void DBClientSafeWriter::safeWrite( DBClientBase* conn,
const BatchItemRef& itemRef,
LastError* error ) {
const BatchedCommandRequest* request = itemRef.getRequest();
try {
// Default settings for checkShardVersion
const bool authoritative = false;
const int tryNum = 1;
// We need to set our version using setShardVersion, managed by checkShardVersionCB
versionManager.checkShardVersionCB( conn,
request->getTargetingNS(),
authoritative,
tryNum );
if ( request->getBatchType() == BatchedCommandRequest::BatchType_Insert ) {
conn->insert( request->getNS(),
request->getInsertRequest()->getDocumentsAt( itemRef.getItemIndex() ),
0 );
}
else if ( request->getBatchType() == BatchedCommandRequest::BatchType_Update ) {
const BatchedUpdateDocument* update =
request->getUpdateRequest()->getUpdatesAt( itemRef.getItemIndex() );
conn->update( request->getNS(),
update->getQuery(),
update->getUpdateExpr(),
update->getUpsert(),
update->getMulti() );
}
else {
dassert( request->getBatchType() == BatchedCommandRequest::BatchType_Delete );
const BatchedDeleteDocument* deleteDoc =
request->getDeleteRequest()->getDeletesAt( itemRef.getItemIndex() );
conn->remove( request->getNS(),
deleteDoc->getQuery(),
deleteDoc->getLimit() == 1 /*just one*/);
}
// Default GLE Options
const bool fsync = false;
const bool j = false;
const int w = 1;
const int wtimeout = 0;
BSONObj result = conn->getLastErrorDetailed( NamespaceString( request->getNS() ).db()
.toString(),
fsync,
j,
w,
wtimeout );
SafeWriter::fillLastError( result, error );
}
catch ( const DBException& ex ) {
error->raiseError( ex.getCode(), ex.toString().c_str() );
}
}
Status DBClientSafeWriter::safeWrite( DBClientBase* conn,
const BatchItemRef& itemRef,
const BSONObj& writeConcern,
BSONObj* gleResponse ) {
const BatchedCommandRequest* request = itemRef.getRequest();
try {
// Default settings for checkShardVersion
const bool authoritative = false;
const int tryNum = 1;
// We need to set our version using setShardVersion, managed by checkShardVersionCB
versionManager.checkShardVersionCB( conn,
request->getTargetingNS(),
authoritative,
tryNum );
if ( request->getBatchType() == BatchedCommandRequest::BatchType_Insert ) {
conn->insert( request->getNS(),
request->getInsertRequest()->getDocumentsAt( itemRef.getItemIndex() ),
0 );
}
else if ( request->getBatchType() == BatchedCommandRequest::BatchType_Update ) {
const BatchedUpdateDocument* update =
request->getUpdateRequest()->getUpdatesAt( itemRef.getItemIndex() );
conn->update( request->getNS(),
update->getQuery(),
update->getUpdateExpr(),
update->getUpsert(),
update->getMulti() );
}
else {
dassert( request->getBatchType() == BatchedCommandRequest::BatchType_Delete );
const BatchedDeleteDocument* deleteDoc =
request->getDeleteRequest()->getDeletesAt( itemRef.getItemIndex() );
conn->remove( request->getNS(),
deleteDoc->getQuery(),
deleteDoc->getLimit() == 1 /*just one*/);
}
const StringData& dbName = NamespaceString( request->getNS() ).db();
BSONObjBuilder gleCmdB;
gleCmdB.append( "getLastError", true );
gleCmdB.appendElements( writeConcern );
conn->runCommand( dbName.toString(), gleCmdB.obj(), *gleResponse );
}
catch ( const DBException& ex ) {
return ex.toStatus();
}
return Status::OK();
}
void WriteBatchExecutor::incOpStats( const BatchItemRef& currWrite ) {
if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Insert ) {
// No-op, for inserts we increment not on the op but once for each write
}
else if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Update ) {
_opCounters->gotUpdate();
}
else {
dassert( currWrite.getOpType() == BatchedCommandRequest::BatchType_Delete );
_opCounters->gotDelete();
}
}
void WriteBatchExecutor::incOpStats( const BatchItemRef& currWrite ) {
if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Insert ) {
_opCounters->gotInsert();
}
else if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Update ) {
_opCounters->gotUpdate();
}
else {
dassert( currWrite.getOpType() == BatchedCommandRequest::BatchType_Delete );
_opCounters->gotDelete();
}
}
/**
* Perform a remove operation, which might remove multiple documents. Dispatches to remove code
* currently to do most of this.
