/**
* @ 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 , 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;
}
/**
* @ 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;
}
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;
}
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 );
}
}
/**
* @ 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)
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;
}
bool ShardingState::trySetVersion( const string& ns , ConfigVersion& version /* IN-OUT */ ) {
// Currently this function is called after a getVersion(), which is the first "check", and the assumption here
// is that we don't do anything nearly as long as a remote query in a thread between then and now.
// Otherwise it may be worth adding an additional check without the _configServerMutex below, since then it
// would be likely that the version may have changed in the meantime without waiting for or fetching config results.
// TODO: Mutex-per-namespace?
LOG( 2 ) << "trying to set shard version of " << version.toString() << " for '" << ns << "'" << endl;
_configServerTickets.waitForTicket();
TicketHolderReleaser needTicketFrom( &_configServerTickets );
// fast path - double-check if requested version is at the same version as this chunk manager before verifying
// against config server
//
// This path will short-circuit the version set if another thread already managed to update the version in the
// meantime. First check is from getVersion().
//
// cases:
// + this shard updated the version for a migrate's commit (FROM side)
// a client reloaded chunk state from config and picked the newest version
// + two clients reloaded
// one triggered the 'slow path' (below)
// when the second's request gets here, the version is already current
ConfigVersion storedVersion;
ShardChunkManagerPtr currManager;
{
scoped_lock lk( _mutex );
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if( it == _chunks.end() ){
// TODO: We need better semantic distinction between *no manager found* and
// *manager of version zero found*
log() << "no current chunk manager found for this shard, will initialize" << endl;
}
else{
currManager = it->second;
if( ( storedVersion = it->second->getVersion() ).isEquivalentTo( version ) )
return true;
}
}
LOG( 2 ) << "verifying cached version " << storedVersion.toString() << " and new version " << version.toString() << " for '" << ns << "'" << endl;
// slow path - requested version is different than the current chunk manager's, if one exists, so must check for
// newest version in the config server
//
// cases:
// + a chunk moved TO here
// (we don't bump up the version on the TO side but the commit to config does use higher version)
// a client reloads from config an issued the request
// + there was a take over from a secondary
// the secondary had no state (managers) at all, so every client request will fall here
// + a stale client request a version that's not current anymore
// Can't lock default mutex while creating ShardChunkManager, b/c may have to create a new connection to myself
const string c = (_configServer == _shardHost) ? "" /* local */ : _configServer;
// If our epochs aren't compatible, it's not useful to use the old manager for chunk diffs
if( currManager && ! currManager->getCollVersion().hasCompatibleEpoch( version ) ){
warning() << "detected incompatible version epoch in new version " << version
<< ", old version was " << currManager->getCollVersion() << endl;
currManager.reset();
}
ShardChunkManagerPtr p( ShardChunkManager::make( c , ns , _shardName, currManager ) );
// Handle the case where the collection isn't sharded more gracefully
if( p->getKey().isEmpty() ){
version = ConfigVersion( 0, OID() );
// There was an error getting any data for this collection, return false
return false;
}
{
// NOTE: This lock prevents the ns version from changing while a write operation occurs.
Lock::DBRead readLk(ns);
// This lock prevents simultaneous metadata changes using the same map
scoped_lock lk( _mutex );
// since we loaded the chunk manager unlocked, other thread may have done the same
// make sure we keep the freshest config info only
ChunkManagersMap::const_iterator it = _chunks.find( ns );
if ( it == _chunks.end() || p->getVersion() >= it->second->getVersion() ) {
_chunks[ns] = p;
}
ChunkVersion oldVersion = version;
version = p->getVersion();
return oldVersion.isEquivalentTo( version );
}
}