virtual bool run(OperationContext* txn,
const string& db,
BSONObj& cmdObj,
int options,
string& errmsg,
BSONObjBuilder& result) {
const std::string ns = parseNs(db, cmdObj);
if (nsToCollectionSubstring(ns).empty()) {
errmsg = "missing collection name";
return false;
}
NamespaceString nss(ns);
// Parse the options for this request.
auto request = AggregationRequest::parseFromBSON(nss, cmdObj);
if (!request.isOK()) {
return appendCommandStatus(result, request.getStatus());
}
// Set up the ExpressionContext.
intrusive_ptr<ExpressionContext> expCtx = new ExpressionContext(txn, request.getValue());
expCtx->tempDir = storageGlobalParams.dbpath + "/_tmp";
// Parse the pipeline.
auto statusWithPipeline = Pipeline::parse(request.getValue().getPipeline(), expCtx);
if (!statusWithPipeline.isOK()) {
return appendCommandStatus(result, statusWithPipeline.getStatus());
}
auto pipeline = std::move(statusWithPipeline.getValue());
auto resolvedNamespaces = resolveInvolvedNamespaces(txn, pipeline, expCtx);
if (!resolvedNamespaces.isOK()) {
return appendCommandStatus(result, resolvedNamespaces.getStatus());
}
expCtx->resolvedNamespaces = std::move(resolvedNamespaces.getValue());
unique_ptr<ClientCursorPin> pin; // either this OR the exec will be non-null
unique_ptr<PlanExecutor> exec;
auto curOp = CurOp::get(txn);
{
// This will throw if the sharding version for this connection is out of date. If the
// namespace is a view, the lock will be released before re-running the aggregation.
// Otherwise, the lock must be held continuously from now until we have we created both
// the output ClientCursor and the input executor. This ensures that both are using the
// same sharding version that we synchronize on here. This is also why we always need to
// create a ClientCursor even when we aren't outputting to a cursor. See the comment on
// ShardFilterStage for more details.
AutoGetCollectionOrViewForRead ctx(txn, nss);
Collection* collection = ctx.getCollection();
// If running $collStats on a view, we do not resolve the view since we want stats
// on this view namespace.
auto startsWithCollStats = [&pipeline]() {
const Pipeline::SourceContainer& sources = pipeline->getSources();
return !sources.empty() &&
dynamic_cast<DocumentSourceCollStats*>(sources.front().get());
};
// If this is a view, resolve it by finding the underlying collection and stitching view
// pipelines and this request's pipeline together. We then release our locks before
// recursively calling run, which will re-acquire locks on the underlying collection.
// (The lock must be released because recursively acquiring locks on the database will
// prohibit yielding.)
auto view = ctx.getView();
if (view && !startsWithCollStats()) {
auto viewDefinition =
ViewShardingCheck::getResolvedViewIfSharded(txn, ctx.getDb(), view);
if (!viewDefinition.isOK()) {
return appendCommandStatus(result, viewDefinition.getStatus());
}
if (!viewDefinition.getValue().isEmpty()) {
ViewShardingCheck::appendShardedViewStatus(viewDefinition.getValue(), &result);
return false;
}
auto resolvedView = ctx.getDb()->getViewCatalog()->resolveView(txn, nss);
if (!resolvedView.isOK()) {
return appendCommandStatus(result, resolvedView.getStatus());
}
// With the view resolved, we can relinquish locks.
ctx.releaseLocksForView();
// Parse the resolved view into a new aggregation request.
auto viewCmd =
resolvedView.getValue().asExpandedViewAggregation(request.getValue());
if (!viewCmd.isOK()) {
return appendCommandStatus(result, viewCmd.getStatus());
}
bool status = this->run(txn, db, viewCmd.getValue(), options, errmsg, result);
{
// Set the namespace of the curop back to the view namespace so ctx records
// stats on this view namespace on destruction.
stdx::lock_guard<Client>(*txn->getClient());
curOp->setNS_inlock(nss.ns());
}
return status;
}
// If the pipeline does not have a user-specified collation, set it from the collection
// default.
if (request.getValue().getCollation().isEmpty() && collection &&
collection->getDefaultCollator()) {
invariant(!expCtx->getCollator());
expCtx->setCollator(collection->getDefaultCollator()->clone());
}
// Propagate the ExpressionContext throughout all of the pipeline's stages and
// expressions.
pipeline->injectExpressionContext(expCtx);
// The pipeline must be optimized after the correct collator has been set on it (by
// injecting the ExpressionContext containing the collator). This is necessary because
// optimization may make string comparisons, e.g. optimizing {$eq: [<str1>, <str2>]} to
// a constant.
pipeline->optimizePipeline();
if (kDebugBuild && !expCtx->isExplain && !expCtx->inShard) {
// Make sure all operations round-trip through Pipeline::serialize() correctly by
// re-parsing every command in debug builds. This is important because sharded
// aggregations rely on this ability. Skipping when inShard because this has
// already been through the transformation (and this un-sets expCtx->inShard).
pipeline = reparsePipeline(pipeline, request.getValue(), expCtx);
}
// This does mongod-specific stuff like creating the input PlanExecutor and adding
// it to the front of the pipeline if needed.
PipelineD::prepareCursorSource(collection, pipeline);
// Create the PlanExecutor which returns results from the pipeline. The WorkingSet
// ('ws') and the PipelineProxyStage ('proxy') will be owned by the created
// PlanExecutor.
auto ws = make_unique<WorkingSet>();
auto proxy = make_unique<PipelineProxyStage>(txn, pipeline, ws.get());
auto statusWithPlanExecutor = (NULL == collection)
? PlanExecutor::make(
txn, std::move(ws), std::move(proxy), nss.ns(), PlanExecutor::YIELD_MANUAL)
: PlanExecutor::make(
txn, std::move(ws), std::move(proxy), collection, PlanExecutor::YIELD_MANUAL);
invariant(statusWithPlanExecutor.isOK());
exec = std::move(statusWithPlanExecutor.getValue());
{
auto planSummary = Explain::getPlanSummary(exec.get());
stdx::lock_guard<Client>(*txn->getClient());
curOp->setPlanSummary_inlock(std::move(planSummary));
}
if (collection) {
PlanSummaryStats stats;
Explain::getSummaryStats(*exec, &stats);
collection->infoCache()->notifyOfQuery(txn, stats.indexesUsed);
}
if (collection) {
const bool isAggCursor = true; // enable special locking behavior
ClientCursor* cursor =
new ClientCursor(collection->getCursorManager(),
exec.release(),
nss.ns(),
txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
0,
cmdObj.getOwned(),
isAggCursor);
pin.reset(new ClientCursorPin(collection->getCursorManager(), cursor->cursorid()));
// Don't add any code between here and the start of the try block.
}
// At this point, it is safe to release the collection lock.
// - In the case where we have a collection: we will need to reacquire the
// collection lock later when cleaning up our ClientCursorPin.
// - In the case where we don't have a collection: our PlanExecutor won't be
// registered, so it will be safe to clean it up outside the lock.
invariant(!exec || !collection);
}
try {
// Unless set to true, the ClientCursor created above will be deleted on block exit.
bool keepCursor = false;
// Use of the aggregate command without specifying to use a cursor is deprecated.
// Applications should migrate to using cursors. Cursors are strictly more useful than
// outputting the results as a single document, since results that fit inside a single
// BSONObj will also fit inside a single batch.
//
// We occasionally log a deprecation warning.
if (!request.getValue().isCursorCommand()) {
RARELY {
warning()
<< "Use of the aggregate command without the 'cursor' "
"option is deprecated. See "
"http://dochub.mongodb.org/core/aggregate-without-cursor-deprecation.";
}
}
// If both explain and cursor are specified, explain wins.
if (expCtx->isExplain) {
result << "stages" << Value(pipeline->writeExplainOps());
} else if (request.getValue().isCursorCommand()) {
keepCursor = handleCursorCommand(txn,
nss.ns(),
pin.get(),
pin ? pin->c()->getExecutor() : exec.get(),
request.getValue(),
result);
} else {
pipeline->run(result);
}
if (!expCtx->isExplain) {
PlanSummaryStats stats;
Explain::getSummaryStats(pin ? *pin->c()->getExecutor() : *exec.get(), &stats);
curOp->debug().setPlanSummaryMetrics(stats);
curOp->debug().nreturned = stats.nReturned;
}
// Clean up our ClientCursorPin, if needed. We must reacquire the collection lock
// in order to do so.
if (pin) {
// We acquire locks here with DBLock and CollectionLock instead of using
// AutoGetCollectionForRead. AutoGetCollectionForRead will throw if the
// sharding version is out of date, and we don't care if the sharding version
// has changed.
Lock::DBLock dbLock(txn->lockState(), nss.db(), MODE_IS);
Lock::CollectionLock collLock(txn->lockState(), nss.ns(), MODE_IS);
if (keepCursor) {
pin->release();
} else {
pin->deleteUnderlying();
}
}
} catch (...) {
virtual bool run(OperationContext* txn, const string& dbname, BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool fromRepl) {
if ( cmdObj.firstElement().type() != Array ) {
errmsg = "ops has to be an array";
return false;
}
BSONObj ops = cmdObj.firstElement().Obj();
{
// check input
BSONObjIterator i( ops );
while ( i.more() ) {
BSONElement e = i.next();
if (!_checkOperation(e, errmsg)) {
return false;
}
}
}
// SERVER-4328 todo : is global ok or does this take a long time? i believe multiple
// ns used so locking individually requires more analysis
ScopedTransaction scopedXact(txn, MODE_X);
Lock::GlobalWrite globalWriteLock(txn->lockState());
if (!fromRepl &&
!repl::getGlobalReplicationCoordinator()->canAcceptWritesForDatabase(dbname)) {
return appendCommandStatus(result, Status(ErrorCodes::NotMaster, str::stream()
<< "Not primary while applying ops to database " << dbname));
}
// Preconditions check reads the database state, so needs to be done locked
if ( cmdObj["preCondition"].type() == Array ) {
BSONObjIterator i( cmdObj["preCondition"].Obj() );
while ( i.more() ) {
BSONObj f = i.next().Obj();
DBDirectClient db( txn );
BSONObj realres = db.findOne( f["ns"].String() , f["q"].Obj() );
// Apply-ops would never have a $where matcher, so use the default callback,
// which will throw an error if $where is found.
Matcher m(f["res"].Obj());
if ( ! m.matches( realres ) ) {
result.append( "got" , realres );
result.append( "whatFailed" , f );
errmsg = "pre-condition failed";
return false;
}
}
}
// apply
int num = 0;
int errors = 0;
BSONObjIterator i( ops );
BSONArrayBuilder ab;
const bool alwaysUpsert = cmdObj.hasField("alwaysUpsert") ?
cmdObj["alwaysUpsert"].trueValue() : true;
while ( i.more() ) {
BSONElement e = i.next();
const BSONObj& temp = e.Obj();
// Ignore 'n' operations.
const char *opType = temp["op"].valuestrsafe();
if (*opType == 'n') continue;
const string ns = temp["ns"].String();
// Run operations under a nested lock as a hack to prevent yielding.
//
// The list of operations is supposed to be applied atomically; yielding
// would break atomicity by allowing an interruption or a shutdown to occur
// after only some operations are applied. We are already locked globally
// at this point, so taking a DBLock on the namespace creates a nested lock,
// and yields are disallowed for operations that hold a nested lock.
//
// We do not have a wrapping WriteUnitOfWork so it is possible for a journal
// commit to happen with a subset of ops applied.
// TODO figure out what to do about this.
Lock::GlobalWrite globalWriteLockDisallowTempRelease(txn->lockState());
// Ensures that yielding will not happen (see the comment above).
DEV {
Locker::LockSnapshot lockSnapshot;
invariant(!txn->lockState()->saveLockStateAndUnlock(&lockSnapshot));
};
OldClientContext ctx(txn, ns);
Status status(ErrorCodes::InternalError, "");
while (true) {
try {
// We assume that in the WriteConflict retry case, either the op rolls back
// any changes it makes or is otherwise safe to rerun.
status =
repl::applyOperation_inlock(txn, ctx.db(), temp, false, alwaysUpsert);
break;
}
catch (const WriteConflictException& wce) {
LOG(2) << "WriteConflictException in applyOps command, retrying.";
txn->recoveryUnit()->commitAndRestart();
continue;
}
}
ab.append(status.isOK());
if (!status.isOK()) {
errors++;
}
num++;
WriteUnitOfWork wuow(txn);
logOpForDbHash(txn, ns.c_str());
wuow.commit();
}
result.append( "applied" , num );
result.append( "results" , ab.arr() );
if ( ! fromRepl ) {
// We want this applied atomically on slaves
// so we re-wrap without the pre-condition for speed
string tempNS = str::stream() << dbname << ".$cmd";
// TODO: possibly use mutable BSON to remove preCondition field
// once it is available
BSONObjIterator iter(cmdObj);
BSONObjBuilder cmdBuilder;
while (iter.more()) {
BSONElement elem(iter.next());
if (strcmp(elem.fieldName(), "preCondition") != 0) {
cmdBuilder.append(elem);
}
}
const BSONObj cmdRewritten = cmdBuilder.done();
// We currently always logOp the command regardless of whether the individial ops
// succeeded and rely on any failures to also happen on secondaries. This isn't
// perfect, but it's what the command has always done and is part of its "correct"
// behavior.