*
* Might fault or error, otherwise populates the result.
*/
static void multiRemove( const BatchItemRef& removeItem,
WriteOpResult* result ) {
const NamespaceString nss( removeItem.getRequest()->getNS() );
DeleteRequest request( nss );
request.setQuery( removeItem.getDelete()->getQuery() );
request.setMulti( removeItem.getDelete()->getLimit() != 1 );
request.setUpdateOpLog(true);
request.setGod( false );
DeleteExecutor executor( &request );
Status status = executor.prepare();
if ( !status.isOK() ) {
result->error = toWriteError( status );
return;
}
// NOTE: Deletes will not fault outside the lock once any data has been written
PageFaultRetryableSection pFaultSection;
///////////////////////////////////////////
Lock::DBWrite writeLock( nss.ns() );
///////////////////////////////////////////
// Check version once we're locked
if ( !checkShardVersion( &shardingState, *removeItem.getRequest(), &result->error ) ) {
// Version error
return;
}
// Context once we're locked, to set more details in currentOp()
// TODO: better constructor?
Client::Context writeContext( nss.ns(),
storageGlobalParams.dbpath,
false /* don't check version */);
try {
result->stats.n = executor.execute();
}
catch ( const PageFaultException& ex ) {
// TODO: An actual data structure that's not an exception for this
result->fault = new PageFaultException( ex );
}
catch ( const DBException& ex ) {
status = ex.toStatus();
if (ErrorCodes::isInterruption(status.code())) {
throw;
}
result->error = toWriteError(status);
}
}
void WriteBatchExecutor::execUpdate( const BatchItemRef& updateItem,
BSONObj* upsertedId,
WriteErrorDetail** error ) {
// Updates currently do a lot of the lock management internally
const BatchedCommandRequest& request = *updateItem.getRequest();
const NamespaceString nss( updateItem.getRequest()->getNS() );
// BEGIN CURRENT OP
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, updateItem ) );
incOpStats( updateItem );
WriteOpResult result;
{
///////////////////////////////////////////
Lock::DBWrite writeLock( nss.ns() );
///////////////////////////////////////////
// Check version once we're locked
if ( checkShardVersion( &shardingState, request, &result.error ) ) {
// Context once we're locked, to set more details in currentOp()
// TODO: better constructor?
Client::Context writeContext( nss.ns(),
storageGlobalParams.dbpath,
false /* don't check version */);
multiUpdate( updateItem, &result );
incWriteStats( updateItem, result.stats, result.error, currentOp.get() );
if ( !result.stats.upsertedID.isEmpty() ) {
*upsertedId = result.stats.upsertedID.getOwned();
}
}
}
// END CURRENT OP
finishCurrentOp( _client, currentOp.get(), result.error );
if ( result.error ) {
result.error->setIndex( updateItem.getItemIndex() );
*error = result.releaseError();
}
}
void WriteBatchExecutor::execRemove( const BatchItemRef& removeItem,
WriteErrorDetail** error ) {
// Removes are similar to updates, but page faults are handled externally
// BEGIN CURRENT OP
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, removeItem ) );
incOpStats( removeItem );
WriteOpResult result;
// NOTE: Deletes will not fault outside the lock once any data has been written
PageFaultRetryableSection pageFaultSection;
while ( true ) {
try {
multiRemove( removeItem, &result );
break;
}
catch (PageFaultException& pfe) {
pfe.touch();
invariant(!result.getError());
continue;
}
fassertFailed(17429);
}
// END CURRENT OP
incWriteStats( removeItem, result.getStats(), result.getError(), currentOp.get() );
finishCurrentOp( _client, currentOp.get(), result.getError() );
if ( result.getError() ) {
result.getError()->setIndex( removeItem.getItemIndex() );
*error = result.releaseError();
}
}
/**
* Perform a single index insert into a collection. Requires the index descriptor be
* preprocessed and the collection already has been created.
