bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
string ns = cmdObj["getShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to specify full namespace";
return false;
}
result.append( "configServer" , shardingState.getConfigServer() );
result.appendTimestamp( "global" , shardingState.getVersion(ns).toLong() );
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
result.appendBool( "inShardedMode" , info != 0 );
if ( info )
result.appendTimestamp( "mine" , info->getVersion(ns).toLong() );
else
result.appendTimestamp( "mine" , 0 );
if ( cmdObj["fullMetadata"].trueValue() ) {
CollectionMetadataPtr metadata = shardingState.getCollectionMetadata( ns );
if ( metadata ) result.append( "metadata", metadata->toBSON() );
else result.append( "metadata", BSONObj() );
}
return true;
}
bool checkConfigOrInit( const string& configdb , bool authoritative , string& errmsg , BSONObjBuilder& result , bool locked=false ) const {
if ( configdb.size() == 0 ) {
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ) {
if ( configdb == shardingState.getConfigServer() )
return true;
result.append( "configdb" , BSON( "stored" << shardingState.getConfigServer() <<
"given" << configdb ) );
errmsg = str::stream() << "mongos specified a different config database string : "
<< "stored : " << shardingState.getConfigServer()
<< " vs given : " << configdb;
return false;
}
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
if ( locked ) {
ShardedConnectionInfo::addHook();
shardingState.enable( configdb );
configServer.init( configdb );
return true;
}
Lock::GlobalWrite lk;
return checkConfigOrInit( configdb , authoritative , errmsg , result , true );
}
bool checkConfigOrInit( const string& configdb , bool authoritative , string& errmsg , BSONObjBuilder& result , bool locked=false ) const {
if ( configdb.size() == 0 ) {
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ) {
if ( configdb == shardingState.getConfigServer() )
return true;
result.append( "configdb" , BSON( "stored" << shardingState.getConfigServer() <<
"given" << configdb ) );
errmsg = "specified a different configdb!";
return false;
}
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
if ( locked ) {
shardingState.enable( configdb );
configServer.init( configdb );
return true;
}
dblock lk;
return checkConfigOrInit( configdb , authoritative , errmsg , result , true );
}
void logDeleteOpForSharding(OperationContext* txn,
const char* ns,
const BSONObj& obj,
bool notInActiveChunk) {
ShardingState* shardingState = ShardingState::get(txn);
if (shardingState->enabled())
shardingState->migrationSourceManager()->logDeleteOp(txn, ns, obj, notInActiveChunk);
}
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , bool isWriteOp , string& errmsg ) {
if ( ! shardingState.enabled() )
return true;
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ) {
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ) {
return true;
}
// TODO
// all collections at some point, be sharded or not, will have a version (and a ShardChunkManager)
// for now, we remove the sharding state of dropped collection
// so delayed request may come in. This has to be fixed.
ConfigVersion clientVersion = info->getVersion(ns);
ConfigVersion version;
if ( ! shardingState.hasVersion( ns , version ) && clientVersion == 0 ) {
return true;
}
if ( version == 0 && clientVersion > 0 ) {
stringstream ss;
ss << "collection was dropped or this shard no longer valied version: " << version << " clientVersion: " << clientVersion;
errmsg = ss.str();
return false;
}
if ( clientVersion >= version )
return true;
if ( clientVersion == 0 ) {
stringstream ss;
ss << "client in sharded mode, but doesn't have version set for this collection: " << ns << " myVersion: " << version;
errmsg = ss.str();
return false;
}
if ( version.majorVersion() == clientVersion.majorVersion() ) {
// this means there was just a split
// since on a split w/o a migrate this server is ok
// going to accept
return true;
}
stringstream ss;
ss << "your version is too old ns: " + ns << " global: " << version << " client: " << clientVersion;
errmsg = ss.str();
return false;
}
bool haveLocalShardingInfo( const string& ns ) {
if ( ! shardingState.enabled() )
return false;
if ( ! shardingState.hasVersion( ns ) )
return false;
return ShardedConnectionInfo::get(false) > 0;
}
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , bool isWriteOp , string& errmsg ){
if ( ! shardingState.enabled() )
return true;
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ){
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ){
return true;
}
ConfigVersion version;
if ( ! shardingState.hasVersion( ns , version ) ){
return true;
}
ConfigVersion clientVersion = info->getVersion(ns);
if ( version == 0 && clientVersion > 0 ){
stringstream ss;
ss << "collection was dropped or this shard no longer valied version: " << version << " clientVersion: " << clientVersion;
errmsg = ss.str();
return false;
}
if ( clientVersion >= version )
return true;
if ( clientVersion == 0 ){
stringstream ss;
ss << "client in sharded mode, but doesn't have version set for this collection: " << ns << " myVersion: " << version;
errmsg = ss.str();
return false;
}
if ( isWriteOp && version.majorVersion() == clientVersion.majorVersion() ){
// this means there was just a split
// since on a split w/o a migrate this server is ok
// going to accept write
return true;
}
stringstream ss;
ss << "your version is too old ns: " + ns << " global: " << version << " client: " << clientVersion;
errmsg = ss.str();
return false;
}
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
string ns = cmdObj["getShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to speciy fully namespace";
return false;
}
result.append( "configServer" , shardingState.getConfigServer() );
result.appendTimestamp( "global" , shardingState.getVersion(ns) );
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( info )
result.appendTimestamp( "mine" , info->getVersion(ns) );
else
result.appendTimestamp( "mine" , 0 );
return true;
}
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool){
// Debugging code for SERVER-1633. Commands have already a coarser timer for
// normal operation.
Timer timer;
vector<int> laps;
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
bool authoritative = cmdObj.getBoolField( "authoritative" );
string configdb = cmdObj["configdb"].valuestrsafe();
{ // configdb checking
if ( configdb.size() == 0 ){
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ){
if ( configdb != shardingState.getConfigServer() ){
errmsg = "specified a different configdb!";
return false;
}
}
else {
if ( ! authoritative ){
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
shardingState.enable( configdb );
configServer.init( configdb );
}
}
// SERVER-1633
laps.push_back( timer.millis() );
if ( cmdObj["shard"].type() == String ){
shardingState.gotShardName( cmdObj["shard"].String() );
shardingState.gotShardHost( cmdObj["shardHost"].String() );
}
{ // setting up ids
if ( cmdObj["serverID"].type() != jstOID ){
// TODO: fix this
//errmsg = "need serverID to be an OID";
//return 0;
}
else {
OID clientId = cmdObj["serverID"].__oid();
if ( ! info->hasID() ){
info->setID( clientId );
}
else if ( clientId != info->getID() ){
errmsg = "server id has changed!";
return 0;
}
}
}
// SERVER-1633
laps.push_back( timer.millis() );
unsigned long long version = extractVersion( cmdObj["version"] , errmsg );
if ( errmsg.size() ){
return false;
}
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ){
errmsg = "need to speciy fully namespace";
return false;
}
ConfigVersion& oldVersion = info->getVersion(ns);
ConfigVersion& globalVersion = shardingState.getVersion(ns);
if ( oldVersion > 0 && globalVersion == 0 ){
// this had been reset
oldVersion = 0;
}
if ( version == 0 && globalVersion == 0 ){
// this connection is cleaning itself
oldVersion = 0;
return 1;
}
// SERVER-1633
laps.push_back( timer.millis() );
if ( version == 0 && globalVersion > 0 ){
if ( ! authoritative ){
result.appendBool( "need_authoritative" , true );
result.appendTimestamp( "globalVersion" , globalVersion );
result.appendTimestamp( "oldVersion" , oldVersion );
errmsg = "dropping needs to be authoritative";
return 0;
}
log() << "wiping data for: " << ns << endl;
result.appendTimestamp( "beforeDrop" , globalVersion );
// only setting global version on purpose
// need clients to re-find meta-data
globalVersion = 0;
oldVersion = 0;
return 1;
}
if ( version < oldVersion ){
errmsg = "you already have a newer version";
result.appendTimestamp( "oldVersion" , oldVersion );
result.appendTimestamp( "newVersion" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
// SERVER-1633
laps.push_back( timer.millis() );
if ( version < globalVersion ){
while ( shardingState.inCriticalMigrateSection() ){
dbtemprelease r;
sleepmillis(2);
log() << "waiting till out of critical section" << endl;
}
errmsg = "going to older version for global";
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( globalVersion == 0 && ! cmdObj.getBoolField( "authoritative" ) ){
// need authoritative for first look
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
errmsg = "first time for this ns";
return false;
}
// SERVER-1633
laps.push_back( timer.millis() );
{
dbtemprelease unlock;
shardingState.getChunkMatcher( ns );
}
result.appendTimestamp( "oldVersion" , oldVersion );
oldVersion = version;
globalVersion = version;
// SERVER-1633
ostringstream lapString;
lapString << name /* command name */ << " partials: " ;
for (size_t i = 1; i<laps.size(); ++i){
lapString << (laps[i] - laps[i-1]) / 1000 << " ";
}
lapString << endl;
logIfSlow( timer, lapString.str() );
result.append( "ok" , 1 );
return 1;
}
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
shardingState.appendInfo( result );
return true;
}
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
bool authoritative = cmdObj.getBoolField( "authoritative" );
string configdb = cmdObj["configdb"].valuestrsafe();
{
// configdb checking
if ( configdb.size() == 0 ) {
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ) {
if ( configdb != shardingState.getConfigServer() ) {
errmsg = "specified a different configdb!";
return false;
}
}
else {
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
shardingState.enable( configdb );
configServer.init( configdb );
}
}
if ( cmdObj["shard"].type() == String ) {
shardingState.gotShardName( cmdObj["shard"].String() );
shardingState.gotShardHost( cmdObj["shardHost"].String() );
}
{
// setting up ids
if ( cmdObj["serverID"].type() != jstOID ) {
// TODO: fix this
//errmsg = "need serverID to be an OID";
//return 0;
}
else {
OID clientId = cmdObj["serverID"].__oid();
if ( ! info->hasID() ) {
info->setID( clientId );
}
else if ( clientId != info->getID() ) {
errmsg = "server id has changed!";
return 0;
}
}
}
unsigned long long version = extractVersion( cmdObj["version"] , errmsg );
if ( errmsg.size() ) {
return false;
}
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to speciy fully namespace";
return false;
}
const ConfigVersion oldVersion = info->getVersion(ns);
const ConfigVersion globalVersion = shardingState.getVersion(ns);
if ( oldVersion > 0 && globalVersion == 0 ) {
// this had been reset
info->setVersion( ns , 0 );
}
if ( version == 0 && globalVersion == 0 ) {
// this connection is cleaning itself
info->setVersion( ns , 0 );
return true;
}
if ( version == 0 && globalVersion > 0 ) {
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
result.appendTimestamp( "globalVersion" , globalVersion );
result.appendTimestamp( "oldVersion" , oldVersion );
errmsg = "dropping needs to be authoritative";
return false;
}
log() << "wiping data for: " << ns << endl;
result.appendTimestamp( "beforeDrop" , globalVersion );
// only setting global version on purpose
// need clients to re-find meta-data
shardingState.resetVersion( ns );
info->setVersion( ns , 0 );
return true;
}
if ( version < oldVersion ) {
errmsg = "you already have a newer version of collection '" + ns + "'";
result.append( "ns" , ns );
result.appendTimestamp( "oldVersion" , oldVersion );
result.appendTimestamp( "newVersion" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( version < globalVersion ) {
while ( shardingState.inCriticalMigrateSection() ) {
dbtemprelease r;
sleepmillis(2);
OCCASIONALLY log() << "waiting till out of critical section" << endl;
}
errmsg = "going to older version for global for collection '" + ns + "'";
result.append( "ns" , ns );
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( globalVersion == 0 && ! cmdObj.getBoolField( "authoritative" ) ) {
// need authoritative for first look
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "first time for collection '" + ns + "'";
return false;
}
{
dbtemprelease unlock;
ShardChunkVersion currVersion = version;
if ( ! shardingState.trySetVersion( ns , currVersion ) ) {
errmsg = str::stream() << "client version differs from config's for colleciton '" << ns << "'";
result.append( "ns" , ns );
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , currVersion );
return false;
}
}
info->setVersion( ns , version );
result.appendTimestamp( "oldVersion" , oldVersion );
result.append( "ok" , 1 );
return true;
}
namespace mongo {
// -----ShardingState START ----
ShardingState::ShardingState()
: _enabled(false) , _mutex( "ShardingState" ) {
}
void ShardingState::enable( const string& server ) {
_enabled = true;
assert( server.size() );
if ( _configServer.size() == 0 )
_configServer = server;
else {
assert( server == _configServer );
}
}
void ShardingState::gotShardName( const string& name ) {
if ( _shardName.size() == 0 ) {
// TODO SERVER-2299 verify the name is sound w.r.t IPs
_shardName = name;
return;
}
if ( _shardName == name )
return;
stringstream ss;
ss << "gotShardName different than what i had before "
<< " before [" << _shardName << "] "
<< " got [" << name << "] "
;
uasserted( 13298 , ss.str() );
}
void ShardingState::gotShardHost( string host ) {
size_t slash = host.find( '/' );
if ( slash != string::npos )
host = host.substr( 0 , slash );
if ( _shardHost.size() == 0 ) {
_shardHost = host;
return;
}
if ( _shardHost == host )
return;
stringstream ss;
ss << "gotShardHost different than what i had before "
<< " before [" << _shardHost << "] "
<< " got [" << host << "] "
;
uasserted( 13299 , ss.str() );
}
void ShardingState::resetShardingState() {
scoped_lock lk(_mutex);
_enabled = false;
_configServer.clear();
_shardName.clear();
_shardHost.clear();
_chunks.clear();
}
// TODO we shouldn't need three ways for checking the version. Fix this.
