Re: [Qemu-devel] [RFC] Introduce vm_stop_permanent()

[Blue Swirl]

On Thu, Jul 28, 2011 at 9:22 PM, Luiz Capitulino <address@hidden> wrote:
> On Thu, 28 Jul 2011 20:04:58 +0200
> Jan Kiszka <address@hidden> wrote:
>
>> On 2011-07-28 20:00, Luiz Capitulino wrote:
>> > On Thu, 28 Jul 2011 19:52:31 +0200
>> > Jan Kiszka <address@hidden> wrote:
>> >
>> >> On 2011-07-28 19:48, Luiz Capitulino wrote:
>> >>> On Thu, 28 Jul 2011 14:39:23 -0300
>> >>> Luiz Capitulino <address@hidden> wrote:
>> >>>
>> >>>> On Thu, 28 Jul 2011 17:20:41 +0200
>> >>>> Jan Kiszka <address@hidden> wrote:
>> >>>>
>> >>>>> On 2011-07-28 17:18, Luiz Capitulino wrote:
>> >>>>>> On Thu, 28 Jul 2011 16:19:19 +0200
>> >>>>>> Jan Kiszka <address@hidden> wrote:
>> >>>>>>
>> >>>>>>> On 2011-07-28 15:37, Avi Kivity wrote:
>> >>>>>>>> On 07/28/2011 04:31 PM, Luiz Capitulino wrote:
>> >>>>>>>>> On Thu, 28 Jul 2011 10:23:22 +0300
>> >>>>>>>>> Avi Kivity<address@hidden>  wrote:
>> >>>>>>>>>
>> >>>>>>>>>>  On 07/28/2011 12:44 AM, Blue Swirl wrote:
>> >>>>>>>>>>  >  On Wed, Jul 27, 2011 at 9:42 PM, Luiz
>> >>>>>>>>> Capitulino<address@hidden>   wrote:
>> >>>>>>>>>>  >  >   This function should be used when the VM is not supposed 
>> >>>>>>>>>> to
>> >>>>>>>>> resume
>> >>>>>>>>>>  >  >   execution (eg. by issuing 'cont' monitor command).
>> >>>>>>>>>>  >  >
>> >>>>>>>>>>  >  >   Today, we allow the user to resume execution even when:
>> >>>>>>>>>>  >  >
>> >>>>>>>>>>  >  >     o the guest shuts down and -no-shutdown is used
>> >>>>>>>>>>  >  >     o there's a kvm internal error
>> >>>>>>>>>>  >  >     o loading the VM state with -loadvm or "loadvm" in the
>> >>>>>>>>> monitor fails
>> >>>>>>>>>>  >  >
>> >>>>>>>>>>  >  >   I think only badness can happen from the cases above.
>> >>>>>>>>>>  >
>> >>>>>>>>>>  >  I'd suppose a system_reset should bring the system back to
>> >>>>>>>>> sanity and
>> >>>>>>>>>>  >  then clear vm_permanent_stopped (where's -ly?)
>> >>>>>>>>>
>> >>>>>>>>> What's -ly?
>> >>>>>>>>>
>> >>>>>>>>
>> >>>>>>>> permanent-ly.
>> >>>>>>>>
>> >>>>>>>>>>  >  except maybe for KVM
>> >>>>>>>>>>  >  internal error if that can't be recovered. Then it would not 
>> >>>>>>>>>> very
>> >>>>>>>>>>  >  permanent anymore, so the name would need adjusting.
>> >>>>>>>>>>
>> >>>>>>>>>>  Currently, all kvm internal errors are recoverable by reset (and
>> >>>>>>>>>>  possibly by fiddling with memory/registers).
>> >>>>>>>>>
>> >>>>>>>>> Ok, but a poweroff in the guest isn't recoverable with system_reset
>> >>>>>>>>> right? Or does it depend on the guest?
>> >>>>>>>>
>> >>>>>>>> Right, it's not recoverable if you shut the power down where the 
>> >>>>>>>> tractor
>> >>>>>>>> beam is coupled to the main reactor.
>> >>>>>>>
>> >>>>>>> system_reset will bring all emulated devices back into their power-on
>> >>>>>>> state - unless we have remaining bugs to fix. Actually, one may 
>> >>>>>>> consider
>> >>>>>>> issuing this reset automatically on vm_start after "permant" vm_stop.
>> >>>>
>> >>>> The only permanent vm_stop we'd have is poweroff when -no-shutdown is 
>> >>>> used.
>> >>>>
>> >>>> Are you saying that system_reset should be able to recover from that 
>> >>>> too?
>> >>>
>> >>> It already does, so we don't have permanent stops.
>> >>
>> >> Exactly. We just have stops over inconsistent states that require a
>> >> reset to continue with anything useful.
>> >
>> > Yes. If I got you right, you suggest that we do the reset automatically.
>> >
>> > I think it's better to let the user do it, because s/he might want to
>> > do something else before resetting. For example, for the kvm error the
>> > user might want to save the vm state.
>>
>> Associating the reset with a cont means requesting an explicit action
>> from the user. I'm not suggesting to do the reset when the stop state is
>> entered.
>
> I see. But automatically resetting on cont might be unexpected to the
> user, even on a bad state.
>
> Another option would be to add a force option to cont, where the reset is
> done when the state is invalid (otherwise cont will return an error).
>
> I still prefer to let the user do it manually though.
>
>> > For the poweroff case with -no-shutdown it's probably fine, but I don't
>> > want to hard code special cases. It's better and easier to treat them all
>> > as "require system_reset to recover".
>>
>> In any case, we need to tag the current state as stopped-and-invalid or
>> so vs. a normal stop. That remains a valuable first step. How to deal
>> with that information is the second one.

I think the right way to fix this is to disable 'cont' until a
system_reset is issued. 'cont' should not perform reset but print an
error message about inconsistent state and suggest issuing a
'system_reset'.