On Thu, Apr 22, 2021, Kenta Ishiguro wrote:
> This commit monitors IPI communication between vCPUs and leverages the
> relationship between vCPUs to select boost candidates.
>
> Cc: David Hildenbrand <david@xxxxxxxxxx>
> Signed-off-by: Kenta Ishiguro <kentaishiguro@xxxxxxxxxxxxxxxxxxxx>
> ---
> arch/x86/kvm/lapic.c | 14 ++++++++++++++
> arch/x86/kvm/vmx/vmx.c | 2 ++
> include/linux/kvm_host.h | 5 +++++
> virt/kvm/kvm_main.c | 26 ++++++++++++++++++++++++--
> 4 files changed, 45 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index cc369b9ad8f1..c8d967ddecf9 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -1269,6 +1269,18 @@ void kvm_apic_set_eoi_accelerated(struct kvm_vcpu *vcpu, int vector)
> }
> EXPORT_SYMBOL_GPL(kvm_apic_set_eoi_accelerated);
>
> +static void mark_ipi_receiver(struct kvm_lapic *apic, struct kvm_lapic_irq *irq)
> +{
> + struct kvm_vcpu *dest_vcpu;
> + u64 prev_ipi_received;
> +
> + dest_vcpu = kvm_get_vcpu_by_id(apic->vcpu->kvm, irq->dest_id);
> + if (READ_ONCE(dest_vcpu->sched_outed)) {
> + prev_ipi_received = READ_ONCE(dest_vcpu->ipi_received);
> + WRITE_ONCE(dest_vcpu->ipi_received, prev_ipi_received | (1 << apic->vcpu->vcpu_id));
The lack of "ull" on '1' actually means this will probably go sideways at >32 vCPUs.
Regardless, there needs to be an actual fallback path when the number of vCPUs
exceeds what the IPI-boost can support. That said, it should be quite easy to
allocate a bitmap to support an arbitrary number of vCPUs.
> + }
> +}
> +
> void kvm_apic_send_ipi(struct kvm_lapic *apic, u32 icr_low, u32 icr_high)
> {
> struct kvm_lapic_irq irq;
> @@ -1287,6 +1299,8 @@ void kvm_apic_send_ipi(struct kvm_lapic *apic, u32 icr_low, u32 icr_high)
>
> trace_kvm_apic_ipi(icr_low, irq.dest_id);
>
> + mark_ipi_receiver(apic, &irq);
> +
> kvm_irq_delivery_to_apic(apic->vcpu->kvm, apic, &irq, NULL);
> }
>
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 29b40e092d13..ced50935a38b 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -6718,6 +6718,8 @@ static fastpath_t vmx_vcpu_run(struct kvm_vcpu *vcpu)
> vmcs_write32(PLE_WINDOW, vmx->ple_window);
> }
>
> + WRITE_ONCE(vcpu->ipi_received, 0);
> +
> /*
> * We did this in prepare_switch_to_guest, because it needs to
> * be within srcu_read_lock.
> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> index 1b65e7204344..6726aeec03e7 100644
> --- a/include/linux/kvm_host.h
> +++ b/include/linux/kvm_host.h
> @@ -320,6 +320,11 @@ struct kvm_vcpu {
> #endif
> bool preempted;
> bool ready;
> + bool sched_outed;
> + /*
> + * The current implementation of strict boost supports up to 64 vCPUs
> + */
> + u64 ipi_received;
...
> struct kvm_vcpu_arch arch;
> struct kvm_dirty_ring dirty_ring;
> };
