bool run(OperationContext* opCtx,
const std::string& dbname,
const BSONObj& cmdObj,
BSONObjBuilder& result) override {
uassert(ErrorCodes::IllegalOperation,
"_configsvrDropCollection can only be run on config servers",
serverGlobalParams.clusterRole == ClusterRole::ConfigServer);
// Set the operation context read concern level to local for reads into the config database.
repl::ReadConcernArgs::get(opCtx) =
repl::ReadConcernArgs(repl::ReadConcernLevel::kLocalReadConcern);
const NamespaceString nss(parseNs(dbname, cmdObj));
uassert(ErrorCodes::InvalidOptions,
str::stream() << "dropCollection must be called with majority writeConcern, got "
<< cmdObj,
opCtx->getWriteConcern().wMode == WriteConcernOptions::kMajority);
Seconds waitFor(DistLockManager::kDefaultLockTimeout);
MONGO_FAIL_POINT_BLOCK(setDropCollDistLockWait, customWait) {
const BSONObj& data = customWait.getData();
waitFor = Seconds(data["waitForSecs"].numberInt());
}
auto const catalogClient = Grid::get(opCtx)->catalogClient();
auto scopedDbLock =
ShardingCatalogManager::get(opCtx)->serializeCreateOrDropDatabase(opCtx, nss.db());
auto scopedCollLock =
ShardingCatalogManager::get(opCtx)->serializeCreateOrDropCollection(opCtx, nss);
auto dbDistLock = uassertStatusOK(
catalogClient->getDistLockManager()->lock(opCtx, nss.db(), "dropCollection", waitFor));
auto collDistLock = uassertStatusOK(
catalogClient->getDistLockManager()->lock(opCtx, nss.ns(), "dropCollection", waitFor));
ON_BLOCK_EXIT(
[opCtx, nss] { Grid::get(opCtx)->catalogCache()->invalidateShardedCollection(nss); });
staleExceptionRetry(
opCtx, "_configsvrDropCollection", [&] { _dropCollection(opCtx, nss); });
return true;
}
bool CatalogManagerReplicaSet::runUserManagementWriteCommand(const std::string& commandName,
const std::string& dbname,
const BSONObj& cmdObj,
BSONObjBuilder* result) {
auto scopedDistLock = getDistLockManager()->lock("authorizationData", commandName, Seconds{5});
if (!scopedDistLock.isOK()) {
return Command::appendCommandStatus(*result, scopedDistLock.getStatus());
}
auto response = _runConfigServerCommandWithNotMasterRetries(dbname, cmdObj);
if (!response.isOK()) {
return Command::appendCommandStatus(*result, response.getStatus());
}
result->appendElements(response.getValue());
return Command::getStatusFromCommandResult(response.getValue()).isOK();
}
Status CatalogManagerReplicaSet::shardCollection(OperationContext* txn,
const string& ns,
const ShardKeyPattern& fieldsAndOrder,
bool unique,
const vector<BSONObj>& initPoints,
const set<ShardId>& initShardIds) {
// Lock the collection globally so that no other mongos can try to shard or drop the collection
// at the same time.
auto scopedDistLock = getDistLockManager()->lock(ns, "shardCollection");
if (!scopedDistLock.isOK()) {
return scopedDistLock.getStatus();
}
StatusWith<DatabaseType> status = getDatabase(nsToDatabase(ns));
if (!status.isOK()) {
return status.getStatus();
}
DatabaseType dbt = status.getValue();
ShardId dbPrimaryShardId = dbt.getPrimary();
const auto primaryShard = grid.shardRegistry()->getShard(dbPrimaryShardId);
{
// In 3.0 and prior we include this extra safety check that the collection is not getting
// sharded concurrently by two different mongos instances. It is not 100%-proof, but it
// reduces the chance that two invocations of shard collection will step on each other's
// toes. Now we take the distributed lock so going forward this check won't be necessary
// but we leave it around for compatibility with other mongoses from 3.0.
