Re: [PATCH 5/7] drm/amdgpu: Make first-level trap state GPUVM-owned

[SRINIVASAN SHANMUGAM <srinivasan.shanmugam@amd.com>]

[-- Attachment #1: Type: text/plain, Size: 1278 bytes --]


On 10/12/2026 8:02 PM, Christian König wrote:
> On 10/8/26 09:05, Srinivasan Shanmugam wrote:
>> The first-level trap state belongs to the GPUVM because the TBA and TMA
>> are mappings in that GPUVM.
> Oh, that was actually not what I thought of.
>
> The idea was to use the VM PDs root reservation object to allocate the trap handler and so simplify the handling.
>
> Moving that into the amdgpu_vm structure was not part of this plan. It's not really a page table and so probably better suited in fpriv after all.
>
> On the other hand we want to modify it through the VM IOCTLs and so making it part of the VM is probably not so bad.
>
> I need to think about that a bit more when I have time.

Hi Christian,

Good Morning!

Thanks for your feed backs and reviews, No rush, but when you have some 
free time, Just a quick follow up on the trap object ownership discussion.

My thinking was that the trap state is tied to the GPUVM mappings 
(TBA/TMA), and queues using the same GPUVM share the same trap state. 
Since the control path is also moving through the VM ioctl interface, 
keeping the trap object in |amdgpu_vm| felt natural to me.

However, I'm happy to follow whichever approach you think is best in the 
long term.

Thanks, Srini




>
> Regards,
> Christian.

[-- Attachment #2: Type: text/html, Size: 2424 bytes --]