, 0, qstate->prefetch_leeway,
2708                                 iq->dp&&iq->dp->has_parent_side_NS,
2709                                 qstate->region, qstate->query_flags);
2710                 /* close down outstanding requests to be discarded */
2711                 outbound_list_clear(&iq->outlist);
2712                 iq->num_current_queries = 0;
2713                 fptr_ok(fptr_whitelist_modenv_detach_subs(
2714                         qstate->env->detach_subs));
2715                 (*qstate->env->detach_subs)(qstate);
2716                 iq->num_target_queries = 0;
2717                 if(qstate->reply)
2718                         sock_list_insert(&qstate->reply_origin,
2719                                 &qstate->reply->addr, qstate->reply->addrlen,
2720                                 qstate->region);
2721                 if(iq->minimisation_state != DONOT_MINIMISE_STATE
2722                         && !(iq->chase_flags & BIT_RD)) {
2723                         if(FLAGS_GET_RCODE(iq->response->rep->flags) !=
2724                                 LDNS_RCODE_NOERROR) {
2725                                 if(qstate->env->cfg->qname_minimisation_strict) {
2726                                         if(FLAGS_GET_RCODE(iq->response->rep->flags) ==
2727                                                 LDNS_RCODE_NXDOMAIN) {
2728                                                 iter_scrub_nxdomain(iq->response);
2729                                                 return final_state(iq);
2730                                         }
2731                                         return error_response(qstate, id,
2732                                                 LDNS_RCODE_SERVFAIL);
2733                                 }
2734                                 /* Best effort qname-minimisation.
2735                                  * Stop minimising and send full query when
2736                                  * RCODE is not NOERROR. */
2737                                 iq->minimisation_state = DONOT_MINIMISE_STATE;
2738                         }
2739                         if(FLAGS_GET_RCODE(iq->response->rep->flags) ==
2740                                 LDNS_RCODE_NXDOMAIN) {
2741                                 /* Stop resolving when NXDOMAIN is DNSSEC
2742                                  * signed. Based on assumption that nameservers
2743                                  * serving signed zones do not return NXDOMAIN
2744                                  * for empty-non-terminals. */
2745                                 if(iq->dnssec_expected)
2746                                         return final_state(iq);
2747                                 /* Make subrequest to validate intermediate
2748                                  * NXDOMAIN if harden-below-nxdomain is
2749                                  * enabled. */
2750                                 if(qstate->env->cfg->harden_below_nxdomain) {
2751                                         struct module_qstate* subq = NULL;
2752                                         log_query_info(VERB_QUERY,
2753                                                 "schedule NXDOMAIN validation:",
2754                                                 &iq->response->qinfo);
2755                                         if(!generate_sub_request(
2756                                                 iq->response->qinfo.qname,
2757                                                 iq->response->qinfo.qname_len,
2758                                                 iq->response->qinfo.qtype,
2759                                                 iq->response->qinfo.qclass,
2760                                                 qstate, id, iq,
2761                                                 INIT_REQUEST_STATE,
2762                                                 FINISHED_STATE, &subq, 1))
2763                                                 verbose(VERB_ALGO,
2764                                                 "could not validate NXDOMAIN "
2765                                                 "response");
2766                                         outbound_list_clear(&iq->outlist);
2767                                         iq->num_current_queries = 0;
2768                                         fptr_ok(fptr_whitelist_modenv_detach_subs(
2769                                                 qstate->env->detach_subs));
2770                                         (*qstate->env->detach_subs)(qstate);
2771                                         iq->num_target_queries = 0;
2772                                 }
2773                         }
2774                         return next_state(iq, QUERYTARGETS_STATE);
2775                 }
2776                 return final_state(iq);
2777         } else if(type == RESPONSE_TYPE_REFERRAL) {
2778                 /* REFERRAL type responses get a reset of the
2779                  * delegation point, and back to the QUERYTARGETS_STATE. */
2780                 verbose(VERB_DETAIL, "query response was REFERRAL");
2781
2782                 if(!(iq->chase_flags & BIT_RD) && !iq->ratelimit_ok) {
2783                         /* we have a referral, no ratelimit, we can send
2784                          * our queries to the given name */
2785                         infra_ratelimit_dec(qstate->env->infra_cache,
2786                                 iq->dp->name, iq->dp->namelen,
2787                                 *qstate->env->now);
2788                 }
2789
2790                 /* if hardened, only store referral if we asked for it */
2791                 if(!qstate->no_cache_store &&
2792                 (!qstate->env->cfg->harden_referral_path ||
2793                     (  qstate->qinfo.qtype == LDNS_RR_TYPE_NS
2794                         && (qstate->query_flags&BIT_RD)
2795                         && !(qstate->query_flags&BIT_CD)
2796                            /* we know that all other NS rrsets are scrubbed
2797                             * away, thus on referral only one is left.
2798                             * see if that equals the query name... */
2799                         && ( /* auth section, but sometimes in answer section*/
2800                           reply_find_rrset_section_ns(iq->response->rep,
2801                                 iq->qchase.qname, iq->qchase.qname_len,
2802                                 LDNS_RR_TYPE_NS, iq->qchase.qclass)
2803                           || reply_find_rrset_section_an(iq->response->rep,
2804                                 iq->qchase.qname, iq->qchase.qname_len,
2805                                 LDNS_RR_TYPE_NS, iq->qchase.qclass)
2806                           )
2807                     ))) {
2808                         /* Store the referral under the current query */
2809                         /* no prefetch-leeway, since its not the answer */
2810                         iter_dns_store(qstate->env, &iq->response->qinfo,
2811                                 iq->response->rep, 1, 0, 0, NULL, 0);
2812                         if(iq->store_parent_NS)
2813                                 iter_store_parentside_NS(qstate->env,
2814                                         iq->response->rep);
2815                         if(qstate->env->neg_cache)
2816                                 val_neg_addreferral(qstate->env->neg_cache,
2817                                         iq->response->rep, iq->dp->name);
2818                 }
2819                 /* store parent-side-in-zone-glue, if directly queried for */
2820                 if(!qstate->no_cache_store && iq->query_for_pside_glue
2821                         && !iq->pside_glue) {
2822                                 iq->pside_glue = reply_find_rrset(iq->response->rep,
2823                                         iq->qchase.qname, iq->qchase.qname_len,
2824                                         iq->qchase.qtype, iq->qchase.qclass);
2825                                 if(iq->pside_glue) {
2826                                         log_rrset_key(VERB_ALGO, "found parent-side "
2827                                                 "glue", iq->pside_glue);
2828                                         iter_store_parentside_rrset(qstate->env,
2829                                                 iq->pside_glue);
2830                                 }
2831                 }
2832
2833                 /* Reset the event state, setting the current delegation
2834                  * point to the referral. */
2835                 iq->deleg_msg = iq->response;
2836                 iq->dp = delegpt_from_message(iq->response, qstate->region);
2837                 if (qstate->env->cfg->qname_minimisation)
2838                         iq->minimisation_state = INIT_MINIMISE_STATE;
2839                 if(!iq->dp) {
2840                         errinf(qstate, "malloc failure, for delegation point");
2841                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2842                 }
2843                 if(!cache_fill_missing(qstate->env, iq->qchase.qclass,
2844                         qstate->region, iq->dp)) {
2845                         errinf(qstate, "malloc failure, copy extra info into delegation point");
2846                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2847                 }
2848                 if(iq->store_parent_NS && query_dname_compare(iq->dp->name,
2849                         iq->store_parent_NS->name) == 0)
2850                         iter_merge_retry_counts(iq->dp, iq->store_parent_NS);
2851                 delegpt_log(VERB_ALGO, iq->dp);
2852                 /* Count this as a referral. */
2853                 iq->referral_count++;
2854                 iq->sent_count = 0;
2855                 /* see if the next dp is a trust anchor, or a DS was sent
2856                  * along, indicating dnssec is expected for next zone */
2857                 iq->dnssec_expected = iter_indicates_dnssec(qstate->env,
2858                         iq->dp, iq->response, iq->qchase.qclass);
2859                 /* if dnssec, validating then also fetch the key for the DS */
2860                 if(iq->dnssec_expected && qstate->env->cfg->prefetch_key &&
2861                         !(qstate->query_flags&BIT_CD))
2862                         generate_dnskey_prefetch(qstate, iq, id);
2863
2864                 /* spawn off NS and addr to auth servers for the NS we just
2865                  * got in the referral. This gets authoritative answer
2866                  * (answer section trust level) rrset.
