void run() {
OldClientWriteContext ctx(&_txn, nss.ns());
insert(BSON("a" << 1 << "b" << 1));
Collection* collection = ctx.getCollection();
BSONObj filterObj = fromjson("{_id: {$gt: 0}, b: {$gt: 0}}");
PlanExecutor* exec = makeCollScanExec(collection, filterObj);
// Make a client cursor from the runner.
ClientCursor* cc =
new ClientCursor(collection->getCursorManager(), exec, nss.ns(), false, 0, BSONObj());
ClientCursorPin ccPin(collection->getCursorManager(), cc->cursorid());
// If the cursor is pinned, it sticks around,
// even after invalidation.
ASSERT_EQUALS(1U, numCursors());
const std::string invalidateReason("InvalidatePinned Test");
collection->getCursorManager()->invalidateAll(false, invalidateReason);
ASSERT_EQUALS(1U, numCursors());
// The invalidation should have killed the runner.
BSONObj objOut;
ASSERT_EQUALS(PlanExecutor::DEAD, exec->getNext(&objOut, NULL));
ASSERT(WorkingSetCommon::isValidStatusMemberObject(objOut));
const Status status = WorkingSetCommon::getMemberObjectStatus(objOut);
ASSERT(status.reason().find(invalidateReason) != string::npos);
// Deleting the underlying cursor should cause the
// number of cursors to return to 0.
ccPin.deleteUnderlying();
ASSERT_EQUALS(0U, numCursors());
}
// static
long long DeleteStage::getNumDeleted(const PlanExecutor& exec) {
invariant(exec.getRootStage()->isEOF());
invariant(exec.getRootStage()->stageType() == STAGE_DELETE);
DeleteStage* deleteStage = static_cast<DeleteStage*>(exec.getRootStage());
const DeleteStats* deleteStats =
static_cast<const DeleteStats*>(deleteStage->getSpecificStats());
return deleteStats->docsDeleted;
}
/**
* Uses 'cursor' and 'request' to fill out 'nextBatch' with the batch of result documents to
* be returned by this getMore.
*
* Returns the number of documents in the batch in *numResults, which must be initialized to
* zero by the caller. Returns the final ExecState returned by the cursor in *state.
*
* Returns an OK status if the batch was successfully generated, and a non-OK status if the
* PlanExecutor encounters a failure.
*/
Status generateBatch(ClientCursor* cursor,
const GetMoreRequest& request,
BSONArrayBuilder* nextBatch,
PlanExecutor::ExecState* state,
int* numResults) {
PlanExecutor* exec = cursor->getExecutor();
const bool isAwaitData = isCursorAwaitData(cursor);
// If an awaitData getMore is killed during this process due to our max time expiring at
// an interrupt point, we just continue as normal and return rather than reporting a
// timeout to the user.
BSONObj obj;
try {
while (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 (nextBatch->len() + obj.objsize() > BSONObjMaxUserSize && *numResults > 0) {
exec->enqueue(obj);
break;
}
// Add result to output buffer.
nextBatch->append(obj);
(*numResults)++;
if (enoughForGetMore(request.batchSize.value_or(0),
*numResults, nextBatch->len())) {
break;
}
}
}
catch (const UserException& except) {
if (isAwaitData && except.getCode() == ErrorCodes::ExceededTimeLimit) {
// We ignore exceptions from interrupt points due to max time expiry for
// awaitData cursors.
}
else {
throw;
}
}
if (PlanExecutor::FAILURE == *state) {
const std::unique_ptr<PlanStageStats> stats(exec->getStats());
error() << "GetMore executor error, stats: " << Explain::statsToBSON(*stats);
return Status(ErrorCodes::OperationFailed,
str::stream() << "GetMore executor error: "
<< WorkingSetCommon::toStatusString(obj));
}
else if (PlanExecutor::DEAD == *state) {
return Status(ErrorCodes::OperationFailed,
str::stream() << "Plan executor killed during getMore command, "
<< "ns: " << request.nss.ns());
}
return Status::OK();
}
// static
void Explain::getSummaryStats(const PlanExecutor& exec, PlanSummaryStats* statsOut) {
invariant(NULL != statsOut);
PlanStage* root = exec.getRootStage();
// We can get some of the fields we need from the common stats stored in the
// root stage of the plan tree.
const CommonStats* common = root->getCommonStats();
statsOut->nReturned = common->advanced;
statsOut->executionTimeMillis = common->executionTimeMillis;
// The other fields are aggregations over the stages in the plan tree. We flatten
// the tree into a list and then compute these aggregations.
std::vector<const PlanStage*> stages;
flattenExecTree(root, &stages);
for (size_t i = 0; i < stages.size(); i++) {
statsOut->totalKeysExamined +=
getKeysExamined(stages[i]->stageType(), stages[i]->getSpecificStats());
statsOut->totalDocsExamined +=
getDocsExamined(stages[i]->stageType(), stages[i]->getSpecificStats());
if (STAGE_IDHACK == stages[i]->stageType()) {
statsOut->isIdhack = true;
}
if (STAGE_SORT == stages[i]->stageType()) {
statsOut->hasSortStage = true;
}
}
}
int countResults(const IndexScanParams& params, BSONObj filterObj = BSONObj()) {
Client::ReadContext ctx(ns());
PlanExecutor runner;
StatusWithMatchExpression swme = MatchExpressionParser::parse(filterObj);
verify(swme.isOK());
auto_ptr<MatchExpression> filterExpr(swme.getValue());
runner.setRoot(new IndexScan(params, runner.getWorkingSet(), filterExpr.get()));
int count = 0;
for (BSONObj obj; Runner::RUNNER_ADVANCED == runner.getNext(&obj); ) {
++count;
}
return count;
}
void endQueryOp(OperationContext* txn,
Collection* collection,
const PlanExecutor& exec,
int dbProfilingLevel,
long long numResults,
CursorId cursorId) {
auto curop = CurOp::get(txn);
// Fill out basic curop query exec properties.
curop->debug().nreturned = numResults;
curop->debug().cursorid = (0 == cursorId ? -1 : cursorId);
curop->debug().cursorExhausted = (0 == cursorId);
// Fill out curop based on explain summary statistics.