*
* Might fault or error, otherwise populates the result.
*/
static void singleCreateIndex( const BatchItemRef& insertItem,
const BSONObj& normalIndexDesc,
Collection* collection,
WriteOpResult* result ) {
const string& indexNS = insertItem.getRequest()->getNS();
Lock::assertWriteLocked( indexNS );
try {
Status status = collection->getIndexCatalog()->createIndex( normalIndexDesc, true );
if ( status.code() == ErrorCodes::IndexAlreadyExists ) {
result->stats.n = 0;
}
else if ( !status.isOK() ) {
result->error = toWriteError( status );
}
else {
logOp( "i", indexNS.c_str(), normalIndexDesc );
result->stats.n = 1;
}
}
catch ( const PageFaultException& ex ) {
// TODO: An actual data structure that's not an exception for this
result->fault = new PageFaultException( ex );
}
catch ( const DBException& ex ) {
result->error = toWriteError( ex.toStatus() );
}
}
void WriteBatchExecutor::execUpdate( const BatchItemRef& updateItem,
BSONObj* upsertedId,
WriteErrorDetail** error ) {
// BEGIN CURRENT OP
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, updateItem ) );
incOpStats( updateItem );
WriteOpResult result;
WriteUnitOfWork wunit(_txn->recoveryUnit());
multiUpdate( _txn, updateItem, &result );
wunit.commit();
if ( !result.getStats().upsertedID.isEmpty() ) {
*upsertedId = result.getStats().upsertedID;
}
// END CURRENT OP
incWriteStats( updateItem, result.getStats(), result.getError(), currentOp.get() );
finishCurrentOp( _txn, _client, currentOp.get(), result.getError() );
if ( result.getError() ) {
result.getError()->setIndex( updateItem.getItemIndex() );
*error = result.releaseError();
}
}
/**
* Perform a single insert into a collection. Requires the insert be preprocessed and the
* collection already has been created.
*
* Might fault or error, otherwise populates the result.
*/
static void singleInsert( const BatchItemRef& insertItem,
const BSONObj& normalInsert,
Collection* collection,
WriteOpResult* result ) {
const string& insertNS = insertItem.getRequest()->getNS();
Lock::assertWriteLocked( insertNS );
try {
// XXX - are we 100% sure that all !OK statuses do not write a document?
StatusWith<DiskLoc> status = collection->insertDocument( normalInsert, true );
if ( !status.isOK() ) {
result->error = toWriteError( status.getStatus() );
}
else {
logOp( "i", insertNS.c_str(), normalInsert );
getDur().commitIfNeeded();
result->stats.n = 1;
}
}
catch ( const PageFaultException& ex ) {
// TODO: An actual data structure that's not an exception for this
result->fault = new PageFaultException( ex );
}
catch ( const DBException& ex ) {
result->error = toWriteError( ex.toStatus() );
}
}
// Does preprocessing of inserts, special casing for indexes
// TODO: Simplify this when indexes aren't here anymore
static StatusWith<BSONObj> normalizeInsert( const BatchItemRef& insertItem ) {
if ( insertItem.getRequest()->isInsertIndexRequest() ) {
StatusWith<BSONObj> normalInsert = fixDocumentForInsert( insertItem.getDocument() );
if ( normalInsert.isOK() && insertItem.getDocument()["ns"].type() != String ) {
return StatusWith<BSONObj>( ErrorCodes::BadValue, "tried to create an index "
"without specifying namespace" );
}
else {
return normalInsert;
}
}
else {
return fixDocumentForInsert( insertItem.getDocument() );
}
}
/**
* Perform a remove operation, which might remove multiple documents. Dispatches to remove code
* currently to do most of this.
