From: Sean Christopherson <seanjc@google.com> Sent: Thursday, February 23, 2023 1:08 PM

On Thu, Feb 23, 2023, Michael Kelley (LINUX) wrote:

quoted

From: Dave Hansen <redacted> Sent: Thursday, February 23, 2023

12:42 PM

quoted

quoted

On 2/23/23 12:26, Dave Hansen wrote:

quoted

quoted

+       if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) {
+               /*
+               * Ensure fixmaps for IOAPIC MMIO respect memory encryption pgprot
+               * bits, just like normal ioremap():
+               */
+               if (x86_platform.hyper.is_private_mmio(phys))
+                       flags = pgprot_encrypted(flags);
+               else
+                       flags = pgprot_decrypted(flags);
+       }

...

quoted

It does seem a bit odd that there's a new CC_ATTR_GUEST_MEM_ENCRYPT
check wrapping this whole thing.  I guess the trip through
pgprot_decrypted() is harmless on normal platforms, though.

Yeah, that's _really_ odd.  Sean, were you trying to optimize away the
indirect call or something?

No, my thought was simply to require platforms that support GUEST_MEM_ENCRYPT
to
implement x86_platform.hyper.is_private_mmio, e.g. to avoid having to check if
is_private_mmio is NULL, to explicit document that non-Hyper-V encrypted guests
don't (yet) support private MMIO, and to add a bit of documentation around the
{de,en}crypted logic.

quoted

quoted

I would just expect the Hyper-V/vTOM code to leave
x86_platform.hyper.is_private_mmio alone unless it *knows* the platform has
private MMIO *and* CC_ATTR_GUEST_MEM_ENCRYPT.

Agreed.

quoted

Is there ever a case where CC_ATTR_GUEST_MEM_ENCRYPT==0 and he
Hyper-V/vTOM code would need to set x86_platform.hyper.is_private_mmio?

There's no such case.

I agree that gating with CC_ATTR_GUEST_MEM_ENCRYPT isn't really necessary.
Current upstream code always does the pgprot_decrypted(), and as you said,
that's a no-op on platforms with no memory encryption.

Right, but since is_private_mmio can be NULL, unless I'm missing something we'll
need an extra check no matter what, i.e. the alternative would be

	if (x86_platform.hyper.is_private_mmio &&
	    x86_platform.hyper.is_private_mmio(phys))
		flags = pgprot_encrypted(flags);
	else
		flags = pgprot_decrypted(flags);

I have no objection to that approach.  It does have the advantage of not needing
an indirect call for encrypted guests that don't support private MMIO, though
I can't imagine this code is performance sensitive.

Or statically set a default stub function for is_private_mmio() that returns "false".
Then there's no need to check for NULL, and only platforms that want to use it
have to code anything.  Several other entries in x86_platform have such defaults.

Michael