Allow win32_handle_metadata to be an arbitrary object?

@leofang

I'd like to have input from the team on this. I feel taking a pointer address here is not the most pleasant design, but the challenge is that there is no official Python binding for the Windows SECURITY_ATTRIBUTES struct, so we either have to take arbitrary object or a pointer.

Originally posted by @leofang in #1242 (comment)

This needs to be addressed before the next cuda.core release is out (v0.5.0), otherwise it'd become a breaking change.

xref: #1264 — possibly related, but NOT certainly

According to the CUDA Driver API docs: https://docs.nvidia.com/cuda/cuda-driver-api/structCUmemAllocationProp__v1.html#structCUmemAllocationProp__v1_1542262ad88e1d00f02b306c641270168

This is a pointer to a OBJECT_ATTRIBUTES struct (https://learn.microsoft.com/en-us/windows/win32/api/ntdef/ns-ntdef-_object_attributes), not a SECURITY_ATTRIBUTES struct. Is there a relationship between these two in some way that makes us care about SECURITY_ATTRIBUTES?

@kkraus14

Right, good catch @kkraus14. Unfortunately I confused myself. Confirmed with Vishnu that we need to use OA (OBJECT_ATTRIBUTES).

I used SA (SECURITY_ATTRIBUTES) because when we started exploring with mempools (#446) SA was what we needed to use:
https://docs.nvidia.com/cuda/cuda-driver-api/structCUmemPoolProps__v1.html#structCUmemPoolProps__v1_1b153a37a3a73be7aaa98f271429b87a7
I forgot that we are now dealing with VMM instead of mempools, and so I did not bother to double check this impression.

FWIW, SA is only 24B whereas OA is 48B (on the Colossus system that I rented), so this is definitely a bug in the test suite that we must fix. This would, for example, answer why I saw weird things like this.

Since OA must be initialized by InitializeObjectAttributes, which unfortunately is a macro function, compiling a bit of C code to wrap it is needed. I confirm this example works:

from cffi import FFI
prog = FFI()
prog.cdef("uintptr_t do_stuff();")
source = r"""
#define _AMD64_
#include <ntdef.h>

static OBJECT_ATTRIBUTES objAttributes;

extern "C"
uintptr_t do_stuff() {
    static bool objAttributesConfigured = false;
    if (!objAttributesConfigured) {
        InitializeObjectAttributes(&objAttributes, NULL, 0, NULL, NULL);
        printf("there you go: %" PRIuPTR "\n",  (uintptr_t)(&objAttributes));
        printf("size of obj attr: %zu \n", sizeof(objAttributes));
        objAttributesConfigured = true;
    }
    return (uintptr_t)(&objAttributes);
}
"""
prog.set_source(
    "cpu_obj",
    source,
    source_extension=".cpp",
    extra_compile_args=["-std=c++11"],
)
import tempfile
temp_dir = tempfile.mkdtemp()
prog.compile()#tmpdir=temp_dir)

import sys
sys.path.append(temp_dir)
from cpu_obj.lib import do_stuff
print(f"{do_stuff()=}")


from cuda.core.experimental import Device, VirtualMemoryResource, VirtualMemoryResourceOptions
dev = Device()
dev.set_current()

options=VirtualMemoryResourceOptions(handle_type="win32", win32_handle_metadata=do_stuff())
mr = VirtualMemoryResource(dev, options)
buf = mr.allocate(1024)
buf.close()
print("DONE!")

I'll integrate this example into the test suite.

Update: on the Thursday team-sync we decided to disable win32 support for now, thereby bypassing this discussion, because the underlying OBJECT_ATTRIBUTES is very hard to produce in Python, and it is unclear to us how to ensure general correctness. If handle_type="win32" is passed we'll raise an exception asking users to reach out.

Navigation Menu

Repository navigation

Allow `win32_handle_metadata` to be an arbitrary object? #1263

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Allow win32_handle_metadata to be an arbitrary object? #1263

Description

Activity

leofang commented on Nov 18, 2025

rwgk commented on Nov 18, 2025

kkraus14 commented on Nov 20, 2025

leofang commented on Nov 20, 2025

leofang commented on Nov 24, 2025

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Allow `win32_handle_metadata` to be an arbitrary object? #1263