GVT-g VM crashes and freezes the host

[dgarfias]

So I had been using GVT-g for over a year with no problems but since last month my Windows 10 VM crashes (and it takes the host with it) when I launch any app that is more demanding than a web browser or a PDF viewer. For example, this now always happens when I launch AutoCAD Electrical, but it didn't happen before.

Here is the output of journalctl -k https://pastebin.com/raw/hEg17TdG but I can't find any error messages that would explain the crash.

I was wondering if anyone has the same problem. Some information about my system:

CPU: Intel i5-10210U, iGPU: Intel UHD (CML GT2)

OS: Arch Linux, DE: GNOME 40, Kernel: 5.12.9-arch1-1, QEMU version: 6.0.0

[sochotnicky]

I started seeing something similar recently (or identical - I'll leave the devs to ponder that based on below info).

I believe it started in 5.12.x kernel (i.e. 5.11.x was working fine). I run a Windows 10 VM and it's basically guaranteed to hard-freeze the host kernel to the point of not even sys-rq can recover/reboot. As such I don't have any logs available but I am more than happy to try and gather some info.

In my cases the used components are:
CPU: i7-9700K
OS: Gentoo, any kernel in 5.12.x line (currently can still reproduce in 5.12.11), QEMU 6.0.0 but I could also reproduce on older 5.2.0.

[dgarfias]

I started seeing something similar recently (or identical - I'll leave the devs to ponder that based on below info).

I believe it started in 5.12.x kernel (i.e. 5.11.x was working fine). I run a Windows 10 VM and it's basically guaranteed to hard-freeze the host kernel to the point of not even sys-rq can recover/reboot. As such I don't have any logs available but I am more than happy to try and gather some info.

In my cases the used components are:
CPU: i7-9700K
OS: Gentoo, any kernel in 5.12.x line (currently can still reproduce in 5.12.11), QEMU 6.0.0 but I could also reproduce on older 5.2.0.

Okay, in that case I will try a kernel downgrade to confirm if we have the same problem.

[TerrenceXu]

The regression is introduced by kernel 5.12, you can use the 5.11, it is stable and validated by us.

[sochotnicky]

Ah, well good to hear that it's a known problem in 5.12 and I wasn't imagining things. Guess we can expect for it to be sorted in 5.13 once it comes out ?

…

On Thu, 17 Jun 2021, 17:42 Terrence Xu, ***@***.***> wrote: The regression is introduced by kernel 5.12, you can use the 5.11, it is stable and validated by us. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#188 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAANTLQVKCZICK6ILCMFGWDTTIJX7ANCNFSM46PZBVKA> .

[dgarfias]

I downgraded the kernel to version 5.11 and can confirm that the issue is no longer present.

[mriedel87]

Hi guys,
it's also crashing with vanilla kernel 5.13.
Do you have a fix? It would be great to know the pull request :)

Best regards

Marc

[dgarfias]

Any news on this issue? What's causing it? I don't want to be stuck with an older kernel.

[TerrenceXu]

@dgarfias , we are working on the bug fix, the regression commit as below, you can revert it in any kernel you want to use first. :)

commit 5858668
Author: Chris Wilson chris@chris-wilson.co.uk
Date: Mon Jan 25 22:01:52 2021 +0000

drm/i915: Disable atomics in L3 for gen9

[dgarfias]

@dgarfias , we are working on the bug fix, the regression commit as below, you can revert it in any kernel you want to use first. :)

commit 5858668
Author: Chris Wilson chris@chris-wilson.co.uk
Date: Mon Jan 25 22:01:52 2021 +0000

drm/i915: Disable atomics in L3 for gen9

Thanks, I'll try it :)

[SimpliFly03]

@dgarfias , we are working on the bug fix, the regression commit as below, you can revert it in any kernel you want to use first. :)

commit 5858668
Author: Chris Wilson chris@chris-wilson.co.uk
Date: Mon Jan 25 22:01:52 2021 +0000

drm/i915: Disable atomics in L3 for gen9

I also had similar problem though not exactly the same. In my case while the Windows guest act broken and display is glitched and frozen, the host only have some momentary freezes every 10-15 seconds. If I kill the VM or shut down through ssh, the host returns to proper functioning as if none of those glitches or freezes happened. Of course starting the VM again problem returns. Around that time I was very busy so I didn't have enough time to debug and report it, and Google wasn't helpful either. Thus, I have just booted to 5.11 kernel every time I need VM with GVT-G.

But today I seen your comment. I applied that commit as a revert to 5.13.1 kernel and tested it. And result is no glitch or freeze anymore and the VM is functioning properly like back in 5.11. The host is Arch Linux, the guest is Windows 11.

Thanks for offering solution.

[mriedel87]

Hi guys,

even with the latest vanilla kernel 5.14 (and also 5.13.13), which includes the mentioned fix "drm/i915/gvt: Fix cached atomics setting for Windows VM", it's still crashing.

Aug 30 10:20:37 kaby kernel: [ 233.546504] i915 0000:00:02.0: [drm] Resetting rcs0 for preemption time out
Aug 30 10:20:37 kaby kernel: [ 233.547026] i915 0000:00:02.0: [drm] GPU HANG: ecode 9:1:8ed9fff2
Aug 30 10:20:48 kaby kernel: [ 244.040421] Asynchronous wait on fence 0000:00:02.0:kwin_x11[2654]:2482 timed out (hint:intel_atomic_commit_ready [i915])
Aug 30 10:20:51 kaby kernel: [ 247.560738] i915 0000:00:02.0: [drm] GPU HANG: ecode 9:0:00000000
Aug 30 10:20:51 kaby kernel: [ 247.561744] i915 0000:00:02.0: [drm] Resetting rcs0 for stopped heartbeat on rcs0
Aug 30 10:20:51 kaby kernel: [ 247.562276] i915 0000:00:02.0: [drm] Resetting chip for stopped heartbeat on rcs0
Aug 30 10:20:51 kaby kernel: [ 247.664819] [drm:__uc_sanitize [i915]] ERROR Failed to reset GuC, ret = -110
Aug 30 10:20:51 kaby kernel: [ 247.736823] i915 0000:00:02.0: [drm] ERROR Failed to reset chip
Aug 30 10:20:51 kaby kernel: [ 247.736825] i915 0000:00:02.0: [drm:add_taint_for_CI [i915]] CI tainted:0x9 by intel_gt_reset+0x2c8/0x2f0 [i915]

Best regards,

Marc

[SimpliFly03]

@mriedel87 it is strange I am currently running 5.14 kernel and GVT-G on Windows 11 VM works just fine. Maybe our problems were different and what they fixed was mine. Also, recently, an Intel driver update inside of VM broken GVT-G and causes Qemu(not Windows) to instantly crash a few seconds after the GPU driver was initialized. I fixed it by booting into the safe mode and downgrading the driver. Currently running the driver 28.20.100.8322 and it works fine.

Edit: I have just switched to driver 30.0.100.9805 it also works just fine.