Imported Upstream version 2.0.4

4 files changed, 30 insertions, 5 deletions

diff --git a/s/d_chunk_manager.cpp b/s/d_chunk_manager.cpp
index 82a06f6..351df15 100644
--- a/s/d_chunk_manager.cpp
+++ b/s/d_chunk_manager.cpp
@@ -40,7 +40,7 @@ namespace mongo {
             conn = direct.get();
         }
         else {
-            scoped.reset( new ScopedDbConnection( configServer ) );
+            scoped.reset( new ScopedDbConnection( configServer, 30.0 ) );
             conn = scoped->get();
         }
 
diff --git a/s/d_logic.h b/s/d_logic.h
index d96f937..dca0ee1 100644
--- a/s/d_logic.h
+++ b/s/d_logic.h
@@ -147,6 +147,9 @@ namespace mongo {
 
         // protects state below
         mutable mongo::mutex _mutex;
+        // protects accessing the config server
+        // Using a ticket holder so we can have multiple redundant tries at any given time
+        mutable TicketHolder _configServerTickets;
 
         // map from a namespace into the ensemble of chunk ranges that are stored in this mongod
         // a ShardChunkManager carries all state we need for a collection at this shard, including its version information
diff --git a/s/d_state.cpp b/s/d_state.cpp
index 929d2a8..638d8c1 100644
--- a/s/d_state.cpp
+++ b/s/d_state.cpp
@@ -45,7 +45,8 @@ namespace mongo {
     // -----ShardingState START ----
 
     ShardingState::ShardingState()
-        : _enabled(false) , _mutex( "ShardingState" ) {
+        : _enabled(false) , _mutex( "ShardingState" ),
+          _configServerTickets( 3 /* max number of concurrent config server refresh threads */ ) {
     }
 
     void ShardingState::enable( const string& server ) {
@@ -183,7 +184,23 @@ namespace mongo {
 
     bool ShardingState::trySetVersion( const string& ns , ConfigVersion& version /* IN-OUT */ ) {
 
-        // fast path - requested version is at the same version as this chunk manager
+        // 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)
@@ -191,12 +208,15 @@ namespace mongo {
         //   + two clients reloaded
         //     one triggered the 'slow path' (below)
         //     when the second's request gets here, the version is already current
+        ConfigVersion storedVersion;
         {
             scoped_lock lk( _mutex );
             ChunkManagersMap::const_iterator it = _chunks.find( ns );
-            if ( it != _chunks.end() && it->second->getVersion() == version )
+            if ( it != _chunks.end() && ( storedVersion = it->second->getVersion() ) == 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
@@ -209,8 +229,10 @@ namespace mongo {
         //     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 ) );
+
         {
             scoped_lock lk( _mutex );
 
diff --git a/s/shardconnection.cpp b/s/shardconnection.cpp
index 89b3839..471e37b 100644
--- a/s/shardconnection.cpp
+++ b/s/shardconnection.cpp
@@ -266,7 +266,7 @@ namespace mongo {
         if ( _conn ) {
             if ( ! _conn->isFailed() ) {
                 /* see done() comments above for why we log this line */
-                log() << "~ScopedDBConnection: _conn != null" << endl;
+                log() << "sharded connection to " << _conn->getServerAddress() << " not being returned to the pool" << endl;
             }
             kill();
         }