Status Database::renameCollection( OperationContext* txn,
const StringData& fromNS,
const StringData& toNS,
bool stayTemp ) {
audit::logRenameCollection( currentClient.get(), fromNS, toNS );
{ // remove anything cached
Collection* coll = getCollection( txn, fromNS );
if ( !coll )
return Status( ErrorCodes::NamespaceNotFound, "collection not found to rename" );
IndexCatalog::IndexIterator ii = coll->getIndexCatalog()->getIndexIterator( true );
while ( ii.more() ) {
IndexDescriptor* desc = ii.next();
_clearCollectionCache( desc->indexNamespace() );
}
{
scoped_lock lk( _collectionLock );
_clearCollectionCache_inlock( fromNS );
_clearCollectionCache_inlock( toNS );
}
Top::global.collectionDropped( fromNS.toString() );
}
return _dbEntry->renameCollection( txn, fromNS, toNS, stayTemp );
}
long long Database::getIndexSizeForCollection(OperationContext* opCtx,
Collection* coll,
BSONObjBuilder* details,
int scale ) {
if ( !coll )
return 0;
IndexCatalog::IndexIterator ii =
coll->getIndexCatalog()->getIndexIterator( true /*includeUnfinishedIndexes*/ );
long long totalSize = 0;
while ( ii.more() ) {
IndexDescriptor* d = ii.next();
string indNS = d->indexNamespace();
// XXX creating a Collection for an index which isn't a Collection
Collection* indColl = getCollection( opCtx, indNS );
if ( ! indColl ) {
log() << "error: have index descriptor [" << indNS
<< "] but no entry in the index collection." << endl;
continue;
}
totalSize += indColl->dataSize();
if ( details ) {
long long const indexSize = indColl->dataSize() / scale;
details->appendNumber( d->indexName() , indexSize );
}
}
return totalSize;
}
IndexAccessMethod* KVDatabaseCatalogEntry::getIndex(OperationContext* opCtx,
const CollectionCatalogEntry* collection,
IndexCatalogEntry* index) {
IndexDescriptor* desc = index->descriptor();
const std::string& type = desc->getAccessMethodName();
std::string ident =
_engine->getCatalog()->getIndexIdent(opCtx, collection->ns().ns(), desc->indexName());
SortedDataInterface* sdi =
_engine->getEngine()->getGroupedSortedDataInterface(opCtx, ident, desc, index->getPrefix());
if ("" == type)
return new BtreeAccessMethod(index, sdi);
if (IndexNames::HASHED == type)
return new HashAccessMethod(index, sdi);
if (IndexNames::GEO_2DSPHERE == type)
return new S2AccessMethod(index, sdi);
if (IndexNames::TEXT == type)
return new FTSAccessMethod(index, sdi);
if (IndexNames::GEO_HAYSTACK == type)
return new HaystackAccessMethod(index, sdi);
if (IndexNames::GEO_2D == type)
return new TwoDAccessMethod(index, sdi);
log() << "Can't find index for keyPattern " << desc->keyPattern();
invariant(false);
}
Status Database::renameCollection(OperationContext* txn,
StringData fromNS,
StringData toNS,
bool stayTemp) {
audit::logRenameCollection(&cc(), fromNS, toNS);
invariant(txn->lockState()->isDbLockedForMode(name(), MODE_X));
BackgroundOperation::assertNoBgOpInProgForNs(fromNS);
BackgroundOperation::assertNoBgOpInProgForNs(toNS);
{ // remove anything cached
Collection* coll = getCollection(fromNS);
if (!coll)
return Status(ErrorCodes::NamespaceNotFound, "collection not found to rename");
string clearCacheReason = str::stream() << "renamed collection '" << fromNS << "' to '"
<< toNS << "'";
IndexCatalog::IndexIterator ii = coll->getIndexCatalog()->getIndexIterator(txn, true);
while (ii.more()) {
IndexDescriptor* desc = ii.next();
_clearCollectionCache(txn, desc->indexNamespace(), clearCacheReason);
}
_clearCollectionCache(txn, fromNS, clearCacheReason);
_clearCollectionCache(txn, toNS, clearCacheReason);
Top::get(txn->getClient()->getServiceContext()).collectionDropped(fromNS.toString());
}
txn->recoveryUnit()->registerChange(new AddCollectionChange(txn, this, toNS));
Status s = _dbEntry->renameCollection(txn, fromNS, toNS, stayTemp);
_collections[toNS] = _getOrCreateCollectionInstance(txn, toNS);
return s;
}
void run() {
OperationContextImpl txn;
Client::WriteContext ctx(&txn, _ns);
int numFinishedIndexesStart = _catalog->numIndexesReady(&txn);
Helpers::ensureIndex(&txn, _coll, BSON("x" << 1), false, "_x_0");
Helpers::ensureIndex(&txn, _coll, BSON("y" << 1), false, "_y_0");
ASSERT_TRUE(_catalog->numIndexesReady(&txn) == numFinishedIndexesStart+2);
IndexCatalog::IndexIterator ii = _catalog->getIndexIterator(&txn,false);
int indexesIterated = 0;
bool foundIndex = false;
while (ii.more()) {
IndexDescriptor* indexDesc = ii.next();
indexesIterated++;
BSONObjIterator boit(indexDesc->infoObj());
while (boit.more() && !foundIndex) {
BSONElement e = boit.next();
if (str::equals(e.fieldName(), "name") &&
str::equals(e.valuestrsafe(), "_y_0")) {
foundIndex = true;
break;
}
}
}
ctx.commit();
ASSERT_TRUE(indexesIterated == _catalog->numIndexesReady(&txn));
ASSERT_TRUE(foundIndex);
}
Status Database::renameCollection( OperationContext* txn,
StringData fromNS,
StringData toNS,
bool stayTemp ) {
audit::logRenameCollection( currentClient.get(), fromNS, toNS );
invariant(txn->lockState()->isDbLockedForMode(name(), MODE_X));
{ // remove anything cached
Collection* coll = getCollection( fromNS );
if ( !coll )
return Status( ErrorCodes::NamespaceNotFound, "collection not found to rename" );
IndexCatalog::IndexIterator ii = coll->getIndexCatalog()->getIndexIterator( txn, true );
while ( ii.more() ) {
IndexDescriptor* desc = ii.next();
_clearCollectionCache( txn, desc->indexNamespace() );
}
_clearCollectionCache( txn, fromNS );
_clearCollectionCache( txn, toNS );
Top::global.collectionDropped( fromNS.toString() );
}
txn->recoveryUnit()->registerChange( new AddCollectionChange(this, toNS) );
Status s = _dbEntry->renameCollection( txn, fromNS, toNS, stayTemp );
_collections[toNS] = _getOrCreateCollectionInstance(txn, toNS);
return s;
}
void run() {
IndexDescriptor* id = addIndexWithInfo();
// Create a SortPhaseOne.
SortPhaseOne phaseOne;
phaseOne.sorter.reset( new BSONObjExternalSorter(_aFirstSort));
// Add index keys to the phaseOne.
int32_t nKeys = 130;
for( int32_t i = 0; i < nKeys; ++i ) {
phaseOne.sorter->add( BSON( "a" << i ), /* dummy disk loc */ DiskLoc(), false );
}
phaseOne.nkeys = phaseOne.n = nKeys;
phaseOne.sorter->sort( false );
// Set up remaining arguments.
set<DiskLoc> dups;
CurOp* op = cc().curop();
ProgressMeterHolder pm (op->setMessage("BuildBottomUp",
"BuildBottomUp Progress",
nKeys,
nKeys));
pm.finished();
Timer timer;
// The index's root has not yet been set.
ASSERT( id->getHead().isNull() );
// Finish building the index.
buildBottomUpPhases2And3<V1>( true,
id,
*phaseOne.sorter,
false,
dups,
op,
&phaseOne,
pm,
timer,
true );
// The index's root is set after the build is complete.
ASSERT( !id->getHead().isNull() );
// Create a cursor over the index.
scoped_ptr<BtreeCursor> cursor(
BtreeCursor::make( nsdetails( _ns ),
id->getOnDisk(),
BSON( "" << -1 ), // startKey below minimum key.
BSON( "" << nKeys ), // endKey above maximum key.
true, // endKeyInclusive true.
1 // direction forward.
) );
// Check that the keys in the index are the expected ones.
int32_t expectedKey = 0;
for( ; cursor->ok(); cursor->advance(), ++expectedKey ) {
ASSERT_EQUALS( expectedKey, cursor->currKey().firstElement().number() );
}
ASSERT_EQUALS( nKeys, expectedKey );
}
void CollectionInfoCache::computeIndexKeys() {
DEV Lock::assertWriteLocked( _collection->ns().ns() );
_indexedPaths.clear();
IndexCatalog::IndexIterator i = _collection->getIndexCatalog()->getIndexIterator(true);
while (i.more()) {
IndexDescriptor* descriptor = i.next();
if (descriptor->getAccessMethodName() != IndexNames::TEXT) {
BSONObj key = descriptor->keyPattern();
BSONObjIterator j(key);
while (j.more()) {
BSONElement e = j.next();
_indexedPaths.addPath(e.fieldName());
}
}
else {
fts::FTSSpec ftsSpec(descriptor->infoObj());
if (ftsSpec.wildcard()) {
_indexedPaths.allPathsIndexed();
}
else {
for (size_t i = 0; i < ftsSpec.numExtraBefore(); ++i) {
_indexedPaths.addPath(ftsSpec.extraBefore(i));
}
for (fts::Weights::const_iterator it = ftsSpec.weights().begin();
it != ftsSpec.weights().end();
++it) {
_indexedPaths.addPath(it->first);
}
for (size_t i = 0; i < ftsSpec.numExtraAfter(); ++i) {
_indexedPaths.addPath(ftsSpec.extraAfter(i));
}
// Any update to a path containing "language" as a component could change the
// language of a subdocument. Add the override field as a path component.
