Re: [Qemu-devel] [PATCH/s390-next 2/3] s390x/cpumodel: add zpci, aen and ais facilities

[Cornelia Huck]

On Thu, 13 Jul 2017 14:29:55 +0200
Christian Borntraeger <address@hidden> wrote:

> On 07/13/2017 02:11 PM, Cornelia Huck wrote:
> diff --git a/target/s390x/kvm.c b/target/s390x/kvm.c
> >> index 78ebe83..1901153 100644
> >> --- a/target/s390x/kvm.c
> >> +++ b/target/s390x/kvm.c
> >> @@ -302,6 +302,9 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
> >>          }
> >>      }
> >>  
> >> +    /* Try to enable AIS facility */
> >> +    kvm_vm_enable_cap(s, KVM_CAP_S390_AIS, 0);  
> > 
> > What happens if you fail to enable it? You probably don't want to allow
> > the feature bit, then?  
> 
> Then this bit is off. This call will enable it in the kernel, if that fails
> the kernel will return this bit as disabled.

Looked at the kernel code again. I thought there was more to it, but I
misremembered. No further complaints here.

> >   
> >> +
> >>      qemu_mutex_init(&qemu_sigp_mutex);
> >>  
> >>      return 0;  
> > 
> > Let's summarize to make sure that I'm not confused:
> > 
> > - Starting with zEC12 GA1, we provide zpci, aen, ais in the full model  
> yes
> 
> > - Starting with zEC12 GA1, we provide zpci and aen in the default model  
> yes. ais has to be enabled manually for z12 and z13.
> The alternative is to have ais as part of the default model. This has the big
> disadvantage that -cpu zEC12 and -cpu z13 will stop working for all available
> distro kernels. I believe that a working -cpu z13 is more important than the
> need to manually enable ais.

Agreed.

> For the future
>  - We can make ais part of the default model for a future system when that 
> happens since
>    the features for a new system will require a new host kernel anyway

Sounds fine.

>  - We can also make ais part of the default model for a future machine type 
> (e.g. 2.13)
>    when we believe that the world has moved on to a newer kernel

I'd rather avoid relying on that.

> 
> 
> > - In the host model, we add zpci and aen; they might be switched off
> >   after applying the found model  
> 
> We also get ais from the kernel, so the host model will have zpci,ais and
> aen
> 
> > - Compat for 2.9 and earlier switches off zpci, aen, ais  
> 
> yes 
> 
> > - We unconditionally enable the kvm part of ais  
> 
> We tell the KVM code in the kernel to enable the facility bit (before the cpu
> model might take it way) and if QEMU really uses ais, that the kernel does
> the right thing then

Yeah, that's fine, see above.

> > 
> > I'm still not sure what's supposed to happen with new qemu + old kernel
> > (no ais) + full zEC12 GA1 or later model.  
> 
> We enable aen and zpci, but disable ais for that guest. In theory a guest
> can drive PCI devices without AIS. This is a valid configuration since zpci
> does not require ais. 

Yes, also fine. Looking at the kernel code cleared things up :)

> The fact that Linux requires ais to use PCI is unfortunate but that could 
> be "fixed" Linux if necessary.

I'm not sure there's a need for this. Once a change like this has
landed in distro kernels, the same distros will also have the ais
changes in their hypervisor kernels.

Thanks for spelling things out again, this stuff always gives me a
headache.