pdns/dnsdist.hh

   1 /*
   2  * This file is part of PowerDNS or dnsdist.
   3  * Copyright -- PowerDNS.COM B.V. and its contributors
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of version 2 of the GNU General Public License as
   7  * published by the Free Software Foundation.
   8  *
   9  * In addition, for the avoidance of any doubt, permission is granted to
  10  * link this program with OpenSSL and to (re)distribute the binaries
  11  * produced as the result of such linking.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
  21  */
  22 #pragma once
  23 #include "config.h"
  24 #include "ext/luawrapper/include/LuaContext.hpp"
  25
  26 #include <atomic>
  27 #include <mutex>
  28 #include <string>
  29 #include <thread>
  30 #include <time.h>
  31 #include <unistd.h>
  32 #include <unordered_map>
  33
  34 #include <boost/variant.hpp>
  35
  36 #include "bpf-filter.hh"
  37 #include "capabilities.hh"
  38 #include "circular_buffer.hh"
  39 #include "dnscrypt.hh"
  40 #include "dnsdist-cache.hh"
  41 #include "dnsdist-dynbpf.hh"
  42 #include "dnsname.hh"
  43 #include "doh.hh"
  44 #include "ednsoptions.hh"
  45 #include "gettime.hh"
  46 #include "iputils.hh"
  47 #include "misc.hh"
  48 #include "mplexer.hh"
  49 #include "sholder.hh"
  50 #include "tcpiohandler.hh"
  51 #include "uuid-utils.hh"
  52
  53 void carbonDumpThread();
  54 uint64_t uptimeOfProcess(const std::string& str);
  55
  56 extern uint16_t g_ECSSourcePrefixV4;
  57 extern uint16_t g_ECSSourcePrefixV6;
  58 extern bool g_ECSOverride;
  59
  60 typedef std::unordered_map<string, string> QTag;
  61
  62 struct DNSQuestion
  63 {
  64   DNSQuestion(const DNSName* name, uint16_t type, uint16_t class_, unsigned int consumed_, const ComboAddress* lc, const ComboAddress* rem, struct dnsheader* header, size_t bufferSize, uint16_t queryLen, bool isTcp, const struct timespec* queryTime_):
  65     qname(name), local(lc), remote(rem), dh(header), queryTime(queryTime_), size(bufferSize), consumed(consumed_), tempFailureTTL(boost::none), qtype(type), qclass(class_), len(queryLen), ecsPrefixLength(rem->sin4.sin_family == AF_INET ? g_ECSSourcePrefixV4 : g_ECSSourcePrefixV6), tcp(isTcp), ecsOverride(g_ECSOverride) {
  66     const uint16_t* flags = getFlagsFromDNSHeader(dh);
  67     origFlags = *flags;
  68   }
  69   DNSQuestion(const DNSQuestion&) = delete;
  70   DNSQuestion& operator=(const DNSQuestion&) = delete;
  71   DNSQuestion(DNSQuestion&&) = default;
  72
  73 #ifdef HAVE_PROTOBUF
  74   boost::optional<boost::uuids::uuid> uniqueId;
  75 #endif
  76   Netmask ecs;
  77   boost::optional<Netmask> subnet;
  78   std::string sni; /* Server Name Indication, if any (DoT or DoH) */
  79   std::string poolname;
  80   const DNSName* qname{nullptr};
  81   const ComboAddress* local{nullptr};
  82   const ComboAddress* remote{nullptr};
  83   std::shared_ptr<QTag> qTag{nullptr};
  84   std::shared_ptr<std::map<uint16_t, EDNSOptionView> > ednsOptions;
  85   std::shared_ptr<DNSCryptQuery> dnsCryptQuery{nullptr};
  86   std::shared_ptr<DNSDistPacketCache> packetCache{nullptr};
  87   struct dnsheader* dh{nullptr};
  88   const struct timespec* queryTime{nullptr};
  89   struct DOHUnit* du{nullptr};
  90   size_t size;
  91   unsigned int consumed{0};
  92   int delayMsec{0};
  93   boost::optional<uint32_t> tempFailureTTL;
  94   uint32_t cacheKeyNoECS;
  95   uint32_t cacheKey;
  96   const uint16_t qtype;
  97   const uint16_t qclass;
  98   uint16_t len;
  99   uint16_t ecsPrefixLength;
 100   uint16_t origFlags;
 101   uint8_t ednsRCode{0};
 102   const bool tcp;
 103   bool skipCache{false};
 104   bool ecsOverride;
 105   bool useECS{true};
 106   bool addXPF{true};
 107   bool ecsSet{false};
 108   bool ecsAdded{false};
 109   bool ednsAdded{false};
 110   bool useZeroScope{false};
 111   bool dnssecOK{false};
 112 };
 113
 114 struct DNSResponse : DNSQuestion
 115 {
 116   DNSResponse(const DNSName* name, uint16_t type, uint16_t class_, unsigned int consumed_, const ComboAddress* lc, const ComboAddress* rem, struct dnsheader* header, size_t bufferSize, uint16_t responseLen, bool isTcp, const struct timespec* queryTime_):
 117     DNSQuestion(name, type, class_, consumed_, lc, rem, header, bufferSize, responseLen, isTcp, queryTime_) { }
 118   DNSResponse(const DNSResponse&) = delete;
 119   DNSResponse& operator=(const DNSResponse&) = delete;
 120   DNSResponse(DNSResponse&&) = default;
 121 };
 122
 123 /* so what could you do:
 124    drop,
 125    fake up nxdomain,
 126    provide actual answer,
 127    allow & and stop processing,
 128    continue processing,
 129    modify header:    (servfail|refused|notimp), set TC=1,
 130    send to pool */
 131
 132 class DNSAction
 133 {
 134 public:
 135   enum class Action { Drop, Nxdomain, Refused, Spoof, Allow, HeaderModify, Pool, Delay, Truncate, ServFail, None, NoOp, NoRecurse };
 136   static std::string typeToString(const Action& action)
 137   {
 138     switch(action) {
 139     case Action::Drop:
 140       return "Drop";
 141     case Action::Nxdomain:
 142       return "Send NXDomain";
 143     case Action::Refused:
 144       return "Send Refused";
 145     case Action::Spoof:
 146       return "Spoof an answer";
 147     case Action::Allow:
 148       return "Allow";
 149     case Action::HeaderModify:
 150       return "Modify the header";
 151     case Action::Pool:
 152       return "Route to a pool";
 153     case Action::Delay:
 154       return "Delay";
 155     case Action::Truncate:
 156       return "Truncate over UDP";
 157     case Action::ServFail:
 158       return "Send ServFail";
 159     case Action::None:
 160     case Action::NoOp:
 161       return "Do nothing";
 162     case Action::NoRecurse:
 163       return "Set rd=0";
 164     }
 165
 166     return "Unknown";
 167   }
 168
 169   virtual Action operator()(DNSQuestion*, string* ruleresult) const =0;
 170   virtual ~DNSAction()
 171   {
 172   }
 173   virtual string toString() const = 0;
 174   virtual std::map<string, double> getStats() const
 175   {
 176     return {{}};
 177   }
 178 };
 179
 180 class DNSResponseAction
 181 {
 182 public:
 183   enum class Action { Allow, Delay, Drop, HeaderModify, ServFail, None };
 184   virtual Action operator()(DNSResponse*, string* ruleresult) const =0;
 185   virtual ~DNSResponseAction()
 186   {
 187   }
 188   virtual string toString() const = 0;
 189 };
 190
 191 struct DynBlock
 192 {
 193   DynBlock(): action(DNSAction::Action::None), warning(false)
