On 2026-11-09 15:55, Han, Qiang wrote:
AMD General
Hi,
I would like to have code review on this kfd fix. I have an
observation: the buggy pageable GPU H2D copy read GPU machine code
from unrelated text from outside the container into the training input
tensor. These bytes have nothing to do with the intended data, and the
copy returned success with no error. This happens silently under host
memory pressure.
The root cause should be a bug in the KFD kernel driver
(kfd_svm.c:svm_range_evict); a workaround and a tested kernel patch
are attached. kfd_svm_evict_fix_2026-11-08_v2.zip
Philip has done the initial code review.
The patch is harder to review as an attachment — I'll resend it inline.
Regards,
Philip
Thanks,
Qiang