_indexedPaths.addPathComponent(ftsSpec.languageOverrideField());
}
}
}
_keysComputed = true;
}
uint64_t Collection::getIndexSize(OperationContext* opCtx, BSONObjBuilder* details, int scale) {
IndexCatalog* idxCatalog = getIndexCatalog();
IndexCatalog::IndexIterator ii = idxCatalog->getIndexIterator(opCtx, true);
uint64_t totalSize = 0;
while (ii.more()) {
IndexDescriptor* d = ii.next();
IndexAccessMethod* iam = idxCatalog->getIndex(d);
long long ds = iam->getSpaceUsedBytes(opCtx);
totalSize += ds;
if (details) {
details->appendNumber(d->indexName(), ds / scale);
}
}
return totalSize;
}
void run() {
Client::WriteContext ctx(_ns);
int numFinishedIndexesStart = _catalog->numIndexesReady();
BSONObjBuilder b1;
b1.append("key", BSON("x" << 1));
b1.append("ns", _ns);
b1.append("name", "_x_0");
_catalog->createIndex(b1.obj(), true);
BSONObjBuilder b2;
b2.append("key", BSON("y" << 1));
b2.append("ns", _ns);
b2.append("name", "_y_0");
_catalog->createIndex(b2.obj(), true);
ASSERT_TRUE(_catalog->numIndexesReady() == numFinishedIndexesStart+2);
IndexCatalog::IndexIterator ii = _catalog->getIndexIterator(false);
int indexesIterated = 0;
bool foundIndex = false;
while (ii.more()) {
IndexDescriptor* indexDesc = ii.next();
indexesIterated++;
BSONObjIterator boit(indexDesc->infoObj());
while (boit.more() && !foundIndex) {
BSONElement e = boit.next();
if (str::equals(e.fieldName(), "name") &&
str::equals(e.valuestrsafe(), "_y_0")) {
foundIndex = true;
break;
}
}
}
ASSERT_TRUE(indexesIterated == _catalog->numIndexesReady());
ASSERT_TRUE(foundIndex);
}
StatusWith<CompactStats> Collection::compact( OperationContext* txn,
const CompactOptions* compactOptions ) {
if ( !_recordStore->compactSupported() )
return StatusWith<CompactStats>( ErrorCodes::BadValue,
str::stream() <<
"cannot compact collection with record store: " <<
_recordStore->name() );
if ( _indexCatalog.numIndexesInProgress() )
return StatusWith<CompactStats>( ErrorCodes::BadValue,
"cannot compact when indexes in progress" );
// same data, but might perform a little different after compact?
_infoCache.reset();
vector<BSONObj> indexSpecs;
{
IndexCatalog::IndexIterator ii( _indexCatalog.getIndexIterator( false ) );
while ( ii.more() ) {
IndexDescriptor* descriptor = ii.next();
const BSONObj spec = _compactAdjustIndexSpec(descriptor->infoObj());
const BSONObj key = spec.getObjectField("key");
const Status keyStatus = validateKeyPattern(key);
if (!keyStatus.isOK()) {
return StatusWith<CompactStats>(
ErrorCodes::CannotCreateIndex,
str::stream() << "Cannot compact collection due to invalid index "
<< spec << ": " << keyStatus.reason() << " For more info see"
<< " http://dochub.mongodb.org/core/index-validation");
}
indexSpecs.push_back(spec);
}
}
// note that the drop indexes call also invalidates all clientcursors for the namespace,
// which is important and wanted here
log() << "compact dropping indexes" << endl;
Status status = _indexCatalog.dropAllIndexes(txn, true);
if ( !status.isOK() ) {
return StatusWith<CompactStats>( status );
}
txn->checkForInterrupt();
CompactStats stats;
MultiIndexBlock multiIndexBlock(txn, this);
status = multiIndexBlock.init( indexSpecs );
if ( !status.isOK() )
return StatusWith<CompactStats>( status );
MyCompactAdaptor adaptor(this, &multiIndexBlock);
_recordStore->compact( txn, &adaptor, compactOptions, &stats );
log() << "starting index commits";
status = multiIndexBlock.commit();
if ( !status.isOK() )
return StatusWith<CompactStats>( status );
return StatusWith<CompactStats>( stats );
}
Status MigrationChunkClonerSourceLegacy::_storeCurrentLocs(OperationContext* txn) {
ScopedTransaction scopedXact(txn, MODE_IS);
AutoGetCollection autoColl(txn, _args.getNss(), MODE_IS);
Collection* const collection = autoColl.getCollection();
if (!collection) {
return {ErrorCodes::NamespaceNotFound,
str::stream() << "Collection " << _args.getNss().ns() << " does not exist."};
}
// Allow multiKey based on the invariant that shard keys must be single-valued. Therefore, any
// multi-key index prefixed by shard key cannot be multikey over the shard key fields.
IndexDescriptor* idx =
collection->getIndexCatalog()->findShardKeyPrefixedIndex(txn,
_shardKeyPattern.toBSON(),
false); // requireSingleKey
if (!idx) {
return {ErrorCodes::IndexNotFound,
str::stream() << "can't find index with prefix " << _shardKeyPattern.toBSON()
<< " in storeCurrentLocs for "
<< _args.getNss().ns()};
}
// Install the stage, which will listen for notifications on the collection
{
stdx::lock_guard<stdx::mutex> sl(_mutex);
invariant(!_deleteNotifyExec);
// Takes ownership of 'ws' and 'dns'.
auto statusWithPlanExecutor =
PlanExecutor::make(txn,
stdx::make_unique<WorkingSet>(),
stdx::make_unique<DeleteNotificationStage>(this, txn),
collection,
PlanExecutor::YIELD_MANUAL);
invariant(statusWithPlanExecutor.isOK());
_deleteNotifyExec = std::move(statusWithPlanExecutor.getValue());
_deleteNotifyExec->registerExec(collection);
}
// Assume both min and max non-empty, append MinKey's to make them fit chosen index
const KeyPattern kp(idx->keyPattern());
BSONObj min = Helpers::toKeyFormat(kp.extendRangeBound(_args.getMinKey(), false));
BSONObj max = Helpers::toKeyFormat(kp.extendRangeBound(_args.getMaxKey(), false));
std::unique_ptr<PlanExecutor> exec(InternalPlanner::indexScan(txn,
collection,
idx,
min,
max,
false, // endKeyInclusive
PlanExecutor::YIELD_MANUAL));
// We can afford to yield here because any change to the base data that we might miss is already
// being queued and will migrate in the 'transferMods' stage.
exec->setYieldPolicy(PlanExecutor::YIELD_AUTO, collection);
// Use the average object size to estimate how many objects a full chunk would carry do that
// while traversing the chunk's range using the sharding index, below there's a fair amount of
// slack before we determine a chunk is too large because object sizes will vary.
unsigned long long maxRecsWhenFull;
long long avgRecSize;
const long long totalRecs = collection->numRecords(txn);
if (totalRecs > 0) {
avgRecSize = collection->dataSize(txn) / totalRecs;
maxRecsWhenFull = _args.getMaxChunkSizeBytes() / avgRecSize;
maxRecsWhenFull = std::min((unsigned long long)(Chunk::MaxObjectPerChunk + 1),
130 * maxRecsWhenFull / 100 /* slack */);
} else {
avgRecSize = 0;
maxRecsWhenFull = Chunk::MaxObjectPerChunk + 1;
}
// Do a full traversal of the chunk and don't stop even if we think it is a large chunk we want
// the number of records to better report, in that case.
bool isLargeChunk = false;
unsigned long long recCount = 0;
BSONObj obj;
RecordId recordId;
PlanExecutor::ExecState state;
while (PlanExecutor::ADVANCED == (state = exec->getNext(&obj, &recordId))) {
if (!isLargeChunk) {
stdx::lock_guard<stdx::mutex> lk(_mutex);
_cloneLocs.insert(recordId);
}
if (++recCount > maxRecsWhenFull) {
isLargeChunk = true;
// Continue on despite knowing that it will fail, just to get the correct value for
// recCount
}
}
if (PlanExecutor::DEAD == state || PlanExecutor::FAILURE == state) {
return {ErrorCodes::InternalError,
str::stream() << "Executor error while scanning for documents belonging to chunk: "
<< WorkingSetCommon::toStatusString(obj)};
}
exec.reset();
if (isLargeChunk) {
return {
ErrorCodes::ChunkTooBig,
str::stream() << "Cannot move chunk: the maximum number of documents for a chunk is "
<< maxRecsWhenFull
<< ", the maximum chunk size is "
<< _args.getMaxChunkSizeBytes()
<< ", average document size is "
<< avgRecSize
<< ". Found "
<< recCount
<< " documents in chunk "
<< " ns: "
<< _args.getNss().ns()
<< " "
<< _args.getMinKey()
<< " -> "
<< _args.getMaxKey()};
}
_averageObjectSizeForCloneLocs = static_cast<uint64_t>(collection->averageObjectSize(txn) + 12);
return Status::OK();
}
StatusWith<RecordId> Collection::updateDocument(OperationContext* txn,
const RecordId& oldLocation,
const Snapshotted<BSONObj>& oldDoc,
const BSONObj& newDoc,
bool enforceQuota,
bool indexesAffected,
OpDebug* debug,
oplogUpdateEntryArgs& args) {
{
auto status = checkValidation(txn, newDoc);
if (!status.isOK()) {
if (_validationLevel == STRICT_V) {
return status;
}
// moderate means we have to check the old doc
auto oldDocStatus = checkValidation(txn, oldDoc.value());
if (oldDocStatus.isOK()) {
// transitioning from good -> bad is not ok
return status;
}
// bad -> bad is ok in moderate mode
}
}
dassert(txn->lockState()->isCollectionLockedForMode(ns().toString(), MODE_IX));
invariant(oldDoc.snapshotId() == txn->recoveryUnit()->getSnapshotId());
if (_needCappedLock) {
// X-lock the metadata resource for this capped collection until the end of the WUOW. This
// prevents the primary from executing with more concurrency than secondaries.
// See SERVER-21646.
Lock::ResourceLock{txn->lockState(), ResourceId(RESOURCE_METADATA, _ns.ns()), MODE_X};
}
SnapshotId sid = txn->recoveryUnit()->getSnapshotId();
BSONElement oldId = oldDoc.value()["_id"];
if (!oldId.eoo() && (oldId != newDoc["_id"]))
return StatusWith<RecordId>(
ErrorCodes::InternalError, "in Collection::updateDocument _id mismatch", 13596);
// The MMAPv1 storage engine implements capped collections in a way that does not allow records
// to grow beyond their original size. If MMAPv1 part of a replicaset with storage engines that
// do not have this limitation, replication could result in errors, so it is necessary to set a
// uniform rule here. Similarly, it is not sufficient to disallow growing records, because this
// happens when secondaries roll back an update shrunk a record. Exactly replicating legacy
// MMAPv1 behavior would require padding shrunk documents on all storage engines. Instead forbid
// all size changes.
const auto oldSize = oldDoc.value().objsize();
if (_recordStore->isCapped() && oldSize != newDoc.objsize())
return {ErrorCodes::CannotGrowDocumentInCappedNamespace,
str::stream() << "Cannot change the size of a document in a capped collection: "
<< oldSize << " != " << newDoc.objsize()};
// At the end of this step, we will have a map of UpdateTickets, one per index, which
// represent the index updates needed to be done, based on the changes between oldDoc and
// newDoc.
OwnedPointerMap<IndexDescriptor*, UpdateTicket> updateTickets;
if (indexesAffected) {
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator(txn, true);
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
IndexCatalogEntry* entry = ii.catalogEntry(descriptor);
IndexAccessMethod* iam = ii.accessMethod(descriptor);
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed =
!(KeyPattern::isIdKeyPattern(descriptor->keyPattern()) || descriptor->unique()) ||
repl::getGlobalReplicationCoordinator()->shouldIgnoreUniqueIndex(descriptor);
UpdateTicket* updateTicket = new UpdateTicket();
updateTickets.mutableMap()[descriptor] = updateTicket;
Status ret = iam->validateUpdate(txn,
oldDoc.value(),
newDoc,
oldLocation,
options,
updateTicket,
entry->getFilterExpression());
if (!ret.isOK()) {
return StatusWith<RecordId>(ret);
}
}
}
// This can call back into Collection::recordStoreGoingToMove. If that happens, the old
// object is removed from all indexes.
