static bool receivedQuery(Client& c, DbResponse& dbresponse, Message& m ){
bool ok = true;
MSGID responseTo = m.header()->id;
DbMessage d(m);
QueryMessage q(d);
auto_ptr< Message > resp( new Message() );
CurOp& op = *(c.curop());
try {
dbresponse.exhaust = runQuery(m, q, op, *resp);
assert( !resp->empty() );
}
catch ( AssertionException& e ) {
ok = false;
op.debug().str << " exception ";
LOGSOME {
log() << "assertion " << e.toString() << " ns:" << q.ns << " query:" <<
(q.query.valid() ? q.query.toString() : "query object is corrupt") << endl;
if( q.ntoskip || q.ntoreturn )
log() << " ntoskip:" << q.ntoskip << " ntoreturn:" << q.ntoreturn << endl;
}
BSONObjBuilder err;
e.getInfo().append( err );
BSONObj errObj = err.done();
BufBuilder b;
b.skip(sizeof(QueryResult));
b.appendBuf((void*) errObj.objdata(), errObj.objsize());
// todo: call replyToQuery() from here instead of this!!! see dbmessage.h
QueryResult * msgdata = (QueryResult *) b.buf();
b.decouple();
QueryResult *qr = msgdata;
qr->_resultFlags() = ResultFlag_ErrSet;
if ( e.getCode() == StaleConfigInContextCode )
qr->_resultFlags() |= ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
resp.reset( new Message() );
resp->setData( msgdata, true );
}
if ( op.shouldDBProfile( 0 ) ){
op.debug().str << " bytes:" << resp->header()->dataLen();
}
dbresponse.response = resp.release();
dbresponse.responseTo = responseTo;
return ok;
}
bool handlePossibleShardedMessage( Message &m, DbResponse &dbresponse ){
if ( shardConfigServer.empty() ){
return false;
}
int op = m.data->operation();
if ( op < 2000 || op >= 3000 )
return false;
const char *ns = m.data->_data + 4;
string errmsg;
if ( shardVersionOk( ns , errmsg ) ){
return false;
}
log() << "shardVersionOk failed ns:" << ns << " " << errmsg << endl;
if ( doesOpGetAResponse( op ) ){
BufBuilder b( 32768 );
b.skip( sizeof( QueryResult ) );
{
BSONObj obj = BSON( "$err" << errmsg );
b.append( obj.objdata() , obj.objsize() );
}
QueryResult *qr = (QueryResult*)b.buf();
qr->_resultFlags() = QueryResult::ResultFlag_ErrSet | QueryResult::ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation( opReply );
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
b.decouple();
Message * resp = new Message();
resp->setData( qr , true );
dbresponse.response = resp;
dbresponse.responseTo = m.data->id;
return true;
}
OID * clientID = clientServerIds.get();
massert( 10422 , "write with bad shard config and no server id!" , clientID );
log() << "got write with an old config - writing back" << endl;
BSONObjBuilder b;
b.appendBool( "writeBack" , true );
b.append( "ns" , ns );
b.appendBinData( "msg" , m.data->len , bdtCustom , (char*)(m.data) );
log() << "writing back msg with len: " << m.data->len << " op: " << m.data->_operation << endl;
clientQueues[clientID->str()]->push( b.obj() );
return true;
}
void replyToQuery(int queryResultFlags, Message &m, DbResponse &dbresponse, BSONObj obj) {
BufBuilder b;
b.skip(sizeof(QueryResult));
b.appendBuf((void*) obj.objdata(), obj.objsize());
QueryResult* msgdata = (QueryResult *) b.buf();
b.decouple();
QueryResult *qr = msgdata;
qr->_resultFlags() = queryResultFlags;
qr->len = b.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
Message *resp = new Message();
resp->setData(msgdata, true); // transport will free
dbresponse.response = resp;
dbresponse.responseTo = m.header()->id;
}
void replyToQuery( int queryResultFlags, Message& response, const BSONObj& resultObj ) {
BufBuilder bufBuilder;
bufBuilder.skip( sizeof( QueryResult ));
bufBuilder.appendBuf( reinterpret_cast< void *>(
const_cast< char* >( resultObj.objdata() )), resultObj.objsize() );
QueryResult* queryResult = reinterpret_cast< QueryResult* >( bufBuilder.buf() );
bufBuilder.decouple();
queryResult->_resultFlags() = queryResultFlags;
queryResult->len = bufBuilder.len();
queryResult->setOperation( opReply );
queryResult->cursorId = 0;
queryResult->startingFrom = 0;
queryResult->nReturned = 1;
response.setData( queryResult, true ); // transport will free
}
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));
b.appendBuf(data, size);
QueryResult *qr = (QueryResult *) b.buf();
qr->_resultFlags() = queryResultFlags;
qr->len = b.len();
qr->setOperation(opReply);
qr->cursorId = cursorId;
qr->startingFrom = startingFrom;
qr->nReturned = nReturned;
b.decouple();
Message resp(qr, true);
p->reply(requestMsg, resp, requestMsg.header()->id);
}
bool handlePossibleShardedMessage( Message &m, DbResponse* dbresponse ) {
if ( ! shardingState.enabled() )
return false;
int op = m.operation();
if ( op < 2000
|| op >= 3000
|| op == dbGetMore // cursors are weird
)
return false;
DbMessage d(m);
const char *ns = d.getns();
string errmsg;
if ( shardVersionOk( ns , opIsWrite( op ) , errmsg ) ) {
