Re: [PATCH 00/15] Intel Xe GPU Debug Support (eudebug) v4

["Christian König" <christian.koenig@amd.com>]

On 12/8/25 09:45, Joonas Lahtinen wrote:
> Quoting Mika Kuoppala (2025-12-03 11:18:45)
>> Christian König <christian.koenig@amd.com> writes:
>>> On 08.08.25 12:43, Mika Kuoppala wrote:
> 
> <SNIP>
> 
>>>> #### 2. ELF binaries not held in kernel memory
>>>>
>>>> In v4, debug data is delivered as a VM bind 'OP_ADD_DEBUG_DATA' extension.
>>>> The ELF binaries are no longer stored within the Xe KMD but are instead
>>>> kept in a file. The file path is passed as part of an extension in
>>>> the newly introduced 'OP_ADD_DEBUG_DATA' VM bind operation. Alternatively
>>>> pseudo-paths can be used to annotate special address ranges similar to
>>>> /proc/<pid>/maps.
>>>>
>>>> #### 3. Debug metadata not carried in VMA struct
>>>>
>>>> Instead of attaching debug data to vma created by 'OP_MAP',
>>>> we introduce separate ops for managing the metadata.
>>>> Debug data is no longer held in the VMA struct. xe_vm contains a
>>>> list of all associated debug data.
>>>
>>> I need to take a closer look but that sounds like it takes a big step into the right direction.
>>
>> Encouraged by this positive feedback, we iterated the series further. VM
>> binds state handling has reworked on top of vm_ops, metada and ufence
>> handling improved and we added pagefaults on top.
>>
>> v6 series can be found here:
>> https://lists.freedesktop.org/archives/intel-xe/2025-December/106405.html
> 
> Christian, would it be possible for you to take a quick look before the
> holiday season? Would be very much appreciated.

I'm just back from a week of sick leave today (stupid flu season) and drowning in work.

But from a ten mile high view it looks sane to me now.

> 
> That way we could start polishing the documentation etc. assuming we
> still agree on the direction. At least from our viewpoint, we're running
> short of ideas where to simplify or improve on the implementation.
> 
> The page-faults are the freshest piece of the code, however it is rather
> well separated patches, so would be great if as a minimum the foundational
> patches before them could be agreed upon initially.

The only requirement I have is that you don't mix page faults, single step debugging and dma_fences for graphics contexts, but you already knew that.

In other words as long as you make sure that you can't attach a debugger to a process which does graphics rendering and uses dma_fences for that everything should be fine from my side.

Regards,
Christian.

> 
> Best Regards, Joonas
> 
>> Thanks,
>> -Mika