Re: [PATCH 1/2] x86/tdx: Extract GET_INFO call from get_cc_mask()

[Dave Hansen]

On 10/31/22 12:27, Andi Kleen wrote:
>> Moving panic() after earlyprintk working is not good idea as it exposes
>> kernel more: by the time we already have full #VE handler.
> 
> It should be fine to move since there is no user land at this point (the
> attack requires user land)

Maybe I'm misunderstanding the exposure.  A normal MMIO #VE goes
something like this:

	1. %rax points to some MMIO
	2. Kernel executes: mov (%rax),%rbx, trying to read MMIO
	3. #VE handler is triggered
	4. Handler emulates the 'mov' with instruction decoding
	5. Handler asks the VMM what the value of %rax should be
	6. Handler puts VMM value in %rax
	7. Return from #VE

I think the attack scenario subverts a normal MMIO to the following
(changes from the normal flow are marked with *):

	*1. %rax points to some private kernel memory, VMM removes
	    Secure-EPT entry for that memory.
	 2. Kernel executes: mov (%rax),%rbx as part of normal kernel
	    execution, not an MMIO read.
	 3. #VE handler is triggered, assuming a MMIO read
	 4. Handler emulates the 'mov' with instruction decoding
	 5. Handler asks the VMM what the value of %rax should be
	*6. Handler puts (malicious) VMM value in %rax
	 7. Return from #VE
	*8. Now the guest kernel is running with an attacker-controlled
	    %rax

This effectively gives the attacker the ability to override the contents
of a memory read.

Am I misunderstanding the attack scenario?  I don't see guest userspace
needing to be involved at all.