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 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 ClusterWriter::write(OperationContext* opCtx,
const BatchedCommandRequest& origRequest,
BatchedCommandResponse* response) {
// Add _ids to insert request if req'd
std::unique_ptr<BatchedCommandRequest> idRequest(
BatchedCommandRequest::cloneWithIds(origRequest));
const BatchedCommandRequest* request = idRequest ? idRequest.get() : &origRequest;
const NamespaceString& nss = request->getNS();
if (!nss.isValid()) {
toBatchError(Status(ErrorCodes::InvalidNamespace, nss.ns() + " is not a valid namespace"),
response);
return;
}
if (!NamespaceString::validCollectionName(nss.coll())) {
toBatchError(
Status(ErrorCodes::BadValue, str::stream() << "invalid collection name " << nss.coll()),
response);
return;
}
if (request->sizeWriteOps() == 0u) {
toBatchError(Status(ErrorCodes::InvalidLength, "no write ops were included in the batch"),
response);
return;
}
if (request->sizeWriteOps() > BatchedCommandRequest::kMaxWriteBatchSize) {
toBatchError(Status(ErrorCodes::InvalidLength,
str::stream() << "exceeded maximum write batch size of "
<< BatchedCommandRequest::kMaxWriteBatchSize),
response);
return;
}
std::string errMsg;
if (request->isInsertIndexRequest() && !request->isValidIndexRequest(&errMsg)) {
toBatchError(Status(ErrorCodes::InvalidOptions, errMsg), response);
return;
}
// 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->getTargetingNSS(), &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->getTargetingNSS().toString()
<< causedBy(targetInitStatus)},
response);
return;
}
BatchWriteExec::executeBatch(opCtx, targeter, *request, response, &_stats);
}
if (_autoSplit) {
splitIfNeeded(opCtx, request->getNS(), targeterStats);
}
}
}