return false;
}
log(1) << "connection meta data too old - will retry ns:(" << ns << ") op:(" << opToString(op) << ") " << errmsg << endl;
if ( doesOpGetAResponse( op ) ) {
assert( dbresponse );
BufBuilder b( 32768 );
b.skip( sizeof( QueryResult ) );
{
BSONObj obj = BSON( "$err" << errmsg );
b.appendBuf( obj.objdata() , obj.objsize() );
}
QueryResult *qr = (QueryResult*)b.buf();
qr->_resultFlags() = ResultFlag_ErrSet | ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation( opReply );
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
b.decouple();
Message * resp = new Message();
resp->setData( qr , true );
dbresponse->response = resp;
dbresponse->responseTo = m.header()->id;
return true;
}
OID writebackID;
writebackID.init();
lastError.getSafe()->writeback( writebackID );
const OID& clientID = ShardedConnectionInfo::get(false)->getID();
massert( 10422 , "write with bad shard config and no server id!" , clientID.isSet() );
log(1) << "got write with an old config - writing back ns: " << ns << endl;
if ( logLevel ) log(1) << debugString( m ) << endl;
BSONObjBuilder b;
b.appendBool( "writeBack" , true );
b.append( "ns" , ns );
b.append( "id" , writebackID );
b.append( "connectionId" , cc().getConnectionId() );
b.appendTimestamp( "version" , shardingState.getVersion( ns ) );
b.appendTimestamp( "yourVersion" , ShardedConnectionInfo::get( true )->getVersion( ns ) );
b.appendBinData( "msg" , m.header()->len , bdtCustom , (char*)(m.singleData()) );
log(2) << "writing back msg with len: " << m.header()->len << " op: " << m.operation() << endl;
writeBackManager.queueWriteBack( clientID.str() , b.obj() );
return true;
}
/**
* Also called by db/ops/query.cpp. This is the new getMore entry point.
*/
QueryResult* newGetMore(const char* ns, int ntoreturn, long long cursorid, CurOp& curop,
int pass, bool& exhaust, bool* isCursorAuthorized) {
exhaust = false;
int bufSize = 512 + sizeof(QueryResult) + MaxBytesToReturnToClientAtOnce;
BufBuilder bb(bufSize);
bb.skip(sizeof(QueryResult));
// This is a read lock. TODO: There is a cursor flag for not needing this. Do we care?
Client::ReadContext ctx(ns);
QLOG() << "running getMore in new system, 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.
replVerifyReadsOk();
// 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(cursorid);
ClientCursor* cc = ccPin.c();
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
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;
// 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());
killCurrentOp.checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// TODO:
// curop.debug().query = BSONForQuery
// curop.setQuery(curop.debug().query);
// TODO: What is pass?
if (0 == pass) { cc->updateSlaveLocation(curop); }
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.
Runner* runner = cc->getRunner();
const int queryOptions = cc->queryOptions();
// Get results out of the runner.
runner->restoreState();
BSONObj obj;
Runner::RunnerState state;
while (Runner::RUNNER_ADVANCED == (state = runner->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;
}
}
if (Runner::RUNNER_EOF == state && 0 == numResults
&& (queryOptions & QueryOption_CursorTailable)
&& (queryOptions & QueryOption_AwaitData) && (pass < 1000)) {
// If the cursor is tailable we don't kill it if it's eof. We let it try to get
// data some # of times first.
return 0;
}
bool saveClientCursor = false;
if (Runner::RUNNER_DEAD == state || Runner::RUNNER_ERROR == state) {
// 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 (Runner::RUNNER_EOF == state) {
// EOF is also end of the line unless it's tailable.
saveClientCursor = queryOptions & QueryOption_CursorTailable;
}
else {
verify(Runner::RUNNER_ADVANCED == state);
saveClientCursor = true;
}
if (!saveClientCursor) {
ccPin.deleteUnderlying();
// cc is now invalid, as is the runner
cursorid = 0;
cc = NULL;
QLOG() << "getMore NOT saving client cursor, ended w/state "
<< Runner::statestr(state)
<< endl;
}
else {
// Continue caching the ClientCursor.
cc->incPos(numResults);
runner->saveState();
QLOG() << "getMore saving client cursor ended w/state "
<< Runner::statestr(state)
<< endl;
// 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* qr = reinterpret_cast<QueryResult*>(bb.buf());
qr->len = bb.len();
qr->setOperation(opReply);
qr->_resultFlags() = resultFlags;
qr->cursorId = cursorid;
qr->startingFrom = startingResult;
qr->nReturned = numResults;
bb.decouple();
QLOG() << "getMore returned " << numResults << " results\n";
return qr;
}
QueryResult* processGetMore(const char* ns,
int ntoreturn,
long long cursorid,
CurOp& curop,
int pass,
bool& exhaust,
bool* isCursorAuthorized ) {
bool hasRunner = false;
// Scoped to kill the pin after seeing if the runner's there.