StatusWith<RecordId> newLocation = _recordStore->updateRecord(
txn, oldLocation, newDoc.objdata(), newDoc.objsize(), _enforceQuota(enforceQuota), this);
if (!newLocation.isOK()) {
return newLocation;
}
// At this point, the old object may or may not still be indexed, depending on if it was
// moved. If the object did move, we need to add the new location to all indexes.
if (newLocation.getValue() != oldLocation) {
if (debug) {
if (debug->nmoved == -1) // default of -1 rather than 0
debug->nmoved = 1;
else
debug->nmoved += 1;
}
std::vector<BsonRecord> bsonRecords;
BsonRecord bsonRecord = {newLocation.getValue(), &newDoc};
bsonRecords.push_back(bsonRecord);
Status s = _indexCatalog.indexRecords(txn, bsonRecords);
if (!s.isOK())
return StatusWith<RecordId>(s);
invariant(sid == txn->recoveryUnit()->getSnapshotId());
args.ns = ns().ns();
getGlobalServiceContext()->getOpObserver()->onUpdate(txn, args);
return newLocation;
}
// Object did not move. We update each index with each respective UpdateTicket.
if (debug)
debug->keyUpdates = 0;
if (indexesAffected) {
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator(txn, true);
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* iam = ii.accessMethod(descriptor);
int64_t updatedKeys;
Status ret = iam->update(txn, *updateTickets.mutableMap()[descriptor], &updatedKeys);
if (!ret.isOK())
return StatusWith<RecordId>(ret);
if (debug)
debug->keyUpdates += updatedKeys;
}
}
invariant(sid == txn->recoveryUnit()->getSnapshotId());
args.ns = ns().ns();
getGlobalServiceContext()->getOpObserver()->onUpdate(txn, args);
return newLocation;
}
Status MMAPV1Engine::repairDatabase( OperationContext* txn,
const std::string& dbName,
bool preserveClonedFilesOnFailure,
bool backupOriginalFiles ) {
// We must hold some form of lock here
invariant(txn->lockState()->threadState());
invariant( dbName.find( '.' ) == string::npos );
scoped_ptr<RepairFileDeleter> repairFileDeleter;
log() << "repairDatabase " << dbName << endl;
BackgroundOperation::assertNoBgOpInProgForDb(dbName);
txn->recoveryUnit()->syncDataAndTruncateJournal(); // Must be done before and after repair
intmax_t totalSize = dbSize( dbName );
intmax_t freeSize = File::freeSpace(storageGlobalParams.repairpath);
if ( freeSize > -1 && freeSize < totalSize ) {
return Status( ErrorCodes::OutOfDiskSpace,
str::stream() << "Cannot repair database " << dbName
<< " having size: " << totalSize
<< " (bytes) because free disk space is: " << freeSize << " (bytes)" );
}
txn->checkForInterrupt();
Path reservedPath =
uniqueReservedPath( ( preserveClonedFilesOnFailure || backupOriginalFiles ) ?
"backup" : "_tmp" );
MONGO_ASSERT_ON_EXCEPTION( boost::filesystem::create_directory( reservedPath ) );
string reservedPathString = reservedPath.string();
if ( !preserveClonedFilesOnFailure )
repairFileDeleter.reset( new RepairFileDeleter( txn,
dbName,
reservedPathString,
reservedPath ) );
{
Database* originalDatabase =
dbHolder().get(txn, dbName);
if (originalDatabase == NULL) {
return Status(ErrorCodes::NamespaceNotFound, "database does not exist to repair");
}
scoped_ptr<MMAPV1DatabaseCatalogEntry> dbEntry;
scoped_ptr<Database> tempDatabase;
{
dbEntry.reset( new MMAPV1DatabaseCatalogEntry( txn,
dbName,
reservedPathString,
storageGlobalParams.directoryperdb,
true ) );
invariant( !dbEntry->exists() );
tempDatabase.reset( new Database( txn,
dbName,
dbEntry.get() ) );
}
map<string,CollectionOptions> namespacesToCopy;
{
string ns = dbName + ".system.namespaces";
Client::Context ctx(txn, ns );
Collection* coll = originalDatabase->getCollection( txn, ns );
if ( coll ) {
scoped_ptr<RecordIterator> it( coll->getIterator( txn,
DiskLoc(),
false,
CollectionScanParams::FORWARD ) );
while ( !it->isEOF() ) {
DiskLoc loc = it->getNext();
BSONObj obj = coll->docFor( loc );
string ns = obj["name"].String();
NamespaceString nss( ns );
if ( nss.isSystem() ) {
if ( nss.isSystemDotIndexes() )
continue;
if ( nss.coll() == "system.namespaces" )
continue;
}
if ( !nss.isNormal() )
continue;
CollectionOptions options;
if ( obj["options"].isABSONObj() ) {
Status status = options.parse( obj["options"].Obj() );
if ( !status.isOK() )
return status;
}
namespacesToCopy[ns] = options;
}
}
}
for ( map<string,CollectionOptions>::const_iterator i = namespacesToCopy.begin();
i != namespacesToCopy.end();
++i ) {
string ns = i->first;
CollectionOptions options = i->second;
Collection* tempCollection = NULL;
{
Client::Context tempContext(txn, ns, tempDatabase );
WriteUnitOfWork wunit(txn);
tempCollection = tempDatabase->createCollection(txn, ns, options, true, false);
wunit.commit();
}
Client::Context readContext(txn, ns, originalDatabase);
Collection* originalCollection = originalDatabase->getCollection( txn, ns );
invariant( originalCollection );
// data
// TODO SERVER-14812 add a mode that drops duplicates rather than failing
MultiIndexBlock indexer(txn, tempCollection );
{
vector<BSONObj> indexes;
IndexCatalog::IndexIterator ii =
originalCollection->getIndexCatalog()->getIndexIterator( false );
while ( ii.more() ) {
IndexDescriptor* desc = ii.next();
indexes.push_back( desc->infoObj() );
}
Client::Context tempContext(txn, ns, tempDatabase);
Status status = indexer.init( indexes );
if ( !status.isOK() )
return status;
}
scoped_ptr<RecordIterator> iterator(
originalCollection->getIterator( txn, DiskLoc(), false,
CollectionScanParams::FORWARD ));
while ( !iterator->isEOF() ) {
DiskLoc loc = iterator->getNext();
invariant( !loc.isNull() );
BSONObj doc = originalCollection->docFor( loc );
Client::Context tempContext(txn, ns, tempDatabase);
WriteUnitOfWork wunit(txn);
StatusWith<DiskLoc> result = tempCollection->insertDocument(txn,
doc,
&indexer,
false);
if ( !result.isOK() )
return result.getStatus();
wunit.commit();
txn->checkForInterrupt(false);
}
Status status = indexer.doneInserting();
if (!status.isOK())
return status;
{
Client::Context tempContext(txn, ns, tempDatabase);
WriteUnitOfWork wunit(txn);
indexer.commit();
wunit.commit();
}
}
txn->recoveryUnit()->syncDataAndTruncateJournal();
globalStorageEngine->flushAllFiles(true); // need both in case journaling is disabled
txn->checkForInterrupt(false);
}
// at this point if we abort, we don't want to delete new files
// as they might be the only copies
if ( repairFileDeleter.get() )
repairFileDeleter->success();
dbHolder().close( txn, dbName );
if ( backupOriginalFiles ) {
_renameForBackup( dbName, reservedPath );
}
else {
// first make new directory before deleting data
Path newDir = Path(storageGlobalParams.dbpath) / dbName;
MONGO_ASSERT_ON_EXCEPTION(boost::filesystem::create_directory(newDir));
// this deletes old files
_deleteDataFiles( dbName );
if ( !boost::filesystem::exists(newDir) ) {
// we deleted because of directoryperdb
// re-create
MONGO_ASSERT_ON_EXCEPTION(boost::filesystem::create_directory(newDir));
}
}
_replaceWithRecovered( dbName, reservedPathString.c_str() );
if ( !backupOriginalFiles )
MONGO_ASSERT_ON_EXCEPTION( boost::filesystem::remove_all( reservedPath ) );
return Status::OK();
}
void run() {
IndexDescriptor* id = addIndexWithInfo();
// Create a SortPhaseOne.
SortPhaseOne phaseOne;
phaseOne.sorter.reset(new BSONObjExternalSorter(_aFirstSort));
// It's necessary to index sufficient keys that a RARELY condition will be triggered,
// but few enough keys that the btree builder will not create an internal node and check
// for an interrupt internally (which would cause this test to pass spuriously).
int32_t nKeys = 130;
// Add index keys to the phaseOne.
for( int32_t i = 0; i < nKeys; ++i ) {
phaseOne.sorter->add( BSON( "a" << i ), /* dummy disk loc */ DiskLoc(), false );
}
phaseOne.nkeys = phaseOne.n = nKeys;
phaseOne.sorter->sort( false );
// Set up remaining arguments.
set<DiskLoc> dups;
CurOp* op = cc().curop();
ProgressMeterHolder pm (op->setMessage("InterruptBuildBottomUp",
"InterruptBuildBottomUp Progress",
nKeys,
nKeys));
pm.finished();
Timer timer;
// The index's root has not yet been set.
ASSERT( id->getHead().isNull() );
// Register a request to kill the current operation.
cc().curop()->kill();
if ( _mayInterrupt ) {
// The build is aborted due to the kill request.
ASSERT_THROWS
( buildBottomUpPhases2And3<V1>( true,
id,
*phaseOne.sorter,
false,
dups,
op,
&phaseOne,
pm,
timer,
_mayInterrupt ),
UserException );
// The root of the index is not set because the build did not complete.
ASSERT( id->getHead().isNull() );
}
else {
// The build is aborted despite the kill request because mayInterrupt == false.
buildBottomUpPhases2And3<V1>( true,
id,
*phaseOne.sorter,
false,
dups,
op,
&phaseOne,
pm,
timer,
_mayInterrupt );
// The index's root is set after the build is complete.
ASSERT( !id->getHead().isNull() );
}
}
void CollectionInfoCache::computeIndexKeys(OperationContext* opCtx) {
_indexedPaths.clear();
bool hadTTLIndex = _hasTTLIndex;
_hasTTLIndex = false;
IndexCatalog::IndexIterator i = _collection->getIndexCatalog()->getIndexIterator(opCtx, true);
while (i.more()) {
IndexDescriptor* descriptor = i.next();
if (descriptor->getAccessMethodName() != IndexNames::TEXT) {
BSONObj key = descriptor->keyPattern();
const BSONObj& infoObj = descriptor->infoObj();
if (infoObj.hasField("expireAfterSeconds")) {
_hasTTLIndex = true;
}
BSONObjIterator j(key);
while (j.more()) {
BSONElement e = j.next();
_indexedPaths.addPath(e.fieldName());
}
} else {
fts::FTSSpec ftsSpec(descriptor->infoObj());
if (ftsSpec.wildcard()) {
_indexedPaths.allPathsIndexed();
} else {
for (size_t i = 0; i < ftsSpec.numExtraBefore(); ++i) {
_indexedPaths.addPath(ftsSpec.extraBefore(i));
}
for (fts::Weights::const_iterator it = ftsSpec.weights().begin();
it != ftsSpec.weights().end();
++it) {
_indexedPaths.addPath(it->first);
}
for (size_t i = 0; i < ftsSpec.numExtraAfter(); ++i) {
_indexedPaths.addPath(ftsSpec.extraAfter(i));
}
// Any update to a path containing "language" as a component could change the
// language of a subdocument. Add the override field as a path component.
