Mesa 21.3 Vulkan causes amdgpu crash
This issue was brought up to me with recent mesa upgrades within distros. The issue is on our Aarch64 HoneyComb LX2K system with Polaris based GPUs. The failure happened with vkcube and certain other vulkan based implementations. The commit that caused the regression "allocate shaders to 32-bit address to skip PGM_HI" was resulting in an access out of bounds error and a GPU reset and failure.
I will be happy to collect any information needed, however I am also interested in how amdgpu deals with 32-bit accesses. It is possible that this bug may also have something to do with issues we have been seeing on the OpenGL side. Are the 32-bit accesses internal to the amdgpu's view of the memory space and if so are they only for memory allocated in the 32-bit BAR space. I ask because our 32-bit MMIO bar space is actually 64-bit addresses with a translation.
Any recommended debugging steps would be great. Thanks
Steps to reproduce
run vkcube on HoneyComb_LX2K system with a Polaris PCIe GPU.
System information
System: Host: buzz Kernel: 5.15.5-200.fc35.aarch64 aarch64 bits: 64 compiler: gcc v: 2.37-10.fc35 Desktop: N/A
wm: gnome-shell dm: GDM Distro: Fedora release 35 (Thirty Five)
CPU: Info: 16-Core model: N/A bits: 64 type: MCP arch: ARMv8 rev: 3
features: Use -f option to see features bogomips: 0
Speed: N/A min/max: N/A Core speeds (MHz): No per core speed data found.
Graphics: Device-1: Advanced Micro Devices [AMD/ATI] Lexa XT [Radeon PRO WX 2100] vendor: Dell driver: amdgpu v: kernel
bus-ID: 0004:01:00.0 chip-ID: 1002:6995
Display: server: X.Org 1.21.1.3 compositor: gnome-shell driver: loaded: amdgpu
note: n/a (using device driver) - try sudo/root resolution: 1920x1080~75Hz s-dpi: 96
OpenGL: renderer: AMD Radeon Pro WX 2100 (POLARIS12 DRM 3.42.0 5.15.5-200.fc35.aarch64 LLVM 13.0.0)
v: 4.6 Mesa 21.3.1 (git-91d2fb81c3) direct render: Yes
- OS:
NAME="Fedora Linux" VERSION_CODENAME="" PRETTY_NAME="Fedora Linux 35 (Workstation Edition)" CPE_NAME="cpe:/o:fedoraproject:fedora:35"
- GPU:
0004:01:00.0 VGA compatible controller [0300]: Advanced Micro Devices, Inc. [AMD/ATI] Lexa XT [Radeon PRO WX 2100] [1002:6995]
- Kernel version:
Linux buzz 5.15.5-200.fc35.aarch64 #1 SMP Fri Nov 26 00:30:50 UTC 2021 aarch64 aarch64 aarch64 GNU/Linux
- Mesa version:
OpenGL version string: 4.6 (Compatibility Profile) Mesa 21.3.1 (git-91d2fb81c3)
- Desktop environment: GNOME on Wayland
If applicable
- Xserver version: Xwayland 21.1.3
Regression
Yes this issue was introduced with git commit 91d2fb81c3b19da33715b3f77607be9a6e31c74b