void DeferredWriter::_worker(InsertStatement stmt) {
auto uniqueOpCtx = Client::getCurrent()->makeOperationContext();
OperationContext* opCtx = uniqueOpCtx.get();
auto result = _getCollection(opCtx);
if (!result.isOK()) {
_logFailure(result.getStatus());
return;
}
auto agc = std::move(result.getValue());
Collection& collection = *agc->getCollection();
Status status = writeConflictRetry(opCtx, "deferred insert", _nss.ns(), [&] {
WriteUnitOfWork wuow(opCtx);
Status status = collection.insertDocument(opCtx, stmt, nullptr, false);
if (!status.isOK()) {
return status;
}
wuow.commit();
return Status::OK();
});
stdx::lock_guard<stdx::mutex> lock(_mutex);
_numBytes -= stmt.doc.objsize();
// If a write to a deferred collection fails, periodically tell the log.
if (!status.isOK()) {
_logFailure(status);
}
}
OpTimeWithHash BackgroundSync::_readLastAppliedOpTimeWithHash(OperationContext* opCtx) {
BSONObj oplogEntry;
try {
bool success = writeConflictRetry(
opCtx, "readLastAppliedHash", NamespaceString::kRsOplogNamespace.ns(), [&] {
Lock::DBLock lk(opCtx, "local", MODE_X);
return Helpers::getLast(
opCtx, NamespaceString::kRsOplogNamespace.ns().c_str(), oplogEntry);
});
if (!success) {
// This can happen when we are to do an initial sync. lastHash will be set
// after the initial sync is complete.
return OpTimeWithHash(0);
}
} catch (const DBException& ex) {
severe() << "Problem reading " << NamespaceString::kRsOplogNamespace.ns() << ": "
<< redact(ex);
fassertFailed(18904);
}
long long hash;
auto status = bsonExtractIntegerField(oplogEntry, kHashFieldName, &hash);
if (!status.isOK()) {
severe() << "Most recent entry in " << NamespaceString::kRsOplogNamespace.ns()
<< " is missing or has invalid \"" << kHashFieldName
<< "\" field. Oplog entry: " << redact(oplogEntry) << ": " << redact(status);
fassertFailed(18902);
}
OplogEntry parsedEntry(oplogEntry);
return OpTimeWithHash(hash, parsedEntry.getOpTime());
}
OpTime ReplicationCoordinatorExternalStateImpl::onTransitionToPrimary(OperationContext* opCtx,
bool isV1ElectionProtocol) {
invariant(opCtx->lockState()->isW());
// Clear the appliedThrough marker so on startup we'll use the top of the oplog. This must be
// done before we add anything to our oplog.
// We record this update at the 'lastAppliedOpTime'. If there are any outstanding
// checkpoints being taken, they should only reflect this write if they see all writes up
// to our 'lastAppliedOpTime'.
invariant(
_replicationProcess->getConsistencyMarkers()->getOplogTruncateAfterPoint(opCtx).isNull());
auto lastAppliedOpTime = repl::ReplicationCoordinator::get(opCtx)->getMyLastAppliedOpTime();
_replicationProcess->getConsistencyMarkers()->clearAppliedThrough(
opCtx, lastAppliedOpTime.getTimestamp());
if (isV1ElectionProtocol) {
writeConflictRetry(opCtx, "logging transition to primary to oplog", "local.oplog.rs", [&] {
WriteUnitOfWork wuow(opCtx);
opCtx->getClient()->getServiceContext()->getOpObserver()->onOpMessage(
opCtx,
BSON("msg"
<< "new primary"));
wuow.commit();
});
}
const auto opTimeToReturn = fassertStatusOK(28665, loadLastOpTime(opCtx));
_shardingOnTransitionToPrimaryHook(opCtx);
_dropAllTempCollections(opCtx);
serverGlobalParams.validateFeaturesAsMaster.store(true);
return opTimeToReturn;
}
void DatabaseHolderImpl::dropDb(OperationContext* opCtx, Database* db) {
invariant(db);
// Store the name so we have if for after the db object is deleted
auto name = db->name();
LOG(1) << "dropDatabase " << name;
invariant(opCtx->lockState()->isDbLockedForMode(name, MODE_X));
BackgroundOperation::assertNoBgOpInProgForDb(name);
audit::logDropDatabase(opCtx->getClient(), name);
auto const serviceContext = opCtx->getServiceContext();
for (auto collIt = db->begin(opCtx); collIt != db->end(opCtx); ++collIt) {
auto coll = *collIt;
if (!coll) {
break;
}
Top::get(serviceContext).collectionDropped(coll->ns().ns(), true);
}
close(opCtx, name);
auto const storageEngine = serviceContext->getStorageEngine();
writeConflictRetry(opCtx, "dropDatabase", name, [&] {
storageEngine->dropDatabase(opCtx, name).transitional_ignore();
});
}
Status ReplicationCoordinatorExternalStateImpl::initializeReplSetStorage(OperationContext* opCtx,
const BSONObj& config) {
try {
createOplog(opCtx);
writeConflictRetry(opCtx,
"initiate oplog entry",
NamespaceString::kRsOplogNamespace.toString(),
[this, &opCtx, &config] {
Lock::GlobalWrite globalWrite(opCtx);
WriteUnitOfWork wuow(opCtx);
Helpers::putSingleton(opCtx, configCollectionName, config);
const auto msgObj = BSON("msg"
<< "initiating set");
_service->getOpObserver()->onOpMessage(opCtx, msgObj);
wuow.commit();
// ReplSetTest assumes that immediately after the replSetInitiate
// command returns, it can allow other nodes to initial sync with no
// retries and they will succeed. Unfortunately, initial sync will
// fail if it finds its sync source has an empty oplog. Thus, we
// need to wait here until the seed document is visible in our oplog.
