[PATCH v2] drm/ttm: Do not put non-struct page memory into PUD/PMDs

[Daniel Vetter]

On Wed, Oct 20, 2021 at 11:09:58AM -0300, Jason Gunthorpe wrote:
> On Wed, Oct 20, 2021 at 08:34:33AM +0200, Thomas Hellström wrote:
> 
> > Follow up question: If we resurrect this in the proper way (and in that case
> > only for x86_64) is there something we need to pay particular attention to
> > WRT the ZONE_DEVICE refcounting fixing you mention above?
> 
> Similar to PTE it should be completely separated from ZONE_DEVICE.
> 
> Seeing the special bit set at any level should trigger all page table
> walkers to never try to get a struct page.
> 
> Today some of the page table walkers are trying to do this with
> vma_is_special(), all of those should end up being the Pxx_SPECIAL
> test instead.

My understanding is that vma_is_special is for platforms which don't have
pte_special, and hence can't do gup_fast. At least I was assuming this is
why vma_is_special is a thing, and why some architectures cannot do
gup_fast.

I think for pmd and higher that approach isn't feasible and so we'll
probably have to disable VM_PFNMAP hugepages if the arch doesn't support
that.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch