Style and IDEA warning cleanups
[lhc/web/wiklou.git] / includes / db / loadbalancer / LBFactory.php
1 <?php
2 /**
3 * Generator of database load balancing objects.
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, write to the Free Software Foundation, Inc.,
17 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
18 * http://www.gnu.org/copyleft/gpl.html
19 *
20 * @file
21 * @ingroup Database
22 */
23
24 use MediaWiki\MediaWikiServices;
25 use MediaWiki\Services\DestructibleService;
26 use Psr\Log\LoggerInterface;
27 use MediaWiki\Logger\LoggerFactory;
28
29 /**
30 * An interface for generating database load balancers
31 * @ingroup Database
32 */
33 abstract class LBFactory implements DestructibleService {
34 /** @var ChronologyProtector */
35 protected $chronProt;
36 /** @var TransactionProfiler */
37 protected $trxProfiler;
38 /** @var LoggerInterface */
39 protected $trxLogger;
40 /** @var BagOStuff */
41 protected $srvCache;
42 /** @var WANObjectCache */
43 protected $wanCache;
44
45 /** @var mixed */
46 protected $ticket;
47 /** @var string|bool String if a requested DBO_TRX transaction round is active */
48 protected $trxRoundId = false;
49 /** @var string|bool Reason all LBs are read-only or false if not */
50 protected $readOnlyReason = false;
51 /** @var callable[] */
52 protected $replicationWaitCallbacks = [];
53
54 const SHUTDOWN_NO_CHRONPROT = 1; // don't save ChronologyProtector positions (for async code)
55
56 /**
57 * Construct a factory based on a configuration array (typically from $wgLBFactoryConf)
58 * @param array $conf
59 * @TODO: inject objects via dependency framework
60 */
61 public function __construct( array $conf ) {
62 if ( isset( $conf['readOnlyReason'] ) && is_string( $conf['readOnlyReason'] ) ) {
63 $this->readOnlyReason = $conf['readOnlyReason'];
64 }
65 $this->chronProt = $this->newChronologyProtector();
66 $this->trxProfiler = Profiler::instance()->getTransactionProfiler();
67 // Use APC/memcached style caching, but avoids loops with CACHE_DB (T141804)
68 $cache = ObjectCache::getLocalServerInstance();
69 if ( $cache->getQoS( $cache::ATTR_EMULATION ) > $cache::QOS_EMULATION_SQL ) {
70 $this->srvCache = $cache;
71 } else {
72 $this->srvCache = new EmptyBagOStuff();
73 }
74 $wCache = ObjectCache::getMainWANInstance();
75 if ( $wCache->getQoS( $wCache::ATTR_EMULATION ) > $wCache::QOS_EMULATION_SQL ) {
76 $this->wanCache = $wCache;
77 } else {
78 $this->wanCache = WANObjectCache::newEmpty();
79 }
80 $this->trxLogger = LoggerFactory::getInstance( 'DBTransaction' );
81 $this->ticket = mt_rand();
82 }
83
84 /**
85 * Disables all load balancers. All connections are closed, and any attempt to
86 * open a new connection will result in a DBAccessError.
87 * @see LoadBalancer::disable()
88 */
89 public function destroy() {
90 $this->shutdown();
91 $this->forEachLBCallMethod( 'disable' );
92 }
93
94 /**
95 * Disables all access to the load balancer, will cause all database access
96 * to throw a DBAccessError
97 */
98 public static function disableBackend() {
99 MediaWikiServices::disableStorageBackend();
100 }
101
102 /**
103 * Get an LBFactory instance
104 *
105 * @deprecated since 1.27, use MediaWikiServices::getDBLoadBalancerFactory() instead.
106 *
107 * @return LBFactory
108 */
109 public static function singleton() {
110 return MediaWikiServices::getInstance()->getDBLoadBalancerFactory();
111 }
112
113 /**
114 * Returns the LBFactory class to use and the load balancer configuration.