PlanSummaryStats summaryStats;
Explain::getSummaryStats(exec, &summaryStats);
curop->debug().hasSortStage = summaryStats.hasSortStage;
curop->debug().keysExamined = summaryStats.totalKeysExamined;
curop->debug().docsExamined = summaryStats.totalDocsExamined;
curop->debug().idhack = summaryStats.isIdhack;
if (collection) {
collection->infoCache()->notifyOfQuery(txn, summaryStats.indexesUsed);
}
const logger::LogComponent commandLogComponent = logger::LogComponent::kCommand;
const logger::LogSeverity logLevelOne = logger::LogSeverity::Debug(1);
// Set debug information for consumption by the profiler and slow query log.
if (dbProfilingLevel > 0 || curop->elapsedMillis() > serverGlobalParams.slowMS ||
logger::globalLogDomain()->shouldLog(commandLogComponent, logLevelOne)) {
// Generate plan summary string.
stdx::lock_guard<Client>(*txn->getClient());
curop->setPlanSummary_inlock(Explain::getPlanSummary(&exec));
}
// Set debug information for consumption by the profiler only.
if (dbProfilingLevel > 0) {
// Get BSON stats.
unique_ptr<PlanStageStats> execStats(exec.getStats());
BSONObjBuilder statsBob;
Explain::statsToBSON(*execStats, &statsBob);
curop->debug().execStats.set(statsBob.obj());
// Replace exec stats with plan summary if stats cannot fit into CachedBSONObj.
if (curop->debug().execStats.tooBig() && !curop->getPlanSummary().empty()) {
BSONObjBuilder bob;
bob.append("summary", curop->getPlanSummary());
curop->debug().execStats.set(bob.done());
}
}
}
/**
* 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;
}
/**
* Called by db/instance.cpp. This is the getMore entry point.
*
* pass - when QueryOption_AwaitData is in use, the caller will make repeated calls
* when this method returns an empty result, incrementing pass on each call.
* Thus, pass == 0 indicates this is the first "attempt" before any 'awaiting'.
*/
QueryResult::View newGetMore(OperationContext* txn,
const char* ns,
int ntoreturn,
long long cursorid,
CurOp& curop,
int pass,
bool& exhaust,
bool* isCursorAuthorized,
bool fromDBDirectClient) {
// For testing, we may want to fail if we receive a getmore.
if (MONGO_FAIL_POINT(failReceivedGetmore)) {
invariant(0);
}
exhaust = false;
// This is a read lock.
const NamespaceString nss(ns);
scoped_ptr<AutoGetCollectionForRead> ctx(new AutoGetCollectionForRead(txn, nss));
Collection* collection = ctx->getCollection();
uassert( 17356, "collection dropped between getMore calls", collection );
QLOG() << "Running getMore, cursorid: " << cursorid << endl;
// This checks to make sure the operation is allowed on a replicated node. Since we are not
// passing in a query object (necessary to check SlaveOK query option), the only state where
// reads are allowed is PRIMARY (or master in master/slave). This function uasserts if
// reads are not okay.
Status status = repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(
txn,
nss,
true);
uassertStatusOK(status);
// A pin performs a CC lookup and if there is a CC, increments the CC's pin value so it
// doesn't time out. Also informs ClientCursor that there is somebody actively holding the
// CC, so don't delete it.
ClientCursorPin ccPin(collection, cursorid);
ClientCursor* cc = ccPin.c();
// If we're not being called from DBDirectClient we want to associate the RecoveryUnit
// used to create the execution machinery inside the cursor with our OperationContext.
// If we throw or otherwise exit this method in a disorderly fashion, we must ensure
// that further calls to getMore won't fail, and that the provided OperationContext
// has a valid RecoveryUnit. As such, we use RAII to accomplish this.
//
// This must be destroyed before the ClientCursor is destroyed.
std::auto_ptr<ScopedRecoveryUnitSwapper> ruSwapper;
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
const int InitialBufSize =
512 + sizeof(QueryResult::Value) + MaxBytesToReturnToClientAtOnce;
BufBuilder bb(InitialBufSize);
bb.skip(sizeof(QueryResult::Value));
if (NULL == cc) {
cursorid = 0;
resultFlags = ResultFlag_CursorNotFound;
}
else {
// Quote: check for spoofing of the ns such that it does not match the one originally
// there for the cursor
uassert(17011, "auth error", str::equals(ns, cc->ns().c_str()));
*isCursorAuthorized = true;
// Restore the RecoveryUnit if we need to.
if (fromDBDirectClient) {
if (cc->hasRecoveryUnit())
invariant(txn->recoveryUnit() == cc->getUnownedRecoveryUnit());
}
else {
if (!cc->hasRecoveryUnit()) {
// Start using a new RecoveryUnit
cc->setOwnedRecoveryUnit(
getGlobalEnvironment()->getGlobalStorageEngine()->newRecoveryUnit(txn));
}
// Swap RecoveryUnit(s) between the ClientCursor and OperationContext.
ruSwapper.reset(new ScopedRecoveryUnitSwapper(cc, txn));
}
// Reset timeout timer on the cursor since the cursor is still in use.
cc->setIdleTime(0);
// TODO: fail point?
// If the operation that spawned this cursor had a time limit set, apply leftover
// time to this getmore.
curop.setMaxTimeMicros(cc->getLeftoverMaxTimeMicros());
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
if (0 == pass) {
cc->updateSlaveLocation(txn, curop);
}
if (cc->isAggCursor) {
// Agg cursors handle their own locking internally.
ctx.reset(); // unlocks
}
CollectionMetadataPtr collMetadata = cc->getCollMetadata();
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// What number result are we starting at? Used to fill out the reply.
startingResult = cc->pos();
// What gives us results.
PlanExecutor* exec = cc->getExecutor();
const int queryOptions = cc->queryOptions();
// Get results out of the executor.
exec->restoreState(txn);
BSONObj obj;
PlanExecutor::ExecState state;
while (PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (queryOptions & QueryOption_OplogReplay) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
if ((ntoreturn && numResults >= ntoreturn)
|| bb.len() > MaxBytesToReturnToClientAtOnce) {
break;
}
}
// We save the client cursor when there might be more results, and hence we may receive
// another getmore. If we receive a EOF or an error, or 'exec' is dead, then we know
// that we will not be producing more results. We indicate that the cursor is closed by
// sending a cursorId of 0 back to the client.