*
* Might fault or error, otherwise populates the result.
*/
static void multiRemove( OperationContext* txn,
const BatchItemRef& removeItem,
WriteOpResult* result ) {
const NamespaceString nss( removeItem.getRequest()->getNS() );
DeleteRequest request( nss );
request.setQuery( removeItem.getDelete()->getQuery() );
request.setMulti( removeItem.getDelete()->getLimit() != 1 );
request.setUpdateOpLog(true);
request.setGod( false );
DeleteExecutor executor( &request );
Status status = executor.prepare();
if ( !status.isOK() ) {
result->setError(toWriteError(status));
return;
}
///////////////////////////////////////////
Lock::DBWrite writeLock(txn->lockState(), nss.ns());
///////////////////////////////////////////
// Check version once we're locked
if (!checkShardVersion(txn, &shardingState, *removeItem.getRequest(), result)) {
// Version error
return;
}
// Context once we're locked, to set more details in currentOp()
// TODO: better constructor?
Client::Context writeContext( nss.ns(),
storageGlobalParams.dbpath,
false /* don't check version */);
try {
result->getStats().n = executor.execute(txn, writeContext.db());
}
catch ( const DBException& ex ) {
status = ex.toStatus();
if (ErrorCodes::isInterruption(status.code())) {
throw;
}
result->setError(toWriteError(status));
}
}
void WriteBatchExecutor::incWriteStats( const BatchItemRef& currWrite,
const WriteOpStats& stats,
const WriteErrorDetail* error,
CurOp* currentOp ) {
if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Insert ) {
// We increment batch inserts like individual inserts
_opCounters->gotInsert();
_stats->numInserted += stats.n;
_le->nObjects = stats.n;
currentOp->debug().ninserted += stats.n;
}
else if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Update ) {
if ( stats.upsertedID.isEmpty() ) {
_stats->numMatched += stats.n;
_stats->numModified += stats.nModified;
}
else {
++_stats->numUpserted;
}
if ( !error ) {
_le->recordUpdate( stats.upsertedID.isEmpty() && stats.n > 0,
stats.n,
stats.upsertedID );
}
}
else {
dassert( currWrite.getOpType() == BatchedCommandRequest::BatchType_Delete );
_stats->numDeleted += stats.n;
if ( !error ) {
_le->recordDelete( stats.n );
}
currentOp->debug().ndeleted += stats.n;
}
// Errors reported in LastError are handled internally in write ops for now
// TODO: Move error reporting out of write op internals?
}
void WriteBatchExecutor::incWriteStats( const BatchItemRef& currWrite,
const WriteOpStats& stats,
const WriteErrorDetail* error,
CurOp* currentOp ) {
if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Insert ) {
_stats->numInserted += stats.n;
_le->nObjects = stats.n;
currentOp->debug().ninserted += stats.n;
}
else if ( currWrite.getOpType() == BatchedCommandRequest::BatchType_Update ) {
if ( stats.upsertedID.isEmpty() ) {
_stats->numMatched += stats.n;
_stats->numModified += stats.nModified;
}
else {
++_stats->numUpserted;
}
if ( !error ) {
_le->recordUpdate( stats.upsertedID.isEmpty() && stats.n > 0,
stats.n,
stats.upsertedID );
}
}
else {
dassert( currWrite.getOpType() == BatchedCommandRequest::BatchType_Delete );
_stats->numDeleted += stats.n;
if ( !error ) {
_le->recordDelete( stats.n );
}
currentOp->debug().ndeleted += stats.n;
}
if (error && !_le->disabled) {
_le->raiseError(error->getErrCode(), error->getErrMessage().c_str());
}
}
/**
* Perform a remove operation, which might remove multiple documents. Dispatches to remove code
* currently to do most of this.
*
* Might fault or error, otherwise populates the result.
*/
static void multiRemove( const BatchItemRef& removeItem,
WriteOpResult* result ) {
Lock::assertWriteLocked( removeItem.getRequest()->getNS() );
try {
long long n = deleteObjects( removeItem.getRequest()->getNS(),
removeItem.getDelete()->getQuery(),
removeItem.getDelete()->getLimit() == 1, // justOne
true, // logOp
false // god
);
result->stats.n = n;
}
catch ( const PageFaultException& ex ) {
// TODO: An actual data structure that's not an exception for this
result->fault = new PageFaultException( ex );
}
catch ( const DBException& ex ) {
result->error = toWriteError( ex.toStatus() );
}
}
static void multiUpdate( OperationContext* txn,
const BatchItemRef& updateItem,
WriteOpResult* result ) {
const NamespaceString nsString(updateItem.getRequest()->getNS());
UpdateRequest request(nsString);
request.setQuery(updateItem.getUpdate()->getQuery());
request.setUpdates(updateItem.getUpdate()->getUpdateExpr());
request.setMulti(updateItem.getUpdate()->getMulti());
request.setUpsert(updateItem.getUpdate()->getUpsert());
request.setUpdateOpLog(true);
UpdateLifecycleImpl updateLifecycle(true, request.getNamespaceString());
request.setLifecycle(&updateLifecycle);
UpdateExecutor executor(&request, &txn->getCurOp()->debug());
Status status = executor.prepare();
if (!status.isOK()) {
result->setError(toWriteError(status));
return;
}
///////////////////////////////////////////
Lock::DBWrite writeLock(txn->lockState(), nsString.ns());
///////////////////////////////////////////
if (!checkShardVersion(txn, &shardingState, *updateItem.getRequest(), result))
return;
Client::Context ctx( nsString.ns(),
storageGlobalParams.dbpath,
false /* don't check version */ );
try {
UpdateResult res = executor.execute(txn, ctx.db());
const long long numDocsModified = res.numDocsModified;
const long long numMatched = res.numMatched;
const BSONObj resUpsertedID = res.upserted;
// We have an _id from an insert
const bool didInsert = !resUpsertedID.isEmpty();
result->getStats().nModified = didInsert ? 0 : numDocsModified;
result->getStats().n = didInsert ? 1 : numMatched;
result->getStats().upsertedID = resUpsertedID;
}
catch (const DBException& ex) {
status = ex.toStatus();
if (ErrorCodes::isInterruption(status.code())) {
throw;
}
result->setError(toWriteError(status));
}
}
/**
* Perform an update operation, which might update multiple documents in the lock. Dispatches
* to update code currently to do most of this.
*
* Might error, otherwise populates the result.
