void DBClientCursor::commandDataReceived() {
int op = batch.m.operation();
invariant(op == opReply || op == dbCommandReply);
batch.nReturned = 1;
batch.pos = 0;
auto commandReply = rpc::makeReply(&batch.m);
auto commandStatus = getStatusFromCommandResult(commandReply->getCommandReply());
if (ErrorCodes::SendStaleConfig == commandStatus) {
throw RecvStaleConfigException("stale config in DBClientCursor::dataReceived()",
commandReply->getCommandReply());
} else if (!commandStatus.isOK()) {
wasError = true;
}
if (_client->getReplyMetadataReader()) {
uassertStatusOK(_client->getReplyMetadataReader()(commandReply->getMetadata(),
_client->getServerAddress()));
}
// HACK: If we got an OP_COMMANDREPLY, take the reply object
// and shove it in to an OP_REPLY message.
if (op == dbCommandReply) {
// Need to take ownership here as we destroy the underlying message.
BSONObj reply = commandReply->getCommandReply().getOwned();
batch.m.reset();
replyToQuery(0, batch.m, reply);
}
QueryResult::View qr = batch.m.singleData().view2ptr();
batch.data = qr.data();
}
/**
* Returns true if request is a query for sharded indexes.
*/
static bool doShardedIndexQuery(OperationContext* txn, Request& r, const QuerySpec& qSpec) {
// Extract the ns field from the query, which may be embedded within the "query" or
// "$query" field.
auto nsField = qSpec.filter()["ns"];
if (nsField.eoo()) {
return false;
}
const NamespaceString indexNSSQuery(nsField.str());
auto status = grid.catalogCache()->getDatabase(txn, indexNSSQuery.db().toString());
if (!status.isOK()) {
return false;
}
shared_ptr<DBConfig> config = status.getValue();
if (!config->isSharded(indexNSSQuery.ns())) {
return false;
}
// if you are querying on system.indexes, we need to make sure we go to a shard
// that actually has chunks. This is not a perfect solution (what if you just
// look at all indexes), but better than doing nothing.
ShardPtr shard;
ChunkManagerPtr cm;
config->getChunkManagerOrPrimary(indexNSSQuery.ns(), cm, shard);
if (cm) {
set<ShardId> shardIds;
cm->getAllShardIds(&shardIds);
verify(shardIds.size() > 0);
shard = grid.shardRegistry()->getShard(*shardIds.begin());
}
ShardConnection dbcon(shard->getConnString(), r.getns());
DBClientBase& c = dbcon.conn();
string actualServer;
Message response;
bool ok = c.call(r.m(), response, true, &actualServer);
uassert(10200, "mongos: error calling db", ok);
{
QueryResult::View qr = response.singleData().view2ptr();
if (qr.getResultFlags() & ResultFlag_ShardConfigStale) {
dbcon.done();
// Version is zero b/c this is deprecated codepath
throw RecvStaleConfigException(r.getns(),
"Strategy::doQuery",
ChunkVersion(0, 0, OID()),
ChunkVersion(0, 0, OID()));
}
}
r.reply(response, actualServer.size() ? actualServer : c.getServerAddress());
dbcon.done();
return true;
}
void DBClientCursor::dataReceived(bool& retry, string& host) {
// If this is a reply to our initial command request.
if (_isCommand && cursorId == 0) {
commandDataReceived();
return;
}
QueryResult::View qr = batch.m.singleData().view2ptr();
resultFlags = qr.getResultFlags();
if (qr.getResultFlags() & ResultFlag_ErrSet) {
wasError = true;
}
if (qr.getResultFlags() & ResultFlag_CursorNotFound) {
// cursor id no longer valid at the server.
invariant(qr.getCursorId() == 0);
if (!(opts & QueryOption_CursorTailable)) {
uasserted(13127,
str::stream() << "cursor id " << cursorId << " didn't exist on server.");
}
// 0 indicates no longer valid (dead)
cursorId = 0;
}
if (cursorId == 0 || !(opts & QueryOption_CursorTailable)) {
// only set initially: we don't want to kill it on end of data
// if it's a tailable cursor
cursorId = qr.getCursorId();
}
batch.nReturned = qr.getNReturned();
batch.pos = 0;
batch.data = qr.data();
batch.remainingBytes = qr.dataLen();
_client->checkResponse(batch.data, batch.nReturned, &retry, &host); // watches for "not master"
if (qr.getResultFlags() & ResultFlag_ShardConfigStale) {
BSONObj error;
verify(peekError(&error));
throw RecvStaleConfigException(
(string) "stale config on lazy receive" + causedBy(getErrField(error)), error);
}
/* this assert would fire the way we currently work:
verify( nReturned || cursorId == 0 );
*/
}
void NetworkInterfaceASIO::_completedWriteCallback(AsyncOp* op) {
// If we were told to send an empty message, toRecv will be empty here.