// TODO(spencer): Remove this after 3.2 ships.
const auto configShard = grid.shardRegistry()->getShard("config");
const auto readHost = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHost.isOK()) {
return readHost.getStatus();
}
auto countStatus = _runCountCommand(
readHost.getValue(), NamespaceString(ChunkType::ConfigNS), BSON(ChunkType::ns(ns)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::AlreadyInitialized,
str::stream() << "collection " << ns << " already sharded with "
<< countStatus.getValue() << " chunks.");
}
}
// Record start in changelog
{
BSONObjBuilder collectionDetail;
collectionDetail.append("shardKey", fieldsAndOrder.toBSON());
collectionDetail.append("collection", ns);
collectionDetail.append("primary", primaryShard->toString());
{
BSONArrayBuilder initialShards(collectionDetail.subarrayStart("initShards"));
for (const ShardId& shardId : initShardIds) {
initialShards.append(shardId);
}
}
collectionDetail.append("numChunks", static_cast<int>(initPoints.size() + 1));
logChange(txn->getClient()->clientAddress(true),
"shardCollection.start",
ns,
collectionDetail.obj());
}
ChunkManagerPtr manager(new ChunkManager(ns, fieldsAndOrder, unique));
manager->createFirstChunks(dbPrimaryShardId, &initPoints, &initShardIds);
manager->loadExistingRanges(nullptr);
CollectionInfo collInfo;
collInfo.useChunkManager(manager);
collInfo.save(ns);
manager->reload(true);
// TODO(spencer) SERVER-19319: Send setShardVersion to primary shard so it knows to start
// rejecting unversioned writes.
BSONObj finishDetail = BSON("version"
<< ""); // TODO(spencer) SERVER-19319 Report actual version used
logChange(txn->getClient()->clientAddress(true), "shardCollection", ns, finishDetail);
return Status::OK();
}
Status CatalogManagerReplicaSet::dropCollection(OperationContext* txn, const NamespaceString& ns) {
logChange(
txn, txn->getClient()->clientAddress(true), "dropCollection.start", ns.ns(), BSONObj());
vector<ShardType> allShards;
Status status = getAllShards(txn, &allShards);
if (!status.isOK()) {
return status;
}
LOG(1) << "dropCollection " << ns << " started";
// Lock the collection globally so that split/migrate cannot run
stdx::chrono::seconds waitFor(2);
MONGO_FAIL_POINT_BLOCK(setDropCollDistLockWait, customWait) {
const BSONObj& data = customWait.getData();
waitFor = stdx::chrono::seconds(data["waitForSecs"].numberInt());
}
const stdx::chrono::milliseconds lockTryInterval(500);
auto scopedDistLock = getDistLockManager()->lock(ns.ns(), "drop", waitFor, lockTryInterval);
if (!scopedDistLock.isOK()) {
return scopedDistLock.getStatus();
}
LOG(1) << "dropCollection " << ns << " locked";
std::map<string, BSONObj> errors;
auto* shardRegistry = grid.shardRegistry();
for (const auto& shardEntry : allShards) {
auto dropResult = shardRegistry->runCommandWithNotMasterRetries(
txn, shardEntry.getName(), ns.db().toString(), BSON("drop" << ns.coll()));
if (!dropResult.isOK()) {
return dropResult.getStatus();
}
auto dropStatus = getStatusFromCommandResult(dropResult.getValue());
if (!dropStatus.isOK()) {
if (dropStatus.code() == ErrorCodes::NamespaceNotFound) {
continue;
}
errors.emplace(shardEntry.getHost(), dropResult.getValue());
}
}
if (!errors.empty()) {
StringBuilder sb;
sb << "Dropping collection failed on the following hosts: ";
for (auto it = errors.cbegin(); it != errors.cend(); ++it) {
if (it != errors.cbegin()) {
sb << ", ";
}
sb << it->first << ": " << it->second;
}
return {ErrorCodes::OperationFailed, sb.str()};
}
LOG(1) << "dropCollection " << ns << " shard data deleted";
// Remove chunk data
Status result = remove(txn, ChunkType::ConfigNS, BSON(ChunkType::ns(ns.ns())), 0, nullptr);
if (!result.isOK()) {
return result;
}
LOG(1) << "dropCollection " << ns << " chunk data deleted";
// Mark the collection as dropped
CollectionType coll;
coll.setNs(ns);
coll.setDropped(true);
coll.setEpoch(ChunkVersion::DROPPED().epoch());
coll.setUpdatedAt(grid.shardRegistry()->getNetwork()->now());
result = updateCollection(txn, ns.ns(), coll);