//
// On the other hand, if we retrieve all results necessary for this batch, then
// 'saveClientCursor' is true and we send a valid cursorId back to the client. In
// this case, there may or may not actually be more results (for example, the next call
// to getNext(...) might just return EOF).
bool saveClientCursor = false;
if (PlanExecutor::DEAD == state || PlanExecutor::EXEC_ERROR == state) {
// Propagate this error to caller.
if (PlanExecutor::EXEC_ERROR == state) {
scoped_ptr<PlanStageStats> stats(exec->getStats());
error() << "Plan executor error, stats: "
<< Explain::statsToBSON(*stats);
uasserted(17406, "getMore executor error: " +
WorkingSetCommon::toStatusString(obj));
}
// If we're dead there's no way to get more results.
saveClientCursor = false;
// In the old system tailable capped cursors would be killed off at the
// cursorid level. If a tailable capped cursor is nuked the cursorid
// would vanish.
//
// In the new system they die and are cleaned up later (or time out).
// So this is where we get to remove the cursorid.
if (0 == numResults) {
resultFlags = ResultFlag_CursorNotFound;
}
}
else if (PlanExecutor::IS_EOF == state) {
// EOF is also end of the line unless it's tailable.
saveClientCursor = queryOptions & QueryOption_CursorTailable;
}
else {
verify(PlanExecutor::ADVANCED == state);
saveClientCursor = true;
}
if (!saveClientCursor) {
ruSwapper.reset();
ccPin.deleteUnderlying();
// cc is now invalid, as is the executor
cursorid = 0;
cc = NULL;
QLOG() << "getMore NOT saving client cursor, ended with state "
<< PlanExecutor::statestr(state)
<< endl;
}
else {
// Continue caching the ClientCursor.
cc->incPos(numResults);
exec->saveState();
QLOG() << "getMore saving client cursor ended with state "
<< PlanExecutor::statestr(state)
<< endl;
if (PlanExecutor::IS_EOF == state && (queryOptions & QueryOption_CursorTailable)) {
if (!fromDBDirectClient) {
// Don't stash the RU. Get a new one on the next getMore.
ruSwapper.reset();
delete cc->releaseOwnedRecoveryUnit();
}
if ((queryOptions & QueryOption_AwaitData)
&& (numResults == 0)
&& (pass < 1000)) {
// Bubble up to the AwaitData handling code in receivedGetMore which will
// try again.
return NULL;
}
}
// Possibly note slave's position in the oplog.
if ((queryOptions & QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
exhaust = (queryOptions & QueryOption_Exhaust);
// If the getmore had a time limit, remaining time is "rolled over" back to the
// cursor (for use by future getmore ops).
cc->setLeftoverMaxTimeMicros( curop.getRemainingMaxTimeMicros() );
}
}
QueryResult::View qr = bb.buf();
qr.msgdata().setLen(bb.len());
qr.msgdata().setOperation(opReply);
qr.setResultFlags(resultFlags);
qr.setCursorId(cursorid);
qr.setStartingFrom(startingResult);
qr.setNReturned(numResults);
bb.decouple();
QLOG() << "getMore returned " << numResults << " results\n";
return qr;
}
virtual bool run(OperationContext* txn, const string& dbname, BSONObj& cmdObj, int options,
string& errmsg, BSONObjBuilder& result,
bool fromRepl = false ) {
NamespaceString ns( dbname, cmdObj[name].String() );
AutoGetCollectionForRead ctx(txn, ns.ns());
Collection* collection = ctx.getCollection();
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(txn));
if (iterators.size() < numCursors) {
numCursors = iterators.size();
}
OwnedPointerVector<PlanExecutor> execs;
for ( size_t i = 0; i < numCursors; i++ ) {
WorkingSet* ws = new WorkingSet();
MultiIteratorStage* mis = new MultiIteratorStage(txn, ws, collection);
PlanExecutor* rawExec;
// Takes ownership of 'ws' and 'mis'.
Status execStatus = PlanExecutor::make(txn, ws, mis, collection,
PlanExecutor::YIELD_AUTO, &rawExec);
invariant(execStatus.isOK());
auto_ptr<PlanExecutor> curExec(rawExec);
// The PlanExecutor was registered on construction due to the YIELD_AUTO policy.
// We have to deregister it, as it will be registered with ClientCursor.
curExec->deregisterExec();
// Need to save state while yielding locks between now and getMore().
curExec->saveState();
execs.push_back(curExec.release());
}
// transfer iterators to executors 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++) {
PlanExecutor* theExec = execs[i % execs.size()];
MultiIteratorStage* mis = static_cast<MultiIteratorStage*>(theExec->getRootStage());
// This wasn't called above as they weren't assigned yet
iterators[i]->saveState();
mis->addIterator(iterators.releaseAt(i));
}
{
BSONArrayBuilder bucketsBuilder;
for (size_t i = 0; i < execs.size(); i++) {
// transfer ownership of an executor to the ClientCursor (which manages its own
// lifetime).
ClientCursor* cc = new ClientCursor( collection->getCursorManager(),
execs.releaseAt(i),
ns.ns() );
BSONObjBuilder threadResult;
appendCursorResponseObject( cc->cursorid(),
ns.ns(),
BSONArray(),
&threadResult );
threadResult.appendBool( "ok", 1 );
bucketsBuilder.append( threadResult.obj() );
}
result.appendArray( "cursors", bucketsBuilder.obj() );
}
return true;
}
virtual bool run(OperationContext* txn, const string& dbname, BSONObj& cmdObj, int options,
string& errmsg, BSONObjBuilder& result,
bool fromRepl = false ) {
NamespaceString ns( dbname, cmdObj[name].String() );
AutoGetCollectionForRead ctx(txn, ns.ns());
Collection* collection = ctx.getCollection();
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(txn));
if (iterators.size() < numCursors) {
numCursors = iterators.size();
}
OwnedPointerVector<PlanExecutor> execs;
for ( size_t i = 0; i < numCursors; i++ ) {
WorkingSet* ws = new WorkingSet();
MultiIteratorStage* mis = new MultiIteratorStage(txn, ws, collection);
// Takes ownership of 'ws' and 'mis'.
auto_ptr<PlanExecutor> curExec(new PlanExecutor(txn, ws, mis, collection));
// Each of the plan executors should yield automatically. We pass "false" to
// indicate that 'curExec' should not register itself, as it will get registered
// by ClientCursor instead.
curExec->setYieldPolicy(PlanExecutor::YIELD_AUTO, false);
// Need to save state while yielding locks between now and newGetMore.
curExec->saveState();
execs.push_back(curExec.release());
}
// transfer iterators to executors 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++) {
PlanExecutor* theExec = execs[i % execs.size()];
MultiIteratorStage* mis = static_cast<MultiIteratorStage*>(theExec->getRootStage());
mis->addIterator(iterators.releaseAt(i));
}
{
BSONArrayBuilder bucketsBuilder;
for (size_t i = 0; i < execs.size(); i++) {
// transfer ownership of an executor to the ClientCursor (which manages its own
// lifetime).
