Re: qemu-kvm: do not allow vcpu stop with in progress PIO

[Avi Kivity <avi@redhat.com>]

On 02/10/2010 06:25 PM, Marcelo Tosatti wrote:
> On Wed, Feb 10, 2010 at 09:02:00AM +0200, Avi Kivity wrote:
>    
>> On 02/09/2010 10:58 PM, Marcelo Tosatti wrote:
>>      
>>> You're right... this should be enough to avoid a stop with uncomplete
>>> PIO (and this is what happens for MMIO already). The signal will not
>>> be dequeued, so KVM will complete_pio and exit before entering with
>>> -EAGAIN. Please review and queue for stable.
>>>
>>>        
>> Not right enough.  This is very fragile, we depend on the kernel
>> noticing the signal after completing pio but before starting
>> execution.  I don't think we guarantee that.
>>      
> As long as the signal is blocked, we do (and for older kernels too).
>    

I meant, that was not an intentional part of the design, but rather a 
side effect of the implementation.  We can pretend it was all part of a 
master plan and document it, though.

>> Maybe we should turn complete_pio/complete_mmio to an ioctl, so that
>> we can control what happens exactly.  Or maybe it's simplest to
>> document it as a feature and guarantee it.  There's some merit in it
>> - only guest execution is the nonatomic part, so we only interrupt
>> that.
>>      
> Right. So would you like a patch to x86.c to comment on this, on top of
> complete_pio / mmio completion?
>    

Documentation/kvm/api.txt.  Note it's not x86 specific.  Alex, can you 
check if ppc complies?

>>> qemu upstream needs a bit more work.
>>>        
>> Could be as simple as raising a blocked exception that is unmasked
>> by kvm, then entering the guest.
>>      
> The vcpu inner loop is not atomic in upstream as it is in qemu-kvm. It
> breaks out to process pending events way too easily.
>    

Hmm.  We can add an explicit call to KVM_RUN.

Note we need to loop there.  My 16-byte mmio patches (which never saw 
the light of day) split 16-byte mmios into two 8-byte mmios issued back 
to back, and we have to be prepared for that eventuality.

-- 
error compiling committee.c: too many arguments to function