SERVER-14306 Make sure mongos never requests more results than needed from the shards

4 files changed, 68 insertions, 0 deletions

diff --git a/jstests/sharding/in_memory_sort_limit.js b/jstests/sharding/in_memory_sort_limit.js
new file mode 100644
index 00000000000..427471e3a9e
--- /dev/null
+++ b/jstests/sharding/in_memory_sort_limit.js
@@ -0,0 +1,50 @@
+// Test SERVER-14306.  Do a query directly against a mongod with an in-memory sort and a limit that
+// doesn't cause the in-memory sort limit to be reached, then make sure the same limit also doesn't
+// cause the in-memory sort limit to be reached when running through a mongos.
+(function() {
+     "use strict";
+
+     var st = new ShardingTest({ shards: 2, chunkSize: 1, other: {separateConfig: true}});
+     // Force write commands mode; this is a backport of a new test.
+     st.s0.forceWriteMode("commands");
+     st.shard0.forceWriteMode("commands");
+     st.shard1.forceWriteMode("commands");
+
+     var db = st.s.getDB('test');
+     var mongosCol = db.getCollection('skip');
+     db.adminCommand({ enableSharding: 'test' });
+     db.adminCommand({ shardCollection: 'test.skip', key: { _id: 1 }});
+     // Disable balancing of this collection
+     assert.writeOK(db.getSiblingDB('config').collections.update({_id: 'test.skip'},
+                                                                 {$set: {noBalance: true}}));
+
+
+     var filler = new Array(10000).toString();
+     var bulk = [];
+     // create enough data to exceed 32MB in-memory sort limit.
+     for (var i = 0; i < 20000; i++) {
+         bulk.push({x:i, str:filler});
+     }
+     assert.writeOK(mongosCol.insert(bulk));
+     // Make sure that at least 1 doc is on another shard so that mongos doesn't treat this as a
+     // single-shard query (which doesn't exercise the bug)
+     assert.commandWorked(db.getSiblingDB('admin').runCommand({moveChunk: 'test.skip',
+                                                               find: {_id:1},
+                                                               to: 'shard0001'}));
+
+     var docCount = mongosCol.count();
+     var shardCol = st.shard0.getDB('test').getCollection('skip');
+     var passLimit = 2000;
+     var failLimit = 4000;
+     jsTestLog("Test no error with limit of " + passLimit + " on mongod");
+     assert.eq(passLimit, shardCol.find().sort({x:1}).limit(passLimit).itcount());
+
+     jsTestLog("Test error with limit of " + failLimit + " on mongod");
+     assert.throws( function() {shardCol.find().sort({x:1}).limit(failLimit).itcount(); } );
+
+     jsTestLog("Test no error with limit of " + passLimit + " on mongos");
+     assert.eq(passLimit, mongosCol.find().sort({x:1}).limit(passLimit).itcount());
+
+     jsTestLog("Test error with limit of " + failLimit + " on mongos");
+     assert.throws( function() {mongosCol.find().sort({x:1}).limit(failLimit).itcount(); } );
+ })();
\ No newline at end of file
diff --git a/src/mongo/client/parallel.cpp b/src/mongo/client/parallel.cpp
index 491094a8033..4316c0ef29a 100644
--- a/src/mongo/client/parallel.cpp
+++ b/src/mongo/client/parallel.cpp
@@ -1456,6 +1456,12 @@ namespace mongo {
         _done = true;
     }
 
+    void ParallelSortClusteredCursor::setBatchSize(int newBatchSize) {
+        for ( int i=0; i<_numServers; i++ ) {
+            _cursors[i].setBatchSize(newBatchSize);
+        }
+    }
+
     bool ParallelSortClusteredCursor::more() {
 
         if ( _needToSkip > 0 ) {
diff --git a/src/mongo/client/parallel.h b/src/mongo/client/parallel.h
index c8c5af13be8..a596bddcf7e 100644
--- a/src/mongo/client/parallel.h
+++ b/src/mongo/client/parallel.h
@@ -192,6 +192,11 @@ namespace mongo {
         bool isExplain(){ return _qSpec.isExplain(); }
         bool isVersioned(){ return _qShards.size() == 0; }
 
+        /**
+         * Sets the batch size on all underlying cursors to 'newBatchSize'.
+         */
+        void setBatchSize(int newBatchSize);
+
         bool isSharded();
         ShardPtr getPrimary();
         void getQueryShards( set<Shard>& shards );
@@ -273,6 +278,8 @@ namespace mongo {
 
         BSONObj peek();
 
+        void setBatchSize(int newBatchSize) { _cursor->setBatchSize(newBatchSize); }
+
         DBClientCursor* raw() { return _cursor.get(); }
         ParallelConnectionMetadata* rawMData(){ return _pcmData; }
 
diff --git a/src/mongo/s/cursors.cpp b/src/mongo/s/cursors.cpp
index 4931c9b9459..f4a9f0e4573 100644
--- a/src/mongo/s/cursors.cpp
+++ b/src/mongo/s/cursors.cpp
@@ -136,6 +136,11 @@ namespace mongo {
 
             buffer.appendBuf( (void*)o.objdata() , o.objsize() );
             docCount++;
+            // Ensure that the next batch will never wind up requesting more docs from the shard
+            // than are remaining to satisfy the initial ntoreturn.
+            if (ntoreturn != 0) {
+                _cursor->setBatchSize(ntoreturn - docCount);
+            }
 
             if ( buffer.len() > maxSize ) {
                 break;