while (true) {
try {
WriteUnitOfWork wunit(txn);
getGlobalEnvironment()->getOpObserver()->onApplyOps(txn,
tempNS,
cmdRewritten);
wunit.commit();
break;
}
catch (const WriteConflictException& wce) {
LOG(2) <<
"WriteConflictException while logging applyOps command, retrying.";
txn->recoveryUnit()->commitAndRestart();
continue;
}
}
}
if (errors != 0) {
return false;
}
return true;
}
bool run(const string& dbname , BSONObj& jsobj, int, string& errmsg, BSONObjBuilder& result, bool /*fromRepl*/) {
static DBDirectClient db;
BSONElement e = jsobj.firstElement();
string toDeleteNs = dbname + '.' + e.valuestr();
MONGO_TLOG(0) << "CMD: reIndex " << toDeleteNs << endl;
Lock::DBWrite dbXLock(dbname);
Client::Context ctx(toDeleteNs);
Collection* collection = cc().database()->getCollection( toDeleteNs );
if ( !collection ) {
errmsg = "ns not found";
return false;
}
BackgroundOperation::assertNoBgOpInProgForNs( toDeleteNs );
std::vector<BSONObj> indexesInProg = stopIndexBuilds(cc().database(), jsobj);
list<BSONObj> all;
auto_ptr<DBClientCursor> i = db.query( dbname + ".system.indexes" , BSON( "ns" << toDeleteNs ) , 0 , 0 , 0 , QueryOption_SlaveOk );
BSONObjBuilder b;
while ( i->more() ) {
const BSONObj spec = i->next().removeField("v").getOwned();
const BSONObj key = spec.getObjectField("key");
const Status keyStatus = validateKeyPattern(key);
if (!keyStatus.isOK()) {
errmsg = str::stream()
<< "Cannot rebuild index " << spec << ": " << keyStatus.reason()
<< " For more info see http://dochub.mongodb.org/core/index-validation";
return false;
}
b.append( BSONObjBuilder::numStr( all.size() ) , spec );
all.push_back( spec );
}
result.appendNumber( "nIndexesWas", collection->getIndexCatalog()->numIndexesTotal() );
Status s = collection->getIndexCatalog()->dropAllIndexes( true );
if ( !s.isOK() ) {
errmsg = "dropIndexes failed";
return appendCommandStatus( result, s );
}
for ( list<BSONObj>::iterator i=all.begin(); i!=all.end(); i++ ) {
BSONObj o = *i;
LOG(1) << "reIndex ns: " << toDeleteNs << " index: " << o << endl;
Status s = collection->getIndexCatalog()->createIndex( o, false );
if ( !s.isOK() )
return appendCommandStatus( result, s );
}
result.append( "nIndexes" , (int)all.size() );
result.appendArray( "indexes" , b.obj() );
IndexBuilder::restoreIndexes(indexesInProg);
return true;
}
void Command::execCommandClientBasic(Command * c ,
ClientBasic& client,
int queryOptions,
const char *ns,
BSONObj& cmdObj,
BSONObjBuilder& result,
bool fromRepl ) {
verify(c);
std::string dbname = nsToDatabase(ns);
// Access control checks
if (!noauth) {
std::vector<Privilege> privileges;
c->addRequiredPrivileges(dbname, cmdObj, &privileges);
AuthorizationManager* authManager = client.getAuthorizationManager();
if (!authManager->checkAuthForPrivileges(privileges).isOK()) {
result.append("note", str::stream() << "not authorized for command: " <<
c->name << " on database " << dbname);
appendCommandStatus(result, false, "unauthorized");
return;
}
}
if (c->adminOnly() && c->localHostOnlyIfNoAuth(cmdObj) && noauth &&
!client.getIsLocalHostConnection()) {
log() << "command denied: " << cmdObj.toString() << endl;
appendCommandStatus(result,
false,
"unauthorized: this command must run from localhost when running db "
"without auth");
return;
}
if (c->adminOnly() && !startsWith(ns, "admin.")) {
log() << "command denied: " << cmdObj.toString() << endl;
appendCommandStatus(result, false, "access denied - use admin db");
return;
}
// End of access control checks
if (cmdObj.getBoolField("help")) {
stringstream help;
help << "help for: " << c->name << " ";
c->help( help );
result.append( "help" , help.str() );
result.append( "lockType" , c->locktype() );
appendCommandStatus(result, true, "");
return;
}
std::string errmsg;
bool ok;
try {
ok = c->run( dbname , cmdObj, queryOptions, errmsg, result, false );
}
catch (DBException& e) {
ok = false;
int code = e.getCode();
if (code == RecvStaleConfigCode) { // code for StaleConfigException
throw;
}
stringstream ss;
ss << "exception: " << e.what();
errmsg = ss.str();
result.append( "code" , code );
}
appendCommandStatus(result, ok, errmsg);
}
virtual bool run( const string& dbname, BSONObj& cmdObj, int options,
string& errmsg, BSONObjBuilder& result,
bool fromRepl = false ) {
// --- parse
NamespaceString ns( dbname, cmdObj[name].String() );
Status status = userAllowedWriteNS( ns );
if ( !status.isOK() )
return appendCommandStatus( result, status );
if ( cmdObj["indexes"].type() != Array ) {
errmsg = "indexes has to be an array";
result.append( "cmdObj", cmdObj );
return false;
}
std::vector<BSONObj> specs;
{
BSONObjIterator i( cmdObj["indexes"].Obj() );
while ( i.more() ) {
BSONElement e = i.next();
if ( e.type() != Object ) {
errmsg = "everything in indexes has to be an Object";
result.append( "cmdObj", cmdObj );
return false;
}
specs.push_back( e.Obj() );
}
}
if ( specs.size() == 0 ) {
errmsg = "no indexes to add";
return false;
}
// check specs
for ( size_t i = 0; i < specs.size(); i++ ) {
BSONObj spec = specs[i];
if ( spec["ns"].eoo() ) {
spec = _addNsToSpec( ns, spec );
specs[i] = spec;
}
if ( spec["ns"].type() != String ) {
errmsg = "spec has no ns";
result.append( "spec", spec );
return false;
}
if ( ns != spec["ns"].String() ) {
errmsg = "namespace mismatch";
result.append( "spec", spec );
return false;
}
}
{
// We first take a read lock to see if we need to do anything
// as many calls are ensureIndex (and hence no-ops), this is good so its a shared
// lock for common calls. We only take write lock if needed.
Client::ReadContext readContext( ns );
const Collection* collection = readContext.ctx().db()->getCollection( ns.ns() );
if ( collection ) {
for ( size_t i = 0; i < specs.size(); i++ ) {
BSONObj spec = specs[i];
StatusWith<BSONObj> statusWithSpec =
collection->getIndexCatalog()->prepareSpecForCreate( spec );
status = statusWithSpec.getStatus();
if ( status.code() == ErrorCodes::IndexAlreadyExists ) {
specs.erase( specs.begin() + i );
i--;
continue;
}
if ( !status.isOK() )
return appendCommandStatus( result, status );
}
if ( specs.size() == 0 ) {
result.append( "numIndexesBefore",
collection->getIndexCatalog()->numIndexesTotal() );
result.append( "note", "all indexes already exist" );
return true;
}
// need to create index
}
}
// now we know we have to create index(es)
Client::WriteContext writeContext( ns.ns() );
Database* db = writeContext.ctx().db();
Collection* collection = db->getCollection( ns.ns() );
result.appendBool( "createdCollectionAutomatically", collection == NULL );
if ( !collection ) {
collection = db->createCollection( ns.ns() );
invariant( collection );
}
result.append( "numIndexesBefore", collection->getIndexCatalog()->numIndexesTotal() );
for ( size_t i = 0; i < specs.size(); i++ ) {
BSONObj spec = specs[i];
status = collection->getIndexCatalog()->createIndex( spec, true );
if ( status.code() == ErrorCodes::IndexAlreadyExists ) {
if ( !result.hasField( "note" ) )
result.append( "note", "index already exists" );
continue;
}
if ( !status.isOK() ) {
appendCommandStatus( result, status );
return false;
}
}
result.append( "numIndexesAfter", collection->getIndexCatalog()->numIndexesTotal() );
if ( !fromRepl ) {
string cmdNs = ns.getCommandNS();
logOp( "c", cmdNs.c_str(), cmdObj );
}
return true;
}
bool run(OperationContext* txn, const string& dbname , BSONObj& jsobj, int, string& errmsg, BSONObjBuilder& result, bool /*fromRepl*/) {
DBDirectClient db(txn);
BSONElement e = jsobj.firstElement();
string toDeleteNs = dbname + '.' + e.valuestr();
LOG(0) << "CMD: reIndex " << toDeleteNs << endl;
Lock::DBLock dbXLock(txn->lockState(), dbname, MODE_X);
Client::Context ctx(txn, toDeleteNs);
Collection* collection = ctx.db()->getCollection( txn, toDeleteNs );
if ( !collection ) {
errmsg = "ns not found";
return false;
}
BackgroundOperation::assertNoBgOpInProgForNs( toDeleteNs );
std::vector<BSONObj> indexesInProg = stopIndexBuilds(txn, ctx.db(), jsobj);
vector<BSONObj> all;
{
vector<string> indexNames;
collection->getCatalogEntry()->getAllIndexes( txn, &indexNames );
for ( size_t i = 0; i < indexNames.size(); i++ ) {
const string& name = indexNames[i];
BSONObj spec = collection->getCatalogEntry()->getIndexSpec( txn, name );
all.push_back(spec.removeField("v").getOwned());
const BSONObj key = spec.getObjectField("key");
const Status keyStatus = validateKeyPattern(key);
if (!keyStatus.isOK()) {
errmsg = str::stream()
<< "Cannot rebuild index " << spec << ": " << keyStatus.reason()
<< " For more info see http://dochub.mongodb.org/core/index-validation";
return false;
}
}
}
result.appendNumber( "nIndexesWas", all.size() );
{
WriteUnitOfWork wunit(txn);
Status s = collection->getIndexCatalog()->dropAllIndexes(txn, true);
if ( !s.isOK() ) {
errmsg = "dropIndexes failed";
return appendCommandStatus( result, s );
}
wunit.commit();
}
MultiIndexBlock indexer(txn, collection);
// do not want interruption as that will leave us without indexes.
Status status = indexer.init(all);
if (!status.isOK())
return appendCommandStatus( result, status );
status = indexer.insertAllDocumentsInCollection();
if (!status.isOK())
return appendCommandStatus( result, status );
{
WriteUnitOfWork wunit(txn);
indexer.commit();
wunit.commit();
}
result.append( "nIndexes", (int)all.size() );
result.append( "indexes", all );
IndexBuilder::restoreIndexes(indexesInProg);
return true;
}
bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int,
string& errmsg,
BSONObjBuilder& result) {
BSONElement first = cmdObj.firstElement();
uassert(28528,
str::stream() << "Argument to listIndexes must be of type String, not "
<< typeName(first.type()),
first.type() == String);
StringData collectionName = first.valueStringData();
uassert(28529,
str::stream() << "Argument to listIndexes must be a collection name, "
<< "not the empty string",
!collectionName.empty());
const NamespaceString ns(dbname, collectionName);
const long long defaultBatchSize = std::numeric_limits<long long>::max();
long long batchSize;
Status parseCursorStatus = parseCommandCursorOptions(cmdObj, defaultBatchSize, &batchSize);
if (!parseCursorStatus.isOK()) {
return appendCommandStatus(result, parseCursorStatus);
}
AutoGetCollectionForRead autoColl(txn, ns);
if (!autoColl.getDb()) {
return appendCommandStatus(result,
Status(ErrorCodes::NamespaceNotFound, "no database"));
}
const Collection* collection = autoColl.getCollection();
if (!collection) {
return appendCommandStatus(result,
Status(ErrorCodes::NamespaceNotFound, "no collection"));
}
const CollectionCatalogEntry* cce = collection->getCatalogEntry();
invariant(cce);
vector<string> indexNames;
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
indexNames.clear();
cce->getAllIndexes(txn, &indexNames);
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "listIndexes", ns.ns());
auto ws = make_unique<WorkingSet>();
auto root = make_unique<QueuedDataStage>(txn, ws.get());
for (size_t i = 0; i < indexNames.size(); i++) {
BSONObj indexSpec;
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
indexSpec = cce->getIndexSpec(txn, indexNames[i]);
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "listIndexes", ns.ns());
WorkingSetID id = ws->allocate();
WorkingSetMember* member = ws->get(id);
member->keyData.clear();
member->loc = RecordId();
member->obj = Snapshotted<BSONObj>(SnapshotId(), indexSpec.getOwned());
member->transitionToOwnedObj();
root->pushBack(id);
}
std::string cursorNamespace = str::stream() << dbname << ".$cmd." << name << "."
<< ns.coll();
dassert(NamespaceString(cursorNamespace).isValid());
dassert(NamespaceString(cursorNamespace).isListIndexesCursorNS());
dassert(ns == NamespaceString(cursorNamespace).getTargetNSForListIndexes());
auto statusWithPlanExecutor = PlanExecutor::make(
txn, std::move(ws), std::move(root), cursorNamespace, PlanExecutor::YIELD_MANUAL);
if (!statusWithPlanExecutor.isOK()) {
return appendCommandStatus(result, statusWithPlanExecutor.getStatus());
}
unique_ptr<PlanExecutor> exec = std::move(statusWithPlanExecutor.getValue());
BSONArrayBuilder firstBatch;
const int byteLimit = FindCommon::kMaxBytesToReturnToClientAtOnce;
for (long long objCount = 0; objCount < batchSize && firstBatch.len() < byteLimit;
objCount++) {
BSONObj next;
PlanExecutor::ExecState state = exec->getNext(&next, NULL);
if (state == PlanExecutor::IS_EOF) {
break;
}
invariant(state == PlanExecutor::ADVANCED);
firstBatch.append(next);
}
CursorId cursorId = 0LL;
if (!exec->isEOF()) {
exec->saveState();
exec->detachFromOperationContext();
ClientCursor* cursor =
new ClientCursor(CursorManager::getGlobalCursorManager(),
exec.release(),
cursorNamespace,
txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot());
cursorId = cursor->cursorid();
}
appendCursorResponseObject(cursorId, cursorNamespace, firstBatch.arr(), &result);
return true;
}
virtual bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int,
string& errmsg,
BSONObjBuilder& result,
bool fromRepl) {
Lock::GlobalWrite globalWriteLock(txn->lockState());
string source = cmdObj.getStringField( name.c_str() );
string target = cmdObj.getStringField( "to" );
// We stay in source context the whole time. This is mostly to set the CurOp namespace.
