void WriteBatchExecutor::bulkExecute( const BatchedCommandRequest& request,
std::vector<BatchedUpsertDetail*>* upsertedIds,
std::vector<WriteErrorDetail*>* errors ) {
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert ) {
execInserts( request, errors );
}
else if ( request.getBatchType() == BatchedCommandRequest::BatchType_Update ) {
for ( size_t i = 0; i < request.sizeWriteOps(); i++ ) {
WriteErrorDetail* error = NULL;
BSONObj upsertedId;
execUpdate( BatchItemRef( &request, i ), &upsertedId, &error );
if ( !upsertedId.isEmpty() ) {
BatchedUpsertDetail* batchUpsertedId = new BatchedUpsertDetail;
batchUpsertedId->setIndex( i );
batchUpsertedId->setUpsertedID( upsertedId );
upsertedIds->push_back( batchUpsertedId );
}
if ( error ) {
errors->push_back( error );
if ( request.getOrdered() )
break;
}
}
}
else {
dassert( request.getBatchType() == BatchedCommandRequest::BatchType_Delete );
for ( size_t i = 0; i < request.sizeWriteOps(); i++ ) {
WriteErrorDetail* error = NULL;
execRemove( BatchItemRef( &request, i ), &error );
if ( error ) {
errors->push_back( error );
if ( request.getOrdered() )
break;
}
}
}
// Fill in stale version errors for unordered batches (update/delete can't do this on own)
if ( !errors->empty() && !request.getOrdered() ) {
const WriteErrorDetail* finalError = errors->back();
if ( finalError->getErrCode() == ErrorCodes::StaleShardVersion ) {
for ( size_t i = finalError->getIndex() + 1; i < request.sizeWriteOps(); i++ ) {
WriteErrorDetail* dupStaleError = new WriteErrorDetail;
finalError->cloneTo( dupStaleError );
errors->push_back( dupStaleError );
}
}
}
}
void BatchSafeWriter::safeWriteBatch( DBClientBase* conn,
const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
// N starts at zero, and we add to it for each item
response->setN( 0 );
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
BatchItemRef itemRef( &request, static_cast<int>( i ) );
LastError lastError;
_safeWriter->safeWrite( conn, itemRef, &lastError );
// Register the error if we need to
BatchedErrorDetail* batchError = lastErrorToBatchError( lastError );
if ( batchError ) {
batchError->setIndex( i );
response->addToErrDetails( batchError );
}
response->setN( response->getN() + lastError.nObjects );
if ( !lastError.upsertedId.isEmpty() ) {
BatchedUpsertDetail* upsertedId = new BatchedUpsertDetail;
upsertedId->setIndex( i );
upsertedId->setUpsertedID( lastError.upsertedId );
response->addToUpsertDetails( upsertedId );
}
// Break on first error if we're ordered
if ( request.getOrdered() && BatchSafeWriter::isFailedOp( lastError ) ) break;
}
if ( request.sizeWriteOps() == 1 && response->isErrDetailsSet()
&& !response->isErrCodeSet() ) {
// Promote single error to batch error
const BatchedErrorDetail* error = response->getErrDetailsAt( 0 );
response->setErrCode( error->getErrCode() );
if ( error->isErrInfoSet() ) response->setErrInfo( error->getErrInfo() );
response->setErrMessage( error->getErrMessage() );
response->unsetErrDetails();
}
if ( request.sizeWriteOps() == 1 && response->isUpsertDetailsSet() ) {
// Promote single upsert to batch upsert
const BatchedUpsertDetail* upsertedId = response->getUpsertDetailsAt( 0 );
response->setSingleUpserted( upsertedId->getUpsertedID() );
response->unsetUpsertDetails();
}
response->setOk( !response->isErrCodeSet() );
dassert( response->isValid( NULL ) );
}
void WriteBatchExecutor::execInserts( const BatchedCommandRequest& request,
std::vector<WriteErrorDetail*>* errors ) {
// Theory of operation:
//
// Instantiates an ExecInsertsState, which represents all of the state involved in the batch
// insert execution algorithm. Most importantly, encapsulates the lock state.
//
// Every iteration of the loop in execInserts() processes one document insertion, by calling
// insertOne() exactly once for a given value of state.currIndex.
//
// If the ExecInsertsState indicates that the requisite write locks are not held, insertOne
// acquires them and performs lock-acquisition-time checks. However, on non-error
// execution, it does not release the locks. Therefore, the yielding logic in the while
// loop in execInserts() is solely responsible for lock release in the non-error case.
//
// Internally, insertOne loops performing the single insert until it completes without a
// PageFaultException, or until it fails with some kind of error. Errors are mostly
// propagated via the request->error field, but DBExceptions or std::exceptions may escape,
// particularly on operation interruption. These kinds of errors necessarily prevent
// further insertOne calls, and stop the batch. As a result, the only expected source of
// such exceptions are interruptions.
