Re: [PATCH v3] target/s390x/kvm: Enable adapter interruption suppression again

[Cornelia Huck]

On Tue, 21 Jan 2020 11:05:18 -0500
Matthew Rosato <address@hidden> wrote:

> On 1/21/20 10:22 AM, Thomas Huth wrote:
> > On 21/01/2020 15.46, Cornelia Huck wrote:  
> >> On Tue, 21 Jan 2020 09:33:02 -0500
> >> Matthew Rosato <address@hidden> wrote:
> >>  
> >>> On 1/20/20 12:27 PM, Cornelia Huck wrote:  
> >>>> On Mon, 20 Jan 2020 14:24:41 +0100
> >>>> Thomas Huth <address@hidden> wrote:
> >>>>      
> >>>>> The AIS feature has been disabled late in the v2.10 development cycle 
> >>>>> since
> >>>>> there were some issues with migration (see commit 3f2d07b3b01ea61126b -
> >>>>> "s390x/ais: for 2.10 stable: disable ais facility"). We originally 
> >>>>> wanted
> >>>>> to enable it again for newer machine types, but apparently we forgot to 
> >>>>> do
> >>>>> this so far. Let's do it for the new s390-ccw-virtio-5.0 machine now.
> >>>>>
> >>>>> While at it, also add a more verbose comment why we need the *_allowed()
> >>>>> wrappers in s390-virtio-ccw.c.
> >>>>>
> >>>>> Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=1756946
> >>>>> Reviewed-by: David Hildenbrand <address@hidden>
> >>>>> Signed-off-by: Thomas Huth <address@hidden>
> >>>>> ---
> >>>>>    v3: Moved "s390mc->kvm_ais_allowed = false" to the end of the 
> >>>>> function
> >>>>>
> >>>>>    hw/s390x/s390-virtio-ccw.c         | 20 +++++++++++++++++---
> >>>>>    include/hw/s390x/s390-virtio-ccw.h |  3 +++
> >>>>>    target/s390x/kvm.c                 |  9 ++++++---
> >>>>>    3 files changed, 26 insertions(+), 6 deletions(-)  
> >>>>      
> >>>>> diff --git a/target/s390x/kvm.c b/target/s390x/kvm.c
> >>>>> index 15260aeb9a..cf4fb4f2d9 100644
> >>>>> --- a/target/s390x/kvm.c
> >>>>> +++ b/target/s390x/kvm.c
> >>>>> @@ -365,10 +365,13 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
> >>>>>        /*
> >>>>>         * The migration interface for ais was introduced with kernel 
> >>>>> 4.13
> >>>>>         * but the capability itself had been active since 4.12. As 
> >>>>> migration
> >>>>> -     * support is considered necessary let's disable ais in the 2.10
> >>>>> -     * machine.
> >>>>> +     * support is considered necessary, we only try to enable this for
> >>>>> +     * newer machine types if KVM_CAP_S390_AIS_MIGRATION is available.
> >>>>>         */
> >>>>> -    /* kvm_vm_enable_cap(s, KVM_CAP_S390_AIS, 0); */
> >>>>> +    if (kvm_ais_allowed() &&
> >>>>> +        kvm_check_extension(s, KVM_CAP_S390_AIS_MIGRATION)) {  
> >>>>
> >>>> Hnm, we actually need a kernel irqchip with the kvm flic to get ais to
> >>>> work; else we'll fail with
> >>>>
> >>>> qemu-system-s390x: Failed to inject airq with AIS supported
> >>>>
> >>>> in the kernel_irqchip=off case, as we won't have an I/O adapter
> >>>> registered.
> >>>>
> >>>> Adding 'kvm_kernel_irqchip_required() &&' seems to do the trick;
> >>>> comments?
> >>>>      
> >>>
> >>> In spirit, I agree with this idea.  But, a quick test shows that putting
> >>> this check here results in ais=off for the 'none' machine case (libvirt
> >>> capabilities detection).  I think we have to only look at
> >>> kvm_kernel_irqchip_required() when working with a real machine.  
> >>
> >> Sigh, I think you're right again. We need to check for the 'none'
> >> machine here; but I can't think of a non-ugly way to do so...  
> > 
> > I think it might work when using kvm_kernel_irqchip_allowed() instead of
> > kvm_kernel_irqchip_required() ... Matthew, could you please give it a
> > try with this patch on top of mine:
> >   
> 
> Sure.
> 
> Libvirt detection works with this patch.

Excellent.

> 
> Alternatively, if I run qemu with kernel_irqchip=off and ais=true, I get:
> qemu-system-s390x: Some features requested in the CPU model are not 
> available in the configuration: ais
> 
> Which was the same result as Connie's proposal.

Yep, that's the expected behaviour.

> 
> It reads a bit odd to me at first, but looking at the code quick I think 
> this is the right answer - kvm_kernel_irqchip_allowed() will only return 
> false when kernel_irqchip has been forced off as above, whereas 
> kernel_irqchip_required will also return false in the case where no 
> setting was specified (this is what tripped libvirt up).
> 
> Looks good to me, thanks Thomas.

Thanks for testing!

Thomas, I guess you'll send a v4?