// TODO: handle metadata SERVER-19156
BSONObj commandReply;
if (op->toRecv()->empty()) {
LOG(3) << "received an empty message";
} else {
QueryResult::View qr = op->toRecv()->singleData().view2ptr();
// unavoidable copy
commandReply = BSONObj(qr.data()).getOwned();
}
_completeOperation(
op, RemoteCommandResponse(std::move(commandReply), BSONObj(), now() - op->start()));
}
std::unique_ptr<Message> LegacyReplyBuilder::done() {
invariant(_state == State::kOutputDocs);
BSONObj reply = uassertStatusOK(rpc::downconvertReplyMetadata(_commandReply, _metadata));
BufBuilder bufBuilder;
bufBuilder.skip(sizeof(QueryResult::Value));
if (_allowAddingOutputDocs) {
BSONObjBuilder topBuilder(bufBuilder);
for (const auto& el : reply) {
if (kCursorTag != el.fieldNameStringData()) {
topBuilder.append(el);
continue;
}
invariant(el.isABSONObj());
BSONObjBuilder curBuilder(topBuilder.subobjStart(kCursorTag));
for (const auto& insideEl : el.Obj()) {
if (kFirstBatchTag != insideEl.fieldNameStringData()) {
curBuilder.append(insideEl);
continue;
}
invariant(insideEl.isABSONObj());
BSONArrayBuilder arrBuilder(curBuilder.subarrayStart(kFirstBatchTag));
for (const auto& doc : _outputDocs) {
arrBuilder.append(doc);
}
arrBuilder.doneFast();
}
curBuilder.doneFast();
}
topBuilder.doneFast();
} else {
reply.appendSelfToBufBuilder(bufBuilder);
}
auto msgHeaderSz = static_cast<std::size_t>(MsgData::MsgDataHeaderSize);
invariant(static_cast<std::size_t>(bufBuilder.len()) + msgHeaderSz <=
mongol::MaxMessageSizeBytes);
QueryResult::View qr = bufBuilder.buf();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bufBuilder.len());
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
_message->setData(qr.view2ptr(), true);
bufBuilder.decouple();
_state = State::kDone;
return std::move(_message);
}
virtual void process(Message& m, AbstractMessagingPort* port) {
while (true) {
if (inShutdown()) {
log() << "got request after shutdown()" << endl;
break;
}
DbResponse dbresponse;
{
auto opCtx = getGlobalServiceContext()->makeOperationContext(&cc());
assembleResponse(opCtx.get(), m, dbresponse, port->remote());
// opCtx must go out of scope here so that the operation cannot show up in currentOp
// results after the response reaches the client
}
if (!dbresponse.response.empty()) {
port->reply(m, dbresponse.response, dbresponse.responseTo);
if (dbresponse.exhaustNS.size() > 0) {
MsgData::View header = dbresponse.response.header();
QueryResult::View qr = header.view2ptr();
long long cursorid = qr.getCursorId();
if (cursorid) {
verify(dbresponse.exhaustNS.size() && dbresponse.exhaustNS[0]);
string ns = dbresponse.exhaustNS; // before reset() free's it...
m.reset();
BufBuilder b(512);
b.appendNum((int)0 /*size set later in appendData()*/);
b.appendNum(header.getId());
b.appendNum(header.getResponseTo());
b.appendNum((int)dbGetMore);
b.appendNum((int)0);
b.appendStr(ns);
b.appendNum((int)0); // ntoreturn
b.appendNum(cursorid);
m.appendData(b.buf(), b.len());
b.decouple();
DEV log() << "exhaust=true sending more";
continue; // this goes back to top loop
}
}
}
break;
}
}
LegacyReply::LegacyReply(const Message* message) : _message(std::move(message)) {
invariant(message->operation() == opReply);
QueryResult::View qr = _message->singleData().view2ptr();
// should be checked by caller.
invariant(qr.msgdata().getNetworkOp() == opReply);
uassert(ErrorCodes::BadValue,
str::stream() << "Got legacy command reply with a bad cursorId field,"
<< " expected a value of 0 but got " << qr.getCursorId(),
qr.getCursorId() == 0);
uassert(ErrorCodes::BadValue,
str::stream() << "Got legacy command reply with a bad nReturned field,"
<< " expected a value of 1 but got " << qr.getNReturned(),
qr.getNReturned() == 1);
uassert(ErrorCodes::BadValue,
str::stream() << "Got legacy command reply with a bad startingFrom field,"
<< " expected a value of 0 but got " << qr.getStartingFrom(),
qr.getStartingFrom() == 0);
std::tie(_commandReply, _metadata) =
uassertStatusOK(rpc::upconvertReplyMetadata(BSONObj(qr.data())));
// Copy the bson array of documents from the message into
// a contiguous area of memory owned by _docBuffer so
// DocumentRange can be used to iterate over documents
auto cursorElem = _commandReply[LegacyReplyBuilder::kCursorTag];
if (cursorElem.eoo())
return;
BSONObj cursorObj = cursorElem.Obj();
auto firstBatchElem = cursorObj[LegacyReplyBuilder::kFirstBatchTag];
if (firstBatchElem.eoo())
return;
for (BSONObjIterator it(firstBatchElem.Obj()); it.more(); it.next()) {
invariant((*it).isABSONObj());
BSONObj doc = (*it).Obj();
doc.appendSelfToBufBuilder(_docBuffer);
}
const char* dataBegin = _docBuffer.buf();
const char* dataEnd = dataBegin + _docBuffer.len();
_outputDocs = DocumentRange(dataBegin, dataEnd);
return;
}
virtual void process( Message& m , AbstractMessagingPort* port , LastError * le) {
OperationContextImpl txn;
while ( true ) {
if ( inShutdown() ) {
log() << "got request after shutdown()" << endl;
break;
}
lastError.startRequest( m , le );
DbResponse dbresponse;
assembleResponse(&txn, m, dbresponse, port->remote());
if ( dbresponse.response ) {
port->reply(m, *dbresponse.response, dbresponse.responseTo);
if( dbresponse.exhaustNS.size() > 0 ) {
MsgData::View header = dbresponse.response->header();
QueryResult::View qr = header.view2ptr();
long long cursorid = qr.getCursorId();
if( cursorid ) {
verify( dbresponse.exhaustNS.size() && dbresponse.exhaustNS[0] );
string ns = dbresponse.exhaustNS; // before reset() free's it...