2867                  * right after, we detach the subs, answer goes to cache. */
2868                 if(qstate->env->cfg->harden_referral_path)
2869                         generate_ns_check(qstate, iq, id);
2870
2871                 /* stop current outstanding queries.
2872                  * FIXME: should the outstanding queries be waited for and
2873                  * handled? Say by a subquery that inherits the outbound_entry.
2874                  */
2875                 outbound_list_clear(&iq->outlist);
2876                 iq->num_current_queries = 0;
2877                 fptr_ok(fptr_whitelist_modenv_detach_subs(
2878                         qstate->env->detach_subs));
2879                 (*qstate->env->detach_subs)(qstate);
2880                 iq->num_target_queries = 0;
2881                 verbose(VERB_ALGO, "cleared outbound list for next round");
2882                 return next_state(iq, QUERYTARGETS_STATE);
2883         } else if(type == RESPONSE_TYPE_CNAME) {
2884                 uint8_t* sname = NULL;
2885                 size_t snamelen = 0;
2886                 /* CNAME type responses get a query restart (i.e., get a
2887                  * reset of the query state and go back to INIT_REQUEST_STATE).
2888                  */
2889                 verbose(VERB_DETAIL, "query response was CNAME");
2890                 if(verbosity >= VERB_ALGO)
2891                         log_dns_msg("cname msg", &iq->response->qinfo,
2892                                 iq->response->rep);
2893                 /* if qtype is DS, check we have the right level of answer,
2894                  * like grandchild answer but we need the middle, reject it */
2895                 if(iq->qchase.qtype == LDNS_RR_TYPE_DS && !iq->dsns_point
2896                         && !(iq->chase_flags&BIT_RD)
2897                         && iter_ds_toolow(iq->response, iq->dp)
2898                         && iter_dp_cangodown(&iq->qchase, iq->dp)) {
2899                         outbound_list_clear(&iq->outlist);
2900                         iq->num_current_queries = 0;
2901                         fptr_ok(fptr_whitelist_modenv_detach_subs(
2902                                 qstate->env->detach_subs));
2903                         (*qstate->env->detach_subs)(qstate);
2904                         iq->num_target_queries = 0;
2905                         return processDSNSFind(qstate, iq, id);
2906                 }
2907                 /* Process the CNAME response. */
2908                 if(!handle_cname_response(qstate, iq, iq->response,
2909                         &sname, &snamelen)) {
2910                         errinf(qstate, "malloc failure, CNAME info");
2911                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
2912                 }
2913                 /* cache the CNAME response under the current query */
2914                 /* NOTE : set referral=1, so that rrsets get stored but not
2915                  * the partial query answer (CNAME only). */
2916                 /* prefetchleeway applied because this updates answer parts */
2917                 if(!qstate->no_cache_store)
2918                         iter_dns_store(qstate->env, &iq->response->qinfo,
2919                                 iq->response->rep, 1, qstate->prefetch_leeway,
2920                                 iq->dp&&iq->dp->has_parent_side_NS, NULL,
2921                                 qstate->query_flags);
2922                 /* set the current request's qname to the new value. */
2923                 iq->qchase.qname = sname;
2924                 iq->qchase.qname_len = snamelen;
2925                 /* Clear the query state, since this is a query restart. */
2926                 iq->deleg_msg = NULL;
2927                 iq->dp = NULL;
2928                 iq->dsns_point = NULL;
2929                 iq->auth_zone_response = 0;
2930                 iq->sent_count = 0;
2931                 if(iq->minimisation_state != MINIMISE_STATE)
2932                         /* Only count as query restart when it is not an extra
2933                          * query as result of qname minimisation. */
2934                         iq->query_restart_count++;
2935                 if(qstate->env->cfg->qname_minimisation)
2936                         iq->minimisation_state = INIT_MINIMISE_STATE;
2937
2938                 /* stop current outstanding queries.
2939                  * FIXME: should the outstanding queries be waited for and
2940                  * handled? Say by a subquery that inherits the outbound_entry.
2941                  */
2942                 outbound_list_clear(&iq->outlist);
2943                 iq->num_current_queries = 0;
2944                 fptr_ok(fptr_whitelist_modenv_detach_subs(
2945                         qstate->env->detach_subs));
2946                 (*qstate->env->detach_subs)(qstate);
2947                 iq->num_target_queries = 0;
2948                 if(qstate->reply)
2949                         sock_list_insert(&qstate->reply_origin,
2950                                 &qstate->reply->addr, qstate->reply->addrlen,
2951                                 qstate->region);
2952                 verbose(VERB_ALGO, "cleared outbound list for query restart");
2953                 /* go to INIT_REQUEST_STATE for new qname. */
2954                 return next_state(iq, INIT_REQUEST_STATE);
2955         } else if(type == RESPONSE_TYPE_LAME) {
2956                 /* Cache the LAMEness. */
2957                 verbose(VERB_DETAIL, "query response was %sLAME",
2958                         dnsseclame?"DNSSEC ":"");
2959                 if(!dname_subdomain_c(iq->qchase.qname, iq->dp->name)) {
2960                         log_err("mark lame: mismatch in qname and dpname");
2961                         /* throwaway this reply below */
2962                 } else if(qstate->reply) {
2963                         /* need addr for lameness cache, but we may have
2964                          * gotten this from cache, so test to be sure */
2965                         if(!infra_set_lame(qstate->env->infra_cache,
2966                                 &qstate->reply->addr, qstate->reply->addrlen,
2967                                 iq->dp->name, iq->dp->namelen,
2968                                 *qstate->env->now, dnsseclame, 0,
2969                                 iq->qchase.qtype))
2970                                 log_err("mark host lame: out of memory");
2971                 }
2972         } else if(type == RESPONSE_TYPE_REC_LAME) {
2973                 /* Cache the LAMEness. */
2974                 verbose(VERB_DETAIL, "query response REC_LAME: "
2975                         "recursive but not authoritative server");
2976                 if(!dname_subdomain_c(iq->qchase.qname, iq->dp->name)) {
2977                         log_err("mark rec_lame: mismatch in qname and dpname");
2978                         /* throwaway this reply below */
2979                 } else if(qstate->reply) {
2980                         /* need addr for lameness cache, but we may have
2981                          * gotten this from cache, so test to be sure */
2982                         verbose(VERB_DETAIL, "mark as REC_LAME");
2983                         if(!infra_set_lame(qstate->env->infra_cache,
2984                                 &qstate->reply->addr, qstate->reply->addrlen,
2985                                 iq->dp->name, iq->dp->namelen,
2986                                 *qstate->env->now, 0, 1, iq->qchase.qtype))
2987                                 log_err("mark host lame: out of memory");
2988                 }
2989         } else if(type == RESPONSE_TYPE_THROWAWAY) {
2990                 /* LAME and THROWAWAY responses are handled the same way.