ExecInsertsState state(&request);
normalizeInserts(request, &state.normalizedInserts, &state.pregeneratedKeys);
ElapsedTracker elapsedTracker(128, 10); // 128 hits or 10 ms, matching RunnerYieldPolicy's
for (state.currIndex = 0;
state.currIndex < state.request->sizeWriteOps();
++state.currIndex) {
if (elapsedTracker.intervalHasElapsed()) {
// Consider yielding between inserts.
if (state.hasLock()) {
int micros = ClientCursor::suggestYieldMicros();
if (micros > 0) {
state.unlock();
killCurrentOp.checkForInterrupt();
sleepmicros(micros);
}
}
killCurrentOp.checkForInterrupt();
elapsedTracker.resetLastTime();
}
WriteErrorDetail* error = NULL;
execOneInsert(&state, &error);
if (error) {
errors->push_back(error);
error->setIndex(state.currIndex);
if (request.getOrdered())
return;
}
}
}
// Goes over the request and preprocesses normalized versions of all the inserts in the request
static void normalizeInserts( const BatchedCommandRequest& request,
vector<StatusWith<BSONObj> >* normalInserts ) {
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
StatusWith<BSONObj> normalInsert = normalizeInsert( BatchItemRef( &request, i ) );
normalInserts->push_back( normalInsert );
if ( request.getOrdered() && !normalInsert.isOK() )
break;
}
}
// Goes over the request and preprocesses normalized versions of all the inserts in the request
static void normalizeInserts( const BatchedCommandRequest& request,
vector<StatusWith<BSONObj> >* normalInserts ) {
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
BSONObj insertDoc = request.getInsertRequest()->getDocumentsAt( i );
StatusWith<BSONObj> normalInsert = fixDocumentForInsert( insertDoc );
normalInserts->push_back( normalInsert );
if ( request.getOrdered() && !normalInsert.isOK() )
break;
}
}
void Strategy::writeOp(OperationContext* txn, int op, Request& request) {
// make sure we have a last error
dassert(&LastError::get(cc()));
OwnedPointerVector<BatchedCommandRequest> commandRequestsOwned;
vector<BatchedCommandRequest*>& commandRequests = commandRequestsOwned.mutableVector();
msgToBatchRequests(request.m(), &commandRequests);
for (vector<BatchedCommandRequest*>::iterator it = commandRequests.begin();
it != commandRequests.end();
++it) {
// Multiple commands registered to last error as multiple requests
if (it != commandRequests.begin())
LastError::get(cc()).startRequest();
BatchedCommandRequest* commandRequest = *it;
// Adjust namespaces for command
NamespaceString fullNS(commandRequest->getNS());
string cmdNS = fullNS.getCommandNS();
// We only pass in collection name to command
commandRequest->setNS(fullNS);
BSONObjBuilder builder;
BSONObj requestBSON = commandRequest->toBSON();
{
// Disable the last error object for the duration of the write cmd
LastError::Disabled disableLastError(&LastError::get(cc()));
Command::runAgainstRegistered(txn, cmdNS.c_str(), requestBSON, builder, 0);
}
BatchedCommandResponse commandResponse;
bool parsed = commandResponse.parseBSON(builder.done(), NULL);
(void)parsed; // for compile
dassert(parsed && commandResponse.isValid(NULL));
// Populate the lastError object based on the write response
LastError::get(cc()).reset();
bool hadError =
batchErrorToLastError(*commandRequest, commandResponse, &LastError::get(cc()));
// Check if this is an ordered batch and we had an error which should stop processing
if (commandRequest->getOrdered() && hadError)
break;
}
}
// Goes over the request and preprocesses normalized versions of all the inserts in the request
static void normalizeInserts( const BatchedCommandRequest& request,
vector<StatusWith<BSONObj> >* normalizedInserts,
vector<PregeneratedKeys>* pregen ) {
normalizedInserts->reserve(request.sizeWriteOps());
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
BSONObj insertDoc = request.getInsertRequest()->getDocumentsAt( i );
StatusWith<BSONObj> normalInsert = fixDocumentForInsert( insertDoc );
normalizedInserts->push_back( normalInsert );
if ( request.getOrdered() && !normalInsert.isOK() )
break;
if ( !normalInsert.getValue().isEmpty() )
insertDoc = normalInsert.getValue();
pregen->push_back( PregeneratedKeys() );
GeneratorHolder::getInstance()->prepare( request.getTargetingNS(),
insertDoc,
&pregen->back() );
}
}
void BatchSafeWriter::safeWriteBatch( DBClientBase* conn,
const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
BatchItemRef itemRef( &request, static_cast<int>( i ) );
LastError lastError;
_safeWriter->safeWrite( conn, itemRef, &lastError );
// Register the error if we need to
BatchedErrorDetail* batchError = lastErrorToBatchError( lastError );
batchError->setIndex( i );
response->addToErrDetails( batchError );
// TODO: Other stats, etc.
