Hello,
please consider upstream commit 342981fff328 ("drm/amdgpu: drop retry loop
in amdgpu_hmm_range_get_pages") for 6.18.y, 6.12.y and 6.6.y.
6.18.57, 6.12.114 and 6.6.160 took 962d684b5dc0 ("drm/amdgpu: fix
amdgpu_hmm_range_get_pages", cherry-picked from c08972f55594). As the commit
message of 342981fff328 explains, that change leaves the -EBUSY retry loop
spinning for the full timeout. 7.0.13 and 7.1.1 took the follow-up. These three
branches did not, and it carries no Fixes: or stable tag. It is not in 6.18.60,
6.12.115 or 6.6.161, and is not in the queue as of 2026-11-05.
The cost on a Radeon AI PRO R9700 with ROCm, a PyTorch speech model loaded
from safetensors:
6.18.56 and before: model load 6-7.5 s
6.18.57: model load 507 s
6.18.59: model load 514 s
7.3.0: model load 16 s with a cold page cache, 8.4 s warm
With the same 679-character text and voice on the same image, it renders
at 0.11x realtime on 6.18.59 and 0.73x on 7.3.0, against about 0.9x before
6.18.57.
The loading thread sits at 99% system time in kfd_ioctl ->
svm_range_set_attr -> svm_range_validate_and_map ->
amdgpu_hmm_range_get_pages -> hmm_range_fault.
https://github.com/ROCm/TheRock/issues/9257 reports the same stall on a gfx1100
card.
The patch does not apply cleanly. On 6.18.60 only the first hunk's context
differs, as hmm_range is still a pointer there. 6.12.115 fails on the same hunk
and 6.6.161 on both. I have not built or run a backport.
#regzbot introduced: v6.18.56..v6.18.57
Thanks,
Trace