m.reset();
BufBuilder b(512);
b.appendNum((int) 0 /*size set later in appendData()*/);
b.appendNum(header.getId());
b.appendNum(header.getResponseTo());
b.appendNum((int) dbGetMore);
b.appendNum((int) 0);
b.appendStr(ns);
b.appendNum((int) 0); // ntoreturn
b.appendNum(cursorid);
m.appendData(b.buf(), b.len());
b.decouple();
DEV log() << "exhaust=true sending more" << endl;
continue; // this goes back to top loop
}
}
}
break;
}
}
void replyToQuery( int queryResultFlags, Message& response, const BSONObj& resultObj ) {
BufBuilder bufBuilder;
bufBuilder.skip( sizeof( QueryResult::Value ));
bufBuilder.appendBuf( reinterpret_cast< void *>(
const_cast< char* >( resultObj.objdata() )), resultObj.objsize() );
QueryResult::View queryResult = bufBuilder.buf();
bufBuilder.decouple();
queryResult.setResultFlags(queryResultFlags);
queryResult.msgdata().setLen(bufBuilder.len());
queryResult.msgdata().setOperation( opReply );
queryResult.setCursorId(0);
queryResult.setStartingFrom(0);
queryResult.setNReturned(1);
response.setData( queryResult.view2ptr(), true ); // transport will free
}
Message LegacyReplyBuilder::done() {
invariant(_state == State::kOutputDocs);
QueryResult::View qr = _builder.buf();
qr.setResultFlagsToOk();
qr.msgdata().setLen(_builder.len());
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
_message.setData(qr.view2ptr(), true);
_builder.decouple();
_state = State::kDone;
return std::move(_message);
}
void OpQueryReplyBuilder::putInMessage(
Message* out, int queryResultFlags, int nReturned, int startingFrom, long long cursorId) {
QueryResult::View qr = _buffer.buf();
qr.setResultFlags(queryResultFlags);
qr.msgdata().setLen(_buffer.len());
qr.msgdata().setOperation(opReply);
qr.setCursorId(cursorId);
qr.setStartingFrom(startingFrom);
qr.setNReturned(nReturned);
out->setData(_buffer.release()); // transport will free
}
void DBClientCursor::dataReceived( bool& retry, string& host ) {
QueryResult::View qr = batch.m->singleData().view2ptr();
resultFlags = qr.getResultFlags();
if ( qr.getResultFlags() & ResultFlag_ErrSet ) {
wasError = true;
}
if ( qr.getResultFlags() & ResultFlag_CursorNotFound ) {
// cursor id no longer valid at the server.
verify( qr.getCursorId() == 0 );
cursorId = 0; // 0 indicates no longer valid (dead)
if ( ! ( opts & QueryOption_CursorTailable ) )
throw UserException( 13127 , "getMore: cursor didn't exist on server, possible restart or timeout?" );
}
if ( cursorId == 0 || ! ( opts & QueryOption_CursorTailable ) ) {
// only set initially: we don't want to kill it on end of data
// if it's a tailable cursor
cursorId = qr.getCursorId();
}
batch.nReturned = qr.getNReturned();
batch.pos = 0;
batch.data = qr.data();
_client->checkResponse( batch.data, batch.nReturned, &retry, &host ); // watches for "not master"
if( qr.getResultFlags() & ResultFlag_ShardConfigStale ) {
BSONObj error;
verify( peekError( &error ) );
throw RecvStaleConfigException( (string)"stale config on lazy receive" + causedBy( getErrField( error ) ), error );
}
/* this assert would fire the way we currently work:
verify( nReturned || cursorId == 0 );
*/
}
void replyToQuery(int queryResultFlags,
AbstractMessagingPort* p, Message& requestMsg,
void *data, int size,
int nReturned, int startingFrom,
long long cursorId
) {
BufBuilder b(32768);
b.skip(sizeof(QueryResult::Value));
b.appendBuf(data, size);
QueryResult::View qr = b.buf();
qr.setResultFlags(queryResultFlags);
qr.msgdata().setLen(b.len());
qr.msgdata().setOperation(opReply);
qr.setCursorId(cursorId);
qr.setStartingFrom(startingFrom);
qr.setNReturned(nReturned);
b.decouple();
Message resp(qr.view2ptr(), true);
p->reply(requestMsg, resp, requestMsg.header().getId());
}
Message LegacyReplyBuilder::done() {
invariant(_haveCommandReply);
QueryResult::View qr = _builder.buf();
if (_staleConfigError) {
// For compatibility with legacy mongos, we need to set this result flag on StaleConfig
qr.setResultFlags(ResultFlag_ErrSet | ResultFlag_ShardConfigStale);
} else {
qr.setResultFlagsToOk();
}
qr.msgdata().setLen(_builder.len());
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
_message.setData(_builder.release());
return std::move(_message);
}
std::string runQuery(OperationContext* txn,
QueryMessage& q,
const NamespaceString& nss,
CurOp& curop,
Message &result) {
// Validate the namespace.
uassert(16256, str::stream() << "Invalid ns [" << nss.ns() << "]", nss.isValid());
invariant(!nss.isCommand());
// Set curop information.
beginQueryOp(nss, q.query, q.ntoreturn, q.ntoskip, &curop);
// Parse the qm into a CanonicalQuery.
std::auto_ptr<CanonicalQuery> cq;
{
CanonicalQuery* cqRaw;
Status canonStatus = CanonicalQuery::canonicalize(q,
&cqRaw,
WhereCallbackReal(txn, nss.db()));
if (!canonStatus.isOK()) {
uasserted(17287, str::stream() << "Can't canonicalize query: "
<< canonStatus.toString());
}
cq.reset(cqRaw);
}
invariant(cq.get());
LOG(5) << "Running query:\n" << cq->toString();
LOG(2) << "Running query: " << cq->toStringShort();
// Parse, canonicalize, plan, transcribe, and get a plan executor.
AutoGetCollectionForRead ctx(txn, nss);
Collection* collection = ctx.getCollection();
const int dbProfilingLevel = ctx.getDb() ? ctx.getDb()->getProfilingLevel() :
serverGlobalParams.defaultProfile;
// We have a parsed query. Time to get the execution plan for it.
std::unique_ptr<PlanExecutor> exec;
{
PlanExecutor* rawExec;
Status execStatus = getExecutorFind(txn,
collection,
nss,
cq.release(),
PlanExecutor::YIELD_AUTO,
&rawExec);
uassertStatusOK(execStatus);
exec.reset(rawExec);
}
const LiteParsedQuery& pq = exec->getCanonicalQuery()->getParsed();
// If it's actually an explain, do the explain and return rather than falling through
// to the normal query execution loop.
if (pq.isExplain()) {
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder explainBob;
Explain::explainStages(exec.get(), ExplainCommon::EXEC_ALL_PLANS, &explainBob);
// Add the resulting object to the return buffer.
BSONObj explainObj = explainBob.obj();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
// Set query result fields.
