On 2026-11-16 23:20, Junhua Shen wrote:
KFD records whether an allocation or SVM range is writable
(KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE for userptr, KFD_IOCTL_SVM_FLAG_GPU_RO
for SVM) and already honors it when building the GPUVM page table. The
userptr registration and the SVM mapping path, however, do not pass that
intent down to the HMM fault, so hmm_range_fault() is asked for write
access even for a read-only buffer.

For a read-only buffer backed by a writable private mapping
(PROT_READ|PROT_WRITE, MAP_PRIVATE) that the GPU only reads, e.g. a
host-to-device copy source, faulting for write breaks COW: every page is
copied into anonymous memory, which doubles the footprint and, because
the copy changes the CPU page tables, fires an MMU-notifier invalidation
that evicts the GPU user-mode queues.

To keep the GPU mapping consistent with the requested permission, this
series makes the HMM fault honor the same read-only intent:

  - Patch 1 (userptr): register the userptr read-only when the allocation
    did not request KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE.
  - Patch 2 (SVM): also honor KFD_IOCTL_SVM_FLAG_GPU_RO instead of deriving
    read-only solely from the VMA write permission.

This does not change what the GPU can access: the page table already maps
a read-only allocation without write permission, so a GPU write to it
already faults, with or without this series. The series only aligns the
HMM fault with that page table and drops the needless write fault and COW.
A buffer the GPU may write must still be allocated writable; there is no
in-place upgrade, so making a buffer writable means freeing and
re-allocating it.

The user-space side that marks such buffers read-only is tracked here:
Link:https://github.com/ROCm/rocm-systems/pull/15545

Junhua Shen (2):
   drm/amdgpu: map KFD userptr read-only when the BO is not writable
   drm/amdkfd: honor KFD_IOCTL_SVM_FLAG_GPU_RO for read-only SVM ranges

  drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 4 +++-
  drivers/gpu/drm/amd/amdkfd/kfd_svm.c             | 3 ++-
  2 files changed, 5 insertions(+), 2 deletions(-)


base-commit: 067c5c24cef23d913b67eaaf71ae67e7169a6620
This change looks safe, svm ranges without the read-only attribute are unaffected, so I don’t expect regressions.

Using GPU_RO to request read-only access on a writable VMA is unusual. Ideally, userspace would use a read-only file-backed mapping, but this provides another way to request read-only GPU access and avoid triggering COW in hmm_range_fault().

Userspace is responsible for setting the read-only flag when appropriate. Failing to do so would be a userspace bug.

I don’t think the Fixes tag suggested by the bot is appropriate, since this supports new userspace functionality rather than fixing a regression. This patch series is

Reviewed-by: Philip Yang <[email protected]>

Reply via email to