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 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 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 );
}
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 ) {
// 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 ) );
}
/**
* The core config write functionality.
*
* Config writes run in two passes - the first is a quick check to ensure the config servers
* are all reachable, the second runs the actual write.
*
* TODO: Upgrade and move this logic to the config servers, a state machine implementation
* is probably the next step.
*/
void ConfigCoordinator::executeBatch( const BatchedCommandRequest& clientRequest,
BatchedCommandResponse* clientResponse,
bool fsyncCheck ) {
NamespaceString nss( clientRequest.getNS() );
dassert( nss.db() == "config" || nss.db() == "admin" );
dassert( clientRequest.sizeWriteOps() == 1u );
if ( fsyncCheck ) {
//
// Sanity check that all configs are still reachable using fsync, preserving legacy
// behavior
//
OwnedPointerVector<ConfigFsyncResponse> fsyncResponsesOwned;
vector<ConfigFsyncResponse*>& fsyncResponses = fsyncResponsesOwned.mutableVector();
//
// Send side
//
for ( vector<ConnectionString>::iterator it = _configHosts.begin();
it != _configHosts.end(); ++it ) {
ConnectionString& configHost = *it;
FsyncRequest fsyncRequest;
_dispatcher->addCommand( configHost, "admin", fsyncRequest );
}
_dispatcher->sendAll();
//
// Recv side
//
bool fsyncError = false;
while ( _dispatcher->numPending() > 0 ) {
fsyncResponses.push_back( new ConfigFsyncResponse() );
ConfigFsyncResponse& fsyncResponse = *fsyncResponses.back();
Status dispatchStatus = _dispatcher->recvAny( &fsyncResponse.configHost,
&fsyncResponse.response );
// We've got to recv everything, no matter what
if ( !dispatchStatus.isOK() ) {
fsyncError = true;
buildFsyncErrorFrom( dispatchStatus, &fsyncResponse.response );
}
else if ( !fsyncResponse.response.getOk() ) {
fsyncError = true;
}
}
if ( fsyncError ) {
combineFsyncErrors( fsyncResponses, clientResponse );
return;
}
else {
fsyncResponsesOwned.clear();
}
}
//
// Do the actual writes
//
BatchedCommandRequest configRequest( clientRequest.getBatchType() );
clientRequest.cloneTo( &configRequest );
configRequest.setNS( nss.coll() );
OwnedPointerVector<ConfigResponse> responsesOwned;
vector<ConfigResponse*>& responses = responsesOwned.mutableVector();
//
// Send the actual config writes
//
// Get as many batches as we can at once
for ( vector<ConnectionString>::iterator it = _configHosts.begin();
it != _configHosts.end(); ++it ) {
ConnectionString& configHost = *it;
_dispatcher->addCommand( configHost, nss.db(), configRequest );
}
// Send them all out
_dispatcher->sendAll();
//
// Recv side
//
while ( _dispatcher->numPending() > 0 ) {
// Get the response
responses.push_back( new ConfigResponse() );
ConfigResponse& configResponse = *responses.back();
Status dispatchStatus = _dispatcher->recvAny( &configResponse.configHost,
&configResponse.response );
if ( !dispatchStatus.isOK() ) {
buildErrorFrom( dispatchStatus, &configResponse.response );
}
}
combineResponses( responses, clientResponse );
}
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 batchErrorToLastError( const BatchedCommandRequest& request,
const BatchedCommandResponse& response,
LastError* error ) {
scoped_ptr<BatchedErrorDetail> topLevelError;
BatchedErrorDetail* lastBatchError = NULL;
if ( !response.getOk() ) {
int code = response.getErrCode();
// Check for batch error
// We don't care about write concern errors, these happen in legacy mode in GLE
if ( code != ErrorCodes::WriteConcernFailed && !response.isErrDetailsSet() ) {
// Top-level error, all writes failed
topLevelError.reset( new BatchedErrorDetail );
buildErrorFromResponse( response, topLevelError.get() );
lastBatchError = topLevelError.get();
}
else if ( response.isErrDetailsSet() ) {
// The last error in the batch is always reported - this matches expected COE
// semantics for insert batches and works for single writes
lastBatchError = response.getErrDetails().back();
}
}
// Record an error if one exists
if ( lastBatchError ) {
error->raiseError( lastBatchError->getErrCode(),
lastBatchError->getErrMessage().c_str() );
return;
}
// 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.isSingleUpsertedSet() ) upsertedId = response.getSingleUpserted();
else 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.isSingleUpsertedSet() )
++numUpserted;
else if ( response.isUpsertDetailsSet() )
numUpserted += response.sizeUpsertDetails();
int numUpdated = response.getN() - numUpserted;
dassert( numUpdated >= 0 );
error->recordUpdate( numUpdated > 0, response.getN(), upsertedId );
}
else if ( request.getBatchType() == BatchedCommandRequest::BatchType_Delete ) {
error->recordDelete( response.getN() );
}
}
/**
* The core config write functionality.
*
* Config writes run in two passes - the first is a quick check to ensure the config servers
* are all reachable, the second runs the actual write.
*
* TODO: Upgrade and move this logic to the config servers, a state machine implementation
* is probably the next step.