QueryResult::View qr = bb.buf();
bb.decouple();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curop.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(qr.view2ptr(), true);
return "";
}
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState.getVersion(nss.ns());
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
bool slaveOK = pq.isSlaveOk() || pq.hasReadPref();
Status serveReadsStatus = repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(
txn,
nss,
slaveOK);
uassertStatusOK(serveReadsStatus);
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(32768);
bb.skip(sizeof(QueryResult::Value));
// How many results have we obtained from the executor?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
Timestamp slaveReadTill;
BSONObj obj;
PlanExecutor::ExecState state;
// uint64_t numMisplacedDocs = 0;
// Get summary info about which plan the executor is using.
curop.debug().planSummary = Explain::getPlanSummary(exec.get());
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 (pq.isOplogReplay()) {
BSONElement e = obj["ts"];
if (Date == e.type() || bsonTimestamp == e.type()) {
slaveReadTill = e.timestamp();
}
}
if (enoughForFirstBatch(pq, numResults, bb.len())) {
LOG(5) << "Enough for first batch, wantMore=" << pq.wantMore()
<< " numToReturn=" << pq.getNumToReturn()
<< " numResults=" << numResults
<< endl;
break;
}
}
// If we cache the executor later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the executor later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the executor.
exec->deregisterExec();
// Caller expects exceptions thrown in certain cases.
if (PlanExecutor::FAILURE == state) {
scoped_ptr<PlanStageStats> stats(exec->getStats());
error() << "Plan executor error, stats: "
<< Explain::statsToBSON(*stats);
uasserted(17144, "Executor error: " + 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)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(nss.ns(), "version changed during initial query",
shardingVersionAtStart,
shardingState.getVersion(nss.ns()));
}
// Fill out curop based on query results. If we have a cursorid, we will fill out curop with
// this cursorid later.
long long ccId = 0;
if (shouldSaveCursor(txn, collection, state, exec.get())) {
// We won't use the executor until it's getMore'd.
exec->saveState();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc = new ClientCursor(collection->getCursorManager(),
exec.release(),
nss.ns(),
pq.getOptions(),
pq.getFilter());
ccId = cc->cursorid();
if (txn->getClient()->isInDirectClient()) {
cc->setUnownedRecoveryUnit(txn->recoveryUnit());
}
else if (state == PlanExecutor::IS_EOF && pq.isTailable()) {
// Don't stash the RU for tailable cursors at EOF, let them get a new RU on their
// next getMore.
}
else {
// We stash away the RecoveryUnit in the ClientCursor. It's used for subsequent
// getMore requests. The calling OpCtx gets a fresh RecoveryUnit.
txn->recoveryUnit()->abandonSnapshot();
cc->setOwnedRecoveryUnit(txn->releaseRecoveryUnit());
StorageEngine* storageEngine = getGlobalServiceContext()->getGlobalStorageEngine();
invariant(txn->setRecoveryUnit(storageEngine->newRecoveryUnit(),
OperationContext::kNotInUnitOfWork)
== OperationContext::kNotInUnitOfWork);
}
LOG(5) << "caching executor with cursorid " << ccId
<< " after returning " << numResults << " results" << endl;
// TODO document
if (pq.isOplogReplay() && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.isExhaust()) {
curop.debug().exhaust = true;
}
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
endQueryOp(cc->getExecutor(), dbProfilingLevel, numResults, ccId, &curop);
}
else {
LOG(5) << "Not caching executor but returning " << numResults << " results.\n";
endQueryOp(exec.get(), dbProfilingLevel, numResults, ccId, &curop);
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult::View qr = result.header().view2ptr();
qr.setCursorId(ccId);
qr.setResultFlagsToOk();
qr.msgdata().setOperation(opReply);
qr.setStartingFrom(0);
qr.setNReturned(numResults);
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? nss.ns() : "";
}
/**
* 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;
}
LegacyReply::LegacyReply(const Message* message) {
invariant(message->operation() == opReply);
QueryResult::View qr = message->singleData().view2ptr();
// should be checked by caller.
invariant(qr.msgdata().getNetworkOp() == opReply);
uassert(ErrorCodes::BadValue,
str::stream() << "Got legacy command reply with a bad cursorId field,"
<< " expected a value of 0 but got "
<< qr.getCursorId(),
qr.getCursorId() == 0);
uassert(ErrorCodes::BadValue,
str::stream() << "Got legacy command reply with a bad nReturned field,"
<< " expected a value of 1 but got "
<< qr.getNReturned(),
qr.getNReturned() == 1);
uassert(ErrorCodes::BadValue,
str::stream() << "Got legacy command reply with a bad startingFrom field,"
<< " expected a value of 0 but got "
<< qr.getStartingFrom(),
qr.getStartingFrom() == 0);
auto status = Validator<BSONObj>::validateLoad(qr.data(), qr.dataLen());
uassert(ErrorCodes::InvalidBSON,
str::stream() << "Got legacy command reply with invalid BSON in the metadata field"
<< causedBy(status),
status.isOK());
_commandReply = BSONObj(qr.data());
_commandReply.shareOwnershipWith(message->sharedBuffer());
if (_commandReply.firstElementFieldName() == "$err"_sd) {
// Upconvert legacy errors.
BSONObjBuilder bob;
bob.appendAs(_commandReply.firstElement(), "errmsg");
bob.append("ok", 0.0);
if (auto code = _commandReply["code"]) {
bob.append(code);
}
_commandReply = bob.obj();
}
return;
}
/**
* 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());
}
std::string runQuery(OperationContext* opCtx,
QueryMessage& q,
const NamespaceString& nss,
Message& result) {
CurOp& curOp = *CurOp::get(opCtx);
curOp.ensureStarted();
uassert(ErrorCodes::InvalidNamespace,
str::stream() << "Invalid ns [" << nss.ns() << "]",
nss.isValid());
invariant(!nss.isCommand());
// Set CurOp information.
const auto upconvertedQuery = upconvertQueryEntry(q.query, nss, q.ntoreturn, q.ntoskip);
beginQueryOp(opCtx, nss, upconvertedQuery, q.ntoreturn, q.ntoskip);
// Parse the qm into a CanonicalQuery.
const boost::intrusive_ptr<ExpressionContext> expCtx;
auto cq = uassertStatusOKWithContext(
CanonicalQuery::canonicalize(opCtx,
q,
expCtx,
ExtensionsCallbackReal(opCtx, &nss),
MatchExpressionParser::kAllowAllSpecialFeatures),
"Can't canonicalize query");
invariant(cq.get());
LOG(5) << "Running query:\n" << redact(cq->toString());
LOG(2) << "Running query: " << redact(cq->toStringShort());
// Parse, canonicalize, plan, transcribe, and get a plan executor.