bool ShardingState::hasVersion( const string& ns ) {
scoped_lock lk(_mutex);
ChunkManagersMap::const_iterator it = _chunks.find(ns);
return it != _chunks.end();
}
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;
}
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;
}
}
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::undoDonateChunk( 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->clonePlus( min , max , version ) );
_chunks[ns] = p;
}
void ShardingState::splitChunk( const string& ns , const BSONObj& min , const BSONObj& max , const vector<BSONObj>& splitKeys ,
ShardChunkVersion version ) {
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
assert( it != _chunks.end() ) ;
ShardChunkManagerPtr p( it->second->cloneSplit( min , max , splitKeys , version ) );
_chunks[ns] = p;
}
void ShardingState::resetVersion( const string& ns ) {
scoped_lock lk( _mutex );
_chunks.erase( ns );
}
bool ShardingState::trySetVersion( const string& ns , ConfigVersion& version /* IN-OUT */ ) {
// fast path - requested version is at the same version as this chunk manager
//
// 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
{
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if ( it != _chunks.end() && it->second->getVersion() == version )
return true;
}
// 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
const string c = (_configServer == _shardHost) ? "" /* local */ : _configServer;
ShardChunkManagerPtr p( new ShardChunkManager( c , ns , _shardName ) );
{
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;
}
ShardChunkVersion oldVersion = version;
version = p->getVersion();
return oldVersion == version;
}
}
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();
}
}
bool ShardingState::needShardChunkManager( const string& ns ) const {
if ( ! _enabled )
return false;
if ( ! ShardedConnectionInfo::get( false ) )
return false;
return true;
}
ShardChunkManagerPtr ShardingState::getShardChunkManager( const string& ns ) {
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if ( it == _chunks.end() ) {
return ShardChunkManagerPtr();
}
else {
return it->second;
}
}
ShardingState shardingState;
// -----ShardingState END ----
// -----ShardedConnectionInfo START ----
boost::thread_specific_ptr<ShardedConnectionInfo> ShardedConnectionInfo::_tl;
ShardedConnectionInfo::ShardedConnectionInfo() {
_forceVersionOk = false;
_id.clear();
}
ShardedConnectionInfo* ShardedConnectionInfo::get( bool create ) {
ShardedConnectionInfo* info = _tl.get();
if ( ! info && create ) {
log(1) << "entering shard mode for connection" << endl;
info = new ShardedConnectionInfo();
_tl.reset( info );
}
return info;
}
void ShardedConnectionInfo::reset() {
_tl.reset();
}
const ConfigVersion ShardedConnectionInfo::getVersion( const string& ns ) const {
NSVersionMap::const_iterator it = _versions.find( ns );
if ( it != _versions.end() ) {
return it->second;
}
else {
return 0;
}
}
void ShardedConnectionInfo::setVersion( const string& ns , const ConfigVersion& version ) {
_versions[ns] = version;
}
void ShardedConnectionInfo::setID( const OID& id ) {
_id = id;
}
// -----ShardedConnectionInfo END ----
unsigned long long extractVersion( BSONElement e , string& errmsg ) {
if ( e.eoo() ) {
errmsg = "no version";
return 0;
}
if ( e.isNumber() )
return (unsigned long long)e.number();
if ( e.type() == Date || e.type() == Timestamp )
return e._numberLong();
errmsg = "version is not a numeric type";
return 0;
}
class MongodShardCommand : public Command {
public:
MongodShardCommand( const char * n ) : Command( n ) {
}
virtual bool slaveOk() const {
return false;
}
virtual bool adminOnly() const {
return true;
}
};
bool haveLocalShardingInfo( const string& ns ) {
if ( ! shardingState.enabled() )
return false;
if ( ! shardingState.hasVersion( ns ) )
return false;
return ShardedConnectionInfo::get(false) > 0;
}
class UnsetShardingCommand : public MongodShardCommand {
public:
UnsetShardingCommand() : MongodShardCommand("unsetSharding") {}
virtual void help( stringstream& help ) const {
help << " example: { unsetSharding : 1 } ";
}
virtual LockType locktype() const { return NONE; }
virtual bool slaveOk() const { return true; }
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
ShardedConnectionInfo::reset();
return true;
}
} unsetShardingCommand;
class SetShardVersion : public MongodShardCommand {
public:
SetShardVersion() : MongodShardCommand("setShardVersion") {}
virtual void help( stringstream& help ) const {
help << " example: { setShardVersion : 'alleyinsider.foo' , version : 1 , configdb : '' } ";
}
virtual LockType locktype() const { return WRITE; } // TODO: figure out how to make this not need to lock
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
bool authoritative = cmdObj.getBoolField( "authoritative" );
string configdb = cmdObj["configdb"].valuestrsafe();
{
// configdb checking
if ( configdb.size() == 0 ) {
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ) {
if ( configdb != shardingState.getConfigServer() ) {
errmsg = "specified a different configdb!";
return false;
}
}
else {
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
shardingState.enable( configdb );
configServer.init( configdb );
}
}
if ( cmdObj["shard"].type() == String ) {
shardingState.gotShardName( cmdObj["shard"].String() );
shardingState.gotShardHost( cmdObj["shardHost"].String() );
}
{
// setting up ids
if ( cmdObj["serverID"].type() != jstOID ) {
// TODO: fix this
//errmsg = "need serverID to be an OID";
//return 0;
}
else {
OID clientId = cmdObj["serverID"].__oid();
if ( ! info->hasID() ) {
info->setID( clientId );
}
else if ( clientId != info->getID() ) {
errmsg = "server id has changed!";
return 0;
}
}
}
unsigned long long version = extractVersion( cmdObj["version"] , errmsg );
if ( errmsg.size() ) {
return false;
}
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to speciy fully namespace";
return false;
}
const ConfigVersion oldVersion = info->getVersion(ns);
const ConfigVersion globalVersion = shardingState.getVersion(ns);
if ( oldVersion > 0 && globalVersion == 0 ) {
// this had been reset
info->setVersion( ns , 0 );
}
if ( version == 0 && globalVersion == 0 ) {
// this connection is cleaning itself
info->setVersion( ns , 0 );
return true;
}
if ( version == 0 && globalVersion > 0 ) {
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
result.appendTimestamp( "globalVersion" , globalVersion );
result.appendTimestamp( "oldVersion" , oldVersion );
errmsg = "dropping needs to be authoritative";
return false;
}
log() << "wiping data for: " << ns << endl;
result.appendTimestamp( "beforeDrop" , globalVersion );
// only setting global version on purpose
// need clients to re-find meta-data
shardingState.resetVersion( ns );
info->setVersion( ns , 0 );
return true;
}
if ( version < oldVersion ) {
errmsg = "you already have a newer version of collection '" + ns + "'";
result.append( "ns" , ns );
result.appendTimestamp( "oldVersion" , oldVersion );
result.appendTimestamp( "newVersion" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( version < globalVersion ) {
while ( shardingState.inCriticalMigrateSection() ) {
dbtemprelease r;
sleepmillis(2);
OCCASIONALLY log() << "waiting till out of critical section" << endl;
}
errmsg = "going to older version for global for collection '" + ns + "'";
result.append( "ns" , ns );
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( globalVersion == 0 && ! cmdObj.getBoolField( "authoritative" ) ) {
// need authoritative for first look
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "first time for collection '" + ns + "'";
return false;
}
{
dbtemprelease unlock;
ShardChunkVersion currVersion = version;
if ( ! shardingState.trySetVersion( ns , currVersion ) ) {
errmsg = str::stream() << "client version differs from config's for colleciton '" << ns << "'";
result.append( "ns" , ns );
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , currVersion );
return false;
}
}
info->setVersion( ns , version );
result.appendTimestamp( "oldVersion" , oldVersion );
result.append( "ok" , 1 );
return true;
}
} setShardVersionCmd;
class GetShardVersion : public MongodShardCommand {
public:
GetShardVersion() : MongodShardCommand("getShardVersion") {}
virtual void help( stringstream& help ) const {
help << " example: { getShardVersion : 'alleyinsider.foo' } ";
}
virtual LockType locktype() const { return NONE; }
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
string ns = cmdObj["getShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to speciy fully namespace";
return false;
}
result.append( "configServer" , shardingState.getConfigServer() );
result.appendTimestamp( "global" , shardingState.getVersion(ns) );
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( info )
result.appendTimestamp( "mine" , info->getVersion(ns) );
else
result.appendTimestamp( "mine" , 0 );
return true;
}
} getShardVersion;
class ShardingStateCmd : public MongodShardCommand {
public:
ShardingStateCmd() : MongodShardCommand( "shardingState" ) {}
virtual LockType locktype() const { return WRITE; } // TODO: figure out how to make this not need to lock
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
shardingState.appendInfo( result );
return true;
}
} shardingStateCmd;
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , bool isWriteOp , string& errmsg ) {
if ( ! shardingState.enabled() )
return true;
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ) {
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ) {
return true;
}
// TODO
// all collections at some point, be sharded or not, will have a version (and a ShardChunkManager)
// for now, we remove the sharding state of dropped collection
// so delayed request may come in. This has to be fixed.
ConfigVersion clientVersion = info->getVersion(ns);
ConfigVersion version;
if ( ! shardingState.hasVersion( ns , version ) && clientVersion == 0 ) {
return true;
}
if ( version == 0 && clientVersion > 0 ) {
stringstream ss;
ss << "collection was dropped or this shard no longer valied version: " << version << " clientVersion: " << clientVersion;
errmsg = ss.str();
return false;
}
if ( clientVersion >= version )
return true;
if ( clientVersion == 0 ) {
stringstream ss;
ss << "client in sharded mode, but doesn't have version set for this collection: " << ns << " myVersion: " << version;
errmsg = ss.str();
return false;
}
if ( version.majorVersion() == clientVersion.majorVersion() ) {
// this means there was just a split
// since on a split w/o a migrate this server is ok
// going to accept
return true;
}
stringstream ss;
ss << "your version is too old ns: " + ns << " global: " << version << " client: " << clientVersion;
errmsg = ss.str();
return false;
}
}
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , string& errmsg, ConfigVersion& received, ConfigVersion& wanted ) {
if ( ! shardingState.enabled() )
return true;
if ( ! isMasterNs( ns.c_str() ) ) {
// right now connections to secondaries aren't versioned at all
return true;
}
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ) {
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ) {
return true;
}
// TODO
// all collections at some point, be sharded or not, will have a version (and a ShardChunkManager)
// for now, we remove the sharding state of dropped collection
// so delayed request may come in. This has to be fixed.
ConfigVersion clientVersion = info->getVersion(ns);
ConfigVersion version;
if ( ! shardingState.hasVersion( ns , version ) && ! clientVersion.isSet() ) {
return true;
}
// The versions we're going to compare, saved for future use
received = clientVersion;
wanted = version;
if ( ! version.isSet() && clientVersion.isSet() ) {
stringstream ss;
ss << "collection was dropped or this shard no longer valid version";
errmsg = ss.str();
return false;
}
if ( clientVersion >= version )
return true;
if ( ! clientVersion.isSet() ) {
stringstream ss;
ss << "client in sharded mode, but doesn't have version set for this collection";
errmsg = ss.str();
return false;
}
if ( version.majorVersion() == clientVersion.majorVersion() ) {
// this means there was just a split
// since on a split w/o a migrate this server is ok
// going to accept
return true;
}
stringstream ss;
ss << "your version is too old";
errmsg = ss.str();
return false;
}
namespace mongo {
// -----ShardingState START ----
ShardingState::ShardingState()
: _enabled(false) , _mutex( "ShardingState" ){
}
void ShardingState::enable( const string& server ){
_enabled = true;
assert( server.size() );
if ( _configServer.size() == 0 )
_configServer = server;
else {
assert( server == _configServer );
}
}
void ShardingState::gotShardName( const string& name ){
if ( _shardName.size() == 0 ){
_shardName = name;
return;
}
if ( _shardName == name )
return;
stringstream ss;
ss << "gotShardName different than what i had before "
<< " before [" << _shardName << "] "
<< " got [" << name << "] "
;
uasserted( 13298 , ss.str() );
}
void ShardingState::gotShardHost( string host ){
size_t slash = host.find( '/' );
if ( slash != string::npos )
host = host.substr( 0 , slash );
if ( _shardHost.size() == 0 ){
_shardHost = host;
return;
}
if ( _shardHost == host )
return;
stringstream ss;
ss << "gotShardHost different than what i had before "
<< " before [" << _shardHost << "] "
<< " got [" << host << "] "
;
uasserted( 13299 , ss.str() );
}
bool ShardingState::hasVersion( const string& ns ){
scoped_lock lk(_mutex);
NSVersionMap::const_iterator i = _versions.find(ns);
return i != _versions.end();
}
bool ShardingState::hasVersion( const string& ns , ConfigVersion& version ){
scoped_lock lk(_mutex);
NSVersionMap::const_iterator i = _versions.find(ns);
if ( i == _versions.end() )
return false;
version = i->second;
return true;
}
const ConfigVersion ShardingState::getVersion( const string& ns ) const {
scoped_lock lk(_mutex);
NSVersionMap::const_iterator it = _versions.find( ns );
if ( it != _versions.end() ) {
return it->second;
} else {
return 0;
}
}
void ShardingState::setVersion( const string& ns , const ConfigVersion& version ){
scoped_lock lk(_mutex);
if ( version != 0 ) {
NSVersionMap::const_iterator it = _versions.find( ns );
// TODO 11-18-2010 as we're bringing chunk boundary information to mongod, it may happen that
// we're setting a version for the ns that the shard knows about already (e.g because it set
// it itself in a chunk migration)
// eventually, the only cases to issue a setVersion would be
// 1) First chunk of a collection, for version 1|0
// 2) Drop of a collection, for version 0|0
// 3) Load of the shard's chunk state, in a primary-secondary failover
assert( it == _versions.end() || version >= it->second );
}
_versions[ns] = version;
}
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 ( NSVersionMap::iterator i=_versions.begin(); i!=_versions.end(); ++i ){
bb.appendTimestamp( i->first , i->second );
}
bb.done();
}
}
bool ShardingState::needShardChunkManager( const string& ns ) const {
if ( ! _enabled )
return false;
if ( ! ShardedConnectionInfo::get( false ) )
return false;
return true;
}
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;
}
ShardingState shardingState;
// -----ShardingState END ----
// -----ShardedConnectionInfo START ----
boost::thread_specific_ptr<ShardedConnectionInfo> ShardedConnectionInfo::_tl;
ShardedConnectionInfo::ShardedConnectionInfo(){
_forceVersionOk = false;
_id.clear();
}
ShardedConnectionInfo* ShardedConnectionInfo::get( bool create ){
ShardedConnectionInfo* info = _tl.get();
if ( ! info && create ){
log(1) << "entering shard mode for connection" << endl;
info = new ShardedConnectionInfo();
_tl.reset( info );
}
return info;
}
void ShardedConnectionInfo::reset(){
_tl.reset();
}
const ConfigVersion ShardedConnectionInfo::getVersion( const string& ns ) const {
NSVersionMap::const_iterator it = _versions.find( ns );
if ( it != _versions.end() ) {
return it->second;
} else {
return 0;
}
}
void ShardedConnectionInfo::setVersion( const string& ns , const ConfigVersion& version ){
_versions[ns] = version;
}
void ShardedConnectionInfo::setID( const OID& id ){
_id = id;
}
// -----ShardedConnectionInfo END ----
unsigned long long extractVersion( BSONElement e , string& errmsg ){
if ( e.eoo() ){
errmsg = "no version";
return 0;
}
if ( e.isNumber() )
return (unsigned long long)e.number();
if ( e.type() == Date || e.type() == Timestamp )
return e._numberLong();
errmsg = "version is not a numeric type";
return 0;
}
class MongodShardCommand : public Command {
public:
MongodShardCommand( const char * n ) : Command( n ){
}
virtual bool slaveOk() const {
return false;
}
virtual bool adminOnly() const {
return true;
}
};
bool haveLocalShardingInfo( const string& ns ){
if ( ! shardingState.enabled() )
return false;
if ( ! shardingState.hasVersion( ns ) )
return false;
return ShardedConnectionInfo::get(false) > 0;
}
class UnsetShardingCommand : public MongodShardCommand {
public:
UnsetShardingCommand() : MongodShardCommand("unsetSharding"){}
virtual void help( stringstream& help ) const {
help << " example: { unsetSharding : 1 } ";
}
virtual LockType locktype() const { return NONE; }
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool){
ShardedConnectionInfo::reset();
return true;
}
} unsetShardingCommand;
class SetShardVersion : public MongodShardCommand {
public:
SetShardVersion() : MongodShardCommand("setShardVersion"){}
virtual void help( stringstream& help ) const {
help << " example: { setShardVersion : 'alleyinsider.foo' , version : 1 , configdb : '' } ";
}
virtual LockType locktype() const { return WRITE; } // TODO: figure out how to make this not need to lock
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool){
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
bool authoritative = cmdObj.getBoolField( "authoritative" );
string configdb = cmdObj["configdb"].valuestrsafe();
{ // configdb checking
if ( configdb.size() == 0 ){
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ){
if ( configdb != shardingState.getConfigServer() ){
errmsg = "specified a different configdb!";
return false;
}
}
else {
if ( ! authoritative ){
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
shardingState.enable( configdb );
configServer.init( configdb );
}
}
if ( cmdObj["shard"].type() == String ){
shardingState.gotShardName( cmdObj["shard"].String() );
shardingState.gotShardHost( cmdObj["shardHost"].String() );
}
{ // setting up ids
if ( cmdObj["serverID"].type() != jstOID ){
// TODO: fix this
//errmsg = "need serverID to be an OID";
//return 0;
}
else {
OID clientId = cmdObj["serverID"].__oid();
if ( ! info->hasID() ){
info->setID( clientId );
}
else if ( clientId != info->getID() ){
errmsg = "server id has changed!";
return 0;
}
}
}
unsigned long long version = extractVersion( cmdObj["version"] , errmsg );
if ( errmsg.size() ){
return false;
}
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ){
errmsg = "need to speciy fully namespace";
return false;
}
const ConfigVersion oldVersion = info->getVersion(ns);
const ConfigVersion globalVersion = shardingState.getVersion(ns);
if ( oldVersion > 0 && globalVersion == 0 ){
// this had been reset
info->setVersion( ns , 0 );
}
if ( version == 0 && globalVersion == 0 ){
// this connection is cleaning itself
info->setVersion( ns , 0 );
return true;
}
if ( version == 0 && globalVersion > 0 ){
if ( ! authoritative ){
result.appendBool( "need_authoritative" , true );
result.appendTimestamp( "globalVersion" , globalVersion );
result.appendTimestamp( "oldVersion" , oldVersion );
errmsg = "dropping needs to be authoritative";
return false;
}
log() << "wiping data for: " << ns << endl;
result.appendTimestamp( "beforeDrop" , globalVersion );
// only setting global version on purpose
// need clients to re-find meta-data
shardingState.setVersion( ns , 0 );
info->setVersion( ns , 0 );
return true;
}
if ( version < oldVersion ){
errmsg = "you already have a newer version";
result.appendTimestamp( "oldVersion" , oldVersion );
result.appendTimestamp( "newVersion" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( version < globalVersion ){
while ( shardingState.inCriticalMigrateSection() ){
dbtemprelease r;
sleepmillis(2);
log() << "waiting till out of critical section" << endl;
}
errmsg = "going to older version for global";
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( globalVersion == 0 && ! cmdObj.getBoolField( "authoritative" ) ){
// need authoritative for first look
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
errmsg = "first time for this ns";
return false;
}
result.appendTimestamp( "oldVersion" , oldVersion );
result.append( "ok" , 1 );
info->setVersion( ns , version );
shardingState.setVersion( ns , version );
// TODO SERVER-1849 pending drop work
// getShardChunkManager is assuming that the setVersion above were valid
// ideally, we'd call getShardChunkManager first, verify that 'version' is sound, and then update
// connection and global state
{
dbtemprelease unlock;
shardingState.getShardChunkManager( ns );
}
return true;
}
} setShardVersionCmd;
class GetShardVersion : public MongodShardCommand {
public:
GetShardVersion() : MongodShardCommand("getShardVersion"){}
virtual void help( stringstream& help ) const {
help << " example: { getShardVersion : 'alleyinsider.foo' } ";
}
virtual LockType locktype() const { return NONE; }
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool){
string ns = cmdObj["getShardVersion"].valuestrsafe();
if ( ns.size() == 0 ){
errmsg = "need to speciy fully namespace";
return false;
}
result.append( "configServer" , shardingState.getConfigServer() );
result.appendTimestamp( "global" , shardingState.getVersion(ns) );
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( info )
result.appendTimestamp( "mine" , info->getVersion(ns) );
else
result.appendTimestamp( "mine" , 0 );
return true;
}
} getShardVersion;
class ShardingStateCmd : public MongodShardCommand {
public:
ShardingStateCmd() : MongodShardCommand( "shardingState" ){}
virtual LockType locktype() const { return WRITE; } // TODO: figure out how to make this not need to lock
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool){
shardingState.appendInfo( result );
return true;
}
} shardingStateCmd;
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , bool isWriteOp , string& errmsg ){
if ( ! shardingState.enabled() )
return true;
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ){
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ){
return true;
}
ConfigVersion version;
if ( ! shardingState.hasVersion( ns , version ) ){
return true;
}
ConfigVersion clientVersion = info->getVersion(ns);
if ( version == 0 && clientVersion > 0 ){
stringstream ss;
ss << "collection was dropped or this shard no longer valied version: " << version << " clientVersion: " << clientVersion;
errmsg = ss.str();
return false;
}
if ( clientVersion >= version )
return true;
if ( clientVersion == 0 ){
stringstream ss;
ss << "client in sharded mode, but doesn't have version set for this collection: " << ns << " myVersion: " << version;
errmsg = ss.str();
return false;
}
if ( isWriteOp && version.majorVersion() == clientVersion.majorVersion() ){
// this means there was just a split
// since on a split w/o a migrate this server is ok
// going to accept write
return true;
}
stringstream ss;
ss << "your version is too old ns: " + ns << " global: " << version << " client: " << clientVersion;
errmsg = ss.str();
return false;
}
}
bool mergeChunks(OperationContext* txn,
const NamespaceString& nss,
const BSONObj& minKey,
const BSONObj& maxKey,
const OID& epoch,
string* errMsg) {
// Get the distributed lock
string whyMessage = stream() << "merging chunks in " << nss.ns() << " from " << minKey << " to "
<< maxKey;
auto scopedDistLock = grid.catalogManager(txn)->distLock(
txn, nss.ns(), whyMessage, DistLockManager::kSingleLockAttemptTimeout);
if (!scopedDistLock.isOK()) {
*errMsg = stream() << "could not acquire collection lock for " << nss.ns()
<< " to merge chunks in [" << minKey << "," << maxKey << ")"
<< causedBy(scopedDistLock.getStatus());
warning() << *errMsg;
return false;
}
ShardingState* shardingState = ShardingState::get(txn);
//
// We now have the collection lock, refresh metadata to latest version and sanity check
//
ChunkVersion shardVersion;
Status status = shardingState->refreshMetadataNow(txn, nss.ns(), &shardVersion);
if (!status.isOK()) {
*errMsg = str::stream() << "could not merge chunks, failed to refresh metadata for "
<< nss.ns() << causedBy(status.reason());
warning() << *errMsg;
return false;
}
if (epoch.isSet() && shardVersion.epoch() != epoch) {
*errMsg = stream() << "could not merge chunks, collection " << nss.ns() << " has changed"
<< " since merge was sent"
<< "(sent epoch : " << epoch.toString()
<< ", current epoch : " << shardVersion.epoch().toString() << ")";
warning() << *errMsg;
return false;
}
shared_ptr<CollectionMetadata> metadata = shardingState->getCollectionMetadata(nss.ns());
if (!metadata || metadata->getKeyPattern().isEmpty()) {
*errMsg = stream() << "could not merge chunks, collection " << nss.ns()
<< " is not sharded";
warning() << *errMsg;
return false;
}
dassert(metadata->getShardVersion().equals(shardVersion));
if (!metadata->isValidKey(minKey) || !metadata->isValidKey(maxKey)) {
*errMsg = stream() << "could not merge chunks, the range " << rangeToString(minKey, maxKey)
<< " is not valid"
<< " for collection " << nss.ns() << " with key pattern "
<< metadata->getKeyPattern();
warning() << *errMsg;
return false;
}
//
// Get merged chunk information
//
ChunkVersion mergeVersion = metadata->getCollVersion();
mergeVersion.incMinor();
std::vector<ChunkType> chunksToMerge;
ChunkType itChunk;
itChunk.setMin(minKey);
itChunk.setMax(minKey);
itChunk.setNS(nss.ns());
itChunk.setShard(shardingState->getShardName());
while (itChunk.getMax().woCompare(maxKey) < 0 &&
metadata->getNextChunk(itChunk.getMax(), &itChunk)) {
chunksToMerge.push_back(itChunk);
}
if (chunksToMerge.empty()) {
*errMsg = stream() << "could not merge chunks, collection " << nss.ns()
<< " range starting at " << minKey << " and ending at " << maxKey
<< " does not belong to shard " << shardingState->getShardName();
warning() << *errMsg;
return false;
}
//
// Validate the range starts and ends at chunks and has no holes, error if not valid
//
BSONObj firstDocMin = chunksToMerge.front().getMin();
BSONObj firstDocMax = chunksToMerge.front().getMax();
// minKey is inclusive
bool minKeyInRange = rangeContains(firstDocMin, firstDocMax, minKey);
if (!minKeyInRange) {
*errMsg = stream() << "could not merge chunks, collection " << nss.ns()
<< " range starting at " << minKey << " does not belong to shard "
<< shardingState->getShardName();
warning() << *errMsg;
return false;
}
BSONObj lastDocMin = chunksToMerge.back().getMin();
BSONObj lastDocMax = chunksToMerge.back().getMax();
// maxKey is exclusive
bool maxKeyInRange = lastDocMin.woCompare(maxKey) < 0 && lastDocMax.woCompare(maxKey) >= 0;
if (!maxKeyInRange) {
*errMsg = stream() << "could not merge chunks, collection " << nss.ns()
<< " range ending at " << maxKey << " does not belong to shard "
<< shardingState->getShardName();
warning() << *errMsg;
return false;
}
bool validRangeStartKey = firstDocMin.woCompare(minKey) == 0;
bool validRangeEndKey = lastDocMax.woCompare(maxKey) == 0;
if (!validRangeStartKey || !validRangeEndKey) {
*errMsg = stream() << "could not merge chunks, collection " << nss.ns()
<< " does not contain a chunk "
<< (!validRangeStartKey ? "starting at " + minKey.toString() : "")
<< (!validRangeStartKey && !validRangeEndKey ? " or " : "")
<< (!validRangeEndKey ? "ending at " + maxKey.toString() : "");
warning() << *errMsg;
return false;
}
if (chunksToMerge.size() == 1) {
*errMsg = stream() << "could not merge chunks, collection " << nss.ns()
<< " already contains chunk for " << rangeToString(minKey, maxKey);
warning() << *errMsg;
return false;
}
// Look for hole in range
for (size_t i = 1; i < chunksToMerge.size(); ++i) {
if (chunksToMerge[i - 1].getMax().woCompare(chunksToMerge[i].getMin()) != 0) {
*errMsg =
stream() << "could not merge chunks, collection " << nss.ns()
<< " has a hole in the range " << rangeToString(minKey, maxKey) << " at "
<< rangeToString(chunksToMerge[i - 1].getMax(), chunksToMerge[i].getMin());
warning() << *errMsg;
return false;
}
}
//
// Run apply ops command
//
Status applyOpsStatus = runApplyOpsCmd(txn, chunksToMerge, shardVersion, mergeVersion);
if (!applyOpsStatus.isOK()) {
warning() << applyOpsStatus;
return false;
}
//
// Install merged chunk metadata
//
{
ScopedTransaction transaction(txn, MODE_IX);
Lock::DBLock writeLk(txn->lockState(), nss.db(), MODE_IX);
Lock::CollectionLock collLock(txn->lockState(), nss.ns(), MODE_X);
shardingState->mergeChunks(txn, nss.ns(), minKey, maxKey, mergeVersion);
}
//
// Log change
//
BSONObj mergeLogEntry = buildMergeLogEntry(chunksToMerge, shardVersion, mergeVersion);
grid.catalogManager(txn)->logChange(txn, "merge", nss.ns(), mergeLogEntry);
return true;
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
// Steps
// 1. check basic config
// 2. extract params from command
// 3. fast check
// 4. slow check (LOCKS)
// step 1
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
// make sure we have the mongos id for writebacks
if ( ! checkMongosID( info , cmdObj["serverID"] , errmsg ) )
return false;
bool authoritative = cmdObj.getBoolField( "authoritative" );
// check config server is ok or enable sharding
if ( ! checkConfigOrInit( cmdObj["configdb"].valuestrsafe() , authoritative , errmsg , result ) )
return false;
// check shard name/hosts are correct
if ( cmdObj["shard"].type() == String ) {
shardingState.gotShardName( cmdObj["shard"].String() );
shardingState.gotShardHost( cmdObj["shardHost"].String() );
}
// Handle initial shard connection
if( cmdObj["version"].eoo() && cmdObj["init"].trueValue() ) {
result.append( "initialized", true );
return true;
}
// we can run on a slave up to here
if ( ! isMaster( "admin" ) ) {
result.append( "errmsg" , "not master" );
result.append( "note" , "from post init in setShardVersion" );
return false;
}
// step 2
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to specify namespace";
return false;
}
const ConfigVersion version = ConfigVersion( extractVersion( cmdObj["version"] , errmsg ), OID() );
if ( errmsg.size() )
return false;
// step 3
const ConfigVersion oldVersion = info->getVersion(ns);
const ConfigVersion globalVersion = shardingState.getVersion(ns);
oldVersion.addToBSON( result, "oldVersion" );
if ( globalVersion.isSet() && version.isSet() ) {
// this means there is no reset going on an either side
// so its safe to make some assumptions
if ( version.isEquivalentTo( globalVersion ) ) {
// mongos and mongod agree!
