static bool checkShardVersion( ShardingState* shardingState,
const BatchedCommandRequest& request,
WriteErrorDetail** error ) {
const NamespaceString nss( request.getTargetingNS() );
Lock::assertWriteLocked( nss.ns() );
ChunkVersion requestShardVersion =
request.isMetadataSet() && request.getMetadata()->isShardVersionSet() ?
request.getMetadata()->getShardVersion() : ChunkVersion::IGNORED();
if ( shardingState->enabled() ) {
CollectionMetadataPtr metadata = shardingState->getCollectionMetadata( nss.ns() );
if ( !ChunkVersion::isIgnoredVersion( requestShardVersion ) ) {
ChunkVersion shardVersion =
metadata ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
if ( !requestShardVersion.isWriteCompatibleWith( shardVersion ) ) {
*error = new WriteErrorDetail;
buildStaleError( requestShardVersion, shardVersion, *error );
return false;
}
}
}
return true;
}
BatchedCommandResponse Shard::runBatchWriteCommandOnConfig(
OperationContext* txn, const BatchedCommandRequest& batchRequest, RetryPolicy retryPolicy) {
invariant(isConfig());
const std::string dbname = batchRequest.getNS().db().toString();
invariant(batchRequest.sizeWriteOps() == 1);
const BSONObj cmdObj = batchRequest.toBSON();
for (int retry = 1; retry <= kOnErrorNumRetries; ++retry) {
auto response = _runCommand(txn,
ReadPreferenceSetting{ReadPreference::PrimaryOnly},
dbname,
kDefaultConfigCommandTimeout,
cmdObj);
BatchedCommandResponse batchResponse;
Status writeStatus =
CommandResponse::processBatchWriteResponse(response.commandResponse, &batchResponse);
if (!writeStatus.isOK() && response.host) {
updateReplSetMonitor(response.host.get(), writeStatus);
}
if (retry < kOnErrorNumRetries && isRetriableError(writeStatus.code(), retryPolicy)) {
LOG(2) << "Batch write command failed with retriable error and will be retried"
<< causedBy(redact(writeStatus));
continue;
}
return batchResponse;
}
MONGO_UNREACHABLE;
}
static bool checkIndexConstraints( ShardingState* shardingState,
const BatchedCommandRequest& request,
WriteErrorDetail** error ) {
const NamespaceString nss( request.getTargetingNS() );
Lock::assertWriteLocked( nss.ns() );
if ( !request.isUniqueIndexRequest() )
return true;
if ( shardingState->enabled() ) {
CollectionMetadataPtr metadata = shardingState->getCollectionMetadata( nss.ns() );
if ( metadata ) {
if ( !isUniqueIndexCompatible( metadata->getKeyPattern(),
request.getIndexKeyPattern() ) ) {
*error = new WriteErrorDetail;
buildUniqueIndexError( metadata->getKeyPattern(),
request.getIndexKeyPattern(),
*error );
return false;
}
}
}
return true;
}
static bool checkShardVersion(OperationContext* txn,
ShardingState* shardingState,
const BatchedCommandRequest& request,
WriteOpResult* result) {
const NamespaceString nss( request.getTargetingNS() );
txn->lockState()->assertWriteLocked( nss.ns() );
ChunkVersion requestShardVersion =
request.isMetadataSet() && request.getMetadata()->isShardVersionSet() ?