*/
static void multiUpdate( const BatchItemRef& updateItem,
WriteOpResult* result ) {
Lock::assertWriteLocked( updateItem.getRequest()->getNS() );
BSONObj queryObj = updateItem.getUpdate()->getQuery();
BSONObj updateObj = updateItem.getUpdate()->getUpdateExpr();
bool multi = updateItem.getUpdate()->getMulti();
bool upsert = updateItem.getUpdate()->getUpsert();
bool didInsert = false;
long long numMatched = 0;
long long numDocsModified = 0;
BSONObj resUpsertedID;
try {
const NamespaceString requestNs( updateItem.getRequest()->getNS() );
UpdateRequest request( requestNs );
request.setQuery( queryObj );
request.setUpdates( updateObj );
request.setUpsert( upsert );
request.setMulti( multi );
request.setUpdateOpLog();
// TODO(greg) We need to send if we are ignoring the shard version below,
// but for now yes
UpdateLifecycleImpl updateLifecycle( true, requestNs );
request.setLifecycle( &updateLifecycle );
UpdateResult res = update( request, &cc().curop()->debug() );
numDocsModified = res.numDocsModified;
numMatched = res.numMatched;
resUpsertedID = res.upserted;
// We have an _id from an insert
didInsert = !resUpsertedID.isEmpty();
result->stats.nModified = didInsert ? 0 : numDocsModified;
result->stats.n = didInsert ? 1 : numMatched;
result->stats.upsertedID = resUpsertedID;
}
catch ( const DBException& ex ) {
result->error = toWriteError( ex.toStatus() );
}
}
void WriteBatchExecutor::execRemove( const BatchItemRef& removeItem,
WriteErrorDetail** error ) {
// Removes are similar to updates, but page faults are handled externally
// BEGIN CURRENT OP
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, removeItem ) );
incOpStats( removeItem );
WriteOpResult result;
multiRemove( _txn, removeItem, &result );
// END CURRENT OP
incWriteStats( removeItem, result.getStats(), result.getError(), currentOp.get() );
finishCurrentOp( _txn, _client, currentOp.get(), result.getError() );
if ( result.getError() ) {
result.getError()->setIndex( removeItem.getItemIndex() );
*error = result.releaseError();
}
}
void WriteBatchExecutor::execRemove( const BatchItemRef& removeItem,
WriteErrorDetail** error ) {
// Removes are similar to updates, but page faults are handled externally
// BEGIN CURRENT OP
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, removeItem ) );
incOpStats( removeItem );
WriteOpResult result;
while ( true ) {
multiRemove( removeItem, &result );
if ( !result.fault ) {
incWriteStats( removeItem, result.stats, result.error, currentOp.get() );
break;
}
//
// Check page fault out of lock
//
dassert( result.fault );
result.fault->touch();
result.reset();
}
// END CURRENT OP
finishCurrentOp( _client, currentOp.get(), result.error );
if ( result.error ) {
result.error->setIndex( removeItem.getItemIndex() );
*error = result.releaseError();
}
}
void WriteBatchExecutor::execUpdate( const BatchItemRef& updateItem,
BSONObj* upsertedId,
WriteErrorDetail** error ) {
// BEGIN CURRENT OP
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, updateItem ) );
incOpStats( updateItem );
WriteOpResult result;
multiUpdate( updateItem, &result );
incWriteStats( updateItem, result.stats, result.error, currentOp.get() );
if ( !result.stats.upsertedID.isEmpty() ) {
*upsertedId = result.stats.upsertedID;
}
// END CURRENT OP
finishCurrentOp( _client, currentOp.get(), result.error );
if ( result.error ) {
result.error->setIndex( updateItem.getItemIndex() );
*error = result.releaseError();
}
}
bool WriteBatchExecutor::applyWriteItem( const BatchItemRef& itemRef,
WriteStats* stats,
BSONObj* upsertedID,
BatchedErrorDetail* error ) {
const BatchedCommandRequest& request = *itemRef.getRequest();
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() );
HostAndPort remote =
_client->hasRemote() ? _client->getRemote() : HostAndPort( "0.0.0.0", 0 );
// TODO Modify CurOp "wrapped" constructor to take an opcode, so calling .reset()
// is unneeded
childOp.reset( remote, getOpCode( request.getBatchType() ) );
childOp.ensureStarted();
OpDebug& opDebug = childOp.debug();
opDebug.ns = ns;
{
Lock::DBWrite dbLock( ns );
Client::Context ctx( ns,
storageGlobalParams.dbpath, // TODO: better constructor?