if ( ! oldVersion.isEquivalentTo( version ) ) {
if ( oldVersion < globalVersion ) {
info->setVersion( ns , version );
}
else if ( authoritative ) {
// this means there was a drop and our version is reset
info->setVersion( ns , version );
}
else {
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "verifying drop on '" + ns + "'";
return false;
}
}
return true;
}
}
// step 4
// this is because of a weird segfault I saw and I can't see why this should ever be set
massert( 13647 , str::stream() << "context should be empty here, is: " << cc().getContext()->ns() , cc().getContext() == 0 );
Lock::GlobalWrite setShardVersionLock; // TODO: can we get rid of this??
if ( oldVersion.isSet() && ! globalVersion.isSet() ) {
// this had been reset
info->setVersion( ns , ShardChunkVersion( 0, OID() ) );
}
if ( ! version.isSet() && ! globalVersion.isSet() ) {
// this connection is cleaning itself
info->setVersion( ns , ShardChunkVersion( 0, OID() ) );
return true;
}
if ( ! version.isSet() && globalVersion.isSet() ) {
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
globalVersion.addToBSON( result, "globalVersion" );
errmsg = "dropping needs to be authoritative";
return false;
}
log() << "wiping data for: " << ns << endl;
globalVersion.addToBSON( result, "beforeDrop" );
// only setting global version on purpose
// need clients to re-find meta-data
shardingState.resetVersion( ns );
info->setVersion( ns , ShardChunkVersion( 0, OID() ) );
return true;
}
if ( version < oldVersion ) {
errmsg = "this connection already had a newer version of collection '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "newVersion" );
globalVersion.addToBSON( result, "globalVersion" );
return false;
}
if ( version < globalVersion ) {
while ( shardingState.inCriticalMigrateSection() ) {
dbtemprelease r;
sleepmillis(2);
OCCASIONALLY log() << "waiting till out of critical section" << endl;
}
errmsg = "shard global version for collection is higher than trying to set to '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "version" );
globalVersion.addToBSON( result, "globalVersion" );
result.appendBool( "reloadConfig" , true );
return false;
}
if ( ! globalVersion.isSet() && ! authoritative ) {
// Needed b/c when the last chunk is moved off a shard, the version gets reset to zero, which
// should require a reload.
// TODO: Maybe a more elegant way of doing this
while ( shardingState.inCriticalMigrateSection() ) {
dbtemprelease r;
sleepmillis(2);
OCCASIONALLY log() << "waiting till out of critical section for version reset" << endl;
}
// need authoritative for first look
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "first time for collection '" + ns + "'";
return false;
}
Timer relockTime;
{
dbtemprelease unlock;
ShardChunkVersion currVersion = version;
if ( ! shardingState.trySetVersion( ns , currVersion ) ) {
errmsg = str::stream() << "client version differs from config's for collection '" << ns << "'";
result.append( "ns" , ns );
version.addToBSON( result, "version" );
globalVersion.addToBSON( result, "globalVersion" );
return false;
}
}
if ( relockTime.millis() >= ( cmdLine.slowMS - 10 ) ) {
log() << "setShardVersion - relocking slow: " << relockTime.millis() << endl;
}
info->setVersion( ns , version );
return true;
}
namespace mongo {
// -----ShardingState START ----
ShardingState::ShardingState()
: _enabled(false) , _mutex( "ShardingState" ),
_configServerTickets( 3 /* max number of concurrent config server refresh threads */ ) {
}
void ShardingState::enable( const string& server ) {
_enabled = true;
verify( server.size() );
if ( _configServer.size() == 0 )
_configServer = server;
else {
verify( server == _configServer );
}
}
void ShardingState::gotShardName( const string& name ) {
scoped_lock lk(_mutex);
if ( _shardName.size() == 0 ) {
// TODO SERVER-2299 verify the name is sound w.r.t IPs
_shardName = name;
return;
}
if ( _shardName == name )
return;
stringstream ss;
ss << "gotShardName different than what i had before "
<< " before [" << _shardName << "] "
<< " got [" << name << "] "
;
msgasserted( 13298 , ss.str() );
}
void ShardingState::gotShardHost( string host ) {
scoped_lock lk(_mutex);
size_t slash = host.find( '/' );
if ( slash != string::npos )
host = host.substr( 0 , slash );
if ( _shardHost.size() == 0 ) {
_shardHost = host;
return;
}
if ( _shardHost == host )
return;
stringstream ss;
ss << "gotShardHost different than what i had before "
<< " before [" << _shardHost << "] "
<< " got [" << host << "] "
;
msgasserted( 13299 , ss.str() );
}
void ShardingState::resetShardingState() {
scoped_lock lk(_mutex);
_enabled = false;
_configServer.clear();
_shardName.clear();
_shardHost.clear();
_chunks.clear();
}
// TODO we shouldn't need three ways for checking the version. Fix this.
bool ShardingState::hasVersion( const string& ns ) {
scoped_lock lk(_mutex);
ChunkManagersMap::const_iterator it = _chunks.find(ns);
return it != _chunks.end();
}
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;
}
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 ConfigVersion( 0, OID() );
}
}
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 );
verify( it != _chunks.end() ) ;
ShardChunkManagerPtr p = it->second;
// empty shards should have version 0
version = ( p->getNumChunks() > 1 ) ? version : ShardChunkVersion( 0 , OID() );
ShardChunkManagerPtr cloned( p->cloneMinus( min , max , version ) );
_chunks[ns] = cloned;
}
void ShardingState::undoDonateChunk( const string& ns , const BSONObj& min , const BSONObj& max , ShardChunkVersion version ) {
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
verify( it != _chunks.end() ) ;
ShardChunkManagerPtr p( it->second->clonePlus( min , max , version ) );
_chunks[ns] = p;
}
void ShardingState::splitChunk( const string& ns , const BSONObj& min , const BSONObj& max , const vector<BSONObj>& splitKeys ,
ShardChunkVersion version ) {
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
verify( it != _chunks.end() ) ;
ShardChunkManagerPtr p( it->second->cloneSplit( min , max , splitKeys , version ) );
_chunks[ns] = p;
}
void ShardingState::resetVersion( const string& ns ) {
scoped_lock lk( _mutex );
_chunks.erase( ns );
}
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() ) currManager = it->second;
if ( it != _chunks.end() && ( 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;
ShardChunkManagerPtr p( new ShardChunkManager( c , ns , _shardName, currManager ) );
{
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;
}
ShardChunkVersion oldVersion = version;
version = p->getVersion();
return oldVersion.isEquivalentTo( version );
}
}
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().toLong() );
}
bb.done();
}
}
bool ShardingState::needShardChunkManager( const string& ns ) const {
if ( ! _enabled )
return false;
if ( ! ShardedConnectionInfo::get( false ) )
return false;
return true;
}
ShardChunkManagerPtr ShardingState::getShardChunkManager( const string& ns ) {
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if ( it == _chunks.end() ) {
return ShardChunkManagerPtr();
}
else {
return it->second;
}
}
ShardingState shardingState;
// -----ShardingState END ----
// -----ShardedConnectionInfo START ----
boost::thread_specific_ptr<ShardedConnectionInfo> ShardedConnectionInfo::_tl;
ShardedConnectionInfo::ShardedConnectionInfo() {
_forceVersionOk = false;
_id.clear();
}
ShardedConnectionInfo* ShardedConnectionInfo::get( bool create ) {
ShardedConnectionInfo* info = _tl.get();
if ( ! info && create ) {
LOG(1) << "entering shard mode for connection" << endl;
info = new ShardedConnectionInfo();
_tl.reset( info );
}
return info;
}
void ShardedConnectionInfo::reset() {
_tl.reset();
}
const ConfigVersion ShardedConnectionInfo::getVersion( const string& ns ) const {
NSVersionMap::const_iterator it = _versions.find( ns );
if ( it != _versions.end() ) {
return it->second;
}
else {
return ConfigVersion( 0, OID() );
}
}
void ShardedConnectionInfo::setVersion( const string& ns , const ConfigVersion& version ) {
_versions[ns] = version;
}
void ShardedConnectionInfo::addHook() {
static bool done = false;
if (!done) {
LOG(1) << "adding sharding hook" << endl;
pool.addHook(new ShardingConnectionHook(false));
shardConnectionPool.addHook(new ShardingConnectionHook(true));
done = true;
}
}
void ShardedConnectionInfo::setID( const OID& id ) {
_id = id;
}
// -----ShardedConnectionInfo END ----
unsigned long long extractVersion( BSONElement e , string& errmsg ) {
if ( e.eoo() ) {
errmsg = "no version";
return 0;
}
if ( e.isNumber() )
return (unsigned long long)e.number();
if ( e.type() == Date || e.type() == Timestamp )
return e._numberLong();
errmsg = "version is not a numeric type";
return 0;
}
class MongodShardCommand : public Command {
public:
MongodShardCommand( const char * n ) : Command( n ) {
}
virtual bool slaveOk() const {
return false;
}
virtual bool adminOnly() const {
return true;
}
};
bool haveLocalShardingInfo( const string& ns ) {
if ( ! shardingState.enabled() )
return false;
if ( ! shardingState.hasVersion( ns ) )
return false;
return ShardedConnectionInfo::get(false) > 0;
}
class UnsetShardingCommand : public MongodShardCommand {
public:
UnsetShardingCommand() : MongodShardCommand("unsetSharding") {}
virtual void help( stringstream& help ) const {
help << " example: { unsetSharding : 1 } ";
}
virtual LockType locktype() const {
return NONE;
}
virtual bool slaveOk() const {
return true;
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
ShardedConnectionInfo::reset();
return true;
}
} unsetShardingCommand;
class SetShardVersion : public MongodShardCommand {
public:
SetShardVersion() : MongodShardCommand("setShardVersion") {}
virtual void help( stringstream& help ) const {
help << " example: { setShardVersion : 'alleyinsider.foo' , version : 1 , configdb : '' } ";
}
virtual bool slaveOk() const {
return true;
}
virtual LockType locktype() const {
return NONE;
}
bool checkConfigOrInit( const string& configdb , bool authoritative , string& errmsg , BSONObjBuilder& result , bool locked=false ) const {
if ( configdb.size() == 0 ) {
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ) {
if ( configdb == shardingState.getConfigServer() )
return true;
result.append( "configdb" , BSON( "stored" << shardingState.getConfigServer() <<
"given" << configdb ) );
errmsg = str::stream() << "mongos specified a different config database string : "
<< "stored : " << shardingState.getConfigServer()
<< " vs given : " << configdb;
return false;
}
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
if ( locked ) {
ShardedConnectionInfo::addHook();
shardingState.enable( configdb );
configServer.init( configdb );
return true;
}
Lock::GlobalWrite lk;
return checkConfigOrInit( configdb , authoritative , errmsg , result , true );
}
bool checkMongosID( ShardedConnectionInfo* info, const BSONElement& id, string& errmsg ) {
if ( id.type() != jstOID ) {
if ( ! info->hasID() ) {
warning() << "bad serverID set in setShardVersion and none in info: " << id << endl;
}
// TODO: fix this
//errmsg = "need serverID to be an OID";
//return 0;
return true;
}
OID clientId = id.__oid();
if ( ! info->hasID() ) {
info->setID( clientId );
return true;
}
if ( clientId != info->getID() ) {
errmsg = "server id has changed!";
return false;
}
return true;
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
// Steps
// 1. check basic config
// 2. extract params from command
// 3. fast check
// 4. slow check (LOCKS)
// step 1
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
// make sure we have the mongos id for writebacks
if ( ! checkMongosID( info , cmdObj["serverID"] , errmsg ) )
return false;
bool authoritative = cmdObj.getBoolField( "authoritative" );
// check config server is ok or enable sharding
if ( ! checkConfigOrInit( cmdObj["configdb"].valuestrsafe() , authoritative , errmsg , result ) )
return false;
// check shard name/hosts are correct
if ( cmdObj["shard"].type() == String ) {
shardingState.gotShardName( cmdObj["shard"].String() );
shardingState.gotShardHost( cmdObj["shardHost"].String() );
}
// Handle initial shard connection
if( cmdObj["version"].eoo() && cmdObj["init"].trueValue() ) {
result.append( "initialized", true );
return true;
}
// we can run on a slave up to here
if ( ! isMaster( "admin" ) ) {
result.append( "errmsg" , "not master" );
result.append( "note" , "from post init in setShardVersion" );
return false;
}
// step 2
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to specify namespace";
return false;
}
const ConfigVersion version = ConfigVersion( extractVersion( cmdObj["version"] , errmsg ), OID() );
if ( errmsg.size() )
return false;
// step 3
const ConfigVersion oldVersion = info->getVersion(ns);
const ConfigVersion globalVersion = shardingState.getVersion(ns);
oldVersion.addToBSON( result, "oldVersion" );
if ( globalVersion.isSet() && version.isSet() ) {
// this means there is no reset going on an either side
// so its safe to make some assumptions
if ( version.isEquivalentTo( globalVersion ) ) {
// mongos and mongod agree!
