Add critical section guards to accessors (NVIDIA#2215)

* Add critical section guards to thread-sensitive accessors

Protect accessors with cython critical sections so free-threaded execution
avoids races when reading cached state.

At least one of these was noticed by free-threaded test runs.  I am not
certain that they will guarantee identity returns, but pretty sure this
will make them thread-safe.
(The reason being that at least on non free-threaded builds they may
release the GIL or, on free-threaded, do API calls that release the
critical section temporarily.)

About the `.pyi` fixes: I think these are incorrect but `cython`
is already used in other places and shouldn't be a runtime/typing requirement
at all.

* Apply suggestion from @seberg

Author: seberg

cuda_core/cuda/core/_memory/_buffer.pyi

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import cython
 from cuda.core._memory._device_memory_resource import DeviceMemoryResource
 from cuda.core._memory._ipc import IPCBufferDescriptor
 from cuda.core._memory._pinned_memory_resource import PinnedMemoryResource
@@ -101,6 +102,7 @@ class Buffer:
         """
 
     @property
+    @cython.critical_section
     def ipc_descriptor(self) -> IPCBufferDescriptor:
         """Descriptor for sharing this buffer with other processes."""
 

cuda_core/cuda/core/_memory/_buffer.pyx

@@ -206,6 +206,7 @@ cdef class Buffer:
         return _ipc.Buffer_from_ipc_descriptor(cls, mr, ipc_descriptor, stream)
 
     @property
+    @cython.critical_section
     def ipc_descriptor(self) -> IPCBufferDescriptor:
         """Descriptor for sharing this buffer with other processes."""
         if self._ipc_data is None:

cuda_core/cuda/core/_memory/_memory_pool.pyi

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import uuid
 
+import cython
 from cuda.core._memory._buffer import Buffer, MemoryResource
 from cuda.core._stream import Stream
 from cuda.core.graph import GraphBuilder
@@ -97,6 +98,7 @@ class _MemPool(MemoryResource):
         """
 
     @property
+    @cython.critical_section
     def attributes(self) -> _MemPoolAttributes:
         """Memory pool attributes."""
 

cuda_core/cuda/core/_memory/_memory_pool.pyx

@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+cimport cython
 from libc.limits cimport ULLONG_MAX
 from libc.stdint cimport uintptr_t
 from libc.string cimport memset
@@ -177,6 +178,7 @@ cdef class _MemPool(MemoryResource):
         _MP_deallocate(self, <uintptr_t>ptr, size, s)
 
     @property
+    @cython.critical_section
     def attributes(self) -> _MemPoolAttributes:
         """Memory pool attributes."""
         if self._attributes is None:

cuda_core/cuda/core/_memoryview.pyx

@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+cimport cython
 from ._dlpack cimport *
 from ._dlpack import classify_dl_device
 from libc.stdint cimport intptr_t
@@ -539,6 +540,7 @@ cdef class StridedMemoryView:
               + f"                  readonly={self.readonly},\n"
               + f"                  exporting_obj={get_simple_repr(self.exporting_obj)})")
 
+    @cython.critical_section
     cdef inline _StridedLayout get_layout(self):
         if self._layout is None:
             if self.dl_tensor:
@@ -549,6 +551,7 @@ cdef class StridedMemoryView:
                 raise ValueError("Cannot infer layout from the exporting object")
         return self._layout
 
+    @cython.critical_section
     cdef inline object get_buffer(self):
         """
         Returns Buffer instance with the underlying data.
@@ -562,6 +565,7 @@ cdef class StridedMemoryView:
                 self._buffer = Buffer.from_handle(self.ptr, 0, owner=self.exporting_obj)
         return self._buffer
 
+    @cython.critical_section
     cdef inline object get_dtype(self):
         if self._dtype is None:
             if self.dl_tensor != NULL:

cuda_core/cuda/core/_module.pyi

@@ -5,6 +5,7 @@ from __future__ import annotations
 from collections import namedtuple
 from os import PathLike
 
+import cython
 from cuda.core._device import Device
 from cuda.core._launch_config import LaunchConfig
 from cuda.core._stream import Stream
@@ -253,6 +254,7 @@ class Kernel:
         ...
 
     @property
+    @cython.critical_section
     def attributes(self) -> KernelAttributes:
         """Get the read-only attributes of this kernel."""
 
@@ -265,6 +267,7 @@ class Kernel:
         """list[ParamInfo]: (offset, size) for each argument of this function"""
 
     @property
+    @cython.critical_section
     def occupancy(self) -> KernelOccupancy:
         """Get the occupancy information for launching this kernel."""
 

cuda_core/cuda/core/_module.pyx

@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+cimport cython
 from libc.stddef cimport size_t
 
 from collections import namedtuple
@@ -454,6 +455,7 @@ cdef class Kernel:
         return ker
 
     @property
+    @cython.critical_section
     def attributes(self) -> KernelAttributes:
         """Get the read-only attributes of this kernel."""
         if self._attributes is None:
@@ -501,6 +503,7 @@ cdef class Kernel:
         return param_info
 
     @property
+    @cython.critical_section
     def occupancy(self) -> KernelOccupancy:
         """Get the occupancy information for launching this kernel."""
         if self._occupancy is None:
@@ -742,6 +745,7 @@ cdef class ObjectCode:
 
     # TODO: do we want to unload in a finalizer? Probably not..
 
+    @cython.critical_section
     cdef int _lazy_load_module(self) except -1:
         if self._h_library:
             return 0