115 *
116 * @todo instead of this, use a ServiceContainer for managing the different implementations.
117 *
118 * @param array $config (e.g. $wgLBFactoryConf)
119 * @return string Class name
120 */
121 public static function getLBFactoryClass( array $config ) {
122 // For configuration backward compatibility after removing
123 // underscores from class names in MediaWiki 1.23.
124 $bcClasses = [
125 'LBFactory_Simple' => 'LBFactorySimple',
126 'LBFactory_Single' => 'LBFactorySingle',
127 'LBFactory_Multi' => 'LBFactoryMulti',
128 'LBFactory_Fake' => 'LBFactoryFake',
129 ];
130
131 $class = $config['class'];
132
133 if ( isset( $bcClasses[$class] ) ) {
134 $class = $bcClasses[$class];
135 wfDeprecated(
136 '$wgLBFactoryConf must be updated. See RELEASE-NOTES for details',
137 '1.23'
138 );
139 }
140
141 return $class;
142 }
143
144 /**
145 * Shut down, close connections and destroy the cached instance.
146 *
147 * @deprecated since 1.27, use LBFactory::destroy()
148 */
149 public static function destroyInstance() {
150 self::singleton()->destroy();
151 }
152
153 /**
154 * Create a new load balancer object. The resulting object will be untracked,
155 * not chronology-protected, and the caller is responsible for cleaning it up.
156 *
157 * @param bool|string $wiki Wiki ID, or false for the current wiki
158 * @return LoadBalancer
159 */
160 abstract public function newMainLB( $wiki = false );
161
162 /**
163 * Get a cached (tracked) load balancer object.
164 *
165 * @param bool|string $wiki Wiki ID, or false for the current wiki
166 * @return LoadBalancer
167 */
168 abstract public function getMainLB( $wiki = false );
169
170 /**
171 * Create a new load balancer for external storage. The resulting object will be
172 * untracked, not chronology-protected, and the caller is responsible for
173 * cleaning it up.
174 *
175 * @param string $cluster External storage cluster, or false for core
176 * @param bool|string $wiki Wiki ID, or false for the current wiki
177 * @return LoadBalancer
178 */
179 abstract protected function newExternalLB( $cluster, $wiki = false );
180
181 /**
182 * Get a cached (tracked) load balancer for external storage
183 *
184 * @param string $cluster External storage cluster, or false for core
185 * @param bool|string $wiki Wiki ID, or false for the current wiki
186 * @return LoadBalancer
187 */
188 abstract public function getExternalLB( $cluster, $wiki = false );
189
190 /**
191 * Execute a function for each tracked load balancer
192 * The callback is called with the load balancer as the first parameter,
193 * and $params passed as the subsequent parameters.
194 *
195 * @param callable $callback
196 * @param array $params
197 */
198 abstract public function forEachLB( $callback, array $params = [] );
199
200 /**
201 * Prepare all tracked load balancers for shutdown
202 * @param integer $flags Supports SHUTDOWN_* flags
203 */
204 public function shutdown( $flags = 0 ) {
205 if ( !( $flags & self::SHUTDOWN_NO_CHRONPROT ) ) {
206 $this->shutdownChronologyProtector( $this->chronProt );
207 }
208 $this->commitMasterChanges( __METHOD__ ); // sanity
209 }
210
211 /**
212 * Call a method of each tracked load balancer
213 *
214 * @param string $methodName
215 * @param array $args
216 */
217 private function forEachLBCallMethod( $methodName, array $args = [] ) {
218 $this->forEachLB(
219 function ( LoadBalancer $loadBalancer, $methodName, array $args ) {
220 call_user_func_array( [ $loadBalancer, $methodName ], $args );
221 },
222 [ $methodName, $args ]
223 );
224 }
225
226 /**
227 * Commit all replica DB transactions so as to flush any REPEATABLE-READ or SSI snapshot
228 *
229 * @param string $fname Caller name
230 * @since 1.28
231 */
232 public function flushReplicaSnapshots( $fname = __METHOD__ ) {
233 $this->forEachLBCallMethod( 'flushReplicaSnapshots', [ $fname ] );
234 }
235
236 /**
237 * Commit on all connections. Done for two reasons:
238 * 1. To commit changes to the masters.