Client::Context ctx(txn, source);
if ( !NamespaceString::validCollectionComponent(target.c_str()) ) {
errmsg = "invalid collection name: " + target;
return false;
}
if ( source.empty() || target.empty() ) {
errmsg = "invalid command syntax";
return false;
}
if (!fromRepl) { // If it got through on the master, need to allow it here too
Status sourceStatus = userAllowedWriteNS(source);
if (!sourceStatus.isOK()) {
errmsg = "error with source namespace: " + sourceStatus.reason();
return false;
}
Status targetStatus = userAllowedWriteNS(target);
if (!targetStatus.isOK()) {
errmsg = "error with target namespace: " + targetStatus.reason();
return false;
}
}
if (NamespaceString(source).coll() == "system.indexes"
|| NamespaceString(target).coll() == "system.indexes") {
errmsg = "renaming system.indexes is not allowed";
return false;
}
Database* const sourceDB = dbHolder().get(txn, nsToDatabase(source));
Collection* const sourceColl = sourceDB ? sourceDB->getCollection(txn, source)
: NULL;
if (!sourceColl) {
errmsg = "source namespace does not exist";
return false;
}
{
// Ensure that collection name does not exceed maximum length.
// Ensure that index names do not push the length over the max.
// Iterator includes unfinished indexes.
IndexCatalog::IndexIterator sourceIndIt =
sourceColl->getIndexCatalog()->getIndexIterator( txn, true );
int longestIndexNameLength = 0;
while ( sourceIndIt.more() ) {
int thisLength = sourceIndIt.next()->indexName().length();
if ( thisLength > longestIndexNameLength )
longestIndexNameLength = thisLength;
}
unsigned int longestAllowed =
min(int(NamespaceString::MaxNsCollectionLen),
int(NamespaceString::MaxNsLen) - 2/*strlen(".$")*/ - longestIndexNameLength);
if (target.size() > longestAllowed) {
StringBuilder sb;
sb << "collection name length of " << target.size()
<< " exceeds maximum length of " << longestAllowed
<< ", allowing for index names";
errmsg = sb.str();
return false;
}
}
const std::vector<BSONObj> indexesInProg = stopIndexBuilds(txn, sourceDB, cmdObj);
// Dismissed on success
ScopeGuard indexBuildRestorer = MakeGuard(IndexBuilder::restoreIndexes, indexesInProg);
Database* const targetDB = dbHolder().openDb(txn, nsToDatabase(target));
{
WriteUnitOfWork wunit(txn);
// Check if the target namespace exists and if dropTarget is true.
// If target exists and dropTarget is not true, return false.
if (targetDB->getCollection(txn, target)) {
if (!cmdObj["dropTarget"].trueValue()) {
errmsg = "target namespace exists";
return false;
}
Status s = targetDB->dropCollection(txn, target);
if ( !s.isOK() ) {
errmsg = s.toString();
return false;
}
}
// If we are renaming in the same database, just
// rename the namespace and we're done.
if (sourceDB == targetDB) {
Status s = targetDB->renameCollection(txn,
source,
target,
cmdObj["stayTemp"].trueValue() );
if (!s.isOK()) {
return appendCommandStatus(result, s);
}
if (!fromRepl) {
repl::logOp(txn, "c", (dbname + ".$cmd").c_str(), cmdObj);
}
wunit.commit();
indexBuildRestorer.Dismiss();
return true;
}
wunit.commit();
}
// If we get here, we are renaming across databases, so we must copy all the data and
// indexes, then remove the source collection.
// Create the target collection. It will be removed if we fail to copy the collection.
// TODO use a temp collection and unset the temp flag on success.
Collection* targetColl = NULL;
{
CollectionOptions options;
options.setNoIdIndex();
if (sourceColl->isCapped()) {
const CollectionOptions sourceOpts =
sourceColl->getCatalogEntry()->getCollectionOptions(txn);
options.capped = true;
options.cappedSize = sourceOpts.cappedSize;
options.cappedMaxDocs = sourceOpts.cappedMaxDocs;
}
WriteUnitOfWork wunit(txn);
// No logOp necessary because the entire renameCollection command is one logOp.
targetColl = targetDB->createCollection(txn, target, options);
if (!targetColl) {
errmsg = "Failed to create target collection.";
return false;
}
wunit.commit();
}
// Dismissed on success
ScopeGuard targetCollectionDropper = MakeGuard(dropCollection, txn, targetDB, target);
MultiIndexBlock indexer(txn, targetColl);
indexer.allowInterruption();
// Copy the index descriptions from the source collection, adjusting the ns field.
{
std::vector<BSONObj> indexesToCopy;
IndexCatalog::IndexIterator sourceIndIt =
sourceColl->getIndexCatalog()->getIndexIterator( txn, true );
while (sourceIndIt.more()) {
const BSONObj currIndex = sourceIndIt.next()->infoObj();
// Process the source index.
BSONObjBuilder newIndex;
newIndex.append("ns", target);
newIndex.appendElementsUnique(currIndex);
indexesToCopy.push_back(newIndex.obj());
}
indexer.init(indexesToCopy);
}
{
// Copy over all the data from source collection to target collection.
boost::scoped_ptr<RecordIterator> sourceIt(sourceColl->getIterator(txn));
while (!sourceIt->isEOF()) {
txn->checkForInterrupt(false);
const BSONObj obj = sourceColl->docFor(txn, sourceIt->getNext());
WriteUnitOfWork wunit(txn);
// No logOp necessary because the entire renameCollection command is one logOp.
Status status = targetColl->insertDocument(txn, obj, &indexer, true).getStatus();
if (!status.isOK())
return appendCommandStatus(result, status);
wunit.commit();
}
}
Status status = indexer.doneInserting();
if (!status.isOK())
return appendCommandStatus(result, status);
{
// Getting here means we successfully built the target copy. We now remove the
// source collection and finalize the rename.
WriteUnitOfWork wunit(txn);
Status status = sourceDB->dropCollection(txn, source);
if (!status.isOK())
return appendCommandStatus(result, status);
indexer.commit();
if (!fromRepl) {
repl::logOp(txn, "c", (dbname + ".$cmd").c_str(), cmdObj);
}
wunit.commit();
}
indexBuildRestorer.Dismiss();
targetCollectionDropper.Dismiss();
return true;
}
virtual bool run(OperationContext* txn, const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool fromRepl) {
Status status = getGlobalReplicationCoordinator()->processReplSetGetRBID(&result);
return appendCommandStatus(result, status);
}
virtual bool run(OperationContext* txn, const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool fromRepl) {
int secs = (int) cmdObj.firstElement().numberInt();
return appendCommandStatus(
result,
getGlobalReplicationCoordinator()->processReplSetFreeze(secs, &result));
}
bool WriteCmd::run(const string& dbName,
BSONObj& cmdObj,
int options,
string& errMsg,
BSONObjBuilder& result,
bool fromRepl) {
// Can't be run on secondaries (logTheOp() == false, slaveOk() == false).
dassert( !fromRepl );
BatchedCommandRequest request( _writeType );
BatchedCommandResponse response;
if ( !request.parseBSON( cmdObj, &errMsg ) || !request.isValid( &errMsg ) ) {
// Batch parse failure
response.setOk( false );
response.setN( 0 );
response.setErrCode( ErrorCodes::FailedToParse );
response.setErrMessage( errMsg );
dassert( response.isValid( &errMsg ) );
result.appendElements( response.toBSON() );
// TODO
// There's a pending issue about how to report response here. If we use
// the command infra-structure, we should reuse the 'errmsg' field. But
// we have already filed that message inside the BatchCommandResponse.
// return response.getOk();
return true;
}
// Note that this is a runCommmand, and therefore, the database and the collection name
// are in different parts of the grammar for the command. But it's more convenient to
// work with a NamespaceString. We built it here and replace it in the parsed command.
// Internally, everything work with the namespace string as opposed to just the
// collection name.
NamespaceString nss(dbName, request.getNS());
request.setNS(nss.ns());
Status status = userAllowedWriteNS( nss );
if ( !status.isOK() )
return appendCommandStatus( result, status );
if ( cc().curop() )
cc().curop()->setNS( nss.ns() );
if ( request.getBatchType() == BatchedCommandRequest::BatchType_Insert ) {
// check all docs
BatchedInsertRequest* insertRequest = request.getInsertRequest();
vector<BSONObj>& docsToInsert = insertRequest->getDocuments();
for ( size_t i = 0; i < docsToInsert.size(); i++ ) {
StatusWith<BSONObj> fixed = fixDocumentForInsert( docsToInsert[i] );
if ( !fixed.isOK() ) {
// we don't return early since each doc can be handled independantly
continue;
}
if ( fixed.getValue().isEmpty() ) {
continue;
}
docsToInsert[i] = fixed.getValue();
}
}
BSONObj defaultWriteConcern;
// This is really bad - it's only safe because we leak the defaults by overriding them with
// new defaults and because we never reset to an empty default.
// TODO: fix this for sane behavior where we query repl set object
if ( getLastErrorDefault ) defaultWriteConcern = *getLastErrorDefault;
if ( defaultWriteConcern.isEmpty() ) {
BSONObjBuilder b;
b.append( "w", 1 );
defaultWriteConcern = b.obj();
}
WriteBatchExecutor writeBatchExecutor(defaultWriteConcern,
&cc(),
&globalOpCounters,
lastError.get());
writeBatchExecutor.executeBatch( request, &response );
result.appendElements( response.toBSON() );
// TODO
// There's a pending issue about how to report response here. If we use
// the command infra-structure, we should reuse the 'errmsg' field. But
// we have already filed that message inside the BatchCommandResponse.
// return response.getOk();
return true;
}
/**
* Generates the next batch of results for a ClientCursor.
*
* TODO: Do we need to support some equivalent of OP_REPLY responseFlags?
*
* TODO: Is it possible to support awaitData?
*/
bool run(OperationContext* txn,
const std::string& dbname,
BSONObj& cmdObj,
int options,
std::string& errmsg,
BSONObjBuilder& result) override {
// Counted as a getMore, not as a command.
globalOpCounters.gotGetMore();
if (txn->getClient()->isInDirectClient()) {
return appendCommandStatus(result,
Status(ErrorCodes::IllegalOperation,
"Cannot run getMore command from eval()"));
}
StatusWith<GetMoreRequest> parseStatus = GetMoreRequest::parseFromBSON(dbname, cmdObj);
if (!parseStatus.isOK()) {
return appendCommandStatus(result, parseStatus.getStatus());
}
const GetMoreRequest& request = parseStatus.getValue();
// Depending on the type of cursor being operated on, we hold locks for the whole
// getMore, or none of the getMore, or part of the getMore. The three cases in detail:
//
// 1) Normal cursor: we lock with "ctx" and hold it for the whole getMore.
// 2) Cursor owned by global cursor manager: we don't lock anything. These cursors
// don't own any collection state.
// 3) Agg cursor: we lock with "ctx", then release, then relock with "unpinDBLock" and
// "unpinCollLock". This is because agg cursors handle locking internally (hence the
// release), but the pin and unpin of the cursor must occur under the collection
// lock. We don't use our AutoGetCollectionForRead "ctx" to relock, because
// AutoGetCollectionForRead checks the sharding version (and we want the relock for
// the unpin to succeed even if the sharding version has changed).
//
// Note that we declare our locks before our ClientCursorPin, in order to ensure that
// the pin's destructor is called before the lock destructors (so that the unpin occurs
// under the lock).
std::unique_ptr<AutoGetCollectionForRead> ctx;
std::unique_ptr<Lock::DBLock> unpinDBLock;
std::unique_ptr<Lock::CollectionLock> unpinCollLock;
CursorManager* cursorManager;
CursorManager* globalCursorManager = CursorManager::getGlobalCursorManager();
if (globalCursorManager->ownsCursorId(request.cursorid)) {
cursorManager = globalCursorManager;
}
else {
ctx.reset(new AutoGetCollectionForRead(txn, request.nss));
Collection* collection = ctx->getCollection();
if (!collection) {
return appendCommandStatus(result,
Status(ErrorCodes::OperationFailed,
"collection dropped between getMore calls"));
}
cursorManager = collection->getCursorManager();
}
ClientCursorPin ccPin(cursorManager, request.cursorid);
ClientCursor* cursor = ccPin.c();
if (!cursor) {
// We didn't find the cursor.
return appendCommandStatus(result, Status(ErrorCodes::CursorNotFound, str::stream()
<< "Cursor not found, cursor id: " << request.cursorid));
}
if (request.nss.ns() != cursor->ns()) {
return appendCommandStatus(result, Status(ErrorCodes::Unauthorized, str::stream()
<< "Requested getMore on namespace '" << request.nss.ns()
<< "', but cursor belongs to a different namespace"));
}
// On early return, get rid of the the cursor.
ScopeGuard cursorFreer = MakeGuard(&ClientCursorPin::deleteUnderlying, ccPin);
if (!cursor->hasRecoveryUnit()) {
// Start using a new RecoveryUnit.
cursor->setOwnedRecoveryUnit(
getGlobalServiceContext()->getGlobalStorageEngine()->newRecoveryUnit());
}
// Swap RecoveryUnit(s) between the ClientCursor and OperationContext.
ScopedRecoveryUnitSwapper ruSwapper(cursor, txn);
// Reset timeout timer on the cursor since the cursor is still in use.
cursor->setIdleTime(0);
// If the operation that spawned this cursor had a time limit set, apply leftover
// time to this getmore.
txn->getCurOp()->setMaxTimeMicros(cursor->getLeftoverMaxTimeMicros());
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
if (cursor->isAggCursor()) {
// Agg cursors handle their own locking internally.
ctx.reset(); // unlocks
}
PlanExecutor* exec = cursor->getExecutor();
exec->restoreState(txn);
// TODO: Handle result sets larger than 16MB.
BSONArrayBuilder nextBatch;
BSONObj obj;
PlanExecutor::ExecState state;
int numResults = 0;
while (PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// Add result to output buffer.
nextBatch.append(obj);
numResults++;
if (enoughForGetMore(request.batchSize, numResults, nextBatch.len())) {
break;
}
}
// If we are operating on an aggregation cursor, then we dropped our collection lock
// earlier and need to reacquire it in order to clean up our ClientCursorPin.
