modules/lmdbbackend/lmdbbackend.cc

   1 /*
   2  * LMDBBackend - a high performance LMDB based backend for PowerDNS written by
   3  * Mark Zealey, 2013
   4  *
   5  * This was originally going to be a backend using BerkeleyDB 5 for high
   6  * performance DNS over massive (millions of zones) databases. However,
   7  * BerkeleyDB had a number of issues to do with locking, contention and
   8  * corruption which made it unsuitable for use. Instead, we use LMDB to perform
   9  * very fast lookups.
  10  *
  11  * See the documentation for more details, and lmdb-example.pl for an example
  12  * script which generates a simple zone.
  13  */
  14
  15 #include "pdns/utility.hh"
  16 #include "pdns/dnsbackend.hh"
  17 #include "pdns/dns.hh"
  18 #include "pdns/dnspacket.hh"
  19 #include "pdns/pdnsexception.hh"
  20 #include "pdns/logger.hh"
  21 #include <signal.h>
  22 #include "lmdbbackend.hh"
  23 #include "pdns/arguments.hh"
  24 #include "pdns/base32.hh"
  25 #include "pdns/lock.hh"
  26
  27 #if 0
  28 #define DEBUGLOG(msg) L<<Logger::Error<<msg
  29 #else
  30 #define DEBUGLOG(msg) do {} while(0)
  31 #endif
  32
  33 int LMDBBackend::s_reloadcount=0;
  34 pthread_mutex_t LMDBBackend::s_initlock = PTHREAD_MUTEX_INITIALIZER;
  35
  36 LMDBBackend::LMDBBackend(const string &suffix)
  37 {
  38     setArgPrefix("lmdb"+suffix);
  39     try {
  40       d_doDnssec = mustDo("experimental-dnssec");
  41     }
  42     catch (ArgException e) {
  43       d_doDnssec = false;
  44     }
  45     d_lastreload = s_reloadcount;
  46     open_db();
  47 }
  48
  49 void LMDBBackend::open_db() {
  50     L<<Logger::Error<<"Loading LMDB database " << getArg("datapath") << endl;
  51
  52     string path = getArg("datapath");
  53     int rc;
  54     int major, minor, patch;
  55
  56     string verstring( mdb_version( &major, &minor, &patch ) );
  57     if( MDB_VERINT( major, minor, patch ) < MDB_VERINT( 0, 9, 8 ) )
  58         throw PDNSException( "LMDB Library version too old (" + verstring + "). Needs to be 0.9.8 or greater" );
  59
  60     Lock l(&s_initlock);
  61
  62     if( (rc = mdb_env_create(&env))  )
  63         throw PDNSException("Couldn't open LMDB database " + path + ": mdb_env_create() returned " + mdb_strerror(rc));
  64
  65     if( (rc = mdb_env_set_maxdbs( env, d_doDnssec ? 5 : 3)) )
  66         throw PDNSException("Couldn't open LMDB database " + path + ": mdb_env_set_maxdbs() returned " + mdb_strerror(rc));
  67
  68     if( (rc = mdb_env_open(env, path.c_str(), MDB_RDONLY, 0)) )
  69         throw PDNSException("Couldn't open LMDB database " + path + ": mdb_env_open() returned " + mdb_strerror(rc));
  70
  71     if( (rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn) ))
  72         throw PDNSException("Couldn't start LMDB txn " + path + ": mdb_txn_begin() returned " + mdb_strerror(rc));
  73
  74     if( (rc = mdb_dbi_open(txn, "zone", 0, &zone_db) ) )
  75         throw PDNSException("Couldn't open LMDB zone database " + path + ": mdb_dbi_open() returned " + mdb_strerror(rc));
  76     if( (rc = mdb_cursor_open(txn, zone_db, &zone_cursor) ))
  77         throw PDNSException("Couldn't open cursor on LMDB zone database " + path + ": mdb_cursor_open() returned " + mdb_strerror(rc));
  78
  79     if( (rc = mdb_dbi_open(txn, "data", MDB_DUPSORT, &data_db) ))
  80         throw PDNSException("Couldn't open LMDB data database " + path + ": mdb_dbi_open() returned " + mdb_strerror(rc));
  81     if( (rc = mdb_cursor_open(txn, data_db, &data_cursor) ))