AutoGetCollectionForReadCommand ctx(opCtx, nss, AutoGetCollection::ViewMode::kViewsForbidden);
Collection* const collection = ctx.getCollection();
{
const QueryRequest& qr = cq->getQueryRequest();
// Allow the query to run on secondaries if the read preference permits it. If no read
// preference was specified, allow the query to run iff slaveOk has been set.
const bool slaveOK = qr.hasReadPref()
? uassertStatusOK(ReadPreferenceSetting::fromContainingBSON(q.query))
.canRunOnSecondary()
: qr.isSlaveOk();
uassertStatusOK(
repl::ReplicationCoordinator::get(opCtx)->checkCanServeReadsFor(opCtx, nss, slaveOK));
}
// We have a parsed query. Time to get the execution plan for it.
auto exec = uassertStatusOK(getExecutorLegacyFind(opCtx, collection, nss, std::move(cq)));
const QueryRequest& qr = exec->getCanonicalQuery()->getQueryRequest();
// If it's actually an explain, do the explain and return rather than falling through
// to the normal query execution loop.
if (qr.isExplain()) {
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder explainBob;
Explain::explainStages(
exec.get(), collection, ExplainOptions::Verbosity::kExecAllPlans, &explainBob);
// Add the resulting object to the return buffer.
BSONObj explainObj = explainBob.obj();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
// Set query result fields.
QueryResult::View qr = bb.buf();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curOp.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(bb.release());
return "";
}
// Handle query option $maxTimeMS (not used with commands).
if (qr.getMaxTimeMS() > 0) {
uassert(40116,
"Illegal attempt to set operation deadline within DBDirectClient",
!opCtx->getClient()->isInDirectClient());
opCtx->setDeadlineAfterNowBy(Milliseconds{qr.getMaxTimeMS()});
}
opCtx->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(FindCommon::kInitReplyBufferSize);
bb.skip(sizeof(QueryResult::Value));
// How many results have we obtained from the executor?
int numResults = 0;
BSONObj obj;
PlanExecutor::ExecState state;
// Get summary info about which plan the executor is using.
{
stdx::lock_guard<Client> lk(*opCtx->getClient());
curOp.setPlanSummary_inlock(Explain::getPlanSummary(exec.get()));
}
while (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, bb.len())) {
exec->enqueue(obj);
break;
}
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
if (FindCommon::enoughForFirstBatch(qr, numResults)) {
LOG(5) << "Enough for first batch, wantMore=" << qr.wantMore()
<< " ntoreturn=" << qr.getNToReturn().value_or(0)
<< " numResults=" << numResults;
break;
}
}
// Caller expects exceptions thrown in certain cases.
if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
error() << "Plan executor error during find: " << PlanExecutor::statestr(state)
<< ", stats: " << redact(Explain::getWinningPlanStats(exec.get()));
uassertStatusOKWithContext(WorkingSetCommon::getMemberObjectStatus(obj),
"Executor error during OP_QUERY find");
MONGO_UNREACHABLE;
}
// Before saving the cursor, ensure that whatever plan we established happened with the expected
// collection version
auto css = CollectionShardingState::get(opCtx, nss);
css->checkShardVersionOrThrow(opCtx);
// Fill out CurOp based on query results. If we have a cursorid, we will fill out CurOp with
// this cursorid later.
long long ccId = 0;
if (shouldSaveCursor(opCtx, collection, state, exec.get())) {
// We won't use the executor until it's getMore'd.
exec->saveState();
exec->detachFromOperationContext();
// Allocate a new ClientCursor and register it with the cursor manager.
ClientCursorPin pinnedCursor = collection->getCursorManager()->registerCursor(
opCtx,
{std::move(exec),
nss,
AuthorizationSession::get(opCtx->getClient())->getAuthenticatedUserNames(),
opCtx->recoveryUnit()->getReadConcernLevel(),
upconvertedQuery});
ccId = pinnedCursor.getCursor()->cursorid();
LOG(5) << "caching executor with cursorid " << ccId << " after returning " << numResults
<< " results";
// TODO document
if (qr.isExhaust()) {
curOp.debug().exhaust = true;
}
pinnedCursor.getCursor()->setPos(numResults);
// 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 query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
pinnedCursor.getCursor()->setLeftoverMaxTimeMicros(opCtx->getRemainingMaxTimeMicros());
}
endQueryOp(opCtx, collection, *pinnedCursor.getCursor()->getExecutor(), numResults, ccId);
} else {
LOG(5) << "Not caching executor but returning " << numResults << " results.";
endQueryOp(opCtx, collection, *exec, numResults, ccId);
}
// Fill out the output buffer's header.
QueryResult::View queryResultView = bb.buf();
queryResultView.setCursorId(ccId);
queryResultView.setResultFlagsToOk();
queryResultView.msgdata().setLen(bb.len());
queryResultView.msgdata().setOperation(opReply);
queryResultView.setStartingFrom(0);
queryResultView.setNReturned(numResults);
// Add the results from the query into the output buffer.
result.setData(bb.release());
// curOp.debug().exhaust is set above.
return curOp.debug().exhaust ? nss.ns() : "";
}
std::string newRunQuery(OperationContext* txn,
Message& m,
QueryMessage& q,
CurOp& curop,
Message &result,
bool fromDBDirectClient) {
// Validate the namespace.
const char *ns = q.ns;
uassert(16332, "can't have an empty ns", ns[0]);
const NamespaceString nsString(ns);
uassert(16256, str::stream() << "Invalid ns [" << ns << "]", nsString.isValid());
// Set curop information.
curop.debug().ns = ns;
curop.debug().ntoreturn = q.ntoreturn;
curop.debug().query = q.query;
curop.setQuery(q.query);
// If the query is really a command, run it.
if (nsString.isCommand()) {
int nToReturn = q.ntoreturn;
uassert(16979, str::stream() << "bad numberToReturn (" << nToReturn
<< ") for $cmd type ns - can only be 1 or -1",
nToReturn == 1 || nToReturn == -1);
curop.markCommand();
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder cmdResBuf;
if (!runCommands(txn, ns, q.query, curop, bb, cmdResBuf, false, q.queryOptions)) {
uasserted(13530, "bad or malformed command request?");
}
curop.debug().iscommand = true;
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
QueryResult::View qr = bb.buf();
bb.decouple();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curop.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(qr.view2ptr(), true);
return "";
}
const NamespaceString nss(q.ns);
// Parse the qm into a CanonicalQuery.
CanonicalQuery* cq;
Status canonStatus = CanonicalQuery::canonicalize(
q, &cq, WhereCallbackReal(txn, StringData(nss.db())));
if (!canonStatus.isOK()) {
uasserted(17287, str::stream() << "Can't canonicalize query: " << canonStatus.toString());
}
QLOG() << "Running query:\n" << cq->toString();
LOG(2) << "Running query: " << cq->toStringShort();
// Parse, canonicalize, plan, transcribe, and get a plan executor.
PlanExecutor* rawExec = NULL;
// We use this a lot below.
const LiteParsedQuery& pq = cq->getParsed();
AutoGetCollectionForRead ctx(txn, nss);
const int dbProfilingLevel = (ctx.getDb() != NULL) ? ctx.getDb()->getProfilingLevel() :
serverGlobalParams.defaultProfile;
Collection* collection = ctx.getCollection();
// We'll now try to get the query executor that will execute this query for us. There
// are a few cases in which we know upfront which executor we should get and, therefore,
// we shortcut the selection process here.
//
// (a) If the query is over a collection that doesn't exist, we use an EOFStage.
//
// (b) if the query is a replication's initial sync one, we use a specifically designed
// stage that skips extents faster (see details in exec/oplogstart.h).
//
// Otherwise we go through the selection of which executor is most suited to the
// query + run-time context at hand.
Status status = Status::OK();
if (NULL != collection && pq.getOptions().oplogReplay) {
// Takes ownership of 'cq'.
status = getOplogStartHack(txn, collection, cq, &rawExec);
}
else {
size_t options = QueryPlannerParams::DEFAULT;
if (shardingState.needCollectionMetadata(pq.ns())) {
options |= QueryPlannerParams::INCLUDE_SHARD_FILTER;
}
// Takes ownership of 'cq'.
status = getExecutor(txn, collection, cq, PlanExecutor::YIELD_AUTO, &rawExec, options);
}
if (!status.isOK()) {
// NOTE: Do not access cq as getExecutor has deleted it.
uasserted(17007, "Unable to execute query: " + status.reason());
}
verify(NULL != rawExec);
auto_ptr<PlanExecutor> exec(rawExec);
// If it's actually an explain, do the explain and return rather than falling through
// to the normal query execution loop.
if (pq.isExplain()) {
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder explainBob;
Explain::explainStages(exec.get(), ExplainCommon::EXEC_ALL_PLANS, &explainBob);
// Add the resulting object to the return buffer.
BSONObj explainObj = explainBob.obj();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
curop.debug().iscommand = true;
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
// Set query result fields.
QueryResult::View qr = bb.buf();
bb.decouple();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curop.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(qr.view2ptr(), true);
return "";
}
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState.getVersion(cq->ns());
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
bool slaveOK = pq.getOptions().slaveOk || pq.hasReadPref();
status = repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(
txn,
NamespaceString(cq->ns()),
slaveOK);
uassertStatusOK(status);
// If this exists, the collection is sharded.
// If it doesn't exist, we can assume we're not sharded.
// If we're sharded, we might encounter data that is not consistent with our sharding state.
// We must ignore this data.
CollectionMetadataPtr collMetadata;
if (!shardingState.needCollectionMetadata(pq.ns())) {
collMetadata = CollectionMetadataPtr();
}
else {
collMetadata = shardingState.getCollectionMetadata(pq.ns());
}
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(32768);
bb.skip(sizeof(QueryResult::Value));
// How many results have we obtained from the executor?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// Do we save the PlanExecutor in a ClientCursor for getMore calls later?
bool saveClientCursor = false;
BSONObj obj;
PlanExecutor::ExecState state;
// uint64_t numMisplacedDocs = 0;
// Get summary info about which plan the executor is using.
curop.debug().planSummary = Explain::getPlanSummary(exec.get());
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 (pq.getOptions().oplogReplay) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
// TODO: only one type of 2d search doesn't support this. We need a way to pull it out
// of CanonicalQuery. :(
const bool supportsGetMore = true;
if (!supportsGetMore && (enough(pq, numResults)
|| bb.len() >= MaxBytesToReturnToClientAtOnce)) {
break;
}
else if (enoughForFirstBatch(pq, numResults, bb.len())) {
QLOG() << "Enough for first batch, wantMore=" << pq.wantMore()
<< " numToReturn=" << pq.getNumToReturn()
<< " numResults=" << numResults
<< endl;
// If only one result requested assume it's a findOne() and don't save the cursor.
if (pq.wantMore() && 1 != pq.getNumToReturn()) {
QLOG() << " executor EOF=" << exec->isEOF() << endl;
saveClientCursor = !exec->isEOF();
}
break;
}
}
// If we cache the executor later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the executor later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the executor.
exec->deregisterExec();
// Caller expects exceptions thrown in certain cases.
if (PlanExecutor::EXEC_ERROR == state) {
scoped_ptr<PlanStageStats> stats(exec->getStats());
error() << "Plan executor error, stats: "
<< Explain::statsToBSON(*stats);
uasserted(17144, "Executor error: " + WorkingSetCommon::toStatusString(obj));
}
// Why save a dead executor?
