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 ) );
}
bool run(OperationContext* txn, const string& dbname,
BSONObj& cmdObj,
int,
string& errmsg,
BSONObjBuilder& result,
bool fromRepl ) {
//
// Correct behavior here is very finicky.
//
// 1. The first step is to append the error that occurred on the previous operation.
// This adds an "err" field to the command, which is *not* the command failing.
//
// 2. Next we parse and validate write concern options. If these options are invalid
// the command fails no matter what, even if we actually had an error earlier. The
// reason for checking here is to match legacy behavior on these kind of failures -
// we'll still get an "err" field for the write error.
//
// 3. If we had an error on the previous operation, we then return immediately.
//
// 4. Finally, we actually enforce the write concern. All errors *except* timeout are
// reported with ok : 0.0, to match legacy behavior.
//
// There is a special case when "wOpTime" and "wElectionId" are explicitly provided by
// the client (mongos) - in this case we *only* enforce the write concern if it is
// valid.
//
// We always need to either report "err" (if ok : 1) or "errmsg" (if ok : 0), even if
// err is null.
//
LastError *le = lastError.disableForCommand();
// Always append lastOp and connectionId
Client& c = cc();
c.appendLastOp( result );
// for sharding; also useful in general for debugging
result.appendNumber( "connectionId" , c.getConnectionId() );
OpTime lastOpTime;
BSONField<OpTime> wOpTimeField("wOpTime");
FieldParser::FieldState extracted = FieldParser::extract(cmdObj, wOpTimeField,
&lastOpTime, &errmsg);
if (!extracted) {
result.append("badGLE", cmdObj);
appendCommandStatus(result, false, errmsg);
return false;
}
bool lastOpTimePresent = extracted != FieldParser::FIELD_NONE;
if (!lastOpTimePresent) {
// Use the client opTime if no wOpTime is specified
lastOpTime = cc().getLastOp();
}
OID electionId;
BSONField<OID> wElectionIdField("wElectionId");
extracted = FieldParser::extract(cmdObj, wElectionIdField,
&electionId, &errmsg);
if (!extracted) {
result.append("badGLE", cmdObj);
appendCommandStatus(result, false, errmsg);
return false;
}
bool electionIdPresent = extracted != FieldParser::FIELD_NONE;
bool errorOccurred = false;
// Errors aren't reported when wOpTime is used
if ( !lastOpTimePresent ) {
if ( le->nPrev != 1 ) {
errorOccurred = LastError::noError.appendSelf( result, false );
le->appendSelfStatus( result );
}
else {
errorOccurred = le->appendSelf( result, false );
}
}
BSONObj writeConcernDoc = cmdObj;
// Use the default options if we have no gle options aside from wOpTime/wElectionId
const int nFields = cmdObj.nFields();
bool useDefaultGLEOptions = (nFields == 1) ||
(nFields == 2 && lastOpTimePresent) ||
(nFields == 3 && lastOpTimePresent && electionIdPresent);
if ( useDefaultGLEOptions && getLastErrorDefault ) {
writeConcernDoc = *getLastErrorDefault;
}
//
// Validate write concern no matter what, this matches 2.4 behavior
//
WriteConcernOptions writeConcern;
Status status = writeConcern.parse( writeConcernDoc );
if ( status.isOK() ) {
// Ensure options are valid for this host
status = validateWriteConcern( writeConcern );
}
if ( !status.isOK() ) {
result.append( "badGLE", writeConcernDoc );
return appendCommandStatus( result, status );
}
// Don't wait for replication if there was an error reported - this matches 2.4 behavior
if ( errorOccurred ) {
dassert( !lastOpTimePresent );
return true;
}
// No error occurred, so we won't duplicate these fields with write concern errors
dassert( result.asTempObj()["err"].eoo() );
dassert( result.asTempObj()["code"].eoo() );
// If we got an electionId, make sure it matches
if (electionIdPresent) {
if (!theReplSet) {
// Ignore electionIds of 0 from mongos.
