All Downloads are FREE. Search and download functionalities are using the official Maven repository.

org.apache.jena.tdb.transaction.TransactionManager Maven / Gradle / Ivy

Go to download

TDB is a storage subsystem for Jena and ARQ, it is a native triple store providing persistent storage of triples/quads.

There is a newer version: 4.10.0
Show newest version
/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

package org.apache.jena.tdb.transaction;

import static java.lang.String.format ;
import static org.apache.jena.tdb.sys.SystemTDB.syslog ;
import static org.apache.jena.tdb.transaction.TransactionManager.TxnPoint.BEGIN ;
import static org.apache.jena.tdb.transaction.TransactionManager.TxnPoint.CLOSE ;

import java.io.File ;
import java.util.ArrayList;
import java.util.List;
import java.util.Objects;
import java.util.Set;
import java.util.concurrent.BlockingQueue ;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.LinkedBlockingDeque ;
import java.util.concurrent.Semaphore ;
import java.util.concurrent.atomic.AtomicLong ;
import java.util.concurrent.atomic.AtomicReference ;
import java.util.concurrent.locks.ReadWriteLock ;
import java.util.concurrent.locks.ReentrantReadWriteLock ;

import org.apache.jena.atlas.lib.Pair ;
import org.apache.jena.atlas.logging.FmtLog;
import org.apache.jena.atlas.logging.Log ;
import org.apache.jena.query.ReadWrite ;
import org.apache.jena.query.TxnType;
import org.apache.jena.shared.Lock ;
import org.apache.jena.sparql.core.Transactional.Promote;
import org.apache.jena.tdb.store.DatasetGraphTDB ;
import org.apache.jena.tdb.sys.SystemTDB ;
import org.slf4j.Logger ;
import org.slf4j.LoggerFactory ;

public class TransactionManager
{
    private static boolean checking = true ;
    
    private static Logger log = LoggerFactory.getLogger(TransactionManager.class) ;
    private Set currentActiveTransactions = ConcurrentHashMap.newKeySet();
    
    // Setting this true cause the TransactionManager to keep lists of transactions
    // and what has happened.  Nothing is thrown away, but eventually it will
    // consume too much memory.
    
    // Record happenings.
    private boolean recordHistory = false ;
    
    /** This controls how many write transactions we batch up before 
     *  deciding to try to flush the journal to the main database.  
     */
    public static /*final*/ int QueueBatchSize = setQueueBatchSize() ; 
    
    // JENA-1222
    /** This controls how large to let the journal get in size 
     * before deciding to flush even if below the queue batch size.
     * -1 means "off".   
     */
    public static /*final*/ int JournalThresholdSize = -1 ;  
    
    // JENA-1224
    /** This controls how large to let the journal in comitted transactions
     * before deciding to pause and flush as soon as possible.
     * -1 means "off".   
     */
    public static /*final*/ int MaxQueueThreshold = 100 ;
    
    private static int setQueueBatchSize() {
        if ( SystemTDB.is64bitSystem )
            return 10 ;
        // On 32bit systems are memory constrained. The Java address space is
        // limited to about 1.5G - the heap can not be bigger.
        // So we don't do batching (change if batching is less memory hungry).
        return 0 ;
    }
    
    // Records the states that a transaction goes though.
    enum TxnPoint { BEGIN, COMMIT, ABORT, CLOSE, QUEUE, UNQUEUE }
    private List> transactionStateTransition ;
    
    private void record(Transaction txn, TxnPoint state) {
        if ( !recordHistory )
            return ;
        initRecordingState() ;
        transactionStateTransition.add(new Pair<>(txn, state)) ;
    }
    
    // Statistic variables to record the maximum length of the flush queue.
    
    int maxQueue = 0 ;
    List commitedAwaitingFlush = new ArrayList<>() ;
    
    static AtomicLong transactionId = new AtomicLong(1) ;
    
    private List additionalComponents = new ArrayList<>();  
    
    // The version of the data starting at 1. The contract is that this is
    // never the same for two version (it may not be monotonic increasing).
    // Currently, it is sequentially increasing.
    // Set on each commit.
    private AtomicLong version = new AtomicLong(0) ;
    
    // This counter is 0 when quiet, else >0 
    private AtomicLong activeTransactions = new AtomicLong(0) ;
    // Not based on activeTransactionsSet.isEmpty() which does a count
    public boolean activeTransactions()     { return activeTransactions.get() != 0; }

