Re: [PATCH, RFC 44/62] x86/mm: Set KeyIDs in encrypted VMAs for MKTME

[Peter Zijlstra]

On Fri, Jun 14, 2019 at 12:11:23PM -0700, Dave Hansen wrote:
> On 6/14/19 11:46 AM, Alison Schofield wrote:
> > On Fri, Jun 14, 2019 at 11:26:10AM -0700, Dave Hansen wrote:
> >> On 6/14/19 10:33 AM, Alison Schofield wrote:
> >>> Preserving the data across encryption key changes has not
> >>> been a requirement. I'm not clear if it was ever considered
> >>> and rejected. I believe that copying in order to preserve
> >>> the data was never considered.
> >>
> >> We could preserve the data pretty easily.  It's just annoying, though.
> >> Right now, our only KeyID conversions happen in the page allocator.  If
> >> we were to convert in-place, we'd need something along the lines of:
> >>
> >> 	1. Allocate a scratch page
> >> 	2. Unmap target page, or at least make it entirely read-only
> >> 	3. Copy plaintext into scratch page
> >> 	4. Do cache KeyID conversion of page being converted:
> >> 	   Flush caches, change page_ext metadata
> >> 	5. Copy plaintext back into target page from scratch area
> >> 	6. Re-establish PTEs with new KeyID
> > 
> > Seems like the 'Copy plaintext' steps might disappoint the user, as
> > much as the 'we don't preserve your data' design. Would users be happy
> > w the plain text steps ?
> 
> Well, it got to be plaintext because they wrote it to memory in
> plaintext in the first place, so it's kinda hard to disappoint them. :)
> 
> IMNHO, the *vast* majority of cases, folks will allocate memory and then
> put a secret in it.  They aren't going to *get* a secret in some
> mysterious fashion and then later decide they want to protect it.  In
> other words, the inability to convert it is pretty academic and not
> worth the complexity.

I'm not saying it is (required to preserve); but I do think it is
somewhat surprising to have an mprotect() call destroy content. It's
traditionally specified to not do that.