{
// See if there's a runner. We do this until agg. is behind a Runner instead of a CC.
ClientCursorPin p(cursorid);
ClientCursor *cc = p.c();
if (NULL != cc && NULL != cc->getRunner()) {
hasRunner = true;
}
}
if (hasRunner) {
return newGetMore(ns, ntoreturn, cursorid, curop, pass, exhaust,
isCursorAuthorized);
}
exhaust = false;
int bufSize = 512 + sizeof( QueryResult ) + MaxBytesToReturnToClientAtOnce;
BufBuilder b( bufSize );
b.skip(sizeof(QueryResult));
int resultFlags = ResultFlag_AwaitCapable;
int start = 0;
int n = 0;
scoped_ptr<Client::ReadContext> ctx(new Client::ReadContext(ns));
// call this readlocked so state can't change
replVerifyReadsOk();
ClientCursorPin p(cursorid);
ClientCursor *cc = p.c();
if ( unlikely(!cc) ) {
LOGSOME << "getMore: cursorid not found " << ns << " " << cursorid << endl;
cursorid = 0;
resultFlags = ResultFlag_CursorNotFound;
}
else {
// Some internal users create a ClientCursor with a Runner. Don't crash if this
// happens. Instead, hand them off to the new framework.
if (NULL != cc->getRunner()) {
p.release();
return newGetMore(ns, ntoreturn, cursorid, curop, pass, exhaust, isCursorAuthorized);
}
// check for spoofing of the ns such that it does not match the one originally there for the cursor
uassert(14833, "auth error", str::equals(ns, cc->ns().c_str()));
*isCursorAuthorized = true;
// This must be done after auth check to ensure proper cleanup.
uassert(16951, "failing getmore due to set failpoint",
!MONGO_FAIL_POINT(getMoreError));
// If the operation that spawned this cursor had a time limit set, apply leftover
// time to this getmore.
curop.setMaxTimeMicros( cc->getLeftoverMaxTimeMicros() );
killCurrentOp.checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
if ( pass == 0 )
cc->updateSlaveLocation( curop );
int queryOptions = cc->queryOptions();
curop.debug().query = cc->query();
curop.setQuery( cc->query() );
start = cc->pos();
Cursor *c = cc->c();
if (!c->requiresLock()) {
// make sure it won't be destroyed under us
fassert(16952, !c->shouldDestroyOnNSDeletion());
fassert(16953, !c->supportYields());
ctx.reset(); // unlocks
}
c->recoverFromYield();
DiskLoc last;
// This metadata may be stale, but it's the state of chunking when the cursor was
// created.
CollectionMetadataPtr metadata = cc->getCollMetadata();
KeyPattern keyPattern( metadata ? metadata->getKeyPattern() : BSONObj() );
while ( 1 ) {
if ( !c->ok() ) {
if ( c->tailable() ) {
// when a tailable cursor hits "EOF", ok() goes false, and current() is
// null. however advance() can still be retries as a reactivation attempt.
// when there is new data, it will return true. that's what we are doing
// here.
if ( c->advance() )
continue;
if( n == 0 && (queryOptions & QueryOption_AwaitData) && pass < 1000 ) {
return 0;
}
break;
}
p.release();
bool ok = ClientCursor::erase(cursorid);
verify(ok);
cursorid = 0;
cc = 0;
break;
}
MatchDetails details;
if ( cc->fields && cc->fields->getArrayOpType() == Projection::ARRAY_OP_POSITIONAL ) {
// field projection specified, and contains an array operator
details.requestElemMatchKey();
}
// in some cases (clone collection) there won't be a matcher
if ( !c->currentMatches( &details ) ) {
}
else if ( metadata && !metadata->keyBelongsToMe( extractKey(c, keyPattern ) ) ) {
LOG(2) << "cursor skipping document in un-owned chunk: " << c->current()
<< endl;
}
else {
if( c->getsetdup(c->currLoc()) ) {
//out() << " but it's a dup \n";
}
else {
last = c->currLoc();
n++;
// Fill out the fields requested by the query.
const Projection::KeyOnly *keyFieldsOnly = c->keyFieldsOnly();
if ( keyFieldsOnly ) {
fillQueryResultFromObj( b, 0, keyFieldsOnly->hydrate(
c->currKey() ), &details );
}
else {
DiskLoc loc = c->currLoc();
fillQueryResultFromObj( b, cc->fields.get(), c->current(), &details,
( ( cc->pq.get() && cc->pq->showDiskLoc() ) ? &loc : 0 ) );
}
if ( ( ntoreturn && n >= ntoreturn ) || b.len() > MaxBytesToReturnToClientAtOnce ) {
c->advance();
cc->incPos( n );
break;
}
}
}
c->advance();
if ( ! cc->yieldSometimes( ( c->ok() && c->keyFieldsOnly() ) ?