 194   {
 195   }
 196
 197   DynBlock(const std::string& reason_, const struct timespec& until_, const DNSName& domain_, DNSAction::Action action_): reason(reason_), until(until_), domain(domain_), action(action_), warning(false)
 198   {
 199   }
 200
 201   DynBlock(const DynBlock& rhs): reason(rhs.reason), until(rhs.until), domain(rhs.domain), action(rhs.action), warning(rhs.warning)
 202   {
 203     blocks.store(rhs.blocks);
 204   }
 205
 206   DynBlock& operator=(const DynBlock& rhs)
 207   {
 208     reason=rhs.reason;
 209     until=rhs.until;
 210     domain=rhs.domain;
 211     action=rhs.action;
 212     blocks.store(rhs.blocks);
 213     warning=rhs.warning;
 214     return *this;
 215   }
 216
 217   string reason;
 218   struct timespec until;
 219   DNSName domain;
 220   DNSAction::Action action;
 221   mutable std::atomic<unsigned int> blocks;
 222   bool warning;
 223 };
 224
 225 extern GlobalStateHolder<NetmaskTree<DynBlock>> g_dynblockNMG;
 226
 227 extern vector<pair<struct timeval, std::string> > g_confDelta;
 228
 229 extern uint64_t getLatencyCount(const std::string&);
 230
 231 struct DNSDistStats
 232 {
 233   using stat_t=std::atomic<uint64_t>; // aww yiss ;-)
 234   stat_t responses{0};
 235   stat_t servfailResponses{0};
 236   stat_t queries{0};
 237   stat_t frontendNXDomain{0};
 238   stat_t frontendServFail{0};
 239   stat_t frontendNoError{0};
 240   stat_t nonCompliantQueries{0};
 241   stat_t nonCompliantResponses{0};
 242   stat_t rdQueries{0};
 243   stat_t emptyQueries{0};
 244   stat_t aclDrops{0};
 245   stat_t dynBlocked{0};
 246   stat_t ruleDrop{0};
 247   stat_t ruleNXDomain{0};
 248   stat_t ruleRefused{0};
 249   stat_t ruleServFail{0};
 250   stat_t selfAnswered{0};
 251   stat_t downstreamTimeouts{0};
 252   stat_t downstreamSendErrors{0};
 253   stat_t truncFail{0};
 254   stat_t noPolicy{0};
 255   stat_t cacheHits{0};
 256   stat_t cacheMisses{0};
 257   stat_t latency0_1{0}, latency1_10{0}, latency10_50{0}, latency50_100{0}, latency100_1000{0}, latencySlow{0}, latencySum{0};
 258   stat_t securityStatus{0};
 259
 260   double latencyAvg100{0}, latencyAvg1000{0}, latencyAvg10000{0}, latencyAvg1000000{0};
 261   typedef std::function<uint64_t(const std::string&)> statfunction_t;
 262   typedef boost::variant<stat_t*, double*, statfunction_t> entry_t;
 263   std::vector<std::pair<std::string, entry_t>> entries{
 264     {"responses", &responses},
 265     {"servfail-responses", &servfailResponses},
 266     {"queries", &queries},
 267     {"frontend-nxdomain", &frontendNXDomain},
 268     {"frontend-servfail", &frontendServFail},
 269     {"frontend-noerror", &frontendNoError},
 270     {"acl-drops", &aclDrops},
 271     {"rule-drop", &ruleDrop},
 272     {"rule-nxdomain", &ruleNXDomain},
 273     {"rule-refused", &ruleRefused},
 274     {"rule-servfail", &ruleServFail},
 275     {"self-answered", &selfAnswered},
 276     {"downstream-timeouts", &downstreamTimeouts},
 277     {"downstream-send-errors", &downstreamSendErrors},
 278     {"trunc-failures", &truncFail},
 279     {"no-policy", &noPolicy},
 280     {"latency0-1", &latency0_1},
 281     {"latency1-10", &latency1_10},
 282     {"latency10-50", &latency10_50},
 283     {"latency50-100", &latency50_100},
 284     {"latency100-1000", &latency100_1000},
 285     {"latency-slow", &latencySlow},
 286     {"latency-avg100", &latencyAvg100},
 287     {"latency-avg1000", &latencyAvg1000},
 288     {"latency-avg10000", &latencyAvg10000},
 289     {"latency-avg1000000", &latencyAvg1000000},
 290     {"uptime", uptimeOfProcess},
 291     {"real-memory-usage", getRealMemoryUsage},
 292     {"special-memory-usage", getSpecialMemoryUsage},
 293     {"noncompliant-queries", &nonCompliantQueries},
 294     {"noncompliant-responses", &nonCompliantResponses},
 295     {"rdqueries", &rdQueries},
 296     {"empty-queries", &emptyQueries},
 297     {"cache-hits", &cacheHits},
 298     {"cache-misses", &cacheMisses},
 299     {"cpu-user-msec", getCPUTimeUser},
 300     {"cpu-sys-msec", getCPUTimeSystem},
 301     {"fd-usage", getOpenFileDescriptors},
 302     {"dyn-blocked", &dynBlocked},
 303     {"dyn-block-nmg-size", [](const std::string&) { return g_dynblockNMG.getLocal()->size(); }},
 304     {"security-status", &securityStatus},
 305     // Latency histogram
 306     {"latency-sum", &latencySum},
 307     {"latency-count", getLatencyCount},
 308   };
 309 };
 310
 311 // Metric types for Prometheus
 312 enum class PrometheusMetricType: int {
 313     counter = 1,
 314     gauge = 2
 315 };
 316
 317 // Keeps additional information about metrics
 318 struct MetricDefinition {
 319   MetricDefinition(PrometheusMetricType _prometheusType, const std::string& _description): description(_description), prometheusType(_prometheusType) {
 320   }
 321
 322   MetricDefinition() = default;
 323
 324   // Metric description
 325   std::string description;
 326   // Metric type for Prometheus
 327   PrometheusMetricType prometheusType;
 328 };
 329
 330 struct MetricDefinitionStorage {
 331   // Return metric definition by name
 332   bool getMetricDetails(std::string metricName, MetricDefinition& metric) {
 333   auto metricDetailsIter = metrics.find(metricName);
 334
 335   if (metricDetailsIter == metrics.end()) {
 336     return false;
 337   }
 338
 339   metric = metricDetailsIter->second;
 340     return true;
 341   };
 342
 343   // Return string representation of Prometheus metric type
 344   std::string getPrometheusStringMetricType(PrometheusMetricType metricType) {
 345     switch (metricType) {
 346       case PrometheusMetricType::counter:
 347         return "counter";
 348         break;
 349       case PrometheusMetricType::gauge:
 350         return "gauge";
 351         break;
 352       default:
 353         return "";
 354         break;
 355     }
 356   };
 357
 358   std::map<std::string, MetricDefinition> metrics = {
 359     { "responses",              MetricDefinition(PrometheusMetricType::counter, "Number of responses received from backends") },
 360     { "servfail-responses",     MetricDefinition(PrometheusMetricType::counter, "Number of SERVFAIL answers received from backends") },
 361     { "queries",                MetricDefinition(PrometheusMetricType::counter, "Number of received queries")},
 362     { "frontend-nxdomain",      MetricDefinition(PrometheusMetricType::counter, "Number of NXDomain answers sent to clients")},
 363     { "frontend-servfail",      MetricDefinition(PrometheusMetricType::counter, "Number of SERVFAIL answers sent to clients")},