  82         throw PDNSException("Couldn't open cursor on LMDB data database " + path + ": mdb_cursor_open() returned " + mdb_strerror(rc));
  83
  84     if( (rc = mdb_dbi_open(txn, "extended_data", 0, &data_extended_db) ))
  85         throw PDNSException("Couldn't open LMDB extended_data database " + path + ": mdb_dbi_open() returned " + mdb_strerror(rc));
  86     if( ( rc = mdb_cursor_open(txn, data_extended_db, &data_extended_cursor)) )
  87         throw PDNSException("Couldn't open cursor on LMDB data_extended database " + path + ": mdb_cursor_open() returned " + mdb_strerror(rc));
  88
  89     if(d_doDnssec) {
  90       DEBUGLOG("Experimental dnssec support enabled"<<endl);
  91       if( (rc = mdb_dbi_open(txn, "rrsig", MDB_DUPSORT, &rrsig_db) ))
  92           throw PDNSException("Couldn't open LMDB rrsig database " + path + ": mdb_dbi_open() returned " + mdb_strerror(rc));
  93       if( ( rc = mdb_cursor_open(txn, rrsig_db, &rrsig_cursor)) )
  94           throw PDNSException("Couldn't open cursor on LMDB rrsig database " + path + ": mdb_cursor_open() returned " + mdb_strerror(rc));
  95
  96       if( (rc = mdb_dbi_open(txn, "nsecx", 0, &nsecx_db) ))
  97           throw PDNSException("Couldn't open LMDB nsecx database " + path + ": mdb_dbi_open() returned " + mdb_strerror(rc));
  98       if( ( rc = mdb_cursor_open(txn, nsecx_db, &nsecx_cursor)) )
  99           throw PDNSException("Couldn't open cursor on LMDB nsecx database " + path + ": mdb_cursor_open() returned " + mdb_strerror(rc));
 100     }
 101 }
 102
 103 void LMDBBackend::close_db() {
 104     L<<Logger::Error<<"Closing LMDB database"<< endl;
 105
 106     mdb_cursor_close(data_cursor);
 107     mdb_cursor_close(zone_cursor);
 108     mdb_cursor_close(data_extended_cursor);
 109     mdb_dbi_close(env, data_db);
 110     mdb_dbi_close(env, zone_db);
 111     mdb_dbi_close(env, data_extended_db);
 112     if (d_doDnssec) {
 113       mdb_cursor_close(rrsig_cursor);
 114       mdb_cursor_close(nsecx_cursor);
 115       mdb_dbi_close(env, rrsig_db);
 116       mdb_dbi_close(env, nsecx_db);
 117     }
 118     mdb_txn_abort(txn);
 119     mdb_env_close(env);
 120 }
 121
 122 LMDBBackend::~LMDBBackend()
 123 {
 124     close_db();
 125 }
 126
 127 void LMDBBackend::reload() {
 128   ++s_reloadcount;
 129 }
 130
 131 void LMDBBackend::needReload() {
 132   if (s_reloadcount > d_lastreload) {
 133     d_lastreload = s_reloadcount;
 134     close_db();
 135     open_db();
 136   }
 137 }
 138
 139 bool LMDBBackend::getDomainMetadata(const string& name, const std::string& kind, std::vector<std::string>& meta)
 140 {
 141   if (!d_doDnssec)
 142     return false;
 143
 144   needReload();
 145
 146   if (kind == "PRESIGNED" || kind == "NSEC3PARAM") {
 147     int rc;
 148     MDB_val key, data;
 149     string key_str, cur_value;
 150     vector<string> valparts;
 151
 152     key_str=bitFlip(labelReverse(toLower(name)))+"\xff";
 153     key.mv_data = (char *)key_str.c_str();
 154     key.mv_size = key_str.length();
 155
 156     if ((rc = mdb_cursor_get(zone_cursor, &key, &data, MDB_SET_KEY)) == 0) {
 157       cur_value.assign((const char *)data.mv_data, data.mv_size);
 158       stringtok(valparts,cur_value,"\t");
 159
 160       if (valparts.size() == 4) {
 161         if (kind == "PRESIGNED")
 162           meta.push_back("1");
 163         else if (valparts[3] != "1")
 164           meta.push_back(valparts[3]);
 165       }
 166     }
 167
 168     if (rc == MDB_NOTFOUND)
 169       DEBUGLOG("Metadata records for zone: '"<<name<<"'' not found. This is impossible !!!"<<endl);
 170   }
 171
 172   return true;
 173 }
 174