AutoGetCollection oplog(
opCtx, NamespaceString::kRsOplogNamespace, MODE_IS);
waitForAllEarlierOplogWritesToBeVisible(opCtx);
});
// Set UUIDs for all non-replicated collections. This is necessary for independent replica
// sets and config server replica sets started with no data files because collections in
// local are created prior to the featureCompatibilityVersion being set to 3.6, so the
// collections are not created with UUIDs. We exclude ShardServers when adding UUIDs to
// non-replicated collections on the primary because ShardServers are started up by default
// with featureCompatibilityVersion 3.4, so we don't want to assign UUIDs to them until the
// cluster's featureCompatibilityVersion is explicitly set to 3.6 by the config server. The
// below UUID addition for non-replicated collections only occurs on the primary; UUIDs are
// added to non-replicated collections on secondaries during InitialSync. When the config
// server sets the featureCompatibilityVersion to 3.6, the shard primary will add UUIDs to
// all the collections that need them. One special case here is if a shard is already in
// featureCompatibilityVersion 3.6 and a new node is started up with --shardsvr and added to
// that shard, the new node will still start up with featureCompatibilityVersion 3.4 and
// need to have UUIDs added to each collection. These UUIDs are added during InitialSync,
// because the new node is a secondary.
if (serverGlobalParams.clusterRole != ClusterRole::ShardServer &&
FeatureCompatibilityVersion::isCleanStartUp()) {
auto schemaStatus = updateUUIDSchemaVersionNonReplicated(opCtx, true);
if (!schemaStatus.isOK()) {
return schemaStatus;
}
}
FeatureCompatibilityVersion::setIfCleanStartup(opCtx, _storageInterface);
} catch (const DBException& ex) {
return ex.toStatus();
}
return Status::OK();
}
Status ReplicationCoordinatorExternalStateImpl::storeLocalConfigDocument(OperationContext* opCtx,
const BSONObj& config) {
try {
writeConflictRetry(opCtx, "save replica set config", configCollectionName, [&] {
Lock::DBLock dbWriteLock(opCtx, configDatabaseName, MODE_X);
Helpers::putSingleton(opCtx, configCollectionName, config);
});
return Status::OK();
} catch (const DBException& ex) {
return ex.toStatus();
}
}
Status dropIndexes(OperationContext* opCtx,
const NamespaceString& nss,
const BSONObj& cmdObj,
BSONObjBuilder* result) {
return writeConflictRetry(opCtx, "dropIndexes", nss.db(), [opCtx, &nss, &cmdObj, result] {
AutoGetCollection autoColl(opCtx, nss, MODE_IX, MODE_X);
bool userInitiatedWritesAndNotPrimary = opCtx->writesAreReplicated() &&
!repl::ReplicationCoordinator::get(opCtx)->canAcceptWritesFor(opCtx, nss);
if (userInitiatedWritesAndNotPrimary) {
return Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while dropping indexes in " << nss);
}
if (!serverGlobalParams.quiet.load()) {
LOG(0) << "CMD: dropIndexes " << nss << ": " << cmdObj[kIndexFieldName].toString(false);
}
// If db/collection does not exist, short circuit and return.
Database* db = autoColl.getDb();
Collection* collection = autoColl.getCollection();
if (!collection) {
if (db && ViewCatalog::get(db)->lookup(opCtx, nss.ns())) {
return Status(ErrorCodes::CommandNotSupportedOnView,
str::stream() << "Cannot drop indexes on view " << nss);
}
return Status(ErrorCodes::NamespaceNotFound, "ns not found");
}
BackgroundOperation::assertNoBgOpInProgForNs(nss);
IndexBuildsCoordinator::get(opCtx)->assertNoIndexBuildInProgForCollection(
collection->uuid().get());
WriteUnitOfWork wunit(opCtx);
OldClientContext ctx(opCtx, nss.ns());
Status status = wrappedRun(opCtx, collection, cmdObj, result);
if (!status.isOK()) {
return status;
}
wunit.commit();
return Status::OK();
});
}
StatusWith<BSONObj> ReplicationCoordinatorExternalStateImpl::loadLocalConfigDocument(
OperationContext* opCtx) {
try {
return writeConflictRetry(opCtx, "load replica set config", configCollectionName, [opCtx] {
BSONObj config;
if (!Helpers::getSingleton(opCtx, configCollectionName, config)) {
return StatusWith<BSONObj>(
ErrorCodes::NoMatchingDocument,
str::stream() << "Did not find replica set configuration document in "
<< configCollectionName);
}
return StatusWith<BSONObj>(config);
});
} catch (const DBException& ex) {
return StatusWith<BSONObj>(ex.toStatus());
}
}
void ServiceContextMongoDTest::_dropAllDBs(OperationContext* opCtx) {
dropAllDatabasesExceptLocal(opCtx);
Lock::GlobalWrite lk(opCtx);
AutoGetDb autoDBLocal(opCtx, "local", MODE_X);
const auto localDB = autoDBLocal.getDb();
if (localDB) {
writeConflictRetry(opCtx, "_dropAllDBs", "local", [&] {
// Do not wrap in a WriteUnitOfWork until SERVER-17103 is addressed.
autoDBLocal.getDb()->dropDatabase(opCtx, localDB);
});
}
// dropAllDatabasesExceptLocal() does not close empty databases. However the holder still
// allocates resources to track these empty databases. These resources not released by
// dropAllDatabasesExceptLocal() will be leaked at exit unless we call DatabaseHolder::closeAll.
dbHolder().closeAll(opCtx, "all databases dropped");
}
StatusWith<LastVote> ReplicationCoordinatorExternalStateImpl::loadLocalLastVoteDocument(
OperationContext* opCtx) {
try {
return writeConflictRetry(
opCtx, "load replica set lastVote", lastVoteCollectionName, [opCtx] {
BSONObj lastVoteObj;
if (!Helpers::getSingleton(opCtx, lastVoteCollectionName, lastVoteObj)) {
return StatusWith<LastVote>(
ErrorCodes::NoMatchingDocument,
str::stream() << "Did not find replica set lastVote document in "
<< lastVoteCollectionName);
}
return LastVote::readFromLastVote(lastVoteObj);
});
} catch (const DBException& ex) {
return StatusWith<LastVote>(ex.toStatus());
}
}
void NoopWriter::_writeNoop(OperationContext* opCtx) {
// Use GlobalLock + lockMMAPV1Flush instead of DBLock to allow return when the lock is not
// available. It may happen when the primary steps down and a shared global lock is acquired.