request.getMetadata()->getShardVersion() : ChunkVersion::IGNORED();
if ( shardingState->enabled() ) {
CollectionMetadataPtr metadata = shardingState->getCollectionMetadata( nss.ns() );
if ( !ChunkVersion::isIgnoredVersion( requestShardVersion ) ) {
ChunkVersion shardVersion =
metadata ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
if ( !requestShardVersion.isWriteCompatibleWith( shardVersion ) ) {
result->setError(new WriteErrorDetail);
buildStaleError(requestShardVersion, shardVersion, result->getError());
return false;
}
}
}
return true;
}
BatchedCommandRequest* msgToBatchUpdate( const Message& updateMsg ) {
// Parsing DbMessage throws
DbMessage dbMsg( updateMsg );
NamespaceString nss( dbMsg.getns() );
int flags = dbMsg.pullInt();
bool upsert = flags & UpdateOption_Upsert;
bool multi = flags & UpdateOption_Multi;
const BSONObj query = dbMsg.nextJsObj();
const BSONObj updateExpr = dbMsg.nextJsObj();
// No exceptions from here on
BatchedUpdateDocument* updateDoc = new BatchedUpdateDocument;
updateDoc->setQuery( query );
updateDoc->setUpdateExpr( updateExpr );
updateDoc->setUpsert( upsert );
updateDoc->setMulti( multi );
BatchedCommandRequest* request =
new BatchedCommandRequest( BatchedCommandRequest::BatchType_Update );
request->setNS( nss.ns() );
request->getUpdateRequest()->addToUpdates( updateDoc );
return request;
}
static bool checkIndexConstraints(OperationContext* txn,
ShardingState* shardingState,
const BatchedCommandRequest& request,
WriteOpResult* result) {
const NamespaceString nss( request.getTargetingNS() );
txn->lockState()->assertWriteLocked( nss.ns() );
if ( !request.isUniqueIndexRequest() )
return true;
if ( shardingState->enabled() ) {
CollectionMetadataPtr metadata = shardingState->getCollectionMetadata( nss.ns() );
if ( metadata ) {
if ( !isUniqueIndexCompatible( metadata->getKeyPattern(),
request.getIndexKeyPattern() ) ) {
result->setError(new WriteErrorDetail);
buildUniqueIndexError(metadata->getKeyPattern(),
request.getIndexKeyPattern(),
result->getError());
return false;
}
}
}
return true;
}
void ClusterWriter::write(OperationContext* opCtx,
const BatchedCommandRequest& request,
BatchWriteExecStats* stats,
BatchedCommandResponse* response) {
const NamespaceString& nss = request.getNS();
LastError::Disabled disableLastError(&LastError::get(opCtx->getClient()));
// Config writes and shard writes are done differently
if (nss.db() == NamespaceString::kConfigDb || nss.db() == NamespaceString::kAdminDb) {
Grid::get(opCtx)->catalogClient()->writeConfigServerDirect(opCtx, request, response);
} else {
TargeterStats targeterStats;
{
ChunkManagerTargeter targeter(request.getTargetingNS(), &targeterStats);
Status targetInitStatus = targeter.init(opCtx);
if (!targetInitStatus.isOK()) {
toBatchError({targetInitStatus.code(),
str::stream() << "unable to target"
<< (request.isInsertIndexRequest() ? " index" : "")
<< " write op for collection "
<< request.getTargetingNS().ns()
<< causedBy(targetInitStatus)},
response);
return;
}
BatchWriteExec::executeBatch(opCtx, targeter, request, response, stats);
}
splitIfNeeded(opCtx, request.getNS(), targeterStats);
}
}
void ClusterWriter::shardWrite( const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
ChunkManagerTargeter targeter;
Status targetInitStatus = targeter.init( NamespaceString( request.getTargetingNS() ) );
if ( !targetInitStatus.isOK() ) {
warning() << "could not initialize targeter for"
<< ( request.isInsertIndexRequest() ? " index" : "" )
<< " write op in collection " << request.getTargetingNS() << endl;
// Errors will be reported in response if we are unable to target
}
DBClientShardResolver resolver;
DBClientMultiCommand dispatcher;
BatchWriteExec exec( &targeter, &resolver, &dispatcher );
exec.executeBatch( request, response );
if ( _autoSplit )
splitIfNeeded( request.getNS(), *targeter.getStats() );
_stats->setShardStats( exec.releaseStats() );
}
BatchedCommandResponse Shard::runBatchWriteCommand(OperationContext* opCtx,
const Milliseconds maxTimeMS,
const BatchedCommandRequest& batchRequest,
RetryPolicy retryPolicy) {
const std::string dbname = batchRequest.getNS().db().toString();
const BSONObj cmdObj = batchRequest.toBSON();
for (int retry = 1; retry <= kOnErrorNumRetries; ++retry) {
// Note: write commands can only be issued against a primary.