ClientCursor* cc = new ClientCursor( collection, execs.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;
}
/**
* Called by db/instance.cpp. This is the getMore entry point.
*
* pass - when QueryOption_AwaitData is in use, the caller will make repeated calls
* when this method returns an empty result, incrementing pass on each call.
* Thus, pass == 0 indicates this is the first "attempt" before any 'awaiting'.
*/
QueryResult::View getMore(OperationContext* txn,
const char* ns,
int ntoreturn,
long long cursorid,
CurOp& curop,
int pass,
bool& exhaust,
bool* isCursorAuthorized) {
// For testing, we may want to fail if we receive a getmore.
if (MONGO_FAIL_POINT(failReceivedGetmore)) {
invariant(0);
}
exhaust = false;
const NamespaceString nss(ns);
// 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).
boost::scoped_ptr<AutoGetCollectionForRead> ctx;
boost::scoped_ptr<Lock::DBLock> unpinDBLock;
boost::scoped_ptr<Lock::CollectionLock> unpinCollLock;
CursorManager* cursorManager;
CursorManager* globalCursorManager = CursorManager::getGlobalCursorManager();
if (globalCursorManager->ownsCursorId(cursorid)) {
cursorManager = globalCursorManager;
}
else {
ctx.reset(new AutoGetCollectionForRead(txn, nss));
Collection* collection = ctx->getCollection();
uassert( 17356, "collection dropped between getMore calls", collection );
cursorManager = collection->getCursorManager();
}
LOG(5) << "Running getMore, cursorid: " << cursorid << endl;
// This checks to make sure the operation is allowed on a replicated node. Since we are not
// passing in a query object (necessary to check SlaveOK query option), the only state where
// reads are allowed is PRIMARY (or master in master/slave). This function uasserts if
// reads are not okay.
Status status = repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(
txn,
nss,
true);
uassertStatusOK(status);
// A pin performs a CC lookup and if there is a CC, increments the CC's pin value so it
// doesn't time out. Also informs ClientCursor that there is somebody actively holding the
// CC, so don't delete it.
ClientCursorPin ccPin(cursorManager, cursorid);
ClientCursor* cc = ccPin.c();
// If we're not being called from DBDirectClient we want to associate the RecoveryUnit
// used to create the execution machinery inside the cursor with our OperationContext.
// If we throw or otherwise exit this method in a disorderly fashion, we must ensure
// that further calls to getMore won't fail, and that the provided OperationContext
// has a valid RecoveryUnit. As such, we use RAII to accomplish this.
//
// This must be destroyed before the ClientCursor is destroyed.
std::auto_ptr<ScopedRecoveryUnitSwapper> ruSwapper;
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
const int InitialBufSize =
512 + sizeof(QueryResult::Value) + MaxBytesToReturnToClientAtOnce;
BufBuilder bb(InitialBufSize);
bb.skip(sizeof(QueryResult::Value));
if (NULL == cc) {
cursorid = 0;
resultFlags = ResultFlag_CursorNotFound;
}
else {
// Check for spoofing of the ns such that it does not match the one originally
// there for the cursor.
uassert(ErrorCodes::Unauthorized,
str::stream() << "Requested getMore on namespace " << ns << ", but cursor "
<< cursorid << " belongs to namespace " << cc->ns(),
ns == cc->ns());
*isCursorAuthorized = true;
// Restore the RecoveryUnit if we need to.
if (txn->getClient()->isInDirectClient()) {
if (cc->hasRecoveryUnit())
invariant(txn->recoveryUnit() == cc->getUnownedRecoveryUnit());
}
else {
if (!cc->hasRecoveryUnit()) {
// Start using a new RecoveryUnit
cc->setOwnedRecoveryUnit(
getGlobalServiceContext()->getGlobalStorageEngine()->newRecoveryUnit());
}
// Swap RecoveryUnit(s) between the ClientCursor and OperationContext.
ruSwapper.reset(new ScopedRecoveryUnitSwapper(cc, txn));
}
// Reset timeout timer on the cursor since the cursor is still in use.
cc->setIdleTime(0);
// If the operation that spawned this cursor had a time limit set, apply leftover
// time to this getmore.
curop.setMaxTimeMicros(cc->getLeftoverMaxTimeMicros());
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
if (0 == pass) {
cc->updateSlaveLocation(txn);
}
if (cc->isAggCursor()) {
// Agg cursors handle their own locking internally.
ctx.reset(); // unlocks
}
// If we're replaying the oplog, we save the last time that we read.
Timestamp slaveReadTill;
// What number result are we starting at? Used to fill out the reply.
startingResult = cc->pos();
// What gives us results.
PlanExecutor* exec = cc->getExecutor();
const int queryOptions = cc->queryOptions();
// Get results out of the executor.
exec->restoreState(txn);
BSONObj obj;
PlanExecutor::ExecState state;
while (PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (queryOptions & QueryOption_OplogReplay) {
BSONElement e = obj["ts"];
if (Date == e.type() || bsonTimestamp == e.type()) {
slaveReadTill = e.timestamp();
}
}
if (enoughForGetMore(ntoreturn, numResults, bb.len())) {
break;
}
}
if (PlanExecutor::DEAD == state || PlanExecutor::FAILURE == state) {
// Propagate this error to caller.
if (PlanExecutor::FAILURE == state) {
scoped_ptr<PlanStageStats> stats(exec->getStats());
error() << "Plan executor error, stats: "
<< Explain::statsToBSON(*stats);
uasserted(17406, "getMore executor error: " +
WorkingSetCommon::toStatusString(obj));
}
// In the old system tailable capped cursors would be killed off at the
// cursorid level. If a tailable capped cursor is nuked the cursorid
// would vanish.