 175 bool LMDBBackend::getDirectNSECx(uint32_t id, const string &hashed, const QType &qtype, string &before, DNSResourceRecord &rr)
 176 {
 177   if (!d_doDnssec)
 178     return false;
 179
 180   needReload();
 181
 182   MDB_val key, data;
 183   string key_str, cur_key, cur_value;
 184   vector<string> keyparts, valparts;
 185
 186   if (qtype == QType::NSEC)
 187     key_str=itoa(id)+"\t"+bitFlip(hashed)+"\xff";
 188   else
 189     key_str=itoa(id)+"\t"+toBase32Hex(bitFlip(hashed));
 190   key.mv_data = (char *)key_str.c_str();
 191   key.mv_size = key_str.length();
 192
 193   before.clear();
 194   if(!mdb_cursor_get(nsecx_cursor, &key, &data, MDB_SET_RANGE)) {
 195     cur_key.assign((const char *)key.mv_data, key.mv_size);
 196     cur_value.assign((const char *)data.mv_data, data.mv_size);
 197     stringtok(keyparts,cur_key,"\t");
 198     stringtok(valparts,cur_value,"\t");
 199
 200     if( keyparts.size() != 2 || valparts.size() != 4 ) {
 201       throw PDNSException("Invalid record in nsecx table: key: '" + cur_key + "'; value: "+ cur_value);
 202     }
 203
 204     // is the key a full match or does the id part match our zone?
 205     // if it does we have a valid answer.
 206     if (!key_str.compare(cur_key) || atoi(keyparts[0].c_str()) == (int) id) // FIXME we need atoui
 207       goto hasnsecx;
 208   }
 209   // no match, now we look for the last record in the NSECx chain.
 210   key_str=itoa(id)+"\t";
 211   key.mv_data = (char *)key_str.c_str();
 212   key.mv_size = key_str.length();
 213
 214   if(!mdb_cursor_get(nsecx_cursor, &key, &data, MDB_NEXT_NODUP )) {
 215     cur_key.assign((const char *)key.mv_data, key.mv_size);
 216     cur_value.assign((const char *)data.mv_data, data.mv_size);
 217     stringtok(keyparts,cur_key,"\t");
 218     stringtok(valparts,cur_value,"\t");
 219
 220     if( keyparts.size() != 2 || valparts.size() != 4 ) {
 221       throw PDNSException("Invalid record in nsecx table: key: '" + cur_key + "'; value: "+ cur_value);
 222     }
 223
 224     if (!key_str.compare(cur_key) || atoi(keyparts[0].c_str()) == (int) id) // FIXME we need atoui
 225       goto hasnsecx;
 226   }
 227
 228   DEBUGLOG("NSECx record for '"<<toBase32Hex(bitFlip(hashed))<<"'' in zone '"<<id<<"' not found"<<endl);
 229   return true;
 230
 231 hasnsecx:
 232   if (qtype == QType::NSEC)
 233     before=bitFlip(keyparts[1]).c_str();
 234   else
 235     before=bitFlip(fromBase32Hex(keyparts[1]));
 236   rr.qname=valparts[0];
 237   rr.ttl=atoi(valparts[1].c_str());
 238   rr.qtype=DNSRecordContent::TypeToNumber(valparts[2]);
 239   rr.content=valparts[3];
 240   rr.d_place=DNSResourceRecord::AUTHORITY;
 241   rr.domain_id=id;
 242   rr.auth=true;
 243
 244   return true;
 245 }
 246
 247 bool LMDBBackend::getDirectRRSIGs(const string &signer, const string &qname, const QType &qtype, vector<DNSResourceRecord> &rrsigs)
 248 {
 249   if (!d_doDnssec)
 250     return false;
 251
 252   needReload();
 253
 254   int rc;
 255   MDB_val key, data;
 256   string key_str, cur_value;
 257   vector<string> valparts;
 258
 259   key_str=signer+"\t"+makeRelative(qname, signer)+"\t"+qtype.getName();
 260   key.mv_data = (char *)key_str.c_str();
 261   key.mv_size = key_str.length();
 262
 263   if ((rc = mdb_cursor_get(rrsig_cursor, &key, &data, MDB_SET_KEY)) == 0) {
 264     DNSResourceRecord rr;
 265     rr.qname=qname;
 266     rr.qtype=QType::RRSIG;
 267     //rr.d_place = (DNSResourceRecord::Place) signPlace;
 268     rr.auth=false;
 269
 270     do {
 271       cur_value.assign((const char *)data.mv_data, data.mv_size);