 364     { "frontend-noerror",       MetricDefinition(PrometheusMetricType::counter, "Number of NoError answers sent to clients")},
 365     { "acl-drops",              MetricDefinition(PrometheusMetricType::counter, "Number of packets dropped because of the ACL")},
 366     { "rule-drop",              MetricDefinition(PrometheusMetricType::counter, "Number of queries dropped because of a rule")},
 367     { "rule-nxdomain",          MetricDefinition(PrometheusMetricType::counter, "Number of NXDomain answers returned because of a rule")},
 368     { "rule-refused",           MetricDefinition(PrometheusMetricType::counter, "Number of Refused answers returned because of a rule")},
 369     { "rule-servfail",          MetricDefinition(PrometheusMetricType::counter, "Number of SERVFAIL answers received because of a rule")},
 370     { "self-answered",          MetricDefinition(PrometheusMetricType::counter, "Number of self-answered responses")},
 371     { "downstream-timeouts",    MetricDefinition(PrometheusMetricType::counter, "Number of queries not answered in time by a backend")},
 372     { "downstream-send-errors", MetricDefinition(PrometheusMetricType::counter, "Number of errors when sending a query to a backend")},
 373     { "trunc-failures",         MetricDefinition(PrometheusMetricType::counter, "Number of errors encountered while truncating an answer")},
 374     { "no-policy",              MetricDefinition(PrometheusMetricType::counter, "Number of queries dropped because no server was available")},
 375     { "latency0-1",             MetricDefinition(PrometheusMetricType::counter, "Number of queries answered in less than 1ms")},
 376     { "latency1-10",            MetricDefinition(PrometheusMetricType::counter, "Number of queries answered in 1-10 ms")},
 377     { "latency10-50",           MetricDefinition(PrometheusMetricType::counter, "Number of queries answered in 10-50 ms")},
 378     { "latency50-100",          MetricDefinition(PrometheusMetricType::counter, "Number of queries answered in 50-100 ms")},
 379     { "latency100-1000",        MetricDefinition(PrometheusMetricType::counter, "Number of queries answered in 100-1000 ms")},
 380     { "latency-slow",           MetricDefinition(PrometheusMetricType::counter, "Number of queries answered in more than 1 second")},
 381     { "latency-avg100",         MetricDefinition(PrometheusMetricType::gauge,   "Average response latency in microseconds of the last 100 packets")},
 382     { "latency-avg1000",        MetricDefinition(PrometheusMetricType::gauge,   "Average response latency in microseconds of the last 1000 packets")},
 383     { "latency-avg10000",       MetricDefinition(PrometheusMetricType::gauge,   "Average response latency in microseconds of the last 10000 packets")},
 384     { "latency-avg1000000",     MetricDefinition(PrometheusMetricType::gauge,   "Average response latency in microseconds of the last 1000000 packets")},
 385     { "uptime",                 MetricDefinition(PrometheusMetricType::gauge,   "Uptime of the dnsdist process in seconds")},
 386     { "real-memory-usage",      MetricDefinition(PrometheusMetricType::gauge,   "Current memory usage in bytes")},
 387     { "noncompliant-queries",   MetricDefinition(PrometheusMetricType::counter, "Number of queries dropped as non-compliant")},
 388     { "noncompliant-responses", MetricDefinition(PrometheusMetricType::counter, "Number of answers from a backend dropped as non-compliant")},
 389     { "rdqueries",              MetricDefinition(PrometheusMetricType::counter, "Number of received queries with the recursion desired bit set")},
 390     { "empty-queries",          MetricDefinition(PrometheusMetricType::counter, "Number of empty queries received from clients")},
 391     { "cache-hits",             MetricDefinition(PrometheusMetricType::counter, "Number of times an answer was retrieved from cache")},
 392     { "cache-misses",           MetricDefinition(PrometheusMetricType::counter, "Number of times an answer not found in the cache")},
 393     { "cpu-user-msec",          MetricDefinition(PrometheusMetricType::counter, "Milliseconds spent by dnsdist in the user state")},
 394     { "cpu-sys-msec",           MetricDefinition(PrometheusMetricType::counter, "Milliseconds spent by dnsdist in the system state")},
 395     { "fd-usage",               MetricDefinition(PrometheusMetricType::gauge,   "Number of currently used file descriptors")},
 396     { "dyn-blocked",            MetricDefinition(PrometheusMetricType::counter, "Number of queries dropped because of a dynamic block")},
 397     { "dyn-block-nmg-size",     MetricDefinition(PrometheusMetricType::gauge,   "Number of dynamic blocks entries") },
 398     { "security-status",        MetricDefinition(PrometheusMetricType::gauge,   "Security status of this software. 0=unknown, 1=OK, 2=upgrade recommended, 3=upgrade mandatory") },
 399   };
 400 };
 401
 402 extern MetricDefinitionStorage g_metricDefinitions;
 403 extern struct DNSDistStats g_stats;
 404 void doLatencyStats(double udiff);
 405
 406
 407 struct StopWatch
 408 {
 409   StopWatch(bool realTime=false): d_needRealTime(realTime)
 410   {
 411   }
 412   struct timespec d_start{0,0};
 413   bool d_needRealTime{false};
 414
 415   void start() {
 416     if(gettime(&d_start, d_needRealTime) < 0)
 417       unixDie("Getting timestamp");
 418
 419   }
 420
 421   void set(const struct timespec& from) {
 422     d_start = from;
 423   }
 424
 425   double udiff() const {
 426     struct timespec now;
 427     if(gettime(&now, d_needRealTime) < 0)
 428       unixDie("Getting timestamp");
 429
 430     return 1000000.0*(now.tv_sec - d_start.tv_sec) + (now.tv_nsec - d_start.tv_nsec)/1000.0;
 431   }
 432
 433   double udiffAndSet() {
 434     struct timespec now;
 435     if(gettime(&now, d_needRealTime) < 0)
 436       unixDie("Getting timestamp");
 437
 438     auto ret= 1000000.0*(now.tv_sec - d_start.tv_sec) + (now.tv_nsec - d_start.tv_nsec)/1000.0;
 439     d_start = now;
 440     return ret;
 441   }
 442
 443 };
 444
 445 class BasicQPSLimiter
 446 {
 447 public:
 448   BasicQPSLimiter()
 449   {
 450   }
 451
 452   BasicQPSLimiter(unsigned int burst): d_tokens(burst)
 453   {
 454     d_prev.start();
 455   }
 456
 457   bool check(unsigned int rate, unsigned int burst) const // this is not quite fair
 458   {
 459     auto delta = d_prev.udiffAndSet();
 460
 461     if(delta > 0.0) // time, frequently, does go backwards..