ClientCursor::DontNeed : ClientCursor::WillNeed ) ) {
ClientCursor::erase(cursorid);
cursorid = 0;
cc = 0;
break;
}
}
if ( cc ) {
if ( c->supportYields() ) {
ClientCursor::YieldData data;
verify( cc->prepareToYield( data ) );
}
else {
cc->c()->noteLocation();
}
cc->storeOpForSlave( last );
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 *qr = (QueryResult *) b.buf();
qr->len = b.len();
qr->setOperation(opReply);
qr->_resultFlags() = resultFlags;
qr->cursorId = cursorid;
qr->startingFrom = start;
qr->nReturned = n;
b.decouple();
return qr;
}
static bool receivedQuery(Client& c, DbResponse& dbresponse, Message& m ) {
bool ok = true;
MSGID responseTo = m.header()->id;
DbMessage d(m);
QueryMessage q(d);
auto_ptr< Message > resp( new Message() );
CurOp& op = *(c.curop());
shared_ptr<AssertionException> ex;
try {
if (!NamespaceString(d.getns()).isCommand()) {
// Auth checking for Commands happens later.
Status status = cc().getAuthorizationManager()->checkAuthForQuery(d.getns());
uassert(16550, status.reason(), status.isOK());
}
dbresponse.exhaustNS = runQuery(m, q, op, *resp);
verify( !resp->empty() );
}
catch ( SendStaleConfigException& e ){
ex.reset( new SendStaleConfigException( e.getns(), e.getInfo().msg, e.getVersionReceived(), e.getVersionWanted() ) );
ok = false;
}
catch ( AssertionException& e ) {
ex.reset( new AssertionException( e.getInfo().msg, e.getCode() ) );
ok = false;
}
if( ex ){
op.debug().exceptionInfo = ex->getInfo();
LOGWITHRATELIMIT {
log() << "assertion " << ex->toString() << " ns:" << q.ns << " query:" <<
(q.query.valid() ? q.query.toString() : "query object is corrupt") << endl;
if( q.ntoskip || q.ntoreturn )
log() << " ntoskip:" << q.ntoskip << " ntoreturn:" << q.ntoreturn << endl;
}
SendStaleConfigException* scex = NULL;
if ( ex->getCode() == SendStaleConfigCode ) scex = static_cast<SendStaleConfigException*>( ex.get() );
BSONObjBuilder err;
ex->getInfo().append( err );
if( scex ){
err.append( "ns", scex->getns() );
scex->getVersionReceived().addToBSON( err, "vReceived" );
scex->getVersionWanted().addToBSON( err, "vWanted" );
}
BSONObj errObj = err.done();
if( scex ){
log() << "stale version detected during query over "
<< q.ns << " : " << errObj << endl;
}
else{
log() << "problem detected during query over "
<< q.ns << " : " << errObj << endl;
}
BufBuilder b;
b.skip(sizeof(QueryResult));
b.appendBuf((void*) errObj.objdata(), errObj.objsize());
// todo: call replyToQuery() from here instead of this!!! see dbmessage.h
QueryResult * msgdata = (QueryResult *) b.buf();
b.decouple();
QueryResult *qr = msgdata;
qr->_resultFlags() = ResultFlag_ErrSet;
if( scex ) qr->_resultFlags() |= ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
resp.reset( new Message() );
resp->setData( msgdata, true );
}
op.debug().responseLength = resp->header()->dataLen();
dbresponse.response = resp.release();
dbresponse.responseTo = responseTo;
return ok;
}
bool _handlePossibleShardedMessage( Message &m, DbResponse* dbresponse ) {
DEV assert( shardingState.enabled() );
int op = m.operation();
if ( op < 2000
|| op >= 3000
|| op == dbGetMore // cursors are weird
)
return false;
DbMessage d(m);
const char *ns = d.getns();
string errmsg;
// We don't care about the version here, since we're returning it later in the writeback
ConfigVersion received, wanted;
if ( shardVersionOk( ns , errmsg, received, wanted ) ) {
return false;
}
LOG(1) << "connection meta data too old - will retry ns:(" << ns << ") op:(" << opToString(op) << ") " << errmsg << endl;
if ( doesOpGetAResponse( op ) ) {
assert( dbresponse );
BufBuilder b( 32768 );
b.skip( sizeof( QueryResult ) );
{
BSONObj obj = BSON( "$err" << errmsg << "ns" << ns );
b.appendBuf( obj.objdata() , obj.objsize() );
}
QueryResult *qr = (QueryResult*)b.buf();
qr->_resultFlags() = ResultFlag_ErrSet | ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation( opReply );
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
b.decouple();
Message * resp = new Message();
resp->setData( qr , true );
dbresponse->response = resp;
dbresponse->responseTo = m.header()->id;
return true;
}
uassert( 9517 , "writeback" , ( d.reservedField() & DbMessage::Reserved_FromWriteback ) == 0 );
OID writebackID;
writebackID.init();
lastError.getSafe()->writeback( writebackID );
const OID& clientID = ShardedConnectionInfo::get(false)->getID();
massert( 10422 , "write with bad shard config and no server id!" , clientID.isSet() );