//
// TODO: We need to ensure that this relock happens if we release the pin above in
// response to PlanExecutor::getNext() throwing an exception.
if (cursor->isAggCursor()) {
invariant(NULL == ctx.get());
unpinDBLock.reset(new Lock::DBLock(txn->lockState(), request.nss.db(), MODE_IS));
unpinCollLock.reset(
new Lock::CollectionLock(txn->lockState(), request.nss.ns(), MODE_IS));
}
// Fail the command if the PlanExecutor reports execution failure.
if (PlanExecutor::FAILURE == state) {
const std::unique_ptr<PlanStageStats> stats(exec->getStats());
error() << "GetMore executor error, stats: " << Explain::statsToBSON(*stats);
return appendCommandStatus(result,
Status(ErrorCodes::OperationFailed,
str::stream() << "GetMore executor error: "
<< WorkingSetCommon::toStatusString(obj)));
}
CursorId respondWithId = 0;
if (shouldSaveCursorGetMore(state, exec, isCursorTailable(cursor))) {
respondWithId = request.cursorid;
exec->saveState();
cursor->setLeftoverMaxTimeMicros(txn->getCurOp()->getRemainingMaxTimeMicros());
cursor->incPos(numResults);
if (isCursorTailable(cursor) && state == PlanExecutor::IS_EOF) {
// Rather than swapping their existing RU into the client cursor, tailable
// cursors should get a new recovery unit.
ruSwapper.dismiss();
}
}
else {
txn->getCurOp()->debug().cursorExhausted = true;
}
appendGetMoreResponseObject(respondWithId, request.nss.ns(), nextBatch.arr(), &result);
if (respondWithId) {
cursorFreer.Dismiss();
}
return true;
}
/**
* Runs a query using the following steps:
* --Parsing.
* --Acquire locks.
* --Plan query, obtaining an executor that can run it.
* --Generate the first batch.
* --Save state for getMore, transferring ownership of the executor to a ClientCursor.
* --Generate response to send to the client.
*/
bool run(OperationContext* txn,
const std::string& dbname,
BSONObj& cmdObj,
int options,
std::string& errmsg,
BSONObjBuilder& result) override {
const NamespaceString nss(parseNs(dbname, cmdObj));
if (!nss.isValid() || nss.isCommand() || nss.isSpecialCommand()) {
return appendCommandStatus(result,
{ErrorCodes::InvalidNamespace,
str::stream() << "Invalid collection name: " << nss.ns()});
}
// Although it is a command, a find command gets counted as a query.
globalOpCounters.gotQuery();
if (txn->getClient()->isInDirectClient()) {
return appendCommandStatus(
result,
Status(ErrorCodes::IllegalOperation, "Cannot run find command from eval()"));
}
// Parse the command BSON to a QueryRequest.
const bool isExplain = false;
auto qrStatus = QueryRequest::makeFromFindCommand(nss, cmdObj, isExplain);
if (!qrStatus.isOK()) {
return appendCommandStatus(result, qrStatus.getStatus());
}
auto& qr = qrStatus.getValue();
// Validate term before acquiring locks, if provided.
if (auto term = qr->getReplicationTerm()) {
auto replCoord = repl::ReplicationCoordinator::get(txn);
Status status = replCoord->updateTerm(txn, *term);
// Note: updateTerm returns ok if term stayed the same.
if (!status.isOK()) {
return appendCommandStatus(result, status);
}
}
// Fill out curop information.
//
// We pass negative values for 'ntoreturn' and 'ntoskip' to indicate that these values
// should be omitted from the log line. Limit and skip information is already present in the
// find command parameters, so these fields are redundant.
const int ntoreturn = -1;
const int ntoskip = -1;
beginQueryOp(txn, nss, cmdObj, ntoreturn, ntoskip);
// Finish the parsing step by using the QueryRequest to create a CanonicalQuery.
ExtensionsCallbackReal extensionsCallback(txn, &nss);
auto statusWithCQ = CanonicalQuery::canonicalize(txn, std::move(qr), extensionsCallback);
if (!statusWithCQ.isOK()) {
return appendCommandStatus(result, statusWithCQ.getStatus());
}
std::unique_ptr<CanonicalQuery> cq = std::move(statusWithCQ.getValue());
// Acquire locks.
AutoGetCollectionForRead ctx(txn, nss);
Collection* collection = ctx.getCollection();
// Get the execution plan for the query.
auto statusWithPlanExecutor =
getExecutorFind(txn, collection, nss, std::move(cq), PlanExecutor::YIELD_AUTO);
if (!statusWithPlanExecutor.isOK()) {
return appendCommandStatus(result, statusWithPlanExecutor.getStatus());
}
std::unique_ptr<PlanExecutor> exec = std::move(statusWithPlanExecutor.getValue());
{
stdx::lock_guard<Client>(*txn->getClient());
CurOp::get(txn)->setPlanSummary_inlock(Explain::getPlanSummary(exec.get()));
}
if (!collection) {
// No collection. Just fill out curop indicating that there were zero results and
// there is no ClientCursor id, and then return.
const long long numResults = 0;
const CursorId cursorId = 0;
endQueryOp(txn, collection, *exec, numResults, cursorId);
appendCursorResponseObject(cursorId, nss.ns(), BSONArray(), &result);
return true;
}
const QueryRequest& originalQR = exec->getCanonicalQuery()->getQueryRequest();
// Stream query results, adding them to a BSONArray as we go.
CursorResponseBuilder firstBatch(/*isInitialResponse*/ true, &result);
BSONObj obj;
PlanExecutor::ExecState state = PlanExecutor::ADVANCED;
long long numResults = 0;
while (!FindCommon::enoughForFirstBatch(originalQR, numResults) &&
PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// If we can't fit this result inside the current batch, then we stash it for later.
if (!FindCommon::haveSpaceForNext(obj, numResults, firstBatch.bytesUsed())) {
exec->enqueue(obj);
break;
}
// Add result to output buffer.
firstBatch.append(obj);
numResults++;
}
// Throw an assertion if query execution fails for any reason.
if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
firstBatch.abandon();
error() << "Plan executor error during find command: " << PlanExecutor::statestr(state)
<< ", stats: " << Explain::getWinningPlanStats(exec.get());
return appendCommandStatus(result,
Status(ErrorCodes::OperationFailed,
str::stream()
<< "Executor error during find command: "
<< WorkingSetCommon::toStatusString(obj)));
}
// Before saving the cursor, ensure that whatever plan we established happened with the
// expected collection version
auto css = CollectionShardingState::get(txn, nss);
css->checkShardVersionOrThrow(txn);
// Set up the cursor for getMore.
CursorId cursorId = 0;
if (shouldSaveCursor(txn, collection, state, exec.get())) {
// Register the execution plan inside a ClientCursor. Ownership of the PlanExecutor is
// transferred to the ClientCursor.
//
// First unregister the PlanExecutor so it can be re-registered with ClientCursor.
exec->deregisterExec();
// Create a ClientCursor containing this plan executor. We don't have to worry about
// leaking it as it's inserted into a global map by its ctor.
ClientCursor* cursor =
new ClientCursor(collection->getCursorManager(),
exec.release(),
nss.ns(),
txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
originalQR.getOptions(),
cmdObj.getOwned());
cursorId = cursor->cursorid();
invariant(!exec);
PlanExecutor* cursorExec = cursor->getExecutor();
// State will be restored on getMore.
cursorExec->saveState();
cursorExec->detachFromOperationContext();
cursor->setLeftoverMaxTimeMicros(txn->getRemainingMaxTimeMicros());
cursor->setPos(numResults);
// Fill out curop based on the results.
endQueryOp(txn, collection, *cursorExec, numResults, cursorId);
} else {
endQueryOp(txn, collection, *exec, numResults, cursorId);
}
// Generate the response object to send to the client.
firstBatch.done(cursorId, nss.ns());
return true;
}
bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int,
string& errmsg,
BSONObjBuilder& result) {
BSONElement first = cmdObj.firstElement();
uassert(
28528,
str::stream() << "Argument to listIndexes must be of type String, not "
<< typeName(first.type()),
first.type() == String);
const NamespaceString ns(parseNs(dbname, cmdObj));
uassert(
28529,
str::stream() << "Argument to listIndexes must be a collection name, "
<< "not the empty string",
!ns.coll().empty());
const long long defaultBatchSize = std::numeric_limits<long long>::max();
long long batchSize;
Status parseCursorStatus = parseCommandCursorOptions(cmdObj,
defaultBatchSize,
&batchSize);
if (!parseCursorStatus.isOK()) {
return appendCommandStatus(result, parseCursorStatus);
}
AutoGetCollectionForRead autoColl(txn, ns);
if (!autoColl.getDb()) {
return appendCommandStatus( result, Status( ErrorCodes::NamespaceNotFound,
"no database" ) );
}
const Collection* collection = autoColl.getCollection();
if (!collection) {
return appendCommandStatus( result, Status( ErrorCodes::NamespaceNotFound,
"no collection" ) );
}
const CollectionCatalogEntry* cce = collection->getCatalogEntry();
invariant(cce);
vector<string> indexNames;
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
indexNames.clear();
cce->getAllIndexes( txn, &indexNames );
} MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "listIndexes", ns.ns());
std::auto_ptr<WorkingSet> ws(new WorkingSet());
std::auto_ptr<QueuedDataStage> root(new QueuedDataStage(ws.get()));
for ( size_t i = 0; i < indexNames.size(); i++ ) {
BSONObj indexSpec;
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
indexSpec = cce->getIndexSpec( txn, indexNames[i] );
} MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "listIndexes", ns.ns());
WorkingSetMember member;
member.state = WorkingSetMember::OWNED_OBJ;
member.keyData.clear();
member.loc = RecordId();
member.obj = Snapshotted<BSONObj>(SnapshotId(), indexSpec.getOwned());
root->pushBack(member);
}
std::string cursorNamespace = str::stream() << dbname << ".$cmd." << name << "."
<< ns.coll();
dassert(NamespaceString(cursorNamespace).isValid());
dassert(NamespaceString(cursorNamespace).isListIndexesGetMore());
dassert(ns == NamespaceString(cursorNamespace).getTargetNSForListIndexesGetMore());
PlanExecutor* rawExec;
Status makeStatus = PlanExecutor::make(txn,
ws.release(),
root.release(),
cursorNamespace,
PlanExecutor::YIELD_MANUAL,
&rawExec);
std::auto_ptr<PlanExecutor> exec(rawExec);
if (!makeStatus.isOK()) {
return appendCommandStatus( result, makeStatus );
}
BSONArrayBuilder firstBatch;
const int byteLimit = MaxBytesToReturnToClientAtOnce;
for (long long objCount = 0;
objCount < batchSize && firstBatch.len() < byteLimit;
objCount++) {
BSONObj next;
PlanExecutor::ExecState state = exec->getNext(&next, NULL);
if ( state == PlanExecutor::IS_EOF ) {
break;
}
invariant( state == PlanExecutor::ADVANCED );
firstBatch.append(next);
}
CursorId cursorId = 0LL;
if ( !exec->isEOF() ) {
exec->saveState();
ClientCursor* cursor = new ClientCursor(CursorManager::getGlobalCursorManager(),
exec.release(),
cursorNamespace);
cursorId = cursor->cursorid();
}
appendCursorResponseObject( cursorId, cursorNamespace, firstBatch.arr(), &result );
return true;
}
bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int,
string& errmsg,
BSONObjBuilder& result,
bool /*fromRepl*/) {
BSONElement first = cmdObj.firstElement();
uassert(
28528,
str::stream() << "Argument to listIndexes must be of type String, not "
<< typeName(first.type()),
first.type() == String);
const NamespaceString ns(parseNs(dbname, cmdObj));
uassert(
28529,
str::stream() << "Argument to listIndexes must be a collection name, "
<< "not the empty string",
!ns.coll().empty());
AutoGetCollectionForRead autoColl(txn, ns);
if (!autoColl.getDb()) {
return appendCommandStatus( result, Status( ErrorCodes::NamespaceNotFound,
"no database" ) );
}
const Collection* collection = autoColl.getCollection();
if (!collection) {
return appendCommandStatus( result, Status( ErrorCodes::NamespaceNotFound,
"no collection" ) );
}
const CollectionCatalogEntry* cce = collection->getCatalogEntry();
invariant(cce);
vector<string> indexNames;
cce->getAllIndexes( txn, &indexNames );
// TODO: Handle options specified in the command request object under the "cursor"
// field.
// TODO: If the full result set does not fit in one batch, allocate a cursor to store
// the remainder of the results.
BSONArrayBuilder arr;
for ( size_t i = 0; i < indexNames.size(); i++ ) {
arr.append( cce->getIndexSpec( txn, indexNames[i] ) );
}
if ( cmdObj["cursor"].type() == mongo::Object ) {
const long long cursorId = 0LL;
std::string cursorNamespace = str::stream()
<< dbname << ".$cmd." << name << "." << ns.coll();
Command::appendCursorResponseObject( cursorId, cursorNamespace, arr.arr(),
&result );
} else {
result.append( "indexes", arr.arr() );
}
return true;
}
/**
* Runs a query using the following steps:
* 1) Parsing.
* 2) Acquire locks.
* 3) Plan query, obtaining an executor that can run it.
* 4) Setup a cursor for the query, which may be used on subsequent getMores.
* 5) Generate the first batch.
* 6) Save state for getMore.
* 7) Generate response to send to the client.
*
* TODO: Rather than using the sharding version available in thread-local storage (i.e. the
* call to ShardingState::needCollectionMetadata() below), shard version information
* should be passed as part of the command parameter.