if ( ! oldVersion.isEquivalentTo( version ) ) {
if ( oldVersion < globalVersion ) {
info->setVersion( ns , version );
}
else if ( authoritative ) {
// this means there was a drop and our version is reset
info->setVersion( ns , version );
}
else {
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "verifying drop on '" + ns + "'";
return false;
}
}
return true;
}
}
// step 4
// this is because of a weird segfault I saw and I can't see why this should ever be set
massert( 13647 , str::stream() << "context should be empty here, is: " << cc().getContext()->ns() , cc().getContext() == 0 );
Lock::GlobalWrite setShardVersionLock; // TODO: can we get rid of this??
if ( oldVersion.isSet() && ! globalVersion.isSet() ) {
// this had been reset
info->setVersion( ns , ShardChunkVersion( 0, OID() ) );
}
if ( ! version.isSet() && ! globalVersion.isSet() ) {
// this connection is cleaning itself
info->setVersion( ns , ShardChunkVersion( 0, OID() ) );
return true;
}
if ( ! version.isSet() && globalVersion.isSet() ) {
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
globalVersion.addToBSON( result, "globalVersion" );
errmsg = "dropping needs to be authoritative";
return false;
}
log() << "wiping data for: " << ns << endl;
globalVersion.addToBSON( result, "beforeDrop" );
// only setting global version on purpose
// need clients to re-find meta-data
shardingState.resetVersion( ns );
info->setVersion( ns , ShardChunkVersion( 0, OID() ) );
return true;
}
if ( version < oldVersion ) {
errmsg = "this connection already had a newer version of collection '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "newVersion" );
globalVersion.addToBSON( result, "globalVersion" );
return false;
}
if ( version < globalVersion ) {
while ( shardingState.inCriticalMigrateSection() ) {
dbtemprelease r;
sleepmillis(2);
OCCASIONALLY log() << "waiting till out of critical section" << endl;
}
errmsg = "shard global version for collection is higher than trying to set to '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "version" );
globalVersion.addToBSON( result, "globalVersion" );
result.appendBool( "reloadConfig" , true );
return false;
}
if ( ! globalVersion.isSet() && ! authoritative ) {
// Needed b/c when the last chunk is moved off a shard, the version gets reset to zero, which
// should require a reload.
// TODO: Maybe a more elegant way of doing this
while ( shardingState.inCriticalMigrateSection() ) {
dbtemprelease r;
sleepmillis(2);
OCCASIONALLY log() << "waiting till out of critical section for version reset" << endl;
}
// need authoritative for first look
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "first time for collection '" + ns + "'";
return false;
}
Timer relockTime;
{
dbtemprelease unlock;
ShardChunkVersion currVersion = version;
if ( ! shardingState.trySetVersion( ns , currVersion ) ) {
errmsg = str::stream() << "client version differs from config's for collection '" << ns << "'";
result.append( "ns" , ns );
version.addToBSON( result, "version" );
globalVersion.addToBSON( result, "globalVersion" );
return false;
}
}
if ( relockTime.millis() >= ( cmdLine.slowMS - 10 ) ) {
log() << "setShardVersion - relocking slow: " << relockTime.millis() << endl;
}
info->setVersion( ns , version );
return true;
}
} setShardVersionCmd;
class GetShardVersion : public MongodShardCommand {
public:
GetShardVersion() : MongodShardCommand("getShardVersion") {}
virtual void help( stringstream& help ) const {
help << " example: { getShardVersion : 'alleyinsider.foo' } ";
}
virtual LockType locktype() const {
return NONE;
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
string ns = cmdObj["getShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to specify full namespace";
return false;
}
result.append( "configServer" , shardingState.getConfigServer() );
result.appendTimestamp( "global" , shardingState.getVersion(ns).toLong() );
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
result.appendBool( "inShardedMode" , info != 0 );
if ( info )
result.appendTimestamp( "mine" , info->getVersion(ns).toLong() );
else
result.appendTimestamp( "mine" , 0 );
return true;
}
} getShardVersion;
class ShardingStateCmd : public MongodShardCommand {
public:
ShardingStateCmd() : MongodShardCommand( "shardingState" ) {}
virtual LockType locktype() const {
return WRITE; // TODO: figure out how to make this not need to lock
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
shardingState.appendInfo( result );
return true;
}
} shardingStateCmd;
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , string& errmsg, ConfigVersion& received, ConfigVersion& wanted ) {
if ( ! shardingState.enabled() )
return true;
if ( ! isMasterNs( ns.c_str() ) ) {
// right now connections to secondaries aren't versioned at all
return true;
}
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ) {
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ) {
return true;
}
// TODO
// all collections at some point, be sharded or not, will have a version (and a ShardChunkManager)
// for now, we remove the sharding state of dropped collection
// so delayed request may come in. This has to be fixed.
ConfigVersion clientVersion = info->getVersion(ns);
ConfigVersion version;
if ( ! shardingState.hasVersion( ns , version ) && ! clientVersion.isSet() ) {
return true;
}
// The versions we're going to compare, saved for future use
received = clientVersion;
wanted = version;
if ( ! version.isSet() && clientVersion.isSet() ) {
stringstream ss;
ss << "collection was dropped or this shard no longer valid version";
errmsg = ss.str();
return false;
}
if ( clientVersion >= version )
return true;
if ( ! clientVersion.isSet() ) {
stringstream ss;
ss << "client in sharded mode, but doesn't have version set for this collection";
errmsg = ss.str();
return false;
}
if ( version.majorVersion() == clientVersion.majorVersion() ) {
// this means there was just a split
// since on a split w/o a migrate this server is ok
// going to accept
return true;
}
stringstream ss;
ss << "your version is too old";
errmsg = ss.str();
return false;
}
void ShardingConnectionHook::onHandedOut( DBClientBase * conn ) {
// no-op for mongod
}
}
bool run(const string& , BSONObj& cmdObj, string& errmsg, BSONObjBuilder& result, bool) {
// Steps
// 1. check basic config
// 2. extract params from command
// 3. fast check
// 4. slow check (LOCKS)
// step 1
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
bool authoritative = cmdObj.getBoolField( "authoritative" );
// check config server is ok or enable sharding
if ( ! checkConfigOrInit( cmdObj["configdb"].valuestrsafe() , authoritative , errmsg , result ) )
return false;
// check shard name/hosts are correct
if ( cmdObj["shard"].type() == String ) {
shardingState.gotShardName( cmdObj["shard"].String() );
shardingState.gotShardHost( cmdObj["shardHost"].String() );
}
// make sure we have the mongos id for writebacks
if ( ! checkMongosID( info , cmdObj["serverID"] , errmsg ) )
return false;
// step 2
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to speciy namespace";
return false;
}
const ConfigVersion version = extractVersion( cmdObj["version"] , errmsg );
if ( errmsg.size() )
return false;
// step 3
const ConfigVersion oldVersion = info->getVersion(ns);
const ConfigVersion globalVersion = shardingState.getVersion(ns);
result.appendTimestamp( "oldVersion" , oldVersion );
if ( globalVersion > 0 && version > 0 ) {
// this means there is no reset going on an either side
// so its safe to make some assuptions
if ( version == globalVersion ) {
// mongos and mongod agree!
if ( oldVersion != version ) {
assert( oldVersion < globalVersion );
info->setVersion( ns , version );
}
return true;
}
}
// step 4
// this is because of a weird segfault I saw and I can't see why this should ever be set
massert( 13647 , str::stream() << "context should be empty here, is: " << cc().getContext()->ns() , cc().getContext() == 0 );
dblock setShardVersionLock; // TODO: can we get rid of this??
if ( oldVersion > 0 && globalVersion == 0 ) {
// this had been reset
info->setVersion( ns , 0 );
}
if ( version == 0 && globalVersion == 0 ) {
// this connection is cleaning itself
info->setVersion( ns , 0 );
return true;
}
if ( version == 0 && globalVersion > 0 ) {
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
result.appendTimestamp( "globalVersion" , globalVersion );
errmsg = "dropping needs to be authoritative";
return false;
}
log() << "wiping data for: " << ns << endl;
result.appendTimestamp( "beforeDrop" , globalVersion );
// only setting global version on purpose
// need clients to re-find meta-data
shardingState.resetVersion( ns );
info->setVersion( ns , 0 );
return true;
}
if ( version < oldVersion ) {
errmsg = "this connection already had a newer version of collection '" + ns + "'";
result.append( "ns" , ns );
result.appendTimestamp( "newVersion" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
return false;
}
if ( version < globalVersion ) {
while ( shardingState.inCriticalMigrateSection() ) {
dbtemprelease r;
sleepmillis(2);
OCCASIONALLY log() << "waiting till out of critical section" << endl;
}
errmsg = "shard global version for collection is higher than trying to set to '" + ns + "'";
result.append( "ns" , ns );
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , globalVersion );
result.appendBool( "reloadConfig" , true );
return false;
}
if ( globalVersion == 0 && ! authoritative ) {
// need authoritative for first look
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "first time for collection '" + ns + "'";
return false;
}
Timer relockTime;
{
dbtemprelease unlock;
ShardChunkVersion currVersion = version;
if ( ! shardingState.trySetVersion( ns , currVersion ) ) {
errmsg = str::stream() << "client version differs from config's for colleciton '" << ns << "'";
result.append( "ns" , ns );
result.appendTimestamp( "version" , version );
result.appendTimestamp( "globalVersion" , currVersion );
return false;
}
}
if ( relockTime.millis() >= ( cmdLine.slowMS - 10 ) ) {
log() << "setShardVersion - relocking slow: " << relockTime.millis() << endl;
}
info->setVersion( ns , version );
return true;
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
// Steps
// 1. check basic config
// 2. extract params from command
// 3. fast check
// 4. slow check (LOCKS)
// step 1
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
// make sure we have the mongos id for writebacks
if ( ! checkMongosID( info , cmdObj["serverID"] , errmsg ) )
return false;
bool authoritative = cmdObj.getBoolField( "authoritative" );
// check config server is ok or enable sharding
if ( ! checkConfigOrInit( cmdObj["configdb"].valuestrsafe() , authoritative , errmsg , result ) )
return false;
// check shard name/hosts are correct
if ( cmdObj["shard"].type() == String ) {
shardingState.gotShardName( cmdObj["shard"].String() );
}
// Handle initial shard connection
if( cmdObj["version"].eoo() && cmdObj["init"].trueValue() ){
result.append( "initialized", true );
// Send back wire version to let mongos know what protocol we can speak
result.append( "minWireVersion", minWireVersion );
result.append( "maxWireVersion", maxWireVersion );
return true;
}
// we can run on a slave up to here
if ( ! isMaster( "admin" ) ) {
result.append( "errmsg" , "not master" );
result.append( "note" , "from post init in setShardVersion" );
return false;
}
// step 2
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to specify namespace";
return false;
}
if( ! ChunkVersion::canParseBSON( cmdObj, "version" ) ){
errmsg = "need to specify version";
return false;
}
const ChunkVersion version = ChunkVersion::fromBSON( cmdObj, "version" );
// step 3
const ChunkVersion oldVersion = info->getVersion(ns);
const ChunkVersion globalVersion = shardingState.getVersion(ns);
oldVersion.addToBSON( result, "oldVersion" );
if ( globalVersion.isSet() && version.isSet() ) {
// this means there is no reset going on an either side
// so its safe to make some assumptions
if ( version.isWriteCompatibleWith( globalVersion ) ) {
// mongos and mongod agree!
if ( ! oldVersion.isWriteCompatibleWith( version ) ) {
if ( oldVersion < globalVersion &&
oldVersion.hasCompatibleEpoch(globalVersion) )
{
info->setVersion( ns , version );
}
else if ( authoritative ) {
// this means there was a drop and our version is reset
info->setVersion( ns , version );
}
else {
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "verifying drop on '" + ns + "'";
return false;
}
}
return true;
}
}
// step 4
// this is because of a weird segfault I saw and I can't see why this should ever be set
massert( 13647 , str::stream() << "context should be empty here, is: " << cc().getContext()->ns() , cc().getContext() == 0 );
if ( oldVersion.isSet() && ! globalVersion.isSet() ) {
// this had been reset
info->setVersion( ns , ChunkVersion( 0, OID() ) );
}
if ( ! version.isSet() && ! globalVersion.isSet() ) {
// this connection is cleaning itself
info->setVersion( ns , ChunkVersion( 0, OID() ) );
return true;
}
// Cases below all either return OR fall-through to remote metadata reload.
if ( version.isSet() || !globalVersion.isSet() ) {
// Not Dropping
// TODO: Refactor all of this
if ( version < oldVersion && version.hasCompatibleEpoch( oldVersion ) ) {
errmsg = "this connection already had a newer version of collection '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "newVersion" );
globalVersion.addToBSON( result, "globalVersion" );
return false;
}
// TODO: Refactor all of this
if ( version < globalVersion && version.hasCompatibleEpoch( globalVersion ) ) {
while ( shardingState.inCriticalMigrateSection() ) {
log() << "waiting till out of critical section" << endl;
shardingState.waitTillNotInCriticalSection( 10 );
}
errmsg = "shard global version for collection is higher than trying to set to '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "version" );
globalVersion.addToBSON( result, "globalVersion" );
result.appendBool( "reloadConfig" , true );
return false;
}
if ( ! globalVersion.isSet() && ! authoritative ) {
// Needed b/c when the last chunk is moved off a shard, the version gets reset to zero, which
// should require a reload.
while ( shardingState.inCriticalMigrateSection() ) {
log() << "waiting till out of critical section" << endl;
shardingState.waitTillNotInCriticalSection( 10 );
}
// need authoritative for first look
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "first time for collection '" + ns + "'";
return false;
}
// Fall through to metadata reload below
}
else {
// Dropping
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
globalVersion.addToBSON( result, "globalVersion" );
errmsg = "dropping needs to be authoritative";
return false;
}
// Fall through to metadata reload below
}
ChunkVersion currVersion;
Status status = shardingState.refreshMetadataIfNeeded( ns, version, &currVersion );
if (!status.isOK()) {
// The reload itself was interrupted or confused here
errmsg = str::stream() << "could not refresh metadata for " << ns
<< " with requested shard version " << version.toString()
<< ", stored shard version is " << currVersion.toString()
<< causedBy( status.reason() );
warning() << errmsg << endl;
result.append( "ns" , ns );
version.addToBSON( result, "version" );
currVersion.addToBSON( result, "globalVersion" );
result.appendBool( "reloadConfig", true );
return false;
}
else if ( !version.isWriteCompatibleWith( currVersion ) ) {
// We reloaded a version that doesn't match the version mongos was trying to
// set.
errmsg = str::stream() << "requested shard version differs from"
<< " config shard version for " << ns
<< ", requested version is " << version.toString()
<< " but found version " << currVersion.toString();
OCCASIONALLY warning() << errmsg << endl;
// WARNING: the exact fields below are important for compatibility with mongos
// version reload.
result.append( "ns" , ns );
currVersion.addToBSON( result, "globalVersion" );
// If this was a reset of a collection or the last chunk moved out, inform mongos to
// do a full reload.
if (currVersion.epoch() != version.epoch() || !currVersion.isSet() ) {
result.appendBool( "reloadConfig", true );
// Zero-version also needed to trigger full mongos reload, sadly
// TODO: Make this saner, and less impactful (full reload on last chunk is bad)
ChunkVersion( 0, 0, OID() ).addToBSON( result, "version" );
// For debugging
version.addToBSON( result, "origVersion" );
}
else {
version.addToBSON( result, "version" );
}
return false;
}
info->setVersion( ns , version );
return true;
}
namespace mongo {
// -----ShardingState START ----
ShardingState::ShardingState()
: _enabled(false) , _mutex( "ShardingState" ),
_configServerTickets( 3 /* max number of concurrent config server refresh threads */ ) {
}
void ShardingState::enable( const string& server ) {
scoped_lock lk(_mutex);
_enabled = true;
verify( server.size() );
if ( _configServer.size() == 0 )
_configServer = server;
else {
verify( server == _configServer );
}
}
void ShardingState::initialize(const string& server) {
ShardedConnectionInfo::addHook();
shardingState.enable(server);
configServer.init(server);
}
bool ShardingState::setShardName( const string& name ) {
scoped_lock lk(_mutex);
if ( _shardName.size() == 0 ) {
// TODO SERVER-2299 remotely verify the name is sound w.r.t IPs
_shardName = name;
string clientAddr = cc().clientAddress(true);
log() << "remote client " << clientAddr << " initialized this host as shard " << name;
return true;
}
if ( _shardName == name )
return true;
string clientAddr = cc().clientAddress(true);
warning() << "remote client " << clientAddr << " tried to initialize this host as shard "
<< name << ", but shard name was previously initialized as " << _shardName;
return false;
}
void ShardingState::gotShardName( const string& name ) {
if ( setShardName( name ) )
return;
string clientAddr = cc().clientAddress(true);
stringstream ss;
// Same error as above, to match for reporting
ss << "remote client " << clientAddr << " tried to initialize this host as shard " << name
<< ", but shard name was previously initialized as " << _shardName;
msgasserted( 13298 , ss.str() );
}
void ShardingState::resetShardingState() {
scoped_lock lk(_mutex);
_enabled = false;
_configServer.clear();
_shardName.clear();
_collMetadata.clear();
}
// TODO we shouldn't need three ways for checking the version. Fix this.
bool ShardingState::hasVersion( const string& ns ) {
scoped_lock lk(_mutex);
CollectionMetadataMap::const_iterator it = _collMetadata.find(ns);
return it != _collMetadata.end();
}
bool ShardingState::hasVersion( const string& ns , ChunkVersion& version ) {
scoped_lock lk(_mutex);
CollectionMetadataMap::const_iterator it = _collMetadata.find(ns);
if ( it == _collMetadata.end() )
return false;
CollectionMetadataPtr p = it->second;
version = p->getShardVersion();
return true;
}
const ChunkVersion ShardingState::getVersion( const string& ns ) const {
scoped_lock lk(_mutex);
CollectionMetadataMap::const_iterator it = _collMetadata.find( ns );
if ( it != _collMetadata.end() ) {
CollectionMetadataPtr p = it->second;
return p->getShardVersion();
}
else {
return ChunkVersion( 0, OID() );
}
}
void ShardingState::donateChunk( const string& ns , const BSONObj& min , const BSONObj& max , ChunkVersion version ) {
scoped_lock lk( _mutex );
CollectionMetadataMap::const_iterator it = _collMetadata.find( ns );
verify( it != _collMetadata.end() ) ;
CollectionMetadataPtr p = it->second;
// empty shards should have version 0
version =
( p->getNumChunks() > 1 ) ?
version : ChunkVersion( 0, 0, p->getCollVersion().epoch() );
ChunkType chunk;
chunk.setMin( min );
chunk.setMax( max );
string errMsg;
CollectionMetadataPtr cloned( p->cloneMigrate( chunk, version, &errMsg ) );
// uassert to match old behavior, TODO: report errors w/o throwing
uassert( 16855, errMsg, NULL != cloned.get() );
// TODO: a bit dangerous to have two different zero-version states - no-metadata and
// no-version
_collMetadata[ns] = cloned;
}
void ShardingState::undoDonateChunk( const string& ns, CollectionMetadataPtr prevMetadata ) {
scoped_lock lk( _mutex );
log() << "ShardingState::undoDonateChunk acquired _mutex" << endl;
CollectionMetadataMap::iterator it = _collMetadata.find( ns );
verify( it != _collMetadata.end() );
it->second = prevMetadata;
}
bool ShardingState::notePending( const string& ns,
const BSONObj& min,
const BSONObj& max,
const OID& epoch,
string* errMsg ) {
scoped_lock lk( _mutex );
CollectionMetadataMap::const_iterator it = _collMetadata.find( ns );
if ( it == _collMetadata.end() ) {
*errMsg = str::stream() << "could not note chunk " << "[" << min << "," << max << ")"
<< " as pending because the local metadata for " << ns
<< " has changed";
return false;
}
CollectionMetadataPtr metadata = it->second;
// This can currently happen because drops aren't synchronized with in-migrations
// The idea for checking this here is that in the future we shouldn't have this problem
if ( metadata->getCollVersion().epoch() != epoch ) {
*errMsg = str::stream() << "could not note chunk " << "[" << min << "," << max << ")"
<< " as pending because the epoch for " << ns
<< " has changed from "
<< epoch << " to " << metadata->getCollVersion().epoch();
return false;
}
ChunkType chunk;
chunk.setMin( min );
chunk.setMax( max );
CollectionMetadataPtr cloned( metadata->clonePlusPending( chunk, errMsg ) );
if ( !cloned ) return false;
_collMetadata[ns] = cloned;
return true;
}
bool ShardingState::forgetPending( const string& ns,
const BSONObj& min,
const BSONObj& max,
const OID& epoch,
string* errMsg ) {
scoped_lock lk( _mutex );
CollectionMetadataMap::const_iterator it = _collMetadata.find( ns );
if ( it == _collMetadata.end() ) {
*errMsg = str::stream() << "no need to forget pending chunk "
<< "[" << min << "," << max << ")"
<< " because the local metadata for " << ns << " has changed";
return false;
}
CollectionMetadataPtr metadata = it->second;
// This can currently happen because drops aren't synchronized with in-migrations
// The idea for checking this here is that in the future we shouldn't have this problem
if ( metadata->getCollVersion().epoch() != epoch ) {
*errMsg = str::stream() << "no need to forget pending chunk "
<< "[" << min << "," << max << ")"
<< " because the epoch for " << ns << " has changed from "
<< epoch << " to " << metadata->getCollVersion().epoch();
return false;
}
ChunkType chunk;
chunk.setMin( min );
chunk.setMax( max );
CollectionMetadataPtr cloned( metadata->cloneMinusPending( chunk, errMsg ) );
if ( !cloned ) return false;
_collMetadata[ns] = cloned;
return true;
}
void ShardingState::splitChunk( const string& ns,
const BSONObj& min,
const BSONObj& max,
const vector<BSONObj>& splitKeys,
ChunkVersion version )
{
scoped_lock lk( _mutex );
CollectionMetadataMap::const_iterator it = _collMetadata.find( ns );
verify( it != _collMetadata.end() ) ;
ChunkType chunk;
chunk.setMin( min );
chunk.setMax( max );
string errMsg;
CollectionMetadataPtr cloned( it->second->cloneSplit( chunk, splitKeys, version, &errMsg ) );
// uassert to match old behavior, TODO: report errors w/o throwing
uassert( 16857, errMsg, NULL != cloned.get() );
_collMetadata[ns] = cloned;
}
void ShardingState::mergeChunks( const string& ns,
const BSONObj& minKey,
const BSONObj& maxKey,
ChunkVersion mergedVersion ) {
scoped_lock lk( _mutex );
CollectionMetadataMap::const_iterator it = _collMetadata.find( ns );
verify( it != _collMetadata.end() );
string errMsg;
CollectionMetadataPtr cloned( it->second->cloneMerge( minKey,
maxKey,
mergedVersion,
&errMsg ) );
// uassert to match old behavior, TODO: report errors w/o throwing
uassert( 17004, errMsg, NULL != cloned.get() );
_collMetadata[ns] = cloned;
}
void ShardingState::resetMetadata( const string& ns ) {
scoped_lock lk( _mutex );
warning() << "resetting metadata for " << ns << ", this should only be used in testing"
<< endl;
_collMetadata.erase( ns );
}
Status ShardingState::refreshMetadataIfNeeded( const string& ns,
const ChunkVersion& reqShardVersion,
ChunkVersion* latestShardVersion )
{
// The _configServerTickets serializes this process such that only a small number of threads
// can try to refresh at the same time.
LOG( 2 ) << "metadata refresh requested for " << ns << " at shard version "
<< reqShardVersion << endl;
//
// Queuing of refresh requests starts here when remote reload is needed. This may take time.
// TODO: Explicitly expose the queuing discipline.
//
_configServerTickets.waitForTicket();
TicketHolderReleaser needTicketFrom( &_configServerTickets );
//
// Fast path - check if the requested version is at a higher version than the current
// metadata version or a different epoch before verifying against config server.
//
CollectionMetadataPtr storedMetadata;
{
scoped_lock lk( _mutex );
CollectionMetadataMap::iterator it = _collMetadata.find( ns );
if ( it != _collMetadata.end() ) storedMetadata = it->second;
}
ChunkVersion storedShardVersion;
if ( storedMetadata ) storedShardVersion = storedMetadata->getShardVersion();
*latestShardVersion = storedShardVersion;
if ( storedShardVersion >= reqShardVersion &&
storedShardVersion.epoch() == reqShardVersion.epoch() ) {
// Don't need to remotely reload if we're in the same epoch with a >= version
return Status::OK();
}
//
// Slow path - remotely reload
//
// Cases:
// A) Initial config load and/or secondary take-over.