_indexedPaths.addPathComponent(ftsSpec.languageOverrideField());
}
}
// handle partial indexes
const IndexCatalogEntry* entry = i.catalogEntry(descriptor);
const MatchExpression* filter = entry->getFilterExpression();
if (filter) {
unordered_set<std::string> paths;
QueryPlannerIXSelect::getFields(filter, "", &paths);
for (auto it = paths.begin(); it != paths.end(); ++it) {
_indexedPaths.addPath(*it);
}
}
}
TTLCollectionCache& ttlCollectionCache = TTLCollectionCache::get(getGlobalServiceContext());
if (_hasTTLIndex != hadTTLIndex) {
if (_hasTTLIndex) {
ttlCollectionCache.registerCollection(_collection->ns());
} else {
ttlCollectionCache.unregisterCollection(_collection->ns());
}
}
_keysComputed = true;
}
bool run(const string& dbname, BSONObj& jsobj, int, string& errmsg, BSONObjBuilder& anObjBuilder, bool /*fromRepl*/) {
BSONElement e = jsobj.firstElement();
const string toDeleteNs = dbname + '.' + e.valuestr();
if (!serverGlobalParams.quiet) {
MONGO_TLOG(0) << "CMD: dropIndexes " << toDeleteNs << endl;
}
Lock::DBWrite dbXLock(dbname);
Client::Context ctx(toDeleteNs);
Collection* collection = cc().database()->getCollection( toDeleteNs );
if ( ! collection ) {
errmsg = "ns not found";
return false;
}
stopIndexBuilds(cc().database(), jsobj);
IndexCatalog* indexCatalog = collection->getIndexCatalog();
anObjBuilder.appendNumber("nIndexesWas", indexCatalog->numIndexesTotal() );
BSONElement f = jsobj.getField("index");
if ( f.type() == String ) {
string indexToDelete = f.valuestr();
if ( indexToDelete == "*" ) {
Status s = indexCatalog->dropAllIndexes( false );
if ( !s.isOK() ) {
appendCommandStatus( anObjBuilder, s );
return false;
}
anObjBuilder.append("msg", "non-_id indexes dropped for collection");
return true;
}
IndexDescriptor* desc = collection->getIndexCatalog()->findIndexByName( indexToDelete );
if ( desc == NULL ) {
errmsg = str::stream() << "index not found with name [" << indexToDelete << "]";
return false;
}
if ( desc->isIdIndex() ) {
errmsg = "cannot drop _id index";
return false;
}
Status s = indexCatalog->dropIndex( desc );
if ( !s.isOK() ) {
appendCommandStatus( anObjBuilder, s );
return false;
}
return true;
}
if ( f.type() == Object ) {
IndexDescriptor* desc = collection->getIndexCatalog()->findIndexByKeyPattern( f.embeddedObject() );
if ( desc == NULL ) {
errmsg = "can't find index with key:";
errmsg += f.embeddedObject().toString();
return false;
}
if ( desc->isIdIndex() ) {
errmsg = "cannot drop _id index";
return false;
}
Status s = indexCatalog->dropIndex( desc );
if ( !s.isOK() ) {
appendCommandStatus( anObjBuilder, s );
return false;
}
return true;
}
errmsg = "invalid index name spec";
return false;
}
bool MigrationSourceManager::storeCurrentLocs(OperationContext* txn,
long long maxChunkSize,
string& errmsg,
BSONObjBuilder& result) {
AutoGetCollection autoColl(txn, _getNS(), MODE_IS);
Collection* collection = autoColl.getCollection();
if (!collection) {
errmsg = "ns not found, should be impossible";
return false;
}
// Allow multiKey based on the invariant that shard keys must be single-valued. Therefore, any
// multi-key index prefixed by shard key cannot be multikey over the shard key fields.
IndexDescriptor* idx =
collection->getIndexCatalog()->findShardKeyPrefixedIndex(txn,
_shardKeyPattern,
false); // requireSingleKey
if (idx == NULL) {
errmsg = str::stream() << "can't find index with prefix " << _shardKeyPattern
<< " in storeCurrentLocs for " << _nss.toString();
return false;
}
// Assume both min and max non-empty, append MinKey's to make them fit chosen index
BSONObj min;
BSONObj max;
KeyPattern kp(idx->keyPattern());
{
// It's alright not to lock _mutex all the way through based on the assumption that this is
// only called by the main thread that drives the migration and only it can start and stop
// the current migration.
stdx::lock_guard<stdx::mutex> sl(_mutex);
invariant(_deleteNotifyExec.get() == NULL);
unique_ptr<WorkingSet> ws = stdx::make_unique<WorkingSet>();
unique_ptr<DeleteNotificationStage> dns = stdx::make_unique<DeleteNotificationStage>(this);
// Takes ownership of 'ws' and 'dns'.
auto statusWithPlanExecutor = PlanExecutor::make(
txn, std::move(ws), std::move(dns), collection, PlanExecutor::YIELD_MANUAL);
invariant(statusWithPlanExecutor.isOK());
_deleteNotifyExec = std::move(statusWithPlanExecutor.getValue());
_deleteNotifyExec->registerExec();
min = Helpers::toKeyFormat(kp.extendRangeBound(_min, false));
max = Helpers::toKeyFormat(kp.extendRangeBound(_max, false));
}
unique_ptr<PlanExecutor> exec(InternalPlanner::indexScan(txn,
collection,
idx,
min,
max,
false, // endKeyInclusive
PlanExecutor::YIELD_MANUAL));
// We can afford to yield here because any change to the base data that we might miss is already
// being queued and will migrate in the 'transferMods' stage.
exec->setYieldPolicy(PlanExecutor::YIELD_AUTO);
// Use the average object size to estimate how many objects a full chunk would carry do that
// while traversing the chunk's range using the sharding index, below there's a fair amount of
// slack before we determine a chunk is too large because object sizes will vary.
unsigned long long maxRecsWhenFull;
long long avgRecSize;
const long long totalRecs = collection->numRecords(txn);
if (totalRecs > 0) {
avgRecSize = collection->dataSize(txn) / totalRecs;
maxRecsWhenFull = maxChunkSize / avgRecSize;
maxRecsWhenFull = std::min((unsigned long long)(Chunk::MaxObjectPerChunk + 1),
130 * maxRecsWhenFull / 100 /* slack */);
} else {
avgRecSize = 0;
maxRecsWhenFull = Chunk::MaxObjectPerChunk + 1;
}
// Do a full traversal of the chunk and don't stop even if we think it is a large chunk we want
// the number of records to better report, in that case
bool isLargeChunk = false;
unsigned long long recCount = 0;
RecordId recordId;
while (PlanExecutor::ADVANCED == exec->getNext(NULL, &recordId)) {
if (!isLargeChunk) {
stdx::lock_guard<stdx::mutex> lk(_cloneLocsMutex);
_cloneLocs.insert(recordId);
}
if (++recCount > maxRecsWhenFull) {
isLargeChunk = true;
// Continue on despite knowing that it will fail, just to get the correct value for
// recCount
}
}
exec.reset();
if (isLargeChunk) {
stdx::lock_guard<stdx::mutex> sl(_mutex);
warning() << "cannot move chunk: the maximum number of documents for a chunk is "
<< maxRecsWhenFull << " , the maximum chunk size is " << maxChunkSize
<< " , average document size is " << avgRecSize << ". Found " << recCount
<< " documents in chunk "
<< " ns: " << _nss << " " << _min << " -> " << _max << migrateLog;
result.appendBool("chunkTooBig", true);
result.appendNumber("estimatedChunkSize", (long long)(recCount * avgRecSize));
errmsg = "chunk too big to move";
return false;
}
log() << "moveChunk number of documents: " << cloneLocsRemaining() << migrateLog;
txn->recoveryUnit()->abandonSnapshot();
return true;
}
StatusWith<RecordId> Collection::updateDocument(OperationContext* txn,
const RecordId& oldLocation,
const Snapshotted<BSONObj>& oldDoc,
const BSONObj& newDoc,
bool enforceQuota,
bool indexesAffected,
OpDebug* debug,
oplogUpdateEntryArgs& args) {
{
auto status = checkValidation(txn, newDoc);
if (!status.isOK()) {
if (_validationLevel == STRICT_V) {
return status;
}
// moderate means we have to check the old doc
auto oldDocStatus = checkValidation(txn, oldDoc.value());
if (oldDocStatus.isOK()) {
// transitioning from good -> bad is not ok
return status;
}
// bad -> bad is ok in moderate mode
}
}
dassert(txn->lockState()->isCollectionLockedForMode(ns().toString(), MODE_IX));
invariant(oldDoc.snapshotId() == txn->recoveryUnit()->getSnapshotId());
SnapshotId sid = txn->recoveryUnit()->getSnapshotId();
BSONElement oldId = oldDoc.value()["_id"];
if (!oldId.eoo() && (oldId != newDoc["_id"]))
return StatusWith<RecordId>(
ErrorCodes::InternalError, "in Collection::updateDocument _id mismatch", 13596);
// At the end of this step, we will have a map of UpdateTickets, one per index, which
// represent the index updates needed to be done, based on the changes between oldDoc and
// newDoc.
OwnedPointerMap<IndexDescriptor*, UpdateTicket> updateTickets;
if (indexesAffected) {
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator(txn, true);
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
IndexCatalogEntry* entry = ii.catalogEntry(descriptor);
IndexAccessMethod* iam = ii.accessMethod(descriptor);
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed =
!(KeyPattern::isIdKeyPattern(descriptor->keyPattern()) || descriptor->unique()) ||
repl::getGlobalReplicationCoordinator()->shouldIgnoreUniqueIndex(descriptor);
UpdateTicket* updateTicket = new UpdateTicket();
updateTickets.mutableMap()[descriptor] = updateTicket;
Status ret = iam->validateUpdate(txn,
oldDoc.value(),
newDoc,
oldLocation,
options,
updateTicket,
entry->getFilterExpression());
if (!ret.isOK()) {
return StatusWith<RecordId>(ret);
}
}
}
// This can call back into Collection::recordStoreGoingToMove. If that happens, the old
// object is removed from all indexes.