if (!result.isOK()) {
return result;
}
LOG(1) << "dropCollection " << ns << " collection marked as dropped";
for (const auto& shardEntry : allShards) {
SetShardVersionRequest ssv = SetShardVersionRequest::makeForVersioningNoPersist(
grid.shardRegistry()->getConfigServerConnectionString(),
shardEntry.getName(),
fassertStatusOK(28781, ConnectionString::parse(shardEntry.getHost())),
ns,
ChunkVersion::DROPPED(),
true);
auto ssvResult = shardRegistry->runCommandWithNotMasterRetries(
txn, shardEntry.getName(), "admin", ssv.toBSON());
if (!ssvResult.isOK()) {
return ssvResult.getStatus();
}
auto ssvStatus = getStatusFromCommandResult(ssvResult.getValue());
if (!ssvStatus.isOK()) {
return ssvStatus;
}
auto unsetShardingStatus = shardRegistry->runCommandWithNotMasterRetries(
txn, shardEntry.getName(), "admin", BSON("unsetSharding" << 1));
if (!unsetShardingStatus.isOK()) {
return unsetShardingStatus.getStatus();
}
auto unsetShardingResult = getStatusFromCommandResult(unsetShardingStatus.getValue());
if (!unsetShardingResult.isOK()) {
return unsetShardingResult;
}
}
LOG(1) << "dropCollection " << ns << " completed";
logChange(txn, txn->getClient()->clientAddress(true), "dropCollection", ns.ns(), BSONObj());
return Status::OK();
}
Status CatalogManagerReplicaSet::shardCollection(OperationContext* txn,
const string& ns,
const ShardKeyPattern& fieldsAndOrder,
bool unique,
const vector<BSONObj>& initPoints,
const set<ShardId>& initShardIds) {
// Lock the collection globally so that no other mongos can try to shard or drop the collection
// at the same time.
auto scopedDistLock = getDistLockManager()->lock(ns, "shardCollection");
if (!scopedDistLock.isOK()) {
return scopedDistLock.getStatus();
}
auto status = getDatabase(txn, nsToDatabase(ns));
if (!status.isOK()) {
return status.getStatus();
}
ShardId dbPrimaryShardId = status.getValue().value.getPrimary();
const auto primaryShard = grid.shardRegistry()->getShard(txn, dbPrimaryShardId);
{
// In 3.0 and prior we include this extra safety check that the collection is not getting
// sharded concurrently by two different mongos instances. It is not 100%-proof, but it
// reduces the chance that two invocations of shard collection will step on each other's
// toes. Now we take the distributed lock so going forward this check won't be necessary
// but we leave it around for compatibility with other mongoses from 3.0.
// TODO(spencer): Remove this after 3.2 ships.
auto countStatus = _runCountCommandOnConfig(
txn, NamespaceString(ChunkType::ConfigNS), BSON(ChunkType::ns(ns)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::AlreadyInitialized,
str::stream() << "collection " << ns << " already sharded with "
<< countStatus.getValue() << " chunks.");
}
}
// Record start in changelog
{
BSONObjBuilder collectionDetail;
collectionDetail.append("shardKey", fieldsAndOrder.toBSON());
collectionDetail.append("collection", ns);
collectionDetail.append("primary", primaryShard->toString());
{
BSONArrayBuilder initialShards(collectionDetail.subarrayStart("initShards"));
for (const ShardId& shardId : initShardIds) {
initialShards.append(shardId);
}
}
collectionDetail.append("numChunks", static_cast<int>(initPoints.size() + 1));
logChange(txn,
txn->getClient()->clientAddress(true),
"shardCollection.start",
ns,
collectionDetail.obj());
}
shared_ptr<ChunkManager> manager(new ChunkManager(ns, fieldsAndOrder, unique));
manager->createFirstChunks(txn, dbPrimaryShardId, &initPoints, &initShardIds);
manager->loadExistingRanges(txn, nullptr);
CollectionInfo collInfo;
collInfo.useChunkManager(manager);
collInfo.save(txn, ns);
manager->reload(txn, true);
// Tell the primary mongod to refresh its data
// TODO: Think the real fix here is for mongos to just
// assume that all collections are sharded, when we get there
SetShardVersionRequest ssv = SetShardVersionRequest::makeForVersioningNoPersist(
grid.shardRegistry()->getConfigServerConnectionString(),
dbPrimaryShardId,
primaryShard->getConnString(),
NamespaceString(ns),
