Status WriteCmd::explain(OperationContext* txn,
const std::string& dbname,
const BSONObj& cmdObj,
ExplainCommon::Verbosity verbosity,
BSONObjBuilder* out) const {
// For now we only explain update and delete write commands.
if ( BatchedCommandRequest::BatchType_Update != _writeType &&
BatchedCommandRequest::BatchType_Delete != _writeType ) {
return Status( ErrorCodes::IllegalOperation,
"Only update and delete write ops can be explained" );
}
// Parse the batch request.
BatchedCommandRequest request( _writeType );
std::string errMsg;
if ( !request.parseBSON( cmdObj, &errMsg ) || !request.isValid( &errMsg ) ) {
return Status( ErrorCodes::FailedToParse, errMsg );
}
// Note that this is a runCommmand, and therefore, the database and the collection name
// are in different parts of the grammar for the command. But it's more convenient to
// work with a NamespaceString. We built it here and replace it in the parsed command.
// Internally, everything work with the namespace string as opposed to just the
// collection name.
NamespaceString nsString(dbname, request.getNS());
request.setNSS(nsString);
// Do the validation of the batch that is shared with non-explained write batches.
Status isValid = WriteBatchExecutor::validateBatch( request );
if (!isValid.isOK()) {
return isValid;
}
// Explain must do one additional piece of validation: For now we only explain
// singleton batches.
if ( request.sizeWriteOps() != 1u ) {
return Status( ErrorCodes::InvalidLength,
"explained write batches must be of size 1" );
}
// Get a reference to the singleton batch item (it's the 0th item in the batch).
BatchItemRef batchItem( &request, 0 );
if ( BatchedCommandRequest::BatchType_Update == _writeType ) {
// Create the update request.
UpdateRequest updateRequest( nsString );
updateRequest.setQuery( batchItem.getUpdate()->getQuery() );
updateRequest.setUpdates( batchItem.getUpdate()->getUpdateExpr() );
updateRequest.setMulti( batchItem.getUpdate()->getMulti() );
updateRequest.setUpsert( batchItem.getUpdate()->getUpsert() );
updateRequest.setUpdateOpLog( true );
UpdateLifecycleImpl updateLifecycle( true, updateRequest.getNamespaceString() );
updateRequest.setLifecycle( &updateLifecycle );
updateRequest.setExplain();
// Explained updates can yield.
updateRequest.setYieldPolicy(PlanExecutor::YIELD_AUTO);
OpDebug* debug = &txn->getCurOp()->debug();
ParsedUpdate parsedUpdate( txn, &updateRequest );
Status parseStatus = parsedUpdate.parseRequest();
if ( !parseStatus.isOK() ) {
return parseStatus;
}
// Explains of write commands are read-only, but we take write locks so
// that timing info is more accurate.
AutoGetDb autoDb( txn, nsString.db(), MODE_IX );
Lock::CollectionLock colLock( txn->lockState(), nsString.ns(), MODE_IX );
// We check the shard version explicitly here rather than using Client::Context,
// as Context can do implicit database creation if the db does not exist. We want
// explain to be a no-op that reports a trivial EOF plan against non-existent dbs
// or collections.
ensureShardVersionOKOrThrow( nsString.ns() );
// Get a pointer to the (possibly NULL) collection.
Collection* collection = NULL;
if ( autoDb.getDb() ) {
collection = autoDb.getDb()->getCollection( txn, nsString.ns() );
}
PlanExecutor* rawExec;
uassertStatusOK(getExecutorUpdate(txn, collection, &parsedUpdate, debug, &rawExec));
boost::scoped_ptr<PlanExecutor> exec(rawExec);
// Explain the plan tree.
Explain::explainStages( exec.get(), verbosity, out );
return Status::OK();
}
else {
invariant( BatchedCommandRequest::BatchType_Delete == _writeType );
// Create the delete request.
DeleteRequest deleteRequest( nsString );
deleteRequest.setQuery( batchItem.getDelete()->getQuery() );
deleteRequest.setMulti( batchItem.getDelete()->getLimit() != 1 );
deleteRequest.setUpdateOpLog(true);
deleteRequest.setGod( false );
deleteRequest.setExplain();
// Explained deletes can yield.
deleteRequest.setYieldPolicy(PlanExecutor::YIELD_AUTO);
ParsedDelete parsedDelete(txn, &deleteRequest);
Status parseStatus = parsedDelete.parseRequest();
if (!parseStatus.isOK()) {
return parseStatus;
}
// Explains of write commands are read-only, but we take write locks so that timing
// info is more accurate.
