Nouveau - Dec 2009 - NV50: the tiled buffer object eviction problem

In short, we move out the low level content of a buffer object. In the
case of textures and such this is utterly useless. Still it is
accessed, because ttm sees no problem in using PL_SYSTEM or PL_TT
memory. What is the best way to let ttm know we don't really
appreciate that? A custom memory to evict to that cannot be mapped
maybe. Or an extension of PL_SYSTEM or PL_TT. Share your ideas.

Maarten.

On Sat, 2009-12-26 at 03:13 +0100, Maarten Maathuis
wrote:
> In short, we move out the low level content of a buffer object. In the
> case of textures and such this is utterly useless. Still it is
> accessed, because ttm sees no problem in using PL_SYSTEM or PL_TT
> memory. What is the best way to let ttm know we don't really
> appreciate that? A custom memory to evict to that cannot be mapped
> maybe. Or an extension of PL_SYSTEM or PL_TT. Share your ideas.
For GPU access it isn't an issue, userspace knows it needs to use VRAM
and won't set GART as an allowed memtype.

For CPU access, we should trap this in the fault handler and revalidate
the buffer into VRAM.

That's my thoughts :)

Merry Xmas!
Ben.
> 
> Maarten.
> _______________________________________________
> Nouveau mailing list
> Nouveau at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/nouveau