Hi Marc,
On 5/26/20 6:11 PM, Marc Zyngier wrote:
> On a system that uses SPIs to implement MSIs (as it would be
> the case on a GICv2 system exposing a GICv2m to its guests),
> we deny the possibility of injecting SPIs on the in-atomic
> fast-path.
>
> This results in a very large amount of context-switches
> (roughly equivalent to twice the interrupt rate) on the host,
> and suboptimal performance for the guest (as measured with
> a test workload involving a virtio interface backed by vhost-net).
> Given that GICv2 systems are usually on the low-end of the spectrum
> performance wise, they could do without the aggravation.
>
> We solved this for GICv3+ITS by having a translation cache. But
> SPIs do not need any extra infrastructure, and can be immediately
> injected in the virtual distributor as the locking is already
> heavy enough that we don't need to worry about anything.
>
> This halves the number of context switches for the same workload.
>
> Signed-off-by: Marc Zyngier <maz@xxxxxxxxxx>
> ---
> arch/arm64/kvm/vgic/vgic-irqfd.c | 20 ++++++++++++++++----
> arch/arm64/kvm/vgic/vgic-its.c | 3 +--
> 2 files changed, 17 insertions(+), 6 deletions(-)
>
> diff --git a/arch/arm64/kvm/vgic/vgic-irqfd.c b/arch/arm64/kvm/vgic/vgic-irqfd.c
> index d8cdfea5cc96..11a9f81115ab 100644
> --- a/arch/arm64/kvm/vgic/vgic-irqfd.c
> +++ b/arch/arm64/kvm/vgic/vgic-irqfd.c
There is still a comment above saying
* Currently only direct MSI injection is supported.
> @@ -107,15 +107,27 @@ int kvm_arch_set_irq_inatomic(struct kvm_kernel_irq_routing_entry *e,
> struct kvm *kvm, int irq_source_id, int level,
> bool line_status)
> {
> - if (e->type == KVM_IRQ_ROUTING_MSI && vgic_has_its(kvm) && level) {
> + if (!level)
> + return -EWOULDBLOCK;
> +
> + switch (e->type) {
> + case KVM_IRQ_ROUTING_MSI: {
> struct kvm_msi msi;
>
> + if (!vgic_has_its(kvm))
> + return -EINVAL;
Shouldn't we return -EWOULDBLOCK by default?
QEMU does not use that path with GICv2m but in kvm_set_routing_entry() I
don't see any check related to the ITS.
> +
> kvm_populate_msi(e, &msi);
> - if (!vgic_its_inject_cached_translation(kvm, &msi))
> - return 0;
> + return vgic_its_inject_cached_translation(kvm, &msi);
> }
>
> - return -EWOULDBLOCK;
> + case KVM_IRQ_ROUTING_IRQCHIP:
> + /* Injecting SPIs is always possible in atomic context */
> + return vgic_irqfd_set_irq(e, kvm, irq_source_id, 1, line_status);
what about the mutex_lock(&kvm->lock) called from within
vgic_irqfd_set_irq/kvm_vgic_inject_irq/vgic_lazy_init
> +
> + default:
> + return -EWOULDBLOCK;
> + }
> }
>
> int kvm_vgic_setup_default_irq_routing(struct kvm *kvm)
> diff --git a/arch/arm64/kvm/vgic/vgic-its.c b/arch/arm64/kvm/vgic/vgic-its.c
> index c012a52b19f5..40cbaca81333 100644
> --- a/arch/arm64/kvm/vgic/vgic-its.c
> +++ b/arch/arm64/kvm/vgic/vgic-its.c
> @@ -757,9 +757,8 @@ int vgic_its_inject_cached_translation(struct kvm *kvm, struct kvm_msi *msi)
>
> db = (u64)msi->address_hi << 32 | msi->address_lo;
> irq = vgic_its_check_cache(kvm, db, msi->devid, msi->data);
> -
> if (!irq)
> - return -1;
> + return -EWOULDBLOCK;
>
> raw_spin_lock_irqsave(&irq->irq_lock, flags);
> irq->pending_latch = true;
>
Thanks
Eric