2991                  * In this case, the event is just sent directly back to
2992                  * the QUERYTARGETS_STATE without resetting anything,
2993                  * because, clearly, the next target must be tried. */
2994                 verbose(VERB_DETAIL, "query response was THROWAWAY");
2995         } else {
2996                 log_warn("A query response came back with an unknown type: %d",
2997                         (int)type);
2998         }
2999
3000         /* LAME, THROWAWAY and "unknown" all end up here.
3001          * Recycle to the QUERYTARGETS state to hopefully try a
3002          * different target. */
3003         if (qstate->env->cfg->qname_minimisation &&
3004                 !qstate->env->cfg->qname_minimisation_strict)
3005                 iq->minimisation_state = DONOT_MINIMISE_STATE;
3006         if(iq->auth_zone_response) {
3007                 /* can we fallback? */
3008                 iq->auth_zone_response = 0;
3009                 if(!auth_zones_can_fallback(qstate->env->auth_zones,
3010                         iq->dp->name, iq->dp->namelen, qstate->qinfo.qclass)) {
3011                         verbose(VERB_ALGO, "auth zone response bad, and no"
3012                                 " fallback possible, servfail");
3013                         errinf_dname(qstate, "response is bad, no fallback, "
3014                                 "for auth zone", iq->dp->name);
3015                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3016                 }
3017                 verbose(VERB_ALGO, "auth zone response was bad, "
3018                         "fallback enabled");
3019                 iq->auth_zone_avoid = 1;
3020                 if(iq->dp->auth_dp) {
3021                         /* we are using a dp for the auth zone, with no
3022                          * nameservers, get one first */
3023                         iq->dp = NULL;
3024                         return next_state(iq, INIT_REQUEST_STATE);
3025                 }
3026         }
3027         return next_state(iq, QUERYTARGETS_STATE);
3028 }
3029
3030 /**
3031  * Return priming query results to interested super querystates.
3032  *
3033  * Sets the delegation point and delegation message (not nonRD queries).
3034  * This is a callback from walk_supers.
3035  *
3036  * @param qstate: priming query state that finished.
3037  * @param id: module id.
3038  * @param forq: the qstate for which priming has been done.
3039  */
3040 static void
3041 prime_supers(struct module_qstate* qstate, int id, struct module_qstate* forq)
3042 {
3043         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
3044         struct delegpt* dp = NULL;
3045
3046         log_assert(qstate->is_priming || foriq->wait_priming_stub);
3047         log_assert(qstate->return_rcode == LDNS_RCODE_NOERROR);
3048         /* Convert our response to a delegation point */
3049         dp = delegpt_from_message(qstate->return_msg, forq->region);
3050         if(!dp) {
3051                 /* if there is no convertable delegation point, then
3052                  * the ANSWER type was (presumably) a negative answer. */
3053                 verbose(VERB_ALGO, "prime response was not a positive "
3054                         "ANSWER; failing");
3055                 foriq->dp = NULL;
3056                 foriq->state = QUERYTARGETS_STATE;
3057                 return;
3058         }
3059
3060         log_query_info(VERB_DETAIL, "priming successful for", &qstate->qinfo);
3061         delegpt_log(VERB_ALGO, dp);
3062         foriq->dp = dp;
3063         foriq->deleg_msg = dns_copy_msg(qstate->return_msg, forq->region);
3064         if(!foriq->deleg_msg) {
3065                 log_err("copy prime response: out of memory");
3066                 foriq->dp = NULL;
3067                 foriq->state = QUERYTARGETS_STATE;
3068                 return;
3069         }
3070
3071         /* root priming responses go to init stage 2, priming stub
3072          * responses to to stage 3. */
3073         if(foriq->wait_priming_stub) {
3074                 foriq->state = INIT_REQUEST_3_STATE;
3075                 foriq->wait_priming_stub = 0;
3076         } else  foriq->state = INIT_REQUEST_2_STATE;
3077         /* because we are finished, the parent will be reactivated */
3078 }
3079
3080 /**
3081  * This handles the response to a priming query. This is used to handle both
3082  * root and stub priming responses. This is basically the equivalent of the
3083  * QUERY_RESP_STATE, but will not handle CNAME responses and will treat
3084  * REFERRALs as ANSWERS. It will also update and reactivate the originating
3085  * event.
3086  *
3087  * @param qstate: query state.
3088  * @param id: module id.
3089  * @return true if the event needs more immediate processing, false if not.
3090  *         This state always returns false.
3091  */
3092 static int
3093 processPrimeResponse(struct module_qstate* qstate, int id)
3094 {
3095         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
3096         enum response_type type;
3097         iq->response->rep->flags &= ~(BIT_RD|BIT_RA); /* ignore rec-lame */
3098         type = response_type_from_server(
3099                 (int)((iq->chase_flags&BIT_RD) || iq->chase_to_rd),
3100                 iq->response, &iq->qchase, iq->dp);
3101         if(type == RESPONSE_TYPE_ANSWER) {
3102                 qstate->return_rcode = LDNS_RCODE_NOERROR;
3103                 qstate->return_msg = iq->response;
3104         } else {
3105                 errinf(qstate, "prime response did not get an answer");
3106                 errinf_dname(qstate, "for", qstate->qinfo.qname);
3107                 qstate->return_rcode = LDNS_RCODE_SERVFAIL;
3108                 qstate->return_msg = NULL;
3109         }
3110
3111         /* validate the root or stub after priming (if enabled).
3112          * This is the same query as the prime query, but with validation.
