dom0 boot loop with kernel-latest-5.14.15

[icequbes1]

Qubes OS release

R4.0, stable + security-testing repos enabled

Brief summary

After kernel-latest-5.14.15 was pushed to stable on November 22, 2021, an AMD Ryzen machine that uses amdgpu boot loops. It is OP's hypothesis that it is related to AMD/amdgpu, but this has not been proven.

Steps to reproduce

1. Power on

Expected behavior

1. Disk passphrase is requested

Actual behavior

1. Disk passphrase is not requested
2. Machine immediately reboots

Additional info

Initial comments were provided 0, where I stated:

1. xen loads
2. dom0 loads, kernel output shown
3. dom0: "Please enter passphrase for disk...." text-mode prompt seem very briefly
4. dom0 kernel: amdgpu: Topology: Add APU mode
5. dom0 kernel: fb0: switching to amdgpudrmfb from EFI VGA
6. reboot

On successful boots with working dom0 kernel (5.13.6-1), lots of amdgpu output immediately after, so my guess is something in amdgpu broke between 5.13.6 and 5.14.15.

After a weak kernel version bisect (with xen 4.8.5-36), the observations are as follows:

1. kernel-latest-5.13.6: no issues
2. kernel-latest-5.14.10: no issues
3. kernel-latest-5.14.15: dom0 boot loop
4. kernel-latest-5.14.16: dom0 boot loop
5. kernel-latest-5.14.17: no issues

Based on the above, the fix was implemented in 5.14.17. The 5.14.17 changelog does reveal two notable changes in amdgpu, one of which mentions a certain feature being "severely broken" and reverting the feature 1. I have not bisected by kernel commit to verify this and cannot state that this is the reason for the boot loop; only that the boot loop was no longer observed in this version.

[ydirson]

@icequbes1 which AMD GPU(s) do you have on this machine ?

[icequbes1]

Raven I believe.

[isodude]

Note:

I've done this trip, there is at least 5 proper broken commits between 5.14 and 5.16. There's

• A patch for shifting early memory allocation
• PCI/MSI broke, fixed somewhere in 5.15-rc4 I think
• Offlining CPUs broke in 5.15-rc4+ solved in 5.16 I think ( affecting both Xen and Non-Xen )
  ..

The PCI-problem is top of my mind and can be isolated via amdgpu.msi=0 as kernel commandline.
The early memory allocation thing needs to be patched.

If you want you can fetch the log by:
Xen cmdline: console=vga vga=keep noreboot
Linux cmdline: console=hvc0

I found that my camera had OTR-support which helped a lot :) I can pinpoint the patches if needed.

[icequbes1]

Closing as a later kernel-latest version [5.16] did not experience the issue.