StatusWith<RecordId> newLocation = _recordStore->updateRecord(
txn, oldLocation, newDoc.objdata(), newDoc.objsize(), _enforceQuota(enforceQuota), this);
if (!newLocation.isOK()) {
return newLocation;
}
// At this point, the old object may or may not still be indexed, depending on if it was
// moved. If the object did move, we need to add the new location to all indexes.
if (newLocation.getValue() != oldLocation) {
if (debug) {
if (debug->nmoved == -1) // default of -1 rather than 0
debug->nmoved = 1;
else
debug->nmoved += 1;
}
Status s = _indexCatalog.indexRecord(txn, newDoc, newLocation.getValue());
if (!s.isOK())
return StatusWith<RecordId>(s);
invariant(sid == txn->recoveryUnit()->getSnapshotId());
args.ns = ns().ns();
getGlobalServiceContext()->getOpObserver()->onUpdate(txn, args);
return newLocation;
}
// Object did not move. We update each index with each respective UpdateTicket.
if (debug)
debug->keyUpdates = 0;
if (indexesAffected) {
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator(txn, true);
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* iam = ii.accessMethod(descriptor);
int64_t updatedKeys;
Status ret = iam->update(txn, *updateTickets.mutableMap()[descriptor], &updatedKeys);
if (!ret.isOK())
return StatusWith<RecordId>(ret);
if (debug)
debug->keyUpdates += updatedKeys;
}
}
invariant(sid == txn->recoveryUnit()->getSnapshotId());
args.ns = ns().ns();
getGlobalServiceContext()->getOpObserver()->onUpdate(txn, args);
return newLocation;
}
StatusWith<DiskLoc> Collection::updateDocument( OperationContext* txn,
const DiskLoc& oldLocation,
const BSONObj& objNew,
bool enforceQuota,
OpDebug* debug ) {
BSONObj objOld = _recordStore->dataFor( txn, oldLocation ).toBson();
if ( objOld.hasElement( "_id" ) ) {
BSONElement oldId = objOld["_id"];
BSONElement newId = objNew["_id"];
if ( oldId != newId )
return StatusWith<DiskLoc>( ErrorCodes::InternalError,
"in Collection::updateDocument _id mismatch",
13596 );
}
if ( ns().coll() == "system.users" ) {
// XXX - andy and spencer think this should go away now
V2UserDocumentParser parser;
Status s = parser.checkValidUserDocument(objNew);
if ( !s.isOK() )
return StatusWith<DiskLoc>( s );
}
/* duplicate key check. we descend the btree twice - once for this check, and once for the actual inserts, further
below. that is suboptimal, but it's pretty complicated to do it the other way without rollbacks...
*/
OwnedPointerMap<IndexDescriptor*,UpdateTicket> updateTickets;
IndexCatalog::IndexIterator ii = _indexCatalog.getIndexIterator( txn, true );
while ( ii.more() ) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed =
!(KeyPattern::isIdKeyPattern(descriptor->keyPattern()) || descriptor->unique())
|| repl::getGlobalReplicationCoordinator()->shouldIgnoreUniqueIndex(descriptor);
UpdateTicket* updateTicket = new UpdateTicket();
updateTickets.mutableMap()[descriptor] = updateTicket;
Status ret = iam->validateUpdate(txn, objOld, objNew, oldLocation, options, updateTicket );
if ( !ret.isOK() ) {
return StatusWith<DiskLoc>( ret );
}
}
// this can callback into Collection::recordStoreGoingToMove
StatusWith<DiskLoc> newLocation = _recordStore->updateRecord( txn,
oldLocation,
objNew.objdata(),
objNew.objsize(),
_enforceQuota( enforceQuota ),
this );
if ( !newLocation.isOK() ) {
return newLocation;
}
_infoCache.notifyOfWriteOp();
if ( newLocation.getValue() != oldLocation ) {
if ( debug ) {
if (debug->nmoved == -1) // default of -1 rather than 0
debug->nmoved = 1;
else
debug->nmoved += 1;
}
_indexCatalog.indexRecord(txn, objNew, newLocation.getValue());
return newLocation;
}
if ( debug )
debug->keyUpdates = 0;
ii = _indexCatalog.getIndexIterator( txn, true );
while ( ii.more() ) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
int64_t updatedKeys;
Status ret = iam->update(txn, *updateTickets.mutableMap()[descriptor], &updatedKeys);
if ( !ret.isOK() )
return StatusWith<DiskLoc>( ret );
if ( debug )
debug->keyUpdates += updatedKeys;
}
// Broadcast the mutation so that query results stay correct.
_cursorCache.invalidateDocument(oldLocation, INVALIDATION_MUTATION);
return newLocation;
}
StatusWith<CompactStats> Collection::compact( const CompactOptions* compactOptions ) {
if ( isCapped() )
return StatusWith<CompactStats>( ErrorCodes::BadValue,
"cannot compact capped collection" );
if ( _indexCatalog.numIndexesInProgress() )
return StatusWith<CompactStats>( ErrorCodes::BadValue,
"cannot compact when indexes in progress" );
NamespaceDetails* d = details();
// this is a big job, so might as well make things tidy before we start just to be nice.
getDur().commitIfNeeded();
list<DiskLoc> extents;
for( DiskLoc L = d->firstExtent(); !L.isNull(); L = L.ext()->xnext )
extents.push_back(L);
log() << "compact " << extents.size() << " extents" << endl;
// same data, but might perform a little different after compact?
_infoCache.reset();
vector<BSONObj> indexSpecs;
{
IndexCatalog::IndexIterator ii( _indexCatalog.getIndexIterator( false ) );
while ( ii.more() ) {
IndexDescriptor* descriptor = ii.next();
indexSpecs.push_back( _compactAdjustIndexSpec( descriptor->infoObj() ) );
}
}
log() << "compact orphan deleted lists" << endl;
d->orphanDeletedList();
// Start over from scratch with our extent sizing and growth
d->setLastExtentSize( 0 );
// before dropping indexes, at least make sure we can allocate one extent!
if ( allocateSpaceForANewRecord( _ns.ns().c_str(),
d,
Record::HeaderSize+1,
false).isNull() ) {
return StatusWith<CompactStats>( ErrorCodes::InternalError,
"compact error no space available to allocate" );
}
// note that the drop indexes call also invalidates all clientcursors for the namespace,
// which is important and wanted here
log() << "compact dropping indexes" << endl;
Status status = _indexCatalog.dropAllIndexes( true );
if ( !status.isOK() ) {
return StatusWith<CompactStats>( status );
}
getDur().commitIfNeeded();
CompactStats stats;
OwnedPointerVector<IndexCatalog::IndexBuildBlock> indexBuildBlocks;
vector<IndexAccessMethod*> indexesToInsertTo;
vector< std::pair<IndexAccessMethod*,IndexAccessMethod*> > bulkToCommit;
for ( size_t i = 0; i < indexSpecs.size(); i++ ) {
killCurrentOp.checkForInterrupt(false);
BSONObj info = indexSpecs[i];
info = _compactAdjustIndexSpec( info );
info = _indexCatalog.fixIndexSpec( info );
auto_ptr<IndexCatalog::IndexBuildBlock> block( new IndexCatalog::IndexBuildBlock( this,info ) );
Status status = block->init();
if ( !status.isOK() )
return StatusWith<CompactStats>(status);
IndexAccessMethod* accessMethod = block->getEntry()->accessMethod();
status = accessMethod->initializeAsEmpty();
if ( !status.isOK() )
return StatusWith<CompactStats>(status);
IndexAccessMethod* bulk = accessMethod->initiateBulk();
if ( bulk ) {
indexesToInsertTo.push_back( bulk );
bulkToCommit.push_back( std::pair<IndexAccessMethod*,IndexAccessMethod*>( accessMethod, bulk ) );
}
else {
indexesToInsertTo.push_back( accessMethod );
}
indexBuildBlocks.mutableVector().push_back( block.release() );
}
// reset data size and record counts to 0 for this namespace
// as we're about to tally them up again for each new extent
d->setStats( 0, 0 );
ProgressMeterHolder pm(cc().curop()->setMessage("compact extent",
"Extent Compacting Progress",
extents.size()));
int extentNumber = 0;
for( list<DiskLoc>::iterator i = extents.begin(); i != extents.end(); i++ ) {
_compactExtent(*i, extentNumber++, indexesToInsertTo, compactOptions, &stats );
pm.hit();
}
verify( d->firstExtent().ext()->xprev.isNull() );
// indexes will do their own progress meter?
pm.finished();
log() << "starting index commits";
for ( size_t i = 0; i < bulkToCommit.size(); i++ ) {
bulkToCommit[i].first->commitBulk( bulkToCommit[i].second, false, NULL );
}
for ( size_t i = 0; i < indexBuildBlocks.size(); i++ ) {
IndexCatalog::IndexBuildBlock* block = indexBuildBlocks.mutableVector()[i];
block->success();
}
return StatusWith<CompactStats>( stats );
}
Status MMAPV1Engine::repairDatabase( OperationContext* txn,
const std::string& dbName,
bool preserveClonedFilesOnFailure,
bool backupOriginalFiles ) {
unique_ptr<RepairFileDeleter> repairFileDeleter;
// Must be done before and after repair
getDur().syncDataAndTruncateJournal(txn);
intmax_t totalSize = dbSize( dbName );
intmax_t freeSize = File::freeSpace(storageGlobalParams.repairpath);
if ( freeSize > -1 && freeSize < totalSize ) {
return Status( ErrorCodes::OutOfDiskSpace,
str::stream() << "Cannot repair database " << dbName
<< " having size: " << totalSize
<< " (bytes) because free disk space is: " << freeSize << " (bytes)" );
}
txn->checkForInterrupt();
Path reservedPath =
uniqueReservedPath( ( preserveClonedFilesOnFailure || backupOriginalFiles ) ?
