ChunkVersion forceShardFilteringMetadataRefresh(OperationContext* opCtx,
const NamespaceString& nss,
bool forceRefreshFromThisThread) {
invariant(!opCtx->lockState()->isLocked());
invariant(!opCtx->getClient()->isInDirectClient());
auto const shardingState = ShardingState::get(opCtx);
invariant(shardingState->canAcceptShardedCommands());
const auto routingInfo =
uassertStatusOK(Grid::get(opCtx)->catalogCache()->getCollectionRoutingInfoWithRefresh(
opCtx, nss, forceRefreshFromThisThread));
const auto cm = routingInfo.cm();
if (!cm) {
// No chunk manager, so unsharded.
// Exclusive collection lock needed since we're now changing the metadata
AutoGetCollection autoColl(opCtx, nss, MODE_IX, MODE_X);
auto css = CollectionShardingState::get(opCtx, nss);
css->refreshMetadata(opCtx, nullptr);
return ChunkVersion::UNSHARDED();
}
{
AutoGetCollection autoColl(opCtx, nss, MODE_IS);
auto metadata = CollectionShardingState::get(opCtx, nss)->getMetadata(opCtx);
// We already have newer version
if (metadata && metadata->getCollVersion().epoch() == cm->getVersion().epoch() &&
metadata->getCollVersion() >= cm->getVersion()) {
LOG(1) << "Skipping refresh of metadata for " << nss << " "
<< metadata->getCollVersion() << " with an older " << cm->getVersion();
return metadata->getShardVersion();
}
}
// Exclusive collection lock needed since we're now changing the metadata
AutoGetCollection autoColl(opCtx, nss, MODE_IX, MODE_X);
auto css = CollectionShardingState::get(opCtx, nss);
auto metadata = css->getMetadata(opCtx);
// We already have newer version
if (metadata && metadata->getCollVersion().epoch() == cm->getVersion().epoch() &&
metadata->getCollVersion() >= cm->getVersion()) {
LOG(1) << "Skipping refresh of metadata for " << nss << " " << metadata->getCollVersion()
<< " with an older " << cm->getVersion();
return metadata->getShardVersion();
}
std::unique_ptr<CollectionMetadata> newCollectionMetadata =
stdx::make_unique<CollectionMetadata>(cm, shardingState->getShardName());
css->refreshMetadata(opCtx, std::move(newCollectionMetadata));
return css->getMetadata(opCtx)->getShardVersion();
}
void ChunkManagerTargeter::noteStaleResponse(const ShardEndpoint& endpoint,
const StaleConfigInfo& staleInfo) {
dassert(!_needsTargetingRefresh);
ChunkVersion remoteShardVersion;
if (!staleInfo.getVersionWanted()) {
// If we don't have a vWanted sent, assume the version is higher than our current version.
remoteShardVersion = getShardVersion(*_routingInfo, endpoint.shardName);
remoteShardVersion.incMajor();
} else {
remoteShardVersion = *staleInfo.getVersionWanted();
}
ShardVersionMap::iterator it = _remoteShardVersions.find(endpoint.shardName);
if (it == _remoteShardVersions.end()) {
_remoteShardVersions.insert(std::make_pair(endpoint.shardName, remoteShardVersion));
} else {
ChunkVersion& previouslyNotedVersion = it->second;
if (previouslyNotedVersion.epoch() == remoteShardVersion.epoch()) {
if (previouslyNotedVersion.isOlderThan(remoteShardVersion)) {
previouslyNotedVersion = remoteShardVersion;
}
} else {
// Epoch changed midway while applying the batch so set the version to something
// unique
// and non-existent to force a reload when refreshIsNeeded is called.
previouslyNotedVersion = ChunkVersion::IGNORED();
}
}
}
Status onShardVersionMismatch(OperationContext* opCtx,
const NamespaceString& nss,
ChunkVersion shardVersionReceived,
bool forceRefreshFromThisThread) noexcept {
invariant(!opCtx->lockState()->isLocked());
invariant(!opCtx->getClient()->isInDirectClient());
auto const shardingState = ShardingState::get(opCtx);
invariant(shardingState->canAcceptShardedCommands());
LOG(2) << "Metadata refresh requested for " << nss.ns() << " at shard version "
<< shardVersionReceived;
ShardingStatistics::get(opCtx).countStaleConfigErrors.addAndFetch(1);
// Ensure any ongoing migrations have completed before trying to do the refresh. This wait is
// just an optimization so that MongoS does not exhaust its maximum number of StaleConfig retry
// attempts while the migration is being committed.
