Re: [PATCH] kvm: Remove KVM_CREATE_IRQCHIP support assumption

[Andrew Jones <ajones@ventanamicro.com>]

On Mon, Jul 24, 2023 at 11:53:39AM +0200, Thomas Huth wrote:
> On 22/07/2023 08.21, Andrew Jones wrote:
> > Since Linux commit 00f918f61c56 ("RISC-V: KVM: Skeletal in-kernel AIA
> > irqchip support") checking KVM_CAP_IRQCHIP returns non-zero when the
> > RISC-V platform has AIA. The cap indicates KVM supports at least one
> > of the following ioctls:
> > 
> >    KVM_CREATE_IRQCHIP
> >    KVM_IRQ_LINE
> >    KVM_GET_IRQCHIP
> >    KVM_SET_IRQCHIP
> >    KVM_GET_LAPIC
> >    KVM_SET_LAPIC
> > 
> > but the cap doesn't imply that KVM must support any of those ioctls
> > in particular. However, QEMU was assuming the KVM_CREATE_IRQCHIP
> > ioctl was supported. Stop making that assumption by introducing a
> > KVM parameter that each architecture which supports KVM_CREATE_IRQCHIP
> > sets. Adding parameters isn't awesome, but given how the
> > KVM_CAP_IRQCHIP isn't very helpful on its own, we don't have a lot of
> > options.
> > 
> > Signed-off-by: Andrew Jones <ajones@ventanamicro.com>
> > ---
> > 
> > While this fixes booting guests on riscv KVM with AIA it's unlikely
> > to get merged before the QEMU support for KVM AIA[1] lands, which
> > would also fix the issue. I think this patch is still worth considering
> > though since QEMU's assumption is wrong.
> > 
> > [1] https://lore.kernel.org/all/20230714084429.22349-1-yongxuan.wang@sifive.com/
> > 
> > 
> >   accel/kvm/kvm-all.c    | 5 ++++-
> >   include/sysemu/kvm.h   | 1 +
> >   target/arm/kvm.c       | 3 +++
> >   target/i386/kvm/kvm.c  | 2 ++
> >   target/s390x/kvm/kvm.c | 3 +++
> >   5 files changed, 13 insertions(+), 1 deletion(-)
> > 
> > diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
> > index 373d876c0580..0f5ff8630502 100644
> > --- a/accel/kvm/kvm-all.c
> > +++ b/accel/kvm/kvm-all.c
> > @@ -86,6 +86,7 @@ struct KVMParkedVcpu {
> >   };
> >   KVMState *kvm_state;
> > +bool kvm_has_create_irqchip;
> >   bool kvm_kernel_irqchip;
> >   bool kvm_split_irqchip;
> >   bool kvm_async_interrupts_allowed;
> > @@ -2377,8 +2378,10 @@ static void kvm_irqchip_create(KVMState *s)
> >           if (s->kernel_irqchip_split == ON_OFF_AUTO_ON) {
> >               error_report("Split IRQ chip mode not supported.");
> >               exit(1);
> > -        } else {
> > +        } else if (kvm_has_create_irqchip) {
> >               ret = kvm_vm_ioctl(s, KVM_CREATE_IRQCHIP);
> > +        } else {
> > +            return;
> >           }
> >       }
> >       if (ret < 0) {
> 
> I think I'd do this differntly... at the beginning of the function, there is
> a check for kvm_check_extension(s, KVM_CAP_IRQCHIP) etc. ... I think you
> could now replace that check with a simple
> 
> 	if (!kvm_has_create_irqchip) {
> 		return;
> 	}
> 
> The "kvm_vm_enable_cap(s, KVM_CAP_S390_IRQCHIP, 0)" of course has to be
> moved to the target/s390x/kvm/kvm.c file, too.

Actually, once we've moved the s390 cap enablement to the s390 file we can
just drop the whole if-else chain. We don't want the
if (!kvm_has_create_irqchip) at the top because we want to try
kvm_arch_irqchip_create() even when kvm_has_create_irqchip is false, and
we don't need to check KVM_CREATE_IRQCHIP for kvm_arch_irqchip_create()
either. Keeping the check, as it is above in this v1, of
kvm_has_create_irqchip for KVM_CREATE_IRQCHIP is still necessary, though.

Thanks,
drew