 462       d_tokens += 1.0 * rate * (delta/1000000.0);
 463
 464     if(d_tokens > burst) {
 465       d_tokens = burst;
 466     }
 467
 468     bool ret=false;
 469     if(d_tokens >= 1.0) { // we need this because burst=1 is weird otherwise
 470       ret=true;
 471       --d_tokens;
 472     }
 473
 474     return ret;
 475   }
 476
 477   bool seenSince(const struct timespec& cutOff) const
 478   {
 479     return cutOff < d_prev.d_start;
 480   }
 481
 482 protected:
 483   mutable StopWatch d_prev;
 484   mutable double d_tokens;
 485 };
 486
 487 class QPSLimiter : public BasicQPSLimiter
 488 {
 489 public:
 490   QPSLimiter(): BasicQPSLimiter()
 491   {
 492   }
 493
 494   QPSLimiter(unsigned int rate, unsigned int burst): BasicQPSLimiter(burst), d_rate(rate), d_burst(burst), d_passthrough(false)
 495   {
 496     d_prev.start();
 497   }
 498
 499   unsigned int getRate() const
 500   {
 501     return d_passthrough ? 0 : d_rate;
 502   }
 503
 504   int getPassed() const
 505   {
 506     return d_passed;
 507   }
 508
 509   int getBlocked() const
 510   {
 511     return d_blocked;
 512   }
 513
 514   bool check() const // this is not quite fair
 515   {
 516     if (d_passthrough) {
 517       return true;
 518     }
 519
 520     bool ret = BasicQPSLimiter::check(d_rate, d_burst);
 521     if (ret) {
 522       d_passed++;
 523     }
 524     else {
 525       d_blocked++;
 526     }
 527
 528     return ret;
 529   }
 530 private:
 531   mutable unsigned int d_passed{0};
 532   mutable unsigned int d_blocked{0};
 533   unsigned int d_rate;
 534   unsigned int d_burst;
 535   bool d_passthrough{true};
 536 };
 537
 538 struct ClientState;
 539
 540 struct IDState
 541 {
 542   IDState(): sentTime(true), delayMsec(0), tempFailureTTL(boost::none) { origDest.sin4.sin_family = 0;}
 543   IDState(const IDState& orig): origRemote(orig.origRemote), origDest(orig.origDest), age(orig.age)
 544   {
 545     usageIndicator.store(orig.usageIndicator.load());
 546     origFD = orig.origFD;
 547     origID = orig.origID;
 548     delayMsec = orig.delayMsec;
 549     tempFailureTTL = orig.tempFailureTTL;
 550   }
 551
 552   static const int64_t unusedIndicator = -1;
 553
 554   static bool isInUse(int64_t usageIndicator)
 555   {
 556     return usageIndicator != unusedIndicator;
 557   }
 558
 559   bool isInUse() const
 560   {
 561     return usageIndicator != unusedIndicator;
 562   }
 563
 564   /* return true if the value has been successfully replaced meaning that
 565      no-one updated the usage indicator in the meantime */
 566   bool tryMarkUnused(int64_t expectedUsageIndicator)
 567   {
 568     return usageIndicator.compare_exchange_strong(expectedUsageIndicator, unusedIndicator);
 569   }
 570
 571   /* mark as unused no matter what, return true if the state was in use before */
 572   bool markAsUsed()
 573   {
 574     auto currentGeneration = generation++;
 575     return markAsUsed(currentGeneration);
 576   }
 577
 578   /* mark as unused no matter what, return true if the state was in use before */
 579   bool markAsUsed(int64_t currentGeneration)
 580   {
 581     int64_t oldUsage = usageIndicator.exchange(currentGeneration);
 582     return oldUsage != unusedIndicator;
 583   }
 584
 585   /* We use this value to detect whether this state is in use.
 586      For performance reasons we don't want to use a lock here, but that means
 587      we need to be very careful when modifying this value. Modifications happen
 588      from:
 589      - one of the UDP or DoH 'client' threads receiving a query, selecting a backend
 590        then picking one of the states associated to this backend (via the idOffset).
 591        Most of the time this state should not be in use and usageIndicator is -1, but we
 592        might not yet have received a response for the query previously associated to this
 593        state, meaning that we will 'reuse' this state and erase the existing state.
 594        If we ever receive a response for this state, it will be discarded. This is
 595        mostly fine for UDP except that we still need to be careful in order to miss
 596        the 'outstanding' counters, which should only be increased when we are picking
 597        an empty state, and not when reusing ;
 598        For DoH, though, we have dynamically allocated a DOHUnit object that needs to
 599        be freed, as well as internal objects internals to libh2o.
 600      - one of the UDP receiver threads receiving a response from a backend, picking
 601        the corresponding state and sending the response to the client ;
 602      - the 'healthcheck' thread scanning the states to actively discover timeouts,
 603        mostly to keep some counters like the 'outstanding' one sane.
 604      We previously based that logic on the origFD (FD on which the query was received,
 605      and therefore from where the response should be sent) but this suffered from an
 606      ABA problem since it was quite likely that a UDP 'client thread' would reset it to the
 607      same value since we only have so much incoming sockets:
 608      - 1/ 'client' thread gets a query and set origFD to its FD, say 5 ;
 609      - 2/ 'receiver' thread gets a response, read the value of origFD to 5, check that the qname,
 610        qtype and qclass match
 611      - 3/ during that time the 'client' thread reuses the state, setting again origFD to 5 ;
 612      - 4/ the 'receiver' thread uses compare_exchange_strong() to only replace the value if it's still
 613        5, except it's not the same 5 anymore and it overrides a fresh state.
 614      We now use a 32-bit unsigned counter instead, which is incremented every time the state is set,
 615      wrapping around if necessary, and we set an atomic signed 64-bit value, so that we still have -1
 616      when the state is unused and the value of our counter otherwise.