 272       stringtok(valparts,cur_value,"\t");
 273
 274       if( valparts.size() != 2 ) {
 275         throw PDNSException("Invalid record in rrsig table: qname: '" + qname + "'; value: "+ cur_value);
 276       }
 277
 278       rr.ttl=atoi(valparts[0].c_str());
 279       rr.content = valparts[1];
 280       rrsigs.push_back(rr);
 281
 282     } while (mdb_cursor_get(rrsig_cursor, &key, &data, MDB_NEXT_DUP) == 0);
 283   }
 284
 285   if (rc == MDB_NOTFOUND)
 286     DEBUGLOG("RRSIG records for qname: '"<<qname<<"'' with type: '"<<qtype.getName()<<"' not found"<<endl);
 287
 288   return true;
 289 }
 290
 291 // Get the zone name of the requested zone (labelReversed) OR the name of the closest parrent zone
 292 bool LMDBBackend::getAuthZone( string &rev_zone )
 293 {
 294     needReload();
 295
 296     MDB_val key, data;
 297     // XXX can do this just using char *
 298
 299     string key_str=bitFlip(rev_zone+" ");
 300     key.mv_data = (char *)key_str.c_str();
 301     key.mv_size = key_str.length();
 302
 303     // Release our transaction and cursors in order to get latest data
 304     mdb_txn_reset( txn );
 305     mdb_txn_renew( txn );
 306     mdb_cursor_renew( txn, zone_cursor );
 307     mdb_cursor_renew( txn, data_cursor );
 308     mdb_cursor_renew( txn, data_extended_cursor );
 309     if (d_doDnssec) {
 310       mdb_cursor_renew( txn, rrsig_cursor );
 311       mdb_cursor_renew( txn, nsecx_cursor );
 312     }
 313
 314     // Find the best record
 315     if( mdb_cursor_get( zone_cursor, &key, &data, MDB_SET_RANGE ) == 0 && key.mv_size <= key_str.length() ) {
 316       // Found a shorter match. Now look if the zones are equal up to key-length-1. If they are check
 317       // if position key-length in key_str is a label separator. If all this is true we have a match.
 318       if( key_str.compare( 0, key.mv_size-1, (const char *) key.mv_data, key.mv_size-1  ) == 0 && key.mv_size && key_str[key.mv_size-1] == ~' ') {
 319         rev_zone.resize( key.mv_size-1 );
 320
 321         DEBUGLOG("Auth key: " << rev_zone <<endl);
 322
 323         return true;
 324       }
 325     }
 326
 327     //reset the cursor the data in it is invallid
 328     mdb_cursor_renew( txn, zone_cursor );
 329
 330     return false;
 331 }
 332
 333 bool LMDBBackend::getAuthData( SOAData &soa, DNSPacket *p )
 334 {
 335     needReload();
 336
 337     MDB_val key, value;
 338     if( mdb_cursor_get(zone_cursor, &key, &value, MDB_GET_CURRENT) )
 339         return false;
 340
 341     string data( (const char *)value.mv_data, value.mv_size );
 342     DEBUGLOG("Auth record data " << data<<endl);
 343
 344 // XXX do this in C too
 345     vector<string>parts;
 346     stringtok(parts,data,"\t");
 347
 348     if(parts.size() < 3)
 349         throw PDNSException("Invalid record in zone table: " + data );
 350
 351     fillSOAData( parts[2], soa );
 352
 353     soa.domain_id = atoi( parts[0].c_str() );
 354     soa.ttl = atoi( parts[1].c_str() );
 355
 356     soa.scopeMask = 0;
 357     soa.db = this;
 358
 359     return true;
 360 }
 361
 362 // Called to start an AXFR then ->get() is called. Return true if the domain exists
 363 bool LMDBBackend::list(const string &target, int zoneId, bool include_disabled) {
 364     DEBUGLOG("list() requested for " <<target << endl);
 365     d_first = true;
 366     d_origdomain = target;
 367     d_domain_id = zoneId;
 368     d_curqtype = QType::AXFR;
 369
 370     // getSOA will have been called first to ensure the domain exists so if
 371     // that's the case then there's no reason we can't AXFR it.