try {
auto& oss = OperationShardingState::get(opCtx);
oss.waitForMigrationCriticalSectionSignal(opCtx);
} catch (const DBException& ex) {
return ex.toStatus();
}
const auto currentShardVersion = [&] {
AutoGetCollection autoColl(opCtx, nss, MODE_IS);
const auto currentMetadata = CollectionShardingState::get(opCtx, nss)->getMetadata(opCtx);
if (currentMetadata) {
return currentMetadata->getShardVersion();
}
return ChunkVersion::UNSHARDED();
}();
if (currentShardVersion.epoch() == shardVersionReceived.epoch() &&
currentShardVersion.majorVersion() >= shardVersionReceived.majorVersion()) {
// Don't need to remotely reload if we're in the same epoch and the requested version is
// smaller than the one we know about. This means that the remote side is behind.
return Status::OK();
}
try {
forceShardFilteringMetadataRefresh(opCtx, nss, forceRefreshFromThisThread);
return Status::OK();
} catch (const DBException& ex) {
log() << "Failed to refresh metadata for collection" << nss << causedBy(redact(ex));
return ex.toStatus();
}
}
bool CollectionShardingState::_checkShardVersionOk(OperationContext* txn,
string* errmsg,
ChunkVersion* expectedShardVersion,
ChunkVersion* actualShardVersion) {
Client* client = txn->getClient();
// Operations using the DBDirectClient are unversioned.
if (client->isInDirectClient()) {
return true;
}
if (!repl::ReplicationCoordinator::get(txn)->canAcceptWritesForDatabase(_nss.db())) {
// Right now connections to secondaries aren't versioned at all.
return true;
}
const auto& oss = OperationShardingState::get(txn);
// If there is a version attached to the OperationContext, use it as the received version.
// Otherwise, get the received version from the ShardedConnectionInfo.
if (oss.hasShardVersion()) {
*expectedShardVersion = oss.getShardVersion(_nss);
} else {
ShardedConnectionInfo* info = ShardedConnectionInfo::get(client, false);
if (!info) {
// There is no shard version information on either 'txn' or 'client'. This means that
// the operation represented by 'txn' is unversioned, and the shard version is always OK
// for unversioned operations.
return true;
}
*expectedShardVersion = info->getVersion(_nss.ns());
}
if (ChunkVersion::isIgnoredVersion(*expectedShardVersion)) {
return true;
}
// Set this for error messaging purposes before potentially returning false.
auto metadata = getMetadata();
*actualShardVersion = metadata ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
if (_sourceMgr && _sourceMgr->getMigrationCriticalSectionSignal()) {
*errmsg = str::stream() << "migration commit in progress for " << _nss.ns();
// Set migration critical section on operation sharding state: operation will wait for the
// migration to finish before returning failure and retrying.
OperationShardingState::get(txn).setMigrationCriticalSectionSignal(
_sourceMgr->getMigrationCriticalSectionSignal());
return false;
}
if (expectedShardVersion->isWriteCompatibleWith(*actualShardVersion)) {
return true;
}
//
// Figure out exactly why not compatible, send appropriate error message
// The versions themselves are returned in the error, so not needed in messages here
//
// Check epoch first, to send more meaningful message, since other parameters probably won't
// match either.
if (actualShardVersion->epoch() != expectedShardVersion->epoch()) {
*errmsg = str::stream() << "version epoch mismatch detected for " << _nss.ns() << ", "
<< "the collection may have been dropped and recreated";
return false;
}
if (!actualShardVersion->isSet() && expectedShardVersion->isSet()) {
*errmsg = str::stream() << "this shard no longer contains chunks for " << _nss.ns() << ", "
<< "the collection may have been dropped";
return false;
}
if (actualShardVersion->isSet() && !expectedShardVersion->isSet()) {
*errmsg = str::stream() << "this shard contains versioned chunks for " << _nss.ns() << ", "
<< "but no version set in request";
return false;
}
if (actualShardVersion->majorVersion() != expectedShardVersion->majorVersion()) {
// Could be > or < - wanted is > if this is the source of a migration, wanted < if this is
// the target of a migration
*errmsg = str::stream() << "version mismatch detected for " << _nss.ns();
return false;
}
// Those are all the reasons the versions can mismatch
MONGO_UNREACHABLE;
}
bool CollectionShardingState::_checkShardVersionOk(OperationContext* opCtx,
string* errmsg,
ChunkVersion* expectedShardVersion,
ChunkVersion* actualShardVersion) {
Client* client = opCtx->getClient();
auto& oss = OperationShardingState::get(opCtx);
// If there is a version attached to the OperationContext, use it as the received version.
