iterator/iterator.c

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