// Break on first error if we're ordered
if ( request.getOrdered() && BatchSafeWriter::isFailedOp( lastError ) ) break;
}
}
void WriteBatchExecutor::executeBatch( const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
Timer commandTimer;
WriteStats stats;
std::auto_ptr<BatchedErrorDetail> error( new BatchedErrorDetail );
BSONObj upsertedID = BSONObj();
bool batchSuccess = true;
bool staleBatch = false;
// Apply each batch item, stopping on an error if we were asked to apply the batch
// sequentially.
size_t numBatchOps = request.sizeWriteOps();
bool verbose = request.isVerboseWC();
for ( size_t i = 0; i < numBatchOps; i++ ) {
if ( applyWriteItem( BatchItemRef( &request, i ),
&stats,
&upsertedID,
error.get() ) ) {
// In case updates turned out to be upserts, the callers may be interested
// in learning what _id was used for that document.
if ( !upsertedID.isEmpty() ) {
if ( numBatchOps == 1 ) {
response->setSingleUpserted(upsertedID);
}
else if ( verbose ) {
std::auto_ptr<BatchedUpsertDetail> upsertDetail(new BatchedUpsertDetail);
upsertDetail->setIndex(i);
upsertDetail->setUpsertedID(upsertedID);
response->addToUpsertDetails(upsertDetail.release());
}
upsertedID = BSONObj();
}
}
else {
// The applyWriteItem did not go thgrou
// If the error is sharding related, we'll have to investigate whether we
// have a stale view of sharding state.
if ( error->getErrCode() == ErrorCodes::StaleShardVersion ) staleBatch = true;
// Don't bother recording if the user doesn't want a verbose answer. We want to
// keep the error if this is a one-item batch, since we already compact the
// response for those.
if (verbose || numBatchOps == 1) {
error->setIndex( static_cast<int>( i ) );
response->addToErrDetails( error.release() );
}
batchSuccess = false;
if ( request.getOrdered() ) break;
error.reset( new BatchedErrorDetail );
}
}
// So far, we may have failed some of the batch's items. So we record
// that. Rergardless, we still need to apply the write concern. If that generates a
// more specific error, we'd replace for the intermediate error here. Note that we
// "compatct" the error messge if this is an one-item batch. (See rationale later in
// this file.)
if ( !batchSuccess ) {
if (numBatchOps > 1) {
// TODO
// Define the final error code here.
// Might be used as a final error, depending on write concern success.
response->setErrCode( 99999 );
response->setErrMessage( "batch op errors occurred" );
}
else {
// Promote the single error.
const BatchedErrorDetail* error = response->getErrDetailsAt( 0 );
response->setErrCode( error->getErrCode() );
if ( error->isErrInfoSet() ) response->setErrInfo( error->getErrInfo() );
response->setErrMessage( error->getErrMessage() );
response->unsetErrDetails();
error = NULL;
}
}
// Apply write concern. Note, again, that we're only assembling a full response if the
// user is interested in it.
BSONObj writeConcern;
if ( request.isWriteConcernSet() ) {
writeConcern = request.getWriteConcern();
}
else {
writeConcern = _defaultWriteConcern;
}
string errMsg;
BSONObjBuilder wcResultsB;
if ( !waitForWriteConcern( writeConcern, !batchSuccess, &wcResultsB, &errMsg ) ) {
// TODO Revisit when user visible family error codes are set
response->setErrCode( ErrorCodes::WriteConcernFailed );
response->setErrMessage( errMsg );
if ( verbose ) {
response->setErrInfo( wcResultsB.obj() );
}
}
// TODO: Audit where we want to queue here
if ( staleBatch ) {
ChunkVersion latestShardVersion;
shardingState.refreshMetadataIfNeeded( request.getTargetingNS(),
request.getShardVersion(),
&latestShardVersion );
}
// Set the main body of the response. We assume that, if there was an error, the error
// code would already be set.