3113          * Now that we are primed, the additional queries that validation
3114          * may need can be resolved, such as DLV. */
3115         if(qstate->env->cfg->harden_referral_path) {
3116                 struct module_qstate* subq = NULL;
3117                 log_nametypeclass(VERB_ALGO, "schedule prime validation",
3118                         qstate->qinfo.qname, qstate->qinfo.qtype,
3119                         qstate->qinfo.qclass);
3120                 if(!generate_sub_request(qstate->qinfo.qname,
3121                         qstate->qinfo.qname_len, qstate->qinfo.qtype,
3122                         qstate->qinfo.qclass, qstate, id, iq,
3123                         INIT_REQUEST_STATE, FINISHED_STATE, &subq, 1)) {
3124                         verbose(VERB_ALGO, "could not generate prime check");
3125                 }
3126                 generate_a_aaaa_check(qstate, iq, id);
3127         }
3128
3129         /* This event is finished. */
3130         qstate->ext_state[id] = module_finished;
3131         return 0;
3132 }
3133
3134 /**
3135  * Do final processing on responses to target queries. Events reach this
3136  * state after the iterative resolution algorithm terminates. This state is
3137  * responsible for reactivating the original event, and housekeeping related
3138  * to received target responses (caching, updating the current delegation
3139  * point, etc).
3140  * Callback from walk_supers for every super state that is interested in
3141  * the results from this query.
3142  *
3143  * @param qstate: query state.
3144  * @param id: module id.
3145  * @param forq: super query state.
3146  */
3147 static void
3148 processTargetResponse(struct module_qstate* qstate, int id,
3149         struct module_qstate* forq)
3150 {
3151         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
3152         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
3153         struct ub_packed_rrset_key* rrset;
3154         struct delegpt_ns* dpns;
3155         log_assert(qstate->return_rcode == LDNS_RCODE_NOERROR);
3156
3157         foriq->state = QUERYTARGETS_STATE;
3158         log_query_info(VERB_ALGO, "processTargetResponse", &qstate->qinfo);
3159         log_query_info(VERB_ALGO, "processTargetResponse super", &forq->qinfo);
3160
3161         /* Tell the originating event that this target query has finished
3162          * (regardless if it succeeded or not). */
3163         foriq->num_target_queries--;
3164
3165         /* check to see if parent event is still interested (in orig name).  */
3166         if(!foriq->dp) {
3167                 verbose(VERB_ALGO, "subq: parent not interested, was reset");
3168                 return; /* not interested anymore */
3169         }
3170         dpns = delegpt_find_ns(foriq->dp, qstate->qinfo.qname,
3171                         qstate->qinfo.qname_len);
3172         if(!dpns) {
3173                 /* If not interested, just stop processing this event */
3174                 verbose(VERB_ALGO, "subq: parent not interested anymore");
3175                 /* could be because parent was jostled out of the cache,
3176                    and a new identical query arrived, that does not want it*/
3177                 return;
3178         }
3179
3180         /* if iq->query_for_pside_glue then add the pside_glue (marked lame) */
3181         if(iq->pside_glue) {
3182                 /* if the pside_glue is NULL, then it could not be found,
3183                  * the done_pside is already set when created and a cache
3184                  * entry created in processFinished so nothing to do here */
3185                 log_rrset_key(VERB_ALGO, "add parentside glue to dp",
3186                         iq->pside_glue);
3187                 if(!delegpt_add_rrset(foriq->dp, forq->region,
3188                         iq->pside_glue, 1))
3189                         log_err("out of memory adding pside glue");
3190         }
3191
3192         /* This response is relevant to the current query, so we
3193          * add (attempt to add, anyway) this target(s) and reactivate
3194          * the original event.
3195          * NOTE: we could only look for the AnswerRRset if the
3196          * response type was ANSWER. */
3197         rrset = reply_find_answer_rrset(&iq->qchase, qstate->return_msg->rep);
3198         if(rrset) {
3199                 /* if CNAMEs have been followed - add new NS to delegpt. */
3200                 /* BTW. RFC 1918 says NS should not have got CNAMEs. Robust. */
3201                 if(!delegpt_find_ns(foriq->dp, rrset->rk.dname,
3202                         rrset->rk.dname_len)) {
3203                         /* if dpns->lame then set newcname ns lame too */
3204                         if(!delegpt_add_ns(foriq->dp, forq->region,
3205                                 rrset->rk.dname, dpns->lame))
3206                                 log_err("out of memory adding cnamed-ns");
3207                 }
3208                 /* if dpns->lame then set the address(es) lame too */
3209                 if(!delegpt_add_rrset(foriq->dp, forq->region, rrset,
3210                         dpns->lame))
3211                         log_err("out of memory adding targets");
3212                 verbose(VERB_ALGO, "added target response");
3213                 delegpt_log(VERB_ALGO, foriq->dp);
3214         } else {
3215                 verbose(VERB_ALGO, "iterator TargetResponse failed");
3216                 dpns->resolved = 1; /* fail the target */
3217         }
3218 }
3219
3220 /**
3221  * Process response for DS NS Find queries, that attempt to find the delegation
3222  * point where we ask the DS query from.
3223  *
3224  * @param qstate: query state.
3225  * @param id: module id.
3226  * @param forq: super query state.
3227  */
3228 static void
3229 processDSNSResponse(struct module_qstate* qstate, int id,
3230         struct module_qstate* forq)
3231 {
3232         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
3233
3234         /* if the finished (iq->response) query has no NS set: continue
3235          * up to look for the right dp; nothing to change, do DPNSstate */
3236         if(qstate->return_rcode != LDNS_RCODE_NOERROR)
3237                 return; /* seek further */
3238         /* find the NS RRset (without allowing CNAMEs) */
3239         if(!reply_find_rrset(qstate->return_msg->rep, qstate->qinfo.qname,
3240                 qstate->qinfo.qname_len, LDNS_RR_TYPE_NS,
3241                 qstate->qinfo.qclass)){
3242                 return; /* seek further */
3243         }
3244
3245         /* else, store as DP and continue at querytargets */
3246         foriq->state = QUERYTARGETS_STATE;
3247         foriq->dp = delegpt_from_message(qstate->return_msg, forq->region);
3248         if(!foriq->dp) {
3249                 log_err("out of memory in dsns dp alloc");
3250                 errinf(qstate, "malloc failure, in DS search");
3251                 return; /* dp==NULL in QUERYTARGETS makes SERVFAIL */
3252         }
3253         /* success, go query the querytargets in the new dp (and go down) */
3254 }
3255
3256 /**
3257  * Process response for qclass=ANY queries for a particular class.
3258  * Append to result or error-exit.
3259  *
3260  * @param qstate: query state.
3261  * @param id: module id.
3262  * @param forq: super query state.
