Re: [PATCH v8 01/23] dma-direct: return struct page from dma_direct_alloc_from_pool()

On Mon, Jul 27, 2026 at 01:23:57AM -0300, Jason Gunthorpe wrote:
> On Sun, Jul 26, 2026 at 11:17:53AM +0300, Leon Romanovsky wrote:
> 
> > Currently, the phys type universally describes memory and can be
> > reliably translated into any required representation. It is the most
> > fundamental type we have.
> 
> You cannot go from phys to page reliably unless you make assumptions
> about the phys, that's the whole issue here. If the callchain expects
> and intends to use a page it should really stay in a page and not take
> a side trip to phys.
> 
> I agree it would help to have some helpers that were more phys based
> for the callchains that want that since we are mixing two fairly
> different activities in these functions..

Yes, we are agreed here.

> 
> But making phys the only output out of the pool isn't going to be an
> improvement.

I don't know, phys was natural suggestion due to multiple
page_to_phys/phys_to_page conversions down the road.

> 
> BTW how does a struct page even work for decrypted memory? You can't
> actually use it that way right? If you try to access it then you'll
> get the wrong IPA and it should explode?
> 
> Maybe we should be blocking the struct page path entirely for CC
> shared?

It will simplify things. Ultimately, this series demonstrates that
CC users must be aware of the memory attribute (CC_SHARED) and struct
page is too generic.

Thanks

> 
> Jason
>