On 2018-02-13 01:15 PM, Christian König wrote:
> Am 13.02.2018 um 18:18 schrieb Felix Kuehling:
>> On 2018-02-13 12:06 PM, Christian König wrote:
>>> Ah, yeah that is also a point I wanted to to talk about with you.
>>> The approach of using the same buffer object with multiple amdgpu
>>> devices doesn't work in general.
>>> We need separate TTM object for each BO in each device or otherwise we
>>> break A+A laptops.
>> I think it broke for VRAM BOs because we enabled P2P on systems that
>> didn't support it properly. But at least system memory BOs can be shared
>> quite well between devices and we do it all the time.
> Sharing VRAM BOs is one issue, but the problem goes deeper than just
> Starting with Carizzo we can scanout from system memory to avoid the
> extra copy on A+A laptops. For this to work we need the BO mapped to
> GART (and I mean a real VMID0 mapping, not just in the GTT domain).
> And for this to work in turn we need a TTM object per device and not a
> global one.
I still don't understand. I think what you're talking about applies only
to BOs used for scan-out. Every BO is allocated from a specific device
and can only be GART-mapped on that device. What we do is map the same
BO in VMs on other devices. It has no effect on GART mappings.
>> I don't see how you can have separate TTM objects referring to the
>> same memory.
> Well that is trivial, we do this all the time with prime and I+A laptops.
As I understand it, you use DMABuf to export/import buffers on multiple
devices. I believe all devices share a single amdgpu_bo, which contains
the ttm_buffer_object. The only way you can have a new TTM buffer object
per device is by using SG tables and pinning the BOs. But I think we
want to avoid pinning BOs.
What we do does not involve pinning of BOs, even when they're shared
between multiple devices' VMs.
>>> That is also the reason we had to disable this feature again in the
>>> hybrid branches.
>> What you disabled on hybrid branches was P2P, which only affects
>> large-BAR systems. Sharing of system memory BOs between devices still
>> works fine.
> No, it doesn't. It completely breaks any scanout on Carizzo, Stoney
> and Raven. Additional to that we found that it breaks some aspects of
> the user space interface.
Let me check that with my current patch series. The patches I submitted
here shouldn't include anything that breaks the use cases you describe.
But I'm quite sure it will support sharing BOs between multiple devices'
> So end result is that we probably need to revert it and find a
> different solution. I'm already working on this for a couple of weeks
> now and should have something ready after I'm done with the PASID
amd-gfx mailing list