auto swResponse = _runCommand(
opCtx, ReadPreferenceSetting{ReadPreference::PrimaryOnly}, dbname, maxTimeMS, cmdObj);
BatchedCommandResponse batchResponse;
auto writeStatus = CommandResponse::processBatchWriteResponse(swResponse, &batchResponse);
if (retry < kOnErrorNumRetries && isRetriableError(writeStatus.code(), retryPolicy)) {
LOG(2) << "Batch write command to " << getId()
<< " failed with retriable error and will be retried"
<< causedBy(redact(writeStatus));
continue;
}
return batchResponse;
}
MONGO_UNREACHABLE;
}
BatchedCommandRequest* msgToBatchInsert( const Message& insertMsg ) {
// Parsing DbMessage throws
DbMessage dbMsg( insertMsg );
NamespaceString nss( dbMsg.getns() );
bool coe = dbMsg.reservedField() & Reserved_InsertOption_ContinueOnError;
vector<BSONObj> docs;
do {
docs.push_back( dbMsg.nextJsObj() );
}
while ( dbMsg.moreJSObjs() );
// Continue-on-error == unordered
bool ordered = !coe;
// No exceptions from here on
BatchedCommandRequest* request =
new BatchedCommandRequest( BatchedCommandRequest::BatchType_Insert );
request->setNS( nss.ns() );
for ( vector<BSONObj>::const_iterator it = docs.begin(); it != docs.end(); ++it ) {
request->getInsertRequest()->addToDocuments( *it );
}
request->setOrdered( ordered );
return request;
}
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 ) );
}
// 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 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 );
}
}
}
}
bool BatchedCommandRequest::containsNoIDUpsert(const BatchedCommandRequest& request) {
if (request.getBatchType() != BatchedCommandRequest::BatchType_Update)
return false;
const vector<BatchedUpdateDocument*>& updates = request.getUpdateRequest()->getUpdates();
for (vector<BatchedUpdateDocument*>::const_iterator it = updates.begin(); it != updates.end();
++it) {
const BatchedUpdateDocument* updateDoc = *it;
if (updateDoc->getUpsert() && updateDoc->getQuery()["_id"].eoo())
return true;
}
return false;
}
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;
}
}
void msgToBatchInserts( const Message& insertMsg,
vector<BatchedCommandRequest*>* insertRequests ) {
// Parsing DbMessage throws
DbMessage dbMsg( insertMsg );
NamespaceString nss( dbMsg.getns() );
// Continue-on-error == unordered
bool coe = dbMsg.reservedField() & Reserved_InsertOption_ContinueOnError;
bool ordered = !coe;
while ( insertRequests->empty() || dbMsg.moreJSObjs() ) {
// Collect docs for next batch, but don't exceed maximum size
int totalInsertSize = 0;
vector<BSONObj> docs;
do {
const char* prevObjMark = dbMsg.markGet();
BSONObj nextObj = dbMsg.nextJsObj();
if ( totalInsertSize + nextObj.objsize() <= BSONObjMaxUserSize ) {
docs.push_back( nextObj );
totalInsertSize += docs.back().objsize();
}
else {
// Size limit exceeded, rollback to previous insert position
dbMsg.markReset( prevObjMark );
break;
}
}
while ( docs.size() < BatchedCommandRequest::kMaxWriteBatchSize
&& dbMsg.moreJSObjs() );
dassert( !docs.empty() );
// No exceptions from here on
BatchedCommandRequest* request =
new BatchedCommandRequest( BatchedCommandRequest::BatchType_Insert );
request->setNSS( nss );
for ( vector<BSONObj>::const_iterator it = docs.begin(); it != docs.end(); ++it ) {
request->getInsertRequest()->addToDocuments( *it );
}
request->setOrdered( ordered );
request->setWriteConcern( WriteConcernOptions::Acknowledged );
insertRequests->push_back( request );
}
}
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;
}
}
}
void CatalogManagerReplicaSet::writeConfigServerDirect(const BatchedCommandRequest& batchRequest,
BatchedCommandResponse* batchResponse) {
std::string dbname = batchRequest.getNS().db().toString();
invariant(dbname == "config" || dbname == "admin");
const BSONObj cmdObj = batchRequest.toBSON();
auto response = _runConfigServerCommandWithNotMasterRetries(dbname, cmdObj);
if (!response.isOK()) {
_toBatchError(response.getStatus(), batchResponse);
return;
}
string errmsg;
if (!batchResponse->parseBSON(response.getValue(), &errmsg)) {
_toBatchError(Status(ErrorCodes::FailedToParse,
str::stream() << "Failed to parse config server response: " << errmsg),
batchResponse);
}
}
BatchedCommandRequest* msgToBatchDelete( const Message& deleteMsg ) {