LOG(1) << "got write with an old config - writing back ns: " << ns << endl;
LOG(1) << m.toString() << endl;
BSONObjBuilder b;
b.appendBool( "writeBack" , true );
b.append( "ns" , ns );
b.append( "id" , writebackID );
b.append( "connectionId" , cc().getConnectionId() );
b.append( "instanceIdent" , prettyHostName() );
b.appendTimestamp( "version" , shardingState.getVersion( ns ) );
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
b.appendTimestamp( "yourVersion" , info ? info->getVersion(ns) : (ConfigVersion)0 );
b.appendBinData( "msg" , m.header()->len , bdtCustom , (char*)(m.singleData()) );
LOG(2) << "writing back msg with len: " << m.header()->len << " op: " << m.operation() << endl;
writeBackManager.queueWriteBack( clientID.str() , b.obj() );
return true;
}
static bool receivedQuery(Client& c, DbResponse& dbresponse, Message& m ) {
bool ok = true;
MSGID responseTo = m.header()->id;
DbMessage d(m);
QueryMessage q(d);
auto_ptr< Message > resp( new Message() );
CurOp& op = *(c.curop());
shared_ptr<AssertionException> ex;
try {
dbresponse.exhaust = runQuery(m, q, op, *resp);
assert( !resp->empty() );
}
catch ( SendStaleConfigException& e ){
ex.reset( new SendStaleConfigException( e.getns(), e.getInfo().msg ) );
ok = false;
}
catch ( AssertionException& e ) {
ex.reset( new AssertionException( e.getInfo().msg, e.getCode() ) );
ok = false;
}
if( ex ){
op.debug().exceptionInfo = ex->getInfo();
LOGWITHRATELIMIT {
log() << "assertion " << ex->toString() << " ns:" << q.ns << " query:" <<
(q.query.valid() ? q.query.toString() : "query object is corrupt") << endl;
if( q.ntoskip || q.ntoreturn )
log() << " ntoskip:" << q.ntoskip << " ntoreturn:" << q.ntoreturn << endl;
}
SendStaleConfigException* scex = NULL;
if ( ex->getCode() == SendStaleConfigCode ) scex = static_cast<SendStaleConfigException*>( ex.get() );
BSONObjBuilder err;
ex->getInfo().append( err );
if( scex ) err.append( "ns", scex->getns() );
BSONObj errObj = err.done();
log() << errObj << endl;
BufBuilder b;
b.skip(sizeof(QueryResult));
b.appendBuf((void*) errObj.objdata(), errObj.objsize());
// todo: call replyToQuery() from here instead of this!!! see dbmessage.h
QueryResult * msgdata = (QueryResult *) b.buf();
b.decouple();
QueryResult *qr = msgdata;
qr->_resultFlags() = ResultFlag_ErrSet;
if( scex ) qr->_resultFlags() |= ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
resp.reset( new Message() );
resp->setData( msgdata, true );
}
op.debug().responseLength = resp->header()->dataLen();
dbresponse.response = resp.release();
dbresponse.responseTo = responseTo;
return ok;
}
bool _handlePossibleShardedMessage( Message &m, DbResponse* dbresponse ) {
DEV verify( shardingState.enabled() );
int op = m.operation();
if ( op < 2000
|| op >= 3000
|| op == dbGetMore // cursors are weird
)
return false;
DbMessage d(m);
const char *ns = d.getns();
string errmsg;
// We don't care about the version here, since we're returning it later in the writeback
ChunkVersion received, wanted;
if ( shardVersionOk( ns , errmsg, received, wanted ) ) {
return false;
}
bool getsAResponse = doesOpGetAResponse( op );
LOG(1) << "connection sharding metadata does not match for collection " << ns
<< ", will retry (wanted : " << wanted << ", received : " << received << ")"
<< ( getsAResponse ? "" : " (queuing writeback)" ) << endl;
if( getsAResponse ){
verify( dbresponse );
BufBuilder b( 32768 );
b.skip( sizeof( QueryResult ) );
{
BSONObjBuilder bob;
bob.append( "$err", errmsg );
bob.append( "ns", ns );
wanted.addToBSON( bob, "vWanted" );
received.addToBSON( bob, "vReceived" );
BSONObj obj = bob.obj();
b.appendBuf( obj.objdata() , obj.objsize() );
}
QueryResult *qr = (QueryResult*)b.buf();
qr->_resultFlags() = ResultFlag_ErrSet | ResultFlag_ShardConfigStale;
qr->len = b.len();
qr->setOperation( opReply );
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
b.decouple();
Message * resp = new Message();
resp->setData( qr , true );
dbresponse->response = resp;
dbresponse->responseTo = m.header()->id;
return true;
}
uassert(9517, "cannot queue a writeback operation to the writeback queue",
(d.reservedField() & Reserved_FromWriteback) == 0);
const OID& clientID = ShardedConnectionInfo::get(false)->getID();
massert( 10422 , "write with bad shard config and no server id!" , clientID.isSet() );
// We need to check this here, since otherwise we'll get errors wrapping the writeback -
// not just here, but also when returning as a command result.