false /* don't check version here */);
opSuccess = doWrite( ns, itemRef, &childOp, stats, upsertedID, error );
}
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 >
serverGlobalParams.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;
}
bool WriteBatchExecutor::doWrite( const string& ns,
const BatchItemRef& itemRef,
CurOp* currentOp,
WriteStats* stats,
BSONObj* upsertedID,
BatchedErrorDetail* error ) {
const BatchedCommandRequest& request = *itemRef.getRequest();
int index = itemRef.getItemIndex();
//
// Check our shard version if we need to (must be in the write lock)
//
if ( shardingState.enabled() && request.isShardVersionSet()
&& !ChunkVersion::isIgnoredVersion( request.getShardVersion() ) ) {
Lock::assertWriteLocked( ns );
CollectionMetadataPtr metadata = shardingState.getCollectionMetadata( ns );
ChunkVersion shardVersion =
metadata ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
if ( !request.getShardVersion() //
.isWriteCompatibleWith( shardVersion ) ) {
// Write stale error to results
error->setErrCode( ErrorCodes::StaleShardVersion );
BSONObjBuilder infoB;
shardVersion.addToBSON( infoB, "vWanted" );
error->setErrInfo( infoB.obj() );
string errMsg = mongoutils::str::stream()
<< "stale shard version detected before write, received "
<< request.getShardVersion().toString() << " but local version is "
<< shardVersion.toString();
error->setErrMessage( errMsg );
return false;
}
}
//
// Not stale, do the actual write
//
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert ) {
// Insert
return doInsert( ns,
request.getInsertRequest()->getDocumentsAt( index ),
currentOp,
stats,
error );
}
else if ( request.getBatchType() == BatchedCommandRequest::BatchType_Update ) {
// Update
return doUpdate( ns,
*request.getUpdateRequest()->getUpdatesAt( index ),
currentOp,
stats,
upsertedID,
error );
}
else {
dassert( request.getBatchType() ==
BatchedCommandRequest::BatchType_Delete );
// Delete
return doDelete( ns,
*request.getDeleteRequest()->getDeletesAt( index ),
currentOp,
stats,
error );
}
}
void WriteBatchExecutor::execRemove( const BatchItemRef& removeItem,
WriteErrorDetail** error ) {
// Removes are similar to updates, but page faults are handled externally
const BatchedCommandRequest& request = *removeItem.getRequest();
const NamespaceString nss( removeItem.getRequest()->getNS() );
// BEGIN CURRENT OP
scoped_ptr<CurOp> currentOp( beginCurrentOp( _client, removeItem ) );
incOpStats( removeItem );
WriteOpResult result;
while ( true ) {
{
// NOTE: Deletes will not fault outside the lock once any data has been written
PageFaultRetryableSection pFaultSection;
///////////////////////////////////////////
Lock::DBWrite writeLock( nss.ns() );
///////////////////////////////////////////
// Check version once we're locked
if ( !checkShardVersion( &shardingState, request, &result.error ) ) {
// Version error
break;
}
// Context once we're locked, to set more details in currentOp()
// TODO: better constructor?
Client::Context writeContext( nss.ns(),
storageGlobalParams.dbpath,
false /* don't check version */);
multiRemove( removeItem, &result );
if ( !result.fault ) {
incWriteStats( removeItem, result.stats, result.error, currentOp.get() );
break;
}
}
//
// Check page fault out of lock
//
dassert( result.fault );
result.fault->touch();
result.reset();
}
// END CURRENT OP
finishCurrentOp( _client, currentOp.get(), result.error );
if ( result.error ) {
result.error->setIndex( removeItem.getItemIndex() );
*error = result.releaseError();
}
}
bool WriteBatchExecutor::doWrite( const string& ns,
const BatchItemRef& itemRef,
CurOp* currentOp,
WriteStats* stats,
BSONObj* upsertedID,
BatchedErrorDetail* error ) {
const BatchedCommandRequest& request = *itemRef.getRequest();
int index = itemRef.getItemIndex();
//
// Check our shard version if we need to (must be in the write lock)