if (electionId != OID()) {
errmsg = "wElectionId passed but no replication active";
result.append("code", ErrorCodes::BadValue);
return false;
}
}
else {
if (electionId != theReplSet->getElectionId()) {
LOG(3) << "oid passed in is " << electionId
<< ", but our id is " << theReplSet->getElectionId();
errmsg = "election occurred after write";
result.append("code", ErrorCodes::WriteConcernFailed);
return false;
}
}
}
cc().curop()->setMessage( "waiting for write concern" );
WriteConcernResult wcResult;
status = waitForWriteConcern( txn, writeConcern, lastOpTime, &wcResult );
wcResult.appendTo( writeConcern, &result );
// For backward compatibility with 2.4, wtimeout returns ok : 1.0
if ( wcResult.wTimedOut ) {
dassert( !wcResult.err.empty() ); // so we always report err
dassert( !status.isOK() );
result.append( "errmsg", "timed out waiting for slaves" );
result.append( "code", status.code() );
return true;
}
return appendCommandStatus( result, status );
}
Status waitForWriteConcern( OperationContext* txn,
const WriteConcernOptions& writeConcern,
const OpTime& replOpTime,
WriteConcernResult* result ) {
// We assume all options have been validated earlier, if not, programming error
dassert( validateWriteConcern( writeConcern ).isOK() );
// Next handle blocking on disk
Timer syncTimer;
switch( writeConcern.syncMode ) {
case WriteConcernOptions::NONE:
break;
case WriteConcernOptions::FSYNC:
if ( !getDur().isDurable() ) {
result->fsyncFiles = globalStorageEngine->flushAllFiles( true );
}
else {
// We only need to commit the journal if we're durable
txn->recoveryUnit()->awaitCommit();
}
break;
case WriteConcernOptions::JOURNAL:
txn->recoveryUnit()->awaitCommit();
break;
}
result->syncMillis = syncTimer.millis();
// Now wait for replication
if (replOpTime.isNull()) {
// no write happened for this client yet
return Status::OK();
}
// needed to avoid incrementing gleWtimeStats SERVER-9005
if (writeConcern.wNumNodes <= 1 && writeConcern.wMode.empty()) {
// no desired replication check
return Status::OK();
}
// Now we wait for replication
// Note that replica set stepdowns and gle mode changes are thrown as errors
repl::ReplicationCoordinator::StatusAndDuration replStatus =
repl::getGlobalReplicationCoordinator()->awaitReplication(txn,
replOpTime,
writeConcern);
if (replStatus.status == ErrorCodes::ExceededTimeLimit) {
gleWtimeouts.increment();
replStatus.status = Status(ErrorCodes::WriteConcernFailed,
"waiting for replication timed out");
result->err = "timeout";
result->wTimedOut = true;
}
// Add stats
result->writtenTo = repl::getGlobalReplicationCoordinator()->getHostsWrittenTo(replOpTime);
gleWtimeStats.recordMillis(replStatus.duration.total_milliseconds());
result->wTime = replStatus.duration.total_milliseconds();
return replStatus.status;
}
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 ) );
}
Status waitForWriteConcern( OperationContext* txn,
const WriteConcernOptions& writeConcern,
const OpTime& replOpTime,
WriteConcernResult* result ) {
// We assume all options have been validated earlier, if not, programming error
dassert( validateWriteConcern( writeConcern ).isOK() );
// Next handle blocking on disk
Timer syncTimer;
switch( writeConcern.syncMode ) {
case WriteConcernOptions::NONE:
break;
case WriteConcernOptions::FSYNC:
if ( !getDur().isDurable() ) {
result->fsyncFiles = MemoryMappedFile::flushAll( true );
}
else {
// We only need to commit the journal if we're durable
txn->recoveryUnit()->awaitCommit();
}
break;
case WriteConcernOptions::JOURNAL:
txn->recoveryUnit()->awaitCommit();
break;
}
result->syncMillis = syncTimer.millis();
// Now wait for replication
if ( replOpTime.isNull() ) {
// no write happened for this client yet
return Status::OK();
}
if ( writeConcern.wNumNodes <= 1 && writeConcern.wMode.empty() ) {
// no desired replication check
return Status::OK();
}
if (!replset::anyReplEnabled() || serverGlobalParams.configsvr) {
// no replication check needed (validated above)
return Status::OK();
}
const bool isMasterSlaveNode = replset::anyReplEnabled() && !replset::theReplSet;
if ( writeConcern.wMode == "majority" && isMasterSlaveNode ) {
// with master/slave, majority is equivalent to w=1
return Status::OK();
}
// We're sure that replication is enabled and that we have more than one node or a wMode
TimerHolder gleTimerHolder( &gleWtimeStats );
// Now we wait for replication
// Note that replica set stepdowns and gle mode changes are thrown as errors
// TODO: Make this cleaner
Status replStatus = Status::OK();
try {
while ( 1 ) {
if ( writeConcern.wNumNodes > 0 ) {
if (replset::opReplicatedEnough(replOpTime, writeConcern.wNumNodes)) {
break;
}
}
else if (replset::opReplicatedEnough(replOpTime, writeConcern.wMode)) {
break;
}
if ( writeConcern.wTimeout > 0 &&
gleTimerHolder.millis() >= writeConcern.wTimeout ) {
gleWtimeouts.increment();
result->err = "timeout";
result->wTimedOut = true;
replStatus = Status( ErrorCodes::WriteConcernFailed,
"waiting for replication timed out" );
break;
}
sleepmillis(1);
txn->checkForInterrupt();
}
}
catch( const AssertionException& ex ) {
// Our replication state changed while enforcing write concern
replStatus = ex.toStatus();
}
// Add stats
result->writtenTo = replset::getHostsWrittenTo(replOpTime);
result->wTime = gleTimerHolder.recordMillis();
return replStatus;
}
