Status enforceLegacyWriteConcern( MultiCommandDispatch* dispatcher,
const StringData& dbName,
const BSONObj& options,
const HostOpTimeMap& hostOpTimes,
vector<LegacyWCResponse>* legacyWCResponses ) {
if ( hostOpTimes.empty() ) {
return Status::OK();
}
for ( HostOpTimeMap::const_iterator it = hostOpTimes.begin(); it != hostOpTimes.end();
++it ) {
const ConnectionString& shardEndpoint = it->first;
const HostOpTime hot = it->second;
const OpTime& opTime = hot.opTime;
const OID& electionId = hot.electionId;
LOG( 3 ) << "enforcing write concern " << options << " on " << shardEndpoint.toString()
<< " at opTime " << opTime.toStringPretty() << " with electionID "
<< electionId;
BSONObj gleCmd = buildGLECmdWithOpTime( options, opTime, electionId );
RawBSONSerializable gleCmdSerial( gleCmd );
dispatcher->addCommand( shardEndpoint, dbName, gleCmdSerial );
}
dispatcher->sendAll();
vector<Status> failedStatuses;
while ( dispatcher->numPending() > 0 ) {
ConnectionString shardEndpoint;
RawBSONSerializable gleResponseSerial;
Status dispatchStatus = dispatcher->recvAny( &shardEndpoint, &gleResponseSerial );
if ( !dispatchStatus.isOK() ) {
// We need to get all responses before returning
failedStatuses.push_back( dispatchStatus );
continue;
}
BSONObj gleResponse = stripNonWCInfo( gleResponseSerial.toBSON() );
// Use the downconversion tools to determine if this GLE response is ok, a
// write concern error, or an unknown error we should immediately abort for.
GLEErrors errors;
Status extractStatus = extractGLEErrors( gleResponse, &errors );
if ( !extractStatus.isOK() ) {
failedStatuses.push_back( extractStatus );
continue;
}
LegacyWCResponse wcResponse;
wcResponse.shardHost = shardEndpoint.toString();
wcResponse.gleResponse = gleResponse;
if ( errors.wcError.get() ) {
wcResponse.errToReport = errors.wcError->getErrMessage();
}
legacyWCResponses->push_back( wcResponse );
}
if ( failedStatuses.empty() ) {
return Status::OK();
}
StringBuilder builder;
builder << "could not enforce write concern";
for ( vector<Status>::const_iterator it = failedStatuses.begin();
it != failedStatuses.end(); ++it ) {
const Status& failedStatus = *it;
if ( it == failedStatuses.begin() ) {
builder << causedBy( failedStatus.toString() );
}
else {
builder << ":: and ::" << failedStatus.toString();
}
}
return Status( failedStatuses.size() == 1u ? failedStatuses.front().code() :
ErrorCodes::MultipleErrorsOccurred,
builder.str() );
}
void BatchSafeWriter::safeWriteBatch( DBClientBase* conn,
const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
const NamespaceString nss( request.getNS() );
// N starts at zero, and we add to it for each item
response->setN( 0 );
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
// Break on first error if we're ordered
if ( request.getOrdered() && response->isErrDetailsSet() )
break;
BatchItemRef itemRef( &request, static_cast<int>( i ) );
bool isLastItem = ( i == request.sizeWriteOps() - 1 );
BSONObj writeConcern;
if ( isLastItem && request.isWriteConcernSet() ) {
writeConcern = request.getWriteConcern();
// Pre-2.4.2 mongods react badly to 'w' being set on config servers
if ( nss.db() == "config" )
writeConcern = fixWCForConfig( writeConcern );
}
BSONObj gleResult;
GLEErrors errors;
Status status = _safeWriter->safeWrite( conn, itemRef, writeConcern, &gleResult );
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
response->clear();
response->setOk( false );
response->setErrCode( status.code() );
response->setErrMessage( status.reason() );
return;
}
//
// STATS HANDLING
//
GLEStats stats;
extractGLEStats( gleResult, &stats );
// Special case for making legacy "n" field result for insert match the write
// command result.
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert
&& !errors.writeError.get() ) {
// n is always 0 for legacy inserts.
dassert( stats.n == 0 );
stats.n = 1;
}
response->setN( response->getN() + stats.n );
if ( !stats.upsertedId.isEmpty() ) {
BatchedUpsertDetail* upsertedId = new BatchedUpsertDetail;
upsertedId->setIndex( i );
upsertedId->setUpsertedID( stats.upsertedId );
response->addToUpsertDetails( upsertedId );
}
response->setLastOp( stats.lastOp );
//
// WRITE ERROR HANDLING
//
// If any error occurs (except stale config) the previous GLE was not enforced
bool enforcedWC = !errors.writeError.get()
|| errors.writeError->getErrCode() == ErrorCodes::StaleShardVersion;
// Save write error
if ( errors.writeError.get() ) {
errors.writeError->setIndex( i );
response->addToErrDetails( errors.writeError.release() );
}
//
// WRITE CONCERN ERROR HANDLING
//
// The last write is weird, since we enforce write concern and check the error through
// the same GLE if possible. If the last GLE was an error, the write concern may not
// have been enforced in that same GLE, so we need to send another after resetting the
// error.
if ( isLastItem ) {
// Try to enforce the write concern if everything succeeded (unordered or ordered)
// OR if something succeeded and we're unordered.