// We choose 1/2 the overhead of the internal maximum so that we can still handle ops of
// 16MB exactly.
massert( 16437, "data size of operation is too large to queue for writeback",
m.dataSize() < BSONObjMaxInternalSize - (8 * 1024));
LOG(1) << "writeback queued for " << m.toString() << endl;
BSONObjBuilder b;
b.appendBool( "writeBack" , true );
b.append( "ns" , ns );
b.append( "connectionId" , cc().getConnectionId() );
b.append( "instanceIdent" , prettyHostName() );
wanted.addToBSON( b );
received.addToBSON( b, "yourVersion" );
b.appendBinData( "msg" , m.header()->len , bdtCustom , (char*)(m.singleData()) );
LOG(2) << "writing back msg with len: " << m.header()->len << " op: " << m.operation() << endl;
// we pass the builder to queueWriteBack so that it can select the writebackId
// this is important so that the id is guaranteed to be ascending
// that is important since mongos assumes if its seen a greater writeback
// that all former have been processed
OID writebackID = writeBackManager.queueWriteBack( clientID.str() , b );
lastError.getSafe()->writeback( writebackID );
return true;
}
/**
* Also called by db/ops/query.cpp. This is the new getMore entry point.
*/
QueryResult* newGetMore(const char* ns, int ntoreturn, long long cursorid, CurOp& curop,
int pass, bool& exhaust, bool* isCursorAuthorized) {
exhaust = false;
int bufSize = 512 + sizeof(QueryResult) + MaxBytesToReturnToClientAtOnce;
BufBuilder bb(bufSize);
bb.skip(sizeof(QueryResult));
// This is a read lock. TODO: There is a cursor flag for not needing this. Do we care?
Client::ReadContext ctx(ns);
log() << "running getMore in new system, cursorid " << cursorid << endl;
// TODO: Document.
// TODO: do this when we can pass in our own parsed query
//replVerifyReadsOk();
// 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(cursorid);
ClientCursor* cc = ccPin.c();
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
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;
// 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());
// TODO:
// curop.debug().query = BSONForQuery
// curop.setQuery(curop.debug().query);
// TODO: What is pass?
if (0 == pass) { cc->updateSlaveLocation(curop); }
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.
Runner* runner = cc->getRunner();
const int queryOptions = cc->queryOptions();
// Get results out of the runner.
// TODO: There may be special handling required for tailable cursors?
runner->restoreState();
BSONObj obj;
Runner::RunnerState state;
while (Runner::RUNNER_ADVANCED == (state = runner->getNext(&obj, NULL))) {
// If we're sharded make sure that we don't return any data that hasn't been
// migrated off of our shard yet.
if (collMetadata) {
KeyPattern kp(collMetadata->getKeyPattern());
if (!collMetadata->keyBelongsToMe(kp.extractSingleKey(obj))) { continue; }
}
// 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 ((numResults && numResults >= ntoreturn)
|| bb.len() > MaxBytesToReturnToClientAtOnce) {
break;
}
}
if (Runner::RUNNER_DEAD == state || Runner::RUNNER_EOF == state) {
log() << "getMore(): runner with id " << cursorid << " EOF/DEAD, state = "
<< static_cast<int>(state) << endl;
// TODO: If the cursor is tailable we don't kill it if it's eof.
ccPin.free();
// cc is now invalid, as is the runner
cursorid = 0;
cc = NULL;
}
else {
// Continue caching the ClientCursor.
cc->incPos(numResults);
runner->saveState();
// 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* qr = reinterpret_cast<QueryResult*>(bb.buf());
qr->len = bb.len();
qr->setOperation(opReply);
qr->_resultFlags() = resultFlags;
qr->cursorId = cursorid;
qr->startingFrom = startingResult;
qr->nReturned = numResults;
bb.decouple();
return qr;
}
QueryResult* processGetMore(const char* ns,
int ntoreturn,
long long cursorid,
CurOp& curop,
int pass,
bool& exhaust,
bool* isCursorAuthorized ) {
exhaust = false;
ClientCursor::Pin p(cursorid);
ClientCursor *client_cursor = p.c();
int bufSize = 512 + sizeof( QueryResult ) + MaxBytesToReturnToClientAtOnce;
BufBuilder b( bufSize );
b.skip(sizeof(QueryResult));
int resultFlags = ResultFlag_AwaitCapable;
int start = 0;
int n = 0;
if ( unlikely(!client_cursor) ) {
LOGSOME << "getMore: cursorid not found " << ns << " " << cursorid << endl;
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(14833, "auth error", str::equals(ns, client_cursor->ns().c_str()));
uassert(16784, "oplog cursor reading data that is too old", !client_cursor->lastOpForSlaveTooOld());
int queryOptions = client_cursor->queryOptions();
OpSettings settings;
settings.setBulkFetch(true);
settings.setQueryCursorMode(DEFAULT_LOCK_CURSOR);
settings.setCappedAppendPK(queryOptions & QueryOption_AddHiddenPK);
cc().setOpSettings(settings);
// Check if the cursor is part of a multi-statement transaction. If it is
// and this is not the right client (meaning the current transaction stack
// does not match that in the cursor), it will uassert. If the cursor is
// not part of a multi-statement transaction, then we need to use the stack
// in the cursor for this scope.