//
// In the new system they die and are cleaned up later (or time out).
// So this is where we get to remove the cursorid.
if (0 == numResults) {
resultFlags = ResultFlag_CursorNotFound;
}
}
const bool shouldSaveCursor =
shouldSaveCursorGetMore(state, exec, isCursorTailable(cc));
// In order to deregister a cursor, we need to be holding the DB + collection lock and
// if the cursor is aggregation, we release these locks.
if (cc->isAggCursor()) {
invariant(NULL == ctx.get());
unpinDBLock.reset(new Lock::DBLock(txn->lockState(), nss.db(), MODE_IS));
unpinCollLock.reset(new Lock::CollectionLock(txn->lockState(), nss.ns(), MODE_IS));
}
// Our two possible ClientCursorPin cleanup paths are:
// 1) If the cursor is not going to be saved, we call deleteUnderlying() on the pin.
// 2) If the cursor is going to be saved, we simply let the pin go out of scope. In
// this case, the pin's destructor will be invoked, which will call release() on the
// pin. Because our ClientCursorPin is declared after our lock is declared, this
// will happen under the lock.
if (!shouldSaveCursor) {
ruSwapper.reset();
ccPin.deleteUnderlying();
// cc is now invalid, as is the executor
cursorid = 0;
cc = NULL;
curop.debug().cursorExhausted = true;
LOG(5) << "getMore NOT saving client cursor, ended with state "
<< PlanExecutor::statestr(state)
<< endl;
}
else {
// Continue caching the ClientCursor.
cc->incPos(numResults);
exec->saveState();
LOG(5) << "getMore saving client cursor ended with state "
<< PlanExecutor::statestr(state)
<< endl;
if (PlanExecutor::IS_EOF == state && (queryOptions & QueryOption_CursorTailable)) {
if (!txn->getClient()->isInDirectClient()) {
// Don't stash the RU. Get a new one on the next getMore.
ruSwapper->dismiss();
}
if ((queryOptions & QueryOption_AwaitData)
&& (numResults == 0)
&& (pass < 1000)) {
// Bubble up to the AwaitData handling code in receivedGetMore which will
// try again.
return NULL;
}
}
// Possibly note slave's position in the oplog.
if ((queryOptions & QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
exhaust = (queryOptions & QueryOption_Exhaust);
// If the getmore had a time limit, remaining time is "rolled over" back to the
// cursor (for use by future getmore ops).
cc->setLeftoverMaxTimeMicros( curop.getRemainingMaxTimeMicros() );
}
}
QueryResult::View qr = bb.buf();
qr.msgdata().setLen(bb.len());
qr.msgdata().setOperation(opReply);
qr.setResultFlags(resultFlags);
qr.setCursorId(cursorid);
qr.setStartingFrom(startingResult);
qr.setNReturned(numResults);
bb.decouple();
LOG(5) << "getMore returned " << numResults << " results\n";
return qr;
}
/**
* Called by db/instance.cpp. This is the getMore entry point.
*/
QueryResult::View getMore(OperationContext* txn,
const char* ns,
int ntoreturn,
long long cursorid,
bool* exhaust,
bool* isCursorAuthorized) {
invariant(ntoreturn >= 0);
CurOp& curop = *CurOp::get(txn);
// For testing, we may want to fail if we receive a getmore.
if (MONGO_FAIL_POINT(failReceivedGetmore)) {
invariant(0);
}
*exhaust = false;
const NamespaceString nss(ns);
// 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. These cursors are generated either by the listCollections or
// listIndexes commands, as these special cursor-generating commands operate over catalog
// data rather than targeting the data within a collection.
// 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).
unique_ptr<AutoGetCollectionForRead> ctx;
unique_ptr<Lock::DBLock> unpinDBLock;
unique_ptr<Lock::CollectionLock> unpinCollLock;
CursorManager* cursorManager;
if (nss.isListIndexesCursorNS() || nss.isListCollectionsCursorNS()) {
// List collections and list indexes are special cursor-generating commands whose
// cursors are managed globally, as they operate over catalog data rather than targeting
// the data within a collection.
cursorManager = CursorManager::getGlobalCursorManager();
} else {
ctx = stdx::make_unique<AutoGetCollectionForRead>(txn, nss);
Collection* collection = ctx->getCollection();
uassert(17356, "collection dropped between getMore calls", collection);
cursorManager = collection->getCursorManager();
}
LOG(5) << "Running getMore, cursorid: " << cursorid << endl;
// This checks to make sure the operation is allowed on a replicated node. Since we are not
// passing in a query object (necessary to check SlaveOK query option), the only state where
// reads are allowed is PRIMARY (or master in master/slave). This function uasserts if
// reads are not okay.
Status status = repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(txn, nss, true);
uassertStatusOK(status);
// A pin performs a CC lookup and if there is a CC, increments the CC's pin value so it
// doesn't time out. Also informs ClientCursor that there is somebody actively holding the
// CC, so don't delete it.
ClientCursorPin ccPin(cursorManager, cursorid);
ClientCursor* cc = ccPin.c();
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
const int InitialBufSize =
512 + sizeof(QueryResult::Value) + FindCommon::kMaxBytesToReturnToClientAtOnce;
BufBuilder bb(InitialBufSize);
bb.skip(sizeof(QueryResult::Value));
if (NULL == cc) {
cursorid = 0;
resultFlags = ResultFlag_CursorNotFound;
} else {
// Check for spoofing of the ns such that it does not match the one originally
// there for the cursor.
uassert(ErrorCodes::Unauthorized,
str::stream() << "Requested getMore on namespace " << ns << ", but cursor "
<< cursorid << " belongs to namespace " << cc->ns(),
ns == cc->ns());
*isCursorAuthorized = true;
if (cc->isReadCommitted())
uassertStatusOK(txn->recoveryUnit()->setReadFromMajorityCommittedSnapshot());
// Reset timeout timer on the cursor since the cursor is still in use.
cc->setIdleTime(0);
// If the operation that spawned this cursor had a time limit set, apply leftover
// time to this getmore.
curop.setMaxTimeMicros(cc->getLeftoverMaxTimeMicros());
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// Ensure that the original query or command object is available in the slow query log,
// profiler, and currentOp.
curop.debug().query = cc->getQuery();
{
stdx::lock_guard<Client> lk(*txn->getClient());
curop.setQuery_inlock(cc->getQuery());
}
cc->updateSlaveLocation(txn);
if (cc->isAggCursor()) {
// Agg cursors handle their own locking internally.
ctx.reset(); // unlocks
}
// If we're replaying the oplog, we save the last time that we read.
