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

[Dave Hansen <dave.hansen@intel.com>]

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?