3263  */
3264 static void
3265 processClassResponse(struct module_qstate* qstate, int id,
3266         struct module_qstate* forq)
3267 {
3268         struct iter_qstate* foriq = (struct iter_qstate*)forq->minfo[id];
3269         struct dns_msg* from = qstate->return_msg;
3270         log_query_info(VERB_ALGO, "processClassResponse", &qstate->qinfo);
3271         log_query_info(VERB_ALGO, "processClassResponse super", &forq->qinfo);
3272         if(qstate->return_rcode != LDNS_RCODE_NOERROR) {
3273                 /* cause servfail for qclass ANY query */
3274                 foriq->response = NULL;
3275                 foriq->state = FINISHED_STATE;
3276                 return;
3277         }
3278         /* append result */
3279         if(!foriq->response) {
3280                 /* allocate the response: copy RCODE, sec_state */
3281                 foriq->response = dns_copy_msg(from, forq->region);
3282                 if(!foriq->response) {
3283                         log_err("malloc failed for qclass ANY response");
3284                         foriq->state = FINISHED_STATE;
3285                         return;
3286                 }
3287                 foriq->response->qinfo.qclass = forq->qinfo.qclass;
3288                 /* qclass ANY does not receive the AA flag on replies */
3289                 foriq->response->rep->authoritative = 0;
3290         } else {
3291                 struct dns_msg* to = foriq->response;
3292                 /* add _from_ this response _to_ existing collection */
3293                 /* if there are records, copy RCODE */
3294                 /* lower sec_state if this message is lower */
3295                 if(from->rep->rrset_count != 0) {
3296                         size_t n = from->rep->rrset_count+to->rep->rrset_count;
3297                         struct ub_packed_rrset_key** dest, **d;
3298                         /* copy appropriate rcode */
3299                         to->rep->flags = from->rep->flags;
3300                         /* copy rrsets */
3301                         if(from->rep->rrset_count > RR_COUNT_MAX ||
3302                                 to->rep->rrset_count > RR_COUNT_MAX) {
3303                                 log_err("malloc failed (too many rrsets) in collect ANY");
3304                                 foriq->state = FINISHED_STATE;
3305                                 return; /* integer overflow protection */
3306                         }
3307                         dest = regional_alloc(forq->region, sizeof(dest[0])*n);
3308                         if(!dest) {
3309                                 log_err("malloc failed in collect ANY");
3310                                 foriq->state = FINISHED_STATE;
3311                                 return;
3312                         }
3313                         d = dest;
3314                         /* copy AN */
3315                         memcpy(dest, to->rep->rrsets, to->rep->an_numrrsets
3316                                 * sizeof(dest[0]));
3317                         dest += to->rep->an_numrrsets;
3318                         memcpy(dest, from->rep->rrsets, from->rep->an_numrrsets
3319                                 * sizeof(dest[0]));
3320                         dest += from->rep->an_numrrsets;
3321                         /* copy NS */
3322                         memcpy(dest, to->rep->rrsets+to->rep->an_numrrsets,
3323                                 to->rep->ns_numrrsets * sizeof(dest[0]));
3324                         dest += to->rep->ns_numrrsets;
3325                         memcpy(dest, from->rep->rrsets+from->rep->an_numrrsets,
3326                                 from->rep->ns_numrrsets * sizeof(dest[0]));
3327                         dest += from->rep->ns_numrrsets;
3328                         /* copy AR */
3329                         memcpy(dest, to->rep->rrsets+to->rep->an_numrrsets+
3330                                 to->rep->ns_numrrsets,
3331                                 to->rep->ar_numrrsets * sizeof(dest[0]));
3332                         dest += to->rep->ar_numrrsets;
3333                         memcpy(dest, from->rep->rrsets+from->rep->an_numrrsets+
3334                                 from->rep->ns_numrrsets,
3335                                 from->rep->ar_numrrsets * sizeof(dest[0]));
3336                         /* update counts */
3337                         to->rep->rrsets = d;
3338                         to->rep->an_numrrsets += from->rep->an_numrrsets;
3339                         to->rep->ns_numrrsets += from->rep->ns_numrrsets;
3340                         to->rep->ar_numrrsets += from->rep->ar_numrrsets;
3341                         to->rep->rrset_count = n;
3342                 }
3343                 if(from->rep->security < to->rep->security) /* lowest sec */
3344                         to->rep->security = from->rep->security;
3345                 if(from->rep->qdcount != 0) /* insert qd if appropriate */
3346                         to->rep->qdcount = from->rep->qdcount;
3347                 if(from->rep->ttl < to->rep->ttl) /* use smallest TTL */
3348                         to->rep->ttl = from->rep->ttl;
3349                 if(from->rep->prefetch_ttl < to->rep->prefetch_ttl)
3350                         to->rep->prefetch_ttl = from->rep->prefetch_ttl;
3351                 if(from->rep->serve_expired_ttl < to->rep->serve_expired_ttl)
3352                         to->rep->serve_expired_ttl = from->rep->serve_expired_ttl;
3353         }
3354         /* are we done? */
3355         foriq->num_current_queries --;
3356         if(foriq->num_current_queries == 0)
3357                 foriq->state = FINISHED_STATE;
3358 }
3359
3360 /**
3361  * Collect class ANY responses and make them into one response.  This
3362  * state is started and it creates queries for all classes (that have
3363  * root hints).  The answers are then collected.
3364  *
3365  * @param qstate: query state.
3366  * @param id: module id.
3367  * @return true if the event needs more immediate processing, false if not.
3368  */
3369 static int
3370 processCollectClass(struct module_qstate* qstate, int id)
3371 {
3372         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
3373         struct module_qstate* subq;
3374         /* If qchase.qclass == 0 then send out queries for all classes.
3375          * Otherwise, do nothing (wait for all answers to arrive and the
3376          * processClassResponse to put them together, and that moves us
3377          * towards the Finished state when done. */
3378         if(iq->qchase.qclass == 0) {
3379                 uint16_t c = 0;
3380                 iq->qchase.qclass = LDNS_RR_CLASS_ANY;
3381                 while(iter_get_next_root(qstate->env->hints,
3382                         qstate->env->fwds, &c)) {
3383                         /* generate query for this class */
3384                         log_nametypeclass(VERB_ALGO, "spawn collect query",
3385                                 qstate->qinfo.qname, qstate->qinfo.qtype, c);
3386                         if(!generate_sub_request(qstate->qinfo.qname,
3387                                 qstate->qinfo.qname_len, qstate->qinfo.qtype,
3388                                 c, qstate, id, iq, INIT_REQUEST_STATE,
3389                                 FINISHED_STATE, &subq,
3390                                 (int)!(qstate->query_flags&BIT_CD))) {
3391                                 errinf(qstate, "could not generate class ANY"
3392                                         " lookup query");
3393                                 return error_response(qstate, id,
3394                                         LDNS_RCODE_SERVFAIL);
3395                         }
3396                         /* ignore subq, no special init required */
3397                         iq->num_current_queries ++;
3398                         if(c == 0xffff)
3399                                 break;
3400                         else c++;
3401                 }
3402                 /* if no roots are configured at all, return */
3403                 if(iq->num_current_queries == 0) {
3404                         verbose(VERB_ALGO, "No root hints or fwds, giving up "
3405                                 "on qclass ANY");
3406                         return error_response(qstate, id, LDNS_RCODE_REFUSED);
3407                 }
3408                 /* return false, wait for queries to return */
3409         }
3410         /* if woke up here because of an answer, wait for more answers */
3411         return 0;
3412 }
3413
3414 /**
3415  * This handles the final state for first-tier responses (i.e., responses to
3416  * externally generated queries).