Lock::GlobalLock lock(
opCtx, MODE_IX, Date_t::now() + Milliseconds(1), Lock::InterruptBehavior::kLeaveUnlocked);
if (!lock.isLocked()) {
LOG(1) << "Global lock is not available skipping noopWrite";
return;
}
opCtx->lockState()->lockMMAPV1Flush();
auto replCoord = ReplicationCoordinator::get(opCtx);
// Its a proxy for being a primary
if (!replCoord->canAcceptWritesForDatabase(opCtx, "admin")) {
LOG(1) << "Not a primary, skipping the noop write";
return;
}
auto lastAppliedOpTime = replCoord->getMyLastAppliedOpTime();
// _lastKnownOpTime is not protected by lock as its used only by one thread.
if (lastAppliedOpTime != _lastKnownOpTime) {
LOG(1) << "Not scheduling a noop write. Last known OpTime: " << _lastKnownOpTime
<< " != last primary OpTime: " << lastAppliedOpTime;
} else {
if (writePeriodicNoops.load()) {
const auto logLevel = getTestCommandsEnabled() ? 0 : 1;
LOG(logLevel)
<< "Writing noop to oplog as there has been no writes to this replica set in over "
<< _writeInterval;
writeConflictRetry(
opCtx, "writeNoop", NamespaceString::kRsOplogNamespace.ns(), [&opCtx] {
WriteUnitOfWork uow(opCtx);
opCtx->getClient()->getServiceContext()->getOpObserver()->onOpMessage(opCtx,
kMsgObj);
uow.commit();
});
}
}
_lastKnownOpTime = replCoord->getMyLastAppliedOpTime();
LOG(1) << "Set last known op time to " << _lastKnownOpTime;
}
SessionCatalogMigrationSource::SessionCatalogMigrationSource(OperationContext* opCtx,
NamespaceString ns)
: _ns(std::move(ns)), _rollbackIdAtInit(repl::ReplicationProcess::get(opCtx)->getRollbackID()) {
// Exclude entries for transaction.
Query query;
// Sort is not needed for correctness. This is just for making it easier to write deterministic
// tests.
query.sort(BSON("_id" << 1));
DBDirectClient client(opCtx);
auto cursor = client.query(NamespaceString::kSessionTransactionsTableNamespace, query);
while (cursor->more()) {
auto nextSession = SessionTxnRecord::parse(
IDLParserErrorContext("Session migration cloning"), cursor->next());
if (!nextSession.getLastWriteOpTime().isNull()) {
_sessionOplogIterators.push_back(
stdx::make_unique<SessionOplogIterator>(std::move(nextSession), _rollbackIdAtInit));
}
}
{
AutoGetCollection autoColl(opCtx, NamespaceString::kRsOplogNamespace, MODE_IX);
writeConflictRetry(
opCtx,
"session migration initialization majority commit barrier",
NamespaceString::kRsOplogNamespace.ns(),
[&] {
const auto message = BSON("sessionMigrateCloneStart" << _ns.ns());
WriteUnitOfWork wuow(opCtx);
opCtx->getClient()->getServiceContext()->getOpObserver()->onInternalOpMessage(
opCtx, _ns, {}, {}, message);
wuow.commit();
});
}
auto opTimeToWait = repl::ReplClientInfo::forClient(opCtx->getClient()).getLastOp();
WriteConcernResult result;
WriteConcernOptions majority(
WriteConcernOptions::kMajority, WriteConcernOptions::SyncMode::UNSET, 0);
uassertStatusOK(waitForWriteConcern(opCtx, opTimeToWait, majority, &result));
}
Status ReplicationCoordinatorExternalStateImpl::storeLocalLastVoteDocument(
OperationContext* opCtx, const LastVote& lastVote) {
BSONObj lastVoteObj = lastVote.toBSON();
try {
Status status =
writeConflictRetry(opCtx, "save replica set lastVote", lastVoteCollectionName, [&] {
Lock::DBLock dbWriteLock(opCtx, lastVoteDatabaseName, MODE_X);
// If there is no last vote document, we want to store one. Otherwise, we only want
// to replace it if the new last vote document would have a higher term. We both
// check the term of the current last vote document and insert the new document
// under the DBLock to synchronize the two operations.
