Re: [PATCH v2] x86/kvm: Disable KVM_ASYNC_PF_SEND_ALWAYS

[Andy Lutomirski]

> On Apr 9, 2020, at 7:32 AM, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> 
> ïOn 09/04/20 16:13, Andrew Cooper wrote:
>>> On 09/04/2020 13:47, Paolo Bonzini wrote:
>>> On 09/04/20 06:50, Andy Lutomirski wrote:
>>>> The small
>>>> (or maybe small) one is that any fancy protocol where the guest
>>>> returns from an exception by doing, logically:
>>>> 
>>>> Hey I'm done;  /* MOV somewhere, hypercall, MOV to CR4, whatever */
>>>> IRET;
>>>> 
>>>> is fundamentally racy.  After we say we're done and before IRET, we
>>>> can be recursively reentered.  Hi, NMI!
>>> That's possible in theory.  In practice there would be only two levels
>>> of nesting, one for the original page being loaded and one for the tail
>>> of the #VE handler.  The nested #VE would see IF=0, resolve the EPT
>>> violation synchronously and both handlers would finish.  For the tail
>>> page to be swapped out again, leading to more nesting, the host's LRU
>>> must be seriously messed up.
>>> 
>>> With IST it would be much messier, and I haven't quite understood why
>>> you believe the #VE handler should have an IST.
>> 
>> Any interrupt/exception which can possibly occur between a SYSCALL and
>> re-establishing a kernel stack (several instructions), must be IST to
>> avoid taking said exception on a user stack and being a trivial
>> privilege escalation.
> 
> Doh, of course.  I always confuse SYSCALL and SYSENTER.
> 
>> Therefore, it doesn't really matter if KVM's paravirt use of #VE does
>> respect the interrupt flag.  It is not sensible to build a paravirt
>> interface using #VE who's safety depends on never turning on
>> hardware-induced #VE's.
> 
> No, I think we wouldn't use a paravirt #VE at this point, we would use
> the real thing if available.
> 
> It would still be possible to switch from the IST to the main kernel
> stack before writing 0 to the reentrancy word.
> 
> 

Almost but not quite. We do this for NMI-from-usermode, and itâs ugly. But we canât do this for NMI-from-kernel or #VE-from-kernel because there might not be a kernel stack.  Trying to hack around this wonât be pretty.

Frankly, I think that we shouldnât even try to report memory failure to the guest if it happens with interrupts off. Just kill the guest cleanly and keep it simple. Or inject an intentionally unrecoverable IST exception.