Use C++ shared_ptr for IPC file descriptor cleanup

Replace Python-level os.close() in IPCAllocationHandle with a C++
shared_ptr custom deleter that calls POSIX close() directly, avoiding
unraisable exception errors during late interpreter shutdown.

Made-with: Cursor

Author: Andy-Jost

cuda_core/cuda/core/_cpp/resource_handles.cpp

@@ -12,6 +12,10 @@
 #include <unordered_map>
 #include <vector>
 
+#ifndef _WIN32
+#include <unistd.h>
+#endif
+
 namespace cuda_core {
 
 // ============================================================================
@@ -1116,4 +1120,21 @@ CuLinkHandle create_culink_handle_ref(CUlinkState state) {
     return CuLinkHandle(box, &box->resource);
 }
 
+// ============================================================================
+// File Descriptor Handles
+// ============================================================================
+
+#ifndef _WIN32
+FileDescriptorHandle create_fd_handle(int fd) {
+    return FileDescriptorHandle(
+        new int(fd),
+        [](const int* p) { ::close(*p); delete p; }
+    );
+}
+#endif
+
+FileDescriptorHandle create_fd_handle_ref(int fd) {
+    return std::make_shared<const int>(fd);
+}
+
 }  // namespace cuda_core

cuda_core/cuda/core/_cpp/resource_handles.hpp

@@ -154,6 +154,7 @@ using NvrtcProgramHandle = std::shared_ptr<const nvrtcProgram>;
 using NvvmProgramHandle = std::shared_ptr<const NvvmProgramValue>;
 using NvJitLinkHandle = std::shared_ptr<const NvJitLinkValue>;
 using CuLinkHandle = std::shared_ptr<const CUlinkState>;
+using FileDescriptorHandle = std::shared_ptr<const int>;
 
 
 // ============================================================================
@@ -477,6 +478,17 @@ CuLinkHandle create_culink_handle(CUlinkState state);
 // The handle will NOT be destroyed when the last reference is released.
 CuLinkHandle create_culink_handle_ref(CUlinkState state);
 
+// ============================================================================
+// File descriptor handle functions
+// ============================================================================
+
+// Create an owning file descriptor handle.
+// When the last reference is released, POSIX close() is called.
+FileDescriptorHandle create_fd_handle(int fd);
+
+// Create a non-owning file descriptor handle (caller manages the fd).
+FileDescriptorHandle create_fd_handle_ref(int fd);
+
 // ============================================================================
 // Overloaded helper functions to extract raw resources from handles
 // ============================================================================
@@ -596,6 +608,10 @@ inline std::intptr_t as_intptr(const CuLinkHandle& h) noexcept {
     return reinterpret_cast<std::intptr_t>(as_cu(h));
 }
 
+inline std::intptr_t as_intptr(const FileDescriptorHandle& h) noexcept {
+    return h ? static_cast<std::intptr_t>(*h) : -1;
+}
+
 // as_py() - convert handle to Python wrapper object (returns new reference)
 #if PY_VERSION_HEX < 0x030D0000
 extern "C" int _Py_IsFinalizing(void);
@@ -687,4 +703,8 @@ inline PyObject* as_py(const GraphicsResourceHandle& h) noexcept {
     return detail::make_py("cuda.bindings.driver", "CUgraphicsResource", as_intptr(h));
 }
 
+inline PyObject* as_py(const FileDescriptorHandle& h) noexcept {
+    return PyLong_FromSsize_t(as_intptr(h));
+}
+
 }  // namespace cuda_core

cuda_core/cuda/core/_memory/_ipc.pxd

@@ -5,6 +5,7 @@
 from cuda.bindings cimport cydriver
 from cuda.core._memory._buffer cimport Buffer
 from cuda.core._memory._memory_pool cimport _MemPool
+from cuda.core._resource_handles cimport FileDescriptorHandle
 
