Re: Changing cache status while validating object

[Jerome G. <gl...@fr...>]

On Wed, 2009-07-29 at 21:53 +0200, Thomas Hellström wrote:
> Jerome Glisse skrev:
> > Hi Thomas,
> >
> > I think ttm is doing somethings wrong when we validate a bo for the
> > same mem type but with different cache flags. Here is my understanding :
> >
> > (Use case i am refering too is object validate with
> > current state GTT | CACHED validate to GTT | WC)
> >
> > if (!ttm_bo_mem_compat(bo->proposed_placement, &bo->mem)
> > Is evaluate as true because cache flags are differents, For instance
> > is it goes from UNCACHED to WC or CACHED.
> >
> > So ttm_bo_move_buffer get call, which call ttm_bo_handle_move_memin there caching change will only happen if there is no ttm allocated
> > which might not be the case if for instance we are dealing with
> > GTT memory (so i think cache transitioning of the object is wrong
> > and we should move ttm_tt_set_placement_caching outside the if
> > in ttm_bo_handle_move_mem).
> >   
> In this particular case, ttm_bo_handle_move_mem will first call 
> ttm_bo_unmap_virtual(), then go ahead and call ttm_bo_move_ttm(). This 
> routine will unbind from GTT, change caching and the rebind to GTT.
> 
> > ttm_bo_handle_move_mem will end up calling driver callback move
> > routine 
> No, it will use ttm_bo_move_ttm().
> 
> > which will call in radeon case the ttm_bo_move_accel_cleanup
> > and which will create a ghost object holding the ttm memory in order
> > to delete it at latter point (which is wrong as this memory is still
> > needed).
> >
> > So shouldn't the ttm_bo_move_buffer catch this and not call device
> > move callback but simply transition cache setup of the object ?
> > Or do i misunderstand somethings ?
> >
> >   
> It should call ttm_bo_move_ttm() and hopefully it does. However, it is a 
> bit wasteful of GTT space since I think it actually allocates a new GTT 
> slot before freeing the old one. Ideally I think it should reuse the old 
> slot. There's some optimization to be done there.
> 
> So why does it bother unbinding from GTT in the first place? This is 
> because translation tables that support both snooped and unsnooped 
> bindings usually need to rewrite the page tables to reflect this change.
> 
> > Cheers,
> > Jerome
> >
> >   
> /Thomas

I need to do more debugging but i am seeing again massive pte corruption
and i thought i spotted the right path.

Jerome