Re: [PATCH v3 02/18] target/s390x: Add ilen to unwind data

[David Hildenbrand]

On 30.09.19 17:03, Richard Henderson wrote:
> On 9/30/19 12:55 AM, David Hildenbrand wrote:
>>> What I think you should do instead is check env_neg(env)->icount_decr, 
>>> exactly
>>> like we do at the start of every basic block, and use that as an indication
>>> that you should exit back to the main loop.
>>
>> The issue is that when we return to the main loop we really have to
>> inject an interrupt - otherwise we might simply skip parts of the
>> (interruptible) instruction and continue with the next one.
> 
> Do we?  We will return to the main loop with psw.addr still pointing at MVCL,
> so the interrupt is delivered, and when the exception returns we restart the
> MVCL.  Or, an interrupt is not delivered and we restart from psw.addr and 
> still
> restart the MVCL.

If that's the case, then this should indeed work.

> 
> We probably would have to take a hard look at EXECUTE of MVCL to see if that 
> is
> actually restartable.  Not that it makes particular sense to want that combo,
> but it's required to work.
> 

Yes, this has to work. I'll look into that once I have some time.

>> However, with I/O interrupts, we can actually race against other VCPUs.
>> So the I/O interrupt might be gone by the time we arrive in the main loop.
> 
> Of course, but it's no different with this case than any other.  If the
> interrupt has already been handled, then we will simply restart the next TB as
> per normal.
Yeah, I was mostly concerned that "the next TB" will be "the next
instruction" and not "the original instruction again".

-- 

Thanks,

David / dhildenb