BSONObj result;
bool exists = Helpers::getSingleton(opCtx, lastVoteCollectionName, result);
if (!exists) {
Helpers::putSingleton(opCtx, lastVoteCollectionName, lastVoteObj);
} else {
StatusWith<LastVote> oldLastVoteDoc = LastVote::readFromLastVote(result);
if (!oldLastVoteDoc.isOK()) {
return oldLastVoteDoc.getStatus();
}
if (lastVote.getTerm() > oldLastVoteDoc.getValue().getTerm()) {
Helpers::putSingleton(opCtx, lastVoteCollectionName, lastVoteObj);
}
}
return Status::OK();
});
if (!status.isOK()) {
return status;
}
opCtx->recoveryUnit()->waitUntilDurable();
return Status::OK();
} catch (const DBException& ex) {
return ex.toStatus();
}
}
StatusWith<int> CollectionRangeDeleter::_doDeletion(OperationContext* opCtx,
Collection* collection,
BSONObj const& keyPattern,
ChunkRange const& range,
int maxToDelete) {
invariant(collection != nullptr);
invariant(!isEmpty());
auto const& nss = collection->ns();
// The IndexChunk has a keyPattern that may apply to more than one index - we need to
// select the index and get the full index keyPattern here.
auto catalog = collection->getIndexCatalog();
const IndexDescriptor* idx = catalog->findShardKeyPrefixedIndex(opCtx, keyPattern, false);
if (!idx) {
std::string msg = str::stream() << "Unable to find shard key index for "
<< keyPattern.toString() << " in " << nss.ns();
LOG(0) << msg;
return {ErrorCodes::InternalError, msg};
}
// Extend bounds to match the index we found
const KeyPattern indexKeyPattern(idx->keyPattern());
const auto extend = [&](const auto& key) {
return Helpers::toKeyFormat(indexKeyPattern.extendRangeBound(key, false));
};
const auto min = extend(range.getMin());
const auto max = extend(range.getMax());
LOG(1) << "begin removal of " << min << " to " << max << " in " << nss.ns();
const auto indexName = idx->indexName();
IndexDescriptor* descriptor = collection->getIndexCatalog()->findIndexByName(opCtx, indexName);
if (!descriptor) {
std::string msg = str::stream() << "shard key index with name " << indexName << " on '"
<< nss.ns() << "' was dropped";
LOG(0) << msg;
return {ErrorCodes::InternalError, msg};
}
boost::optional<Helpers::RemoveSaver> saver;
if (serverGlobalParams.moveParanoia) {
saver.emplace("moveChunk", nss.ns(), "cleaning");
}
auto halfOpen = BoundInclusion::kIncludeStartKeyOnly;
auto manual = PlanExecutor::YIELD_MANUAL;
auto forward = InternalPlanner::FORWARD;
auto fetch = InternalPlanner::IXSCAN_FETCH;
auto exec = InternalPlanner::indexScan(
opCtx, collection, descriptor, min, max, halfOpen, manual, forward, fetch);
int numDeleted = 0;
do {
RecordId rloc;
BSONObj obj;
PlanExecutor::ExecState state = exec->getNext(&obj, &rloc);
if (state == PlanExecutor::IS_EOF) {
break;
}
if (state == PlanExecutor::FAILURE || state == PlanExecutor::DEAD) {
warning() << PlanExecutor::statestr(state) << " - cursor error while trying to delete "
<< redact(min) << " to " << redact(max) << " in " << nss << ": "
<< redact(WorkingSetCommon::toStatusString(obj))
<< ", stats: " << Explain::getWinningPlanStats(exec.get());
break;
}
invariant(PlanExecutor::ADVANCED == state);
exec->saveState();
writeConflictRetry(opCtx, "delete range", nss.ns(), [&] {
WriteUnitOfWork wuow(opCtx);
if (saver) {
uassertStatusOK(saver->goingToDelete(obj));
}
collection->deleteDocument(opCtx, kUninitializedStmtId, rloc, nullptr, true);
wuow.commit();
});
try {
exec->restoreState();
} catch (const DBException& ex) {
warning() << "error restoring cursor state while trying to delete " << redact(min)
<< " to " << redact(max) << " in " << nss
<< ", stats: " << Explain::getWinningPlanStats(exec.get()) << ": "
<< redact(ex.toStatus());
break;
}
ShardingStatistics::get(opCtx).countDocsDeletedOnDonor.addAndFetch(1);
} while (++numDeleted < maxToDelete);
return numDeleted;
}
Status createCollectionForApplyOps(OperationContext* opCtx,
const std::string& dbName,
const BSONElement& ui,
const BSONObj& cmdObj,
const BSONObj& idIndex) {
invariant(opCtx->lockState()->isDbLockedForMode(dbName, MODE_X));
auto db = dbHolder().get(opCtx, dbName);
const NamespaceString newCollName(Command::parseNsCollectionRequired(dbName, cmdObj));
auto newCmd = cmdObj;
// If a UUID is given, see if we need to rename a collection out of the way, and whether the
// collection already exists under a different name. If so, rename it into place. As this is
// done during replay of the oplog, the operations do not need to be atomic, just idempotent.
// We need to do the renaming part in a separate transaction, as we cannot transactionally
// create a database on MMAPv1, which could result in createCollection failing if the database
// does not yet exist.
if (ui.ok()) {
// Return an optional, indicating whether we need to early return (if the collection already
// exists, or in case of an error).
using Result = boost::optional<Status>;
auto result =
writeConflictRetry(opCtx, "createCollectionForApplyOps", newCollName.ns(), [&] {
WriteUnitOfWork wunit(opCtx);
// Options need the field to be named "uuid", so parse/recreate.
auto uuid = uassertStatusOK(UUID::parse(ui));
uassert(ErrorCodes::InvalidUUID,
"Invalid UUID in applyOps create command: " + uuid.toString(),
uuid.isRFC4122v4());
auto& catalog = UUIDCatalog::get(opCtx);
auto currentName = catalog.lookupNSSByUUID(uuid);
OpObserver* opObserver = getGlobalServiceContext()->getOpObserver();
if (currentName == newCollName)
return Result(Status::OK());
// In the case of oplog replay, a future command may have created or renamed a
// collection with that same name. In that case, renaming this future collection to
// a random temporary name is correct: once all entries are replayed no temporary
// names will remain. On MMAPv1 the rename can result in index names that are too
// long. However this should only happen for initial sync and "resync collection"
// for rollback, so we can let the error propagate resulting in an abort and restart
// of the initial sync or result in rollback to fassert, requiring a resync of that
// node.
