Merge pull request #486 from kitsudaiki/fix/fix-cluster-resize

fix: fix cluster-resize
kitsudaiki · Oct 24, 2024 · f2d4d68 · f2d4d68
2 parents 4539446 + 89621b8
commit f2d4d68
Show file tree

Hide file tree

Showing 3 changed files with 13 additions and 5 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -47,6 +47,7 @@
 - fixed memory violations found by the new added ASan-memory-check
 - some cases of invalid http-requests had returned an 200 reponse with empty body, which was fixed
 - changed token-key in helm-chart from configmap to secret
+- fixed memory-corruption in cluster-resize, because a pointer was used after free
 
 ### Removed
 

diff --git a/src/hanami/src/core/processing/cluster_resize.h b/src/hanami/src/core/processing/cluster_resize.h
@@ -115,7 +115,7 @@ resizeBlocks(Hexagon* targetHexagon)
  * @return true, if successful, else false
  */
 inline bool
-createNewSection(Cluster& cluster, Connection* sourceConnection)
+splitSection(Cluster& cluster, Connection* sourceConnection)
 {
     // get origin object
     SourceLocation sourceLoc = getSourceNeuron(sourceConnection->origin, &cluster.hexagons[0]);
@@ -133,9 +133,6 @@ createNewSection(Cluster& cluster, Connection* sourceConnection)
     ItemBuffer<SynapseBlock>* synapseBlockBuffer = &targetHexagon->attachedHost->synapseBlocks;
 
     // get target-connection
-    if (targetHexagon->header.numberOfFreeSections < NUMBER_OF_SYNAPSESECTION / 2) {
-        resizeBlocks(targetHexagon);
-    }
     Connection* targetConnection = searchTargetInHexagon(targetHexagon, *synapseBlockBuffer);
     if (targetConnection == nullptr) {
         Hanami::ErrorContainer error;
@@ -201,6 +198,7 @@ createNewSection(Cluster& cluster,
     }
     targetHexagon->header.numberOfFreeSections--;
     targetHexagon->wasResized = true;
+    cluster.metrics.numberOfSections++;
 
     // initialize new connection
     targetConnection->origin.blockId = blockId;
@@ -248,10 +246,17 @@ updateCluster(Cluster& cluster, Hexagon* hexagon)
 
             if (connection->splitValue > 0.0f) {
                 found = true;
-                createNewSection(cluster, connection);
+                splitSection(cluster, connection);
             }
             connection->splitValue = 0.0f;
         }
+
+        // resize if necessary
+        // IMPORTANT: this must be done at the end, because the resize change the target of the
+        // pointer
+        if (hexagon->header.numberOfFreeSections < NUMBER_OF_SYNAPSESECTION / 2) {
+            resizeBlocks(hexagon);
+        }
     }
     return found;
 }

diff --git a/testing/python_sdk_api/sdk_api_test.py b/testing/python_sdk_api/sdk_api_test.py
@@ -300,6 +300,8 @@ def test_workflow():
                          length=50)
 
         print("\n")
+        result = cluster.get_cluster(token, address, cluster_uuid, False)
+        print(json.dumps(json.loads(result), indent=4))
         result = task.delete_task(token, address, task_uuid, cluster_uuid, False)
 
     # save and reload checkpoint