Timestamp slaveReadTill;
// What number result are we starting at? Used to fill out the reply.
startingResult = cc->pos();
uint64_t notifierVersion = 0;
std::shared_ptr<CappedInsertNotifier> notifier;
if (isCursorAwaitData(cc)) {
invariant(ctx->getCollection()->isCapped());
// 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.
notifier = ctx->getCollection()->getCappedInsertNotifier();
// Must get the version before we call generateBatch in case a write comes in after
// that call and before we call wait on the notifier.
notifierVersion = notifier->getVersion();
}
PlanExecutor* exec = cc->getExecutor();
exec->reattachToOperationContext(txn);
exec->restoreState();
PlanExecutor::ExecState state;
generateBatch(ntoreturn, cc, &bb, &numResults, &slaveReadTill, &state);
// If this is an await data cursor, and we hit EOF without generating any results, then
// we block waiting for new data to arrive.
if (isCursorAwaitData(cc) && state == PlanExecutor::IS_EOF && numResults == 0) {
// Save the PlanExecutor and drop our locks.
exec->saveState();
ctx.reset();
// Block waiting for data for up to 1 second.
Seconds timeout(1);
notifier->wait(notifierVersion, timeout);
notifier.reset();
// Set expected latency to match wait time. This makes sure the logs aren't spammed
// by awaitData queries that exceed slowms due to blocking on the CappedInsertNotifier.
curop.setExpectedLatencyMs(durationCount<Milliseconds>(timeout));
// Reacquiring locks.
ctx = make_unique<AutoGetCollectionForRead>(txn, nss);
exec->restoreState();
// We woke up because either the timed_wait expired, or there was more data. Either
// way, attempt to generate another batch of results.
generateBatch(ntoreturn, cc, &bb, &numResults, &slaveReadTill, &state);
}
// We have to do this before re-acquiring locks in the agg case because
// shouldSaveCursorGetMore() can make a network call for agg cursors.
//
// TODO: Getting rid of PlanExecutor::isEOF() in favor of PlanExecutor::IS_EOF would mean
// that this network operation is no longer necessary.
const bool shouldSaveCursor = shouldSaveCursorGetMore(state, exec, isCursorTailable(cc));
// In order to deregister a cursor, we need to be holding the DB + collection lock and
// if the cursor is aggregation, we release these locks.
if (cc->isAggCursor()) {
invariant(NULL == ctx.get());
unpinDBLock = make_unique<Lock::DBLock>(txn->lockState(), nss.db(), MODE_IS);
unpinCollLock = make_unique<Lock::CollectionLock>(txn->lockState(), nss.ns(), MODE_IS);
}
// Our two possible ClientCursorPin cleanup paths are:
// 1) If the cursor is not going to be saved, we call deleteUnderlying() on the pin.
// 2) If the cursor is going to be saved, we simply let the pin go out of scope. In
// this case, the pin's destructor will be invoked, which will call release() on the
// pin. Because our ClientCursorPin is declared after our lock is declared, this
// will happen under the lock.
if (!shouldSaveCursor) {
ccPin.deleteUnderlying();
// cc is now invalid, as is the executor
cursorid = 0;
cc = NULL;
curop.debug().cursorExhausted = true;
LOG(5) << "getMore NOT saving client cursor, ended with state "
<< PlanExecutor::statestr(state) << endl;
} else {
// Continue caching the ClientCursor.
cc->incPos(numResults);
exec->saveState();
exec->detachFromOperationContext();
LOG(5) << "getMore saving client cursor ended with state "
<< PlanExecutor::statestr(state) << endl;
// Possibly note slave's position in the oplog.
if ((cc->queryOptions() & QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
*exhaust = cc->queryOptions() & QueryOption_Exhaust;
// If the getmore had a time limit, remaining time is "rolled over" back to the
// cursor (for use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
}
}
QueryResult::View qr = bb.buf();
qr.msgdata().setLen(bb.len());
qr.msgdata().setOperation(opReply);
qr.setResultFlags(resultFlags);
qr.setCursorId(cursorid);
qr.setStartingFrom(startingResult);
qr.setNReturned(numResults);
bb.decouple();
LOG(5) << "getMore returned " << numResults << " results\n";
return qr;
}
/**
* 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;
}
/**
* Called by db/instance.cpp. This is the getMore entry point.
*/
Message getMore(OperationContext* opCtx,
const char* ns,
int ntoreturn,
long long cursorid,
bool* exhaust,
bool* isCursorAuthorized) {
invariant(ntoreturn >= 0);
CurOp& curOp = *CurOp::get(opCtx);
curOp.ensureStarted();
// For testing, we may want to fail if we receive a getmore.
if (MONGO_FAIL_POINT(failReceivedGetmore)) {
MONGO_UNREACHABLE;
}
*exhaust = false;
const NamespaceString nss(ns);
// Cursors come in one of two flavors:
// - Cursors owned by the collection cursor manager, such as those generated via the find
// command. For these cursors, we hold the appropriate collection lock for the duration of the
// getMore using AutoGetCollectionForRead.
// - Cursors owned by the global cursor manager, such as those generated via the aggregate
// command. These cursors either hold no collection state or manage their collection state
// internally, so we acquire no locks.
//
// While we only need to acquire locks in the case of a cursor which is *not* globally owned, we
// need to create an AutoStatsTracker in either case. This is responsible for updating
// statistics in CurOp and Top. We avoid using AutoGetCollectionForReadCommand because we may
// need to drop and reacquire locks when the cursor is awaitData, but we don't want to update
// the stats twice.
//
// Note that we acquire our locks before our ClientCursorPin, in order to ensure that the pin's
// destructor is called before the lock's destructor (if there is one) so that the cursor
// cleanup can occur under the lock.