response->setOk( !response->isErrCodeSet() );
response->setN( stats.numInserted + stats.numUpserted + stats.numUpdated
+ stats.numDeleted );
dassert( response->isValid( NULL ) );
}
void WriteBatchExecutor::executeBatch( const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
// Validate namespace
const NamespaceString nss = NamespaceString( request.getNS() );
if ( !nss.isValid() ) {
toBatchError( Status( ErrorCodes::InvalidNamespace,
nss.ns() + " is not a valid namespace" ),
response );
return;
}
// Make sure we can write to the namespace
Status allowedStatus = userAllowedWriteNS( nss );
if ( !allowedStatus.isOK() ) {
toBatchError( allowedStatus, response );
return;
}
// Validate insert index requests
// TODO: Push insert index requests through createIndex once all upgrade paths support it
string errMsg;
if ( request.isInsertIndexRequest() && !request.isValidIndexRequest( &errMsg ) ) {
toBatchError( Status( ErrorCodes::InvalidOptions, errMsg ), response );
return;
}
// Validate write concern
// TODO: Lift write concern parsing out of this entirely
WriteConcernOptions writeConcern;
BSONObj wcDoc;
if ( request.isWriteConcernSet() ) {
wcDoc = request.getWriteConcern();
}
Status wcStatus = Status::OK();
if ( wcDoc.isEmpty() ) {
// The default write concern if empty is w : 1
// Specifying w : 0 is/was allowed, but is interpreted identically to w : 1
wcStatus = writeConcern.parse(
_defaultWriteConcern.isEmpty() ?
WriteConcernOptions::Acknowledged : _defaultWriteConcern );
if ( writeConcern.wNumNodes == 0 && writeConcern.wMode.empty() ) {
writeConcern.wNumNodes = 1;
}
}
else {
wcStatus = writeConcern.parse( wcDoc );
}
if ( wcStatus.isOK() ) {
wcStatus = validateWriteConcern( writeConcern );
}
if ( !wcStatus.isOK() ) {
toBatchError( wcStatus, response );
return;
}
if ( request.sizeWriteOps() == 0u ) {
toBatchError( Status( ErrorCodes::InvalidLength,
"no write ops were included in the batch" ),
response );
return;
}
// Validate batch size
if ( request.sizeWriteOps() > BatchedCommandRequest::kMaxWriteBatchSize ) {
toBatchError( Status( ErrorCodes::InvalidLength,
stream() << "exceeded maximum write batch size of "
<< BatchedCommandRequest::kMaxWriteBatchSize ),
response );
return;
}
//
// End validation
//
bool silentWC = writeConcern.wMode.empty() && writeConcern.wNumNodes == 0
&& writeConcern.syncMode == WriteConcernOptions::NONE;
Timer commandTimer;
OwnedPointerVector<WriteErrorDetail> writeErrorsOwned;
vector<WriteErrorDetail*>& writeErrors = writeErrorsOwned.mutableVector();
OwnedPointerVector<BatchedUpsertDetail> upsertedOwned;
vector<BatchedUpsertDetail*>& upserted = upsertedOwned.mutableVector();
//
// Apply each batch item, possibly bulking some items together in the write lock.
// Stops on error if batch is ordered.
//
bulkExecute( request, &upserted, &writeErrors );
//
// Try to enforce the write concern if everything succeeded (unordered or ordered)
// OR if something succeeded and we're unordered.
//
auto_ptr<WCErrorDetail> wcError;
bool needToEnforceWC = writeErrors.empty()
|| ( !request.getOrdered()
&& writeErrors.size() < request.sizeWriteOps() );
if ( needToEnforceWC ) {
_client->curop()->setMessage( "waiting for write concern" );
WriteConcernResult res;
Status status = waitForWriteConcern( _txn, writeConcern, _client->getLastOp(), &res );
if ( !status.isOK() ) {
wcError.reset( toWriteConcernError( status, res ) );
}
}
//
// Refresh metadata if needed
//
bool staleBatch = !writeErrors.empty()
&& writeErrors.back()->getErrCode() == ErrorCodes::StaleShardVersion;
if ( staleBatch ) {
const BatchedRequestMetadata* requestMetadata = request.getMetadata();
dassert( requestMetadata );
// Make sure our shard name is set or is the same as what was set previously
if ( shardingState.setShardName( requestMetadata->getShardName() ) ) {
//
// First, we refresh metadata if we need to based on the requested version.
//
ChunkVersion latestShardVersion;
shardingState.refreshMetadataIfNeeded( request.getTargetingNS(),
requestMetadata->getShardVersion(),
&latestShardVersion );
// Report if we're still changing our metadata
// TODO: Better reporting per-collection
if ( shardingState.inCriticalMigrateSection() ) {
noteInCriticalSection( writeErrors.back() );
}
if ( queueForMigrationCommit ) {
//
// Queue up for migration to end - this allows us to be sure that clients will
// not repeatedly try to refresh metadata that is not yet written to the config
// server. Not necessary for correctness.
// Exposed as optional parameter to allow testing of queuing behavior with
// different network timings.