 372
 373     return true;
 374 }
 375
 376 void LMDBBackend::lookup(const QType &type, const string &inQdomain, DNSPacket *p, int zoneId)
 377 {
 378     DEBUGLOG("lookup: " <<inQdomain << " " << type.getName() << endl);
 379
 380     needReload();
 381
 382     d_first = true;
 383     d_origdomain = inQdomain;
 384     d_curqtype = type;
 385 }
 386
 387 inline bool LMDBBackend::get_finished()
 388 {
 389     d_origdomain = "";
 390
 391     return false;
 392 }
 393
 394 bool LMDBBackend::get(DNSResourceRecord &rr)
 395 {
 396     MDB_val key, value;
 397     bool is_axfr = (d_curqtype == QType::AXFR);
 398     bool is_full_key = ( ! is_axfr && d_curqtype != QType::ANY );
 399
 400     DEBUGLOG("get : " <<d_origdomain<< endl);
 401     if( !d_origdomain.length() )
 402         return false;
 403
 404     DEBUGLOG("Starting Q " << d_first<< endl);
 405
 406     if( d_first ) {
 407         d_first = false;
 408
 409         // Reverse the query string
 410         string lowerq = toLower( d_origdomain );
 411         d_querykey = string( lowerq.rbegin(), lowerq.rend() );
 412         d_searchkey = d_querykey;
 413
 414         // For normal queries ensure that we are only trying to get the exact
 415         // record and also try to specify the type too to make negatives a lot
 416         // quicker
 417         if( ! is_axfr ) {
 418             d_searchkey += "\t";
 419
 420             // Search by query type too to easily exclude anything that doesn't
 421             // belong to us
 422             if( is_full_key )
 423                 d_searchkey += d_curqtype.getName();
 424         }
 425
 426         key.mv_size = d_searchkey.length();
 427         key.mv_data = (char *)d_searchkey.c_str();
 428         if( mdb_cursor_get(data_cursor, &key, &value, is_full_key ? MDB_SET_KEY : MDB_SET_RANGE ) )
 429             return get_finished();
 430     } else {
 431 next_record:
 432         key.mv_size = 0;
 433         if( mdb_cursor_get(data_cursor, &key, &value, is_full_key ? MDB_NEXT_DUP : MDB_NEXT ) )
 434             return get_finished();
 435     }
 436
 437     // Some buggy versions of lmdb will do this. Should be caught in opendb above though.
 438     if( key.mv_size == 0 ) {
 439         DEBUGLOG("No key returned. Error" << endl);
 440         return get_finished();
 441     }
 442
 443     string cur_value((const char *)value.mv_data, value.mv_size);
 444     string cur_key((const char *)key.mv_data, key.mv_size);
 445
 446     DEBUGLOG("querykey: " << d_querykey << "; cur_key: " <<cur_key<< "; cur_value: '" << cur_value << "'" << endl);
 447
 448     vector<string> keyparts, valparts;
 449
 450     stringtok(keyparts,cur_key,"\t");
 451     stringtok(valparts,cur_value,"\t");
 452
 453     if( valparts.size() == 2 && valparts[0] == "REF" ) {
 454         MDB_val extended_key, extended_val;
 455
 456         // XXX parse into an int and have extended table as MDB_INTEGER to have
 457         // a bit better performance/smaller space?