const bool cursorPartOfMultiStatementTxn = client_cursor->checkMultiStatementTxn();
scoped_ptr<Client::WithTxnStack> wts;
if (!cursorPartOfMultiStatementTxn) {
// For simplicity, prevent multi-statement transactions from
// reading cursors it didn't create.
uassert(16813, "Cannot getMore() on a cursor not created by this multi-statement transaction",
!cc().hasTxn());
wts.reset(new Client::WithTxnStack(client_cursor->transactions));
}
*isCursorAuthorized = true;
if (pass == 0) {
client_cursor->updateSlaveLocation( curop );
}
curop.debug().query = client_cursor->query();
start = client_cursor->pos();
Cursor *c = client_cursor->c();
// This manager may be stale, but it's the state of chunking when the cursor was created.
ShardChunkManagerPtr manager = client_cursor->getChunkManager();
while ( 1 ) {
if ( !c->ok() ) {
if ( c->tailable() ) {
/* when a tailable cursor hits "EOF", ok() goes false, and current() is null. however
advance() can still be retries as a reactivation attempt. when there is new data, it will
return true. that's what we are doing here.
*/
if ( c->advance() )
continue;
if( n == 0 && (queryOptions & QueryOption_AwaitData) && pass < 1000 ) {
return 0;
}
break;
}
p.release();
// Done with this cursor, steal transaction stack back to commit or abort it here.
bool ok = ClientCursor::erase(cursorid);
verify(ok);
cursorid = 0;
client_cursor = 0;
break;
}
MatchDetails details;
if ( client_cursor->fields && client_cursor->fields->getArrayOpType() == Projection::ARRAY_OP_POSITIONAL ) {
// field projection specified, and contains an array operator
details.requestElemMatchKey();
}
// in some cases (clone collection) there won't be a matcher
if ( !c->currentMatches( &details ) ) {
}
else if ( manager && ! manager->belongsToMe( client_cursor ) ){
LOG(2) << "cursor skipping document in un-owned chunk: " << c->current() << endl;
}
else {
if( c->getsetdup(c->currPK()) ) {
//out() << " but it's a dup \n";
}
else {
// save this so that at the end of the loop,
// we can update the location for write concern
// in replication. Note that if this cursor is not
// doing replication, this is pointless
if ( client_cursor->queryOptions() & QueryOption_OplogReplay ) {
client_cursor->storeOpForSlave( c->current() );
}
n++;
client_cursor->fillQueryResultFromObj( b, &details );
if ( ( ntoreturn && n >= ntoreturn ) || b.len() > MaxBytesToReturnToClientAtOnce ) {
c->advance();
client_cursor->incPos( n );
break;
}
}
}
c->advance();
}
if ( client_cursor ) {
client_cursor->resetIdleAge();
exhaust = client_cursor->queryOptions() & QueryOption_Exhaust;
} else if (!cursorPartOfMultiStatementTxn) {
// This cursor is done and it wasn't part of a multi-statement
// transaction. We can commit the transaction now.
cc().commitTopTxn();
wts->release();
}
}
QueryResult *qr = (QueryResult *) b.buf();
qr->len = b.len();
qr->setOperation(opReply);
qr->_resultFlags() = resultFlags;
qr->cursorId = cursorid;
qr->startingFrom = start;
qr->nReturned = n;
b.decouple();
return qr;
}
/**
* Also called by db/ops/query.cpp. This is the new getMore entry point.
*/
QueryResult* newGetMore(const char* ns, int ntoreturn, long long cursorid, CurOp& curop,
int pass, bool& exhaust, bool* isCursorAuthorized) {
exhaust = false;
int bufSize = 512 + sizeof(QueryResult) + MaxBytesToReturnToClientAtOnce;
BufBuilder bb(bufSize);
bb.skip(sizeof(QueryResult));
// This is a read lock. TODO: There is a cursor flag for not needing this. Do we care?
Client::ReadContext ctx(ns);
// TODO: Document.
replVerifyReadsOk();
ClientCursorPin ccPin(cursorid);
ClientCursor* cc = ccPin.c();
// These are set in the QueryResult msg we return.
int resultFlags = ResultFlag_AwaitCapable;
int numResults = 0;
int startingResult = 0;
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;
// 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());
// TODO:
// curop.debug().query = BSONForQuery
// curop.setQuery(curop.debug().query);
// TODO: What is pass?
if (0 == pass) { cc->updateSlaveLocation(curop); }
CollectionMetadataPtr collMetadata = cc->getCollMetadata();
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
startingResult = cc->pos();
Runner* runner = cc->getRunner();
const ParsedQuery& pq = runner->getQuery().getParsed();
// Get results out of the runner.