UninterruptibleLockGuard noInterrupt(opCtx->lockState());
boost::optional<AutoGetCollectionForRead> readLock;
boost::optional<AutoStatsTracker> statsTracker;
CursorManager* cursorManager;
if (CursorManager::isGloballyManagedCursor(cursorid)) {
cursorManager = CursorManager::getGlobalCursorManager();
if (boost::optional<NamespaceString> nssForCurOp = nss.isGloballyManagedNamespace()
? nss.getTargetNSForGloballyManagedNamespace()
: nss) {
AutoGetDb autoDb(opCtx, nssForCurOp->db(), MODE_IS);
const auto profilingLevel = autoDb.getDb()
? boost::optional<int>{autoDb.getDb()->getProfilingLevel()}
: boost::none;
statsTracker.emplace(opCtx, *nssForCurOp, Top::LockType::NotLocked, profilingLevel);
auto view = autoDb.getDb()
? autoDb.getDb()->getViewCatalog()->lookup(opCtx, nssForCurOp->ns())
: nullptr;
uassert(
ErrorCodes::CommandNotSupportedOnView,
str::stream() << "Namespace " << nssForCurOp->ns()
<< " is a view. OP_GET_MORE operations are not supported on views. "
<< "Only clients which support the getMore command can be used to "
"query views.",
!view);
}
} else {
readLock.emplace(opCtx, nss);
const int doNotChangeProfilingLevel = 0;
statsTracker.emplace(opCtx,
nss,
Top::LockType::ReadLocked,
readLock->getDb() ? readLock->getDb()->getProfilingLevel()
: doNotChangeProfilingLevel);
Collection* collection = readLock->getCollection();
uassert(
ErrorCodes::OperationFailed, "collection dropped between getMore calls", collection);
cursorManager = collection->getCursorManager();
// This checks to make sure the operation is allowed on a replicated node. Since we are not
// passing in a query object (necessary to check SlaveOK query option), we allow reads
// whether we are PRIMARY or SECONDARY.
uassertStatusOK(
repl::ReplicationCoordinator::get(opCtx)->checkCanServeReadsFor(opCtx, nss, true));
}
LOG(5) << "Running getMore, cursorid: " << cursorid;
// A pin performs a CC lookup and if there is a CC, increments the CC's pin value so it
// doesn't time out. Also informs ClientCursor that there is somebody actively holding the
// CC, so don't delete it.
auto ccPin = cursorManager->pinCursor(opCtx, cursorid);
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
const int InitialBufSize =
512 + sizeof(QueryResult::Value) + FindCommon::kMaxBytesToReturnToClientAtOnce;
BufBuilder bb(InitialBufSize);
bb.skip(sizeof(QueryResult::Value));
if (!ccPin.isOK()) {
if (ccPin == ErrorCodes::CursorNotFound) {
cursorid = 0;
resultFlags = ResultFlag_CursorNotFound;
} else {
uassertStatusOK(ccPin.getStatus());
}
} else {
ClientCursor* cc = ccPin.getValue().getCursor();
// Check for spoofing of the ns such that it does not match the one originally
// there for the cursor.
uassert(ErrorCodes::Unauthorized,
str::stream() << "Requested getMore on namespace " << ns << ", but cursor "
<< cursorid
<< " belongs to namespace "
<< cc->nss().ns(),
nss == cc->nss());
// A user can only call getMore on their own cursor. If there were multiple users
// authenticated when the cursor was created, then at least one of them must be
// authenticated in order to run getMore on the cursor.
uassert(ErrorCodes::Unauthorized,
str::stream() << "cursor id " << cursorid
<< " was not created by the authenticated user",
AuthorizationSession::get(opCtx->getClient())
->isCoauthorizedWith(cc->getAuthenticatedUsers()));
*isCursorAuthorized = true;
const auto replicationMode = repl::ReplicationCoordinator::get(opCtx)->getReplicationMode();
opCtx->recoveryUnit()->setReadConcernLevelAndReplicationMode(cc->getReadConcernLevel(),
replicationMode);
// TODO SERVER-33698: Remove kSnapshotReadConcern clause once we can guarantee that a
// readConcern level snapshot getMore will have an established point-in-time WiredTiger
// snapshot.
if (replicationMode == repl::ReplicationCoordinator::modeReplSet &&
(cc->getReadConcernLevel() == repl::ReadConcernLevel::kMajorityReadConcern ||
cc->getReadConcernLevel() == repl::ReadConcernLevel::kSnapshotReadConcern)) {
uassertStatusOK(opCtx->recoveryUnit()->obtainMajorityCommittedSnapshot());
}
uassert(40548,
"OP_GET_MORE operations are not supported on tailable aggregations. Only clients "
"which support the getMore command can be used on tailable aggregations.",
readLock || !cc->isAwaitData());
// If the operation that spawned this cursor had a time limit set, apply leftover
// time to this getmore.
if (cc->getLeftoverMaxTimeMicros() < Microseconds::max()) {
uassert(40136,
"Illegal attempt to set operation deadline within DBDirectClient",
!opCtx->getClient()->isInDirectClient());
opCtx->setDeadlineAfterNowBy(cc->getLeftoverMaxTimeMicros());
}
opCtx->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// What number result are we starting at? Used to fill out the reply.
startingResult = cc->pos();
uint64_t notifierVersion = 0;
std::shared_ptr<CappedInsertNotifier> notifier;
if (cc->isAwaitData()) {
invariant(readLock->getCollection()->isCapped());
// 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.
notifier = readLock->getCollection()->getCappedInsertNotifier();
// Must get the version before we call generateBatch in case a write comes in after
// that call and before we call wait on the notifier.
notifierVersion = notifier->getVersion();
}
PlanExecutor* exec = cc->getExecutor();
exec->reattachToOperationContext(opCtx);
uassertStatusOK(exec->restoreState());
auto planSummary = Explain::getPlanSummary(exec);
{
stdx::lock_guard<Client> lk(*opCtx->getClient());
curOp.setPlanSummary_inlock(planSummary);
// Ensure that the original query object is available in the slow query log, profiler
// and currentOp. Upconvert _query to resemble a getMore command, and set the original
// command or upconverted legacy query in the originatingCommand field.
curOp.setOpDescription_inlock(upconvertGetMoreEntry(nss, cursorid, ntoreturn));
curOp.setOriginatingCommand_inlock(cc->getOriginatingCommandObj());
}
PlanExecutor::ExecState state;
// We report keysExamined and docsExamined to OpDebug for a given getMore operation. To
// obtain these values we need to take a diff of the pre-execution and post-execution
// metrics, as they accumulate over the course of a cursor's lifetime.