*/
bool run(OperationContext* txn,
const std::string& dbname,
BSONObj& cmdObj,
int options,
std::string& errmsg,
BSONObjBuilder& result) override {
const std::string fullns = parseNs(dbname, cmdObj);
const NamespaceString nss(fullns);
if (!nss.isValid()) {
return appendCommandStatus(result,
{ErrorCodes::InvalidNamespace,
str::stream() << "Invalid collection name: " << nss.ns()});
}
// Although it is a command, a find command gets counted as a query.
globalOpCounters.gotQuery();
if (txn->getClient()->isInDirectClient()) {
return appendCommandStatus(
result,
Status(ErrorCodes::IllegalOperation, "Cannot run find command from eval()"));
}
// 1a) Parse the command BSON to a LiteParsedQuery.
const bool isExplain = false;
auto lpqStatus = LiteParsedQuery::makeFromFindCommand(nss, cmdObj, isExplain);
if (!lpqStatus.isOK()) {
return appendCommandStatus(result, lpqStatus.getStatus());
}
auto& lpq = lpqStatus.getValue();
// Validate term, if provided.
if (auto term = lpq->getReplicationTerm()) {
auto replCoord = repl::ReplicationCoordinator::get(txn);
Status status = replCoord->updateTerm(*term);
// Note: updateTerm returns ok if term stayed the same.
if (!status.isOK()) {
return appendCommandStatus(result, status);
}
}
// Fill out curop information.
long long ntoreturn = lpq->getBatchSize().value_or(0);
beginQueryOp(txn, nss, cmdObj, ntoreturn, lpq->getSkip());
// 1b) Finish the parsing step by using the LiteParsedQuery to create a CanonicalQuery.
WhereCallbackReal whereCallback(txn, nss.db());
auto statusWithCQ = CanonicalQuery::canonicalize(lpq.release(), whereCallback);
if (!statusWithCQ.isOK()) {
return appendCommandStatus(result, statusWithCQ.getStatus());
}
std::unique_ptr<CanonicalQuery> cq = std::move(statusWithCQ.getValue());
// 2) Acquire locks.
AutoGetCollectionForRead ctx(txn, nss);
Collection* collection = ctx.getCollection();
const int dbProfilingLevel =
ctx.getDb() ? ctx.getDb()->getProfilingLevel() : serverGlobalParams.defaultProfile;
ShardingState* const shardingState = ShardingState::get(txn);
// It is possible that the sharding version will change during yield while we are
// retrieving a plan executor. If this happens we will throw an error and mongos will
// retry.
const ChunkVersion shardingVersionAtStart = shardingState->getVersion(nss.ns());
// 3) Get the execution plan for the query.
auto statusWithPlanExecutor =
getExecutorFind(txn, collection, nss, std::move(cq), PlanExecutor::YIELD_AUTO);
if (!statusWithPlanExecutor.isOK()) {
return appendCommandStatus(result, statusWithPlanExecutor.getStatus());
}
std::unique_ptr<PlanExecutor> exec = std::move(statusWithPlanExecutor.getValue());
// TODO: Currently, chunk ranges are kept around until all ClientCursors created while
// the chunk belonged on this node are gone. Separating chunk lifetime management from
// ClientCursor should allow this check to go away.
if (!shardingState->getVersion(nss.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// Version changed while retrieving a PlanExecutor. Terminate the operation,
// signaling that mongos should retry.
throw SendStaleConfigException(nss.ns(),
"version changed during find command",
shardingVersionAtStart,
shardingState->getVersion(nss.ns()));
}
if (!collection) {
// No collection. Just fill out curop indicating that there were zero results and
// there is no ClientCursor id, and then return.
const long long numResults = 0;
const CursorId cursorId = 0;
endQueryOp(txn, *exec, dbProfilingLevel, numResults, cursorId);
appendCursorResponseObject(cursorId, nss.ns(), BSONArray(), &result);
return true;
}
const LiteParsedQuery& pq = exec->getCanonicalQuery()->getParsed();
// 4) If possible, register the execution plan inside a ClientCursor, and pin that
// cursor. In this case, ownership of the PlanExecutor is transferred to the
// ClientCursor, and 'exec' becomes null.
//
// First unregister the PlanExecutor so it can be re-registered with ClientCursor.
exec->deregisterExec();
// Create a ClientCursor containing this plan executor. We don't have to worry
// about leaking it as it's inserted into a global map by its ctor.
ClientCursor* cursor =
new ClientCursor(collection->getCursorManager(),
exec.release(),
nss.ns(),
txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
pq.getOptions(),
pq.getFilter());
CursorId cursorId = cursor->cursorid();
ClientCursorPin ccPin(collection->getCursorManager(), cursorId);
// On early return, get rid of the the cursor.
ScopeGuard cursorFreer = MakeGuard(&ClientCursorPin::deleteUnderlying, ccPin);
invariant(!exec);
PlanExecutor* cursorExec = cursor->getExecutor();
// 5) Stream query results, adding them to a BSONArray as we go.
BSONArrayBuilder firstBatch;
BSONObj obj;
PlanExecutor::ExecState state;
long long numResults = 0;
while (!enoughForFirstBatch(pq, numResults, firstBatch.len()) &&
PlanExecutor::ADVANCED == (state = cursorExec->getNext(&obj, NULL))) {
// If adding this object will cause us to exceed the BSON size limit, then we stash
// it for later.
if (firstBatch.len() + obj.objsize() > BSONObjMaxUserSize && numResults > 0) {
cursorExec->enqueue(obj);
break;
}
// Add result to output buffer.
firstBatch.append(obj);
numResults++;
}
// Throw an assertion if query execution fails for any reason.
if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
const std::unique_ptr<PlanStageStats> stats(cursorExec->getStats());
error() << "Plan executor error during find command: " << PlanExecutor::statestr(state)
<< ", stats: " << Explain::statsToBSON(*stats);
return appendCommandStatus(result,
Status(ErrorCodes::OperationFailed,
str::stream()
<< "Executor error during find command: "
<< WorkingSetCommon::toStatusString(obj)));
}
// 6) Set up the cursor for getMore.
if (shouldSaveCursor(txn, collection, state, cursorExec)) {
// State will be restored on getMore.
cursorExec->saveState();
cursorExec->detachFromOperationContext();
cursor->setLeftoverMaxTimeMicros(CurOp::get(txn)->getRemainingMaxTimeMicros());
cursor->setPos(numResults);
} else {
cursorId = 0;
}
// Fill out curop based on the results.
endQueryOp(txn, *cursorExec, dbProfilingLevel, numResults, cursorId);
// 7) Generate the response object to send to the client.
appendCursorResponseObject(cursorId, nss.ns(), firstBatch.arr(), &result);
if (cursorId) {
cursorFreer.Dismiss();
}
return true;
}
bool run(OperationContext* txn,
const std::string& dbname,
BSONObj& cmdObj,
int options,
std::string& errmsg,
BSONObjBuilder& result) override {
// Counted as a getMore, not as a command.
globalOpCounters.gotGetMore();
if (txn->getClient()->isInDirectClient()) {
return appendCommandStatus(result,
Status(ErrorCodes::IllegalOperation,
"Cannot run getMore command from eval()"));
}
StatusWith<GetMoreRequest> parseStatus = GetMoreRequest::parseFromBSON(dbname, cmdObj);
if (!parseStatus.isOK()) {
return appendCommandStatus(result, parseStatus.getStatus());
}
const GetMoreRequest& request = parseStatus.getValue();
// Depending on the type of cursor being operated on, we hold locks for the whole
// getMore, or none of the getMore, or part of the getMore. The three cases in detail:
//
// 1) Normal cursor: we lock with "ctx" and hold it for the whole getMore.
// 2) Cursor owned by global cursor manager: we don't lock anything. These cursors
// don't own any collection state.
// 3) Agg cursor: we lock with "ctx", then release, then relock with "unpinDBLock" and
// "unpinCollLock". This is because agg cursors handle locking internally (hence the
// release), but the pin and unpin of the cursor must occur under the collection
// lock. We don't use our AutoGetCollectionForRead "ctx" to relock, because
// AutoGetCollectionForRead checks the sharding version (and we want the relock for
// the unpin to succeed even if the sharding version has changed).
//
// Note that we declare our locks before our ClientCursorPin, in order to ensure that
// the pin's destructor is called before the lock destructors (so that the unpin occurs
// under the lock).
std::unique_ptr<AutoGetCollectionForRead> ctx;
std::unique_ptr<Lock::DBLock> unpinDBLock;
std::unique_ptr<Lock::CollectionLock> unpinCollLock;
CursorManager* cursorManager;
CursorManager* globalCursorManager = CursorManager::getGlobalCursorManager();
if (globalCursorManager->ownsCursorId(request.cursorid)) {
cursorManager = globalCursorManager;
}
else {
ctx.reset(new AutoGetCollectionForRead(txn, request.nss));
Collection* collection = ctx->getCollection();
if (!collection) {
return appendCommandStatus(result,
Status(ErrorCodes::OperationFailed,
"collection dropped between getMore calls"));
}
cursorManager = collection->getCursorManager();
}
ClientCursorPin ccPin(cursorManager, request.cursorid);
ClientCursor* cursor = ccPin.c();
if (!cursor) {
// We didn't find the cursor.
return appendCommandStatus(result, Status(ErrorCodes::CursorNotFound, str::stream()
<< "Cursor not found, cursor id: " << request.cursorid));
}
if (request.nss.ns() != cursor->ns()) {
return appendCommandStatus(result, Status(ErrorCodes::Unauthorized, str::stream()
<< "Requested getMore on namespace '" << request.nss.ns()
<< "', but cursor belongs to a different namespace"));
}
const bool hasOwnMaxTime = CurOp::get(txn)->isMaxTimeSet();
// Validation related to awaitData.
if (isCursorAwaitData(cursor)) {
invariant(isCursorTailable(cursor));
if (!hasOwnMaxTime) {
Status status(ErrorCodes::BadValue,
str::stream() << "Must set maxTimeMS on a getMore if the initial "
<< "query had 'awaitData' set: " << cmdObj);
return appendCommandStatus(result, status);
}
if (cursor->isAggCursor()) {
Status status(ErrorCodes::BadValue,
"awaitData cannot be set on an aggregation cursor");
return appendCommandStatus(result, status);
}
}
// On early return, get rid of the cursor.
ScopeGuard cursorFreer = MakeGuard(&GetMoreCmd::cleanupCursor, txn, &ccPin, request);
if (!cursor->hasRecoveryUnit()) {
// Start using a new RecoveryUnit.
cursor->setOwnedRecoveryUnit(
getGlobalServiceContext()->getGlobalStorageEngine()->newRecoveryUnit());
}
// Swap RecoveryUnit(s) between the ClientCursor and OperationContext.
ScopedRecoveryUnitSwapper ruSwapper(cursor, txn);
// Reset timeout timer on the cursor since the cursor is still in use.
cursor->setIdleTime(0);
// If there is no time limit set directly on this getMore command, but the operation
// that spawned this cursor had a time limit set, then we have to apply any leftover
// time to this getMore.
if (!hasOwnMaxTime) {
CurOp::get(txn)->setMaxTimeMicros(cursor->getLeftoverMaxTimeMicros());
}
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
if (cursor->isAggCursor()) {
// Agg cursors handle their own locking internally.
ctx.reset(); // unlocks
}
PlanExecutor* exec = cursor->getExecutor();
exec->restoreState(txn);
// If we're tailing a capped collection, retrieve a monotonically increasing insert
// counter.
uint64_t lastInsertCount = 0;
if (isCursorAwaitData(cursor)) {
invariant(ctx->getCollection()->isCapped());
lastInsertCount = ctx->getCollection()->getCappedInsertNotifier()->getCount();
}
CursorId respondWithId = 0;
BSONArrayBuilder nextBatch;
BSONObj obj;
PlanExecutor::ExecState state;
int numResults = 0;
Status batchStatus = generateBatch(cursor, request, &nextBatch, &state, &numResults);
if (!batchStatus.isOK()) {
return appendCommandStatus(result, batchStatus);
}
// If this is an await data cursor, and we hit EOF without generating any results, then
// we block waiting for new oplog data to arrive.
if (isCursorAwaitData(cursor) && state == PlanExecutor::IS_EOF && numResults == 0) {
// Retrieve the notifier which we will wait on until new data arrives. We make sure
// to do this in the lock because once we drop the lock it is possible for the
// collection to become invalid. The notifier itself will outlive the collection if
// the collection is dropped, as we keep a shared_ptr to it.
auto notifier = ctx->getCollection()->getCappedInsertNotifier();
// Save the PlanExecutor and drop our locks.
exec->saveState();
ctx.reset();
// Block waiting for data.
Microseconds timeout(CurOp::get(txn)->getRemainingMaxTimeMicros());
notifier->waitForInsert(lastInsertCount, timeout);
notifier.reset();
ctx.reset(new AutoGetCollectionForRead(txn, request.nss));
exec->restoreState(txn);
// We woke up because either the timed_wait expired, or there was more data. Either
// way, attempt to generate another batch of results.
batchStatus = generateBatch(cursor, request, &nextBatch, &state, &numResults);
if (!batchStatus.isOK()) {
return appendCommandStatus(result, batchStatus);
}
}
if (shouldSaveCursorGetMore(state, exec, isCursorTailable(cursor))) {
respondWithId = request.cursorid;
exec->saveState();
// If maxTimeMS was set directly on the getMore rather than being rolled over
// from a previous find, then don't roll remaining micros over to the next
// getMore.
if (!hasOwnMaxTime) {
cursor->setLeftoverMaxTimeMicros(CurOp::get(txn)->getRemainingMaxTimeMicros());
}
cursor->incPos(numResults);
if (isCursorTailable(cursor) && state == PlanExecutor::IS_EOF) {
// Rather than swapping their existing RU into the client cursor, tailable
// cursors should get a new recovery unit.
ruSwapper.dismiss();
}
}
else {
CurOp::get(txn)->debug().cursorExhausted = true;
}
appendGetMoreResponseObject(respondWithId, request.nss.ns(), nextBatch.arr(), &result);
if (respondWithId) {
cursorFreer.Dismiss();
// If we are operating on an aggregation cursor, then we dropped our collection lock
// earlier and need to reacquire it in order to clean up our ClientCursorPin.
if (cursor->isAggCursor()) {
invariant(NULL == ctx.get());
unpinDBLock.reset(
new Lock::DBLock(txn->lockState(), request.nss.db(), MODE_IS));
unpinCollLock.reset(
new Lock::CollectionLock(txn->lockState(), request.nss.ns(), MODE_IS));
}
}
return true;
}
virtual bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int options,
string& errmsg,
BSONObjBuilder& result,
bool fromRepl) {
const std::string ns = parseNsCollectionRequired(dbname, cmdObj);
const BSONObj query = cmdObj.getObjectField("query");
const BSONObj fields = cmdObj.getObjectField("fields");
const BSONObj update = cmdObj.getObjectField("update");
const BSONObj sort = cmdObj.getObjectField("sort");
bool upsert = cmdObj["upsert"].trueValue();
bool returnNew = cmdObj["new"].trueValue();
bool remove = cmdObj["remove"].trueValue();
if ( remove ) {
if ( upsert ) {
errmsg = "remove and upsert can't co-exist";
return false;
}
if ( !update.isEmpty() ) {
errmsg = "remove and update can't co-exist";
return false;
}
if ( returnNew ) {
errmsg = "remove and returnNew can't co-exist";
return false;
}
}
else if ( !cmdObj.hasField("update") ) {
errmsg = "need remove or update";
return false;
}
bool ok = false;
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
errmsg = "";
// We can always retry because we only ever modify one document
ok = runImpl(txn,
dbname,
ns,
query,
fields,
update,
sort,
upsert,
returnNew,
remove,
result,
errmsg);
} MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "findAndModify", ns);
if ( !ok && errmsg == "no-collection" ) {
// Take X lock so we can create collection, then re-run operation.