239 * 2. To release the snapshot on all connections, master and replica DB.
240 * @param string $fname Caller name
241 * @param array $options Options map:
242 * - maxWriteDuration: abort if more than this much time was spent in write queries
243 */
244 public function commitAll( $fname = __METHOD__, array $options = [] ) {
245 $this->commitMasterChanges( $fname, $options );
246 $this->forEachLBCallMethod( 'commitAll', [ $fname ] );
247 }
248
249 /**
250 * Flush any master transaction snapshots and set DBO_TRX (if DBO_DEFAULT is set)
251 *
252 * The DBO_TRX setting will be reverted to the default in each of these methods:
253 * - commitMasterChanges()
254 * - rollbackMasterChanges()
255 * - commitAll()
256 *
257 * This allows for custom transaction rounds from any outer transaction scope.
258 *
259 * @param string $fname
260 * @throws DBTransactionError
261 * @since 1.28
262 */
263 public function beginMasterChanges( $fname = __METHOD__ ) {
264 if ( $this->trxRoundId !== false ) {
265 throw new DBTransactionError(
266 null,
267 "$fname: transaction round '{$this->trxRoundId}' already started."
268 );
269 }
270 $this->trxRoundId = $fname;
271 // Set DBO_TRX flags on all appropriate DBs
272 $this->forEachLBCallMethod( 'beginMasterChanges', [ $fname ] );
273 }
274
275 /**
276 * Commit changes on all master connections
277 * @param string $fname Caller name
278 * @param array $options Options map:
279 * - maxWriteDuration: abort if more than this much time was spent in write queries
280 * @throws Exception
281 */
282 public function commitMasterChanges( $fname = __METHOD__, array $options = [] ) {
283 if ( $this->trxRoundId !== false && $this->trxRoundId !== $fname ) {
284 throw new DBTransactionError(
285 null,
286 "$fname: transaction round '{$this->trxRoundId}' still running."
287 );
288 }
289 // Run pre-commit callbacks and suppress post-commit callbacks, aborting on failure
290 $this->forEachLBCallMethod( 'finalizeMasterChanges' );
291 $this->trxRoundId = false;
292 // Perform pre-commit checks, aborting on failure
293 $this->forEachLBCallMethod( 'approveMasterChanges', [ $options ] );
294 // Log the DBs and methods involved in multi-DB transactions
295 $this->logIfMultiDbTransaction();
296 // Actually perform the commit on all master DB connections and revert DBO_TRX
297 $this->forEachLBCallMethod( 'commitMasterChanges', [ $fname ] );
298 // Run all post-commit callbacks
299 /** @var Exception $e */
300 $e = null; // first callback exception
301 $this->forEachLB( function ( LoadBalancer $lb ) use ( &$e ) {
302 $ex = $lb->runMasterPostTrxCallbacks( IDatabase::TRIGGER_COMMIT );
303 $e = $e ?: $ex;
304 } );
305 // Commit any dangling DBO_TRX transactions from callbacks on one DB to another DB
306 $this->forEachLBCallMethod( 'commitMasterChanges', [ $fname ] );
307 // Throw any last post-commit callback error
308 if ( $e instanceof Exception ) {
309 throw $e;
310 }
311 }
312
313 /**
314 * Rollback changes on all master connections
315 * @param string $fname Caller name
316 * @since 1.23
317 */
318 public function rollbackMasterChanges( $fname = __METHOD__ ) {
319 $this->trxRoundId = false;
320 $this->forEachLBCallMethod( 'suppressTransactionEndCallbacks' );
321 $this->forEachLBCallMethod( 'rollbackMasterChanges', [ $fname ] );
322 // Run all post-rollback callbacks
323 $this->forEachLB( function ( LoadBalancer $lb ) {
324 $lb->runMasterPostTrxCallbacks( IDatabase::TRIGGER_ROLLBACK );
325 } );
326 }
327
328 /**
329 * Log query info if multi DB transactions are going to be committed now
330 */
331 private function logIfMultiDbTransaction() {
332 $callersByDB = [];
333 $this->forEachLB( function ( LoadBalancer $lb ) use ( &$callersByDB ) {
334 $masterName = $lb->getServerName( $lb->getWriterIndex() );
335 $callers = $lb->pendingMasterChangeCallers();
336 if ( $callers ) {
337 $callersByDB[$masterName] = $callers;
338 }
339 } );
340
341 if ( count( $callersByDB ) >= 2 ) {
342 $dbs = implode( ', ', array_keys( $callersByDB ) );
343 $msg = "Multi-DB transaction [{$dbs}]:\n";
344 foreach ( $callersByDB as $db => $callers ) {
345 $msg .= "$db: " . implode( '; ', $callers ) . "\n";
346 }
347 $this->trxLogger->info( $msg );
348 }
349 }
350
351 /**
352 * Determine if any master connection has pending changes
353 * @return bool
354 * @since 1.23
355 */
356 public function hasMasterChanges() {
357 $ret = false;
358 $this->forEachLB( function ( LoadBalancer $lb ) use ( &$ret ) {
359 $ret = $ret || $lb->hasMasterChanges();
360 } );
361
362 return $ret;
363 }
364
365 /**
366 * Detemine if any lagged replica DB connection was used
367 * @return bool
368 * @since 1.28
369 */
370 public function laggedReplicaUsed() {
371 $ret = false;
372 $this->forEachLB( function ( LoadBalancer $lb ) use ( &$ret ) {
373 $ret = $ret || $lb->laggedReplicaUsed();
374 } );
375
376 return $ret;
377 }
378
379 /**
380 * @return bool
381 * @since 1.27
382 * @deprecated Since 1.28; use laggedReplicaUsed()
383 */
384 public function laggedSlaveUsed() {
385 return $this->laggedReplicaUsed();
386 }
387
388 /**
389 * Determine if any master connection has pending/written changes from this request
390 * @return bool
391 * @since 1.27
392 */
393 public function hasOrMadeRecentMasterChanges() {
394 $ret = false;
395 $this->forEachLB( function ( LoadBalancer $lb ) use ( &$ret ) {
396 $ret = $ret || $lb->hasOrMadeRecentMasterChanges();
397 } );
398 return $ret;
399 }
400
401 /**
402 * Waits for the replica DBs to catch up to the current master position
403 *
404 * Use this when updating very large numbers of rows, as in maintenance scripts,
405 * to avoid causing too much lag. Of course, this is a no-op if there are no replica DBs.
406 *
407 * By default this waits on all DB clusters actually used in this request.
408 * This makes sense when lag being waiting on is caused by the code that does this check.
409 * In that case, setting "ifWritesSince" can avoid the overhead of waiting for clusters
410 * that were not changed since the last wait check. To forcefully wait on a specific cluster
411 * for a given wiki, use the 'wiki' parameter. To forcefully wait on an "external" cluster,
412 * use the "cluster" parameter.
413 *
414 * Never call this function after a large DB write that is *still* in a transaction.
415 * It only makes sense to call this after the possible lag inducing changes were committed.
