Public bug reported:

## Summary

This is an Ubuntu kernel regression affecting a headless Radeon AI PRO
R9700 in a hybrid-GPU system. The workload was stable with
`7.0.0-38-generic`. Starting with `7.0.0-42-generic`, the R9700 reports
`Using BOCO for runtime pm`, and large RADV/Vulkan compute/model-loading
workloads can create very large `gpu_active` memory pressure followed by
a global OOM cascade. The problem remains present in `7.0.0-43-generic`.

Keeping only the R9700 awake with `power/control=on` completely prevents
the failure and allows the identical workload to load and run
successfully.

Upstream AMDGPU report:
https://gitlab.freedesktop.org/drm/amd/-/work_items/6147

## Hardware and software

- Ubuntu 26.04, amd64
- CPU/iGPU: AMD Ryzen 7 PRO 6850H / Radeon 680M
- Discrete GPU: Radeon AI PRO R9700, PCI ID `1002:7551`, revision c0, 32 GiB 
VRAM
- R9700 used as a headless compute GPU over PCIe/OCuLink
- System RAM: 24 GB; swap: 12 GB
- Vulkan: Mesa/RADV 26.0.8
- Firmware package: `20260319.git217ca6e4.1ubuntu`
- Last observed good kernel: `7.0.0-38-generic` (`7.0.0-38.38`)
- First observed bad kernel: `7.0.0-42-generic` (`7.0.0-42.42`)
- Current affected kernel: `7.0.0-43-generic` (`7.0.0-43.43`)

## Regression evidence

- Six recorded boots with `7.0.0-38-generic`: no global OOM event from this 
workload.
- First boot with `7.0.0-42-generic`: the R9700 reports `Using BOCO for runtime 
pm`; the first GPU-associated global OOM cascade occurred during that boot.
- `7.0.0-43-generic`: BOCO runtime PM remains enabled and the failure remains 
reproducible.
- Mesa/RADV and AMD firmware did not change across the start of the regression.

## Sanitized log excerpts

```text
amdgpu 0000:03:00.0: Using BOCO for runtime pm

Node 0 ... all_unreclaimable? yes ... gpu_active:11002768kB gpu_reclaim:0kB
Out of memory: Killed process ...

Node 0 ... all_unreclaimable? yes ... gpu_active:18338004kB gpu_reclaim:0kB
Out of memory: Killed process ...
systemd-oomd: Killed ... org.gnome.Shell ... due to memory pressure
```

Repeated around affected loads:

```text
amdgpu 0000:03:00.0: SMU is resuming...
amdgpu 0000:03:00.0: SMU is resumed successfully!
```

## Reproduction

1. Leave the R9700 runtime-PM policy at its default:
   ```text
   /sys/bus/pci/devices/0000:03:00.0/power/control = auto
   ```
2. Use RADV/Vulkan to load an approximately 20 GiB workload with full GPU 
offload and a large GPU-resident context.
3. Observe repeated R9700 runtime resume events and rapidly increasing 
`gpu_active` in kernel OOM diagnostics.
4. Global memory pressure kills unrelated user processes and can terminate the 
desktop session.

## Workaround and controlled A/B verification

Applying:

```text
echo on > /sys/bus/pci/devices/0000:03:00.0/power/control
```

prevents runtime suspend only for the R9700.

The identical workload was then tested successfully:

- model load and real inference completed successfully
- peak R9700 VRAM use: 27.16 GiB
- minimum available system RAM: 14.34 GiB
- minimum free swap: 7.95 GiB
- `runtime_status` remained `active`
- no OOM event
- no GPU reset, timeout, or SMU resume event during the test

This demonstrates that normal RAM/VRAM capacity is sufficient and
isolates the failure to the runtime-PM/BOCO path.

## Expected result

The R9700 should enter and leave runtime suspend without creating enough
unreclaimable memory pressure to trigger global OOM, or runtime PM
should be disabled for this device/topology if BOCO is unsafe.

## Actual result

Starting with Ubuntu kernel `7.0.0-42`, runtime PM on the headless R9700
can cause large `gpu_active` memory pressure and global OOM kills under
Vulkan compute load. Keeping the GPU active prevents the issue entirely.

Please correlate the Ubuntu changes between `7.0.0-38.38` and
`7.0.0-42.42` with the upstream AMDGPU report and advise whether a test
kernel or additional narrowly scoped diagnostics would help.

** Affects: linux (Ubuntu)
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2174670

Title:
  [amdgpu][R9700] Regression in 7.0.0-42: BOCO runtime PM causes global
  OOM during Vulkan compute loads

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2174670/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to