if (PlanExecutor::DEAD == state) {
saveClientCursor = false;
}
else if (pq.getOptions().tailable) {
// If we're tailing a capped collection, we don't bother saving the cursor if the
// collection is empty. Otherwise, the semantics of the tailable cursor is that the
// client will keep trying to read from it. So we'll keep it around.
if (collection && collection->numRecords(txn) != 0 && pq.getNumToReturn() != 1) {
saveClientCursor = true;
}
}
// TODO(greg): This will go away soon.
if (!shardingState.getVersion(pq.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(pq.ns(), "version changed during initial query",
shardingVersionAtStart,
shardingState.getVersion(pq.ns()));
}
const logger::LogComponent queryLogComponent = logger::LogComponent::kQuery;
const logger::LogSeverity logLevelOne = logger::LogSeverity::Debug(1);
PlanSummaryStats summaryStats;
Explain::getSummaryStats(exec.get(), &summaryStats);
curop.debug().ntoskip = pq.getSkip();
curop.debug().nreturned = numResults;
curop.debug().scanAndOrder = summaryStats.hasSortStage;
curop.debug().nscanned = summaryStats.totalKeysExamined;
curop.debug().nscannedObjects = summaryStats.totalDocsExamined;
curop.debug().idhack = summaryStats.isIdhack;
// Set debug information for consumption by the profiler.
if (dbProfilingLevel > 0 ||
curop.elapsedMillis() > serverGlobalParams.slowMS ||
logger::globalLogDomain()->shouldLog(queryLogComponent, logLevelOne)) {
// Get BSON stats.
scoped_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.debug().planSummary.empty()) {
BSONObjBuilder bob;
bob.append("summary", curop.debug().planSummary.toString());
curop.debug().execStats.set(bob.done());
}
}
long long ccId = 0;
if (saveClientCursor) {
// We won't use the executor until it's getMore'd.
exec->saveState();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc = new ClientCursor(collection, exec.get(),
cq->getParsed().getOptions().toInt(),
cq->getParsed().getFilter());
ccId = cc->cursorid();
if (fromDBDirectClient) {
cc->setUnownedRecoveryUnit(txn->recoveryUnit());
}
else if (state == PlanExecutor::IS_EOF && pq.getOptions().tailable) {
// Don't stash the RU for tailable cursors at EOF, let them get a new RU on their
// next getMore.
}
else {
// We stash away the RecoveryUnit in the ClientCursor. It's used for subsequent
// getMore requests. The calling OpCtx gets a fresh RecoveryUnit.
cc->setOwnedRecoveryUnit(txn->releaseRecoveryUnit());
StorageEngine* storageEngine = getGlobalEnvironment()->getGlobalStorageEngine();
txn->setRecoveryUnit(storageEngine->newRecoveryUnit(txn));
}
QLOG() << "caching executor with cursorid " << ccId
<< " after returning " << numResults << " results" << endl;
// ClientCursor takes ownership of executor. Release to make sure it's not deleted.
exec.release();
// TODO document
if (pq.getOptions().oplogReplay && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.getOptions().exhaust) {
curop.debug().exhaust = true;
}
// Set attributes for getMore.
cc->setCollMetadata(collMetadata);
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
}
else {
QLOG() << "Not caching executor but returning " << numResults << " results.\n";
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult::View qr = result.header().view2ptr();
qr.setCursorId(ccId);
curop.debug().cursorid = (0 == ccId ? -1 : ccId);
qr.setResultFlagsToOk();
qr.msgdata().setOperation(opReply);
qr.setStartingFrom(0);
qr.setNReturned(numResults);
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? pq.ns() : "";
}
/**
* 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;
}
std::string runQuery(OperationContext* txn,
QueryMessage& q,
const NamespaceString& nss,
Message& result) {
CurOp& curop = *CurOp::get(txn);
uassert(ErrorCodes::InvalidNamespace,
str::stream() << "Invalid ns [" << nss.ns() << "]",
nss.isValid());
invariant(!nss.isCommand());
// Set curop information.
beginQueryOp(txn, nss, q.query, q.ntoreturn, q.ntoskip);
// Parse the qm into a CanonicalQuery.
auto statusWithCQ = CanonicalQuery::canonicalize(q, ExtensionsCallbackReal(txn, &nss));
if (!statusWithCQ.isOK()) {
uasserted(
17287,
str::stream() << "Can't canonicalize query: " << statusWithCQ.getStatus().toString());
}
unique_ptr<CanonicalQuery> cq = std::move(statusWithCQ.getValue());
invariant(cq.get());
LOG(5) << "Running query:\n" << cq->toString();
LOG(2) << "Running query: " << cq->toStringShort();
// Parse, canonicalize, plan, transcribe, and get a plan executor.
AutoGetCollectionForRead ctx(txn, nss);
Collection* collection = ctx.getCollection();
const int dbProfilingLevel =
ctx.getDb() ? ctx.getDb()->getProfilingLevel() : serverGlobalParams.defaultProfile;
// We have a parsed query. Time to get the execution plan for it.
std::unique_ptr<PlanExecutor> exec = uassertStatusOK(
getExecutorFind(txn, collection, nss, std::move(cq), PlanExecutor::YIELD_AUTO));
const LiteParsedQuery& pq = exec->getCanonicalQuery()->getParsed();
// If it's actually an explain, do the explain and return rather than falling through
// to the normal query execution loop.
if (pq.isExplain()) {
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder explainBob;
Explain::explainStages(exec.get(), ExplainCommon::EXEC_ALL_PLANS, &explainBob);
// Add the resulting object to the return buffer.
BSONObj explainObj = explainBob.obj();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
// Set query result fields.
QueryResult::View qr = bb.buf();
bb.decouple();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curop.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(qr.view2ptr(), true);
return "";
}
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
bool slaveOK = pq.isSlaveOk() || pq.hasReadPref();
Status serveReadsStatus =
repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(txn, nss, slaveOK);
uassertStatusOK(serveReadsStatus);
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(FindCommon::kInitReplyBufferSize);
bb.skip(sizeof(QueryResult::Value));
// How many results have we obtained from the executor?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
Timestamp slaveReadTill;
BSONObj obj;
PlanExecutor::ExecState state;
// Get summary info about which plan the executor is using.
