WSL2 / dxgkrnl: GPU virtual address & memory drift under ephemeral CUDA process churn (Blackwell ARM64)

Open
#42,238 3 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
5/5
Estimated time
Over a week
Newbie friendliness
25/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
linux

Research direction

Start by reviewing the WSL kernel's dxgkrnl GPU memory and process teardown paths; the issue provides a standalone CUDA reproducer and commands for tracking free VRAM and guest processes. Run repeated allocation and abrupt-exit cycles, then compare memory and dxg ioctl logs. Done means allocations are reclaimed after process exit without requiring wsl.exe --shutdown.

Written by the indexing model from the issue text.

Description

GPU wsl2
Windows Version

Microsoft Windows 11 ARM64 (Build 26100+)

WSL Version

WSL 2

Are you using WSL 1, WSL 2, or WSLC?
  • WSL 2
Kernel Version

6.18.58.1-microsoft-standard-WSL2 (aarch64)

Distro Version

Ubuntu 26.04.1 LTS (Resolute Raccoon)

Other Software
  • GPU: NVIDIA RTX Spark N1X (6144-core Blackwell RTX GPU, GB10, sm_121a)
  • NVIDIA Windows Display Driver: 621.14 (KMD)
  • WSL2 NVIDIA Driver: libcuda.so.1 (CUDA UMD 13.5, nvidia-smi 620.70.03)
  • Application: llama.cpp CUDA build sm_121a, PyTorch / CUDA 13.x
Repro Steps
  1. On a WSL2 instance with an NVIDIA GPU (reproduced on unified memory Blackwell ARM64), record initial baseline free VRAM:
    nvidia-smi --query-gpu=memory.free --format=csv,noheader,nounits
  2. Run an ephemeral CUDA process that allocates GPU memory and unified memory buffers (e.g. llama-server or a standalone CUDA C program allocating 4GB device + 512MB UVM).
  3. Terminate the process abruptly via SIGTERM or _exit(0) without an explicit synchronous cudaDeviceReset().
  4. Observe free VRAM after process termination and zero surviving guest processes.
  5. Repeat 10-20 times.

Minimal standalone C reproducer (cuda_dirty_exit.cu):

// Compile with: nvcc -O2 -o cuda_dirty_exit cuda_dirty_exit.cu
#include <cuda_runtime.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>

int main(int argc, char** argv) {
    void* d_ptr = NULL;
    size_t bytes = 1024ULL * 1024 * 1024 * 4; // 4 GB GPU device allocation
    cudaError_t err = cudaMalloc(&d_ptr, bytes);
    if (err != cudaSuccess) return 1;

    void* uvm_ptr = NULL;
    size_t uvm_bytes = 1024ULL * 1024 * 512; // 512 MB Unified Memory
    err = cudaMallocManaged(&uvm_ptr, uvm_bytes, cudaMemAttachGlobal);
    if (err != cudaSuccess) return 1;

    cudaMemset(d_ptr, 0xAA, 1024 * 1024 * 64);
    _exit(0); // Exit without cudaDeviceReset() to replicate server termination
}
Expected Behavior

When an ephemeral CUDA process terminates, all paravirtualized GPU virtual address mappings, page tables, and WDDM memory residency allocations should be synchronously reclaimed by dxgkrnl and host WDDM, restoring free VRAM back to the initial baseline.

Actual Behavior
  • After 10 cycles, baseline free VRAM degraded from 2,352 MiB down to 2,233 MiB (a permanent net loss of 127 MiB, peaking at 190 MiB cumulative unreturned drift), despite zero surviving processes in pgrep or /proc/[0-9]*/cmdline.
  • Under extended lifecycle sweeps (e.g. automated model autotuning / benchmarking), this degradation compounds at ~32 MiB unreturned per spawn until large contiguous allocations fail with cudaErrorMemoryAllocation and the VM becomes unresponsive.
  • dmesg accumulates thousands of misc dxg: dxgk: dxgkio_escape: Ioctl failed: -22 entries.
  • The only recovery boundary is wsl.exe --shutdown from Windows; guest-side process kills, dropping caches, or running clean-up scripts recover 0 bytes.
Diagnostic Logs

Empirical run log across 10 ungraceful process terminations:

=== WSL2 GPU Virtual Address / Memory Drift Reproducer ===
Kernel: 6.18.58.1-microsoft-standard-WSL2
GPU: NVIDIA RTX Spark N1X (6144-core Blackwell RTX GPU)
Driver / CUDA UMD: 621.14
Initial Baseline VRAM: 2352 MiB free
---------------------------------------------------------
Cycle 01: Free=2192 MiB | Delta=+166 MiB | Cumulative Drift=166 MiB
Cycle 02: Free=2217 MiB | Delta=-25 MiB  | Cumulative Drift=141 MiB
Cycle 03: Free=2285 MiB | Delta=-68 MiB  | Cumulative Drift=73 MiB
Cycle 04: Free=2329 MiB | Delta=-44 MiB  | Cumulative Drift=29 MiB
Cycle 05: Free=2325 MiB | Delta=+4 MiB   | Cumulative Drift=33 MiB
Cycle 06: Free=2339 MiB | Delta=-14 MiB  | Cumulative Drift=19 MiB
Cycle 07: Free=2259 MiB | Delta=+80 MiB  | Cumulative Drift=99 MiB
Cycle 08: Free=2168 MiB | Delta=+91 MiB  | Cumulative Drift=190 MiB
Cycle 09: Free=2176 MiB | Delta=-8 MiB   | Cumulative Drift=182 MiB
Cycle 10: Free=2233 MiB | Delta=-57 MiB  | Cumulative Drift=125 MiB
---------------------------------------------------------
Final Baseline Free: 2233 MiB
Total Drift across 10 cycles: 127 MiB unreturned
Active Linux Processes: 0
Dominant language
C++
Stars
34.4k
Forks
1.9k
Avg merge
1d 23h
Merged PRs (30d)
166

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from microsoft/WSL

All issues in microsoft/WSL

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.