[PATCH 1/3] drm/amdgpu: fix VA hole handling on Vega10 v2

Christian König ckoenig.leichtzumerken at gmail.com
Sat Nov 18 14:31:52 UTC 2017


Am 17.11.2017 um 17:09 schrieb Michel Dänzer:
> On 17/11/17 11:28 AM, Christian König wrote:
>> Ping? Michel, Alex can somebody take a look?
> Patch 2 is
>
> Reviewed-by: Michel Dänzer <michel.daenzer at amd.com>
>
>
> With patches 1 & 3, it's not 100% clear to me what the idea is behind
> the handling of the hole on the kernel and userspace side. Maybe you can
> add some explanation in code comments or the commit logs?
Yeah, that is actually a bit of a mess because the hardware 
documentation wasn't very clear on how this works.

How about this as extra code comment on patch 1 to the assignment of 
dev_info.virtual_address_max:

/*
  * Old userspace isn't aware of the VA hole on Vega10. So in theory an 
client could get invalid VA addresses assigned.
  * To fix this and keep backward compatibility we limit the VA space 
reported in this field to the range below the hole.
  */

The last patch is then to report the VA space above the hole, cause that 
is actually what libdrm should use.

The crux is when I put the VA space above the hole directly into the old 
fields older versions of libdrm would break and we can't do that.

Regards,
Christian.


More information about the amd-gfx mailing list