 
 # Holds _MemPool objects imported by this process.  This enables
@@ -46,8 +47,8 @@ cdef class IPCBufferDescriptor:
 
 cdef class IPCAllocationHandle:
     cdef:
-        int    _handle
-        object _uuid
+        FileDescriptorHandle _h_fd
+        object               _uuid
 
     cpdef close(self)
 

cuda_core/cuda/core/_memory/_ipc.pyx

@@ -10,10 +10,13 @@ from cuda.core._memory._memory_pool cimport _MemPool
 from cuda.core._stream cimport Stream
 from cuda.core._resource_handles cimport (
     DevicePtrHandle,
+    create_fd_handle,
     create_mempool_handle_ipc,
     deviceptr_import_ipc,
     get_last_error,
     as_cu,
+    as_intptr,
+    as_py,
 )
 
 from cuda.core._stream cimport default_stream
@@ -110,31 +113,24 @@ cdef class IPCAllocationHandle:
     def _init(cls, handle: int, uuid):  # no-cython-lint
         cdef IPCAllocationHandle self = IPCAllocationHandle.__new__(cls)
         assert handle >= 0
-        self._handle = handle
+        self._h_fd = create_fd_handle(handle)
         self._uuid = uuid
         return self
 
     cpdef close(self):
         """Close the handle."""
-        if self._handle >= 0:
-            try:
-                os.close(self._handle)
-            finally:
-                self._handle = -1
-
-    def __dealloc__(self):
-        self.close()
+        self._h_fd.reset()
 
     def __int__(self) -> int:
-        if self._handle < 0:
+        if not self._h_fd or as_intptr(self._h_fd) < 0:
             raise ValueError(
                 f"Cannot convert IPCAllocationHandle to int: the handle (id={id(self)}) is closed."
             )
-        return self._handle
+        return as_py(self._h_fd)
 
     @property
     def handle(self) -> int:
-        return self._handle
+        return as_py(self._h_fd)
 
     @property
     def uuid(self) -> uuid.UUID:

cuda_core/cuda/core/_resource_handles.pxd

@@ -41,6 +41,7 @@ cdef extern from "_cpp/resource_handles.hpp" namespace "cuda_core":
     ctypedef shared_ptr[const NvJitLinkValue] NvJitLinkHandle
 
     ctypedef shared_ptr[const cydriver.CUlinkState] CuLinkHandle
+    ctypedef shared_ptr[const int] FileDescriptorHandle
 
     # as_cu() - extract the raw CUDA handle (inline C++)
     cydriver.CUcontext as_cu(ContextHandle h) noexcept nogil
@@ -73,6 +74,7 @@ cdef extern from "_cpp/resource_handles.hpp" namespace "cuda_core":
     intptr_t as_intptr(NvvmProgramHandle h) noexcept nogil
     intptr_t as_intptr(NvJitLinkHandle h) noexcept nogil
     intptr_t as_intptr(CuLinkHandle h) noexcept nogil
+    intptr_t as_intptr(FileDescriptorHandle h) noexcept nogil
 
     # as_py() - convert handle to Python wrapper object (inline C++; requires GIL)
     object as_py(ContextHandle h)
@@ -89,6 +91,7 @@ cdef extern from "_cpp/resource_handles.hpp" namespace "cuda_core":
     object as_py(NvvmProgramHandle h)
     object as_py(NvJitLinkHandle h)
     object as_py(CuLinkHandle h)
+    object as_py(FileDescriptorHandle h)
 
 
 # =============================================================================
@@ -203,3 +206,7 @@ cdef NvJitLinkHandle create_nvjitlink_handle_ref(cynvjitlink.nvJitLinkHandle han
 # cuLink handles
 cdef CuLinkHandle create_culink_handle(cydriver.CUlinkState state) except+ nogil
 cdef CuLinkHandle create_culink_handle_ref(cydriver.CUlinkState state) except+ nogil
+
+# File descriptor handles
+cdef FileDescriptorHandle create_fd_handle(int fd) except+ nogil
+cdef FileDescriptorHandle create_fd_handle_ref(int fd) except+ nogil

cuda_core/cuda/core/_resource_handles.pyx

@@ -191,6 +191,12 @@ cdef extern from "_cpp/resource_handles.hpp" namespace "cuda_core":
     CuLinkHandle create_culink_handle_ref "cuda_core::create_culink_handle_ref" (
         cydriver.CUlinkState state) except+ nogil
 
+    # File descriptor handles
+    FileDescriptorHandle create_fd_handle "cuda_core::create_fd_handle" (
+        int fd) except+ nogil
+    FileDescriptorHandle create_fd_handle_ref "cuda_core::create_fd_handle_ref" (
+        int fd) except+ nogil
+
 
 # =============================================================================
 # CUDA Driver API capsule