const bool stayTemp = true;
if (auto futureColl = db ? db->getCollection(opCtx, newCollName) : nullptr) {
auto tmpNameResult = db->makeUniqueCollectionNamespace(opCtx, "tmp%%%%%");
if (!tmpNameResult.isOK()) {
return Result(Status(tmpNameResult.getStatus().code(),
str::stream() << "Cannot generate temporary "
"collection namespace for applyOps "
"create command: collection: "
<< newCollName.ns()
<< ". error: "
<< tmpNameResult.getStatus().reason()));
}
const auto& tmpName = tmpNameResult.getValue();
Status status =
db->renameCollection(opCtx, newCollName.ns(), tmpName.ns(), stayTemp);
if (!status.isOK())
return Result(status);
opObserver->onRenameCollection(opCtx,
newCollName,
tmpName,
futureColl->uuid(),
/*dropTarget*/ false,
/*dropTargetUUID*/ {},
stayTemp);
}
// If the collection with the requested UUID already exists, but with a different
// name, just rename it to 'newCollName'.
if (catalog.lookupCollectionByUUID(uuid)) {
Status status =
db->renameCollection(opCtx, currentName.ns(), newCollName.ns(), stayTemp);
if (!status.isOK())
return Result(status);
opObserver->onRenameCollection(opCtx,
currentName,
newCollName,
uuid,
/*dropTarget*/ false,
/*dropTargetUUID*/ {},
stayTemp);
wunit.commit();
return Result(Status::OK());
}
// A new collection with the specific UUID must be created, so add the UUID to the
// creation options. Regular user collection creation commands cannot do this.
auto uuidObj = uuid.toBSON();
newCmd = cmdObj.addField(uuidObj.firstElement());
wunit.commit();
return Result(boost::none);
});
if (result) {
return *result;
}
}
return createCollection(
opCtx, newCollName, newCmd, idIndex, CollectionOptions::parseForStorage);
}
Status dropCollection(OperationContext* opCtx,
const NamespaceString& collectionName,
BSONObjBuilder& result,
const repl::OpTime& dropOpTime,
DropCollectionSystemCollectionMode systemCollectionMode) {
if (!serverGlobalParams.quiet.load()) {
log() << "CMD: drop " << collectionName;
}
return writeConflictRetry(opCtx, "drop", collectionName.ns(), [&] {
AutoGetDb autoDb(opCtx, collectionName.db(), MODE_X);
Database* const db = autoDb.getDb();
Collection* coll = db ? db->getCollection(opCtx, collectionName) : nullptr;
auto view =
db && !coll ? db->getViewCatalog()->lookup(opCtx, collectionName.ns()) : nullptr;
if (MONGO_FAIL_POINT(hangDuringDropCollection)) {
log() << "hangDuringDropCollection fail point enabled. Blocking until fail point is "
"disabled.";
MONGO_FAIL_POINT_PAUSE_WHILE_SET(hangDuringDropCollection);
}
if (!db || (!coll && !view)) {
return Status(ErrorCodes::NamespaceNotFound, "ns not found");
}
const bool shardVersionCheck = true;
OldClientContext context(opCtx, collectionName.ns(), shardVersionCheck);
bool userInitiatedWritesAndNotPrimary = opCtx->writesAreReplicated() &&
!repl::ReplicationCoordinator::get(opCtx)->canAcceptWritesFor(opCtx, collectionName);
if (userInitiatedWritesAndNotPrimary) {
return Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while dropping collection "
<< collectionName);
}
WriteUnitOfWork wunit(opCtx);
if (!result.hasField("ns")) {
result.append("ns", collectionName.ns());
}
if (coll) {
invariant(!view);
int numIndexes = coll->getIndexCatalog()->numIndexesTotal(opCtx);
BackgroundOperation::assertNoBgOpInProgForNs(collectionName.ns());
Status s = systemCollectionMode ==
DropCollectionSystemCollectionMode::kDisallowSystemCollectionDrops
? db->dropCollection(opCtx, collectionName.ns(), dropOpTime)
: db->dropCollectionEvenIfSystem(opCtx, collectionName, dropOpTime);
if (!s.isOK()) {
return s;
}
result.append("nIndexesWas", numIndexes);
} else {
invariant(view);
Status status = db->dropView(opCtx, collectionName.ns());
if (!status.isOK()) {
return status;
}
}
wunit.commit();
return Status::OK();
});
}
Status dropDatabase(OperationContext* opCtx, const std::string& dbName) {
uassert(ErrorCodes::IllegalOperation,
"Cannot drop a database in read-only mode",
!storageGlobalParams.readOnly);
// TODO (Kal): OldClientContext legacy, needs to be removed
{
CurOp::get(opCtx)->ensureStarted();
stdx::lock_guard<Client> lk(*opCtx->getClient());
CurOp::get(opCtx)->setNS_inlock(dbName);
}
auto replCoord = repl::ReplicationCoordinator::get(opCtx);
std::size_t numCollectionsToDrop = 0;
// We have to wait for the last drop-pending collection to be removed if there are no
// collections to drop.
repl::OpTime latestDropPendingOpTime;
using Result = boost::optional<Status>;
// Get an optional result--if it's there, early return; otherwise, wait for collections to drop.
auto result = writeConflictRetry(opCtx, "dropDatabase_collection", dbName, [&] {
Lock::GlobalWrite lk(opCtx);
AutoGetDb autoDB(opCtx, dbName, MODE_X);
Database* const db = autoDB.getDb();
if (!db) {
return Result(Status(ErrorCodes::NamespaceNotFound,
str::stream() << "Could not drop database " << dbName
<< " because it does not exist"));
}
bool userInitiatedWritesAndNotPrimary =
opCtx->writesAreReplicated() && !replCoord->canAcceptWritesForDatabase(opCtx, dbName);
if (userInitiatedWritesAndNotPrimary) {
return Result(
Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while dropping database " << dbName));
}
log() << "dropDatabase " << dbName << " - starting";
db->setDropPending(opCtx, true);
// If Database::dropCollectionEventIfSystem() fails, we should reset the drop-pending state
// on Database.