 617   */
 618   std::atomic<int64_t> usageIndicator{unusedIndicator};  // set to unusedIndicator to indicate this state is empty   // 8
 619   std::atomic<uint32_t> generation{0}; // increased every time a state is used, to be able to detect an ABA issue    // 4
 620   ComboAddress origRemote;                                    // 28
 621   ComboAddress origDest;                                      // 28
 622   StopWatch sentTime;                                         // 16
 623   DNSName qname;                                              // 80
 624   std::shared_ptr<DNSCryptQuery> dnsCryptQuery{nullptr};
 625 #ifdef HAVE_PROTOBUF
 626   boost::optional<boost::uuids::uuid> uniqueId;
 627 #endif
 628   boost::optional<Netmask> subnet{boost::none};
 629   std::shared_ptr<DNSDistPacketCache> packetCache{nullptr};
 630   std::shared_ptr<QTag> qTag{nullptr};
 631   const ClientState* cs{nullptr};
 632   DOHUnit* du{nullptr};
 633   uint32_t cacheKey;                                          // 4
 634   uint32_t cacheKeyNoECS;                                     // 4
 635   uint16_t age;                                               // 4
 636   uint16_t qtype;                                             // 2
 637   uint16_t qclass;                                            // 2
 638   uint16_t origID;                                            // 2
 639   uint16_t origFlags;                                         // 2
 640   int origFD{-1};
 641   int delayMsec;
 642   boost::optional<uint32_t> tempFailureTTL;
 643   bool ednsAdded{false};
 644   bool ecsAdded{false};
 645   bool skipCache{false};
 646   bool destHarvested{false}; // if true, origDest holds the original dest addr, otherwise the listening addr
 647   bool dnssecOK{false};
 648   bool useZeroScope;
 649 };
 650
 651 typedef std::unordered_map<string, unsigned int> QueryCountRecords;
 652 typedef std::function<std::tuple<bool, string>(const DNSQuestion* dq)> QueryCountFilter;
 653 struct QueryCount {
 654   QueryCount()
 655   {
 656     pthread_rwlock_init(&queryLock, nullptr);
 657   }
 658   QueryCountRecords records;
 659   QueryCountFilter filter;
 660   pthread_rwlock_t queryLock;
 661   bool enabled{false};
 662 };
 663
 664 extern QueryCount g_qcount;
 665
 666 struct ClientState
 667 {
 668   ClientState(const ComboAddress& local_, bool isTCP_, bool doReusePort, int fastOpenQueue, const std::string& itfName, const std::set<int>& cpus_): cpus(cpus_), local(local_), interface(itfName), fastOpenQueueSize(fastOpenQueue), tcp(isTCP_), reuseport(doReusePort)
 669   {
 670   }
 671
 672   std::set<int> cpus;
 673   ComboAddress local;
 674   std::shared_ptr<DNSCryptContext> dnscryptCtx{nullptr};
 675   std::shared_ptr<TLSFrontend> tlsFrontend{nullptr};
 676   std::shared_ptr<DOHFrontend> dohFrontend{nullptr};
 677   std::string interface;
 678   std::atomic<uint64_t> queries{0};
 679   mutable std::atomic<uint64_t> responses{0};
 680   std::atomic<uint64_t> tcpDiedReadingQuery{0};
 681   std::atomic<uint64_t> tcpDiedSendingResponse{0};
 682   std::atomic<uint64_t> tcpGaveUp{0};
 683   std::atomic<uint64_t> tcpClientTimeouts{0};
 684   std::atomic<uint64_t> tcpDownstreamTimeouts{0};
 685   std::atomic<uint64_t> tcpCurrentConnections{0};
 686   std::atomic<uint64_t> tlsNewSessions{0}; // A new TLS session has been negotiated, no resumption
 687   std::atomic<uint64_t> tlsResumptions{0}; // A TLS session has been resumed, either via session id or via a TLS ticket
 688   std::atomic<uint64_t> tlsUnknownTicketKey{0}; // A TLS ticket has been presented but we don't have the associated key (might have expired)
 689   std::atomic<uint64_t> tlsInactiveTicketKey{0}; // A TLS ticket has been successfully resumed but the key is no longer active, we should issue a new one
 690   std::atomic<uint64_t> tls10queries{0};   // valid DNS queries received via TLSv1.0
 691   std::atomic<uint64_t> tls11queries{0};   // valid DNS queries received via TLSv1.1
 692   std::atomic<uint64_t> tls12queries{0};   // valid DNS queries received via TLSv1.2
 693   std::atomic<uint64_t> tls13queries{0};   // valid DNS queries received via TLSv1.3
 694   std::atomic<uint64_t> tlsUnknownqueries{0};   // valid DNS queries received via unknown TLS version
 695   std::atomic<double> tcpAvgQueriesPerConnection{0.0};
 696   /* in ms */
 697   std::atomic<double> tcpAvgConnectionDuration{0.0};
 698   int udpFD{-1};
 699   int tcpFD{-1};
 700   int fastOpenQueueSize{0};
 701   bool muted{false};
 702   bool tcp;
 703   bool reuseport;
 704   bool ready{false};
 705
 706   int getSocket() const
 707   {
 708     return udpFD != -1 ? udpFD : tcpFD;
 709   }
 710
 711   bool isUDP() const
 712   {
 713     return udpFD != -1;
 714   }
 715
 716   bool isTCP() const
 717   {
 718     return udpFD == -1;
 719   }
 720
 721   bool hasTLS() const
 722   {
 723     return tlsFrontend != nullptr || dohFrontend != nullptr;
 724   }
 725
 726   std::string getType() const
 727   {
 728     std::string result = udpFD != -1 ? "UDP" : "TCP";
 729
 730     if (dohFrontend) {
 731       result += " (DNS over HTTPS)";
 732     }
 733     else if (tlsFrontend) {
 734       result += " (DNS over TLS)";
 735     }
 736     else if (dnscryptCtx) {
 737       result += " (DNSCrypt)";
 738     }
 739
 740     return result;
 741   }
 742
 743 #ifdef HAVE_EBPF
 744   shared_ptr<BPFFilter> d_filter;
 745
 746   void detachFilter()
 747   {
 748     if (d_filter) {
 749       d_filter->removeSocket(getSocket());
 750       d_filter = nullptr;
 751     }
 752   }
 753
 754   void attachFilter(shared_ptr<BPFFilter> bpf)
 755   {
 756     detachFilter();
 757
 758     bpf->addSocket(getSocket());
 759     d_filter = bpf;
 760   }
 761 #endif /* HAVE_EBPF */
 762
 763   void updateTCPMetrics(size_t nbQueries, uint64_t durationMs)
 764   {
 765     tcpAvgQueriesPerConnection = (99.0 * tcpAvgQueriesPerConnection / 100.0) + (nbQueries / 100.0);
 766     tcpAvgConnectionDuration = (99.0 * tcpAvgConnectionDuration / 100.0) + (durationMs / 100.0);
 767   }
 768 };
 769
 770 class TCPClientCollection {
 771   std::vector<int> d_tcpclientthreads;
 772   std::atomic<uint64_t> d_numthreads{0};
 773   std::atomic<uint64_t> d_pos{0};
 774   std::atomic<uint64_t> d_queued{0};
 775   const uint64_t d_maxthreads{0};
 776   std::mutex d_mutex;
 777   int d_singlePipe[2];
 778   const bool d_useSinglePipe;