//
const ChunkVersion& requestShardVersion = requestMetadata->getShardVersion();
//
// Only wait if we're an older version (in the current collection epoch) and
// we're not write compatible, implying that the current migration is affecting
// writes.
//
if ( requestShardVersion.isOlderThan( latestShardVersion ) &&
!requestShardVersion.isWriteCompatibleWith( latestShardVersion ) ) {
while ( shardingState.inCriticalMigrateSection() ) {
log() << "write request to old shard version "
<< requestMetadata->getShardVersion().toString()
<< " waiting for migration commit" << endl;
shardingState.waitTillNotInCriticalSection( 10 /* secs */);
}
}
}
}
else {
// If our shard name is stale, our version must have been stale as well
dassert( writeErrors.size() == request.sizeWriteOps() );
}
}
//
// Construct response
//
response->setOk( true );
if ( !silentWC ) {
if ( upserted.size() ) {
response->setUpsertDetails( upserted );
}
if ( writeErrors.size() ) {
response->setErrDetails( writeErrors );
}
if ( wcError.get() ) {
response->setWriteConcernError( wcError.release() );
}
const repl::ReplicationCoordinator::Mode replMode =
repl::getGlobalReplicationCoordinator()->getReplicationMode();
if (replMode != repl::ReplicationCoordinator::modeNone) {
response->setLastOp( _client->getLastOp() );
if (replMode == repl::ReplicationCoordinator::modeReplSet) {
response->setElectionId(repl::theReplSet->getElectionId());
}
}
// Set the stats for the response
response->setN( _stats->numInserted + _stats->numUpserted + _stats->numMatched
+ _stats->numDeleted );
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Update )
response->setNModified( _stats->numModified );
}
dassert( response->isValid( NULL ) );
}
void WriteBatchExecutor::execInserts( const BatchedCommandRequest& request,
std::vector<WriteErrorDetail*>* errors ) {
// Bulk insert is a bit different from other bulk operations in that multiple request docs
// can be processed at once inside the write lock.
const NamespaceString nss( request.getTargetingNS() );
scoped_ptr<BatchItemRef> currInsertItem( new BatchItemRef( &request, 0 ) );
// Go through our request and do some preprocessing on insert documents outside the lock to
// validate and put them in a normalized form - i.e. put _id in front and fill in
// timestamps. The insert document may also be invalid.
// TODO: Might be more efficient to do in batches.
vector<StatusWith<BSONObj> > normalInserts;
normalizeInserts( request, &normalInserts );
while ( currInsertItem->getItemIndex() < static_cast<int>( request.sizeWriteOps() ) ) {
WriteOpResult currResult;
// Don't (re-)acquire locks and create database until it's necessary
if ( !normalInserts[currInsertItem->getItemIndex()].isOK() ) {
currResult.error =
toWriteError( normalInserts[currInsertItem->getItemIndex()].getStatus() );
}
else {
PageFaultRetryableSection pFaultSection;
////////////////////////////////////
Lock::DBWrite writeLock( nss.ns() );
////////////////////////////////////
// Check version inside of write lock
if ( checkIsMasterForCollection( nss, &currResult.error )
&& checkShardVersion( &shardingState, request, &currResult.error )
&& checkIndexConstraints( &shardingState, request, &currResult.error ) ) {
//
// Get the collection for the insert
//
scoped_ptr<Client::Context> writeContext;
Collection* collection = NULL;
try {
// Context once we're locked, to set more details in currentOp()
// TODO: better constructor?
writeContext.reset( new Client::Context( request.getNS(),
storageGlobalParams.dbpath,
false /* don't check version */) );
Database* database = writeContext->db();
dassert( database );
collection = database->getCollection( nss.ns() );
if ( !collection ) {
// Implicitly create if it doesn't exist
collection = database->createCollection( nss.ns() );
if ( !collection ) {
currResult.error =
toWriteError( Status( ErrorCodes::InternalError,
"could not create collection" ) );
}
}
}
catch ( const DBException& ex ) {
Status status(ex.toStatus());
if (ErrorCodes::isInterruption(status.code())) {
throw;
}
currResult.error = toWriteError(status);
}
//
// Perform writes inside write lock
//
while ( collection
&& currInsertItem->getItemIndex()
< static_cast<int>( request.sizeWriteOps() ) ) {
//
// BEGIN CURRENT OP
//
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, *currInsertItem ) );
incOpStats( *currInsertItem );
// Get the actual document we want to write, assuming it's valid
const StatusWith<BSONObj>& normalInsert = //
normalInserts[currInsertItem->getItemIndex()];
const BSONObj& normalInsertDoc =
normalInsert.getValue().isEmpty() ?