auto dropPendingGuard = MakeGuard([&db, opCtx] { db->setDropPending(opCtx, false); });
for (auto collection : *db) {
const auto& nss = collection->ns();
if (nss.isDropPendingNamespace() && replCoord->isReplEnabled() &&
opCtx->writesAreReplicated()) {
log() << "dropDatabase " << dbName << " - found drop-pending collection: " << nss;
latestDropPendingOpTime = std::max(
latestDropPendingOpTime, uassertStatusOK(nss.getDropPendingNamespaceOpTime()));
continue;
}
if (replCoord->isOplogDisabledFor(opCtx, nss) || nss.isSystemDotIndexes()) {
continue;
}
log() << "dropDatabase " << dbName << " - dropping collection: " << nss;
WriteUnitOfWork wunit(opCtx);
fassertStatusOK(40476, db->dropCollectionEvenIfSystem(opCtx, nss));
wunit.commit();
numCollectionsToDrop++;
}
dropPendingGuard.Dismiss();
// If there are no collection drops to wait for, we complete the drop database operation.
if (numCollectionsToDrop == 0U && latestDropPendingOpTime.isNull()) {
return Result(_finishDropDatabase(opCtx, dbName, db));
}
return Result(boost::none);
});
if (result) {
return *result;
}
// If waitForWriteConcern() returns an error or throws an exception, we should reset the
// drop-pending state on Database.
auto dropPendingGuardWhileAwaitingReplication = MakeGuard([dbName, opCtx] {
Lock::GlobalWrite lk(opCtx);
AutoGetDb autoDB(opCtx, dbName, MODE_X);
if (auto db = autoDB.getDb()) {
db->setDropPending(opCtx, false);
}
});
{
// Holding of any locks is disallowed while awaiting replication because this can
// potentially block for long time while doing network activity.
//
// Even though dropDatabase() does not explicitly acquire any locks before awaiting
// replication, it is possible that the caller of this function may already have acquired
// a lock. The applyOps command is an example of a dropDatabase() caller that does this.
// Therefore, we have to release any locks using a TempRelease RAII object.
//
// TODO: Remove the use of this TempRelease object when SERVER-29802 is completed.
// The work in SERVER-29802 will adjust the locking rules around applyOps operations and
// dropDatabase is expected to be one of the operations where we expect to no longer acquire
// the global lock.
Lock::TempRelease release(opCtx->lockState());
if (numCollectionsToDrop > 0U) {
auto status =
replCoord->awaitReplicationOfLastOpForClient(opCtx, kDropDatabaseWriteConcern)
.status;
if (!status.isOK()) {
return Status(status.code(),
str::stream() << "dropDatabase " << dbName << " failed waiting for "
<< numCollectionsToDrop
<< " collection drops to replicate: "
<< status.reason());
}
log() << "dropDatabase " << dbName << " - successfully dropped " << numCollectionsToDrop
<< " collections. dropping database";
} else {
invariant(!latestDropPendingOpTime.isNull());
auto status =
replCoord
->awaitReplication(opCtx, latestDropPendingOpTime, kDropDatabaseWriteConcern)
.status;
if (!status.isOK()) {
return Status(
status.code(),
str::stream()
<< "dropDatabase "
<< dbName
<< " failed waiting for pending collection drops (most recent drop optime: "
<< latestDropPendingOpTime.toString()
<< ") to replicate: "
<< status.reason());
}
log() << "dropDatabase " << dbName
<< " - pending collection drops completed. dropping database";
}
}
dropPendingGuardWhileAwaitingReplication.Dismiss();
return writeConflictRetry(opCtx, "dropDatabase_database", dbName, [&] {
Lock::GlobalWrite lk(opCtx);
AutoGetDb autoDB(opCtx, dbName, MODE_X);
if (auto db = autoDB.getDb()) {
return _finishDropDatabase(opCtx, dbName, db);
}
return Status(ErrorCodes::NamespaceNotFound,
str::stream() << "Could not drop database " << dbName
<< " because it does not exist after dropping "
<< numCollectionsToDrop
<< " collection(s).");
});
}
Status IndexAccessMethod::commitBulk(OperationContext* opCtx,
std::unique_ptr<BulkBuilder> bulk,
bool mayInterrupt,
bool dupsAllowed,
set<RecordId>* dupsToDrop,
bool assignTimestamp) {
// Do not track multikey path info for index builds.
ScopeGuard restartTracker =
MakeGuard([opCtx] { MultikeyPathTracker::get(opCtx).startTrackingMultikeyPathInfo(); });
if (!MultikeyPathTracker::get(opCtx).isTrackingMultikeyPathInfo()) {
restartTracker.Dismiss();
}
MultikeyPathTracker::get(opCtx).stopTrackingMultikeyPathInfo();
Timer timer;
std::unique_ptr<BulkBuilder::Sorter::Iterator> i(bulk->_sorter->done());
stdx::unique_lock<Client> lk(*opCtx->getClient());
ProgressMeterHolder pm(
CurOp::get(opCtx)->setMessage_inlock("Index Bulk Build: (2/3) btree bottom up",
"Index: (2/3) BTree Bottom Up Progress",
bulk->_keysInserted,
10));
lk.unlock();
std::unique_ptr<SortedDataBuilderInterface> builder;
writeConflictRetry(opCtx, "setting index multikey flag", "", [&] {
WriteUnitOfWork wunit(opCtx);
if (bulk->_everGeneratedMultipleKeys || isMultikeyFromPaths(bulk->_indexMultikeyPaths)) {
_btreeState->setMultikey(opCtx, bulk->_indexMultikeyPaths);
}
builder.reset(_newInterface->getBulkBuilder(opCtx, dupsAllowed));
if (assignTimestamp) {
fassertStatusOK(50705,
opCtx->recoveryUnit()->setTimestamp(
LogicalClock::get(opCtx)->getClusterTime().asTimestamp()));
}
wunit.commit();
});
while (i->more()) {
if (mayInterrupt) {
opCtx->checkForInterrupt();
}
WriteUnitOfWork wunit(opCtx);
// Improve performance in the btree-building phase by disabling rollback tracking.