bool needToEnforceWC =
!response->isErrDetailsSet()
|| ( !request.getOrdered()
&& response->sizeErrDetails() < request.sizeWriteOps() );
if ( !enforcedWC && needToEnforceWC ) {
dassert( !errors.writeError.get() ); // emptied above
// Might have gotten a write concern validity error earlier, these are
// enforced even if the wc isn't applied, so we ignore.
errors.wcError.reset();
Status status = _safeWriter->enforceWriteConcern( conn,
nss.db().toString(),
writeConcern,
&gleResult );
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
response->clear();
response->setOk( false );
response->setErrCode( status.code() );
response->setErrMessage( status.reason() );
return;
}
}
// END Write concern retry
if ( errors.wcError.get() ) {
response->setWriteConcernError( errors.wcError.release() );
}
}
}
response->setOk( true );
dassert( response->isValid( NULL ) );
}
void BatchSafeWriter::safeWriteBatch( DBClientBase* conn,
const BatchedCommandRequest& request,
BatchedCommandResponse* response ) {
const NamespaceString nss( request.getNS() );
// N starts at zero, and we add to it for each item
response->setN( 0 );
// GLE path always sets nModified to -1 (sentinel) to indicate we should omit it later.
response->setNModified(-1);
for ( size_t i = 0; i < request.sizeWriteOps(); ++i ) {
// Break on first error if we're ordered
if ( request.getOrdered() && response->isErrDetailsSet() )
break;
BatchItemRef itemRef( &request, static_cast<int>( i ) );
BSONObj gleResult;
GLEErrors errors;
Status status = _safeWriter->safeWrite( conn,
itemRef,
WriteConcernOptions::Acknowledged,
&gleResult );
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
response->clear();
response->setOk( false );
response->setErrCode( ErrorCodes::RemoteResultsUnavailable );
StringBuilder builder;
builder << "could not get write error from safe write";
builder << causedBy( status.toString() );
response->setErrMessage( builder.str() );
return;
}
if ( errors.wcError.get() ) {
response->setWriteConcernError( errors.wcError.release() );
}
//
// STATS HANDLING
//
GLEStats stats;
extractGLEStats( gleResult, &stats );
// Special case for making legacy "n" field result for insert match the write
// command result.
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert
&& !errors.writeError.get() ) {
// n is always 0 for legacy inserts.
dassert( stats.n == 0 );
stats.n = 1;
}
response->setN( response->getN() + stats.n );
if ( !stats.upsertedId.isEmpty() ) {
BatchedUpsertDetail* upsertedId = new BatchedUpsertDetail;
upsertedId->setIndex( i );
upsertedId->setUpsertedID( stats.upsertedId );
response->addToUpsertDetails( upsertedId );
}
response->setLastOp( stats.lastOp );
// Save write error
if ( errors.writeError.get() ) {
errors.writeError->setIndex( i );
response->addToErrDetails( errors.writeError.release() );
}
}
//
// WRITE CONCERN ERROR HANDLING
//
// The last write is weird, since we enforce write concern and check the error through
// the same GLE if possible. If the last GLE was an error, the write concern may not
// have been enforced in that same GLE, so we need to send another after resetting the
// error.
BSONObj writeConcern;
if ( request.isWriteConcernSet() ) {
writeConcern = request.getWriteConcern();
// Pre-2.4.2 mongods react badly to 'w' being set on config servers
if ( nss.db() == "config" )
writeConcern = fixWCForConfig( writeConcern );
}
bool needToEnforceWC = WriteConcernOptions::Acknowledged.woCompare(writeConcern) != 0 &&
WriteConcernOptions::Unacknowledged.woCompare(writeConcern) != 0;
if ( needToEnforceWC &&
( !response->isErrDetailsSet() ||
( !request.getOrdered() &&
// Not all errored. Note: implicit response->isErrDetailsSet().
response->sizeErrDetails() < request.sizeWriteOps() ))) {
// Might have gotten a write concern validity error earlier, these are
// enforced even if the wc isn't applied, so we ignore.
response->unsetWriteConcernError();
const string dbName( nss.db().toString() );
Status status( Status::OK() );
if ( response->isErrDetailsSet() ) {
const WriteErrorDetail* lastError = response->getErrDetails().back();
// If last write op was an error.
if ( lastError->getIndex() == static_cast<int>( request.sizeWriteOps() - 1 )) {
// Reset previous errors so we can apply the write concern no matter what
// as long as it is valid.
status = _safeWriter->clearErrors( conn, dbName );
}
}
BSONObj gleResult;
if ( status.isOK() ) {
status = _safeWriter->enforceWriteConcern( conn,
dbName,
writeConcern,
&gleResult );
}
GLEErrors errors;
if ( status.isOK() ) {
status = extractGLEErrors( gleResult, &errors );
}
if ( !status.isOK() ) {
auto_ptr<WCErrorDetail> wcError( new WCErrorDetail );
wcError->setErrCode( status.code() );
wcError->setErrMessage( status.reason() );
response->setWriteConcernError( wcError.release() );
}
else if ( errors.wcError.get() ) {
response->setWriteConcernError( errors.wcError.release() );
}
}
response->setOk( true );
dassert( response->isValid( NULL ) );
}