currInsertItem->getDocument() : normalInsert.getValue();
if ( !normalInsert.isOK() ) {
// This insert failed on preprocessing
currResult.error = toWriteError( normalInsert.getStatus() );
}
else if ( !request.isInsertIndexRequest() ) {
// Try the insert
singleInsert( *currInsertItem,
normalInsertDoc,
collection,
&currResult );
}
else {
// Try the create index
singleCreateIndex( *currInsertItem,
normalInsertDoc,
collection,
&currResult );
}
//
// END CURRENT OP
//
finishCurrentOp( _client, currentOp.get(), currResult.error );
// Faults release the write lock
if ( currResult.fault )
break;
// In general, we might have stats and errors
incWriteStats( *currInsertItem,
currResult.stats,
currResult.error,
currentOp.get() );
// Errors release the write lock
if ( currResult.error )
break;
// Increment in the write lock and reset the stats for next time
currInsertItem.reset( new BatchItemRef( &request,
currInsertItem->getItemIndex()
+ 1 ) );
currResult.reset();
// Destruct curop so that our parent curop is restored, so that we
// record the yield count in the parent.
currentOp.reset(NULL);
// yield sometimes
int micros = ClientCursor::suggestYieldMicros();
if (micros > 0) {
ClientCursor::staticYield(micros, "", NULL);
}
}
}
} // END WRITE LOCK
//
// Store the current error if it exists
//
if ( currResult.error ) {
errors->push_back( currResult.releaseError() );
errors->back()->setIndex( currInsertItem->getItemIndex() );
// Break early for ordered batches
if ( request.getOrdered() )
break;
}
//
// Fault or increment
//
if ( currResult.fault ) {
// Check page fault out of lock
currResult.fault->touch();
}
else {
// Increment if not a fault
currInsertItem.reset( new BatchItemRef( &request,
currInsertItem->getItemIndex() + 1 ) );
}
}
}
void WriteBatchExecutor::executeBatch( const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
// TODO: Lift write concern parsing out of this entirely.
WriteConcernOptions writeConcern;
Status status = Status::OK();
BSONObj wcDoc;
if ( request.isWriteConcernSet() ) {
wcDoc = request.getWriteConcern();
}
if ( wcDoc.isEmpty() ) {
status = writeConcern.parse( _defaultWriteConcern );
}
else {
status = writeConcern.parse( wcDoc );
}
if ( status.isOK() ) {
status = validateWriteConcern( writeConcern );
}
if ( !status.isOK() ) {
response->setErrCode( status.code() );
response->setErrMessage( status.reason() );
response->setOk( false );
dassert( response->isValid(NULL) );
return;
}
bool silentWC = writeConcern.wMode.empty() && writeConcern.wNumNodes == 0
&& writeConcern.syncMode == WriteConcernOptions::NONE;
Timer commandTimer;
OwnedPointerVector<WriteErrorDetail> writeErrorsOwned;
vector<WriteErrorDetail*>& writeErrors = writeErrorsOwned.mutableVector();
OwnedPointerVector<BatchedUpsertDetail> upsertedOwned;
vector<BatchedUpsertDetail*>& upserted = upsertedOwned.mutableVector();
//
// Apply each batch item, possibly bulking some items together in the write lock.
// Stops on error if batch is ordered.
//
bulkExecute( request, &upserted, &writeErrors );
//
// Try to enforce the write concern if everything succeeded (unordered or ordered)
// OR if something succeeded and we're unordered.
//
auto_ptr<WCErrorDetail> wcError;
bool needToEnforceWC = writeErrors.empty()
|| ( !request.getOrdered()
&& writeErrors.size() < request.sizeWriteOps() );
if ( needToEnforceWC ) {
_client->curop()->setMessage( "waiting for write concern" );
WriteConcernResult res;
status = waitForWriteConcern( writeConcern, _client->getLastOp(), &res );
if ( !status.isOK() ) {
wcError.reset( toWriteConcernError( status, res ) );
}
}
//
// Refresh metadata if needed
//
bool staleBatch = !writeErrors.empty()
&& writeErrors.back()->getErrCode() == ErrorCodes::StaleShardVersion;
if ( staleBatch ) {
const BatchedRequestMetadata* requestMetadata = request.getMetadata();
dassert( requestMetadata );
// Make sure our shard name is set or is the same as what was set previously
if ( shardingState.setShardName( requestMetadata->getShardName() ) ) {
//
// First, we refresh metadata if we need to based on the requested version.
//
ChunkVersion latestShardVersion;
shardingState.refreshMetadataIfNeeded( request.getTargetingNS(),
requestMetadata->getShardVersion(),
&latestShardVersion );
// Report if we're still changing our metadata
// TODO: Better reporting per-collection
if ( shardingState.inCriticalMigrateSection() ) {
noteInCriticalSection( writeErrors.back() );
}
if ( queueForMigrationCommit ) {
//
// Queue up for migration to end - this allows us to be sure that clients will
// not repeatedly try to refresh metadata that is not yet written to the config
// server. Not necessary for correctness.