*/
void ConfigCoordinator::executeBatch(const BatchedCommandRequest& clientRequest,
BatchedCommandResponse* clientResponse) {
const NamespaceString nss(clientRequest.getNS());
// Should never use it for anything other than DBs residing on the config server
dassert(nss.db() == "config" || nss.db() == "admin");
dassert(clientRequest.sizeWriteOps() == 1u);
// This is an opportunistic check that all config servers look healthy by calling
// getLastError on each one of them. If there was some form of write/journaling error, get
// last error would fail.
{
for (vector<ConnectionString>::iterator it = _configHosts.begin();
it != _configHosts.end();
++it) {
_dispatcher->addCommand(*it,
"admin",
RawBSONSerializable(BSON("getLastError" << true <<
"fsync" << true)));
}
_dispatcher->sendAll();
bool error = false;
while (_dispatcher->numPending()) {
ConnectionString host;
RawBSONSerializable response;
Status status = _dispatcher->recvAny(&host, &response);
if (status.isOK()) {
BSONObj obj = response.toBSON();
LOG(3) << "Response " << obj.toString();
// If the ok field is anything other than 1, count it as error
if (!obj["ok"].trueValue()) {
error = true;
log() << "Config server check for host " << host
<< " returned error: " << response;
}
}
else {
error = true;
log() << "Config server check for host " << host
<< " failed with status: " << status;
}
}
// All responses should have been gathered by this point
if (error) {
clientResponse->setOk(false);
clientResponse->setErrCode(ErrorCodes::RemoteValidationError);
clientResponse->setErrMessage("Could not verify that config servers were active"
" and reachable before write");
return;
}
}
if (!_checkConfigString(clientResponse)) {
return;
}
//
// Do the actual writes
//
BatchedCommandRequest configRequest( clientRequest.getBatchType() );
clientRequest.cloneTo( &configRequest );
configRequest.setNS( nss.coll() );
OwnedPointerVector<ConfigResponse> responsesOwned;
vector<ConfigResponse*>& responses = responsesOwned.mutableVector();
//
// Send the actual config writes
//
// Get as many batches as we can at once
for (vector<ConnectionString>::const_iterator it = _configHosts.begin();
it != _configHosts.end();
++it) {
const ConnectionString& configHost = *it;
_dispatcher->addCommand(configHost, nss.db(), configRequest);
}
// Send them all out
_dispatcher->sendAll();
//
// Recv side
//
while (_dispatcher->numPending() > 0) {
// Get the response
responses.push_back(new ConfigResponse());
ConfigResponse& configResponse = *responses.back();
Status dispatchStatus = _dispatcher->recvAny(&configResponse.configHost,
&configResponse.response);
if (!dispatchStatus.isOK()) {
buildErrorFrom(dispatchStatus, &configResponse.response);
}
}
combineResponses(responses, clientResponse);
}
bool WriteBatchExecutor::applyWriteItem( const BatchedCommandRequest& request,
int index,
WriteStats* stats,
BatchedErrorDetail* error ) {
const string& ns = request.getNS();
// Clear operation's LastError before starting.
_le->reset( true );
//uint64_t itemTimeMicros = 0;
bool opSuccess = true;
// Each write operation executes in its own PageFaultRetryableSection. This means that
// a single batch can throw multiple PageFaultException's, which is not the case for
// other operations.
PageFaultRetryableSection s;
while ( true ) {
try {
// Execute the write item as a child operation of the current operation.
CurOp childOp( _client, _client->curop() );
// TODO Modify CurOp "wrapped" constructor to take an opcode, so calling .reset()
// is unneeded
childOp.reset( _client->getRemote(), getOpCode( request.getBatchType() ) );
childOp.ensureStarted();
OpDebug& opDebug = childOp.debug();
opDebug.ns = ns;
{
Client::WriteContext ctx( ns );
switch ( request.getBatchType() ) {
case BatchedCommandRequest::BatchType_Insert:
opSuccess =
applyInsert( ns,
request.getInsertRequest()->getDocumentsAt( index ),
&childOp,
stats,
error );
break;
case BatchedCommandRequest::BatchType_Update:
opSuccess = applyUpdate( ns,
*request.getUpdateRequest()->getUpdatesAt( index ),
&childOp,
stats,
error );
break;
default:
dassert( request.getBatchType() ==
BatchedCommandRequest::BatchType_Delete );
opSuccess = applyDelete( ns,
*request.getDeleteRequest()->getDeletesAt( index ),
&childOp,
stats,
error );
break;
}
}
childOp.done();
//itemTimeMicros = childOp.totalTimeMicros();
opDebug.executionTime = childOp.totalTimeMillis();
opDebug.recordStats();
// Log operation if running with at least "-v", or if exceeds slow threshold.
if ( logger::globalLogDomain()->shouldLog( logger::LogSeverity::Debug( 1 ) )
|| opDebug.executionTime > cmdLine.slowMS + childOp.getExpectedLatencyMs() ) {
MONGO_TLOG(1) << opDebug.report( childOp ) << endl;
}
// TODO Log operation if logLevel >= 3 and assertion thrown (as assembleResponse()
// does).
// Save operation to system.profile if shouldDBProfile().
if ( childOp.shouldDBProfile( opDebug.executionTime ) ) {
profile( *_client, getOpCode( request.getBatchType() ), childOp );
}
break;
}
catch ( PageFaultException& e ) {
e.touch();
}
}
return opSuccess;
}
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 ) );
}