// Parsing DbMessage throws
DbMessage dbMsg( deleteMsg );
NamespaceString nss( dbMsg.getns() );
int flags = dbMsg.pullInt();
const BSONObj query = dbMsg.nextJsObj();
int limit = ( flags & RemoveOption_JustOne ) ? 1 : 0;
// No exceptions from here on
BatchedDeleteDocument* deleteDoc = new BatchedDeleteDocument;
deleteDoc->setLimit( limit );
deleteDoc->setQuery( query );
BatchedCommandRequest* request =
new BatchedCommandRequest( BatchedCommandRequest::BatchType_Delete );
request->setNS( nss.ns() );
request->getDeleteRequest()->addToDeletes( deleteDoc );
return request;
}
// 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;
}
}
// static
Status WriteBatchExecutor::validateBatch( const BatchedCommandRequest& request ) {
// Validate namespace
const NamespaceString nss = NamespaceString( request.getNS() );
if ( !nss.isValid() ) {
return Status( ErrorCodes::InvalidNamespace,
nss.ns() + " is not a valid namespace" );
}
// Make sure we can write to the namespace
Status allowedStatus = userAllowedWriteNS( nss );
if ( !allowedStatus.isOK() ) {
return allowedStatus;
}
// 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 ) ) {
return Status( ErrorCodes::InvalidOptions, errMsg );
}
return Status::OK();
}
void clusterWrite( const BatchedCommandRequest& request,
BatchedCommandResponse* response,
bool autoSplit ) {
// App-level validation of a create index insert
if ( request.isInsertIndexRequest() ) {
if ( request.sizeWriteOps() != 1 || request.isWriteConcernSet() ) {
// Invalid request to create index
response->setOk( false );
response->setErrCode( ErrorCodes::InvalidOptions );
response->setErrMessage( "invalid batch request for index creation" );
dassert( response->isValid( NULL ) );
return;
}
}
// Config writes and shard writes are done differently
string dbName = NamespaceString( request.getNS() ).db().toString();
if ( dbName == "config" || dbName == "admin" ) {
bool verboseWC = request.isVerboseWC();
// We only support batch sizes of one and {w:0} write concern for config writes
if ( request.sizeWriteOps() != 1 || ( verboseWC && request.isWriteConcernSet() ) ) {
// Invalid config server write
response->setOk( false );
response->setErrCode( ErrorCodes::InvalidOptions );
response->setErrMessage( "invalid batch request for config write" );
dassert( response->isValid( NULL ) );
return;
}
// We need to support "best-effort" writes for pings to the config server.
// {w:0} (!verbose) writes are interpreted as best-effort in this case - they may still
// error, but do not do the initial fsync check.
configWrite( request, response, verboseWC );
}
else {
shardWrite( request, response, autoSplit );
}
}
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 BatchWriteExec::executeBatch( const BatchedCommandRequest& clientRequest,
BatchedCommandResponse* clientResponse ) {
BatchWriteOp batchOp;
batchOp.initClientRequest( &clientRequest );
// Current batch status
bool refreshedTargeter = false;
int rounds = 0;
int numCompletedOps = 0;
int numRoundsWithoutProgress = 0;
while ( !batchOp.isFinished() ) {
//
// Get child batches to send using the targeter
//
// Targeting errors can be caused by remote metadata changing (the collection could have
// been dropped and recreated, for example with a new shard key). If a remote metadata
// change occurs *before* a client sends us a batch, we need to make sure that we don't
// error out just because we're staler than the client - otherwise mongos will be have
// unpredictable behavior.
//
// (If a metadata change happens *during* or *after* a client sends us a batch, however,
// we make no guarantees about delivery.)
//
// For this reason, we don't record targeting errors until we've refreshed our targeting
// metadata at least once *after* receiving the client batch - at that point, we know:
//
// 1) our new metadata is the same as the metadata when the client sent a batch, and so
// targeting errors are real.
// OR
// 2) our new metadata is a newer version than when the client sent a batch, and so
// the metadata must have changed after the client batch was sent. We don't need to
// deliver in this case, since for all the client knows we may have gotten the batch
// exactly when the metadata changed.