PlanSummaryStats preExecutionStats;
Explain::getSummaryStats(*exec, &preExecutionStats);
generateBatch(ntoreturn, cc, &bb, &numResults, &state);
// If this is an await data cursor, and we hit EOF without generating any results, then
// we block waiting for new data to arrive.
if (cc->isAwaitData() && state == PlanExecutor::IS_EOF && numResults == 0) {
// Save the PlanExecutor and drop our locks.
exec->saveState();
readLock.reset();
// Block waiting for data for up to 1 second. Time spent blocking is not counted towards
// the total operation latency.
curOp.pauseTimer();
Seconds timeout(1);
notifier->waitUntil(notifierVersion,
opCtx->getServiceContext()->getPreciseClockSource()->now() +
timeout);
notifier.reset();
curOp.resumeTimer();
// Reacquiring locks.
readLock.emplace(opCtx, nss);
uassertStatusOK(exec->restoreState());
// We woke up because either the timed_wait expired, or there was more data. Either
// way, attempt to generate another batch of results.
generateBatch(ntoreturn, cc, &bb, &numResults, &state);
}
PlanSummaryStats postExecutionStats;
Explain::getSummaryStats(*exec, &postExecutionStats);
postExecutionStats.totalKeysExamined -= preExecutionStats.totalKeysExamined;
postExecutionStats.totalDocsExamined -= preExecutionStats.totalDocsExamined;
curOp.debug().setPlanSummaryMetrics(postExecutionStats);
// We do not report 'execStats' for aggregation or other globally managed cursors, both in
// the original request and subsequent getMore. It would be useful to have this information
// for an aggregation, but the source PlanExecutor could be destroyed before we know whether
// we need execStats and we do not want to generate for all operations due to cost.
if (!CursorManager::isGloballyManagedCursor(cursorid) && curOp.shouldDBProfile()) {
BSONObjBuilder execStatsBob;
Explain::getWinningPlanStats(exec, &execStatsBob);
curOp.debug().execStats = execStatsBob.obj();
}
// Our two possible ClientCursorPin cleanup paths are:
// 1) If the cursor is not going to be saved, we call deleteUnderlying() on the pin.
// 2) If the cursor is going to be saved, we simply let the pin go out of scope. In this
// case, the pin's destructor will be invoked, which will call release() on the pin.
// Because our ClientCursorPin is declared after our lock is declared, this will happen
// under the lock if any locking was necessary.
if (!shouldSaveCursorGetMore(state, exec, cc->isTailable())) {
ccPin.getValue().deleteUnderlying();
// cc is now invalid, as is the executor
cursorid = 0;
cc = nullptr;
curOp.debug().cursorExhausted = true;
LOG(5) << "getMore NOT saving client cursor, ended with state "
<< PlanExecutor::statestr(state);
} else {
// Continue caching the ClientCursor.
cc->incPos(numResults);
exec->saveState();
exec->detachFromOperationContext();
LOG(5) << "getMore saving client cursor ended with state "
<< PlanExecutor::statestr(state);
*exhaust = cc->queryOptions() & QueryOption_Exhaust;
// We assume that cursors created through a DBDirectClient are always used from their
// original OperationContext, so we do not need to move time to and from the cursor.
if (!opCtx->getClient()->isInDirectClient()) {
// If the getmore had a time limit, remaining time is "rolled over" back to the
// cursor (for use by future getmore ops).
cc->setLeftoverMaxTimeMicros(opCtx->getRemainingMaxTimeMicros());
}
}
}
QueryResult::View qr = bb.buf();
qr.msgdata().setLen(bb.len());
qr.msgdata().setOperation(opReply);
qr.setResultFlags(resultFlags);
qr.setCursorId(cursorid);
qr.setStartingFrom(startingResult);
qr.setNReturned(numResults);
LOG(5) << "getMore returned " << numResults << " results\n";
return Message(bb.release());
}
/**
* 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;
}
/**
* 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;
}
PlanStage::StageState SubplanStage::work(WorkingSetID* out) {
++_commonStats.works;
// Adds the amount of time taken by work() to executionTimeMillis.
ScopedTimer timer(&_commonStats.executionTimeMillis);
if (_killed) {
return PlanStage::DEAD;
}
if (isEOF()) { return PlanStage::IS_EOF; }
if (SubplanStage::PLANNING == _state) {
// Try to run as sub-plans.
if (runSubplans()) {
// If runSubplans returns true we expect something here.
invariant(_child.get());
}
else if (!_killed) {
// Couldn't run as subplans so we'll just call normal getExecutor.
PlanExecutor* exec;
Status status = getExecutorAlwaysPlan(_collection, _query, _plannerParams, &exec);
if (!status.isOK()) {
// We utterly failed.
_killed = true;
// Propagate the error to the user wrapped in a BSONObj
WorkingSetID id = _ws->allocate();
WorkingSetMember* member = _ws->get(id);
member->state = WorkingSetMember::OWNED_OBJ;
member->keyData.clear();
member->loc = DiskLoc();
BSONObjBuilder bob;
bob.append("ok", status.isOK() ? 1.0 : 0.0);
bob.append("code", status.code());
bob.append("errmsg", status.reason());
member->obj = bob.obj();
*out = id;
return PlanStage::FAILURE;
}
else {
scoped_ptr<PlanExecutor> cleanupExec(exec);
_child.reset(exec->releaseStages());
}
}
// We can change state when we're either killed or we have an underlying runner.
invariant(_killed || NULL != _child.get());
_state = SubplanStage::RUNNING;
}
if (_killed) {
return PlanStage::DEAD;
}
if (isEOF()) {
return PlanStage::IS_EOF;
}
// If we're here we should have planned already.
invariant(SubplanStage::RUNNING == _state);
invariant(_child.get());
return _child->work(out);
}
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;
}