AutoGetDb autoDb(txn, nsString.db(), MODE_IX);
Lock::CollectionLock colLock(txn->lockState(), nsString.ns(), MODE_IX);
// We check the shard version explicitly here rather than using Client::Context,
// as Context can do implicit database creation if the db does not exist. We want
// explain to be a no-op that reports a trivial EOF plan against non-existent dbs
// or collections.
ensureShardVersionOKOrThrow( nsString.ns() );
// Get a pointer to the (possibly NULL) collection.
Collection* collection = NULL;
if (autoDb.getDb()) {
collection = autoDb.getDb()->getCollection(txn, nsString.ns());
}
PlanExecutor* rawExec;
uassertStatusOK(getExecutorDelete(txn, collection, &parsedDelete, &rawExec));
boost::scoped_ptr<PlanExecutor> exec(rawExec);
// Explain the plan tree.
Explain::explainStages(exec.get(), verbosity, out);
return Status::OK();
}
}
virtual bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int options,
string& errmsg,
BSONObjBuilder& result) {
const NamespaceString ns(parseNs(dbname, cmdObj));
Status status = userAllowedWriteNS(ns);
if (!status.isOK())
return appendCommandStatus(result, status);
if (cmdObj["indexes"].type() != Array) {
errmsg = "indexes has to be an array";
result.append("cmdObj", cmdObj);
return false;
}
std::vector<BSONObj> specs;
{
BSONObjIterator i(cmdObj["indexes"].Obj());
while (i.more()) {
BSONElement e = i.next();
if (e.type() != Object) {
errmsg = "everything in indexes has to be an Object";
result.append("cmdObj", cmdObj);
return false;
}
specs.push_back(e.Obj());
}
}
if (specs.size() == 0) {
errmsg = "no indexes to add";
return false;
}
// check specs
for (size_t i = 0; i < specs.size(); i++) {
BSONObj spec = specs[i];
if (spec["ns"].eoo()) {
spec = _addNsToSpec(ns, spec);
specs[i] = spec;
}
if (spec["ns"].type() != String) {
errmsg = "ns field must be a string";
result.append("spec", spec);
return false;
}
std::string nsFromUser = spec["ns"].String();
if (nsFromUser.empty()) {
errmsg = "ns field cannot be an empty string";
result.append("spec", spec);
return false;
}
if (ns != nsFromUser) {
errmsg = str::stream() << "value of ns field '" << nsFromUser
<< "' doesn't match namespace " << ns.ns();
result.append("spec", spec);
return false;
}
}
// now we know we have to create index(es)
// Note: createIndexes command does not currently respect shard versioning.
ScopedTransaction transaction(txn, MODE_IX);
Lock::DBLock dbLock(txn->lockState(), ns.db(), MODE_X);
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns)) {
return appendCommandStatus(
result,
Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while creating indexes in " << ns.ns()));
}
Database* db = dbHolder().get(txn, ns.db());
if (!db) {
db = dbHolder().openDb(txn, ns.db());
}
Collection* collection = db->getCollection(ns.ns());
if (collection) {
result.appendBool("createdCollectionAutomatically", false);
} else {
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
WriteUnitOfWork wunit(txn);
collection = db->createCollection(txn, ns.ns(), CollectionOptions());
invariant(collection);
wunit.commit();
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "createIndexes", ns.ns());
result.appendBool("createdCollectionAutomatically", true);
}
const int numIndexesBefore = collection->getIndexCatalog()->numIndexesTotal(txn);
result.append("numIndexesBefore", numIndexesBefore);
auto client = txn->getClient();
ScopeGuard lastOpSetterGuard =
MakeObjGuard(repl::ReplClientInfo::forClient(client),
&repl::ReplClientInfo::setLastOpToSystemLastOpTime,
txn);
MultiIndexBlock indexer(txn, collection);
indexer.allowBackgroundBuilding();
indexer.allowInterruption();
const size_t origSpecsSize = specs.size();
indexer.removeExistingIndexes(&specs);
if (specs.size() == 0) {
result.append("numIndexesAfter", numIndexesBefore);
result.append("note", "all indexes already exist");
return true;
}
if (specs.size() != origSpecsSize) {
result.append("note", "index already exists");
}
for (size_t i = 0; i < specs.size(); i++) {
const BSONObj& spec = specs[i];
if (spec["unique"].trueValue()) {
status = checkUniqueIndexConstraints(txn, ns.ns(), spec["key"].Obj());
if (!status.isOK()) {
return appendCommandStatus(result, status);
}
}
if (spec["v"].isNumber() && spec["v"].numberInt() == 0) {
return appendCommandStatus(
result,
Status(ErrorCodes::CannotCreateIndex,
str::stream() << "illegal index specification: " << spec << ". "
<< "The option v:0 cannot be passed explicitly"));
}
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
uassertStatusOK(indexer.init(specs));
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "createIndexes", ns.ns());
// If we're a background index, replace exclusive db lock with an intent lock, so that
// other readers and writers can proceed during this phase.