ScopedTransaction transaction(txn, MODE_IX);
Lock::DBLock lk(txn->lockState(), dbname, MODE_X);
Client::Context ctx(txn, ns, false /* don't check version */);
if (!fromRepl &&
!repl::getGlobalReplicationCoordinator()->canAcceptWritesForDatabase(dbname)) {
return appendCommandStatus(result, Status(ErrorCodes::NotMaster, str::stream()
<< "Not primary while creating collection " << ns
<< " during findAndModify"));
}
Database* db = ctx.db();
if ( db->getCollection( ns ) ) {
// someone else beat us to it, that's ok
// we might race while we unlock if someone drops
// but that's ok, we'll just do nothing and error out
}
else {
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
WriteUnitOfWork wuow(txn);
uassertStatusOK( userCreateNS( txn, db,
ns, BSONObj(),
!fromRepl ) );
wuow.commit();
} MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "findAndModify", ns);
}
errmsg = "";
ok = runImpl(txn,
dbname,
ns,
query,
fields,
update,
sort,
upsert,
returnNew,
remove,
result,
errmsg);
}
virtual bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int options,
string& errmsg,
BSONObjBuilder& result) {
// --- parse
NamespaceString ns(dbname, cmdObj[name].String());
Status status = userAllowedWriteNS(ns);
if (!status.isOK())
return appendCommandStatus(result, status);
if (cmdObj["indexes"].type() != Array) {
errmsg = "indexes has to be an array";
result.append("cmdObj", cmdObj);
return false;
}
std::vector<BSONObj> specs;
{
BSONObjIterator i(cmdObj["indexes"].Obj());
while (i.more()) {
BSONElement e = i.next();
if (e.type() != Object) {
errmsg = "everything in indexes has to be an Object";
result.append("cmdObj", cmdObj);
return false;
}
specs.push_back(e.Obj());
}
}
if (specs.size() == 0) {
errmsg = "no indexes to add";
return false;
}
// check specs
for (size_t i = 0; i < specs.size(); i++) {
BSONObj spec = specs[i];
if (spec["ns"].eoo()) {
spec = _addNsToSpec(ns, spec);
specs[i] = spec;
}
if (spec["ns"].type() != String) {
errmsg = "ns field must be a string";
result.append("spec", spec);
return false;
}
std::string nsFromUser = spec["ns"].String();
if (nsFromUser.empty()) {
errmsg = "ns field cannot be an empty string";
result.append("spec", spec);
return false;
}
if (ns != nsFromUser) {
errmsg = str::stream() << "value of ns field '" << nsFromUser
<< "' doesn't match namespace " << ns.ns();
result.append("spec", spec);
return false;
}
}
// now we know we have to create index(es)
// Note: createIndexes command does not currently respect shard versioning.
ScopedTransaction transaction(txn, MODE_IX);
Lock::DBLock dbLock(txn->lockState(), ns.db(), MODE_X);
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns)) {
return appendCommandStatus(
result,
Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while creating indexes in " << ns.ns()));
}
Database* db = dbHolder().get(txn, ns.db());
if (!db) {
db = dbHolder().openDb(txn, ns.db());
}
Collection* collection = db->getCollection(ns.ns());
if (collection) {
result.appendBool("createdCollectionAutomatically", false);
} else {
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
WriteUnitOfWork wunit(txn);
collection = db->createCollection(txn, ns.ns(), CollectionOptions());
invariant(collection);
wunit.commit();
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "createIndexes", ns.ns());
result.appendBool("createdCollectionAutomatically", true);
}
const int numIndexesBefore = collection->getIndexCatalog()->numIndexesTotal(txn);
result.append("numIndexesBefore", numIndexesBefore);
auto client = txn->getClient();
ScopeGuard lastOpSetterGuard =
MakeObjGuard(repl::ReplClientInfo::forClient(client),
&repl::ReplClientInfo::setLastOpToSystemLastOpTime,
txn);
MultiIndexBlock indexer(txn, collection);
indexer.allowBackgroundBuilding();
indexer.allowInterruption();
const size_t origSpecsSize = specs.size();
indexer.removeExistingIndexes(&specs);
if (specs.size() == 0) {
result.append("numIndexesAfter", numIndexesBefore);
result.append("note", "all indexes already exist");
return true;
}
if (specs.size() != origSpecsSize) {
result.append("note", "index already exists");
}
for (size_t i = 0; i < specs.size(); i++) {
const BSONObj& spec = specs[i];
if (spec["unique"].trueValue()) {
status = checkUniqueIndexConstraints(txn, ns.ns(), spec["key"].Obj());
if (!status.isOK()) {
return appendCommandStatus(result, status);
}
}
if (spec["v"].isNumber() && spec["v"].numberInt() == 0) {
return appendCommandStatus(
result,
Status(ErrorCodes::CannotCreateIndex,
str::stream() << "illegal index specification: " << spec << ". "
<< "The option v:0 cannot be passed explicitly"));
}
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
uassertStatusOK(indexer.init(specs));
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "createIndexes", ns.ns());
// If we're a background index, replace exclusive db lock with an intent lock, so that
// other readers and writers can proceed during this phase.
if (indexer.getBuildInBackground()) {
txn->recoveryUnit()->abandonSnapshot();
dbLock.relockWithMode(MODE_IX);
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns)) {
return appendCommandStatus(
result,
Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while creating background indexes in "
<< ns.ns()));
}
}
try {
Lock::CollectionLock colLock(txn->lockState(), ns.ns(), MODE_IX);
uassertStatusOK(indexer.insertAllDocumentsInCollection());
} catch (const DBException& e) {
invariant(e.getCode() != ErrorCodes::WriteConflict);
// Must have exclusive DB lock before we clean up the index build via the
// destructor of 'indexer'.
if (indexer.getBuildInBackground()) {
try {
// This function cannot throw today, but we will preemptively prepare for
// that day, to avoid data corruption due to lack of index cleanup.
txn->recoveryUnit()->abandonSnapshot();
dbLock.relockWithMode(MODE_X);
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns)) {
return appendCommandStatus(
result,
Status(ErrorCodes::NotMaster,
str::stream()
<< "Not primary while creating background indexes in "
<< ns.ns() << ": cleaning up index build failure due to "
<< e.toString()));
}
} catch (...) {
std::terminate();
}
}
throw;
}
// Need to return db lock back to exclusive, to complete the index build.
if (indexer.getBuildInBackground()) {
txn->recoveryUnit()->abandonSnapshot();
dbLock.relockWithMode(MODE_X);
uassert(ErrorCodes::NotMaster,
str::stream() << "Not primary while completing index build in " << dbname,
repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns));
Database* db = dbHolder().get(txn, ns.db());
uassert(28551, "database dropped during index build", db);
uassert(28552, "collection dropped during index build", db->getCollection(ns.ns()));
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
WriteUnitOfWork wunit(txn);
indexer.commit();
for (size_t i = 0; i < specs.size(); i++) {
std::string systemIndexes = ns.getSystemIndexesCollection();
getGlobalServiceContext()->getOpObserver()->onCreateIndex(
txn, systemIndexes, specs[i]);
}
wunit.commit();
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "createIndexes", ns.ns());
result.append("numIndexesAfter", collection->getIndexCatalog()->numIndexesTotal(txn));
lastOpSetterGuard.Dismiss();
return true;
}
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;
}
static bool runImpl(OperationContext* txn,
const string& dbname,
const string& ns,
const BSONObj& query,
const BSONObj& fields,
const BSONObj& update,
const BSONObj& sort,
bool upsert,
bool returnNew,
bool remove ,
BSONObjBuilder& result,
string& errmsg) {
AutoGetOrCreateDb autoDb(txn, dbname, MODE_IX);
Lock::CollectionLock collLock(txn->lockState(), ns, MODE_IX);
Client::Context ctx(txn, ns, autoDb.getDb(), autoDb.justCreated());
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesForDatabase(dbname)) {
return appendCommandStatus(result, Status(ErrorCodes::NotMaster, str::stream()
<< "Not primary while running findAndModify in " << ns));
}
Collection* collection = ctx.db()->getCollection(ns);
const WhereCallbackReal whereCallback(txn, StringData(ns));
if ( !collection ) {
if ( !upsert ) {
// no collectio and no upsert, so can't possible do anything
_appendHelper( result, BSONObj(), false, fields, whereCallback );
return true;
}
// no collection, but upsert, so we want to create it
// problem is we only have IX on db and collection :(
// so we tell our caller who can do it
errmsg = "no-collection";
return false;
}
Snapshotted<BSONObj> snapshotDoc;
RecordId loc;
bool found = false;
{
CanonicalQuery* cq;
const BSONObj projection;
const long long skip = 0;
const long long limit = -1; // 1 document requested; negative indicates hard limit.
uassertStatusOK(CanonicalQuery::canonicalize(ns,
query,
sort,
projection,
skip,
limit,
&cq,
whereCallback));
PlanExecutor* rawExec;
uassertStatusOK(getExecutor(txn,
collection,
cq,
PlanExecutor::YIELD_AUTO,
&rawExec,
QueryPlannerParams::DEFAULT));
scoped_ptr<PlanExecutor> exec(rawExec);
PlanExecutor::ExecState state = exec->getNextSnapshotted(&snapshotDoc, &loc);
if (PlanExecutor::ADVANCED == state) {
found = true;
}
else if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
if (PlanExecutor::FAILURE == state &&
WorkingSetCommon::isValidStatusMemberObject(snapshotDoc.value())) {
const Status errorStatus =
WorkingSetCommon::getMemberObjectStatus(snapshotDoc.value());
invariant(!errorStatus.isOK());
uasserted(errorStatus.code(), errorStatus.reason());
}
uasserted(ErrorCodes::OperationFailed,
str::stream() << "executor returned " << PlanExecutor::statestr(state)
<< " while finding document to update");
}
else {
invariant(PlanExecutor::IS_EOF == state);
}
}
WriteUnitOfWork wuow(txn);
if (found) {
// We found a doc, but it might not be associated with the active snapshot.
// If the doc has changed or is no longer in the collection, we will throw a
// write conflict exception and start again from the beginning.
if (txn->recoveryUnit()->getSnapshotId() != snapshotDoc.snapshotId()) {
BSONObj oldObj = snapshotDoc.value();
if (!collection->findDoc(txn, loc, &snapshotDoc)) {
// Got deleted in the new snapshot.
throw WriteConflictException();
}
if (!oldObj.binaryEqual(snapshotDoc.value())) {
// Got updated in the new snapshot.
throw WriteConflictException();
}
}
// If we get here without throwing, then we should have the copy of the doc from
// the latest snapshot.
invariant(txn->recoveryUnit()->getSnapshotId() == snapshotDoc.snapshotId());
}
BSONObj doc = snapshotDoc.value();
BSONObj queryModified = query;
if (found && !doc["_id"].eoo() && !CanonicalQuery::isSimpleIdQuery(query)) {
// we're going to re-write the query to be more efficient
// we have to be a little careful because of positional operators
// maybe we can pass this all through eventually, but right now isn't an easy way
bool hasPositionalUpdate = false;
{
// if the update has a positional piece ($)
// then we need to pull all query parts in
// so here we check for $
// a little hacky
BSONObjIterator i( update );
while ( i.more() ) {
const BSONElement& elem = i.next();
if ( elem.fieldName()[0] != '$' || elem.type() != Object )
continue;
BSONObjIterator j( elem.Obj() );
while ( j.more() ) {
if ( str::contains( j.next().fieldName(), ".$" ) ) {
hasPositionalUpdate = true;
break;
}
}
}
}
BSONObjBuilder b(query.objsize() + 10);
b.append( doc["_id"] );
bool addedAtomic = false;
BSONObjIterator i(query);
while ( i.more() ) {
const BSONElement& elem = i.next();
if ( str::equals( "_id" , elem.fieldName() ) ) {
// we already do _id
continue;
}
if ( ! hasPositionalUpdate ) {
// if there is a dotted field, accept we may need more query parts
continue;
}
if ( ! addedAtomic ) {
b.appendBool( "$atomic" , true );
addedAtomic = true;
}
b.append( elem );
}
queryModified = b.obj();
}
if ( remove ) {
_appendHelper(result, doc, found, fields, whereCallback);
if ( found ) {
deleteObjects(txn, ctx.db(), ns, queryModified, PlanExecutor::YIELD_MANUAL,
true, true);
BSONObjBuilder le( result.subobjStart( "lastErrorObject" ) );
le.appendNumber( "n" , 1 );
le.done();
}
}
else {
// update
if ( ! found && ! upsert ) {
// didn't have it, and am not upserting
_appendHelper(result, doc, found, fields, whereCallback);
}
else {
// we found it or we're updating
if ( ! returnNew ) {
_appendHelper(result, doc, found, fields, whereCallback);
}
const NamespaceString requestNs(ns);
UpdateRequest request(requestNs);
request.setQuery(queryModified);
request.setUpdates(update);
request.setUpsert(upsert);
request.setUpdateOpLog();
request.setStoreResultDoc(returnNew);
request.setYieldPolicy(PlanExecutor::YIELD_MANUAL);
// TODO(greg) We need to send if we are ignoring
// the shard version below, but for now no
UpdateLifecycleImpl updateLifecycle(false, requestNs);
request.setLifecycle(&updateLifecycle);
UpdateResult res = mongo::update(txn,
ctx.db(),
request,
&txn->getCurOp()->debug());
if (!found && res.existing) {
// No match was found during the read part of this find and modify, which
// means that we're here doing an upsert. But the update also told us that
// we modified an *already existing* document. This probably means that
// the query reported EOF based on an out-of-date snapshot. This should be
// a rare event, so we handle it by throwing a write conflict.
