linux-amdgpu-testing-6.13.6.arch1-1

[SeryogaBrigada]

Build: linux-amdgpu-testing-6.13.6.arch1-1, linux-amdgpu-testing-headers-6.13.6.arch1-1
Included patches:

• Revert "drm/amdgpu/gfx9: manually control gfxoff for CS
• Revert "drm/amdgpu: workaround for TLB seq race"
• Add PG workaround for Raven Compute rings
• Add PG workaround for GFX10.1.x and 11.0.x
• Optimize mutex protected blocks in amdgpu_vm_flush
• Refactor GFXOFF handling v4 (direct GFXOFF handling without delayed work SeryogaBrigada/linux@d5a7fc4)

Kernel option to keep during testing period: fsck.mode=force
Source

This build is similar to linux-amdgpu-testing-6.13.2.arch1-6 which was stable, but with non-blocking reference counting.

[SeryogaBrigada]

This build probably might freeze, because it uses the same memory order for reference counting as linux-amdgpu-testing-6.13.5.arch1-16 and linux-amdgpu-testing-6.13.5.arch1-17
Build with stronger memory order 6.13.6.arch1-2 prepared.

[fleming-2]

Test	amdgoup-testing-6.13.6.arch1-1	linux-6.13.5arch1-1
System Idle (PkgWatt)	~0.75W	~1.35W
glxgears (vblank_mode=0)	~4800 fps, ~6.7W	~5200 fps, ~7W
WebGL aquarium 15,000 fish	~48 fps, ~9.rW	45-50 fps, ~9.4W

• Ryzen 5 2500U (Raven Ridge)
• Chromium 134.0.6998.35 Arch Linux
• Mesa 1:24.3.4-1
• KDE Plasma 6.3.2

No freezes so far for me, I can continue to test stability if you would like.

[SeryogaBrigada]

@fleming-2 Could you please check is there any difference between this build and 6.13.6.arch1-2 in performance? I'm wondering about the cost of strong memory order for reference counting.
Extended stability test is not necessary until the performance will be on par or better compared to repo kernel.