Re: [PATCH v8 4/5] x86/mm: Add noalias variants of set_memory_*crypted() functions

[Kirill A. Shutemov]

On Fri, Jun 24, 2022 at 06:19:23AM -0700, Dave Hansen wrote:
> On 6/8/22 19:52, Kuppuswamy Sathyanarayanan wrote:
> > set_memory_*crypted() functions are used to modify the "shared" page
> > attribute of the given memory. Using these APIs will modify the page
> > attributes of the aliased mappings (which also includes the direct
> > mapping).
> > 
> > But such aliased mappings modification is not desirable in use cases
> > like TDX guest, where the requirement is to create the shared mapping
> > without touching the direct map. It is used when allocating VMM shared
> > buffers using alloc_pages()/vmap()/set_memory_*crypted() API
> > combinations.
> > 
> > So to support such use cases, add support for noalias variants of
> > set_memory_*crypted() functions.
> 
> This changelog has a lot of words, but doesn't tell us much.
> 
> It basically says, we don't want to touch the direct map in use cases
> where we don't want to touch the direct map.  Not helpful.
> 
> The alias processing is there for a reason.  What is it?  Why don't you
> need alias processing for TDX?  Sure, you decided you don't want to
> touch the direct map, but *WHY*?  Why is it normally OK to touch the
> direct map, but not OK in this case?  Even better, why is it *DESIRABLE*
> to leave the direct map alone?  Lastly, why is this safe?  If alias
> processing was to protect us from something, why is losing that
> protection OK here?

The whole idea of alloc_pages()/vmap()/set_memory_decrypted() exercise is
to avoid direct map fragmentation.

Alias processing modifies direct mapping and highmap in addition to the
mapping user requested the modification for. Alias processing is required
to keep different mappings in sync, but here we want to avoid this because
it leads to direct mapping fragmentation.

Normally, direct mapping modifications are done for long-lasting
allocation where fragmentation is lesser concern. Here we have transient
allocation which may repeat over and over trashing the direct mapping.

Speaking of safety, I wanted initially claim that it is safe as only owner
of the allocation will touch the memory and it only does via the vmap
mapping.

*BUT* 

It made me thing about my recent story with load_unaligned_zeropad(). 
If we leave the page in direct mapping mapped as private and
load_unaligned_zeropad() will roll off to it, we will get SEPT violation
that will terminate the TD as it is considered unaccepted.

I think we must keep aliases in think. And vmap() doesn't make much sense
in this case :/

I urge you folks to consider DMA API again. Or have any other way to tap
into swiotlb pool.

-- 
 Kirill A. Shutemov