Restore previous behavior for DMR.from_allocation_handle when passed a file descriptor (caller closes the fd).

Author: Andy-Jost

cuda_core/cuda/core/experimental/_memory/_dmr.pyx

@@ -276,7 +276,9 @@ cdef class DeviceMemoryResource(MemoryResource):
             resource is created.
 
         alloc_handle : int | IPCAllocationHandle
-            The shareable handle of the device memory resource to import.
+            The shareable handle of the device memory resource to import. If an
+            integer is supplied, it must represent a valid platform-specific
+            handle. It is the caller's responsibility to close that handle.
 
         Returns
         -------

cuda_core/cuda/core/experimental/_memory/_ipc.pyx

@@ -180,13 +180,21 @@ cdef Buffer Buffer_from_ipc_descriptor(
 
 cdef DeviceMemoryResource DMR_from_allocation_handle(cls, device_id, alloc_handle):
     # Quick exit for registry hits.
-    if isinstance(alloc_handle, int):
-        alloc_handle = IPCAllocationHandle._init(alloc_handle, None)
     uuid = getattr(alloc_handle, 'uuid', None)  # no-cython-lint
     mr = registry.get(uuid)
     if mr is not None:
         return mr
 
+    # Ensure we have an allocation handle. Duplicate the file descriptor, if
+    # necessary.
+    if isinstance(alloc_handle, int):
+        fd = os.dup(alloc_handle)
+        try:
+            alloc_handle = IPCAllocationHandle._init(fd, None)
+        except:
+            os.close(fd)
+            raise
+
     # Construct a new DMR.
     cdef DeviceMemoryResource self = DeviceMemoryResource.__new__(cls)
     self._dev_id = getattr(device_id, 'device_id', device_id)
@@ -205,7 +213,7 @@ cdef DeviceMemoryResource DMR_from_allocation_handle(cls, device_id, alloc_handl
         registered = self.register(uuid)
         assert registered is self
 
-    # Always close the file handle (caller can dup it, if needed).
+    # Always close the file handle.
     alloc_handle.close()
 
     return self

cuda_core/tests/memory_ipc/test_serialize.py

@@ -3,6 +3,7 @@
 
 import multiprocessing as mp
 import multiprocessing.reduction
+import os
 
 from cuda.core.experimental import Buffer, Device, DeviceMemoryResource
 from helpers.buffers import PatternGen
@@ -59,6 +60,7 @@ def child_main(self, conn):
         # Receive the memory resource.
         handle = mp.reduction.recv_handle(conn)
         mr = DeviceMemoryResource.from_allocation_handle(device, handle)
+        os.close(handle)
 
         # Receive the buffers.
         buffer1 = conn.recv()  # directly