3417  *
3418  * @param qstate: query state.
3419  * @param iq: iterator query state.
3420  * @param id: module id.
3421  * @return true if the event needs more processing, false if not. Since this
3422  *         is the final state for an event, it always returns false.
3423  */
3424 static int
3425 processFinished(struct module_qstate* qstate, struct iter_qstate* iq,
3426         int id)
3427 {
3428         log_query_info(VERB_QUERY, "finishing processing for",
3429                 &qstate->qinfo);
3430
3431         /* store negative cache element for parent side glue. */
3432         if(!qstate->no_cache_store && iq->query_for_pside_glue
3433                 && !iq->pside_glue)
3434                         iter_store_parentside_neg(qstate->env, &qstate->qinfo,
3435                                 iq->deleg_msg?iq->deleg_msg->rep:
3436                                 (iq->response?iq->response->rep:NULL));
3437         if(!iq->response) {
3438                 verbose(VERB_ALGO, "No response is set, servfail");
3439                 errinf(qstate, "(no response found at query finish)");
3440                 return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3441         }
3442
3443         /* Make sure that the RA flag is set (since the presence of
3444          * this module means that recursion is available) */
3445         iq->response->rep->flags |= BIT_RA;
3446
3447         /* Clear the AA flag */
3448         /* FIXME: does this action go here or in some other module? */
3449         iq->response->rep->flags &= ~BIT_AA;
3450
3451         /* make sure QR flag is on */
3452         iq->response->rep->flags |= BIT_QR;
3453
3454         /* we have finished processing this query */
3455         qstate->ext_state[id] = module_finished;
3456
3457         /* TODO:  we are using a private TTL, trim the response. */
3458         /* if (mPrivateTTL > 0){IterUtils.setPrivateTTL(resp, mPrivateTTL); } */
3459
3460         /* prepend any items we have accumulated */
3461         if(iq->an_prepend_list || iq->ns_prepend_list) {
3462                 if(!iter_prepend(iq, iq->response, qstate->region)) {
3463                         log_err("prepend rrsets: out of memory");
3464                         return error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3465                 }
3466                 /* reset the query name back */
3467                 iq->response->qinfo = qstate->qinfo;
3468                 /* the security state depends on the combination */
3469                 iq->response->rep->security = sec_status_unchecked;
3470                 /* store message with the finished prepended items,
3471                  * but only if we did recursion. The nonrecursion referral
3472                  * from cache does not need to be stored in the msg cache. */
3473                 if(!qstate->no_cache_store && qstate->query_flags&BIT_RD) {
3474                         iter_dns_store(qstate->env, &qstate->qinfo,
3475                                 iq->response->rep, 0, qstate->prefetch_leeway,
3476                                 iq->dp&&iq->dp->has_parent_side_NS,
3477                                 qstate->region, qstate->query_flags);
3478                 }
3479         }
3480         qstate->return_rcode = LDNS_RCODE_NOERROR;
3481         qstate->return_msg = iq->response;
3482         return 0;
3483 }
3484
3485 /*
3486  * Return priming query results to interested super querystates.
3487  *
3488  * Sets the delegation point and delegation message (not nonRD queries).
3489  * This is a callback from walk_supers.
3490  *
3491  * @param qstate: query state that finished.
3492  * @param id: module id.
3493  * @param super: the qstate to inform.
3494  */
3495 void
3496 iter_inform_super(struct module_qstate* qstate, int id,
3497         struct module_qstate* super)
3498 {
3499         if(!qstate->is_priming && super->qinfo.qclass == LDNS_RR_CLASS_ANY)
3500                 processClassResponse(qstate, id, super);
3501         else if(super->qinfo.qtype == LDNS_RR_TYPE_DS && ((struct iter_qstate*)
3502                 super->minfo[id])->state == DSNS_FIND_STATE)
3503                 processDSNSResponse(qstate, id, super);
3504         else if(qstate->return_rcode != LDNS_RCODE_NOERROR)
3505                 error_supers(qstate, id, super);
3506         else if(qstate->is_priming)
3507                 prime_supers(qstate, id, super);
3508         else    processTargetResponse(qstate, id, super);
3509 }
3510
3511 /**
3512  * Handle iterator state.
3513  * Handle events. This is the real processing loop for events, responsible
3514  * for moving events through the various states. If a processing method
3515  * returns true, then it will be advanced to the next state. If false, then
3516  * processing will stop.
3517  *
3518  * @param qstate: query state.
3519  * @param ie: iterator shared global environment.
3520  * @param iq: iterator query state.
3521  * @param id: module id.
3522  */
3523 static void
3524 iter_handle(struct module_qstate* qstate, struct iter_qstate* iq,
3525         struct iter_env* ie, int id)
3526 {
3527         int cont = 1;
3528         while(cont) {
3529                 verbose(VERB_ALGO, "iter_handle processing q with state %s",
3530                         iter_state_to_string(iq->state));
3531                 switch(iq->state) {
3532                         case INIT_REQUEST_STATE:
3533                                 cont = processInitRequest(qstate, iq, ie, id);
3534                                 break;
3535                         case INIT_REQUEST_2_STATE:
3536                                 cont = processInitRequest2(qstate, iq, id);
3537                                 break;
3538                         case INIT_REQUEST_3_STATE:
3539                                 cont = processInitRequest3(qstate, iq, id);
3540                                 break;
3541                         case QUERYTARGETS_STATE:
3542                                 cont = processQueryTargets(qstate, iq, ie, id);
3543                                 break;
3544                         case QUERY_RESP_STATE:
3545                                 cont = processQueryResponse(qstate, iq, id);
3546                                 break;
3547                         case PRIME_RESP_STATE:
3548                                 cont = processPrimeResponse(qstate, id);
3549                                 break;
3550                         case COLLECT_CLASS_STATE:
3551                                 cont = processCollectClass(qstate, id);
3552                                 break;
3553                         case DSNS_FIND_STATE:
3554                                 cont = processDSNSFind(qstate, iq, id);
3555                                 break;
3556                         case FINISHED_STATE:
3557                                 cont = processFinished(qstate, iq, id);
3558                                 break;
3559                         default:
3560                                 log_warn("iterator: invalid state: %d",
3561                                         iq->state);
3562                                 cont = 0;
3563                                 break;
3564                 }
3565         }
3566 }
3567
3568 /**
3569  * This is the primary entry point for processing request events. Note that
3570  * this method should only be used by external modules.
3571  * @param qstate: query state.
3572  * @param ie: iterator shared global environment.
3573  * @param iq: iterator query state.
3574  * @param id: module id.
