Re: [PATCH v6 13/18] KVM: arm64: Support SDEI_EVENT_{COMPLETE,COMPLETE_AND_RESUME} hypercall

[Gavin Shan <gshan@redhat.com>]

Hi Oliver,

On 5/2/22 3:38 PM, Oliver Upton wrote:
> On Mon, May 02, 2022 at 02:19:30PM +0800, Gavin Shan wrote:
>> On 5/1/22 2:50 PM, Oliver Upton wrote:
>>> On Sun, Apr 03, 2022 at 11:39:06PM +0800, Gavin Shan wrote:
>>>> This supports SDEI_EVENT_{COMPLETE, COMPLETE_AND_RESUME} hypercall.
>>>> They are used by guest to notify the completion of event in its
>>>> handler. The previously interrupted or preempted context is restored
>>>> like below.
>>>>
>>>>      * x0 - x17, PC and PState are restored to what values we had in
>>>>        the interrupted or preempted context.
>>>>
>>>>      * If it's SDEI_EVENT_COMPLETE_AND_RESUME hypercall, IRQ exception
>>>>        is injected.
>>>
>>> I don't think that's how COMPLETE_AND_RESUME works. The caller specifies an
>>> address at which it would like to begin execution within the client
>>> exception level.
>>>
>>> SDEI spec suggests this behaves like a synchronous exception. DEN 0054C
>>> 5.2.2 'Event Resume Context' speaks more about how it is supposed to
>>> work.
>>>
>>
>> It's actually the linux convention. If the event handler, which was
>> specified in previous hypercall to EVENT_REGISTER, returns success,
>> the (linux) client calls into COMPLETE_AND_RESUME and the resume
>> address is specified with FIQ vector offset. More details can be
>> found from arch/arm64/kernel::sdei.c::do_sdei_event().
> 
> Right -- but look at what its doing. It returns the address at which it
> wants to resume execution.
> 
> arch/arm64/kernel.entry.S::__sdei_asm_handler winds up passing this as
> an argument to COMPLETE_AND_RESUME. Also, what would happen if we run
> something that isn't Linux inside of KVM? This is why I suggested
> implementing COMPLETE_AND_RESUME in line with the specification, not
> based on what the kernel is presently doing.
> 

Indeed. The address for the resumed execution is passed by x1 when
COMPLETE_AND_RESUME is called. I will figure this out in next revision.
I don't think we can have the assumption that the guest is linux.

Thanks again for your review and comments :)

Thanks,
Gavin