// B) Migration TO this shard finished, notified by mongos.
// C) Dropping a collection, notified (currently) by mongos.
// D) Stale client wants to reload metadata with a different *epoch*, so we aren't sure.
if ( storedShardVersion.epoch() != reqShardVersion.epoch() ) {
// Need to remotely reload if our epochs aren't the same, to verify
LOG( 1 ) << "metadata change requested for " << ns << ", from shard version "
<< storedShardVersion << " to " << reqShardVersion
<< ", need to verify with config server" << endl;
}
else {
// Need to remotely reload since our epochs aren't the same but our version is greater
LOG( 1 ) << "metadata version update requested for " << ns
<< ", from shard version " << storedShardVersion << " to " << reqShardVersion
<< ", need to verify with config server" << endl;
}
return doRefreshMetadata( ns, reqShardVersion, true, latestShardVersion );
}
Status ShardingState::refreshMetadataNow( const string& ns, ChunkVersion* latestShardVersion )
{
return doRefreshMetadata( ns, ChunkVersion( 0, 0, OID() ), false, latestShardVersion );
}
Status ShardingState::doRefreshMetadata( const string& ns,
const ChunkVersion& reqShardVersion,
bool useRequestedVersion,
ChunkVersion* latestShardVersion )
{
// The idea here is that we're going to reload the metadata from the config server, but
// we need to do so outside any locks. When we get our result back, if the current metadata
// has changed, we may not be able to install the new metadata.
//
// Get the initial metadata
// No DBLock is needed since the metadata is expected to change during reload.
//
CollectionMetadataPtr beforeMetadata;
string shardName;
{
scoped_lock lk( _mutex );
CollectionMetadataMap::iterator it = _collMetadata.find( ns );
if ( it != _collMetadata.end() ) beforeMetadata = it->second;
shardName = _shardName;
}
ChunkVersion beforeShardVersion;
ChunkVersion beforeCollVersion;
if ( beforeMetadata ) {
beforeShardVersion = beforeMetadata->getShardVersion();
beforeCollVersion = beforeMetadata->getCollVersion();
}
*latestShardVersion = beforeShardVersion;
// We can't reload without a shard name. Must check here before loading, since shard name
// may have changed if we checked it earlier and released the _mutex.
if ( shardName.empty() ) {
string errMsg = str::stream() << "cannot refresh metadata for " << ns
<< " before shard name has been set";
LOG( 0 ) << errMsg << endl;
return Status( ErrorCodes::IllegalOperation, errMsg );
}
//
// Determine whether we need to diff or fully reload
//
bool fullReload = false;
if ( !beforeMetadata ) {
// We don't have any metadata to reload from
fullReload = true;
}
else if ( useRequestedVersion && reqShardVersion.epoch() != beforeShardVersion.epoch() ) {
// It's not useful to use the metadata as a base because we think the epoch will differ
fullReload = true;
}
//
// Load the metadata from the remote server, start construction
//
LOG( 0 ) << "remotely refreshing metadata for " << ns
<< ( useRequestedVersion ?
string( " with requested shard version " ) + reqShardVersion.toString() : "" )
<< ( fullReload ?
", current shard version is " : " based on current shard version " )
<< beforeShardVersion
<< ", current metadata version is " << beforeCollVersion << endl;
string errMsg;
ConnectionString configServerLoc = ConnectionString::parse( _configServer, errMsg );
MetadataLoader mdLoader( configServerLoc );
CollectionMetadata* remoteMetadataRaw = new CollectionMetadata();
CollectionMetadataPtr remoteMetadata( remoteMetadataRaw );
Timer refreshTimer;
Status status =
mdLoader.makeCollectionMetadata( ns,
shardName,
( fullReload ? NULL : beforeMetadata.get() ),
remoteMetadataRaw );
long long refreshMillis = refreshTimer.millis();
if ( status.code() == ErrorCodes::NamespaceNotFound ) {
remoteMetadata.reset();
remoteMetadataRaw = NULL;
}
else if ( !status.isOK() ) {
warning() << "could not remotely refresh metadata for " << ns
<< causedBy( status.reason() ) << endl;
return status;
}
ChunkVersion remoteShardVersion;
ChunkVersion remoteCollVersion;
if ( remoteMetadata ) {
remoteShardVersion = remoteMetadata->getShardVersion();
remoteCollVersion = remoteMetadata->getCollVersion();
}
//
// Get ready to install loaded metadata if needed
//
CollectionMetadataPtr afterMetadata;
ChunkVersion afterShardVersion;
ChunkVersion afterCollVersion;
ChunkVersion::VersionChoice choice;
// If we choose to install the new metadata, this describes the kind of install
enum InstallType {
InstallType_New, InstallType_Update, InstallType_Replace, InstallType_Drop,
InstallType_None
} installType = InstallType_None; // compiler complains otherwise
{
// DBLock needed since we're now potentially changing the metadata, and don't want
// reads/writes to be ongoing.
Lock::DBWrite writeLk( ns );
//
// Get the metadata now that the load has completed
//
scoped_lock lk( _mutex );
CollectionMetadataMap::iterator it = _collMetadata.find( ns );
if ( it != _collMetadata.end() ) afterMetadata = it->second;
if ( afterMetadata ) {
afterShardVersion = afterMetadata->getShardVersion();
afterCollVersion = afterMetadata->getCollVersion();
}
*latestShardVersion = afterShardVersion;
//
// Resolve newer pending chunks with the remote metadata, finish construction
//
status = mdLoader.promotePendingChunks( afterMetadata.get(), remoteMetadataRaw );
if ( !status.isOK() ) {
warning() << "remote metadata for " << ns
<< " is inconsistent with current pending chunks"
<< causedBy( status.reason() ) << endl;
return status;
}
//
// Compare the 'before', 'after', and 'remote' versions/epochs and choose newest
// Zero-epochs (sentinel value for "dropped" collections), are tested by
// !epoch.isSet().
//
choice = ChunkVersion::chooseNewestVersion( beforeCollVersion,
afterCollVersion,
remoteCollVersion );
if ( choice == ChunkVersion::VersionChoice_Remote ) {
dassert(!remoteCollVersion.epoch().isSet() ||
remoteShardVersion >= beforeShardVersion);
if ( !afterCollVersion.epoch().isSet() ) {
// First metadata load
installType = InstallType_New;
dassert( it == _collMetadata.end() );
_collMetadata.insert( make_pair( ns, remoteMetadata ) );
}
else if ( remoteCollVersion.epoch().isSet() &&
remoteCollVersion.epoch() == afterCollVersion.epoch() ) {
// Update to existing metadata
installType = InstallType_Update;
// Invariant: If CollMetadata was not found, version should be have been 0.
dassert( it != _collMetadata.end() );
it->second = remoteMetadata;
}
else if ( remoteCollVersion.epoch().isSet() ) {
// New epoch detected, replacing metadata
installType = InstallType_Replace;
// Invariant: If CollMetadata was not found, version should be have been 0.
dassert( it != _collMetadata.end() );
it->second = remoteMetadata;
}
else {
dassert( !remoteCollVersion.epoch().isSet() );
// Drop detected
installType = InstallType_Drop;
_collMetadata.erase( it );
}
*latestShardVersion = remoteShardVersion;
}
}
// End _mutex
// End DBWrite
//
// Do messaging based on what happened above
//
string versionMsg = str::stream()
<< " (loaded metadata version : " << remoteCollVersion.toString()
<< ( beforeCollVersion.epoch() == afterCollVersion.epoch() ?
string( ", stored version : " ) + afterCollVersion.toString() :
string( ", stored versions : " ) +
beforeCollVersion.toString() + " / " + afterCollVersion.toString() )
<< ", took " << refreshMillis << "ms)";
if ( choice == ChunkVersion::VersionChoice_Unknown ) {
string errMsg =
str::stream() << "need to retry loading metadata for " << ns
<< ", collection may have been dropped or recreated during load"
<< versionMsg;
warning() << errMsg << endl;
return Status( ErrorCodes::RemoteChangeDetected, errMsg );
}
if ( choice == ChunkVersion::VersionChoice_Local ) {
LOG( 0 ) << "newer metadata not found for " << ns << versionMsg << endl;
return Status::OK();
}
dassert( choice == ChunkVersion::VersionChoice_Remote );
switch( installType ) {
case InstallType_New:
LOG( 0 ) << "loaded new metadata for " << ns << versionMsg << endl;
break;
case InstallType_Update:
LOG( 0 ) << "loaded newer metadata for " << ns << versionMsg << endl;
break;
case InstallType_Replace:
LOG( 0 ) << "replacing metadata for " << ns << versionMsg << endl;
break;
case InstallType_Drop:
LOG( 0 ) << "dropping metadata for " << ns << versionMsg << endl;
break;
default:
verify( false );
break;
}
return Status::OK();
}
void ShardingState::appendInfo( BSONObjBuilder& b ) {
b.appendBool( "enabled" , _enabled );
if ( ! _enabled )
return;
b.append( "configServer" , _configServer );
b.append( "shardName" , _shardName );
{
BSONObjBuilder bb( b.subobjStart( "versions" ) );
scoped_lock lk(_mutex);
for ( CollectionMetadataMap::iterator it = _collMetadata.begin(); it != _collMetadata.end(); ++it ) {
CollectionMetadataPtr p = it->second;
bb.appendTimestamp( it->first , p->getShardVersion().toLong() );
}
bb.done();
}
}
bool ShardingState::needCollectionMetadata( const string& ns ) const {
if ( ! _enabled )
return false;
if ( ! ShardedConnectionInfo::get( false ) )
return false;
return true;
}
CollectionMetadataPtr ShardingState::getCollectionMetadata( const string& ns ) {
scoped_lock lk( _mutex );
CollectionMetadataMap::const_iterator it = _collMetadata.find( ns );
if ( it == _collMetadata.end() ) {
return CollectionMetadataPtr();
}
else {
return it->second;
}
}
ShardingState shardingState;
// -----ShardingState END ----
// -----ShardedConnectionInfo START ----
boost::thread_specific_ptr<ShardedConnectionInfo> ShardedConnectionInfo::_tl;
ShardedConnectionInfo::ShardedConnectionInfo() {
_forceVersionOk = false;
_id.clear();
}
ShardedConnectionInfo* ShardedConnectionInfo::get( bool create ) {
ShardedConnectionInfo* info = _tl.get();
if ( ! info && create ) {
LOG(1) << "entering shard mode for connection" << endl;
info = new ShardedConnectionInfo();
_tl.reset( info );
}
return info;
}
void ShardedConnectionInfo::reset() {
_tl.reset();
}
const ChunkVersion ShardedConnectionInfo::getVersion( const string& ns ) const {
NSVersionMap::const_iterator it = _versions.find( ns );
if ( it != _versions.end() ) {
return it->second;
}
else {
return ChunkVersion( 0, OID() );
}
}
void ShardedConnectionInfo::setVersion( const string& ns , const ChunkVersion& version ) {
_versions[ns] = version;
}
void ShardedConnectionInfo::addHook() {
static mongo::mutex lock("ShardedConnectionInfo::addHook mutex");
static bool done = false;
scoped_lock lk(lock);
if (!done) {
log() << "first cluster operation detected, adding sharding hook to enable versioning "
"and authentication to remote servers" << endl;
pool.addHook(new ShardingConnectionHook(false));
shardConnectionPool.addHook(new ShardingConnectionHook(true));
done = true;
}
}
void ShardedConnectionInfo::setID( const OID& id ) {
_id = id;
}
class MongodShardCommand : public Command {
public:
MongodShardCommand( const char * n ) : Command( n ) {
}
virtual bool slaveOk() const {
return false;
}
virtual bool adminOnly() const {
return true;
}
};
bool haveLocalShardingInfo( const string& ns ) {
if ( ! shardingState.enabled() )
return false;
if ( ! shardingState.hasVersion( ns ) )
return false;
return ShardedConnectionInfo::get(false) > 0;
}
class UnsetShardingCommand : public MongodShardCommand {
public:
UnsetShardingCommand() : MongodShardCommand("unsetSharding") {}
virtual void help( stringstream& help ) const {
help << "internal";
}
virtual LockType locktype() const { return NONE; }
virtual bool slaveOk() const { return true; }
virtual void addRequiredPrivileges(const std::string& dbname,
const BSONObj& cmdObj,
std::vector<Privilege>* out) {
ActionSet actions;
actions.addAction(ActionType::internal);
out->push_back(Privilege(ResourcePattern::forClusterResource(), actions));
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
ShardedConnectionInfo::reset();
return true;
}
} unsetShardingCommand;
class SetShardVersion : public MongodShardCommand {
public:
SetShardVersion() : MongodShardCommand("setShardVersion") {}
virtual void help( stringstream& help ) const {
help << "internal";
}
virtual bool slaveOk() const { return true; }
virtual LockType locktype() const { return NONE; }
virtual void addRequiredPrivileges(const std::string& dbname,
const BSONObj& cmdObj,
std::vector<Privilege>* out) {
ActionSet actions;
actions.addAction(ActionType::internal);
out->push_back(Privilege(ResourcePattern::forClusterResource(), actions));
}
bool checkConfigOrInit( const string& configdb , bool authoritative , string& errmsg , BSONObjBuilder& result , bool locked=false ) const {
if ( configdb.size() == 0 ) {
errmsg = "no configdb";
return false;
}
if ( shardingState.enabled() ) {
if ( configdb == shardingState.getConfigServer() )
return true;
result.append( "configdb" , BSON( "stored" << shardingState.getConfigServer() <<
"given" << configdb ) );
errmsg = str::stream() << "mongos specified a different config database string : "
<< "stored : " << shardingState.getConfigServer()
<< " vs given : " << configdb;
return false;
}
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
errmsg = "first setShardVersion";
return false;
}
if ( locked ) {
ShardingState::initialize(configdb);
return true;
}
Lock::GlobalWrite lk;
return checkConfigOrInit( configdb , authoritative , errmsg , result , true );
}
bool checkMongosID( ShardedConnectionInfo* info, const BSONElement& id, string& errmsg ) {
if ( id.type() != jstOID ) {
if ( ! info->hasID() ) {
warning() << "bad serverID set in setShardVersion and none in info: " << id << endl;
}
// TODO: fix this
//errmsg = "need serverID to be an OID";
//return 0;
return true;
}
OID clientId = id.__oid();
if ( ! info->hasID() ) {
info->setID( clientId );
return true;
}
if ( clientId != info->getID() ) {
errmsg = "server id has changed!";
return false;
}
return true;
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
// Steps
// 1. check basic config
// 2. extract params from command
// 3. fast check
// 4. slow check (LOCKS)
// step 1
lastError.disableForCommand();
ShardedConnectionInfo* info = ShardedConnectionInfo::get( true );
// make sure we have the mongos id for writebacks
if ( ! checkMongosID( info , cmdObj["serverID"] , errmsg ) )
return false;
bool authoritative = cmdObj.getBoolField( "authoritative" );
// check config server is ok or enable sharding
if ( ! checkConfigOrInit( cmdObj["configdb"].valuestrsafe() , authoritative , errmsg , result ) )
return false;
// check shard name/hosts are correct
if ( cmdObj["shard"].type() == String ) {
shardingState.gotShardName( cmdObj["shard"].String() );
}
// Handle initial shard connection
if( cmdObj["version"].eoo() && cmdObj["init"].trueValue() ){
result.append( "initialized", true );
// Send back wire version to let mongos know what protocol we can speak
result.append( "minWireVersion", minWireVersion );
result.append( "maxWireVersion", maxWireVersion );
return true;
}
// we can run on a slave up to here
if ( ! isMaster( "admin" ) ) {
result.append( "errmsg" , "not master" );
result.append( "note" , "from post init in setShardVersion" );
return false;
}
// step 2
string ns = cmdObj["setShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to specify namespace";
return false;
}
if( ! ChunkVersion::canParseBSON( cmdObj, "version" ) ){
errmsg = "need to specify version";
return false;
}
const ChunkVersion version = ChunkVersion::fromBSON( cmdObj, "version" );
// step 3
const ChunkVersion oldVersion = info->getVersion(ns);
const ChunkVersion globalVersion = shardingState.getVersion(ns);
oldVersion.addToBSON( result, "oldVersion" );
if ( globalVersion.isSet() && version.isSet() ) {
// this means there is no reset going on an either side
// so its safe to make some assumptions
if ( version.isWriteCompatibleWith( globalVersion ) ) {
// mongos and mongod agree!