"backup" : "_tmp" );
bool created = false;
MONGO_ASSERT_ON_EXCEPTION( created = boost::filesystem::create_directory( reservedPath ) );
invariant( created );
string reservedPathString = reservedPath.string();
if ( !preserveClonedFilesOnFailure )
repairFileDeleter.reset( new RepairFileDeleter( txn,
dbName,
reservedPathString,
reservedPath ) );
{
Database* originalDatabase = dbHolder().openDb(txn, dbName);
if (originalDatabase == NULL) {
return Status(ErrorCodes::NamespaceNotFound, "database does not exist to repair");
}
unique_ptr<MMAPV1DatabaseCatalogEntry> dbEntry;
unique_ptr<Database> tempDatabase;
// Must call this before MMAPV1DatabaseCatalogEntry's destructor closes the DB files
ON_BLOCK_EXIT(&dur::DurableInterface::syncDataAndTruncateJournal, &getDur(), txn);
{
dbEntry.reset(new MMAPV1DatabaseCatalogEntry(txn,
dbName,
reservedPathString,
storageGlobalParams.directoryperdb,
true));
tempDatabase.reset( new Database(txn, dbName, dbEntry.get()));
}
map<string,CollectionOptions> namespacesToCopy;
{
string ns = dbName + ".system.namespaces";
OldClientContext ctx(txn, ns );
Collection* coll = originalDatabase->getCollection( ns );
if ( coll ) {
auto cursor = coll->getCursor(txn);
while (auto record = cursor->next()) {
BSONObj obj = record->data.releaseToBson();
string ns = obj["name"].String();
NamespaceString nss( ns );
if ( nss.isSystem() ) {
if ( nss.isSystemDotIndexes() )
continue;
if ( nss.coll() == "system.namespaces" )
continue;
}
if ( !nss.isNormal() )
continue;
CollectionOptions options;
if ( obj["options"].isABSONObj() ) {
Status status = options.parse( obj["options"].Obj() );
if ( !status.isOK() )
return status;
}
namespacesToCopy[ns] = options;
}
}
}
for ( map<string,CollectionOptions>::const_iterator i = namespacesToCopy.begin();
i != namespacesToCopy.end();
++i ) {
string ns = i->first;
CollectionOptions options = i->second;
Collection* tempCollection = NULL;
{
WriteUnitOfWork wunit(txn);
tempCollection = tempDatabase->createCollection(txn, ns, options, false);
wunit.commit();
}
OldClientContext readContext(txn, ns, originalDatabase);
Collection* originalCollection = originalDatabase->getCollection( ns );
invariant( originalCollection );
// data
// TODO SERVER-14812 add a mode that drops duplicates rather than failing
MultiIndexBlock indexer(txn, tempCollection );
{
vector<BSONObj> indexes;
IndexCatalog::IndexIterator ii =
originalCollection->getIndexCatalog()->getIndexIterator( txn, false );
while ( ii.more() ) {
IndexDescriptor* desc = ii.next();
indexes.push_back( desc->infoObj() );
}
Status status = indexer.init( indexes );
if (!status.isOK()) {
return status;
}
}
auto cursor = originalCollection->getCursor(txn);
while (auto record = cursor->next()) {
BSONObj doc = record->data.releaseToBson();
WriteUnitOfWork wunit(txn);
StatusWith<RecordId> result = tempCollection->insertDocument(txn,
doc,
&indexer,
false);
if ( !result.isOK() )
return result.getStatus();
wunit.commit();
txn->checkForInterrupt();
}
Status status = indexer.doneInserting();
if (!status.isOK())
return status;
{
WriteUnitOfWork wunit(txn);
indexer.commit();
wunit.commit();
}
}
getDur().syncDataAndTruncateJournal(txn);
// need both in case journaling is disabled
MongoFile::flushAll(true);
txn->checkForInterrupt();
}
// at this point if we abort, we don't want to delete new files
// as they might be the only copies
if ( repairFileDeleter.get() )
repairFileDeleter->success();
// Close the database so we can rename/delete the original data files
dbHolder().close(txn, dbName);
if ( backupOriginalFiles ) {
_renameForBackup( dbName, reservedPath );
}
else {
// first make new directory before deleting data
Path newDir = Path(storageGlobalParams.dbpath) / dbName;
MONGO_ASSERT_ON_EXCEPTION(boost::filesystem::create_directory(newDir));
// this deletes old files
_deleteDataFiles( dbName );
if ( !boost::filesystem::exists(newDir) ) {
// we deleted because of directoryperdb
// re-create
MONGO_ASSERT_ON_EXCEPTION(boost::filesystem::create_directory(newDir));
}
}
_replaceWithRecovered( dbName, reservedPathString.c_str() );
if (!backupOriginalFiles) {
MONGO_ASSERT_ON_EXCEPTION(boost::filesystem::remove_all(reservedPath));
}
// Reopen the database so it's discoverable
dbHolder().openDb(txn, dbName);
return Status::OK();
}
StatusWith<CompactStats> Collection::compact( const CompactOptions* compactOptions ) {
if ( isCapped() )
return StatusWith<CompactStats>( ErrorCodes::BadValue,
"cannot compact capped collection" );
if ( _indexCatalog.numIndexesInProgress() )
return StatusWith<CompactStats>( ErrorCodes::BadValue,
"cannot compact when indexes in progress" );
NamespaceDetails* d = details();
// this is a big job, so might as well make things tidy before we start just to be nice.
getDur().commitIfNeeded();
list<DiskLoc> extents;
for( DiskLoc L = d->firstExtent(); !L.isNull(); L = L.ext()->xnext )
extents.push_back(L);
log() << "compact " << extents.size() << " extents" << endl;
// same data, but might perform a little different after compact?
_infoCache.reset();
vector<BSONObj> indexSpecs;
{
IndexCatalog::IndexIterator ii( _indexCatalog.getIndexIterator( false ) );
while ( ii.more() ) {
IndexDescriptor* descriptor = ii.next();
const BSONObj spec = _compactAdjustIndexSpec(descriptor->infoObj());
const BSONObj key = spec.getObjectField("key");
const Status keyStatus = validateKeyPattern(key);
if (!keyStatus.isOK()) {
return StatusWith<CompactStats>(
ErrorCodes::CannotCreateIndex,
str::stream() << "Cannot rebuild index " << spec << ": "
<< keyStatus.reason()
<< " For more info see"
<< " http://dochub.mongodb.org/core/index-validation");
}
indexSpecs.push_back(spec);
}
}
log() << "compact orphan deleted lists" << endl;
d->orphanDeletedList();
// Start over from scratch with our extent sizing and growth
d->setLastExtentSize( 0 );
// before dropping indexes, at least make sure we can allocate one extent!
// this will allocate an extent and add to free list
// if it cannot, it will throw an exception
increaseStorageSize( _details->lastExtentSize(), true );
// note that the drop indexes call also invalidates all clientcursors for the namespace,
// which is important and wanted here
log() << "compact dropping indexes" << endl;
Status status = _indexCatalog.dropAllIndexes( true );
if ( !status.isOK() ) {
return StatusWith<CompactStats>( status );
}
getDur().commitIfNeeded();
killCurrentOp.checkForInterrupt();
CompactStats stats;
MultiIndexBlock multiIndexBlock( this );
status = multiIndexBlock.init( indexSpecs );
if ( !status.isOK() )
return StatusWith<CompactStats>( status );
// reset data size and record counts to 0 for this namespace
// as we're about to tally them up again for each new extent
d->setStats( 0, 0 );
ProgressMeterHolder pm(cc().curop()->setMessage("compact extent",
"Extent Compacting Progress",
extents.size()));
int extentNumber = 0;
for( list<DiskLoc>::iterator i = extents.begin(); i != extents.end(); i++ ) {
_compactExtent(*i, extentNumber++, multiIndexBlock, compactOptions, &stats );
pm.hit();
}
verify( d->firstExtent().ext()->xprev.isNull() );
// indexes will do their own progress meter?
pm.finished();
log() << "starting index commits";
status = multiIndexBlock.commit();
if ( !status.isOK() )
return StatusWith<CompactStats>( status );
return StatusWith<CompactStats>( stats );
}
StatusWith<CompactStats> compactCollection(OperationContext* opCtx,
Collection* collection,
const CompactOptions* compactOptions) {
dassert(opCtx->lockState()->isCollectionLockedForMode(collection->ns().toString(), MODE_X));
DisableDocumentValidation validationDisabler(opCtx);
auto recordStore = collection->getRecordStore();
auto indexCatalog = collection->getIndexCatalog();
if (!recordStore->compactSupported())
return StatusWith<CompactStats>(ErrorCodes::CommandNotSupported,
str::stream()
<< "cannot compact collection with record store: "
<< recordStore->name());
if (recordStore->compactsInPlace()) {
CompactStats stats;
Status status = recordStore->compact(opCtx);
if (!status.isOK())
return StatusWith<CompactStats>(status);
// Compact all indexes (not including unfinished indexes)
std::unique_ptr<IndexCatalog::IndexIterator> ii(
indexCatalog->getIndexIterator(opCtx, false));
while (ii->more()) {
IndexCatalogEntry* entry = ii->next();
IndexDescriptor* descriptor = entry->descriptor();
IndexAccessMethod* iam = entry->accessMethod();
LOG(1) << "compacting index: " << descriptor->toString();
Status status = iam->compact(opCtx);
if (!status.isOK()) {
error() << "failed to compact index: " << descriptor->toString();
return status;
}
}
return StatusWith<CompactStats>(stats);
}
if (indexCatalog->numIndexesInProgress(opCtx))
return StatusWith<CompactStats>(ErrorCodes::BadValue,
"cannot compact when indexes in progress");
std::vector<BSONObj> indexSpecs;
{
std::unique_ptr<IndexCatalog::IndexIterator> ii(
indexCatalog->getIndexIterator(opCtx, false));
while (ii->more()) {
IndexDescriptor* descriptor = ii->next()->descriptor();
// Compact always creates the new index in the foreground.
const BSONObj spec =
descriptor->infoObj().removeField(IndexDescriptor::kBackgroundFieldName);
const BSONObj key = spec.getObjectField("key");
const Status keyStatus =
index_key_validate::validateKeyPattern(key, descriptor->version());
if (!keyStatus.isOK()) {
return StatusWith<CompactStats>(
ErrorCodes::CannotCreateIndex,
str::stream() << "Cannot compact collection due to invalid index " << spec
<< ": "
<< keyStatus.reason()
<< " For more info see"
<< " http://dochub.mongodb.org/core/index-validation");
}
indexSpecs.push_back(spec);
}
}
// Give a chance to be interrupted *before* we drop all indexes.
opCtx->checkForInterrupt();
{
// note that the drop indexes call also invalidates all clientcursors for the namespace,
// which is important and wanted here
WriteUnitOfWork wunit(opCtx);
log() << "compact dropping indexes";
indexCatalog->dropAllIndexes(opCtx, true);
wunit.commit();
}
CompactStats stats;
MultiIndexBlockImpl indexer(opCtx, collection);
indexer.allowInterruption();
indexer.ignoreUniqueConstraint(); // in compact we should be doing no checking
Status status = indexer.init(indexSpecs).getStatus();
if (!status.isOK())
return StatusWith<CompactStats>(status);
status = recordStore->compact(opCtx);
if (!status.isOK())
return StatusWith<CompactStats>(status);
log() << "starting index commits";
status = indexer.dumpInsertsFromBulk();
if (!status.isOK())
return StatusWith<CompactStats>(status);
{
WriteUnitOfWork wunit(opCtx);
status = indexer.commit();
if (!status.isOK()) {
return StatusWith<CompactStats>(status);
}
wunit.commit();
}
return StatusWith<CompactStats>(stats);
}
Status repairDatabase( string dbName,
bool preserveClonedFilesOnFailure,
bool backupOriginalFiles ) {
scoped_ptr<RepairFileDeleter> repairFileDeleter;
doingRepair dr;
dbName = nsToDatabase( dbName );
log() << "repairDatabase " << dbName << endl;
invariant( cc().database()->name() == dbName );
invariant( cc().database()->path() == storageGlobalParams.dbpath );
BackgroundOperation::assertNoBgOpInProgForDb(dbName);
getDur().syncDataAndTruncateJournal(); // Must be done before and after repair
intmax_t totalSize = dbSize( dbName );
intmax_t freeSize = File::freeSpace(storageGlobalParams.repairpath);
if ( freeSize > -1 && freeSize < totalSize ) {
return Status( ErrorCodes::OutOfDiskSpace,
str::stream() << "Cannot repair database " << dbName
<< " having size: " << totalSize
<< " (bytes) because free disk space is: " << freeSize << " (bytes)" );
}
killCurrentOp.checkForInterrupt();
Path reservedPath =
uniqueReservedPath( ( preserveClonedFilesOnFailure || backupOriginalFiles ) ?
