Mesa 19.3 Adds Support For New AMDGPU Reset Interface

[phoronix]

Phoronix: Mesa 19.3 Adds Support For New AMDGPU Reset Interface

AMD open-source developer Marek Olšák is landing the last of his changes for the Mesa 19.3 imminent feature freeze...

http://www.phoronix.com/scan.php?page=news_item&px=Mesa-19.3-New-AMDGPU-Reset2

[ThoreauHD]

Should help with vfio.

[boxie]

One step closer to a smooth reset experience

[Nille]

Should help with vfio.

Fix the IOMMU Bugs in AGESA would be a start. And in the Future Generations the PCIe Function to reset the Device.

[Nille_kungen]

Fix the IOMMU Bugs in AGESA would be a start. And in the Future Generations the PCIe Function to reset the Device.

What IOMMU bugs are still there and what hardware?

[Nille]

What IOMMU bugs are still there and what hardware?

Since some AGESA Versions (Somewhere between Zen and Zen+ are they Intruded) you get

Unknown header type 7f

or

internal error: Unknown PCI header type '127'

Errors for PCIe Devices. As far as i Know all AMD Platforms are Affected by it.

There are Kernel Patches as a work around.

And there is a Problem with the AMD GPUs. The GPUs are not Supporting a PCIe Reset function. So, after a VM Reboot with an AMD GPU the GPU is in a unrecoverable state and looks up after a short while (a work around is to shutdown the card before the shutdown of the guest) The Windows Driver sometimes Succeed to reset the GPU on the Guest but the Linux driver did fail and crash the System. For Navi and Vega there are Patches to reset the GPUs.

The VFIO Community has a lot of stuff about it if you ever think about a gpu pass though config. But one conclusion currently is to avoid AMD CPUs and GPUs.

[dwagner]

I have yet to see an amdgpu crash that is survived by the rest of the kernel, so this seems to be targeted solely to GPU-to-VM-passthrough-users, but apart from that it reminds me of the sad habit of some programmers writing scripts restarting their server binaries in an endless loop, instead of fixing their bugs in the binary. It's like raising a white flag to signal that the coders where defeated, and the bugs won.

[DMJC]

Is there anyone working on smooth unmount support for amdgpu when in an egpu enclosure? I find I can't remove the amdgpu module, the system just spins and does nothing. As a hotplug device it's pretty terrible, usually requires rebooting everytime you plug in/disconnect the GPU. NVIDIA worked a bit better, not that that was perfect either.