 458         extended_key.mv_data = (char *)valparts[1].c_str();
 459         extended_key.mv_size = valparts[1].length();
 460
 461         if( int rc = mdb_cursor_get( data_extended_cursor, &extended_key, &extended_val, MDB_SET_KEY ) )
 462             throw PDNSException("Record " + cur_key + " references extended record " + cur_value + " but this doesn't exist: " + mdb_strerror( rc ));
 463
 464         cur_value.assign((const char *)extended_val.mv_data, extended_val.mv_size);
 465         valparts.clear();
 466         stringtok(valparts, cur_value, "\t");
 467     }
 468
 469     if (valparts.size() != 3) // FIXME
 470       valparts.push_back(".");
 471
 472     if( keyparts.size() != 2 || valparts.size() != 3 )
 473         throw PDNSException("Invalid record in record table: key: '" + cur_key + "'; value: '"+ cur_value+"'");
 474
 475     string compare_string = cur_key.substr(0, d_searchkey.length());
 476     DEBUGLOG( "searchkey: " << d_searchkey << "; compare: " << compare_string << ";" << endl);
 477
 478     // If we're onto records not beginning with this search prefix, then we
 479     // must be past the end
 480     if( compare_string.compare( d_searchkey ) )
 481         return get_finished();
 482
 483     int domain_id = atoi( valparts[0].c_str() );
 484
 485     // If we are doing an AXFR and the record fetched has been outside of our domain then end the transfer
 486     if( is_axfr ) {
 487         // Check it's not a subdomain ie belongs to this record
 488         if( domain_id != d_domain_id )
 489             goto next_record;
 490
 491         // If it's under the main domain then append the . to the comparison to
 492         // ensure items outside our zone don't enter
 493         if( keyparts[0].length() > d_querykey.length() ) {
 494             string test = d_querykey;
 495             test.append(".");
 496
 497             compare_string = cur_key.substr(0, d_querykey.length() + 1);
 498
 499             DEBUGLOG("test: " << test << "; compare: " << compare_string << ";" << endl);
 500
 501             if( test.compare( compare_string ) )
 502                 goto next_record;
 503         }
 504
 505         // We need to maintain query casing so strip off domain (less dot) and append originial query
 506         string sub = keyparts[0].substr( d_origdomain.length(), string::npos );
 507         rr.qname = string( sub.rbegin(), sub.rend() ) + d_origdomain;
 508     } else
 509         rr.qname = d_origdomain; // use cached and original casing
 510
 511     DEBUGLOG("Found record: " <<cur_key << ": "<<valparts.size() << endl);
 512
 513     DEBUGLOG("pass! " << rr.qname << ";" << endl);
 514     rr.qtype = keyparts[1];
 515
 516     /* Filter records to only match query type */
 517     if( d_curqtype != QType::ANY && !is_axfr && rr.qtype != d_curqtype )
 518         goto next_record;
 519
 520     DEBUGLOG("Correct record type" << endl);
 521     rr.auth = 1;
 522
 523     rr.domain_id = domain_id;
 524     rr.ttl = atoi( valparts[1].c_str() );
 525     rr.content = valparts[2];
 526
 527     return true;
 528 }
 529
 530 class LMDBFactory : public BackendFactory
 531 {
 532 public:
 533   LMDBFactory() : BackendFactory("lmdb") {}
 534   void declareArguments(const string &suffix="")
 535   {
 536     declare(suffix,"datapath","Path to the directory containing the lmdb files","/etc/pdns/data");
 537     declare(suffix,"experimental-dnssec","Enable experimental DNSSEC processing","no");
 538   }
 539   DNSBackend *make(const string &suffix="")
 540   {
 541     return new LMDBBackend(suffix);
 542   }
 543 };
 544
 545 /* THIRD PART */
 546
 547 class LMDBLoader
 548 {
 549 public:
 550   LMDBLoader()
 551   {
 552     BackendMakers().report(new LMDBFactory);
 553     L << Logger::Info << "[lmdbbackend] This is the lmdb backend version " VERSION " (" __DATE__ ", " __TIME__ ") reporting" << endl;
 554   }
 555 };
 556
 557 static LMDBLoader lmdbLoader;
 558