 779 public:
 780
 781   TCPClientCollection(size_t maxThreads, bool useSinglePipe=false): d_maxthreads(maxThreads), d_singlePipe{-1,-1}, d_useSinglePipe(useSinglePipe)
 782
 783   {
 784     d_tcpclientthreads.reserve(maxThreads);
 785
 786     if (d_useSinglePipe) {
 787       if (pipe(d_singlePipe) < 0) {
 788         int err = errno;
 789         throw std::runtime_error("Error creating the TCP single communication pipe: " + stringerror(err));
 790       }
 791
 792       if (!setNonBlocking(d_singlePipe[0])) {
 793         int err = errno;
 794         close(d_singlePipe[0]);
 795         close(d_singlePipe[1]);
 796         throw std::runtime_error("Error setting the TCP single communication pipe non-blocking: " + stringerror(err));
 797       }
 798
 799       if (!setNonBlocking(d_singlePipe[1])) {
 800         int err = errno;
 801         close(d_singlePipe[0]);
 802         close(d_singlePipe[1]);
 803         throw std::runtime_error("Error setting the TCP single communication pipe non-blocking: " + stringerror(err));
 804       }
 805     }
 806   }
 807   int getThread()
 808   {
 809     uint64_t pos = d_pos++;
 810     ++d_queued;
 811     return d_tcpclientthreads[pos % d_numthreads];
 812   }
 813   bool hasReachedMaxThreads() const
 814   {
 815     return d_numthreads >= d_maxthreads;
 816   }
 817   uint64_t getThreadsCount() const
 818   {
 819     return d_numthreads;
 820   }
 821   uint64_t getQueuedCount() const
 822   {
 823     return d_queued;
 824   }
 825   void decrementQueuedCount()
 826   {
 827     --d_queued;
 828   }
 829   void addTCPClientThread();
 830 };
 831
 832 extern std::unique_ptr<TCPClientCollection> g_tcpclientthreads;
 833
 834 struct DownstreamState
 835 {
 836    typedef std::function<std::tuple<DNSName, uint16_t, uint16_t>(const DNSName&, uint16_t, uint16_t, dnsheader*)> checkfunc_t;
 837
 838   DownstreamState(const ComboAddress& remote_, const ComboAddress& sourceAddr_, unsigned int sourceItf, const std::string& sourceItfName, size_t numberOfSockets);
 839   DownstreamState(const ComboAddress& remote_): DownstreamState(remote_, ComboAddress(), 0, std::string(), 1) {}
 840   ~DownstreamState()
 841   {
 842     for (auto& fd : sockets) {
 843       if (fd >= 0) {
 844         close(fd);
 845         fd = -1;
 846       }
 847     }
 848   }
 849   boost::uuids::uuid id;
 850   std::set<unsigned int> hashes;
 851   mutable pthread_rwlock_t d_lock;
 852   std::vector<int> sockets;
 853   const std::string sourceItfName;
 854   std::mutex socketsLock;
 855   std::mutex connectLock;
 856   std::unique_ptr<FDMultiplexer> mplexer{nullptr};
 857   std::thread tid;
 858   const ComboAddress remote;
 859   QPSLimiter qps;
 860   vector<IDState> idStates;
 861   const ComboAddress sourceAddr;
 862   checkfunc_t checkFunction;
 863   DNSName checkName{"a.root-servers.net."};
 864   QType checkType{QType::A};
 865   uint16_t checkClass{QClass::IN};
 866   std::atomic<uint64_t> idOffset{0};
 867   std::atomic<uint64_t> sendErrors{0};
 868   std::atomic<uint64_t> outstanding{0};
 869   std::atomic<uint64_t> reuseds{0};
 870   std::atomic<uint64_t> queries{0};
 871   std::atomic<uint64_t> responses{0};
 872   struct {
 873     std::atomic<uint64_t> sendErrors{0};
 874     std::atomic<uint64_t> reuseds{0};
 875     std::atomic<uint64_t> queries{0};
 876   } prev;
 877   std::atomic<uint64_t> tcpDiedSendingQuery{0};
 878   std::atomic<uint64_t> tcpDiedReadingResponse{0};
 879   std::atomic<uint64_t> tcpGaveUp{0};
 880   std::atomic<uint64_t> tcpReadTimeouts{0};
 881   std::atomic<uint64_t> tcpWriteTimeouts{0};
 882   std::atomic<uint64_t> tcpCurrentConnections{0};
 883   std::atomic<double> tcpAvgQueriesPerConnection{0.0};
 884   /* in ms */
 885   std::atomic<double> tcpAvgConnectionDuration{0.0};
 886   string name;
 887   size_t socketsOffset{0};
 888   double queryLoad{0.0};
 889   double dropRate{0.0};
 890   double latencyUsec{0.0};
 891   int order{1};
 892   int weight{1};
 893   int tcpConnectTimeout{5};
 894   int tcpRecvTimeout{30};
 895   int tcpSendTimeout{30};
 896   unsigned int checkInterval{1};
 897   unsigned int lastCheck{0};
 898   const unsigned int sourceItf{0};
 899   uint16_t retries{5};
 900   uint16_t xpfRRCode{0};
 901   uint16_t checkTimeout{1000}; /* in milliseconds */
 902   uint8_t currentCheckFailures{0};
 903   uint8_t consecutiveSuccessfulChecks{0};
 904   uint8_t maxCheckFailures{1};
 905   uint8_t minRiseSuccesses{1};
 906   StopWatch sw;
 907   set<string> pools;
 908   enum class Availability { Up, Down, Auto} availability{Availability::Auto};
 909   bool mustResolve{false};
 910   bool upStatus{false};
 911   bool useECS{false};
 912   bool setCD{false};
 913   bool disableZeroScope{false};
 914   std::atomic<bool> connected{false};
 915   std::atomic_flag threadStarted;
 916   bool tcpFastOpen{false};
 917   bool ipBindAddrNoPort{true};
 918
 919   bool isUp() const
 920   {
 921     if(availability == Availability::Down)
 922       return false;
 923     if(availability == Availability::Up)
 924       return true;
 925     return upStatus;
 926   }
 927   void setUp() { availability = Availability::Up; }
 928   void setDown() { availability = Availability::Down; }
 929   void setAuto() { availability = Availability::Auto; }
 930   string getName() const {
 931     if (name.empty()) {
 932       return remote.toStringWithPort();
 933     }
 934     return name;
 935   }
 936   string getNameWithAddr() const {
 937     if (name.empty()) {
 938       return remote.toStringWithPort();
 939     }
 940     return name + " (" + remote.toStringWithPort()+ ")";
 941   }
 942   string getStatus() const
 943   {
 944     string status;
 945     if(availability == DownstreamState::Availability::Up)
 946       status = "UP";
 947     else if(availability == DownstreamState::Availability::Down)
 948       status = "DOWN";
 949     else
 950       status = (upStatus ? "up" : "down");
 951     return status;
 952   }
 953   bool reconnect();
 954   void hash();
 955   void setId(const boost::uuids::uuid& newId);
 956   void setWeight(int newWeight);
 957
 958   void updateTCPMetrics(size_t nbQueries, uint64_t durationMs)
 959   {
 960     tcpAvgQueriesPerConnection = (99.0 * tcpAvgQueriesPerConnection / 100.0) + (nbQueries / 100.0);
 961     tcpAvgConnectionDuration = (99.0 * tcpAvgConnectionDuration / 100.0) + (durationMs / 100.0);