manager->getVersion(),
true);
auto ssvStatus = grid.shardRegistry()->runCommandWithNotMasterRetries(
txn, dbPrimaryShardId, "admin", ssv.toBSON());
if (!ssvStatus.isOK()) {
warning() << "could not update initial version of " << ns << " on shard primary "
<< dbPrimaryShardId << ssvStatus.getStatus();
}
logChange(txn,
txn->getClient()->clientAddress(true),
"shardCollection",
ns,
BSON("version" << manager->getVersion().toString()));
return Status::OK();
}
Status CatalogManagerReplicaSet::dropCollection(OperationContext* txn, const NamespaceString& ns) {
logChange(
txn, txn->getClient()->clientAddress(true), "dropCollection.start", ns.ns(), BSONObj());
vector<ShardType> allShards;
Status status = getAllShards(txn, &allShards);
if (!status.isOK()) {
return status;
}
LOG(1) << "dropCollection " << ns << " started";
// Lock the collection globally so that split/migrate cannot run
auto scopedDistLock = getDistLockManager()->lock(ns.ns(), "drop");
if (!scopedDistLock.isOK()) {
return scopedDistLock.getStatus();
}
LOG(1) << "dropCollection " << ns << " locked";
std::map<string, BSONObj> errors;
auto* shardRegistry = grid.shardRegistry();
for (const auto& shardEntry : allShards) {
auto dropResult = shardRegistry->runCommandWithNotMasterRetries(
txn, shardEntry.getName(), ns.db().toString(), BSON("drop" << ns.coll()));
if (!dropResult.isOK()) {
return dropResult.getStatus();
}
auto dropStatus = getStatusFromCommandResult(dropResult.getValue());
if (!dropStatus.isOK()) {
if (dropStatus.code() == ErrorCodes::NamespaceNotFound) {
continue;
}
errors.emplace(shardEntry.getHost(), dropResult.getValue());
}
}
if (!errors.empty()) {
StringBuilder sb;
sb << "Dropping collection failed on the following hosts: ";
for (auto it = errors.cbegin(); it != errors.cend(); ++it) {
if (it != errors.cbegin()) {
sb << ", ";
}
sb << it->first << ": " << it->second;
}
return {ErrorCodes::OperationFailed, sb.str()};
}
LOG(1) << "dropCollection " << ns << " shard data deleted";
// Remove chunk data
Status result = remove(txn, ChunkType::ConfigNS, BSON(ChunkType::ns(ns.ns())), 0, nullptr);
if (!result.isOK()) {
return result;
}
LOG(1) << "dropCollection " << ns << " chunk data deleted";
// Mark the collection as dropped
CollectionType coll;
coll.setNs(ns);
coll.setDropped(true);
coll.setEpoch(ChunkVersion::DROPPED().epoch());
coll.setUpdatedAt(grid.shardRegistry()->getNetwork()->now());
result = updateCollection(txn, ns.ns(), coll);
if (!result.isOK()) {
return result;
}
LOG(1) << "dropCollection " << ns << " collection marked as dropped";
// We just called updateCollection above and this would have advanced the config op time, so use
// the latest value. On the MongoD side, we need to load the latest config metadata, which
// indicates that the collection was dropped.
const ChunkVersionAndOpTime droppedVersion(ChunkVersion::DROPPED(),
grid.shardRegistry()->getConfigOpTime());
for (const auto& shardEntry : allShards) {
SetShardVersionRequest ssv = SetShardVersionRequest::makeForVersioningNoPersist(
grid.shardRegistry()->getConfigServerConnectionString(),
shardEntry.getName(),
fassertStatusOK(28781, ConnectionString::parse(shardEntry.getHost())),
ns,
droppedVersion,
true);
auto ssvResult = shardRegistry->runCommandWithNotMasterRetries(
txn, shardEntry.getName(), "admin", ssv.toBSON());
if (!ssvResult.isOK()) {
return ssvResult.getStatus();
}
auto ssvStatus = getStatusFromCommandResult(ssvResult.getValue());
if (!ssvStatus.isOK()) {
return ssvStatus;
}
auto unsetShardingStatus = shardRegistry->runCommandWithNotMasterRetries(
txn, shardEntry.getName(), "admin", BSON("unsetSharding" << 1));
if (!unsetShardingStatus.isOK()) {
return unsetShardingStatus.getStatus();
}
auto unsetShardingResult = getStatusFromCommandResult(unsetShardingStatus.getValue());
if (!unsetShardingResult.isOK()) {
return unsetShardingResult;
}
}
LOG(1) << "dropCollection " << ns << " completed";
logChange(txn, txn->getClient()->clientAddress(true), "dropCollection", ns.ns(), BSONObj());
return Status::OK();
}