throw WriteConflictException();
}
if ( !collection ) {
// collection created by an upsert
collection = ctx.db()->getCollection(ns);
}
LOG(3) << "update result: " << res ;
if (returnNew) {
dassert(!res.newObj.isEmpty());
_appendHelper(result, res.newObj, true, fields, whereCallback);
}
BSONObjBuilder le( result.subobjStart( "lastErrorObject" ) );
le.appendBool( "updatedExisting" , res.existing );
le.appendNumber( "n" , res.numMatched );
if ( !res.upserted.isEmpty() ) {
le.append( res.upserted[kUpsertedFieldName] );
}
le.done();
}
}
// Committing the WUOW can close the current snapshot. Until this happens, the
// snapshot id should not have changed.
if (found) {
invariant(txn->recoveryUnit()->getSnapshotId() == snapshotDoc.snapshotId());
}
wuow.commit();
return true;
}
virtual bool run( const string& dbname, BSONObj& cmdObj, int options,
string& errmsg, BSONObjBuilder& result,
bool fromRepl = false ) {
NamespaceString ns( dbname, cmdObj[name].String() );
Client::ReadContext ctx(ns.ns());
Database* db = ctx.ctx().db();
Collection* collection = db->getCollection( ns );
if ( !collection )
return appendCommandStatus( result,
Status( ErrorCodes::NamespaceNotFound,
str::stream() <<
"ns does not exist: " << ns.ns() ) );
size_t numCursors = static_cast<size_t>( cmdObj["numCursors"].numberInt() );
if ( numCursors == 0 || numCursors > 10000 )
return appendCommandStatus( result,
Status( ErrorCodes::BadValue,
str::stream() <<
"numCursors has to be between 1 and 10000" <<
" was: " << numCursors ) );
OwnedPointerVector<RecordIterator> iterators(collection->getManyIterators());
if (iterators.size() < numCursors) {
numCursors = iterators.size();
}
OwnedPointerVector<MultiIteratorRunner> runners;
for ( size_t i = 0; i < numCursors; i++ ) {
runners.push_back(new MultiIteratorRunner(ns.ns(), collection));
}
// transfer iterators to runners using a round-robin distribution.
// TODO consider using a common work queue once invalidation issues go away.
for (size_t i = 0; i < iterators.size(); i++) {
runners[i % runners.size()]->addIterator(iterators.releaseAt(i));
}
{
BSONArrayBuilder bucketsBuilder;
for (size_t i = 0; i < runners.size(); i++) {
// transfer ownership of a runner to the ClientCursor (which manages its own
// lifetime).
ClientCursor* cc = new ClientCursor( collection, runners.releaseAt(i) );
// we are mimicking the aggregation cursor output here
// that is why there are ns, ok and empty firstBatch
BSONObjBuilder threadResult;
{
BSONObjBuilder cursor;
cursor.appendArray( "firstBatch", BSONObj() );
cursor.append( "ns", ns );
cursor.append( "id", cc->cursorid() );
threadResult.append( "cursor", cursor.obj() );
}
threadResult.appendBool( "ok", 1 );
bucketsBuilder.append( threadResult.obj() );
}
result.appendArray( "cursors", bucketsBuilder.obj() );
}
return true;
}
/**
* Runs a query using the following steps:
* --Parsing.
* --Acquire locks.
* --Plan query, obtaining an executor that can run it.
* --Generate the first batch.
* --Save state for getMore, transferring ownership of the executor to a ClientCursor.
* --Generate response to send to the client.
*/
bool run(OperationContext* txn,
const std::string& dbname,
BSONObj& cmdObj,
int options,
std::string& errmsg,
BSONObjBuilder& result) override {
const std::string fullns = parseNs(dbname, cmdObj);
const NamespaceString nss(fullns);
if (!nss.isValid() || nss.isCommand() || nss.isSpecialCommand()) {
return appendCommandStatus(result,
{ErrorCodes::InvalidNamespace,
str::stream() << "Invalid collection name: " << nss.ns()});
}
// Although it is a command, a find command gets counted as a query.
globalOpCounters.gotQuery();
if (txn->getClient()->isInDirectClient()) {
return appendCommandStatus(
result,
Status(ErrorCodes::IllegalOperation, "Cannot run find command from eval()"));
}
// Parse the command BSON to a LiteParsedQuery.
const bool isExplain = false;
auto lpqStatus = LiteParsedQuery::makeFromFindCommand(nss, cmdObj, isExplain);
if (!lpqStatus.isOK()) {
return appendCommandStatus(result, lpqStatus.getStatus());
}
auto& lpq = lpqStatus.getValue();
// Validate term before acquiring locks, if provided.
if (auto term = lpq->getReplicationTerm()) {
auto replCoord = repl::ReplicationCoordinator::get(txn);
Status status = replCoord->updateTerm(txn, *term);
// Note: updateTerm returns ok if term stayed the same.
if (!status.isOK()) {
return appendCommandStatus(result, status);
}
}
// Fill out curop information.
//
// We pass negative values for 'ntoreturn' and 'ntoskip' to indicate that these values
// should be omitted from the log line. Limit and skip information is already present in the
// find command parameters, so these fields are redundant.
const int ntoreturn = -1;
const int ntoskip = -1;
beginQueryOp(txn, nss, cmdObj, ntoreturn, ntoskip);
// Finish the parsing step by using the LiteParsedQuery to create a CanonicalQuery.
ExtensionsCallbackReal extensionsCallback(txn, &nss);
auto statusWithCQ = CanonicalQuery::canonicalize(lpq.release(), extensionsCallback);
if (!statusWithCQ.isOK()) {
return appendCommandStatus(result, statusWithCQ.getStatus());
}
std::unique_ptr<CanonicalQuery> cq = std::move(statusWithCQ.getValue());
ShardingState* const shardingState = ShardingState::get(txn);
if (OperationShardVersion::get(txn).hasShardVersion() && shardingState->enabled()) {
ChunkVersion receivedVersion = OperationShardVersion::get(txn).getShardVersion(nss);
ChunkVersion latestVersion;
// Wait for migration completion to get the correct chunk version.
const int maxTimeoutSec = 30;
int timeoutSec = cq->getParsed().getMaxTimeMS() / 1000;
if (!timeoutSec || timeoutSec > maxTimeoutSec) {
timeoutSec = maxTimeoutSec;
}
if (!shardingState->waitTillNotInCriticalSection(timeoutSec)) {
uasserted(ErrorCodes::LockTimeout, "Timeout while waiting for migration commit");
}
// If the received version is newer than the version cached in 'shardingState', then we
// have to refresh 'shardingState' from the config servers. We do this before acquiring
// locks so that we don't hold locks while waiting on the network.
uassertStatusOK(shardingState->refreshMetadataIfNeeded(
txn, nss.ns(), receivedVersion, &latestVersion));
}
// Acquire locks.
AutoGetCollectionForRead ctx(txn, nss);
Collection* collection = ctx.getCollection();
const int dbProfilingLevel =
ctx.getDb() ? ctx.getDb()->getProfilingLevel() : serverGlobalParams.defaultProfile;
// It is possible that the sharding version will change during yield while we are
// retrieving a plan executor. If this happens we will throw an error and mongos will
// retry.
const ChunkVersion shardingVersionAtStart = shardingState->getVersion(nss.ns());
// Get the execution plan for the query.
auto statusWithPlanExecutor =
getExecutorFind(txn, collection, nss, std::move(cq), PlanExecutor::YIELD_AUTO);
if (!statusWithPlanExecutor.isOK()) {
return appendCommandStatus(result, statusWithPlanExecutor.getStatus());
}
std::unique_ptr<PlanExecutor> exec = std::move(statusWithPlanExecutor.getValue());
if (!collection) {
// No collection. Just fill out curop indicating that there were zero results and
// there is no ClientCursor id, and then return.
const long long numResults = 0;
const CursorId cursorId = 0;
endQueryOp(txn, collection, *exec, dbProfilingLevel, numResults, cursorId);
appendCursorResponseObject(cursorId, nss.ns(), BSONArray(), &result);
return true;
}
const LiteParsedQuery& pq = exec->getCanonicalQuery()->getParsed();
// Stream query results, adding them to a BSONArray as we go.
BSONArrayBuilder firstBatch;
BSONObj obj;
PlanExecutor::ExecState state = PlanExecutor::ADVANCED;
long long numResults = 0;
while (!FindCommon::enoughForFirstBatch(pq, numResults, firstBatch.len()) &&
PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// If adding this object will cause us to exceed the BSON size limit, then we stash
// it for later.
if (firstBatch.len() + obj.objsize() > BSONObjMaxUserSize && numResults > 0) {
exec->enqueue(obj);
break;
}
// Add result to output buffer.
firstBatch.append(obj);
numResults++;
}
// Throw an assertion if query execution fails for any reason.
if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
const std::unique_ptr<PlanStageStats> stats(exec->getStats());
error() << "Plan executor error during find command: " << PlanExecutor::statestr(state)
<< ", stats: " << Explain::statsToBSON(*stats);
return appendCommandStatus(result,
Status(ErrorCodes::OperationFailed,
str::stream()
<< "Executor error during find command: "
<< WorkingSetCommon::toStatusString(obj)));
}
// TODO: Currently, chunk ranges are kept around until all ClientCursors created while the
// chunk belonged on this node are gone. Separating chunk lifetime management from
// ClientCursor should allow this check to go away.
if (!shardingState->getVersion(nss.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// Version changed while retrieving a PlanExecutor. Terminate the operation,
// signaling that mongos should retry.
throw SendStaleConfigException(nss.ns(),
"version changed during find command",
shardingVersionAtStart,
shardingState->getVersion(nss.ns()));
}
// Set up the cursor for getMore.
CursorId cursorId = 0;
if (shouldSaveCursor(txn, collection, state, exec.get())) {
// Register the execution plan inside a ClientCursor. Ownership of the PlanExecutor is
// transferred to the ClientCursor.
//
// First unregister the PlanExecutor so it can be re-registered with ClientCursor.
exec->deregisterExec();
// Create a ClientCursor containing this plan executor. We don't have to worry about
// leaking it as it's inserted into a global map by its ctor.
ClientCursor* cursor =
new ClientCursor(collection->getCursorManager(),
exec.release(),
nss.ns(),
txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
pq.getOptions(),
pq.getFilter());
cursorId = cursor->cursorid();
invariant(!exec);
PlanExecutor* cursorExec = cursor->getExecutor();
// State will be restored on getMore.
cursorExec->saveState();
cursorExec->detachFromOperationContext();
cursor->setLeftoverMaxTimeMicros(CurOp::get(txn)->getRemainingMaxTimeMicros());
cursor->setPos(numResults);
// Fill out curop based on the results.
endQueryOp(txn, collection, *cursorExec, dbProfilingLevel, numResults, cursorId);
} else {
endQueryOp(txn, collection, *exec, dbProfilingLevel, numResults, cursorId);
}
// Generate the response object to send to the client.
appendCursorResponseObject(cursorId, nss.ns(), firstBatch.arr(), &result);
return 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;
}
virtual bool run(const string& db, BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool fromRepl) {
string coll = cmdObj.firstElement().valuestr();
if( coll.empty() || db.empty() ) {
errmsg = "no collection name specified";
return false;
}
if( isCurrentlyAReplSetPrimary() && !cmdObj["force"].trueValue() ) {
errmsg = "will not run compact on an active replica set primary as this is a slow blocking operation. use force:true to force";
return false;
}
NamespaceString ns(db,coll);
if ( !ns.isNormal() ) {
errmsg = "bad namespace name";
return false;
}
if ( ns.isSystem() ) {
// items in system.* cannot be moved as there might be pointers to them
// i.e. system.indexes entries are pointed to from NamespaceDetails
errmsg = "can't compact a system namespace";
return false;
}
CompactOptions compactOptions;
if ( cmdObj["preservePadding"].trueValue() ) {
compactOptions.paddingMode = CompactOptions::PRESERVE;
if ( cmdObj.hasElement( "paddingFactor" ) ||
cmdObj.hasElement( "paddingBytes" ) ) {
errmsg = "cannot mix preservePadding and paddingFactor|paddingBytes";
return false;
}
}
else if ( cmdObj.hasElement( "paddingFactor" ) || cmdObj.hasElement( "paddingBytes" ) ) {
compactOptions.paddingMode = CompactOptions::MANUAL;
if ( cmdObj.hasElement("paddingFactor") ) {
compactOptions.paddingFactor = cmdObj["paddingFactor"].Number();
if ( compactOptions.paddingFactor < 1 ||
compactOptions.paddingFactor > 4 ){
errmsg = "invalid padding factor";
return false;
}
}
if ( cmdObj.hasElement("paddingBytes") ) {
compactOptions.paddingBytes = cmdObj["paddingBytes"].numberInt();
if ( compactOptions.paddingBytes < 0 ||
compactOptions.paddingBytes > ( 1024 * 1024 ) ) {
errmsg = "invalid padding bytes";
return false;
}
}
}
if ( cmdObj.hasElement("validate") )
compactOptions.validateDocuments = cmdObj["validate"].trueValue();
Lock::DBWrite lk(ns.ns());
BackgroundOperation::assertNoBgOpInProgForNs(ns.ns());
Client::Context ctx(ns);
Collection* collection = ctx.db()->getCollection(ns.ns());
if( ! collection ) {
errmsg = "namespace does not exist";
return false;
}
if ( collection->isCapped() ) {
errmsg = "cannot compact a capped collection";
return false;
}
log() << "compact " << ns << " begin, options: " << compactOptions.toString();
std::vector<BSONObj> indexesInProg = stopIndexBuilds(db, cmdObj);
StatusWith<CompactStats> status = collection->compact( &compactOptions );
if ( !status.isOK() )
return appendCommandStatus( result, status.getStatus() );
if ( status.getValue().corruptDocuments > 0 )
result.append("invalidObjects", status.getValue().corruptDocuments );
log() << "compact " << ns << " end";
IndexBuilder::restoreIndexes(indexesInProg);
return true;
}
virtual bool run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj,
int options,
string& errmsg,
BSONObjBuilder& result) {
// --- parse
NamespaceString ns(dbname, cmdObj[name].String());
Status status = userAllowedWriteNS(ns);
if (!status.isOK())
return appendCommandStatus(result, status);
if (cmdObj["indexes"].type() != Array) {
errmsg = "indexes has to be an array";
result.append("cmdObj", cmdObj);
return false;
}
std::vector<BSONObj> specs;
{
BSONObjIterator i(cmdObj["indexes"].Obj());
while (i.more()) {
BSONElement e = i.next();
if (e.type() != Object) {
errmsg = "everything in indexes has to be an Object";
result.append("cmdObj", cmdObj);
return false;
}
specs.push_back(e.Obj());
}
}
if (specs.size() == 0) {
errmsg = "no indexes to add";
return false;
}
// check specs
for (size_t i = 0; i < specs.size(); i++) {
BSONObj spec = specs[i];
if (spec["ns"].eoo()) {
spec = _addNsToSpec(ns, spec);
specs[i] = spec;
}
if (spec["ns"].type() != String) {
errmsg = "spec has no ns";
result.append("spec", spec);
return false;
}
if (ns != spec["ns"].String()) {
errmsg = "namespace mismatch";
result.append("spec", spec);
return false;
}
}
// now we know we have to create index(es)
// Note: createIndexes command does not currently respect shard versioning.