//
CollectionMetadataPtr metadata;
if ( shardingState.enabled() ) {
// Index inserts make the namespace nontrivial for versioning
string targetingNS = itemRef.getRequest()->getTargetingNS();
Lock::assertWriteLocked( targetingNS );
metadata = shardingState.getCollectionMetadata( targetingNS );
if ( request.isShardVersionSet()
&& !ChunkVersion::isIgnoredVersion( request.getShardVersion() ) ) {
ChunkVersion shardVersion =
metadata ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
if ( !request.getShardVersion() //
.isWriteCompatibleWith( shardVersion ) ) {
buildStaleError( request.getShardVersion(), shardVersion, error );
return false;
}
}
}
//
// Not stale, do the actual write
//
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert ) {
// Need to check for unique index problems
if ( metadata && request.isUniqueIndexRequest() ) {
if ( !isUniqueIndexCompatible( metadata->getKeyPattern(),
request.getIndexKeyPattern() ) ) {
buildUniqueIndexError( metadata->getKeyPattern(),
request.getIndexKeyPattern(),
error );
return false;
}
}
// Insert
return doInsert( ns,
request.getInsertRequest()->getDocumentsAt( index ),
currentOp,
stats,
error );
}
else if ( request.getBatchType() == BatchedCommandRequest::BatchType_Update ) {
// TODO: Pass down immutable shard key fields
// Update
return doUpdate( ns,
*request.getUpdateRequest()->getUpdatesAt( index ),
currentOp,
stats,
upsertedID,
error );
}
else {
dassert( request.getBatchType() ==
BatchedCommandRequest::BatchType_Delete );
// Delete
return doDelete( ns,
*request.getDeleteRequest()->getDeletesAt( index ),
currentOp,
stats,
error );
}
}
Status Strategy::commandOpWrite(const std::string& dbName,
const BSONObj& command,
BatchItemRef targetingBatchItem,
std::vector<CommandResult>* results) {
// Note that this implementation will not handle targeting retries and does not completely
// emulate write behavior
ChunkManagerTargeter targeter(NamespaceString(
targetingBatchItem.getRequest()->getTargetingNS()));
Status status = targeter.init();
if (!status.isOK())
return status;
OwnedPointerVector<ShardEndpoint> endpointsOwned;
vector<ShardEndpoint*>& endpoints = endpointsOwned.mutableVector();
if (targetingBatchItem.getOpType() == BatchedCommandRequest::BatchType_Insert) {
ShardEndpoint* endpoint;
Status status = targeter.targetInsert(targetingBatchItem.getDocument(), &endpoint);
if (!status.isOK())
return status;
endpoints.push_back(endpoint);
}
else if (targetingBatchItem.getOpType() == BatchedCommandRequest::BatchType_Update) {
Status status = targeter.targetUpdate(*targetingBatchItem.getUpdate(), &endpoints);
if (!status.isOK())
return status;
}
else {
invariant(targetingBatchItem.getOpType() == BatchedCommandRequest::BatchType_Delete);
Status status = targeter.targetDelete(*targetingBatchItem.getDelete(), &endpoints);
if (!status.isOK())
return status;
}
DBClientShardResolver resolver;
DBClientMultiCommand dispatcher;
// Assemble requests
for (vector<ShardEndpoint*>::const_iterator it = endpoints.begin(); it != endpoints.end();
++it) {
const ShardEndpoint* endpoint = *it;
ConnectionString host;
Status status = resolver.chooseWriteHost(endpoint->shardName, &host);
if (!status.isOK())
return status;
RawBSONSerializable request(command);
dispatcher.addCommand(host, dbName, request);
}
// Errors reported when recv'ing responses
dispatcher.sendAll();
Status dispatchStatus = Status::OK();
// Recv responses
while (dispatcher.numPending() > 0) {
ConnectionString host;
RawBSONSerializable response;
Status status = dispatcher.recvAny(&host, &response);
if (!status.isOK()) {
// We always need to recv() all the sent operations
dispatchStatus = status;
continue;
}
CommandResult result;
result.target = host;
result.shardTarget = Shard::make(host.toString());
result.result = response.toBSON();
results->push_back(result);
}
return dispatchStatus;
}