On Wed, Jul 18, 2018 at 10:36:24AM -0700, Dave Hansen wrote:

On 07/17/2018 04:20 AM, Kirill A. Shutemov wrote:

quoted

Zero page is not encrypted and putting it into encrypted VMA produces
garbage.

We can map zero page with KeyID-0 into an encrypted VMA, but this would
be violation security boundary between encryption domains.

Why?  How is it a violation?

It only matters if they write secrets.  They can't write secrets to the
zero page.

I believe usage of zero page is wrong here. It would indirectly reveal
content of supposedly encrypted memory region.

I can see argument why it should be okay and I don't have very strong
opinion on this.

If folks see it's okay to use zero page in encrypted VMAs I can certainly
make it work.

Is this only because you accidentally inherited ->vm_page_prot on the
zero page PTE?

Yes, in previous patchset I mapped zero page with wrong KeyID. This is one
of possible fixes for this.

-- 
 Kirill A. Shutemov