const ConfigVersion ShardingState::getVersion( const string& ns ) const {
scoped_lock lk(_mutex);
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if ( it != _chunks.end() ) {
ShardChunkManagerPtr p = it->second;
return p->getVersion();
} else {
return 0;
}
}
bool ShardingState::hasVersion( const string& ns , ConfigVersion& version ) {
scoped_lock lk(_mutex);
ChunkManagersMap::const_iterator it = _chunks.find(ns);
if ( it == _chunks.end() )
return false;
ShardChunkManagerPtr p = it->second;
version = p->getVersion();
return true;
}
void ShardingState::donateChunk( const string& ns , const BSONObj& min , const BSONObj& max , ShardChunkVersion version ) {
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
assert( it != _chunks.end() ) ;
ShardChunkManagerPtr p = it->second;
// empty shards should have version 0
version = ( p->getNumChunks() > 1 ) ? version : ShardChunkVersion( 0 , 0 );
ShardChunkManagerPtr cloned( p->cloneMinus( min , max , version ) );
_chunks[ns] = cloned;
}
void ShardingState::donateChunk( const string& ns , const BSONObj& min , const BSONObj& max , ChunkVersion version ) {
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
verify( it != _chunks.end() ) ;
ShardChunkManagerPtr p = it->second;
// empty shards should have version 0
version = ( p->getNumChunks() > 1 ) ? version : ChunkVersion( 0 , OID() );
ShardChunkManagerPtr cloned( p->cloneMinus( min , max , version ) );
// TODO: a bit dangerous to have two different zero-version states - no-manager and
// no-version
_chunks[ns] = cloned;
}
bool _tryQueryByPKHack(const char *ns, const BSONObj &query,
const ParsedQuery &pq, CurOp &curop, Message &result) {
BSONObj resObject;
bool found = false;
Collection *cl = getCollection(ns);
if (cl == NULL) {
return false; // ns doesn't exist, fall through to optimizer for legacy reasons
}
const BSONObj &pk = cl->getSimplePKFromQuery(query);
if (pk.isEmpty()) {
return false; // unable to query by PK - resort to using the optimizer
}
found = queryByPKHack(cl, pk, query, resObject);
if ( shardingState.needShardChunkManager( ns ) ) {
ShardChunkManagerPtr m = shardingState.getShardChunkManager( ns );
if ( m && ! m->belongsToMe( resObject ) ) {
// I have something for this _id
// but it doesn't belong to me
// so return nothing
resObject = BSONObj();
found = false;
}
}
BufBuilder bb(sizeof(QueryResult)+resObject.objsize()+32);
bb.skip(sizeof(QueryResult));
if ( found ) {
fillQueryResultFromObj( bb , pq.getFields() , resObject );
}
auto_ptr< QueryResult > qr( (QueryResult *) bb.buf() );
bb.decouple();
qr->setResultFlagsToOk();
qr->len = bb.len();
curop.debug().responseLength = bb.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = found ? 1 : 0;
result.setData( qr.release(), true );
return true;
}
void ShardingState::appendInfo( BSONObjBuilder& b ) {
b.appendBool( "enabled" , _enabled );
if ( ! _enabled )
return;
b.append( "configServer" , _configServer );
b.append( "shardName" , _shardName );
b.append( "shardHost" , _shardHost );
{
BSONObjBuilder bb( b.subobjStart( "versions" ) );
scoped_lock lk(_mutex);
for ( ChunkManagersMap::iterator it = _chunks.begin(); it != _chunks.end(); ++it ) {
ShardChunkManagerPtr p = it->second;
bb.appendTimestamp( it->first , p->getVersion() );
}
bb.done();
}
}
ShardChunkManagerPtr ShardingState::getShardChunkManager( const string& ns ){
ConfigVersion version;
{
// check cache
scoped_lock lk( _mutex );
NSVersionMap::const_iterator it = _versions.find( ns );
if ( it == _versions.end() ) {
return ShardChunkManagerPtr();
}
version = it->second;
// TODO SERVER-1849 pending drop work
// the manager should use the cached version only if the versions match exactly
ShardChunkManagerPtr p = _chunks[ns];
if ( p && p->getVersion() >= version ){
// our cached version is good, so just return
return p;
}
}
// load the chunk information for this shard from the config database
// a reminder: ShardChunkManager may throw on construction
const string c = (_configServer == _shardHost) ? "" /* local */ : _configServer;
ShardChunkManagerPtr p( new ShardChunkManager( c , ns , _shardName ) );
// TODO SERVER-1849 verify that the manager's version is exactly the one requested
// If not, do update _chunks, but fail the request.