3575  */
3576 static void
3577 process_request(struct module_qstate* qstate, struct iter_qstate* iq,
3578         struct iter_env* ie, int id)
3579 {
3580         /* external requests start in the INIT state, and finish using the
3581          * FINISHED state. */
3582         iq->state = INIT_REQUEST_STATE;
3583         iq->final_state = FINISHED_STATE;
3584         verbose(VERB_ALGO, "process_request: new external request event");
3585         iter_handle(qstate, iq, ie, id);
3586 }
3587
3588 /** process authoritative server reply */
3589 static void
3590 process_response(struct module_qstate* qstate, struct iter_qstate* iq,
3591         struct iter_env* ie, int id, struct outbound_entry* outbound,
3592         enum module_ev event)
3593 {
3594         struct msg_parse* prs;
3595         struct edns_data edns;
3596         sldns_buffer* pkt;
3597
3598         verbose(VERB_ALGO, "process_response: new external response event");
3599         iq->response = NULL;
3600         iq->state = QUERY_RESP_STATE;
3601         if(event == module_event_noreply || event == module_event_error) {
3602                 if(event == module_event_noreply && iq->sent_count >= 3 &&
3603                         qstate->env->cfg->use_caps_bits_for_id &&
3604                         !iq->caps_fallback && !is_caps_whitelisted(ie, iq)) {
3605                         /* start fallback */
3606                         iq->caps_fallback = 1;
3607                         iq->caps_server = 0;
3608                         iq->caps_reply = NULL;
3609                         iq->caps_response = NULL;
3610                         iq->caps_minimisation_state = DONOT_MINIMISE_STATE;
3611                         iq->state = QUERYTARGETS_STATE;
3612                         iq->num_current_queries--;
3613                         /* need fresh attempts for the 0x20 fallback, if
3614                          * that was the cause for the failure */
3615                         iter_dec_attempts(iq->dp, 3);
3616                         verbose(VERB_DETAIL, "Capsforid: timeouts, starting fallback");
3617                         goto handle_it;
3618                 }
3619                 goto handle_it;
3620         }
3621         if( (event != module_event_reply && event != module_event_capsfail)
3622                 || !qstate->reply) {
3623                 log_err("Bad event combined with response");
3624                 outbound_list_remove(&iq->outlist, outbound);
3625                 errinf(qstate, "module iterator received wrong internal event with a response message");
3626                 (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3627                 return;
3628         }
3629
3630         /* parse message */
3631         prs = (struct msg_parse*)regional_alloc(qstate->env->scratch,
3632                 sizeof(struct msg_parse));
3633         if(!prs) {
3634                 log_err("out of memory on incoming message");
3635                 /* like packet got dropped */
3636                 goto handle_it;
3637         }
3638         memset(prs, 0, sizeof(*prs));
3639         memset(&edns, 0, sizeof(edns));
3640         pkt = qstate->reply->c->buffer;
3641         sldns_buffer_set_position(pkt, 0);
3642         if(parse_packet(pkt, prs, qstate->env->scratch) != LDNS_RCODE_NOERROR) {
3643                 verbose(VERB_ALGO, "parse error on reply packet");
3644                 goto handle_it;
3645         }
3646         /* edns is not examined, but removed from message to help cache */
3647         if(parse_extract_edns(prs, &edns, qstate->env->scratch) !=
3648                 LDNS_RCODE_NOERROR)
3649                 goto handle_it;
3650
3651         /* Copy the edns options we may got from the back end */
3652         if(edns.opt_list) {
3653                 qstate->edns_opts_back_in = edns_opt_copy_region(edns.opt_list,
3654                         qstate->region);
3655                 if(!qstate->edns_opts_back_in) {
3656                         log_err("out of memory on incoming message");
3657                         /* like packet got dropped */
3658                         goto handle_it;
3659                 }
3660                 if(!inplace_cb_edns_back_parsed_call(qstate->env, qstate)) {
3661                         log_err("unable to call edns_back_parsed callback");
3662                         goto handle_it;
3663                 }
3664         }
3665
3666         /* remove CD-bit, we asked for in case we handle validation ourself */
3667         prs->flags &= ~BIT_CD;
3668
3669         /* normalize and sanitize: easy to delete items from linked lists */
3670         if(!scrub_message(pkt, prs, &iq->qinfo_out, iq->dp->name,
3671                 qstate->env->scratch, qstate->env, ie)) {
3672                 /* if 0x20 enabled, start fallback, but we have no message */
3673                 if(event == module_event_capsfail && !iq->caps_fallback) {
3674                         iq->caps_fallback = 1;
3675                         iq->caps_server = 0;
3676                         iq->caps_reply = NULL;
3677                         iq->caps_response = NULL;
3678                         iq->caps_minimisation_state = DONOT_MINIMISE_STATE;
3679                         iq->state = QUERYTARGETS_STATE;
3680                         iq->num_current_queries--;
3681                         verbose(VERB_DETAIL, "Capsforid: scrub failed, starting fallback with no response");
3682                 }
3683                 goto handle_it;
3684         }
3685
3686         /* allocate response dns_msg in region */
3687         iq->response = dns_alloc_msg(pkt, prs, qstate->region);
3688         if(!iq->response)
3689                 goto handle_it;
3690         log_query_info(VERB_DETAIL, "response for", &qstate->qinfo);
3691         log_name_addr(VERB_DETAIL, "reply from", iq->dp->name,
3692                 &qstate->reply->addr, qstate->reply->addrlen);
3693         if(verbosity >= VERB_ALGO)
3694                 log_dns_msg("incoming scrubbed packet:", &iq->response->qinfo,
3695                         iq->response->rep);
3696
3697         if(event == module_event_capsfail || iq->caps_fallback) {
3698                 if(qstate->env->cfg->qname_minimisation &&
3699                         iq->minimisation_state != DONOT_MINIMISE_STATE) {
3700                         /* Skip QNAME minimisation for next query, since that
3701                          * one has to match the current query. */
3702                         iq->minimisation_state = SKIP_MINIMISE_STATE;
3703                 }
3704                 /* for fallback we care about main answer, not additionals */
3705                 /* removing that makes comparison more likely to succeed */
3706                 caps_strip_reply(iq->response->rep);
3707
3708                 if(iq->caps_fallback &&
3709                         iq->caps_minimisation_state != iq->minimisation_state) {
3710                         /* QNAME minimisation state has changed, restart caps
3711                          * fallback. */
3712                         iq->caps_fallback = 0;
3713                 }
3714
3715                 if(!iq->caps_fallback) {
3716                         /* start fallback */
3717                         iq->caps_fallback = 1;
3718                         iq->caps_server = 0;
3719                         iq->caps_reply = iq->response->rep;
3720                         iq->caps_response = iq->response;
3721                         iq->caps_minimisation_state = iq->minimisation_state;
3722                         iq->state = QUERYTARGETS_STATE;
3723                         iq->num_current_queries--;
3724                         verbose(VERB_DETAIL, "Capsforid: starting fallback");
3725                         goto handle_it;
3726                 } else {
3727                         /* check if reply is the same, otherwise, fail */
