Re: [PATCH v3 1/2] KVM: s390: Don't indicate suppression on dirtying, failing memop

[David Hildenbrand]

On 12.05.22 15:51, Christian Borntraeger wrote:
> 
> 
> Am 12.05.22 um 15:22 schrieb David Hildenbrand:
>> On 12.05.22 15:10, Janis Schoetterl-Glausch wrote:
>>> If user space uses a memop to emulate an instruction and that
>>> memop fails, the execution of the instruction ends.
>>> Instruction execution can end in different ways, one of which is
>>> suppression, which requires that the instruction execute like a no-op.
>>> A writing memop that spans multiple pages and fails due to key
>>> protection may have modified guest memory, as a result, the likely
>>> correct ending is termination. Therefore, do not indicate a
>>> suppressing instruction ending in this case.
>>
>> I think that is possibly problematic handling.
>>
>> In TCG we stumbled in similar issues in the past for MVC when crossing
>> page boundaries. Failing after modifying the first page already
>> seriously broke some user space, because the guest would retry the
>> instruction after fixing up the fault reason on the second page: if
>> source and destination operands overlap, you'll be in trouble because
>> the input parameters already changed.
>>
>> For this reason, in TCG we make sure that all accesses are valid before
>> starting modifications.
>>
>> See target/s390x/tcg/mem_helper.c:do_helper_mvc with access_prepare()
>> and friends as an example.
>>
>> Now, I don't know how to tackle that for KVM, I just wanted to raise
>> awareness that injecting an interrupt after modifying page content is
>> possible dodgy and dangerous.
> 
> this is really special and only for key protection crossing pages.
> Its been done since the 70ies in that way on z/VM. The architecture
> is and was always written in a way to allow termination for this
> case for hypervisors.

Just so I understand correctly: all instructions that a hypervisor with
hardware virtualization is supposed to emulate are "written in a way to
allow termination", correct? That makes things a lot easier.

-- 
Thanks,

David / dhildenb