//
vector<TargetedWriteBatch*> childBatches;
// If we've already had a targeting error, we've refreshed the metadata once and can
// record target errors definitively.
bool recordTargetErrors = refreshedTargeter;
Status targetStatus = batchOp.targetBatch( *_targeter,
recordTargetErrors,
&childBatches );
if ( !targetStatus.isOK() ) {
// Don't do anything until a targeter refresh
_targeter->noteCouldNotTarget();
refreshedTargeter = true;
++_stats->numTargetErrors;
dassert( childBatches.size() == 0u );
}
//
// Send all child batches
//
size_t numSent = 0;
size_t numToSend = childBatches.size();
bool remoteMetadataChanging = false;
while ( numSent != numToSend ) {
// Collect batches out on the network, mapped by endpoint
HostBatchMap pendingBatches;
//
// Send side
//
// Get as many batches as we can at once
for ( vector<TargetedWriteBatch*>::iterator it = childBatches.begin();
it != childBatches.end(); ++it ) {
//
// Collect the info needed to dispatch our targeted batch
//
TargetedWriteBatch* nextBatch = *it;
// If the batch is NULL, we sent it previously, so skip
if ( nextBatch == NULL ) continue;
// Figure out what host we need to dispatch our targeted batch
ConnectionString shardHost;
Status resolveStatus = _resolver->chooseWriteHost( nextBatch->getEndpoint()
.shardName,
&shardHost );
if ( !resolveStatus.isOK() ) {
++_stats->numResolveErrors;
// Record a resolve failure
// TODO: It may be necessary to refresh the cache if stale, or maybe just
// cancel and retarget the batch
WriteErrorDetail error;
buildErrorFrom( resolveStatus, &error );
batchOp.noteBatchError( *nextBatch, error );
// We're done with this batch
*it = NULL;
--numToSend;
continue;
}
// If we already have a batch for this host, wait until the next time
HostBatchMap::iterator pendingIt = pendingBatches.find( shardHost );
if ( pendingIt != pendingBatches.end() ) continue;
//
// We now have all the info needed to dispatch the batch
//
BatchedCommandRequest request( clientRequest.getBatchType() );
batchOp.buildBatchRequest( *nextBatch, &request );
// Internally we use full namespaces for request/response, but we send the
// command to a database with the collection name in the request.
NamespaceString nss( request.getNS() );
request.setNS( nss.coll() );
_dispatcher->addCommand( shardHost, nss.db(), request );
// Indicate we're done by setting the batch to NULL
// We'll only get duplicate hostEndpoints if we have broadcast and non-broadcast
// endpoints for the same host, so this should be pretty efficient without
// moving stuff around.
*it = NULL;
// Recv-side is responsible for cleaning up the nextBatch when used
pendingBatches.insert( make_pair( shardHost, nextBatch ) );
}
// Send them all out
_dispatcher->sendAll();
numSent += pendingBatches.size();
//
// Recv side
//
while ( _dispatcher->numPending() > 0 ) {
// Get the response
ConnectionString shardHost;
BatchedCommandResponse response;
Status dispatchStatus = _dispatcher->recvAny( &shardHost, &response );
// Get the TargetedWriteBatch to find where to put the response
dassert( pendingBatches.find( shardHost ) != pendingBatches.end() );
TargetedWriteBatch* batchRaw = pendingBatches.find( shardHost )->second;
scoped_ptr<TargetedWriteBatch> batch( batchRaw );
if ( dispatchStatus.isOK() ) {
TrackedErrors trackedErrors;
trackedErrors.startTracking( ErrorCodes::StaleShardVersion );
// Dispatch was ok, note response
batchOp.noteBatchResponse( *batch, response, &trackedErrors );
// Note if anything was stale
const vector<ShardError*>& staleErrors =
trackedErrors.getErrors( ErrorCodes::StaleShardVersion );
if ( staleErrors.size() > 0 ) {
noteStaleResponses( staleErrors, _targeter );
++_stats->numStaleBatches;
}
// Remember if the shard is actively changing metadata right now
if ( isShardMetadataChanging( staleErrors ) ) {
remoteMetadataChanging = true;
}
// Remember that we successfully wrote to this shard
// NOTE: This will record lastOps for shards where we actually didn't update
// or delete any documents, which preserves old behavior but is conservative
_stats->noteWriteAt( shardHost,
response.isLastOpSet() ?
response.getLastOp() : OpTime(),
response.isElectionIdSet() ?
response.getElectionId() : OID());
}
else {
// Error occurred dispatching, note it
WriteErrorDetail error;
buildErrorFrom( dispatchStatus, &error );
batchOp.noteBatchError( *batch, error );
}
}
}
++rounds;
++_stats->numRounds;
// If we're done, get out
if ( batchOp.isFinished() )
break;
// MORE WORK TO DO
//
// Refresh the targeter if we need to (no-op if nothing stale)
//
bool targeterChanged = false;
Status refreshStatus = _targeter->refreshIfNeeded( &targeterChanged );
if ( !refreshStatus.isOK() ) {
// It's okay if we can't refresh, we'll just record errors for the ops if
// needed.