    // Accessed by SysTxnState
    // Because there are two counters, these can be slightly out of step.
    // The test (activeReaders==0 & activeWriters==0) isn't atomic overall.
    /*package*/ AtomicLong activeReaders = new AtomicLong(0) ; 
    /*package*/ AtomicLong activeWriters = new AtomicLong(0) ; // 0 or 1
    public long getCountActiveReaders()     { return activeReaders.get() ; }
    public long getCountActiveWriters()     { return activeWriters.get() ; }
    
    // Misc stats (should be LongAdder / Java8?)
    /*package*/ AtomicLong finishedReaders = new AtomicLong(0) ;
    /*package*/ AtomicLong committedWriters = new AtomicLong(0) ;
    /*package*/ AtomicLong abortedWriters = new AtomicLong(0) ;
    
    // This is the DatasetGraphTDB for the first read-transaction created for
    // a particular view. This read DatasetGraphTDB can be used by all the readers
    // seeing the same view. A write transaction clears this when it commits; the first
    // reader of the next state creates a new view DatasetGraph.
    private AtomicReference currentReaderView = new AtomicReference<>(null) ;
    
    // Ensure single writer. A writer calling begin(WRITE) blocks.  
    private Semaphore writerPermits = new Semaphore(1, true) ;
    
    // All transactions need a "read" lock throughout their lifetime. 
    // Do not confuse with read/write transactions.  We need a 
    // "one exclusive, or many other" lock which happens to be called a ReadWriteLock.
    // Fair lock - approximately arrival order. 
    // Stops "readers" (normal transactions, READ or WRITE) from locking
    // out a "writer" (exclusive mode).  
    private ReadWriteLock exclusivitylock = new ReentrantReadWriteLock(true) ;
    
    // Delays enacting transactions.
    private BlockingQueue queue = new LinkedBlockingDeque<>() ;
    public long getQueueLength() { return queue.size() ; }

    private DatasetGraphTDB baseDataset ;
    private Journal journal ;
    
    /*
     * The order of calls is: 
     * 1/ transactionStarts
     * 2/ readerStarts or writerStarts
     * 3/ readerFinishes or writerCommits or writerAborts
     * 4/ transactionFinishes
     * 5/ transactionCloses
     */
    
    private interface TSM {
        void transactionStarts(Transaction txn) ;
        void transactionFinishes(Transaction txn) ;
        void transactionCloses(Transaction txn) ;
        void readerStarts(Transaction txn) ;
        void readerFinishes(Transaction txn) ;
        void transactionPromotes(Transaction txnOld, Transaction txnNew) ;
        void writerStarts(Transaction txn) ;
        void writerCommits(Transaction txn) ;
        void writerAborts(Transaction txn) ;
    }
    
    class TSM_Base implements TSM {
        @Override public void transactionStarts(Transaction txn)    {}
        @Override public void transactionFinishes(Transaction txn)  {}
        @Override public void transactionCloses(Transaction txn)    {}
        @Override public void readerStarts(Transaction txn)         {}
        @Override public void readerFinishes(Transaction txn)       {}
        @Override public void transactionPromotes(Transaction txnOld, Transaction txnNew) {}
        @Override public void writerStarts(Transaction txn)         {}
        @Override public void writerCommits(Transaction txn)        {}
        @Override public void writerAborts(Transaction txn)         {}
    }
    
    class TSM_Logger extends TSM_Base {
        TSM_Logger() {}
        @Override public void readerStarts(Transaction txn)         { log("start", txn) ; }
        @Override public void readerFinishes(Transaction txn)       { log("finish", txn) ; }
        @Override public void transactionPromotes(Transaction txnOld, Transaction txnNew)
        { log("promote(old)", txnOld) ; log("promote(new)", txnNew) ; }  
        @Override public void writerStarts(Transaction txn)         { log("begin", txn) ; }
        @Override public void writerCommits(Transaction txn)        { log("commit", txn) ; }
        @Override public void writerAborts(Transaction txn)         { log("abort", txn) ; }
    }