3728                         if(!iq->caps_reply) {
3729                                 iq->caps_reply = iq->response->rep;
3730                                 iq->caps_response = iq->response;
3731                                 iq->caps_server = -1; /*become zero at ++,
3732                                 so that we start the full set of trials */
3733                         } else if(caps_failed_rcode(iq->caps_reply) &&
3734                                 !caps_failed_rcode(iq->response->rep)) {
3735                                 /* prefer to upgrade to non-SERVFAIL */
3736                                 iq->caps_reply = iq->response->rep;
3737                                 iq->caps_response = iq->response;
3738                         } else if(!caps_failed_rcode(iq->caps_reply) &&
3739                                 caps_failed_rcode(iq->response->rep)) {
3740                                 /* if we have non-SERVFAIL as answer then
3741                                  * we can ignore SERVFAILs for the equality
3742                                  * comparison */
3743                                 /* no instructions here, skip other else */
3744                         } else if(caps_failed_rcode(iq->caps_reply) &&
3745                                 caps_failed_rcode(iq->response->rep)) {
3746                                 /* failure is same as other failure in fallbk*/
3747                                 /* no instructions here, skip other else */
3748                         } else if(!reply_equal(iq->response->rep, iq->caps_reply,
3749                                 qstate->env->scratch)) {
3750                                 verbose(VERB_DETAIL, "Capsforid fallback: "
3751                                         "getting different replies, failed");
3752                                 outbound_list_remove(&iq->outlist, outbound);
3753                                 errinf(qstate, "0x20 failed, then got different replies in fallback");
3754                                 (void)error_response(qstate, id,
3755                                         LDNS_RCODE_SERVFAIL);
3756                                 return;
3757                         }
3758                         /* continue the fallback procedure at next server */
3759                         iq->caps_server++;
3760                         iq->state = QUERYTARGETS_STATE;
3761                         iq->num_current_queries--;
3762                         verbose(VERB_DETAIL, "Capsforid: reply is equal. "
3763                                 "go to next fallback");
3764                         goto handle_it;
3765                 }
3766         }
3767         iq->caps_fallback = 0; /* if we were in fallback, 0x20 is OK now */
3768
3769 handle_it:
3770         outbound_list_remove(&iq->outlist, outbound);
3771         iter_handle(qstate, iq, ie, id);
3772 }
3773
3774 void
3775 iter_operate(struct module_qstate* qstate, enum module_ev event, int id,
3776         struct outbound_entry* outbound)
3777 {
3778         struct iter_env* ie = (struct iter_env*)qstate->env->modinfo[id];
3779         struct iter_qstate* iq = (struct iter_qstate*)qstate->minfo[id];
3780         verbose(VERB_QUERY, "iterator[module %d] operate: extstate:%s event:%s",
3781                 id, strextstate(qstate->ext_state[id]), strmodulevent(event));
3782         if(iq) log_query_info(VERB_QUERY, "iterator operate: query",
3783                 &qstate->qinfo);
3784         if(iq && qstate->qinfo.qname != iq->qchase.qname)
3785                 log_query_info(VERB_QUERY, "iterator operate: chased to",
3786                         &iq->qchase);
3787
3788         /* perform iterator state machine */
3789         if((event == module_event_new || event == module_event_pass) &&
3790                 iq == NULL) {
3791                 if(!iter_new(qstate, id)) {
3792                         errinf(qstate, "malloc failure, new iterator module allocation");
3793                         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3794                         return;
3795                 }
3796                 iq = (struct iter_qstate*)qstate->minfo[id];
3797                 process_request(qstate, iq, ie, id);
3798                 return;
3799         }
3800         if(iq && event == module_event_pass) {
3801                 iter_handle(qstate, iq, ie, id);
3802                 return;
3803         }
3804         if(iq && outbound) {
3805                 process_response(qstate, iq, ie, id, outbound, event);
3806                 return;
3807         }
3808         if(event == module_event_error) {
3809                 verbose(VERB_ALGO, "got called with event error, giving up");
3810                 errinf(qstate, "iterator module got the error event");
3811                 (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3812                 return;
3813         }
3814
3815         log_err("bad event for iterator");
3816         errinf(qstate, "iterator module received wrong event");
3817         (void)error_response(qstate, id, LDNS_RCODE_SERVFAIL);
3818 }
3819
3820 void
3821 iter_clear(struct module_qstate* qstate, int id)
3822 {
3823         struct iter_qstate* iq;
3824         if(!qstate)
3825                 return;
3826         iq = (struct iter_qstate*)qstate->minfo[id];
3827         if(iq) {
3828                 outbound_list_clear(&iq->outlist);
3829                 if(iq->target_count && --iq->target_count[0] == 0)
3830                         free(iq->target_count);
3831                 iq->num_current_queries = 0;
3832         }
3833         qstate->minfo[id] = NULL;
3834 }
3835
3836 size_t
3837 iter_get_mem(struct module_env* env, int id)
3838 {
3839         struct iter_env* ie = (struct iter_env*)env->modinfo[id];
3840         if(!ie)
3841                 return 0;
3842         return sizeof(*ie) + sizeof(int)*((size_t)ie->max_dependency_depth+1)
3843                 + donotq_get_mem(ie->donotq) + priv_get_mem(ie->priv);
3844 }
3845
3846 /**
3847  * The iterator function block
3848  */
3849 static struct module_func_block iter_block = {
3850         "iterator",
3851         &iter_init, &iter_deinit, &iter_operate, &iter_inform_super,
3852         &iter_clear, &iter_get_mem
3853 };
3854
3855 struct module_func_block*
3856 iter_get_funcblock(void)
3857 {
3858         return &iter_block;
3859 }
3860
3861 const char*
3862 iter_state_to_string(enum iter_state state)
3863 {
3864         switch (state)
3865         {
3866         case INIT_REQUEST_STATE :
3867                 return "INIT REQUEST STATE";
3868         case INIT_REQUEST_2_STATE :
3869                 return "INIT REQUEST STATE (stage 2)";
3870         case INIT_REQUEST_3_STATE:
3871                 return "INIT REQUEST STATE (stage 3)";
3872         case QUERYTARGETS_STATE :
3873                 return "QUERY TARGETS STATE";
3874         case PRIME_RESP_STATE :
3875                 return "PRIME RESPONSE STATE";
3876         case COLLECT_CLASS_STATE :
3877                 return "COLLECT CLASS STATE";
3878         case DSNS_FIND_STATE :
3879                 return "DSNS FIND STATE";
3880         case QUERY_RESP_STATE :
3881                 return "QUERY RESPONSE STATE";
3882         case FINISHED_STATE :
3883                 return "FINISHED RESPONSE STATE";
3884         default :
3885                 return "UNKNOWN ITER STATE";
3886         }
3887 }
3888
3889 int
3890 iter_state_is_responsestate(enum iter_state s)
3891 {
3892         switch(s) {
3893                 case INIT_REQUEST_STATE :
3894                 case INIT_REQUEST_2_STATE :
3895                 case INIT_REQUEST_3_STATE :
3896                 case QUERYTARGETS_STATE :
3897                 case COLLECT_CLASS_STATE :
3898                         return 0;
3899                 default:
3900                         break;
3901         }
3902         return 1;
3903 }