if ( ! oldVersion.isWriteCompatibleWith( version ) ) {
if ( oldVersion < globalVersion &&
oldVersion.hasCompatibleEpoch(globalVersion) )
{
info->setVersion( ns , version );
}
else if ( authoritative ) {
// this means there was a drop and our version is reset
info->setVersion( ns , version );
}
else {
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "verifying drop on '" + ns + "'";
return false;
}
}
return true;
}
}
// step 4
// this is because of a weird segfault I saw and I can't see why this should ever be set
massert( 13647 , str::stream() << "context should be empty here, is: " << cc().getContext()->ns() , cc().getContext() == 0 );
if ( oldVersion.isSet() && ! globalVersion.isSet() ) {
// this had been reset
info->setVersion( ns , ChunkVersion( 0, OID() ) );
}
if ( ! version.isSet() && ! globalVersion.isSet() ) {
// this connection is cleaning itself
info->setVersion( ns , ChunkVersion( 0, OID() ) );
return true;
}
// Cases below all either return OR fall-through to remote metadata reload.
if ( version.isSet() || !globalVersion.isSet() ) {
// Not Dropping
// TODO: Refactor all of this
if ( version < oldVersion && version.hasCompatibleEpoch( oldVersion ) ) {
errmsg = "this connection already had a newer version of collection '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "newVersion" );
globalVersion.addToBSON( result, "globalVersion" );
return false;
}
// TODO: Refactor all of this
if ( version < globalVersion && version.hasCompatibleEpoch( globalVersion ) ) {
while ( shardingState.inCriticalMigrateSection() ) {
log() << "waiting till out of critical section" << endl;
shardingState.waitTillNotInCriticalSection( 10 );
}
errmsg = "shard global version for collection is higher than trying to set to '" + ns + "'";
result.append( "ns" , ns );
version.addToBSON( result, "version" );
globalVersion.addToBSON( result, "globalVersion" );
result.appendBool( "reloadConfig" , true );
return false;
}
if ( ! globalVersion.isSet() && ! authoritative ) {
// Needed b/c when the last chunk is moved off a shard, the version gets reset to zero, which
// should require a reload.
while ( shardingState.inCriticalMigrateSection() ) {
log() << "waiting till out of critical section" << endl;
shardingState.waitTillNotInCriticalSection( 10 );
}
// need authoritative for first look
result.append( "ns" , ns );
result.appendBool( "need_authoritative" , true );
errmsg = "first time for collection '" + ns + "'";
return false;
}
// Fall through to metadata reload below
}
else {
// Dropping
if ( ! authoritative ) {
result.appendBool( "need_authoritative" , true );
result.append( "ns" , ns );
globalVersion.addToBSON( result, "globalVersion" );
errmsg = "dropping needs to be authoritative";
return false;
}
// Fall through to metadata reload below
}
ChunkVersion currVersion;
Status status = shardingState.refreshMetadataIfNeeded( ns, version, &currVersion );
if (!status.isOK()) {
// The reload itself was interrupted or confused here
errmsg = str::stream() << "could not refresh metadata for " << ns
<< " with requested shard version " << version.toString()
<< ", stored shard version is " << currVersion.toString()
<< causedBy( status.reason() );
warning() << errmsg << endl;
result.append( "ns" , ns );
version.addToBSON( result, "version" );
currVersion.addToBSON( result, "globalVersion" );
result.appendBool( "reloadConfig", true );
return false;
}
else if ( !version.isWriteCompatibleWith( currVersion ) ) {
// We reloaded a version that doesn't match the version mongos was trying to
// set.
errmsg = str::stream() << "requested shard version differs from"
<< " config shard version for " << ns
<< ", requested version is " << version.toString()
<< " but found version " << currVersion.toString();
OCCASIONALLY warning() << errmsg << endl;
// WARNING: the exact fields below are important for compatibility with mongos
// version reload.
result.append( "ns" , ns );
currVersion.addToBSON( result, "globalVersion" );
// If this was a reset of a collection or the last chunk moved out, inform mongos to
// do a full reload.
if (currVersion.epoch() != version.epoch() || !currVersion.isSet() ) {
result.appendBool( "reloadConfig", true );
// Zero-version also needed to trigger full mongos reload, sadly
// TODO: Make this saner, and less impactful (full reload on last chunk is bad)
ChunkVersion( 0, 0, OID() ).addToBSON( result, "version" );
// For debugging
version.addToBSON( result, "origVersion" );
}
else {
version.addToBSON( result, "version" );
}
return false;
}
info->setVersion( ns , version );
return true;
}
} setShardVersionCmd;
class GetShardVersion : public MongodShardCommand {
public:
GetShardVersion() : MongodShardCommand("getShardVersion") {}
virtual void help( stringstream& help ) const {
help << " example: { getShardVersion : 'alleyinsider.foo' } ";
}
virtual LockType locktype() const { return NONE; }
virtual Status checkAuthForCommand(ClientBasic* client,
const std::string& dbname,
const BSONObj& cmdObj) {
if (!client->getAuthorizationSession()->isAuthorizedForActionsOnResource(
ResourcePattern::forExactNamespace(NamespaceString(parseNs(dbname, cmdObj))),
ActionType::getShardVersion)) {
return Status(ErrorCodes::Unauthorized, "Unauthorized");
}
return Status::OK();
}
virtual std::string parseNs(const std::string& dbname, const BSONObj& cmdObj) const {
return parseNsFullyQualified(dbname, cmdObj);
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
string ns = cmdObj["getShardVersion"].valuestrsafe();
if ( ns.size() == 0 ) {
errmsg = "need to specify full namespace";
return false;
}
result.append( "configServer" , shardingState.getConfigServer() );
result.appendTimestamp( "global" , shardingState.getVersion(ns).toLong() );
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
result.appendBool( "inShardedMode" , info != 0 );
if ( info )
result.appendTimestamp( "mine" , info->getVersion(ns).toLong() );
else
result.appendTimestamp( "mine" , 0 );
if ( cmdObj["fullMetadata"].trueValue() ) {
CollectionMetadataPtr metadata = shardingState.getCollectionMetadata( ns );
if ( metadata ) result.append( "metadata", metadata->toBSON() );
else result.append( "metadata", BSONObj() );
}
return true;
}
} getShardVersion;
class ShardingStateCmd : public MongodShardCommand {
public:
ShardingStateCmd() : MongodShardCommand( "shardingState" ) {}
virtual LockType locktype() const { return WRITE; } // TODO: figure out how to make this not need to lock
virtual void addRequiredPrivileges(const std::string& dbname,
const BSONObj& cmdObj,
std::vector<Privilege>* out) {
ActionSet actions;
actions.addAction(ActionType::shardingState);
out->push_back(Privilege(ResourcePattern::forClusterResource(), actions));
}
bool run(const string& , BSONObj& cmdObj, int, string& errmsg, BSONObjBuilder& result, bool) {
shardingState.appendInfo( result );
return true;
}
} shardingStateCmd;
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , string& errmsg, ChunkVersion& received, ChunkVersion& wanted ) {
if ( ! shardingState.enabled() )
return true;
if ( ! isMasterNs( ns.c_str() ) ) {
// right now connections to secondaries aren't versioned at all
return true;
}
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ) {
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ) {
return true;
}
// TODO : all collections at some point, be sharded or not, will have a version
// (and a CollectionMetadata)
received = info->getVersion( ns );
wanted = shardingState.getVersion( ns );
if( received.isWriteCompatibleWith( wanted ) ) return true;
//
// Figure out exactly why not compatible, send appropriate error message
// The versions themselves are returned in the error, so not needed in messages here
//
// Check epoch first, to send more meaningful message, since other parameters probably
// won't match either
if( ! wanted.hasCompatibleEpoch( received ) ){
errmsg = str::stream() << "version epoch mismatch detected for " << ns << ", "
<< "the collection may have been dropped and recreated";
return false;
}
if( ! wanted.isSet() && received.isSet() ){
errmsg = str::stream() << "this shard no longer contains chunks for " << ns << ", "
<< "the collection may have been dropped";
return false;
}
if( wanted.isSet() && ! received.isSet() ){
errmsg = str::stream() << "this shard contains versioned chunks for " << ns << ", "
<< "but no version set in request";
return false;
}
if( wanted.majorVersion() != received.majorVersion() ){
//
// Could be > or < - wanted is > if this is the source of a migration,
// wanted < if this is the target of a migration
//
errmsg = str::stream() << "version mismatch detected for " << ns << ", "
<< "stored major version " << wanted.majorVersion()
<< " does not match received " << received.majorVersion();
return false;
}
// Those are all the reasons the versions can mismatch
verify( false );
return false;
}
void usingAShardConnection( const string& addr ) {
}
}
/**
* @ return true if not in sharded mode
or if version for this client is ok
*/
bool shardVersionOk( const string& ns , string& errmsg, ChunkVersion& received, ChunkVersion& wanted ) {
if ( ! shardingState.enabled() )
return true;
if ( ! isMasterNs( ns.c_str() ) ) {
// right now connections to secondaries aren't versioned at all
return true;
}
ShardedConnectionInfo* info = ShardedConnectionInfo::get( false );
if ( ! info ) {
// this means the client has nothing sharded
// so this allows direct connections to do whatever they want
// which i think is the correct behavior
return true;
}
if ( info->inForceVersionOkMode() ) {
return true;
}
// TODO : all collections at some point, be sharded or not, will have a version
// (and a CollectionMetadata)
received = info->getVersion( ns );
wanted = shardingState.getVersion( ns );
if( received.isWriteCompatibleWith( wanted ) ) return true;
//
// Figure out exactly why not compatible, send appropriate error message
// The versions themselves are returned in the error, so not needed in messages here
//
// Check epoch first, to send more meaningful message, since other parameters probably
// won't match either
if( ! wanted.hasCompatibleEpoch( received ) ){
errmsg = str::stream() << "version epoch mismatch detected for " << ns << ", "
<< "the collection may have been dropped and recreated";
return false;
}
if( ! wanted.isSet() && received.isSet() ){
errmsg = str::stream() << "this shard no longer contains chunks for " << ns << ", "
<< "the collection may have been dropped";
return false;
}
if( wanted.isSet() && ! received.isSet() ){
errmsg = str::stream() << "this shard contains versioned chunks for " << ns << ", "
<< "but no version set in request";
return false;
}
if( wanted.majorVersion() != received.majorVersion() ){
//
// Could be > or < - wanted is > if this is the source of a migration,
// wanted < if this is the target of a migration
//
errmsg = str::stream() << "version mismatch detected for " << ns << ", "
<< "stored major version " << wanted.majorVersion()
<< " does not match received " << received.majorVersion();
return false;
}
// Those are all the reasons the versions can mismatch
verify( false );
return false;
}