    /** More detailed */
    class TSM_LoggerDebug extends TSM_Base {
        TSM_LoggerDebug() {}
        @Override public void readerStarts(Transaction txn)         { logInternal("start",  txn) ; }
        @Override public void readerFinishes(Transaction txn)       { logInternal("finish", txn) ; }
        @Override public void transactionPromotes(Transaction txnOld, Transaction txnNew)
        { logInternal("promote(old)", txnOld) ; logInternal("promote(new)", txnNew) ; }
        @Override public void writerStarts(Transaction txn)         { logInternal("begin",  txn) ; }
        @Override public void writerCommits(Transaction txn)        { logInternal("commit", txn) ; }
        @Override public void writerAborts(Transaction txn)         { logInternal("abort",  txn) ; }
    }
    
    class TSM_Counters implements TSM {
        // activeTransactions must be in read/writer ending operation, not transactionFinishes
        // because the check in checkReplaySafe happens after readerFinishes/writerCommits/writerAborts
        // but before transactionFinishes.
        TSM_Counters() {}
        @Override public void transactionStarts(Transaction txn)    { currentActiveTransactions.add(txn) ; }
        @Override public void transactionFinishes(Transaction txn)  { currentActiveTransactions.remove(txn) ; }
        @Override public void transactionCloses(Transaction txn)    { }
        @Override public void readerStarts(Transaction txn)         { inc(activeTransactions); inc(activeReaders) ; }
        @Override public void readerFinishes(Transaction txn)       { dec(activeReaders) ; inc(finishedReaders); dec(activeTransactions); }
        
        @Override public void transactionPromotes(Transaction txnOld, Transaction txnNew) {
            // Add first so the set is not momentarily empty.
            currentActiveTransactions.add(txnNew); currentActiveTransactions.remove(txnOld) ;
            dec(activeReaders) ; inc(finishedReaders); inc(activeWriters);
        }
        
        @Override public void writerStarts(Transaction txn)         { inc(activeTransactions); inc(activeWriters) ; }
        @Override public void writerCommits(Transaction txn)        { dec(activeWriters) ; inc(committedWriters) ; dec(activeTransactions); }
        @Override public void writerAborts(Transaction txn)         { dec(activeWriters) ; inc(abortedWriters) ; dec(activeTransactions); }
    }
    
    // Short name: x++
    static long inc(AtomicLong x)   { return x.getAndIncrement() ; }
    // Short name: --x
    static long dec(AtomicLong x)   { return x.decrementAndGet() ; }
    
    // Transaction policy:
    // TSM + WriterEnters, WriterLeaves which may use the semaphore. (+ReaderEnters, ReaderLeaves ??)
    
    // Policy for writing back journal'ed data to the base datasetgraph
    // Writes if no reader at end of writer, else queues.
    // Queue cleared at end of any transaction finding itself the only transaction.
    class TSM_WriteBackEndTxn extends TSM_Base {
        // Safe mode.
        // Take a READ lock over the base dataset.
        // Write-back takes a WRITE lock.
        @Override public void readerStarts(Transaction txn)     { txn.getBaseDataset().getLock().enterCriticalSection(Lock.READ) ; }
        
        @Override public void writerStarts(Transaction txn)     { txn.getBaseDataset().getLock().enterCriticalSection(Lock.READ) ; }
        
        // Currently, the writer semaphore is managed explicitly in the main code.
        
        @Override public void transactionPromotes(Transaction txnOld, Transaction txnNew) {
            // Switch locks.
            txnOld.getBaseDataset().getLock().leaveCriticalSection() ;
            txnNew.getBaseDataset().getLock().enterCriticalSection(Lock.READ) ;
        }
        
        @Override
        public void readerFinishes(Transaction txn) {
            txn.getBaseDataset().getLock().leaveCriticalSection() ;
            readerFinishesWorker(txn) ;
        }

        @Override
        public void writerCommits(Transaction txn) {
            txn.getBaseDataset().getLock().leaveCriticalSection() ;
            writerCommitsWorker(txn) ;
        }

        @Override
        public void writerAborts(Transaction txn) {
            txn.getBaseDataset().getLock().leaveCriticalSection() ;
            writerAbortsWorker(txn) ;
        }
    }
    
    // Policy for writing back that always writes from the writer by using a
    // MRSW lock, with a write step at the end of the writer.
    // Always a read loc for any active transaction (reader or the writer)
    // Still use semaphore for writer entry control.
    class TSM_WriterWriteBack extends TSM_Base
    {
        // Not implemented
    }
    
    // Policy for writing where a transaction takes an MRSW at the start.
    // Semaphore for writer entry unnecessary.
    class TSM_MRSW_Writer extends TSM_Base
    {
        // Not implemented
    }
    

    class TSM_Record extends TSM_Base {
        // Later - record on one list the state transition.
        @Override
        public void transactionStarts(Transaction txn)      { record(txn, BEGIN) ; }
        @Override
        public void transactionFinishes(Transaction txn)    { record(txn, CLOSE) ; }
    }
    
    private TSM[] actions = new TSM[] { 
        new TSM_Counters() ,           // Must be first.
        //new TSM_LoggerDebug() ,
        new TSM_Logger() ,
        (recordHistory ? new TSM_Record() : null ) ,
        new TSM_WriteBackEndTxn()        // Write back policy. Must be last.
    } ;

    public TransactionManager(DatasetGraphTDB dsg){
        this.baseDataset = dsg ; 
        this.journal = Journal.create(dsg.getLocation()) ;
        // LATER
//        Committer c = new Committer() ;
//        this.committerThread = new Thread(c) ;
//        committerThread.setDaemon(true) ;
//        committerThread.start() ;
    }

    synchronized
    public void closedown() {
        processDelayedReplayQueue(null) ;
        journal.close() ;
    }
    
    /** Add a {@link TransactionLifecycle} to transactions createed from this point onwards. */  
    public void addAdditionComponent(TransactionLifecycle txnLifecycle) {
        additionalComponents.add(txnLifecycle);
    }

    public /*for testing only*/ static final boolean DEBUG = false ;
    
    public DatasetGraphTxn begin(TxnType txnType, String label) {
        return beginInternal(txnType, txnType, label);
    }
    
    private DatasetGraphTxn beginInternal(TxnType txnType, TxnType originalTxnType, String label) {   
        // The exclusivitylock surrounds the entire transaction cycle.
        // Paired with notifyCommit, notifyAbort.
        startNonExclusive();
        
        // Not synchronized (else blocking on semaphore will never wake up
        // because Semaphore.release is inside synchronized).
        // Allow only one active writer. 
        if ( txnType == TxnType.WRITE ) {
            // Writers take a WRITE permit from the semaphore to ensure there
            // is at most one active writer, else the attempt to start the
            // transaction blocks.
            acquireWriterLock(true) ;
        }
        // entry synchronized part
        DatasetGraphTxn dsgtxn = begin$(txnType, originalTxnType, label) ;
        noteTxnStart(dsgtxn.getTransaction()) ;
        return dsgtxn;
    }
    
    /** Ensure a DatasetGraphTxn is for a write transaction.
     * 

* If the transaction is already a write transaction, this is an efficient no-op. *

* If the transaction is a read transaction then promotion can either respect the transactions current * view of the data where no changes from other writers that started after this transaction are visible * ("serialized" or "fully isolated") or the promotion can include changes by other such writers ("read committed"). *

* However, "serialized" can fail, in which case an exception {@link TDBTransactionException} * is thrown. The transactions can continue as a read transaction. * There is no point retrying - later committed changes have been made and will remain. *

* "read committed" will always succeed but the app needs to be aware that data access before the promotion * is no longer valid. It may need to check it. *

* Return null for "no promote" due to intermediate commits. */ /*package*/ DatasetGraphTxn promote(DatasetGraphTxn dsgtxn, TxnType originalTxnType, Promote promoteType) throws TDBTransactionException { Transaction txn = dsgtxn.getTransaction() ; if ( txn.getState() != TxnState.ACTIVE ) throw new TDBTransactionException("promote: transaction is not active") ; if ( txn.getTxnMode() == ReadWrite.WRITE ) return dsgtxn ; if ( txn.getTxnType() == TxnType.READ ) return null; // Did no promote. //txn.abort(); //throw new TDBTransactionException("promote: transaction is a READ transaction") ; // Read commit - pick up whatever is current at the point setup. // Can also promote - may need to wait for active writers. // Go through begin for the writers lock. if ( promoteType == Promote.READ_COMMITTED ) { acquireWriterLock(true); // No need to sync - we just queue as a writer. return promoteExec$(dsgtxn, originalTxnType); } // First check, without the writer lock. Fast fail. // Have any finished writers run and commited since this transaction began? // This is a "fast fail" test of whether we can promote. Passing it does not gauarantee // we can promote but failing it does mean we can't. // There are active writers running at this moment, (or any time up to // acquireWriterLock returning. It catches many cases without needing // to acquire the writer lock. if ( txn.getVersion() != version.get() ) return null; // Put ourselves in the serialization timeline of the dataset - that is, grab a writer // lock as a step toward promotion. We can then test properly because no other writer // can start; readers can between acquireWriterLock and the synchronized in promoteSync$ // but they don't matter. // Potentially blocking - must be outside 'synchronized' so that any active writer // can commit/abort. Otherwise, we have deadlock. acquireWriterLock(true) ; // Do the synchronized stuff. return promoteSync$(dsgtxn, originalTxnType) ; } synchronized private DatasetGraphTxn promoteSync$(DatasetGraphTxn dsgtxn, TxnType originalTxnType) { Transaction txn = dsgtxn.getTransaction() ; // Writers may have happened between the first check of the active writers may have committed. if ( txn.getVersion() != version.get() ) { releaseWriterLock(); return null; } return promoteExec$(dsgtxn, originalTxnType); } private DatasetGraphTxn promoteExec$(DatasetGraphTxn dsgtxn, TxnType originalTxnType) { // Use begin$ (not beginInternal) // We have the writers lock. // We keep the exclusivity lock. Transaction txn = dsgtxn.getTransaction() ; DatasetGraphTxn dsgtxn2 = begin$(TxnType.WRITE, originalTxnType, txn.getLabel()) ; noteTxnPromote(txn, dsgtxn2.getTransaction()); return dsgtxn2 ; } // If DatasetGraphTransaction has a sync lock on sConn, this // does not need to be sync'ed. But it's possible to use some // of the low level objects directly so we'll play safe. synchronized private DatasetGraphTxn begin$(TxnType txnType, TxnType originalTxnType, String label) { Objects.requireNonNull(txnType); if ( txnType == TxnType.WRITE && activeWriters.get() > 0 ) // Guard throw new TDBTransactionException("Existing active write transaction") ; if ( DEBUG ) switch ( txnType ) { case READ : System.out.print("r") ; break ; case WRITE : System.out.print("w") ; break ; case READ_COMMITTED_PROMOTE : System.out.print("r(cp)") ; break ; case READ_PROMOTE : System.out.print("rp") ; break ; } DatasetGraphTDB dsg = determineBaseDataset() ; Transaction txn = createTransaction(dsg, txnType, originalTxnType, label) ; log("begin$", txn) ; ReadWrite mode = initialMode(txnType); DatasetGraphTxn dsgTxn = createDSGTxn(dsg, txn, mode); txn.setActiveDataset(dsgTxn) ; dsgTxn.getTransaction().forAllComponents(component->component.begin(dsgTxn.getTransaction())) ; return dsgTxn ; } private DatasetGraphTDB determineBaseDataset() { // if ( DEBUG ) { // if ( !commitedAwaitingFlush.isEmpty() ) // System.out.print(commitedAwaitingFlush.size()) ; // } else { // System.out.print('_') ; // } DatasetGraphTDB dsg = baseDataset ; // But, if there are pending, committed transactions, use latest. if ( !commitedAwaitingFlush.isEmpty() ) dsg = commitedAwaitingFlush.get(commitedAwaitingFlush.size() - 1).getActiveDataset().getView() ; return dsg ; } private Transaction createTransaction(DatasetGraphTDB dsg, TxnType txnType, TxnType originalTxnType, String label) { if ( originalTxnType == null ) originalTxnType = txnType; Transaction txn = new Transaction(dsg, version.get(), txnType, initialMode(txnType), transactionId.getAndIncrement(), originalTxnType, label, this) ; // Add the extras. additionalComponents.forEach(txn::addAdditionaComponent); return txn ; } // State. private static ReadWrite initialMode(TxnType txnType) { return TxnType.initial(txnType); } private DatasetGraphTxn createDSGTxn(DatasetGraphTDB dsg, Transaction txn, ReadWrite mode) { // A read transaction (if it has no lifecycle components) can be shared over all // read transactions at the same commit level. // lastreader if ( mode == ReadWrite.READ ) { // If a READ transaction, and a previously built one is cached, use it. DatasetGraphTDB dsgCached = currentReaderView.get() ; if ( dsgCached != null ) { // No components so we don't need to notify them. // We can just reuse the storage dataset. return new DatasetGraphTxn(dsgCached, txn) ; } } DatasetGraphTxn dsgTxn = new DatasetBuilderTxn(this, dsg).build(txn, mode); if ( mode == ReadWrite.READ ) { // If a READ transaction, cache the storage view. // This is cleared when a WRITE commits currentReaderView.set(dsgTxn.getView()) ; } return dsgTxn ; } /* Signal a transaction has committed. The journal has a commit record * and a sync to disk. The code here manages the inter-transaction stage * of deciding how to play the changes back to the base data * together with general recording of transaction details and status. */ /*package*/ void notifyCommit(Transaction transaction) { boolean excessiveQueue = false ; synchronized(this) { if ( ! currentActiveTransactions.contains(transaction) ) SystemTDB.errlog.warn("Transaction not active: "+transaction.getTxnId()) ; noteTxnCommit(transaction) ; switch ( transaction.getTxnMode() ) { case READ: break ; case WRITE: version.incrementAndGet() ; currentReaderView.set(null) ; // Clear the READ transaction cache. // JENA-1224 excessiveQueue = ( MaxQueueThreshold >= 0 && queue.size() > MaxQueueThreshold ) ; releaseWriterLock(); } // Paired with begin() finishNonExclusive(); } // Force the queue to flush if it is getting excessively long. // Note that writers may happen between releaseWriterLock above // and exclusiveFlushQueue and so several writers may try to flush // the queue; this is safe. if ( excessiveQueue ) { // Check again. This is an imperfect filter for multiple attempts to flush the queue // and avoid going into exclusive mode. // excessiveQueue = ( MaxQueueThreshold >= 0 && queue.size() > MaxQueueThreshold ) ; // if ( excessiveQueue ) // Must have released the exclusivity lock (a read-MRSW lock) otherwise // taking the exclusivity lock (a write-MRSW lock) in // exclusiveFlushQueue->startExclusiveMode will block. // Java MRSW locks do not allow "hold reader -> take writer". exclusiveFlushQueue(); } } synchronized /*package*/ void notifyAbort(Transaction transaction) { // Transaction has done the abort on all the transactional elements. if ( ! currentActiveTransactions.contains(transaction) ) SystemTDB.errlog.warn("Transaction not active: "+transaction.getTxnId()) ; noteTxnAbort(transaction) ; switch ( transaction.getTxnMode() ) { case READ: break ; case WRITE: releaseWriterLock(); } // Paired with begin() finishNonExclusive(); } synchronized /*package*/ void notifyClose(Transaction txn) { // Caution - not called if "Transactional.end() is not called." if ( txn.getState() == TxnState.ACTIVE ) { // The application error case for begin(WRITE)...end() is handled in Trasnaction.close(). // This is internal checking. String x = txn.getBaseDataset().getLocation().getDirectoryPath() ; syslog.warn("close: Transaction not commited or aborted: Transaction: "+txn.getTxnId()+" @ "+x) ; // Force abort then close txn.abort() ; txn.close() ; return ; } noteTxnClose(txn) ; } private void releaseWriterLock() { int x = writerPermits.availablePermits() ; if ( x != 0 ) throw new TDBTransactionException("TransactionCoordinator: Probably mismatch of enableWriters/blockWriters calls") ; writerPermits.release() ; } private boolean acquireWriterLock(boolean canBlock) { if ( ! canBlock ) return writerPermits.tryAcquire() ; try { writerPermits.acquire() ; return true; } catch (InterruptedException e) { throw new TDBTransactionException(e) ; } } /** Block until no writers are active. * When this returns, it guarantees that the database is not changing * and the journal is flush to disk. *

* The application must call {@link #enableWriters} later. *

* This operation must not be nested (it will block). * * @see #tryBlockWriters() * @see #enableWriters() * */ public void blockWriters() { acquireWriterLock(true) ; } /** Block until no writers are active or, optionally, return if can't at the moment. * Return 'true' if the operation succeeded. *

* If it returns true, the application must call {@link #enableWriters} later. * * @see #blockWriters() * @see #enableWriters() */ public boolean tryBlockWriters() { return acquireWriterLock(false) ; } /** Allow writers. * This must be used in conjunction with {@link #blockWriters()} or {@link #tryBlockWriters()} * * @see #blockWriters() * @see #tryBlockWriters() */ public void enableWriters() { releaseWriterLock(); } /** Force flushing the queue by entering exclusive mode.*/ private void exclusiveFlushQueue() { startExclusiveMode(true) ; finishExclusiveMode(); } /** Start non-exclusive mode - this is the normal transaction mode. * This does not relate to reader/writer. */ private void startNonExclusive() { exclusivitylock.readLock().lock(); } /** Finish non-exclusive mode - this is the normal transaction mode. * This does not relate to reader/writer. */ private void finishNonExclusive() { exclusivitylock.readLock().unlock(); } /** Enter exclusive mode. *

* There are no active transactions on return; new transactions will be held up in 'begin'. * Return to normal (release waiting transactions, allow new transactions) * with {@link #finishExclusiveMode}. *

* The caller must not be inside a transaction associated with this TransactionManager. * (The call will block waiting for that transaction to finish.) */ public void startExclusiveMode() { startExclusiveMode(true); } /** Try to enter exclusive mode. * If return is true, then are no active transactions on return and new transactions will be held up in 'begin'. * If false, there is an in-progress transactions. * Return to normal (release waiting transactions, allow new transactions) * with {@link #finishExclusiveMode}. *

* The call must not itself be in a transaction (this call will return false). */ public boolean tryExclusiveMode() { return startExclusiveMode(false); } private boolean startExclusiveMode(boolean canBlock) { if ( canBlock ) { exclusivitylock.writeLock().lock() ; processDelayedReplayQueue(null); return true ; } boolean b = exclusivitylock.writeLock().tryLock() ; if ( ! b ) return false ; processDelayedReplayQueue(null); return true ; } /** Return the exclusivity lock. Testing and internal use only. */ public ReadWriteLock getExclusivityLock$() { return exclusivitylock ; } /** Return to normal (release waiting transactions, allow new transactions). * Must be paired with an earlier {@link #startExclusiveMode}. */ public void finishExclusiveMode() { exclusivitylock.writeLock().unlock() ; } /** The stage in a commit after committing - make the changes permanent in the base data */ private void enactTransaction(Transaction transaction) { transaction.forAllComponents(x->x.enactCommitted(transaction)); transaction.forAllComponents(x->x.clearupCommitted(transaction)); transaction.signalEnacted() ; } /** Try to flush the delayed write queue - only happens if there are no active transactions */ synchronized public void flush() { processDelayedReplayQueue(null) ; } // -- The main operations to undertake when a transaction finishes. // Called from TSM_WriteBackEndTxn but the worker code is here so all // related code, including queue flushing is close together. private void readerFinishesWorker(Transaction txn) { if ( checkForJournalFlush() ) processDelayedReplayQueue(txn) ; } private void writerAbortsWorker(Transaction txn) { if ( checkForJournalFlush() ) processDelayedReplayQueue(txn) ; } // Whether to try to flush the journal. We may still find that we are blocked // from doing so by another transaction (checked in processDelayedReplayQueue). // MaxQueueThreshold is handled in notifyCommit (writer exit). private boolean checkForJournalFlush() { if ( queue.size() >= QueueBatchSize ) return true ; boolean journalSizeFlush = (JournalThresholdSize > 0 && journal.size() > JournalThresholdSize ) ; if ( journalSizeFlush ) // JENA-1222 // Based on Journal file growing large in terms of bytes return true ; // No test here for excessive queue length (MaxQueueThreshold). return false ; } private void writerCommitsWorker(Transaction txn) { if ( activeReaders.get() == 0 && checkForJournalFlush() ) { // Can commit immediately. // Ensure the queue is empty though. // Could simply add txn to the commit queue and do it that way. if ( log() ) log("Commit immediately", txn) ; // Currently, all we need is // JournalControl.replay(txn) ; // because that plays queued transactions. // But for long term generallity, at the cost of one check of the journal size // we do this sequence. processDelayedReplayQueue(txn) ; enactTransaction(txn) ; JournalControl.replay(txn) ; } else { // Can't write back to the base database at the moment. commitedAwaitingFlush.add(txn) ; maxQueue = Math.max(commitedAwaitingFlush.size(), maxQueue) ; if ( log() ) log("Add to pending queue", txn) ; queue.add(txn) ; } } private void processDelayedReplayQueue(Transaction txn) { // JENA-1224: Are there too many wrapper layers? // This is handled in notifyCommit. // Can we do work? if ( activeReaders.get() != 0 || activeWriters.get() != 0 ) { if ( queue.size() > 0 && log() ) log(format("Pending transactions: R=%s / W=%s", activeReaders, activeWriters), txn) ; return ; } if ( DEBUG ) { if ( queue.size() > 0 ) System.out.print("!"+queue.size()+"!") ; } if ( DEBUG ) checkNodesDatJrnl("1", txn) ; if ( queue.size() == 0 && txn != null ) // Nothing to do - journal should be empty. return ; if ( log() ) log("Start flush delayed commits", txn) ; // Drop the cached reader view so that next time it is recreated // against the updated database. currentReaderView.set(null) ; while (queue.size() > 0) { // Currently, replay is replay everything // so looping on a per-transaction basis is // pointless but harmless. try { Transaction txn2 = queue.take() ; if ( txn2.getTxnMode() == ReadWrite.READ ) continue ; if ( log() ) log(" Flush delayed commit of "+txn2.getLabel(), txn) ; if ( DEBUG ) checkNodesDatJrnl("2", txn) ; checkReplaySafe() ; enactTransaction(txn2) ; commitedAwaitingFlush.remove(txn2) ; } catch (InterruptedException ex) { Log.error(this, "Interruped!", ex) ; } } checkReplaySafe() ; if ( DEBUG ) checkNodesDatJrnl("3", txn) ; // Whole journal to base database JournalControl.replay(journal, baseDataset) ; if ( DEBUG ) checkNodesDatJrnl("4", txn) ; checkReplaySafe() ; if ( log() ) log("End flush delayed commits", txn) ; } private void checkNodesDatJrnl(String label, Transaction txn) { if ( txn != null ) { String x = txn.getBaseDataset().getLocation().getPath(label + ": nodes.dat-jrnl") ; long len = new File(x).length() ; if ( len != 0 ) log("nodes.dat-jrnl: not empty", txn) ; } } private void checkReplaySafe() { if ( ! checking ) return ; long x = activeTransactions.get(); if ( x == 0 ) return ; // Not a consistent view. long r = activeReaders.get(); long w = activeWriters.get(); FmtLog.error(log, "checkReplaySafe: There are currently active transactions. C=%d (R=%d, W=%d)", x, r, w) ; } private void noteTxnStart(Transaction transaction) { switch (transaction.getTxnMode()) { case READ : readerStarts(transaction) ; break ; case WRITE : writerStarts(transaction) ; break ; } transactionStarts(transaction) ; } private void noteTxnPromote(Transaction transaction, Transaction transaction2) { transactionPromotes(transaction, transaction2) ; } private void noteTxnCommit(Transaction transaction) { switch (transaction.getTxnMode()) { case READ : readerFinishes(transaction) ; break ; case WRITE : writerCommits(transaction) ; break ; } transactionFinishes(transaction) ; } private void noteTxnAbort(Transaction transaction) { switch (transaction.getTxnMode()) { case READ : readerFinishes(transaction) ; break ; case WRITE : writerAborts(transaction) ; break ; } transactionFinishes(transaction) ; } private void noteTxnClose(Transaction transaction) { transactionCloses(transaction) ; } // ---- Recording /** Get recording state */ public boolean recording() { return recordHistory ; } /** Set recording on or off */ public void recording(boolean flag) { recordHistory = flag ; if ( recordHistory ) initRecordingState() ; } /** Clear all recording state - does not clear stats */ public void clearRecordingState() { initRecordingState() ; transactionStateTransition.clear() ; } private void initRecordingState() { if ( transactionStateTransition == null ) transactionStateTransition = new ArrayList<>() ; } public Journal getJournal() { return journal ; } // ---- Logging // Choose log output once when this object is created. private final boolean logstate = (syslog.isDebugEnabled() || log.isDebugEnabled()) ; private boolean log() { return logstate ; } private void log(String msg, Transaction txn) { if ( ! log() ) return ; if ( txn == null ) logger().debug(": "+msg) ; else logger().debug(txn.getLabel()+": "+msg) ; } private void logInternal(String action, Transaction txn) { if ( ! log() ) return ; String txnStr = ( txn == null ) ? "" : txn.getLabel() ; //System.err.printf(format("%6s %s -- %s", action, txnStr, state())) ; logger().debug(format("%6s %s -- %s", action, txnStr, state())) ; } private static Logger logger() { if ( syslog.isDebugEnabled() ) return syslog ; else return log ; } synchronized public SysTxnState state() { return new SysTxnState(this) ; } // LATER. class Committer implements Runnable { @Override public void run() { for ( ;; ) { // Wait until the reader count goes to zero. // This wakes up for every transation but maybe // able to play several transactions at once (later). try { Transaction txn = queue.take() ; // This takes a Write lock on the DSG - this is where it blocks. JournalControl.replay(txn) ; synchronized (TransactionManager.this) { commitedAwaitingFlush.remove(txn) ; } } catch (InterruptedException ex) { Log.error(this, "Interruped!", ex) ; } } } } private void transactionStarts(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.transactionStarts(txn) ; } private void transactionFinishes(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.transactionFinishes(txn) ; } private void transactionCloses(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.transactionCloses(txn) ; } private void readerStarts(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.readerStarts(txn) ; } private void readerFinishes(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.readerFinishes(txn) ; } private void transactionPromotes(Transaction txnOld, Transaction txnNew) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.transactionPromotes(txnOld, txnNew); } private void writerStarts(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.writerStarts(txn) ; } private void writerCommits(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.writerCommits(txn) ; } private void writerAborts(Transaction txn) { for ( TSM tsm : actions ) if ( tsm != null ) tsm.writerAborts(txn) ; } }





© 2015 - 2024 Weber Informatics LLC | Privacy Policy