// TODO: There may be special handling required for tailable cursors?
runner->restoreState();
BSONObj obj;
// TODO: Differentiate EOF from error.
while (runner->getNext(&obj)) {
// If we're sharded make sure that we don't return any data that hasn't been
// migrated off of our shard yet.
if (collMetadata) {
KeyPattern kp(collMetadata->getKeyPattern());
if (!collMetadata->keyBelongsToMe(kp.extractSingleKey(obj))) { continue; }
}
// 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.hasOption(QueryOption_OplogReplay)) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
if ((numResults && numResults >= ntoreturn)
|| bb.len() > MaxBytesToReturnToClientAtOnce) {
break;
}
}
cc->incPos(numResults);
runner->saveState();
// Possibly note slave's position in the oplog.
if (pq.hasOption(QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
exhaust = pq.hasOption(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* qr = reinterpret_cast<QueryResult*>(bb.buf());
qr->len = bb.len();
qr->setOperation(opReply);
qr->_resultFlags() = resultFlags;
qr->cursorId = cursorid;
qr->startingFrom = startingResult;
qr->nReturned = numResults;
bb.decouple();
return qr;
}
QueryResult* processGetMore(const char *ns, int ntoreturn, long long cursorid , CurOp& curop, int pass, bool& exhaust ) {
exhaust = false;
ClientCursor::Pointer p(cursorid);
ClientCursor *cc = p.c();
int bufSize = 512 + sizeof( QueryResult ) + MaxBytesToReturnToClientAtOnce;
BufBuilder b( bufSize );
b.skip(sizeof(QueryResult));
int resultFlags = ResultFlag_AwaitCapable;
int start = 0;
int n = 0;
if ( unlikely(!cc) ) {
LOGSOME << "getMore: cursorid not found " << ns << " " << cursorid << endl;
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(14833, "auth error", str::equals(ns, cc->ns().c_str()));
if ( pass == 0 )
cc->updateSlaveLocation( curop );
int queryOptions = cc->queryOptions();
curop.debug().query = cc->query();
start = cc->pos();
Cursor *c = cc->c();
c->recoverFromYield();
DiskLoc last;
scoped_ptr<Projection::KeyOnly> keyFieldsOnly;
if ( cc->modifiedKeys() == false && cc->isMultiKey() == false && cc->fields )
keyFieldsOnly.reset( cc->fields->checkKey( cc->indexKeyPattern() ) );
// This manager may be stale, but it's the state of chunking when the cursor was created.
ShardChunkManagerPtr manager = cc->getChunkManager();
while ( 1 ) {
if ( !c->ok() ) {
if ( c->tailable() ) {
/* when a tailable cursor hits "EOF", ok() goes false, and current() is null. however
advance() can still be retries as a reactivation attempt. when there is new data, it will
return true. that's what we are doing here.
*/
if ( c->advance() )
continue;
if( n == 0 && (queryOptions & QueryOption_AwaitData) && pass < 1000 ) {
return 0;
}
break;
}
p.release();
bool ok = ClientCursor::erase(cursorid);
verify(ok);
cursorid = 0;
cc = 0;
break;
}
// in some cases (clone collection) there won't be a matcher
if ( !c->currentMatches() ) {
}
else if ( manager && ! manager->belongsToMe( cc ) ){
LOG(2) << "cursor skipping document in un-owned chunk: " << c->current() << endl;
}
else {
if( c->getsetdup(c->currLoc()) ) {
//out() << " but it's a dup \n";
}
else {
last = c->currLoc();
n++;
if ( keyFieldsOnly ) {
fillQueryResultFromObj(b, 0, keyFieldsOnly->hydrate( c->currKey() ) );
}
else {
BSONObj js = c->current();
// show disk loc should be part of the main query, not in an $or clause, so this should be ok
fillQueryResultFromObj(b, cc->fields.get(), js, ( cc->pq.get() && cc->pq->showDiskLoc() ? &last : 0));
}
if ( ( ntoreturn && n >= ntoreturn ) || b.len() > MaxBytesToReturnToClientAtOnce ) {
c->advance();
cc->incPos( n );
break;
}
}
}
c->advance();
if ( ! cc->yieldSometimes( ClientCursor::MaybeCovered ) ) {
ClientCursor::erase(cursorid);
cursorid = 0;
cc = 0;
p.deleted();
break;
}
}
if ( cc ) {
if ( c->supportYields() ) {
ClientCursor::YieldData data;
verify( cc->prepareToYield( data ) );
}
else {
cc->c()->noteLocation();
}
cc->mayUpgradeStorage();
cc->storeOpForSlave( last );
exhaust = cc->queryOptions() & QueryOption_Exhaust;
}
}
QueryResult *qr = (QueryResult *) b.buf();
qr->len = b.len();
qr->setOperation(opReply);
qr->_resultFlags() = resultFlags;
qr->cursorId = cursorid;
qr->startingFrom = start;
qr->nReturned = n;
b.decouple();
return qr;
}