// This avoids copying all the written bytes to a buffer that is only used to roll back.
// Note that this is safe to do, as this entire index-build-in-progress will be cleaned
// up by the index system.
opCtx->recoveryUnit()->setRollbackWritesDisabled();
// Get the next datum and add it to the builder.
BulkBuilder::Sorter::Data d = i->next();
Status status = builder->addKey(d.first, d.second);
if (!status.isOK()) {
// Overlong key that's OK to skip?
if (status.code() == ErrorCodes::KeyTooLong && ignoreKeyTooLong(opCtx)) {
continue;
}
// Check if this is a duplicate that's OK to skip
if (status.code() == ErrorCodes::DuplicateKey) {
invariant(!dupsAllowed); // shouldn't be getting DupKey errors if dupsAllowed.
if (dupsToDrop) {
dupsToDrop->insert(d.second);
continue;
}
}
return status;
}
// If we're here either it's a dup and we're cool with it or the addKey went just
// fine.
pm.hit();
if (assignTimestamp) {
fassertStatusOK(50704,
opCtx->recoveryUnit()->setTimestamp(
LogicalClock::get(opCtx)->getClusterTime().asTimestamp()));
}
wunit.commit();
}
pm.finished();
{
stdx::lock_guard<Client> lk(*opCtx->getClient());
CurOp::get(opCtx)->setMessage_inlock("Index Bulk Build: (3/3) btree-middle",
"Index: (3/3) BTree Middle Progress");
}
LOG(timer.seconds() > 10 ? 0 : 1) << "\t done building bottom layer, going to commit";
std::unique_ptr<TimestampBlock> tsBlock;
if (assignTimestamp) {
tsBlock = stdx::make_unique<TimestampBlock>(
opCtx, LogicalClock::get(opCtx)->getClusterTime().asTimestamp());
}
builder->commit(mayInterrupt);
return Status::OK();
}
Status IndexBuildInterceptor::drainWritesIntoIndex(OperationContext* opCtx,
const InsertDeleteOptions& options,
RecoveryUnit::ReadSource readSource) {
invariant(!opCtx->lockState()->inAWriteUnitOfWork());
// Callers may request to read at a specific timestamp so that no drained writes are timestamped
// earlier than their original write timestamp. Also ensure that leaving this function resets
// the ReadSource to its original value.
auto resetReadSourceGuard =
makeGuard([ opCtx, prevReadSource = opCtx->recoveryUnit()->getTimestampReadSource() ] {
opCtx->recoveryUnit()->abandonSnapshot();
opCtx->recoveryUnit()->setTimestampReadSource(prevReadSource);
});
if (readSource != RecoveryUnit::ReadSource::kUnset) {
opCtx->recoveryUnit()->abandonSnapshot();
opCtx->recoveryUnit()->setTimestampReadSource(readSource);
} else {
resetReadSourceGuard.dismiss();
}
// These are used for logging only.
int64_t totalDeleted = 0;
int64_t totalInserted = 0;
Timer timer;
const int64_t appliedAtStart = _numApplied;
// Set up the progress meter. This will never be completely accurate, because more writes can be
// read from the side writes table than are observed before draining.
static const char* curopMessage = "Index Build: draining writes received during build";
ProgressMeterHolder progress;
{
stdx::unique_lock<Client> lk(*opCtx->getClient());
progress.set(CurOp::get(opCtx)->setProgress_inlock(curopMessage));
}
// Force the progress meter to log at the end of every batch. By default, the progress meter
// only logs after a large number of calls to hit(), but since we batch inserts by up to
// 1000 records, progress would rarely be displayed.
progress->reset(_sideWritesCounter.load() - appliedAtStart /* total */,
3 /* secondsBetween */,
1 /* checkInterval */);
// Buffer operations into batches to insert per WriteUnitOfWork. Impose an upper limit on the
// number of documents and the total size of the batch.
const int32_t kBatchMaxSize = 1000;
const int64_t kBatchMaxBytes = BSONObjMaxInternalSize;
int64_t batchSizeBytes = 0;
std::vector<SideWriteRecord> batch;
batch.reserve(kBatchMaxSize);
// Hold on to documents that would exceed the per-batch memory limit. Always insert this first
// into the next batch.
boost::optional<SideWriteRecord> stashed;
auto cursor = _sideWritesTable->rs()->getCursor(opCtx);
bool atEof = false;
while (!atEof) {
opCtx->checkForInterrupt();
// Stashed records should be inserted into a batch first.
if (stashed) {
invariant(batch.empty());
batch.push_back(std::move(stashed.get()));
stashed.reset();
}
auto record = cursor->next();
if (record) {
RecordId currentRecordId = record->id;
BSONObj docOut = record->data.toBson().getOwned();
// If the total batch size in bytes would be too large, stash this document and let the
// current batch insert.
int objSize = docOut.objsize();
if (batchSizeBytes + objSize > kBatchMaxBytes) {
invariant(!stashed);
// Stash this document to be inserted in the next batch.