if (indexer.getBuildInBackground()) {
txn->recoveryUnit()->abandonSnapshot();
dbLock.relockWithMode(MODE_IX);
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns)) {
return appendCommandStatus(
result,
Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while creating background indexes in "
<< ns.ns()));
}
}
try {
Lock::CollectionLock colLock(txn->lockState(), ns.ns(), MODE_IX);
uassertStatusOK(indexer.insertAllDocumentsInCollection());
} catch (const DBException& e) {
invariant(e.getCode() != ErrorCodes::WriteConflict);
// Must have exclusive DB lock before we clean up the index build via the
// destructor of 'indexer'.
if (indexer.getBuildInBackground()) {
try {
// This function cannot throw today, but we will preemptively prepare for
// that day, to avoid data corruption due to lack of index cleanup.
txn->recoveryUnit()->abandonSnapshot();
dbLock.relockWithMode(MODE_X);
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns)) {
return appendCommandStatus(
result,
Status(ErrorCodes::NotMaster,
str::stream()
<< "Not primary while creating background indexes in "
<< ns.ns() << ": cleaning up index build failure due to "
<< e.toString()));
}
} catch (...) {
std::terminate();
}
}
throw;
}
// Need to return db lock back to exclusive, to complete the index build.
if (indexer.getBuildInBackground()) {
txn->recoveryUnit()->abandonSnapshot();
dbLock.relockWithMode(MODE_X);
uassert(ErrorCodes::NotMaster,
str::stream() << "Not primary while completing index build in " << dbname,
repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns));
Database* db = dbHolder().get(txn, ns.db());
uassert(28551, "database dropped during index build", db);
uassert(28552, "collection dropped during index build", db->getCollection(ns.ns()));
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
WriteUnitOfWork wunit(txn);
indexer.commit();
for (size_t i = 0; i < specs.size(); i++) {
std::string systemIndexes = ns.getSystemIndexesCollection();
getGlobalServiceContext()->getOpObserver()->onCreateIndex(
txn, systemIndexes, specs[i]);
}
wunit.commit();
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "createIndexes", ns.ns());
result.append("numIndexesAfter", collection->getIndexCatalog()->numIndexesTotal(txn));
lastOpSetterGuard.Dismiss();
return true;
}
Status IndexBuilder::_build(OperationContext* txn,
Database* db,
bool allowBackgroundBuilding,
Lock::DBLock* dbLock) const {
const NamespaceString ns(_index["ns"].String());
Collection* c = db->getCollection( txn, ns.ns() );
if ( !c ) {
WriteUnitOfWork wunit(txn);
c = db->getOrCreateCollection( txn, ns.ns() );
verify(c);
wunit.commit();
}
// Show which index we're building in the curop display.
txn->getCurOp()->setQuery(_index);
MultiIndexBlock indexer(txn, c);
indexer.allowInterruption();
if (allowBackgroundBuilding)
indexer.allowBackgroundBuilding();
Status status = Status::OK();
IndexDescriptor* descriptor(NULL);
try {
status = indexer.init(_index);
if ( status.code() == ErrorCodes::IndexAlreadyExists )
return Status::OK();
if (status.isOK()) {
if (allowBackgroundBuilding) {
descriptor = indexer.registerIndexBuild();
invariant(dbLock);
dbLock->relockWithMode(MODE_IX);
}
Lock::CollectionLock colLock(txn->lockState(), ns.ns(), MODE_IX);
status = indexer.insertAllDocumentsInCollection();
}
if (status.isOK()) {
if (allowBackgroundBuilding) {
dbLock->relockWithMode(MODE_X);
}
WriteUnitOfWork wunit(txn);
indexer.commit();
wunit.commit();
}
}
catch (const DBException& e) {
status = e.toStatus();
}
if (allowBackgroundBuilding) {
dbLock->relockWithMode(MODE_X);
Database* db = dbHolder().get(txn, ns.db());
fassert(28553, db);
fassert(28554, db->getCollection(txn, ns.ns()));
indexer.unregisterIndexBuild(descriptor);
}
if (status.code() == ErrorCodes::InterruptedAtShutdown) {
// leave it as-if kill -9 happened. This will be handled on restart.
indexer.abortWithoutCleanup();
}
return status;
}