warning() << "could not refresh targeter" << causedBy( refreshStatus.reason() )
<< endl;
}
//
// Ensure progress is being made toward completing the batch op
//
int currCompletedOps = batchOp.numWriteOpsIn( WriteOpState_Completed );
if ( currCompletedOps == numCompletedOps && !targeterChanged
&& !remoteMetadataChanging ) {
++numRoundsWithoutProgress;
}
else {
numRoundsWithoutProgress = 0;
}
numCompletedOps = currCompletedOps;
if ( numRoundsWithoutProgress > kMaxRoundsWithoutProgress ) {
stringstream msg;
msg << "no progress was made executing batch write op in " << clientRequest.getNS()
<< " after " << kMaxRoundsWithoutProgress << " rounds (" << numCompletedOps
<< " ops completed in " << rounds << " rounds total)";
WriteErrorDetail error;
buildErrorFrom( Status( ErrorCodes::NoProgressMade, msg.str() ), &error );
batchOp.setBatchError( error );
break;
}
}
batchOp.buildClientResponse( clientResponse );
}
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 ) );
}
}
}
bool batchErrorToLastError(const BatchedCommandRequest& request,
const BatchedCommandResponse& response,
LastError* error) {
unique_ptr<WriteErrorDetail> commandError;
WriteErrorDetail* lastBatchError = NULL;
if (!response.getOk()) {
// Command-level error, all writes failed
commandError.reset(new WriteErrorDetail);
buildErrorFromResponse(response, commandError.get());
lastBatchError = commandError.get();
} else if (response.isErrDetailsSet()) {
// The last error in the batch is always reported - this matches expected COE
// semantics for insert batches. For updates and deletes, error is only reported
// if the error was on the last item.
const bool lastOpErrored = response.getErrDetails().back()->getIndex() ==
static_cast<int>(request.sizeWriteOps() - 1);
if (request.getBatchType() == BatchedCommandRequest::BatchType_Insert || lastOpErrored) {
lastBatchError = response.getErrDetails().back();
}
} else {
// We don't care about write concern errors, these happen in legacy mode in GLE.
}
// Record an error if one exists
if (lastBatchError) {
string errMsg = lastBatchError->getErrMessage();
error->setLastError(lastBatchError->getErrCode(),
errMsg.empty() ? "see code for details" : errMsg.c_str());
return true;
}
// Record write stats otherwise
// NOTE: For multi-write batches, our semantics change a little because we don't have
// un-aggregated "n" stats.
if (request.getBatchType() == BatchedCommandRequest::BatchType_Update) {
BSONObj upsertedId;
if (response.isUpsertDetailsSet()) {
// Only report the very last item's upserted id if applicable
if (response.getUpsertDetails().back()->getIndex() + 1 ==
static_cast<int>(request.sizeWriteOps())) {
upsertedId = response.getUpsertDetails().back()->getUpsertedID();
}
}
int numUpserted = 0;
if (response.isUpsertDetailsSet())
numUpserted = response.sizeUpsertDetails();
int numMatched = response.getN() - numUpserted;
dassert(numMatched >= 0);
// Wrap upserted id in "upserted" field
BSONObj leUpsertedId;
if (!upsertedId.isEmpty())
leUpsertedId = upsertedId.firstElement().wrap(kUpsertedFieldName);
error->recordUpdate(numMatched > 0, response.getN(), leUpsertedId);
} else if (request.getBatchType() == BatchedCommandRequest::BatchType_Delete) {
error->recordDelete(response.getN());
}
return false;
}
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 BatchWriteExec::executeBatch( const BatchedCommandRequest& clientRequest,
BatchedCommandResponse* clientResponse ) {
BatchWriteOp batchOp;
batchOp.initClientRequest( &clientRequest );
int numTargetErrors = 0;
int numStaleBatches = 0;
for ( int rounds = 0; !batchOp.isFinished(); rounds++ ) {
//
// Refresh the targeter if we need to (no-op if nothing stale)
//
Status refreshStatus = _targeter->refreshIfNeeded();
if ( !refreshStatus.isOK() ) {
// It's okay if we can't refresh, we'll just record errors for the ops if
// needed.
warning() << "could not refresh targeter" << causedBy( refreshStatus.reason() )
<< endl;
}
//
// Get child batches to send
//
vector<TargetedWriteBatch*> childBatches;
//
// Targeting errors can be caused by remote metadata changing (the collection could have
// been dropped and recreated, for example with a new shard key). If a remote metadata
// change occurs *before* a client sends us a batch, we need to make sure that we don't
// error out just because we're staler than the client - otherwise mongos will be have
// unpredictable behavior.