stashed.emplace(currentRecordId, std::move(docOut));
} else {
batchSizeBytes += objSize;
batch.emplace_back(currentRecordId, std::move(docOut));
// Continue if there is more room in the batch.
if (batch.size() < kBatchMaxSize) {
continue;
}
}
} else {
atEof = true;
if (batch.empty())
break;
}
invariant(!batch.empty());
cursor->save();
// If we are here, either we have reached the end of the table or the batch is full, so
// insert everything in one WriteUnitOfWork, and delete each inserted document from the side
// writes table.
auto status = writeConflictRetry(opCtx, "index build drain", _indexCatalogEntry->ns(), [&] {
WriteUnitOfWork wuow(opCtx);
for (auto& operation : batch) {
auto status =
_applyWrite(opCtx, operation.second, options, &totalInserted, &totalDeleted);
if (!status.isOK()) {
return status;
}
// Delete the document from the table as soon as it has been inserted into the
// index. This ensures that no key is ever inserted twice and no keys are skipped.
_sideWritesTable->rs()->deleteRecord(opCtx, operation.first);
}
// For rollback to work correctly, these writes need to be timestamped. The actual time
// is not important, as long as it not older than the most recent visible side write.
IndexTimestampHelper::setGhostCommitTimestampForWrite(
opCtx, NamespaceString(_indexCatalogEntry->ns()));
wuow.commit();
return Status::OK();
});
if (!status.isOK()) {
return status;
}
progress->hit(batch.size());
// Lock yielding will only happen if we are holding intent locks.
_tryYield(opCtx);
cursor->restore();
// Account for more writes coming in during a batch.
progress->setTotalWhileRunning(_sideWritesCounter.loadRelaxed() - appliedAtStart);
_numApplied += batch.size();
batch.clear();
batchSizeBytes = 0;
}
progress->finished();
int logLevel = (_numApplied - appliedAtStart > 0) ? 0 : 1;
LOG(logLevel) << "index build: drain applied " << (_numApplied - appliedAtStart)
<< " side writes (inserted: " << totalInserted << ", deleted: " << totalDeleted
<< ") for '" << _indexCatalogEntry->descriptor()->indexName() << "' in "
<< timer.millis() << " ms";
return Status::OK();
}
Status applyOps(OperationContext* opCtx,
const std::string& dbName,
const BSONObj& applyOpCmd,
BSONObjBuilder* result) {
bool allowAtomic = false;
uassertStatusOK(
bsonExtractBooleanFieldWithDefault(applyOpCmd, "allowAtomic", true, &allowAtomic));
auto areOpsCrudOnly = _areOpsCrudOnly(applyOpCmd);
auto isAtomic = allowAtomic && areOpsCrudOnly;
auto hasPrecondition = _hasPrecondition(applyOpCmd);
boost::optional<Lock::GlobalWrite> globalWriteLock;
boost::optional<Lock::DBLock> dbWriteLock;
// There's only one case where we are allowed to take the database lock instead of the global
// lock - no preconditions; only CRUD ops; and non-atomic mode.
if (!hasPrecondition && areOpsCrudOnly && !allowAtomic) {
dbWriteLock.emplace(opCtx, dbName, MODE_X);
} else {
globalWriteLock.emplace(opCtx);
}
bool userInitiatedWritesAndNotPrimary = opCtx->writesAreReplicated() &&
!repl::getGlobalReplicationCoordinator()->canAcceptWritesForDatabase(opCtx, dbName);
if (userInitiatedWritesAndNotPrimary)
return Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while applying ops to database " << dbName);
if (hasPrecondition) {
auto status = _checkPrecondition(opCtx, applyOpCmd, result);
if (!status.isOK()) {
return status;
}
}
int numApplied = 0;
if (!isAtomic)
return _applyOps(opCtx, dbName, applyOpCmd, result, &numApplied);
// Perform write ops atomically
invariant(globalWriteLock);
try {
writeConflictRetry(opCtx, "applyOps", dbName, [&] {
BSONObjBuilder intermediateResult;
WriteUnitOfWork wunit(opCtx);
numApplied = 0;
{
// Suppress replication for atomic operations until end of applyOps.
repl::UnreplicatedWritesBlock uwb(opCtx);
uassertStatusOK(
_applyOps(opCtx, dbName, applyOpCmd, &intermediateResult, &numApplied));
}
// Generate oplog entry for all atomic ops collectively.
if (opCtx->writesAreReplicated()) {
// We want this applied atomically on slaves so we rewrite the oplog entry without
// the pre-condition for speed.
BSONObjBuilder cmdBuilder;
for (auto elem : applyOpCmd) {
auto name = elem.fieldNameStringData();
if (name == kPreconditionFieldName)
continue;
if (name == "bypassDocumentValidation")
continue;
cmdBuilder.append(elem);
}
const BSONObj cmdRewritten = cmdBuilder.done();
auto opObserver = getGlobalServiceContext()->getOpObserver();
invariant(opObserver);
opObserver->onApplyOps(opCtx, dbName, cmdRewritten);
}
wunit.commit();
result->appendElements(intermediateResult.obj());
});
} catch (const DBException& ex) {
if (ex.getCode() == ErrorCodes::NamespaceNotFound) {
// Retry in non-atomic mode, since MMAP cannot implicitly create a new database
// within an active WriteUnitOfWork.
return _applyOps(opCtx, dbName, applyOpCmd, result, &numApplied);
}
BSONArrayBuilder ab;
++numApplied;
for (int j = 0; j < numApplied; j++)
ab.append(false);
result->append("applied", numApplied);
result->append("code", ex.getCode());
result->append("codeName", ErrorCodes::errorString(ErrorCodes::fromInt(ex.getCode())));
result->append("errmsg", ex.what());
result->append("results", ab.arr());
return Status(ErrorCodes::UnknownError, ex.what());
}
return Status::OK();
}
