[Intel-gfx] Sparsely populated TTM bos

[Intel-gfx] Sparsely populated TTM bos

[Thomas Hellström]

Hi, Christian,

We have an upcoming use-case in i915 where one solution would be 
sparsely populated TTM bos.

We had that at one point where ttm_tt pages were allocated on demand, 
but this time we'd rather be looking at multiple struct ttm_resources 
per bo and those resources could be from different managers.

There might theoretically be other ways we can handle this use-case but 
I wanted to check with you whether this is something AMD is already 
looking into and if not, your general opinion.

Thanks,
Thomas

Re: [Intel-gfx] Sparsely populated TTM bos

[Christian König]

Hi Thomas,

Am 19.11.21 um 15:28 schrieb Thomas Hellström:
> Hi, Christian,
>
> We have an upcoming use-case in i915 where one solution would be 
> sparsely populated TTM bos.
>
> We had that at one point where ttm_tt pages were allocated on demand, 
> but this time we'd rather be looking at multiple struct ttm_resources 
> per bo and those resources could be from different managers.
>
> There might theoretically be other ways we can handle this use-case 
> but I wanted to check with you whether this is something AMD is 
> already looking into and if not, your general opinion.

oh, yes I've looked into this as well a very long time ago.

At that point the basic blocker was that we couldn't have different 
cache setting for the same VMA, but I think that's fixed by now.

Another thing is that you essentially need to move the LRU handling into 
the resource like I already planned to do anyway.

Regards,
Christian.

>
> Thanks,
> Thomas
>
>

Re: [Intel-gfx] Sparsely populated TTM bos

[Daniel Vetter]

On Fri, Nov 19, 2021 at 05:35:53PM +0100, Christian König wrote:
> Hi Thomas,
> 
> Am 19.11.21 um 15:28 schrieb Thomas Hellström:
> > Hi, Christian,
> > 
> > We have an upcoming use-case in i915 where one solution would be
> > sparsely populated TTM bos.
> > 
> > We had that at one point where ttm_tt pages were allocated on demand,
> > but this time we'd rather be looking at multiple struct ttm_resources
> > per bo and those resources could be from different managers.
> > 
> > There might theoretically be other ways we can handle this use-case but
> > I wanted to check with you whether this is something AMD is already
> > looking into and if not, your general opinion.
> 
> oh, yes I've looked into this as well a very long time ago.
> 
> At that point the basic blocker was that we couldn't have different cache
> setting for the same VMA, but I think that's fixed by now.

I think for cpu mmap we might just disallow them. Or we just migrate them
back into so that cpu access is always done in the same (or at least a
compatible) cache domain.

We can't really talk yet about what this thing is for, but "entire ttm_bo
cpu mmap must have same caching mode" shouldn't be a real limitation for
what we want to do here.

> Another thing is that you essentially need to move the LRU handling into the
> resource like I already planned to do anyway.

Yeah, hence why I suggested going ttm_bo : ttm_resource 1:N might be a
good idea in general, and we could piggy-pack on top of this. If you're
all on board then I guess we'll try to prototype something and maybe if
you're bored we could resurrect some of the patches to move lru/dma_resv
and whatever else from ttm_bo to ttm_resource? Just to see how much this
would impact.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch