Reviewed-by: Chunming Zhou for series.
-David
在 2019/9/3 17:09, Christian König 写道:
> Trying to evict things from the current working set doesn't work that
> well anymore because of per VM BOs.
>
> Rely on reserving VRAM for page tables to avoid contention.
>
> Signed-off-by: Christian König
>
Trying to evict things from the current working set doesn't work that
well anymore because of per VM BOs.
Rely on reserving VRAM for page tables to avoid contention.
Signed-off-by: Christian König
---
drivers/gpu/drm/amd/amdgpu/amdgpu.h| 1 -
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 71
PTE means Page Table Entry.
Starting with Vega10 we have multi level page tables and when you
continuously allocate 2MB (or even 1GB) you can set a bit in the
hierarchy that the walker should stop and take the address and flags for
the whole 2MB block.
The result is that with 4K pages you
Ah ok, thanks for the explanation. About the last bit, what is the
reason behind the differences in page size? (I assume that's what you
meant by PTE? Or is that something else?)
Regards,
Kenny
On Mon, Sep 2, 2019 at 10:31 AM Christian König
wrote:
>
> Hi Kenny,
>
> When we do a CS we have a
Hi Kenny,
When we do a CS we have a certain set of buffers which the submission is
working with and are locked down while we prepare the submission.
This working set contains of the buffers in the BO list as well as the
one in the VM plus one or two for CSA and user fences etc..
Now what
Hey Christian,
Can you go into details a bit more on the how and why this doesn't
work well anymore? (such as its relationship with per VM BOs?) I am
curious to learn more because I was reading into this chunk of code
earlier. Is this something that the Shrinker API can help with?
Regards,
Trying to evict things from the current working set doesn't work that
well anymore because of per VM BOs.
Rely on reserving VRAM for page tables to avoid contention.
Signed-off-by: Christian König
---
drivers/gpu/drm/amd/amdgpu/amdgpu.h| 1 -
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 71