"backup" : "_tmp" );
MONGO_ASSERT_ON_EXCEPTION( boost::filesystem::create_directory( reservedPath ) );
string reservedPathString = reservedPath.string();
if ( !preserveClonedFilesOnFailure )
repairFileDeleter.reset( new RepairFileDeleter( dbName,
reservedPathString,
reservedPath ) );
{
Database* originalDatabase = dbHolder().get( dbName, storageGlobalParams.dbpath );
if ( originalDatabase == NULL )
return Status( ErrorCodes::NamespaceNotFound, "database does not exist to repair" );
Database* tempDatabase = NULL;
{
bool justCreated = false;
tempDatabase = dbHolderW().getOrCreate( dbName, reservedPathString, justCreated );
invariant( justCreated );
}
map<string,CollectionOptions> namespacesToCopy;
{
string ns = dbName + ".system.namespaces";
Client::Context ctx( ns );
Collection* coll = originalDatabase->getCollection( ns );
if ( coll ) {
scoped_ptr<CollectionIterator> it( coll->getIterator( DiskLoc(),
false,
CollectionScanParams::FORWARD ) );
while ( !it->isEOF() ) {
DiskLoc loc = it->getNext();
BSONObj obj = coll->docFor( loc );
string ns = obj["name"].String();
NamespaceString nss( ns );
if ( nss.isSystem() ) {
if ( nss.isSystemDotIndexes() )
continue;
if ( nss.coll() == "system.namespaces" )
continue;
}
if ( !nss.isNormal() )
continue;
CollectionOptions options;
if ( obj["options"].isABSONObj() ) {
Status status = options.parse( obj["options"].Obj() );
if ( !status.isOK() )
return status;
}
namespacesToCopy[ns] = options;
}
}
}
for ( map<string,CollectionOptions>::const_iterator i = namespacesToCopy.begin();
i != namespacesToCopy.end();
++i ) {
string ns = i->first;
CollectionOptions options = i->second;
Collection* tempCollection = NULL;
{
Client::Context tempContext( ns, tempDatabase );
tempCollection = tempDatabase->createCollection( ns, options, true, false );
}
Client::Context readContext( ns, originalDatabase );
Collection* originalCollection = originalDatabase->getCollection( ns );
invariant( originalCollection );
// data
MultiIndexBlock indexBlock( tempCollection );
{
vector<BSONObj> indexes;
IndexCatalog::IndexIterator ii =
originalCollection->getIndexCatalog()->getIndexIterator( false );
while ( ii.more() ) {
IndexDescriptor* desc = ii.next();
indexes.push_back( desc->infoObj() );
}
Client::Context tempContext( ns, tempDatabase );
Status status = indexBlock.init( indexes );
if ( !status.isOK() )
return status;
}
scoped_ptr<CollectionIterator> iterator( originalCollection->getIterator( DiskLoc(),
false,
CollectionScanParams::FORWARD ) );
while ( !iterator->isEOF() ) {
DiskLoc loc = iterator->getNext();
invariant( !loc.isNull() );
BSONObj doc = originalCollection->docFor( loc );
Client::Context tempContext( ns, tempDatabase );
StatusWith<DiskLoc> result = tempCollection->insertDocument( doc, indexBlock );
if ( !result.isOK() )
return result.getStatus();
getDur().commitIfNeeded();
killCurrentOp.checkForInterrupt(false);
}
{
Client::Context tempContext( ns, tempDatabase );
Status status = indexBlock.commit();
if ( !status.isOK() )
return status;
}
}
getDur().syncDataAndTruncateJournal();
MongoFile::flushAll(true); // need both in case journaling is disabled
killCurrentOp.checkForInterrupt(false);
Client::Context tempContext( dbName, reservedPathString );
Database::closeDatabase( dbName, reservedPathString );
}
// at this point if we abort, we don't want to delete new files
// as they might be the only copies
if ( repairFileDeleter.get() )
repairFileDeleter->success();
Client::Context ctx( dbName );
Database::closeDatabase(dbName, storageGlobalParams.dbpath);
if ( backupOriginalFiles ) {
_renameForBackup( dbName, reservedPath );
}
else {
// first make new directory before deleting data
Path newDir = Path(storageGlobalParams.dbpath) / dbName;
MONGO_ASSERT_ON_EXCEPTION(boost::filesystem::create_directory(newDir));
// this deletes old files
_deleteDataFiles( dbName );
if ( !boost::filesystem::exists(newDir) ) {
// we deleted because of directoryperdb
// re-create
MONGO_ASSERT_ON_EXCEPTION(boost::filesystem::create_directory(newDir));
}
}
_replaceWithRecovered( dbName, reservedPathString.c_str() );
if ( !backupOriginalFiles )
MONGO_ASSERT_ON_EXCEPTION( boost::filesystem::remove_all( reservedPath ) );
return Status::OK();
}
StatusWith<CompactStats> Collection::compact(OperationContext* txn,
const CompactOptions* compactOptions) {
dassert(txn->lockState()->isCollectionLockedForMode(ns().toString(), MODE_X));
DisableDocumentValidation validationDisabler(txn);
if (!_recordStore->compactSupported())
return StatusWith<CompactStats>(ErrorCodes::CommandNotSupported,
str::stream()
<< "cannot compact collection with record store: "
<< _recordStore->name());
if (_recordStore->compactsInPlace()) {
CompactStats stats;
Status status = _recordStore->compact(txn, NULL, compactOptions, &stats);
if (!status.isOK())
return StatusWith<CompactStats>(status);
// Compact all indexes (not including unfinished indexes)
IndexCatalog::IndexIterator ii(_indexCatalog.getIndexIterator(txn, false));
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
IndexAccessMethod* index = _indexCatalog.getIndex(descriptor);
LOG(1) << "compacting index: " << descriptor->toString();
Status status = index->compact(txn);
if (!status.isOK()) {
error() << "failed to compact index: " << descriptor->toString();
return status;
}
}
return StatusWith<CompactStats>(stats);
}
if (_indexCatalog.numIndexesInProgress(txn))
return StatusWith<CompactStats>(ErrorCodes::BadValue,
"cannot compact when indexes in progress");
vector<BSONObj> indexSpecs;
{
IndexCatalog::IndexIterator ii(_indexCatalog.getIndexIterator(txn, false));
while (ii.more()) {
IndexDescriptor* descriptor = ii.next();
const BSONObj spec = _compactAdjustIndexSpec(descriptor->infoObj());
const BSONObj key = spec.getObjectField("key");
const Status keyStatus = validateKeyPattern(key);
if (!keyStatus.isOK()) {
return StatusWith<CompactStats>(
ErrorCodes::CannotCreateIndex,
str::stream() << "Cannot compact collection due to invalid index " << spec
<< ": "
<< keyStatus.reason()
<< " For more info see"
<< " http://dochub.mongodb.org/core/index-validation");
}
indexSpecs.push_back(spec);
}
}
// Give a chance to be interrupted *before* we drop all indexes.
txn->checkForInterrupt();
{
// note that the drop indexes call also invalidates all clientcursors for the namespace,
// which is important and wanted here
WriteUnitOfWork wunit(txn);
log() << "compact dropping indexes";
Status status = _indexCatalog.dropAllIndexes(txn, true);
if (!status.isOK()) {
return StatusWith<CompactStats>(status);
}
wunit.commit();
}
CompactStats stats;
MultiIndexBlock indexer(txn, this);
indexer.allowInterruption();
indexer.ignoreUniqueConstraint(); // in compact we should be doing no checking
Status status = indexer.init(indexSpecs);
if (!status.isOK())
return StatusWith<CompactStats>(status);
MyCompactAdaptor adaptor(this, &indexer);
status = _recordStore->compact(txn, &adaptor, compactOptions, &stats);
if (!status.isOK())
return StatusWith<CompactStats>(status);
log() << "starting index commits";
status = indexer.doneInserting();
if (!status.isOK())
return StatusWith<CompactStats>(status);
{
WriteUnitOfWork wunit(txn);
indexer.commit();
wunit.commit();
}
return StatusWith<CompactStats>(stats);
}
/**
* For a given query, get a runner. The runner could be a SingleSolutionRunner, a
* CachedQueryRunner, or a MultiPlanRunner, depending on the cache/query solver/etc.
*/
Status getRunner(CanonicalQuery* rawCanonicalQuery, Runner** out, size_t plannerOptions) {
verify(rawCanonicalQuery);
auto_ptr<CanonicalQuery> canonicalQuery(rawCanonicalQuery);
// Try to look up a cached solution for the query.
// TODO: Can the cache have negative data about a solution?
PlanCache* localCache = PlanCache::get(canonicalQuery->ns());
if (NULL != localCache) {
CachedSolution* cs = localCache->get(*canonicalQuery);
if (NULL != cs) {
// We have a cached solution. Hand the canonical query and cached solution off to
// the cached plan runner, which takes ownership of both.
WorkingSet* ws;
PlanStage* root;
verify(StageBuilder::build(*cs->solution, &root, &ws));
*out = new CachedPlanRunner(canonicalQuery.release(), cs, root, ws);
return Status::OK();
}
}
// No entry in cache for the query. We have to solve the query ourself.
// Get the indices that we could possibly use.
Database* db = cc().database();
verify( db );
Collection* collection = db->getCollection( canonicalQuery->ns() );
// This can happen as we're called by internal clients as well.
if (NULL == collection) {
const string& ns = canonicalQuery->ns();
*out = new EOFRunner(canonicalQuery.release(), ns);
return Status::OK();
}
// If we have an _id index we can use the idhack runner.
if (canUseIDHack(*canonicalQuery) && collection->getIndexCatalog()->findIdIndex()) {
*out = new IDHackRunner(collection, canonicalQuery.release());
return Status::OK();
}
// If it's not NULL, we may have indices. Access the catalog and fill out IndexEntry(s)
QueryPlannerParams plannerParams;
for (int i = 0; i < collection->getIndexCatalog()->numIndexesReady(); ++i) {
IndexDescriptor* desc = collection->getIndexCatalog()->getDescriptor( i );
plannerParams.indices.push_back(IndexEntry(desc->keyPattern(),
desc->isMultikey(),
desc->isSparse(),
desc->indexName()));
}
// Tailable: If the query requests tailable the collection must be capped.
if (canonicalQuery->getParsed().hasOption(QueryOption_CursorTailable)) {
if (!collection->isCapped()) {
return Status(ErrorCodes::BadValue,
"tailable cursor requested on non capped collection");
}
// If a sort is specified it must be equal to expectedSort.
const BSONObj expectedSort = BSON("$natural" << 1);
const BSONObj& actualSort = canonicalQuery->getParsed().getSort();
if (!actualSort.isEmpty() && !(actualSort == expectedSort)) {
return Status(ErrorCodes::BadValue,
"invalid sort specified for tailable cursor: "
+ actualSort.toString());
}
}
// Process the planning options.
plannerParams.options = plannerOptions;
if (storageGlobalParams.noTableScan) {
const string& ns = canonicalQuery->ns();
// There are certain cases where we ignore this restriction:
bool ignore = canonicalQuery->getQueryObj().isEmpty()
|| (string::npos != ns.find(".system."))
|| (0 == ns.find("local."));
if (!ignore) {
plannerParams.options |= QueryPlannerParams::NO_TABLE_SCAN;
}
}
if (!(plannerParams.options & QueryPlannerParams::NO_TABLE_SCAN)) {
plannerParams.options |= QueryPlannerParams::INCLUDE_COLLSCAN;
}
// If the caller wants a shard filter, make sure we're actually sharded.
if (plannerParams.options & QueryPlannerParams::INCLUDE_SHARD_FILTER) {
CollectionMetadataPtr collMetadata = shardingState.getCollectionMetadata(canonicalQuery->ns());
if (collMetadata) {
plannerParams.shardKey = collMetadata->getKeyPattern();
}
else {
// If there's no metadata don't bother w/the shard filter since we won't know what
// the key pattern is anyway...
plannerParams.options &= ~QueryPlannerParams::INCLUDE_SHARD_FILTER;
}
}
vector<QuerySolution*> solutions;
QueryPlanner::plan(*canonicalQuery, plannerParams, &solutions);
/*
for (size_t i = 0; i < solutions.size(); ++i) {
QLOG() << "solution " << i << " is " << solutions[i]->toString() << endl;
}
*/
// We cannot figure out how to answer the query. Should this ever happen?
if (0 == solutions.size()) {
return Status(ErrorCodes::BadValue, "No query solutions");
}
if (1 == solutions.size()) {
// Only one possible plan. Run it. Build the stages from the solution.
