On Tue, 31 Jul 2018 13:27:15 +1000
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Mon, Jul 30, 2018 at 11:59:00AM +0200, Greg Kurz wrote:
> > On Mon, 30 Jul 2018 10:41:45 +0200
> > Greg Kurz <groug@kaod.org> wrote:
> >   
> > > On Mon, 30 Jul 2018 15:57:15 +1000
> > > David Gibson <david@gibson.dropbear.id.au> wrote:  
> > [...]  
> > > > > > I'm pretty sure trying to change the accelerator on a qtest test just
> > > > > > doesn't make sense.  We'd need a different approach for testing cpu
> > > > > > hotplug against kvm & tcg backends.
> > > > > >       
> > > > > 
> > > > > The test starts QEMU, triggers the CPU hotplug code with a QMP command
> > > > > and checks the command didn't fail (or QEMU didn't crash, as it would
> > > > > have before commit b585395b655a)... I really don't understand what
> > > > > is wrong with that... Please elaborate.      
> > > > 
> > > > Well, ok, let me turn that around.  A test that doesn't rely on
> > > > controlling the guest side behaviour at all probably shouldn't be a
> > > > qtest based test, since that's what qtest is all about.
> > > >     
> > > 
> > > The CPU hotplug test doesn't seem to do anything on the guest side: it
> > > just checks that 'device_add' returns a response that isn't an error.
> > > I'm not aware that the guest is expected to have a specific behavior
> > > during 'device_add', apart from not crashing or hanging. That was the
> > > initial idea behind passing '-S' to ensure the guest doesn't run.
> > > 
> > > Your remark seems to be more general though... are you meaning that
> > > doing something like qtest_start("-machine accel=kvm:tcg") is just
> > > wrong ?  
> > 
> > The purpose of this test is simply to exercise a path in QEMU that
> > is only used with KVM, but it can also be achieved the other way
> > around:
> > 
> > @@ -189,7 +190,7 @@ static void xics_system_init(MachineState *machine, int nr_irqs, Error **errp)
> >      sPAPRMachineState *spapr = SPAPR_MACHINE(machine);
> >      Error *local_err = NULL;
> >  
> > -    if (kvm_enabled()) {
> > +    if (kvm_enabled() || qtest_enabled()) {
> >          if (machine_kernel_irqchip_allowed(machine) &&
> >              !xics_kvm_init(spapr, &local_err)) {
> > 
> > This will test the setup of the in-kernel XICS when run on a book3s host,
> > and fallback to emulated XICS otherwise (eg, travis).
> > 
> > Would this be more acceptable ?  
> 
> No, I don't think that will work.  With this we call into kvm related
> code via machine_kernel_irqchip_allowed() and xics_kvm_init() even in
> the qtest case.  If they work on a host which doesn't have KVM (say
> x86) it will only be by sheer accident.
> 

It's the other way around actually. The expected behaviour would be
for machine_kernel_irqchip_allowed(machine) and/or xics_kvm_init()
to fail and to fallback to emulated XICS if run without a proper KVM.
This means no behavior change for this test when run on a x86 host.

The issue is when we run this with KVM actually, because the XICS KVM
code obviously needs... KVM to be initialized, which won't happen
with qtest :)