 962   }
 963 };
 964 using servers_t =vector<std::shared_ptr<DownstreamState>>;
 965
 966 template <class T> using NumberedVector = std::vector<std::pair<unsigned int, T> >;
 967
 968 void responderThread(std::shared_ptr<DownstreamState> state);
 969 extern std::mutex g_luamutex;
 970 extern LuaContext g_lua;
 971 extern std::string g_outputBuffer; // locking for this is ok, as locked by g_luamutex
 972
 973 class DNSRule
 974 {
 975 public:
 976   virtual ~DNSRule ()
 977   {
 978   }
 979   virtual bool matches(const DNSQuestion* dq) const =0;
 980   virtual string toString() const = 0;
 981   mutable std::atomic<uint64_t> d_matches{0};
 982 };
 983
 984 using NumberedServerVector = NumberedVector<shared_ptr<DownstreamState>>;
 985 typedef std::function<shared_ptr<DownstreamState>(const NumberedServerVector& servers, const DNSQuestion*)> policyfunc_t;
 986
 987 struct ServerPolicy
 988 {
 989   string name;
 990   policyfunc_t policy;
 991   bool isLua;
 992   std::string toString() const {
 993     return string("ServerPolicy") + (isLua ? " (Lua)" : "") + " \"" + name + "\"";
 994   }
 995 };
 996
 997 struct ServerPool
 998 {
 999   ServerPool()
1000   {
1001     pthread_rwlock_init(&d_lock, nullptr);
1002   }
1003
1004   const std::shared_ptr<DNSDistPacketCache> getCache() const { return packetCache; };
1005
1006   bool getECS() const
1007   {
1008     return d_useECS;
1009   }
1010
1011   void setECS(bool useECS)
1012   {
1013     d_useECS = useECS;
1014   }
1015
1016   std::shared_ptr<DNSDistPacketCache> packetCache{nullptr};
1017   std::shared_ptr<ServerPolicy> policy{nullptr};
1018
1019   size_t countServers(bool upOnly)
1020   {
1021     size_t count = 0;
1022     ReadLock rl(&d_lock);
1023     for (const auto& server : d_servers) {
1024       if (!upOnly || std::get<1>(server)->isUp() ) {
1025         count++;
1026       }
1027     }
1028     return count;
1029   }
1030
1031   NumberedVector<shared_ptr<DownstreamState>> getServers()
1032   {
1033     NumberedVector<shared_ptr<DownstreamState>> result;
1034     {
1035       ReadLock rl(&d_lock);
1036       result = d_servers;
1037     }
1038     return result;
1039   }
1040
1041   void addServer(shared_ptr<DownstreamState>& server)
1042   {
1043     WriteLock wl(&d_lock);
1044     unsigned int count = (unsigned int) d_servers.size();
1045     d_servers.push_back(make_pair(++count, server));
1046     /* we need to reorder based on the server 'order' */
1047     std::stable_sort(d_servers.begin(), d_servers.end(), [](const std::pair<unsigned int,std::shared_ptr<DownstreamState> >& a, const std::pair<unsigned int,std::shared_ptr<DownstreamState> >& b) {
1048       return a.second->order < b.second->order;
1049     });
1050     /* and now we need to renumber for Lua (custom policies) */
1051     size_t idx = 1;
1052     for (auto& serv : d_servers) {
1053       serv.first = idx++;
1054     }
1055   }
1056
1057   void removeServer(shared_ptr<DownstreamState>& server)
1058   {
1059     WriteLock wl(&d_lock);
1060     size_t idx = 1;
1061     bool found = false;
1062     for (auto it = d_servers.begin(); it != d_servers.end();) {
1063       if (found) {
1064         /* we need to renumber the servers placed
1065            after the removed one, for Lua (custom policies) */
1066         it->first = idx++;
1067         it++;
1068       }
1069       else if (it->second == server) {
1070         it = d_servers.erase(it);
1071         found = true;
1072       } else {
1073         idx++;
1074         it++;
1075       }
1076     }
1077   }
1078
1079 private:
1080   NumberedVector<shared_ptr<DownstreamState>> d_servers;
1081   pthread_rwlock_t d_lock;
1082   bool d_useECS{false};
1083 };
1084 using pools_t=map<std::string,std::shared_ptr<ServerPool>>;
1085 void setPoolPolicy(pools_t& pools, const string& poolName, std::shared_ptr<ServerPolicy> policy);
1086 void addServerToPool(pools_t& pools, const string& poolName, std::shared_ptr<DownstreamState> server);
1087 void removeServerFromPool(pools_t& pools, const string& poolName, std::shared_ptr<DownstreamState> server);
1088
1089 struct CarbonConfig
1090 {
1091   ComboAddress server;
1092   std::string namespace_name;
1093   std::string ourname;
1094   std::string instance_name;
1095   unsigned int interval;
1096 };
1097
1098 enum ednsHeaderFlags {
1099   EDNS_HEADER_FLAG_NONE = 0,
1100   EDNS_HEADER_FLAG_DO = 32768
1101 };
1102
1103 struct DNSDistRuleAction
1104 {
1105   std::shared_ptr<DNSRule> d_rule;
1106   std::shared_ptr<DNSAction> d_action;
1107   boost::uuids::uuid d_id;
1108   uint64_t d_creationOrder;
1109 };
1110
1111 struct DNSDistResponseRuleAction
1112 {
1113   std::shared_ptr<DNSRule> d_rule;
1114   std::shared_ptr<DNSResponseAction> d_action;
1115   boost::uuids::uuid d_id;
1116   uint64_t d_creationOrder;
1117 };
1118
1119 extern GlobalStateHolder<SuffixMatchTree<DynBlock>> g_dynblockSMT;
1120 extern DNSAction::Action g_dynBlockAction;
1121
1122 extern GlobalStateHolder<vector<CarbonConfig> > g_carbon;
1123 extern GlobalStateHolder<ServerPolicy> g_policy;
1124 extern GlobalStateHolder<servers_t> g_dstates;
1125 extern GlobalStateHolder<pools_t> g_pools;
1126 extern GlobalStateHolder<vector<DNSDistRuleAction> > g_rulactions;
1127 extern GlobalStateHolder<vector<DNSDistResponseRuleAction> > g_resprulactions;
1128 extern GlobalStateHolder<vector<DNSDistResponseRuleAction> > g_cachehitresprulactions;
1129 extern GlobalStateHolder<vector<DNSDistResponseRuleAction> > g_selfansweredresprulactions;
1130 extern GlobalStateHolder<NetmaskGroup> g_ACL;
1131
1132 extern ComboAddress g_serverControl; // not changed during runtime
1133
1134 extern std::vector<std::tuple<ComboAddress, bool, bool, int, std::string, std::set<int>>> g_locals; // not changed at runtime (we hope XXX)
1135 extern std::vector<shared_ptr<TLSFrontend>> g_tlslocals;
1136 extern std::vector<shared_ptr<DOHFrontend>> g_dohlocals;
1137 extern std::vector<std::unique_ptr<ClientState>> g_frontends;
1138 extern bool g_truncateTC;
1139 extern bool g_fixupCase;
1140 extern int g_tcpRecvTimeout;
1141 extern int g_tcpSendTimeout;
1142 extern int g_udpTimeout;
1143 extern uint16_t g_maxOutstanding;
1144 extern std::atomic<bool> g_configurationDone;
1145 extern uint64_t g_maxTCPClientThreads;
1146 extern uint64_t g_maxTCPQueuedConnections;
1147 extern size_t g_maxTCPQueriesPerConn;