{
scoped_lock lk( _mutex );
_chunks[ns] = p;
}
return p;
}
bool ShardingState::trySetVersion( const string& ns , ConfigVersion& version /* IN-OUT */ ) {
// Currently this function is called after a getVersion(), which is the first "check", and the assumption here
// is that we don't do anything nearly as long as a remote query in a thread between then and now.
// Otherwise it may be worth adding an additional check without the _configServerMutex below, since then it
// would be likely that the version may have changed in the meantime without waiting for or fetching config results.
// TODO: Mutex-per-namespace?
LOG( 2 ) << "trying to set shard version of " << version.toString() << " for '" << ns << "'" << endl;
_configServerTickets.waitForTicket();
TicketHolderReleaser needTicketFrom( &_configServerTickets );
// fast path - double-check if requested version is at the same version as this chunk manager before verifying
// against config server
//
// This path will short-circuit the version set if another thread already managed to update the version in the
// meantime. First check is from getVersion().
//
// cases:
// + this shard updated the version for a migrate's commit (FROM side)
// a client reloaded chunk state from config and picked the newest version
// + two clients reloaded
// one triggered the 'slow path' (below)
// when the second's request gets here, the version is already current
ConfigVersion storedVersion;
ShardChunkManagerPtr currManager;
{
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if( it == _chunks.end() ){
// TODO: We need better semantic distinction between *no manager found* and
// *manager of version zero found*
log() << "no current chunk manager found for this shard, will initialize" << endl;
}
else{
currManager = it->second;
if( ( storedVersion = it->second->getVersion() ).isEquivalentTo( version ) )
return true;
}
}
LOG( 2 ) << "verifying cached version " << storedVersion.toString() << " and new version " << version.toString() << " for '" << ns << "'" << endl;
// slow path - requested version is different than the current chunk manager's, if one exists, so must check for
// newest version in the config server
//
// cases:
// + a chunk moved TO here
// (we don't bump up the version on the TO side but the commit to config does use higher version)
// a client reloads from config an issued the request
// + there was a take over from a secondary
// the secondary had no state (managers) at all, so every client request will fall here
// + a stale client request a version that's not current anymore
// Can't lock default mutex while creating ShardChunkManager, b/c may have to create a new connection to myself
const string c = (_configServer == _shardHost) ? "" /* local */ : _configServer;
// If our epochs aren't compatible, it's not useful to use the old manager for chunk diffs
if( currManager && ! currManager->getCollVersion().hasCompatibleEpoch( version ) ){
warning() << "detected incompatible version epoch in new version " << version
<< ", old version was " << currManager->getCollVersion() << endl;
currManager.reset();
}
ShardChunkManagerPtr p( ShardChunkManager::make( c , ns , _shardName, currManager ) );
// Handle the case where the collection isn't sharded more gracefully
if( p->getKey().isEmpty() ){
version = ConfigVersion( 0, OID() );
// There was an error getting any data for this collection, return false
return false;
}
{
// NOTE: This lock prevents the ns version from changing while a write operation occurs.
Lock::DBRead readLk(ns);
// This lock prevents simultaneous metadata changes using the same map
scoped_lock lk( _mutex );
// since we loaded the chunk manager unlocked, other thread may have done the same
// make sure we keep the freshest config info only
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if ( it == _chunks.end() || p->getVersion() >= it->second->getVersion() ) {
_chunks[ns] = p;
}
ChunkVersion oldVersion = version;
version = p->getVersion();
return oldVersion.isEquivalentTo( version );
}
}
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;
}
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;
}
/**
* Run a query -- includes checking for and running a Command.
* @return points to ns if exhaust mode. 0=normal mode
* @locks the db mutex for reading (and potentially for writing temporarily to create a new db).
* @yields the db mutex periodically after acquiring it.
* @asserts on scan and order memory exhaustion and other cases.