ScopedTransaction transaction(txn, MODE_IX);
Lock::DBLock dbLock(txn->lockState(), ns.db(), MODE_X);
if (!repl::getGlobalReplicationCoordinator()->canAcceptWritesFor(ns)) {
return appendCommandStatus(
result,
Status(ErrorCodes::NotMaster,
str::stream() << "Not primary while creating indexes in " << ns.ns()));
}
Database* db = dbHolder().get(txn, ns.db());
if (!db) {
db = dbHolder().openDb(txn, ns.db());
}
Collection* collection = db->getCollection(ns.ns());
result.appendBool("createdCollectionAutomatically", collection == NULL);
if (!collection) {
MONGO_WRITE_CONFLICT_RETRY_LOOP_BEGIN {
WriteUnitOfWork wunit(txn);
collection = db->createCollection(txn, ns.ns(), CollectionOptions());
invariant(collection);
wunit.commit();
}
MONGO_WRITE_CONFLICT_RETRY_LOOP_END(txn, "createIndexes", ns.ns());
}
bool run(OperationContext* txn, const string& dbname,
BSONObj& cmdObj,
int,
string& errmsg,
BSONObjBuilder& result,
bool fromRepl ) {
//
// Correct behavior here is very finicky.
//
// 1. The first step is to append the error that occurred on the previous operation.
// This adds an "err" field to the command, which is *not* the command failing.
//
// 2. Next we parse and validate write concern options. If these options are invalid
// the command fails no matter what, even if we actually had an error earlier. The
// reason for checking here is to match legacy behavior on these kind of failures -
// we'll still get an "err" field for the write error.
//
// 3. If we had an error on the previous operation, we then return immediately.
//
// 4. Finally, we actually enforce the write concern. All errors *except* timeout are
// reported with ok : 0.0, to match legacy behavior.
//
// There is a special case when "wOpTime" and "wElectionId" are explicitly provided by
// the client (mongos) - in this case we *only* enforce the write concern if it is
// valid.
//
// We always need to either report "err" (if ok : 1) or "errmsg" (if ok : 0), even if
// err is null.
//
LastError *le = lastError.disableForCommand();
// Always append lastOp and connectionId
Client& c = cc();
c.appendLastOp( result );
// for sharding; also useful in general for debugging
result.appendNumber( "connectionId" , c.getConnectionId() );
OpTime lastOpTime;
BSONField<OpTime> wOpTimeField("wOpTime");
FieldParser::FieldState extracted = FieldParser::extract(cmdObj, wOpTimeField,
&lastOpTime, &errmsg);
if (!extracted) {
result.append("badGLE", cmdObj);
appendCommandStatus(result, false, errmsg);
return false;
}
bool lastOpTimePresent = extracted != FieldParser::FIELD_NONE;
if (!lastOpTimePresent) {
// Use the client opTime if no wOpTime is specified
lastOpTime = cc().getLastOp();
}
OID electionId;
BSONField<OID> wElectionIdField("wElectionId");
extracted = FieldParser::extract(cmdObj, wElectionIdField,
&electionId, &errmsg);
if (!extracted) {
result.append("badGLE", cmdObj);
appendCommandStatus(result, false, errmsg);
return false;
}
bool electionIdPresent = extracted != FieldParser::FIELD_NONE;
bool errorOccurred = false;
// Errors aren't reported when wOpTime is used
if ( !lastOpTimePresent ) {
if ( le->nPrev != 1 ) {
errorOccurred = LastError::noError.appendSelf( result, false );
le->appendSelfStatus( result );
}
else {
errorOccurred = le->appendSelf( result, false );
}
}
BSONObj writeConcernDoc = cmdObj;
// Use the default options if we have no gle options aside from wOpTime/wElectionId
const int nFields = cmdObj.nFields();
bool useDefaultGLEOptions = (nFields == 1) ||
(nFields == 2 && lastOpTimePresent) ||
(nFields == 3 && lastOpTimePresent && electionIdPresent);
if ( useDefaultGLEOptions && getLastErrorDefault ) {
writeConcernDoc = *getLastErrorDefault;
}
//
// Validate write concern no matter what, this matches 2.4 behavior
//
WriteConcernOptions writeConcern;
Status status = writeConcern.parse( writeConcernDoc );
if ( status.isOK() ) {
// Ensure options are valid for this host
status = validateWriteConcern( writeConcern );
}
if ( !status.isOK() ) {
result.append( "badGLE", writeConcernDoc );
return appendCommandStatus( result, status );
}
// Don't wait for replication if there was an error reported - this matches 2.4 behavior
if ( errorOccurred ) {
dassert( !lastOpTimePresent );
return true;
}
// No error occurred, so we won't duplicate these fields with write concern errors
dassert( result.asTempObj()["err"].eoo() );
dassert( result.asTempObj()["code"].eoo() );
// If we got an electionId, make sure it matches
if (electionIdPresent) {
if (!theReplSet) {
// Ignore electionIds of 0 from mongos.
if (electionId != OID()) {
errmsg = "wElectionId passed but no replication active";
result.append("code", ErrorCodes::BadValue);
return false;
}
}
else {
if (electionId != theReplSet->getElectionId()) {
LOG(3) << "oid passed in is " << electionId
<< ", but our id is " << theReplSet->getElectionId();
errmsg = "election occurred after write";
result.append("code", ErrorCodes::WriteConcernFailed);
return false;
}
}
}
cc().curop()->setMessage( "waiting for write concern" );
WriteConcernResult wcResult;
status = waitForWriteConcern( txn, writeConcern, lastOpTime, &wcResult );
wcResult.appendTo( writeConcern, &result );
// For backward compatibility with 2.4, wtimeout returns ok : 1.0
if ( wcResult.wTimedOut ) {
dassert( !wcResult.err.empty() ); // so we always report err
dassert( !status.isOK() );
result.append( "errmsg", "timed out waiting for slaves" );
result.append( "code", status.code() );
return true;
}
return appendCommandStatus( result, status );
}
virtual bool run(OperationContext* txn, const string& dbname, BSONObj& cmdObj, int options,
string& errmsg, BSONObjBuilder& result,
bool fromRepl = false ) {
// --- parse
NamespaceString ns( dbname, cmdObj[name].String() );
Status status = userAllowedWriteNS( ns );
if ( !status.isOK() )
return appendCommandStatus( result, status );
if ( cmdObj["indexes"].type() != Array ) {
errmsg = "indexes has to be an array";
result.append( "cmdObj", cmdObj );
return false;
}
std::vector<BSONObj> specs;
{
BSONObjIterator i( cmdObj["indexes"].Obj() );
while ( i.more() ) {
BSONElement e = i.next();
if ( e.type() != Object ) {
errmsg = "everything in indexes has to be an Object";
result.append( "cmdObj", cmdObj );
return false;
}
specs.push_back( e.Obj() );
}
}
if ( specs.size() == 0 ) {
errmsg = "no indexes to add";
return false;
}
// check specs
for ( size_t i = 0; i < specs.size(); i++ ) {
BSONObj spec = specs[i];
if ( spec["ns"].eoo() ) {
spec = _addNsToSpec( ns, spec );
specs[i] = spec;
}
if ( spec["ns"].type() != String ) {
errmsg = "spec has no ns";
result.append( "spec", spec );
return false;
}
if ( ns != spec["ns"].String() ) {
errmsg = "namespace mismatch";
result.append( "spec", spec );
return false;
}
}
// now we know we have to create index(es)
// Note: createIndexes command does not currently respect shard versioning.
Lock::DBLock lk(txn->lockState(), ns.db(), MODE_X);
Client::Context ctx(txn, ns.ns(), false /* doVersion */ );
Database* db = ctx.db();
Collection* collection = db->getCollection( txn, ns.ns() );
result.appendBool( "createdCollectionAutomatically", collection == NULL );
if ( !collection ) {
WriteUnitOfWork wunit(txn);
collection = db->createCollection( txn, ns.ns() );
invariant( collection );
if (!fromRepl) {
repl::logOp(txn, "c", (dbname + ".$cmd").c_str(), BSON("create" << ns.coll()));
}
wunit.commit();
}
result.append( "numIndexesBefore", collection->getIndexCatalog()->numIndexesTotal(txn) );
MultiIndexBlock indexer(txn, collection);
indexer.allowBackgroundBuilding();
indexer.allowInterruption();
const size_t origSpecsSize = specs.size();
indexer.removeExistingIndexes(&specs);
if (specs.size() == 0) {
result.append( "note", "all indexes already exist" );
return true;
}
if (specs.size() != origSpecsSize) {
result.append( "note", "index already exists" );
}
for ( size_t i = 0; i < specs.size(); i++ ) {
const BSONObj& spec = specs[i];
if ( spec["unique"].trueValue() ) {
status = checkUniqueIndexConstraints(txn, ns.ns(), spec["key"].Obj());
if ( !status.isOK() ) {
appendCommandStatus( result, status );
return false;
}
}
}
uassertStatusOK(indexer.init(specs));
uassertStatusOK(indexer.insertAllDocumentsInCollection());
{
WriteUnitOfWork wunit(txn);
indexer.commit();
if ( !fromRepl ) {
for ( size_t i = 0; i < specs.size(); i++ ) {
std::string systemIndexes = ns.getSystemIndexesCollection();
repl::logOp(txn, "i", systemIndexes.c_str(), specs[i]);
}
}
wunit.commit();
}
result.append( "numIndexesAfter", collection->getIndexCatalog()->numIndexesTotal(txn) );
return true;
}
bool CmdExplain::run(OperationContext* txn,
const string& dbname,
BSONObj& cmdObj, int options,
string& errmsg,
BSONObjBuilder& result,
bool fromRepl) {
// Should never get explain commands issued from replication.
if (fromRepl) {
Status commandStat(ErrorCodes::IllegalOperation,
"explain command should not be from repl");
appendCommandStatus(result, commandStat);
return false;
}
// Get the verbosity. We use the executionStats verbosity by default.
Explain::Verbosity verbosity = Explain::EXEC_STATS;
if (!cmdObj["verbosity"].eoo()) {
const char* verbStr = cmdObj["verbosity"].valuestrsafe();
if (mongoutils::str::equals(verbStr, "queryPlanner")) {
verbosity = Explain::QUERY_PLANNER;
}
else if (mongoutils::str::equals(verbStr, "allPlansExecution")) {
verbosity = Explain::EXEC_ALL_PLANS;
}
else if (mongoutils::str::equals(verbStr, "full")) {
verbosity = Explain::FULL;
}
else if (!mongoutils::str::equals(verbStr, "executionStats")) {
Status commandStat(ErrorCodes::BadValue,
"verbosity string must be one of "
"{'queryPlanner', 'executionStats', 'allPlansExecution'}");
appendCommandStatus(result, commandStat);
return false;
}
}
if (Object != cmdObj.firstElement().type()) {
Status commandStat(ErrorCodes::BadValue,
"explain command requires a nested object");
appendCommandStatus(result, commandStat);
return false;
}
// This is the nested command which we are explaining.
BSONObj explainObj = cmdObj.firstElement().Obj();
Command* commToExplain = Command::findCommand(explainObj.firstElementFieldName());
if (NULL == commToExplain) {
mongoutils::str::stream ss;
ss << "unknown command: " << explainObj.firstElementFieldName();
Status explainStatus(ErrorCodes::CommandNotFound, ss);
return appendCommandStatus(result, explainStatus);
}
// Check whether the child command is allowed to run here. TODO: this logic is
// copied from Command::execCommand and should be abstracted. Until then, make
// sure to keep it up to date.
repl::ReplicationCoordinator* replCoord = repl::getGlobalReplicationCoordinator();
bool canRunHere =
replCoord->canAcceptWritesForDatabase(dbname) ||
commToExplain->slaveOk() ||
(commToExplain->slaveOverrideOk() && (options & QueryOption_SlaveOk));
if (!canRunHere) {
mongoutils::str::stream ss;
ss << "Explain's child command cannot run on this node. "
<< "Are you explaining a write command on a secondary?";
appendCommandStatus(result, false, ss);
return false;
}
// Actually call the nested command's explain(...) method.
Status explainStatus = commToExplain->explain(txn, dbname, explainObj, verbosity, &result);
if (!explainStatus.isOK()) {
return appendCommandStatus(result, explainStatus);
}
return true;
}