// Exposed as optional parameter to allow testing of queuing behavior with
// different network timings.
//
const ChunkVersion& requestShardVersion = requestMetadata->getShardVersion();
//
// Only wait if we're an older version (in the current collection epoch) and
// we're not write compatible, implying that the current migration is affecting
// writes.
//
if ( requestShardVersion.isOlderThan( latestShardVersion ) &&
!requestShardVersion.isWriteCompatibleWith( latestShardVersion ) ) {
while ( shardingState.inCriticalMigrateSection() ) {
log() << "write request to old shard version "
<< requestMetadata->getShardVersion().toString()
<< " waiting for migration commit" << endl;
shardingState.waitTillNotInCriticalSection( 10 /* secs */);
}
}
}
}
else {
// If our shard name is stale, our version must have been stale as well
dassert( writeErrors.size() == request.sizeWriteOps() );
}
}
//
// Construct response
//
response->setOk( true );
if ( !silentWC ) {
if ( upserted.size() ) {
response->setUpsertDetails( upserted );
upserted.clear();
}
if ( writeErrors.size() ) {
response->setErrDetails( writeErrors );
writeErrors.clear();
}
if ( wcError.get() ) {
response->setWriteConcernError( wcError.release() );
}
if ( anyReplEnabled() ) {
response->setLastOp( _client->getLastOp() );
if (theReplSet) {
response->setElectionId( theReplSet->getElectionId() );
}
}
// Set the stats for the response
response->setN( _stats->numInserted + _stats->numUpserted + _stats->numMatched
+ _stats->numDeleted );
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Update )
response->setNModified( _stats->numModified );
}
dassert( response->isValid( NULL ) );
}
void BatchSafeWriter::safeWriteBatch( DBClientBase* conn,
const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
const NamespaceString nss( request.getNS() );
// N starts at zero, and we add to it for each item
response->setN( 0 );
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
// Break on first error if we're ordered
if ( request.getOrdered() && response->isErrDetailsSet() )
break;
BatchItemRef itemRef( &request, static_cast<int>( i ) );
bool isLastItem = ( i == request.sizeWriteOps() - 1 );
BSONObj writeConcern;
if ( isLastItem && request.isWriteConcernSet() ) {
writeConcern = request.getWriteConcern();
// Pre-2.4.2 mongods react badly to 'w' being set on config servers
if ( nss.db() == "config" )
writeConcern = fixWCForConfig( writeConcern );
}
BSONObj gleResult;
GLEErrors errors;
Status status = _safeWriter->safeWrite( conn, itemRef, writeConcern, &gleResult );
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
response->clear();
response->setOk( false );
response->setErrCode( status.code() );
response->setErrMessage( status.reason() );
return;
}
//
// STATS HANDLING
//
GLEStats stats;
extractGLEStats( gleResult, &stats );
// Special case for making legacy "n" field result for insert match the write
// command result.
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert
&& !errors.writeError.get() ) {
// n is always 0 for legacy inserts.
dassert( stats.n == 0 );
stats.n = 1;
}
response->setN( response->getN() + stats.n );
if ( !stats.upsertedId.isEmpty() ) {
BatchedUpsertDetail* upsertedId = new BatchedUpsertDetail;
upsertedId->setIndex( i );
upsertedId->setUpsertedID( stats.upsertedId );
response->addToUpsertDetails( upsertedId );
}
response->setLastOp( stats.lastOp );
//
// WRITE ERROR HANDLING
//
// If any error occurs (except stale config) the previous GLE was not enforced
bool enforcedWC = !errors.writeError.get()
|| errors.writeError->getErrCode() == ErrorCodes::StaleShardVersion;
// Save write error
if ( errors.writeError.get() ) {
errors.writeError->setIndex( i );
response->addToErrDetails( errors.writeError.release() );
}
//
// WRITE CONCERN ERROR HANDLING
//
// The last write is weird, since we enforce write concern and check the error through
// the same GLE if possible. If the last GLE was an error, the write concern may not
// have been enforced in that same GLE, so we need to send another after resetting the
// error.
if ( isLastItem ) {
// Try to enforce the write concern if everything succeeded (unordered or ordered)
// OR if something succeeded and we're unordered.
bool needToEnforceWC =
!response->isErrDetailsSet()
|| ( !request.getOrdered()
&& response->sizeErrDetails() < request.sizeWriteOps() );
if ( !enforcedWC && needToEnforceWC ) {
dassert( !errors.writeError.get() ); // emptied above
// Might have gotten a write concern validity error earlier, these are
// enforced even if the wc isn't applied, so we ignore.