WorkingSet* ws;
PlanStage* root;
verify(StageBuilder::build(*solutions[0], &root, &ws));
// And, run the plan.
*out = new SingleSolutionRunner(canonicalQuery.release(), solutions[0], root, ws);
return Status::OK();
}
else {
// Many solutions. Let the MultiPlanRunner pick the best, update the cache, and so on.
auto_ptr<MultiPlanRunner> mpr(new MultiPlanRunner(canonicalQuery.release()));
for (size_t i = 0; i < solutions.size(); ++i) {
WorkingSet* ws;
PlanStage* root;
verify(StageBuilder::build(*solutions[i], &root, &ws));
// Takes ownership of all arguments.
mpr->addPlan(solutions[i], root, ws);
}
*out = mpr.release();
return Status::OK();
}
}
StatusWith<DiskLoc> Collection::updateDocument( const DiskLoc& oldLocation,
const BSONObj& objNew,
bool enforceQuota,
OpDebug* debug ) {
Record* oldRecord = getExtentManager()->recordFor( oldLocation );
BSONObj objOld = BSONObj::make( oldRecord );
if ( objOld.hasElement( "_id" ) ) {
BSONElement oldId = objOld["_id"];
BSONElement newId = objNew["_id"];
if ( oldId != newId )
return StatusWith<DiskLoc>( ErrorCodes::InternalError,
"in Collection::updateDocument _id mismatch",
13596 );
}
if ( ns().coll() == "system.users" ) {
// XXX - andy and spencer think this should go away now
V2UserDocumentParser parser;
Status s = parser.checkValidUserDocument(objNew);
if ( !s.isOK() )
return StatusWith<DiskLoc>( s );
}
/* duplicate key check. we descend the btree twice - once for this check, and once for the actual inserts, further
below. that is suboptimal, but it's pretty complicated to do it the other way without rollbacks...
*/
OwnedPointerVector<UpdateTicket> updateTickets;
updateTickets.mutableVector().resize(_indexCatalog.numIndexesTotal());
for (int i = 0; i < _indexCatalog.numIndexesTotal(); ++i) {
IndexDescriptor* descriptor = _indexCatalog.getDescriptor( i );
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
InsertDeleteOptions options;
options.logIfError = false;
options.dupsAllowed =
!(KeyPattern::isIdKeyPattern(descriptor->keyPattern()) || descriptor->unique())
|| ignoreUniqueIndex(descriptor);
updateTickets.mutableVector()[i] = new UpdateTicket();
Status ret = iam->validateUpdate(objOld, objNew, oldLocation, options,
updateTickets.mutableVector()[i]);
if ( !ret.isOK() ) {
return StatusWith<DiskLoc>( ret );
}
}
if ( oldRecord->netLength() < objNew.objsize() ) {
// doesn't fit, have to move to new location
if ( _details->isCapped() )
return StatusWith<DiskLoc>( ErrorCodes::InternalError,
"failing update: objects in a capped ns cannot grow",
10003 );
moveCounter.increment();
_details->paddingTooSmall();
// unindex old record, don't delete
// this way, if inserting new doc fails, we can re-index this one
ClientCursor::aboutToDelete(_ns.ns(), _details, oldLocation);
_indexCatalog.unindexRecord( objOld, oldLocation, true );
if ( debug ) {
if (debug->nmoved == -1) // default of -1 rather than 0
debug->nmoved = 1;
else
debug->nmoved += 1;
}
StatusWith<DiskLoc> loc = insertDocument( objNew, enforceQuota );
if ( loc.isOK() ) {
// insert successful, now lets deallocate the old location
// remember its already unindexed
_recordStore.deallocRecord( oldLocation, oldRecord );
}
else {
// new doc insert failed, so lets re-index the old document and location
_indexCatalog.indexRecord( objOld, oldLocation );
}
return loc;
}
_infoCache.notifyOfWriteOp();
_details->paddingFits();
if ( debug )
debug->keyUpdates = 0;
for (int i = 0; i < _indexCatalog.numIndexesTotal(); ++i) {
IndexDescriptor* descriptor = _indexCatalog.getDescriptor( i );
IndexAccessMethod* iam = _indexCatalog.getIndex( descriptor );
int64_t updatedKeys;
Status ret = iam->update(*updateTickets.vector()[i], &updatedKeys);
if ( !ret.isOK() )
return StatusWith<DiskLoc>( ret );
if ( debug )
debug->keyUpdates += updatedKeys;
}
// update in place
int sz = objNew.objsize();
memcpy(getDur().writingPtr(oldRecord->data(), sz), objNew.objdata(), sz);
return StatusWith<DiskLoc>( oldLocation );
}
bool wrappedRun(OperationContext* txn,
const string& dbname,
BSONObj& jsobj,
string& errmsg,
BSONObjBuilder& anObjBuilder) {
const std::string coll = jsobj.firstElement().valuestrsafe();
if (coll.empty()) {
errmsg = "no collection name specified";
return false;
}
const std::string toDeleteNs = dbname + '.' + coll;
if (!serverGlobalParams.quiet) {
LOG(0) << "CMD: dropIndexes " << toDeleteNs << endl;
}
Client::Context ctx(txn, toDeleteNs);
Database* db = ctx.db();
Collection* collection = db->getCollection( txn, toDeleteNs );
if ( ! collection ) {
errmsg = "ns not found";
return false;
}
stopIndexBuilds(txn, db, jsobj);
IndexCatalog* indexCatalog = collection->getIndexCatalog();
anObjBuilder.appendNumber("nIndexesWas", indexCatalog->numIndexesTotal(txn) );
BSONElement f = jsobj.getField("index");
if ( f.type() == String ) {
string indexToDelete = f.valuestr();
if ( indexToDelete == "*" ) {
Status s = indexCatalog->dropAllIndexes(txn, false);
if ( !s.isOK() ) {
appendCommandStatus( anObjBuilder, s );
return false;
}
anObjBuilder.append("msg", "non-_id indexes dropped for collection");
return true;
}
IndexDescriptor* desc = collection->getIndexCatalog()->findIndexByName( txn,
indexToDelete );
if ( desc == NULL ) {
errmsg = str::stream() << "index not found with name [" << indexToDelete << "]";
return false;
}
if ( desc->isIdIndex() ) {
errmsg = "cannot drop _id index";
return false;
}
Status s = indexCatalog->dropIndex(txn, desc);
if ( !s.isOK() ) {
appendCommandStatus( anObjBuilder, s );
return false;
}
return true;
}
if ( f.type() == Object ) {
IndexDescriptor* desc =
collection->getIndexCatalog()->findIndexByKeyPattern( txn, f.embeddedObject() );
if ( desc == NULL ) {
errmsg = "can't find index with key:";
errmsg += f.embeddedObject().toString();
return false;
}
if ( desc->isIdIndex() ) {
errmsg = "cannot drop _id index";
return false;
}
Status s = indexCatalog->dropIndex(txn, desc);
if ( !s.isOK() ) {
appendCommandStatus( anObjBuilder, s );
return false;
}
return true;
}
errmsg = "invalid index name spec";
return false;
}
shared_ptr<Cursor> QueryPlan::newCursor( const DiskLoc& startLoc,
bool requestIntervalCursor ) const {
if ( _type ) {
// hopefully safe to use original query in these contexts - don't think we can mix type
// with $or clause separation yet
int numWanted = 0;
if ( _parsedQuery ) {
// SERVER-5390
numWanted = _parsedQuery->getSkip() + _parsedQuery->getNumToReturn();
}
IndexDescriptor* descriptor = CatalogHack::getDescriptor(_d, _idxNo);
IndexAccessMethod* iam = CatalogHack::getIndex(descriptor);
return shared_ptr<Cursor>(EmulatedCursor::make(descriptor, iam, _originalQuery,
_order, numWanted,
descriptor->keyPattern()));
}
if ( _utility == Impossible ) {
// Dummy table scan cursor returning no results. Allowed in --notablescan mode.
return shared_ptr<Cursor>( new BasicCursor( DiskLoc() ) );
}
if ( willScanTable() ) {
checkTableScanAllowed();
return findTableScan( _frs.ns(), _order, startLoc );
}
massert( 10363,
"newCursor() with start location not implemented for indexed plans",
startLoc.isNull() );
if ( _startOrEndSpec ) {
// we are sure to spec _endKeyInclusive
return shared_ptr<Cursor>( BtreeCursor::make( _d,
*_index,
_startKey,
_endKey,
_endKeyInclusive,
_direction >= 0 ? 1 : -1 ) );
}
if ( _index->getSpec().getType() ) {
return shared_ptr<Cursor>( BtreeCursor::make( _d,
*_index,
_frv->startKey(),
_frv->endKey(),
true,
_direction >= 0 ? 1 : -1 ) );
}
// An IntervalBtreeCursor is returned if explicitly requested AND _frv is exactly
// represented by a single interval within the btree.
if ( // If an interval cursor is requested and ...
requestIntervalCursor &&
// ... equalities come before ranges (a requirement of Optimal) and ...
_utility == Optimal &&
// ... the field range vector exactly represents a single interval ...
_frv->isSingleInterval() ) {
// ... and an interval cursor can be created ...
shared_ptr<Cursor> ret( IntervalBtreeCursor::make( _d,
*_index,
_frv->startKey(),
_frv->startKeyInclusive(),
_frv->endKey(),
_frv->endKeyInclusive() ) );
if ( ret ) {
// ... then return the interval cursor.
return ret;
}
}
return shared_ptr<Cursor>( BtreeCursor::make( _d,
*_index,
_frv,
independentRangesSingleIntervalLimit(),
_direction >= 0 ? 1 : -1 ) );
}