lkml.org 
[lkml]   [2017]   [Aug]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v3 53/59] KVM: arm/arm64: GICv4: Hook vPE scheduling into vgic flush/sync
From
Date
On 28/08/17 19:17, Christoffer Dall wrote:
> On Mon, Jul 31, 2017 at 06:26:31PM +0100, Marc Zyngier wrote:
>> The redistributor needs to be told which vPE is about to be run,
>> and tells us whether there is any pending VLPI on exit.
>>
>> Let's add the scheduling calls to the vgic flush/sync functions,
>> allowing the VLPIs to be delivered to the guest.
>>
>> Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
>> ---
>> virt/kvm/arm/vgic/vgic-v4.c | 24 ++++++++++++++++++++++++
>> virt/kvm/arm/vgic/vgic.c | 4 ++++
>> virt/kvm/arm/vgic/vgic.h | 1 +
>> 3 files changed, 29 insertions(+)
>>
>> diff --git a/virt/kvm/arm/vgic/vgic-v4.c b/virt/kvm/arm/vgic/vgic-v4.c
>> index 50721c4e3da5..0a8deefbcf1c 100644
>> --- a/virt/kvm/arm/vgic/vgic-v4.c
>> +++ b/virt/kvm/arm/vgic/vgic-v4.c
>> @@ -119,6 +119,30 @@ void vgic_v4_teardown(struct kvm *kvm)
>> its_vm->vpes = NULL;
>> }
>>
>> +int vgic_v4_schedule(struct kvm_vcpu *vcpu, bool on)
>> +{
>> + int irq = vcpu->arch.vgic_cpu.vgic_v3.its_vpe.irq;
>> +
>> + if (!vgic_is_v4_capable(vcpu->kvm) || !irq)
>> + return 0;
>
> why do we need to check the its_vpe.irq here? This check is certainly
> not untuitive, as I don't understand what happened on a v4 capable
> system that somehow failed. Is it because a specific VM is configured
> to not use VLPIs, or?

Hmm. I think that's a debug leftover from my early attempt at making
things work with QEMU, which initializes things in the opposite order
as kvmtool. It should be removed (or replaced by a fat WARN_ON).

>> +
>> + /*
>> + * Before making the VPE resident, make sure the redistributor
>> + * expects us here.
>> + */
>> + if (on) {
>> + int err;
>> +
>> + err = irq_set_affinity(irq, cpumask_of(smp_processor_id()));
>
> This is pretty unintuitive, and coming here without having read your
> documentation may make people completely puzzled. Could we provide a
> pointer to the documentation that explains how the vpe irq hooks this
> all together?

Sure, will do.

>
>> + if (err) {
>> + kvm_err("failed irq_set_affinity IRQ%d (%d)\n", irq, err);
>> + return err;
>> + }
>> + }
>> +
>> + return its_schedule_vpe(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe, on);
>> +}
>> +
>
> I'd prefer this function be split into two and follow the vgic notation
> of having a flush and a sync function.

Yes, makes sense.

>> static struct vgic_its *vgic_get_its(struct kvm *kvm,
>> struct kvm_kernel_irq_routing_entry *irq_entry)
>> {
>> diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
>> index dfac894f6f03..9ab52108989d 100644
>> --- a/virt/kvm/arm/vgic/vgic.c
>> +++ b/virt/kvm/arm/vgic/vgic.c
>> @@ -721,6 +721,8 @@ void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
>> {
>> struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
>>
>> + WARN_ON(vgic_v4_schedule(vcpu, false));
>> +
>
> This is in the critical path, so would it be worth considering a static
> key to cater for non-GICv4 systems here?

Hey, for once I wasn't trying to optimize early! ;-) This would be
useful indeed, as I expect GICv4 systems to be the absolute minority for
the foreseeable future.

>
>> /* An empty ap_list_head implies used_lrs == 0 */
>> if (list_empty(&vcpu->arch.vgic_cpu.ap_list_head))
>> return;
>> @@ -733,6 +735,8 @@ void kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
>> /* Flush our emulation state into the GIC hardware before entering the guest. */
>> void kvm_vgic_flush_hwstate(struct kvm_vcpu *vcpu)
>> {
>> + WARN_ON(vgic_v4_schedule(vcpu, true));
>> +
>> /*
>> * If there are no virtual interrupts active or pending for this
>> * VCPU, then there is no work to do and we can bail out without
>> diff --git a/virt/kvm/arm/vgic/vgic.h b/virt/kvm/arm/vgic/vgic.h
>> index 1210bf4681dc..693b654acf4d 100644
>> --- a/virt/kvm/arm/vgic/vgic.h
>> +++ b/virt/kvm/arm/vgic/vgic.h
>> @@ -234,5 +234,6 @@ int update_lpi_config(struct kvm *kvm, struct vgic_irq *irq,
>> bool vgic_is_v4_capable(struct kvm *kvm);
>> int vgic_v4_init(struct kvm *kvm);
>> void vgic_v4_teardown(struct kvm *kvm);
>> +int vgic_v4_schedule(struct kvm_vcpu *vcpu, bool on);
>>
>> #endif
>> --
>> 2.11.0
>>
> Functionally, this looks correct.
>
> Thanks,
> -Christoffer
>

Thanks,

M.
--
Jazz is not dead. It just smells funny...

\
 
 \ /
  Last update: 2017-08-30 12:00    [W:0.100 / U:0.216 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site