*/
const char *runQuery(Message& m, QueryMessage& q, CurOp& curop, Message &result) {
shared_ptr<ParsedQuery> pq_shared( new ParsedQuery(q) );
ParsedQuery& pq( *pq_shared );
BSONObj jsobj = q.query;
int queryOptions = q.queryOptions;
const char *ns = q.ns;
if( logLevel >= 2 )
log() << "runQuery called " << ns << " " << jsobj << endl;
curop.debug().ns = ns;
curop.debug().ntoreturn = pq.getNumToReturn();
curop.debug().query = jsobj;
curop.setQuery(jsobj);
// Run a command.
if ( pq.couldBeCommand() ) {
BufBuilder bb;
bb.skip(sizeof(QueryResult));
BSONObjBuilder cmdResBuf;
if ( runCommands(ns, jsobj, curop, bb, cmdResBuf, false, queryOptions) ) {
curop.debug().iscommand = true;
curop.debug().query = jsobj;
curop.markCommand();
auto_ptr< QueryResult > qr;
qr.reset( (QueryResult *) bb.buf() );
bb.decouple();
qr->setResultFlagsToOk();
qr->len = bb.len();
curop.debug().responseLength = bb.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
result.setData( qr.release(), true );
}
else {
uasserted(13530, "bad or malformed command request?");
}
return 0;
}
bool explain = pq.isExplain();
BSONObj order = pq.getOrder();
BSONObj query = pq.getFilter();
/* The ElemIter will not be happy if this isn't really an object. So throw exception
here when that is true.
(Which may indicate bad data from client.)
*/
if ( query.objsize() == 0 ) {
out() << "Bad query object?\n jsobj:";
out() << jsobj.toString() << "\n query:";
out() << query.toString() << endl;
uassert( 10110 , "bad query object", false);
}
Client::ReadContext ctx( ns , dbpath ); // read locks
const ConfigVersion shardingVersionAtStart = shardingState.getVersion( ns );
replVerifyReadsOk(&pq);
if ( pq.hasOption( QueryOption_CursorTailable ) ) {
NamespaceDetails *d = nsdetails( ns );
uassert( 13051, "tailable cursor requested on non capped collection", d && d->isCapped() );
const BSONObj nat1 = BSON( "$natural" << 1 );
if ( order.isEmpty() ) {
order = nat1;
}
else {
uassert( 13052, "only {$natural:1} order allowed for tailable cursor", order == nat1 );
}
}
// Run a simple id query.
if ( ! (explain || pq.showDiskLoc()) && isSimpleIdQuery( query ) && !pq.hasOption( QueryOption_CursorTailable ) ) {
int n = 0;
bool nsFound = false;
bool indexFound = false;
BSONObj resObject;
Client& c = cc();
bool found = Helpers::findById( c, ns , query , resObject , &nsFound , &indexFound );
if ( nsFound == false || indexFound == true ) {
if ( shardingState.needShardChunkManager( ns ) ) {
ShardChunkManagerPtr m = shardingState.getShardChunkManager( ns );
if ( m && ! m->belongsToMe( resObject ) ) {
// I have something this _id
// but it doesn't belong to me
// so return nothing
resObject = BSONObj();
found = false;
}
}
BufBuilder bb(sizeof(QueryResult)+resObject.objsize()+32);
bb.skip(sizeof(QueryResult));
curop.debug().idhack = true;
if ( found ) {
n = 1;
fillQueryResultFromObj( bb , pq.getFields() , resObject );
}
auto_ptr< QueryResult > qr;
qr.reset( (QueryResult *) bb.buf() );
bb.decouple();
qr->setResultFlagsToOk();
qr->len = bb.len();
curop.debug().responseLength = bb.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = n;
result.setData( qr.release(), true );
return NULL;
}
}
// Run a regular query.
BSONObj oldPlan;
if ( explain && ! pq.hasIndexSpecifier() ) {
MultiPlanScanner mps( ns, query, order );
if ( mps.usingCachedPlan() ) {
oldPlan =
mps.oldExplain().firstElement().embeddedObject()
.firstElement().embeddedObject().getOwned();
}
}
// In some cases the query may be retried if there is an in memory sort size assertion.
for( int retry = 0; retry < 2; ++retry ) {
try {
return queryWithQueryOptimizer( m, queryOptions, ns, jsobj, curop, query, order,
pq_shared, oldPlan, shardingVersionAtStart, result );
} catch ( const QueryRetryException & ) {
verify( retry == 0 );
}
}
verify( false );
return 0;
}