{
stdx::lock_guard<Client> lk(*txn->getClient());
curop.setPlanSummary_inlock(Explain::getPlanSummary(exec.get()));
}
while (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, bb.len())) {
exec->enqueue(obj);
break;
}
// 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 (pq.isOplogReplay()) {
BSONElement e = obj["ts"];
if (Date == e.type() || bsonTimestamp == e.type()) {
slaveReadTill = e.timestamp();
}
}
if (FindCommon::enoughForFirstBatch(pq, numResults)) {
LOG(5) << "Enough for first batch, wantMore=" << pq.wantMore()
<< " ntoreturn=" << pq.getNToReturn().value_or(0) << " numResults=" << numResults
<< endl;
break;
}
}
// If we cache the executor later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the executor later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the executor.
exec->deregisterExec();
// Caller expects exceptions thrown in certain cases.
if (PlanExecutor::FAILURE == state || PlanExecutor::DEAD == state) {
error() << "Plan executor error during find: " << PlanExecutor::statestr(state)
<< ", stats: " << Explain::getWinningPlanStats(exec.get());
uasserted(17144, "Executor error: " + 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);
// Fill out curop based on query results. If we have a cursorid, we will fill out curop with
// this cursorid later.
long long ccId = 0;
if (shouldSaveCursor(txn, collection, state, exec.get())) {
// We won't use the executor until it's getMore'd.
exec->saveState();
exec->detachFromOperationContext();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc =
new ClientCursor(collection->getCursorManager(),
exec.release(),
nss.ns(),
txn->recoveryUnit()->isReadingFromMajorityCommittedSnapshot(),
pq.getOptions(),
pq.getFilter());
ccId = cc->cursorid();
LOG(5) << "caching executor with cursorid " << ccId << " after returning " << numResults
<< " results" << endl;
// TODO document
if (pq.isOplogReplay() && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.isExhaust()) {
curop.debug().exhaust = true;
}
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
endQueryOp(txn, collection, *cc->getExecutor(), dbProfilingLevel, numResults, ccId);
} else {
LOG(5) << "Not caching executor but returning " << numResults << " results.\n";
endQueryOp(txn, collection, *exec, dbProfilingLevel, numResults, ccId);
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult::View qr = result.header().view2ptr();
qr.setCursorId(ccId);
qr.setResultFlagsToOk();
qr.msgdata().setOperation(opReply);
qr.setStartingFrom(0);
qr.setNReturned(numResults);
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? nss.ns() : "";
}
void generateLegacyQueryErrorResponse(const AssertionException* exception,
const QueryMessage& queryMessage,
CurOp* curop,
Message* response) {
curop->debug().exceptionInfo = exception->getInfo();
log(LogComponent::kQuery) << "assertion " << exception->toString() << " ns:" << queryMessage.ns
<< " query:" << (queryMessage.query.valid()
? queryMessage.query.toString()
: "query object is corrupt");
if (queryMessage.ntoskip || queryMessage.ntoreturn) {
log(LogComponent::kQuery) << " ntoskip:" << queryMessage.ntoskip
<< " ntoreturn:" << queryMessage.ntoreturn;
}
const SendStaleConfigException* scex = (exception->getCode() == ErrorCodes::SendStaleConfig)
? static_cast<const SendStaleConfigException*>(exception)
: NULL;
BSONObjBuilder err;
exception->getInfo().append(err);
if (scex) {
err.append("ok", 0.0);
err.append("ns", scex->getns());
scex->getVersionReceived().addToBSON(err, "vReceived");
scex->getVersionWanted().addToBSON(err, "vWanted");
}
BSONObj errObj = err.done();
if (scex) {
log(LogComponent::kQuery) << "stale version detected during query over " << queryMessage.ns
<< " : " << errObj;
}
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
bb.appendBuf((void*)errObj.objdata(), errObj.objsize());
// TODO: call replyToQuery() from here instead of this!!! see dbmessage.h
QueryResult::View msgdata = bb.buf();
bb.decouple();
QueryResult::View qr = msgdata;
qr.setResultFlags(ResultFlag_ErrSet);
if (scex)
qr.setResultFlags(qr.getResultFlags() | ResultFlag_ShardConfigStale);
qr.msgdata().setLen(bb.len());
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
response->setData(msgdata.view2ptr(), true);
}
void operator()() const {
DBClientConnection dest;
string errmsg;
Timer connectTimer;
while (!dest.connect(HostAndPort(mongoBridgeGlobalParams.destUri), errmsg)) {
// If we can't connect for the configured timeout, give up
//
if (connectTimer.seconds() >= mongoBridgeGlobalParams.connectTimeoutSec) {
cout << "Unable to establish connection from " << mp_.psock->remoteString()
<< " to " << mongoBridgeGlobalParams.destUri
<< " after " << connectTimer.seconds() << " seconds. Giving up." << endl;
mp_.shutdown();
return;
}
sleepmillis(500);
}
Message m;
while( 1 ) {
try {
m.reset();
if ( !mp_.recv( m ) ) {
cout << "end connection " << mp_.psock->remoteString() << endl;
mp_.shutdown();
break;
}
sleepmillis(mongoBridgeGlobalParams.delay);
int oldId = m.header().getId();
if ( m.operation() == dbQuery || m.operation() == dbMsg || m.operation() == dbGetMore ) {
bool exhaust = false;
if ( m.operation() == dbQuery ) {
DbMessage d( m );
QueryMessage q( d );
exhaust = q.queryOptions & QueryOption_Exhaust;
}
Message response;
dest.port().call( m, response );
// nothing to reply with?
if ( response.empty() ) cleanup(0);
mp_.reply( m, response, oldId );
while ( exhaust ) {
MsgData::View header = response.header();
QueryResult::View qr = header.view2ptr();
if ( qr.getCursorId() ) {
response.reset();
dest.port().recv( response );
mp_.reply( m, response ); // m argument is ignored anyway
}
else {
exhaust = false;
}
}
}
else {
dest.port().say( m, oldId );
}
}
catch ( ... ) {
log() << "caught exception in Forwarder, continuing" << endl;
}
}
}