416 *
417 * @param array $opts Optional fields that include:
418 * - wiki : wait on the load balancer DBs that handles the given wiki
419 * - cluster : wait on the given external load balancer DBs
420 * - timeout : Max wait time. Default: ~60 seconds
421 * - ifWritesSince: Only wait if writes were done since this UNIX timestamp
422 * @throws DBReplicationWaitError If a timeout or error occured waiting on a DB cluster
423 * @since 1.27
424 */
425 public function waitForReplication( array $opts = [] ) {
426 $opts += [
427 'wiki' => false,
428 'cluster' => false,
429 'timeout' => 60,
430 'ifWritesSince' => null
431 ];
432
433 foreach ( $this->replicationWaitCallbacks as $callback ) {
434 $callback();
435 }
436
437 // Figure out which clusters need to be checked
438 /** @var LoadBalancer[] $lbs */
439 $lbs = [];
440 if ( $opts['cluster'] !== false ) {
441 $lbs[] = $this->getExternalLB( $opts['cluster'] );
442 } elseif ( $opts['wiki'] !== false ) {
443 $lbs[] = $this->getMainLB( $opts['wiki'] );
444 } else {
445 $this->forEachLB( function ( LoadBalancer $lb ) use ( &$lbs ) {
446 $lbs[] = $lb;
447 } );
448 if ( !$lbs ) {
449 return; // nothing actually used
450 }
451 }
452
453 // Get all the master positions of applicable DBs right now.
454 // This can be faster since waiting on one cluster reduces the
455 // time needed to wait on the next clusters.
456 $masterPositions = array_fill( 0, count( $lbs ), false );
457 foreach ( $lbs as $i => $lb ) {
458 if ( $lb->getServerCount() <= 1 ) {
459 // Bug 27975 - Don't try to wait for replica DBs if there are none
460 // Prevents permission error when getting master position
461 continue;
462 } elseif ( $opts['ifWritesSince']
463 && $lb->lastMasterChangeTimestamp() < $opts['ifWritesSince']
464 ) {
465 continue; // no writes since the last wait
466 }
467 $masterPositions[$i] = $lb->getMasterPos();
468 }
469
470 $failed = [];
471 foreach ( $lbs as $i => $lb ) {
472 if ( $masterPositions[$i] ) {
473 // The DBMS may not support getMasterPos() or the whole
474 // load balancer might be fake (e.g. $wgAllDBsAreLocalhost).
475 if ( !$lb->waitForAll( $masterPositions[$i], $opts['timeout'] ) ) {
476 $failed[] = $lb->getServerName( $lb->getWriterIndex() );
477 }
478 }
479 }
480
481 if ( $failed ) {
482 throw new DBReplicationWaitError(
483 "Could not wait for replica DBs to catch up to " .
484 implode( ', ', $failed )
485 );
486 }
487 }
488
489 /**
490 * Add a callback to be run in every call to waitForReplication() before waiting
491 *
492 * Callbacks must clear any transactions that they start
493 *
494 * @param string $name Callback name
495 * @param callable|null $callback Use null to unset a callback
496 * @since 1.28
497 */
498 public function setWaitForReplicationListener( $name, callable $callback = null ) {
499 if ( $callback ) {
500 $this->replicationWaitCallbacks[$name] = $callback;
501 } else {
502 unset( $this->replicationWaitCallbacks[$name] );
503 }
504 }
505
506 /**
507 * Get a token asserting that no transaction writes are active
508 *
509 * @param string $fname Caller name (e.g. __METHOD__)
510 * @return mixed A value to pass to commitAndWaitForReplication()
511 * @since 1.28
512 */
513 public function getEmptyTransactionTicket( $fname ) {
514 if ( $this->hasMasterChanges() ) {
515 $this->trxLogger->error( __METHOD__ . ": $fname does not have outer scope." );
516 return null;
517 }
518
519 return $this->ticket;
520 }
521
522 /**
523 * Convenience method for safely running commitMasterChanges()/waitForReplication()
524 *
525 * This will commit and wait unless $ticket indicates it is unsafe to do so
526 *
527 * @param string $fname Caller name (e.g. __METHOD__)
528 * @param mixed $ticket Result of getEmptyTransactionTicket()
529 * @param array $opts Options to waitForReplication()
530 * @throws DBReplicationWaitError
531 * @since 1.28
532 */
533 public function commitAndWaitForReplication( $fname, $ticket, array $opts = [] ) {
534 if ( $ticket !== $this->ticket ) {
535 $logger = LoggerFactory::getInstance( 'DBPerformance' );
536 $logger->error( __METHOD__ . ": cannot commit; $fname does not have outer scope." );
537 return;
538 }
539
540 // The transaction owner and any caller with the empty transaction ticket can commit
541 // so that getEmptyTransactionTicket() callers don't risk seeing DBTransactionError.
