On 5/20/26 10:07, Mikhail Gavrilov wrote:
> On Wed, May 20, 2026 at 12:08 PM Christian König
> <[email protected]> wrote:
>>
>> That whole infrastructure is superflous. You just need to modify 
>> amdgpu_vm_lock_by_pasid() to take a drm_exec object to lock the root BO.
>>
> 
> Christian, modifying amdgpu_vm_lock_by_pasid() to take a drm_exec turns
> out to also require converting its other caller, amdgpu_vm_handle_fault(),
> to drm_exec — most of the diff is that conversion, not the helper itself.
> 
> I can:
>  (a) convert both in a 2-patch series (handle_fault becomes
>      drm_exec_init + drm_exec_until_all_locked + drm_exec_fini, ~30 lines),
>      or
>  (b) keep the loop inside amdgpu_vm_lock_by_pasid() so handle_fault stays
>      a one-liner — but then the devcoredump caller can't add the IB BOs
>      to the same ticket, which is the whole point.
> 
> (a) seems unavoidable if we want one helper. Is that what you had in mind,
> or did you intend something lighter — e.g. a separate
> amdgpu_vm_lock_by_pasid_exec() leaving handle_fault untouched?
> 

Just make that two patches, first switching over amdgpu_vm_lock_by_pasid() to 
using drm_exec() on both use cases.

And then changing the one for device core dumping to lock all BOs at once.

Thanks,
Christian.

Reply via email to