1148 extern size_t g_maxTCPConnectionDuration;
1149 extern size_t g_maxTCPConnectionsPerClient;
1150 extern std::atomic<uint16_t> g_cacheCleaningDelay;
1151 extern std::atomic<uint16_t> g_cacheCleaningPercentage;
1152 extern bool g_verboseHealthChecks;
1153 extern uint32_t g_staleCacheEntriesTTL;
1154 extern bool g_apiReadWrite;
1155 extern std::string g_apiConfigDirectory;
1156 extern bool g_servFailOnNoPolicy;
1157 extern uint32_t g_hashperturb;
1158 extern bool g_useTCPSinglePipe;
1159 extern uint16_t g_downstreamTCPCleanupInterval;
1160 extern size_t g_udpVectorSize;
1161 extern bool g_preserveTrailingData;
1162 extern bool g_allowEmptyResponse;
1163 extern bool g_roundrobinFailOnNoServer;
1164
1165 #ifdef HAVE_EBPF
1166 extern shared_ptr<BPFFilter> g_defaultBPFFilter;
1167 extern std::vector<std::shared_ptr<DynBPFFilter> > g_dynBPFFilters;
1168 #endif /* HAVE_EBPF */
1169
1170 struct LocalHolders
1171 {
1172   LocalHolders(): acl(g_ACL.getLocal()), policy(g_policy.getLocal()), rulactions(g_rulactions.getLocal()), cacheHitRespRulactions(g_cachehitresprulactions.getLocal()), selfAnsweredRespRulactions(g_selfansweredresprulactions.getLocal()), servers(g_dstates.getLocal()), dynNMGBlock(g_dynblockNMG.getLocal()), dynSMTBlock(g_dynblockSMT.getLocal()), pools(g_pools.getLocal())
1173   {
1174   }
1175
1176   LocalStateHolder<NetmaskGroup> acl;
1177   LocalStateHolder<ServerPolicy> policy;
1178   LocalStateHolder<vector<DNSDistRuleAction> > rulactions;
1179   LocalStateHolder<vector<DNSDistResponseRuleAction> > cacheHitRespRulactions;
1180   LocalStateHolder<vector<DNSDistResponseRuleAction> > selfAnsweredRespRulactions;
1181   LocalStateHolder<servers_t> servers;
1182   LocalStateHolder<NetmaskTree<DynBlock> > dynNMGBlock;
1183   LocalStateHolder<SuffixMatchTree<DynBlock> > dynSMTBlock;
1184   LocalStateHolder<pools_t> pools;
1185 };
1186
1187 struct dnsheader;
1188
1189 void controlThread(int fd, ComboAddress local);
1190 vector<std::function<void(void)>> setupLua(bool client, const std::string& config);
1191 std::shared_ptr<ServerPool> getPool(const pools_t& pools, const std::string& poolName);
1192 std::shared_ptr<ServerPool> createPoolIfNotExists(pools_t& pools, const string& poolName);
1193 NumberedServerVector getDownstreamCandidates(const pools_t& pools, const std::string& poolName);
1194
1195 std::shared_ptr<DownstreamState> firstAvailable(const NumberedServerVector& servers, const DNSQuestion* dq);
1196
1197 std::shared_ptr<DownstreamState> leastOutstanding(const NumberedServerVector& servers, const DNSQuestion* dq);
1198 std::shared_ptr<DownstreamState> wrandom(const NumberedServerVector& servers, const DNSQuestion* dq);
1199 std::shared_ptr<DownstreamState> whashed(const NumberedServerVector& servers, const DNSQuestion* dq);
1200 std::shared_ptr<DownstreamState> chashed(const NumberedServerVector& servers, const DNSQuestion* dq);
1201 std::shared_ptr<DownstreamState> roundrobin(const NumberedServerVector& servers, const DNSQuestion* dq);
1202
1203 struct WebserverConfig
1204 {
1205   std::string password;
1206   std::string apiKey;
1207   boost::optional<std::map<std::string, std::string> > customHeaders;
1208   std::mutex lock;
1209 };
1210
1211 void setWebserverAPIKey(const boost::optional<std::string> apiKey);
1212 void setWebserverPassword(const std::string& password);
1213 void setWebserverCustomHeaders(const boost::optional<std::map<std::string, std::string> > customHeaders);
1214
1215 void dnsdistWebserverThread(int sock, const ComboAddress& local);
1216 void tcpAcceptorThread(void* p);
1217 #ifdef HAVE_DNS_OVER_HTTPS
1218 void dohThread(ClientState* cs);
1219 #endif /* HAVE_DNS_OVER_HTTPS */
1220
1221 void setLuaNoSideEffect(); // if nothing has been declared, set that there are no side effects
1222 void setLuaSideEffect();   // set to report a side effect, cancelling all _no_ side effect calls
1223 bool getLuaNoSideEffect(); // set if there were only explicit declarations of _no_ side effect
1224 void resetLuaSideEffect(); // reset to indeterminate state
1225
1226 bool responseContentMatches(const char* response, const uint16_t responseLen, const DNSName& qname, const uint16_t qtype, const uint16_t qclass, const ComboAddress& remote, unsigned int& consumed);
1227 bool processResponse(char** response, uint16_t* responseLen, size_t* responseSize, LocalStateHolder<vector<DNSDistResponseRuleAction> >& localRespRulactions, DNSResponse& dr, size_t addRoom, std::vector<uint8_t>& rewrittenResponse, bool muted);
1228 bool processRulesResult(const DNSAction::Action& action, DNSQuestion& dq, std::string& ruleresult, bool& drop);
1229
1230 bool checkQueryHeaders(const struct dnsheader* dh);
1231
1232 extern std::vector<std::shared_ptr<DNSCryptContext>> g_dnsCryptLocals;
1233 int handleDNSCryptQuery(char* packet, uint16_t len, std::shared_ptr<DNSCryptQuery> query, uint16_t* decryptedQueryLen, bool tcp, time_t now, std::vector<uint8_t>& response);
1234 boost::optional<std::vector<uint8_t>> checkDNSCryptQuery(const ClientState& cs, const char* query, uint16_t& len, std::shared_ptr<DNSCryptQuery>& dnsCryptQuery, time_t now, bool tcp);
1235
1236 bool addXPF(DNSQuestion& dq, uint16_t optionCode);
1237
1238 uint16_t getRandomDNSID();
1239
1240 #include "dnsdist-snmp.hh"
1241
1242 extern bool g_snmpEnabled;
1243 extern bool g_snmpTrapsEnabled;
1244 extern DNSDistSNMPAgent* g_snmpAgent;
1245 extern bool g_addEDNSToSelfGeneratedResponses;
1246
1247 extern std::set<std::string> g_capabilitiesToRetain;
1248 static const uint16_t s_udpIncomingBufferSize{1500}; // don't accept UDP queries larger than this value
1249 static const size_t s_maxPacketCacheEntrySize{4096}; // don't cache responses larger than this value
1250
1251 enum class ProcessQueryResult { Drop, SendAnswer, PassToBackend };
1252 ProcessQueryResult processQuery(DNSQuestion& dq, ClientState& cs, LocalHolders& holders, std::shared_ptr<DownstreamState>& selectedBackend);
1253
1254 DNSResponse makeDNSResponseFromIDState(IDState& ids, struct dnsheader* dh, size_t bufferSize, uint16_t responseLen, bool isTCP);
1255 void setIDStateFromDNSQuestion(IDState& ids, DNSQuestion& dq, DNSName&& qname);
1256
1257 int pickBackendSocketForSending(std::shared_ptr<DownstreamState>& state);
1258 ssize_t udpClientSendRequestToBackend(const std::shared_ptr<DownstreamState>& ss, const int sd, const char* request, const size_t requestLen, bool healthCheck=false);