542 if ( $this->trxRoundId !== false && $fname !== $this->trxRoundId ) {
543 $this->trxLogger->info( "$fname: committing on behalf of {$this->trxRoundId}." );
544 $fnameEffective = $this->trxRoundId;
545 } else {
546 $fnameEffective = $fname;
547 }
548
549 $this->commitMasterChanges( $fnameEffective );
550 $this->waitForReplication( $opts );
551 // If a nested caller committed on behalf of $fname, start another empty $fname
552 // transaction, leaving the caller with the same empty transaction state as before.
553 if ( $fnameEffective !== $fname ) {
554 $this->beginMasterChanges( $fnameEffective );
555 }
556 }
557
558 /**
559 * Disable the ChronologyProtector for all load balancers
560 *
561 * This can be called at the start of special API entry points
562 *
563 * @since 1.27
564 */
565 public function disableChronologyProtection() {
566 $this->chronProt->setEnabled( false );
567 }
568
569 /**
570 * @return ChronologyProtector
571 */
572 protected function newChronologyProtector() {
573 $request = RequestContext::getMain()->getRequest();
574 $chronProt = new ChronologyProtector(
575 ObjectCache::getMainStashInstance(),
576 [
577 'ip' => $request->getIP(),
578 'agent' => $request->getHeader( 'User-Agent' )
579 ]
580 );
581 if ( PHP_SAPI === 'cli' ) {
582 $chronProt->setEnabled( false );
583 } elseif ( $request->getHeader( 'ChronologyProtection' ) === 'false' ) {
584 // Request opted out of using position wait logic. This is useful for requests
585 // done by the job queue or background ETL that do not have a meaningful session.
586 $chronProt->setWaitEnabled( false );
587 }
588
589 return $chronProt;
590 }
591
592 /**
593 * @param ChronologyProtector $cp
594 */
595 protected function shutdownChronologyProtector( ChronologyProtector $cp ) {
596 // Get all the master positions needed
597 $this->forEachLB( function ( LoadBalancer $lb ) use ( $cp ) {
598 $cp->shutdownLB( $lb );
599 } );
600 // Write them to the stash
601 $unsavedPositions = $cp->shutdown();
602 // If the positions failed to write to the stash, at least wait on local datacenter
603 // replica DBs to catch up before responding. Even if there are several DCs, this increases
604 // the chance that the user will see their own changes immediately afterwards. As long
605 // as the sticky DC cookie applies (same domain), this is not even an issue.
606 $this->forEachLB( function ( LoadBalancer $lb ) use ( $unsavedPositions ) {
607 $masterName = $lb->getServerName( $lb->getWriterIndex() );
608 if ( isset( $unsavedPositions[$masterName] ) ) {
609 $lb->waitForAll( $unsavedPositions[$masterName] );
610 }
611 } );
612 }
613
614 /**
615 * @param LoadBalancer $lb
616 */
617 protected function initLoadBalancer( LoadBalancer $lb ) {
618 if ( $this->trxRoundId !== false ) {
619 $lb->beginMasterChanges( $this->trxRoundId ); // set DBO_TRX
620 }
621 }
622
623 /**
624 * Close all open database connections on all open load balancers.
625 * @since 1.28
626 */
627 public function closeAll() {
628 $this->forEachLBCallMethod( 'closeAll', [] );
629 }
630
631 }