On Wed, 2022-08-17 at 00:07 +0000, Sean Christopherson wrote:
> On Thu, Aug 11, 2022, Paolo Bonzini wrote:
> > Do not return true from kvm_apic_has_events, and consequently from
> > kvm_vcpu_has_events, if the vCPU is not going to process an INIT.
> >
> > Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > ---
> > arch/x86/include/asm/kvm_host.h | 1 +
> > arch/x86/kvm/i8259.c | 2 +-
> > arch/x86/kvm/lapic.h | 2 +-
> > arch/x86/kvm/x86.c | 5 +++++
> > arch/x86/kvm/x86.h | 5 -----
> > 5 files changed, 8 insertions(+), 7 deletions(-)
> >
> > diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> > index 293ff678fff5..1ce4ebc41118 100644
> > --- a/arch/x86/include/asm/kvm_host.h
> > +++ b/arch/x86/include/asm/kvm_host.h
> > @@ -2042,6 +2042,7 @@ void __user *__x86_set_memory_region(struct kvm *kvm, int id, gpa_t gpa,
> > u32 size);
> > bool kvm_vcpu_is_reset_bsp(struct kvm_vcpu *vcpu);
> > bool kvm_vcpu_is_bsp(struct kvm_vcpu *vcpu);
> > +bool kvm_vcpu_latch_init(struct kvm_vcpu *vcpu);
> >
> > bool kvm_intr_is_single_vcpu(struct kvm *kvm, struct kvm_lapic_irq *irq,
> > struct kvm_vcpu **dest_vcpu);
> > diff --git a/arch/x86/kvm/i8259.c b/arch/x86/kvm/i8259.c
> > index e1bb6218bb96..177555eea54e 100644
> > --- a/arch/x86/kvm/i8259.c
> > +++ b/arch/x86/kvm/i8259.c
> > @@ -29,9 +29,9 @@
> > #include <linux/mm.h>
> > #include <linux/slab.h>
> > #include <linux/bitops.h>
> > -#include "irq.h"
> > +#include <linux/kvm_host.h>
> >
> > -#include <linux/kvm_host.h>
> > +#include "irq.h"
> > #include "trace.h"
> >
> > #define pr_pic_unimpl(fmt, ...) \
> > diff --git a/arch/x86/kvm/lapic.h b/arch/x86/kvm/lapic.h
> > index 117a46df5cc1..12577ddccdfc 100644
> > --- a/arch/x86/kvm/lapic.h
> > +++ b/arch/x86/kvm/lapic.h
> > @@ -225,7 +225,7 @@ static inline bool kvm_vcpu_apicv_active(struct kvm_vcpu *vcpu)
> >
> > static inline bool kvm_apic_has_events(struct kvm_vcpu *vcpu)
> > {
> > - return lapic_in_kernel(vcpu) && vcpu->arch.apic->pending_events;
> > + return lapic_in_kernel(vcpu) && vcpu->arch.apic->pending_events && !kvm_vcpu_latch_init(vcpu);
>
> Blech, the kvm_apic_has_events() name is awful (as is pending_events), e.g. it
> really should be kvm_apic_has_pending_sipi_or_init().
110% agree.
>
> To avoid the odd kvm_vcpu_latch_init() declaration and the long line above, what
> if we open code this in kvm_vcpu_has_events() like we do for NMI, SMI, etc...?
>
> And as follow-up, I would love to rename kvm_vcpu_latch_init() => kvm_vcpu_init_blocked(),
> kvm_apic_has_events(), and pending_events.
>
> E.g. for this patch just do:
>
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 9f11b505cbee..559900736a71 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -12533,7 +12533,8 @@ static inline bool kvm_vcpu_has_events(struct kvm_vcpu *vcpu)
> if (!list_empty_careful(&vcpu->async_pf.done))
> return true;
>
> - if (kvm_apic_has_events(vcpu))
> + /* comment explaning that SIPIs are dropped in this case. */
> + if (kvm_apic_has_events(vcpu) && !kvm_vcpu_latch_init(vcpu))
> return true;
I personally don't know if I prefer this or the original patch.
>
> if (vcpu->arch.pv.pv_unhalted)
>
While reviwing this, I noticed that we have this code:
static bool svm_apic_init_signal_blocked(struct kvm_vcpu *vcpu)
{
struct vcpu_svm *svm = to_svm(vcpu);
/*
* TODO: Last condition latch INIT signals on vCPU when
* vCPU is in guest-mode and vmcb12 defines intercept on INIT.
* To properly emulate the INIT intercept,
* svm_check_nested_events() should call nested_svm_vmexit()
* if an INIT signal is pending.
*/
return !gif_set(svm) ||
(vmcb_is_intercept(&svm->vmcb->control, INTERCEPT_INIT));
}
Is this workaround still needed? svm_check_nested_events does check the apic's INIT/SIPI status.
Currently the '.apic_init_signal_blocked' is called from kvm_vcpu_latch_init which itself is
currently called from kvm_vcpu_latch_init which happens after we would vmexit if INIT is intercepted by nested
hypervisor.
Best regards,
Maxim Levitsky
