Re: [PATCH v2 1/3] drm/amdgpu: reset SI ASIC on MacBookPros with VRAM at 0

[Brad Drury <drury.brad@gmail.com>]

Hi Maciej,

I tested this patch on a second MacBookPro11,5 (Radeon R9 M370X, Verde,
1002:6821, subsystem 106b:0149). Here the dGPU is a headless PRIME
offload GPU and the panel is on i915. The kernel is linux 7.3-rc7, with
amdgpu rebuilt from that tree with this patch on top and nothing else
changed in amdgpu (amdgpu.si_support=1 radeon.si_support=0,
mem_sleep=deep).

Without the patch, VRAM is at 0 after boot and every S3 resume fails
with "ring gfx test failed (-110)".

With it, amdgpu logs "PCI CONFIG reset" and "GPU posting now..." at
load, and VRAM ends up at 0x000000F400000000. I then ran four S3 cycles:
two on battery and two on AC, from 9 minutes to 13 h 19 min asleep.
After each resume the GART table was still at 0x000000F400000000 and
the kernel log showed no ring timeouts, -110 errors or GPU recovery.
After the last (overnight) resume, DRI_PRIME=1 glxgears (radeonsi) and
vkcube (RADV) ran, and VA-API H.264 encode (VCE) and decode (UVD) both
worked through ffmpeg.

One thing I noticed: on 2 of the 4 resumes, about a second after
resume, vce_v1_0_stop() logged:

  amdgpu 0000:01:00.0: VCE busy: VCE_STATUS=0x2, SRBM_STATUS2=0x80

VCE encode worked straight afterwards, so it doesn't seem to break
anything. I didn't see it in about 11 resumes with a local workaround
that kept VRAM at 0, so it may come with the new layout, but I haven't
looked further. Mentioning it in case it's known.

This supersedes the gmc6 patch I sent on 20 Oct for the same failure:
https://lore.kernel.org/amd-gfx/20261020-gmc6-writes-only-v3-v1-1-70730a302504@gmail.com/

Tested-by: Brad Drury <drury.brad@gmail.com> # MacBookPro11,5, Radeon R9 M370X

Thanks,
Brad