errors.wcError.reset();
Status status = _safeWriter->enforceWriteConcern( conn,
nss.db().toString(),
writeConcern,
&gleResult );
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
response->clear();
response->setOk( false );
response->setErrCode( status.code() );
response->setErrMessage( status.reason() );
return;
}
}
// END Write concern retry
if ( errors.wcError.get() ) {
response->setWriteConcernError( errors.wcError.release() );
}
}
}
response->setOk( true );
dassert( response->isValid( NULL ) );
}
void BatchSafeWriter::safeWriteBatch( DBClientBase* conn,
const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
const NamespaceString nss( request.getNS() );
// N starts at zero, and we add to it for each item
response->setN( 0 );
// GLE path always sets nModified to -1 (sentinel) to indicate we should omit it later.
response->setNModified(-1);
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
// Break on first error if we're ordered
if ( request.getOrdered() && response->isErrDetailsSet() )
break;
BatchItemRef itemRef( &request, static_cast<int>( i ) );
BSONObj gleResult;
GLEErrors errors;
Status status = _safeWriter->safeWrite( conn,
itemRef,
WriteConcernOptions::Acknowledged,
&gleResult );
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
response->clear();
response->setOk( false );
response->setErrCode( ErrorCodes::RemoteResultsUnavailable );
StringBuilder builder;
builder << "could not get write error from safe write";
builder << causedBy( status.toString() );
response->setErrMessage( builder.str() );
return;
}
if ( errors.wcError.get() ) {
response->setWriteConcernError( errors.wcError.release() );
}
//
// STATS HANDLING
//
GLEStats stats;
extractGLEStats( gleResult, &stats );
// Special case for making legacy "n" field result for insert match the write
// command result.
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert
&& !errors.writeError.get() ) {
// n is always 0 for legacy inserts.
dassert( stats.n == 0 );
stats.n = 1;
}
response->setN( response->getN() + stats.n );
if ( !stats.upsertedId.isEmpty() ) {
BatchedUpsertDetail* upsertedId = new BatchedUpsertDetail;
upsertedId->setIndex( i );
upsertedId->setUpsertedID( stats.upsertedId );
response->addToUpsertDetails( upsertedId );
}
response->setLastOp( stats.lastOp );
// Save write error
if ( errors.writeError.get() ) {
errors.writeError->setIndex( i );
response->addToErrDetails( errors.writeError.release() );
}
}
//
// WRITE CONCERN ERROR HANDLING
//
// The last write is weird, since we enforce write concern and check the error through
// the same GLE if possible. If the last GLE was an error, the write concern may not
// have been enforced in that same GLE, so we need to send another after resetting the
// error.
BSONObj writeConcern;
if ( request.isWriteConcernSet() ) {
writeConcern = request.getWriteConcern();
// Pre-2.4.2 mongods react badly to 'w' being set on config servers
if ( nss.db() == "config" )
writeConcern = fixWCForConfig( writeConcern );
}
bool needToEnforceWC = WriteConcernOptions::Acknowledged.woCompare(writeConcern) != 0 &&
WriteConcernOptions::Unacknowledged.woCompare(writeConcern) != 0;
if ( needToEnforceWC &&
( !response->isErrDetailsSet() ||
( !request.getOrdered() &&
// Not all errored. Note: implicit response->isErrDetailsSet().
response->sizeErrDetails() < request.sizeWriteOps() ))) {
// Might have gotten a write concern validity error earlier, these are
// enforced even if the wc isn't applied, so we ignore.
response->unsetWriteConcernError();
const string dbName( nss.db().toString() );
Status status( Status::OK() );
if ( response->isErrDetailsSet() ) {
const WriteErrorDetail* lastError = response->getErrDetails().back();
// If last write op was an error.
if ( lastError->getIndex() == static_cast<int>( request.sizeWriteOps() - 1 )) {
// Reset previous errors so we can apply the write concern no matter what
// as long as it is valid.
status = _safeWriter->clearErrors( conn, dbName );
}
}
BSONObj gleResult;
if ( status.isOK() ) {
status = _safeWriter->enforceWriteConcern( conn,
dbName,
writeConcern,
&gleResult );
}
GLEErrors errors;
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
auto_ptr<WCErrorDetail> wcError( new WCErrorDetail );
wcError->setErrCode( status.code() );
wcError->setErrMessage( status.reason() );
response->setWriteConcernError( wcError.release() );
}
else if ( errors.wcError.get() ) {
response->setWriteConcernError( errors.wcError.release() );
}
}
response->setOk( true );
dassert( response->isValid( NULL ) );
}