don't delete old chunk data until there aren't cursors iterating it SERVER-937

db/client.cpp

@@ -193,7 +193,7 @@ namespace mongo {
         if ( doauth )
             _auth( lockState );
 
-        {
+        if ( _client->_curOp->getOp() != dbGetMore ){ // getMore's are special and should be handled else where
             string errmsg;
             if ( ! shardVersionOk( _ns , errmsg ) ){
                 msgasserted( StaleConfigInContextCode , (string)"shard version not ok in Client::Context: " + errmsg );

db/clientcursor.cpp

@@ -357,6 +357,16 @@ namespace mongo {
         client.shutdown();
     }
 
+    void ClientCursor::find( const string& ns , set<CursorId>& all ){
+        recursive_scoped_lock lock(ccmutex);
+        
+        for ( CCById::iterator i=clientCursorsById.begin(); i!=clientCursorsById.end(); ++i ){
+            if ( i->second->ns == ns )
+                all.insert( i->first );
+        }
+    }
+
+
     ClientCursorMonitor clientCursorMonitor;
 
 } // namespace mongo

db/clientcursor.h

@@ -338,6 +338,8 @@ public:
 
         static void informAboutToDeleteBucket(const DiskLoc& b);
         static void aboutToDelete(const DiskLoc& dl);
+
+        static void find( const string& ns , set<CursorId>& all );
     };
 
     class ClientCursorMonitor : public BackgroundJob {

jstests/sharding/cursor1.js

@@ -38,8 +38,8 @@ s.adminCommand( { movechunk : "test.foo" , find : { _id : 5 } , to : secondary.g
 assert.eq( 2, s.config.chunks.count() );
  
 // the cursors should not have been affected
-assert.eq( numObjs , cursor1.itcount() );
-assert.eq( numObjs , cursor2.itcount() );
-assert.eq( numObjs , cursor3.itcount() );
+assert.eq( numObjs , cursor1.itcount() , "c1" );
+assert.eq( numObjs , cursor2.itcount() , "c2" );
+assert.eq( numObjs , cursor3.itcount() , "c3" );
 
-s.stop()
+s.stop()

s/d_migrate.cpp

@@ -133,7 +133,47 @@ namespace mongo {
         ofstream* _out;
         
     };
+    
+    struct OldDataCleanup {
+        string ns;
+        BSONObj min;
+        BSONObj max;
+    };
+    
+    void cleanupOldData( OldDataCleanup cleanup ){
+        Client::initThread( "moveChunkPostClean" );
+        
+        set<CursorId> initial;
+        ClientCursor::find( cleanup.ns , initial );
+        
+        Timer t;
+        while ( t.seconds() < 600 ){ // 10 minutes
+            sleepmillis( 20 );
 
+            set<CursorId> now;
+            ClientCursor::find( cleanup.ns , now );            
+
+            set<CursorId> left;
+            for ( set<CursorId>::iterator i=initial.begin(); i!=initial.end(); ++i ){
+                CursorId id = *i;
+                if ( now.count(id) )
+                    left.insert( id );
+            }
+            
+            if ( left.size() == 0 )
+                break;
+            initial = left;
+        }
+        
+        ShardForceModeBlock sf;
+        writelock lk(cleanup.ns);
+        RemoveSaver rs(cleanup.ns,"post-cleanup");
+        long long num = Helpers::removeRange( cleanup.ns , cleanup.min , cleanup.max , true , false , &rs );
+        log() << "moveChunk deleted: " << num << endl;
+
+        cc().shutdown();
+    }
+    
     class ChunkCommandHelper : public Command {
     public:
         ChunkCommandHelper( const char * name ) 
@@ -598,17 +638,15 @@ namespace mongo {
             log( LL_WARNING ) << " deleting data before ensuring no more cursors TODO" << endl;
             
             timing.done(6);
-            // 7
-            {
-                ShardForceModeBlock sf;
-                writelock lk(ns);
-                RemoveSaver rs(ns,"post-cleanup");
-                long long num = Helpers::removeRange( ns , min , max , true , false , &rs );
-                log() << "moveChunk deleted: " << num << endl;
-                result.appendNumber( "numDeleted" , num );
+
+            { // 7.
+                OldDataCleanup c;
+                c.ns = ns;
+                c.min = min.getOwned();
+                c.max = max.getOwned();
+                boost::thread t( boost::bind( &cleanupOldData , c ) );
             }
             
-            timing.done(7);
 
             return true;