// Otherwise, get the received version from the ShardedConnectionInfo.
if (oss.hasShardVersion()) {
*expectedShardVersion = oss.getShardVersion(_nss);
} else {
ShardedConnectionInfo* info = ShardedConnectionInfo::get(client, false);
if (!info) {
// There is no shard version information on either 'opCtx' or 'client'. This means that
// the operation represented by 'opCtx' is unversioned, and the shard version is always
// OK for unversioned operations.
return true;
}
*expectedShardVersion = info->getVersion(_nss.ns());
}
// An operation with read concern 'available' should never have shardVersion set.
invariant(repl::ReadConcernArgs::get(opCtx).getLevel() !=
repl::ReadConcernLevel::kAvailableReadConcern);
if (ChunkVersion::isIgnoredVersion(*expectedShardVersion)) {
return true;
}
// Set this for error messaging purposes before potentially returning false.
auto metadata = getMetadata();
*actualShardVersion = metadata ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
if (_sourceMgr) {
const bool isReader = !opCtx->lockState()->isWriteLocked();
auto criticalSectionSignal = _sourceMgr->getMigrationCriticalSectionSignal(isReader);
if (criticalSectionSignal) {
*errmsg = str::stream() << "migration commit in progress for " << _nss.ns();
// Set migration critical section on operation sharding state: operation will wait for
// the migration to finish before returning failure and retrying.
oss.setMigrationCriticalSectionSignal(criticalSectionSignal);
return false;
}
}
if (expectedShardVersion->isWriteCompatibleWith(*actualShardVersion)) {
return true;
}
//
// Figure out exactly why not compatible, send appropriate error message
// The versions themselves are returned in the error, so not needed in messages here
//
// Check epoch first, to send more meaningful message, since other parameters probably won't
// match either.
if (actualShardVersion->epoch() != expectedShardVersion->epoch()) {
*errmsg = str::stream() << "version epoch mismatch detected for " << _nss.ns() << ", "
<< "the collection may have been dropped and recreated";
return false;
}
if (!actualShardVersion->isSet() && expectedShardVersion->isSet()) {
*errmsg = str::stream() << "this shard no longer contains chunks for " << _nss.ns() << ", "
<< "the collection may have been dropped";
return false;
}
if (actualShardVersion->isSet() && !expectedShardVersion->isSet()) {
*errmsg = str::stream() << "this shard contains versioned chunks for " << _nss.ns() << ", "
<< "but no version set in request";
return false;
}
if (actualShardVersion->majorVersion() != expectedShardVersion->majorVersion()) {
// Could be > or < - wanted is > if this is the source of a migration, wanted < if this is
// the target of a migration
*errmsg = str::stream() << "version mismatch detected for " << _nss.ns();
return false;
}
// Those are all the reasons the versions can mismatch
MONGO_UNREACHABLE;
}
void CollectionShardingState::checkShardVersionOrThrow(OperationContext* opCtx) {
const auto optReceivedShardVersion = getOperationReceivedVersion(opCtx, _nss);
if (!optReceivedShardVersion)
return;
const auto& receivedShardVersion = *optReceivedShardVersion;
if (ChunkVersion::isIgnoredVersion(receivedShardVersion)) {
return;
}
// An operation with read concern 'available' should never have shardVersion set.
invariant(repl::ReadConcernArgs::get(opCtx).getLevel() !=
repl::ReadConcernLevel::kAvailableReadConcern);
const auto metadata = getCurrentMetadata();
const auto wantedShardVersion =
metadata->isSharded() ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
auto criticalSectionSignal = [&] {
auto csrLock = CSRLock::lock(opCtx, this);
return _critSec.getSignal(opCtx->lockState()->isWriteLocked()
? ShardingMigrationCriticalSection::kWrite
: ShardingMigrationCriticalSection::kRead);
}();
if (criticalSectionSignal) {
// Set migration critical section on operation sharding state: operation will wait for the
// migration to finish before returning failure and retrying.
auto& oss = OperationShardingState::get(opCtx);
oss.setMigrationCriticalSectionSignal(criticalSectionSignal);
uasserted(StaleConfigInfo(_nss, receivedShardVersion, wantedShardVersion),
str::stream() << "migration commit in progress for " << _nss.ns());
}
if (receivedShardVersion.isWriteCompatibleWith(wantedShardVersion)) {
return;
}
//
// Figure out exactly why not compatible, send appropriate error message
// The versions themselves are returned in the error, so not needed in messages here
//
StaleConfigInfo sci(_nss, receivedShardVersion, wantedShardVersion);
uassert(std::move(sci),
str::stream() << "epoch mismatch detected for " << _nss.ns() << ", "
<< "the collection may have been dropped and recreated",
wantedShardVersion.epoch() == receivedShardVersion.epoch());
if (!wantedShardVersion.isSet() && receivedShardVersion.isSet()) {
uasserted(std::move(sci),
str::stream() << "this shard no longer contains chunks for " << _nss.ns() << ", "
<< "the collection may have been dropped");
}
if (wantedShardVersion.isSet() && !receivedShardVersion.isSet()) {
uasserted(std::move(sci),
str::stream() << "this shard contains chunks for " << _nss.ns() << ", "
<< "but the client expects unsharded collection");
}
if (wantedShardVersion.majorVersion() != receivedShardVersion.majorVersion()) {
// Could be > or < - wanted is > if this is the source of a migration, wanted < if this is
// the target of a migration
uasserted(std::move(sci), str::stream() << "version mismatch detected for " << _nss.ns());
}
// Those are all the reasons the versions can mismatch
MONGO_UNREACHABLE;
}
