Fix reference counting in Python

src/IO/AbstractIOHandlerImpl.cpp

@@ -268,7 +268,10 @@ std::future<void> AbstractIOHandlerImpl::flush()
                     i.writable->parent,
                     "->",
                     i.writable,
-                    "] WRITE_DATASET");
+                    "] WRITE_DATASET, offset=",
+                    [&parameter]() { return vec_as_string(parameter.offset); },
+                    ", extent=",
+                    [&parameter]() { return vec_as_string(parameter.extent); });
                 writeDataset(i.writable, parameter);
                 break;
             }

src/binding/python/RecordComponent.cpp

@@ -22,6 +22,7 @@
 #include <pybind11/detail/common.h>
 #include <pybind11/numpy.h>
 #include <pybind11/pybind11.h>
+#include <pybind11/pytypes.h>
 #include <pybind11/stl.h>
 
 #include "openPMD/Dataset.hpp"
@@ -322,13 +323,16 @@ struct StoreChunkFromPythonArray
         Offset const &offset,
         Extent const &extent)
     {
-        // here, we increase a reference on the user-passed data so that
+        a.inc_ref();
+        void *data = a.mutable_data();
+        // here, we store an owning handle in the lambda capture so that
         // temporary and lost-scope variables stay alive until we flush
         // note: this does not yet prevent the user, as in C++, to build
         // a race condition by manipulating the data that was passed
-        a.inc_ref();
-        void *data = a.mutable_data();
-        std::shared_ptr<T> shared((T *)data, [a](T *) { a.dec_ref(); });
+        std::shared_ptr<T> shared(
+            (T *)data, [owning_handle = a.cast<py::object>()](T *) {
+                // no-op
+            });
         r.storeChunk(std::move(shared), offset, extent);
     }
 
@@ -343,13 +347,15 @@ struct LoadChunkIntoPythonArray
         Offset const &offset,
         Extent const &extent)
     {
-        // here, we increase a reference on the user-passed data so that
+        void *data = a.mutable_data();
+        // here, we store an owning handle in the lambda capture so that
         // temporary and lost-scope variables stay alive until we flush
         // note: this does not yet prevent the user, as in C++, to build
         // a race condition by manipulating the data that was passed
-        a.inc_ref();
-        void *data = a.mutable_data();
-        std::shared_ptr<T> shared((T *)data, [a](T *) { a.dec_ref(); });
+        std::shared_ptr<T> shared(
+            (T *)data, [owning_handle = a.cast<py::object>()](T *) {
+                // no-op
+            });
         r.loadChunk(std::move(shared), offset, extent);
     }
 
@@ -365,14 +371,15 @@ struct LoadChunkIntoPythonBuffer
         Offset const &offset,
         Extent const &extent)
     {
-        // here, we increase a reference on the user-passed data so that
+        void *data = buffer_info.ptr;
+        // here, we store an owning handle in the lambda capture so that
         // temporary and lost-scope variables stay alive until we flush
         // note: this does not yet prevent the user, as in C++, to build
         // a race condition by manipulating the data that was passed
-        buffer.inc_ref();
-        void *data = buffer_info.ptr;
         std::shared_ptr<T> shared(
-            (T *)data, [buffer](T *) { buffer.dec_ref(); });
+            (T *)data, [owning_handle = buffer.cast<py::object>()](T *) {
+                // no-op
+            });
         r.loadChunk(std::move(shared), offset, extent);
     }
 

test/python/unittest/API/APITest.py

@@ -82,6 +82,52 @@ def tearDown(self):
         del self.__particle_series
         del self.__series
 
+    # This function exhibits a bug in the old use of refcounting.
+    def refcountingCreateData(self):
+        series = io.Series(
+            "../samples/refcounting.json",
+            io.Access.create_linear,
+        )
+
+        for i in range(10):
+            current_iteration = series.snapshots()[i]
+
+            # First, write an E mesh.
+            E = current_iteration.meshes["E"]
+            E.axis_labels = ["x", "y"]
+            for dim in ["x", "y"]:
+                component = E[dim]
+                component.reset_dataset(
+                    io.Dataset(np.dtype("float"), [10, 10]))
+                component[:, :] = np.reshape(
+                    np.arange(i * 100, (i + 1) * 100, dtype=np.dtype("float")),
+                    [10, 10],
+                )
+
+            # Now, write some e particles.
+            e = current_iteration.particles["e"]
+            for dim in ["x", "y"]:
+                # Do not bother with a positionOffset
+                position_offset = e["positionOffset"][dim]
+                position_offset.make_constant(0)
+
+                position = e["position"][dim]
+                position.reset_dataset(io.Dataset(np.dtype("float"), [100]))
+                position[:] = np.arange(
+                    i * 100, (i + 1) * 100, dtype=np.dtype("float")
+                )
+
+    def testRefCounting(self):
+        self.refcountingCreateData()
+
+        read = io.Series("../samples/refcounting.json", io.Access.read_linear)
+        iteration = read.snapshots()[0]
+        pos_x = iteration.particles["e"]["position"]["x"]
+        loaded = pos_x[:]
+        read.flush()
+        self.assertTrue(np.allclose(
+            loaded, np.arange(0, 100, dtype=np.dtype("float"))))
+
     def testFieldData(self):
         """ Testing serial IO on a pure field dataset. """