//
// (If a metadata change happens *during* or *after* a client sends us a batch, however,
// we make no guarantees about delivery.)
//
// For this reason, we don't record targeting errors until we've refreshed our targeting
// metadata at least once *after* receiving the client batch - at that point, we know:
//
// 1) our new metadata is the same as the metadata when the client sent a batch, and so
// targeting errors are real.
// OR
// 2) our new metadata is a newer version than when the client sent a batch, and so
// the metadata must have changed after the client batch was sent. We don't need to
// deliver in this case, since for all the client knows we may have gotten the batch
// exactly when the metadata changed.
//
// If we've had a targeting error or stale error, we've refreshed the metadata once and
// can record target errors.
bool recordTargetErrors = numTargetErrors > 0 || numStaleBatches > 0;
Status targetStatus = batchOp.targetBatch( *_targeter,
recordTargetErrors,
&childBatches );
if ( !targetStatus.isOK() ) {
_targeter->noteCouldNotTarget();
++numTargetErrors;
continue;
}
//
// Send all child batches
//
size_t numSent = 0;
while ( numSent != childBatches.size() ) {
// Collect batches out on the network, mapped by endpoint
EndpointBatchMap pendingBatches;
//
// Send side
//
// Get as many batches as we can at once
for ( vector<TargetedWriteBatch*>::iterator it = childBatches.begin();
it != childBatches.end(); ++it ) {
TargetedWriteBatch* nextBatch = *it;
// If the batch is NULL, we sent it previously, so skip
if ( nextBatch == NULL ) continue;
const ConnectionString& hostEndpoint = nextBatch->getEndpoint().shardHost;
EndpointBatchMap::iterator pendingIt = pendingBatches.find( &hostEndpoint );
// If we already have a batch for this endpoint, continue
if ( pendingIt != pendingBatches.end() ) continue;
// Otherwise send it out to the endpoint via a command to a database
BatchedCommandRequest request( clientRequest.getBatchType() );
batchOp.buildBatchRequest( *nextBatch, &request );
// Internally we use full namespaces for request/response, but we send the
// command to a database with the collection name in the request.
NamespaceString nss( request.getNS() );
request.setNS( nss.coll() );
_dispatcher->addCommand( hostEndpoint, nss.db(), request );
// Indicate we're done by setting the batch to NULL
// We'll only get duplicate hostEndpoints if we have broadcast and non-broadcast
// endpoints for the same host, so this should be pretty efficient without
// moving stuff around.
*it = NULL;
// Recv-side is responsible for cleaning up the nextBatch when used
pendingBatches.insert( make_pair( &hostEndpoint, nextBatch ) );
}
// Send them all out
_dispatcher->sendAll();
numSent += pendingBatches.size();
//
// Recv side
//
while ( _dispatcher->numPending() > 0 ) {
// Get the response
ConnectionString endpoint;
BatchedCommandResponse response;
Status dispatchStatus = _dispatcher->recvAny( &endpoint, &response );
// Get the TargetedWriteBatch to find where to put the response
TargetedWriteBatch* batchRaw = pendingBatches.find( &endpoint )->second;
scoped_ptr<TargetedWriteBatch> batch( batchRaw );
if ( dispatchStatus.isOK() ) {
TrackedErrors trackedErrors;
trackedErrors.startTracking( ErrorCodes::StaleShardVersion );
// Dispatch was ok, note response
batchOp.noteBatchResponse( *batch, response, &trackedErrors );
// Note if anything was stale
const vector<ShardError*>& staleErrors =
trackedErrors.getErrors( ErrorCodes::StaleShardVersion );
if ( staleErrors.size() > 0 ) {
noteStaleResponses( staleErrors, _targeter );
++numStaleBatches;
}
}
else {
// Error occurred dispatching, note it
BatchedErrorDetail error;
buildErrorFrom( dispatchStatus, &error );
batchOp.noteBatchError( *batch, error );
}
}
}
}
batchOp.buildClientResponse( clientResponse );
}