On 24/09/15 23:31, Shannon Zhao wrote:
> When calling perf_event_create_kernel_counter to create perf_event,
> assign a overflow handler. Then when perf event overflows, set
> irq_pending and call kvm_vcpu_kick() to sync the interrupt.
>
> Signed-off-by: Shannon Zhao <shannon.zhao at linaro.org>
> ---
> arch/arm/kvm/arm.c | 4 ++++
> include/kvm/arm_pmu.h | 2 ++
> virt/kvm/arm/pmu.c | 54 ++++++++++++++++++++++++++++++++++++++++++++++++++-
> 3 files changed, 59 insertions(+), 1 deletion(-)
>
> diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c
> index ce404a5..3fca263 100644
> --- a/arch/arm/kvm/arm.c
> +++ b/arch/arm/kvm/arm.c
> @@ -28,6 +28,7 @@
> #include <linux/sched.h>
> #include <linux/kvm.h>
> #include <trace/events/kvm.h>
> +#include <kvm/arm_pmu.h>
>
> #define CREATE_TRACE_POINTS
> #include "trace.h"
> @@ -554,6 +555,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
> kvm_vgic_sync_hwstate(vcpu);
> preempt_enable();
> kvm_timer_sync_hwstate(vcpu);
> + kvm_pmu_sync_hwstate(vcpu);
> continue;
> }
>
> @@ -604,6 +606,8 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run)
>
> kvm_timer_sync_hwstate(vcpu);
>
> + kvm_pmu_sync_hwstate(vcpu);
> +
> ret = handle_exit(vcpu, run, ret);
> }
The code around here is about to change with Christopher's patches. Most
importantly, virtual devices must signal their changes before we touch
the vgic. I suspect this will have some impacts, see below.
>
> diff --git a/include/kvm/arm_pmu.h b/include/kvm/arm_pmu.h
> index 9293133..953c400 100644
> --- a/include/kvm/arm_pmu.h
> +++ b/include/kvm/arm_pmu.h
> @@ -38,6 +38,7 @@ struct kvm_pmu {
> };
>
> #ifdef CONFIG_KVM_ARM_PMU
> +void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu);
> unsigned long kvm_pmu_get_counter_value(struct kvm_vcpu *vcpu, u32 select_idx);
> void kvm_pmu_disable_counter(struct kvm_vcpu *vcpu, u32 val);
> void kvm_pmu_enable_counter(struct kvm_vcpu *vcpu, u32 val);
> @@ -45,6 +46,7 @@ void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u32 val);
> void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u32 data,
> u32 select_idx);
> #else
> +void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu) {}
> unsigned long kvm_pmu_get_counter_value(struct kvm_vcpu *vcpu, u32 select_idx)
> {
> return 0;
> diff --git a/virt/kvm/arm/pmu.c b/virt/kvm/arm/pmu.c
> index 18637c9..ca7e849 100644
> --- a/virt/kvm/arm/pmu.c
> +++ b/virt/kvm/arm/pmu.c
> @@ -21,6 +21,7 @@
> #include <linux/perf_event.h>
> #include <asm/kvm_emulate.h>
> #include <kvm/arm_pmu.h>
> +#include <kvm/arm_vgic.h>
>
> static void kvm_pmu_set_evttyper(struct kvm_vcpu *vcpu, u32 idx, u32 val)
> {
> @@ -62,6 +63,56 @@ static void kvm_pmu_stop_counter(struct kvm_vcpu *vcpu, u32 select_idx)
> }
>
> /**
> + * kvm_pmu_sync_hwstate - sync pmu state for cpu
> + * @vcpu: The vcpu pointer
> + *
> + * Inject virtual PMU IRQ if IRQ is pending for this cpu.
> + */
> +void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
> +{
> + struct kvm_pmu *pmu = &vcpu->arch.pmu;
> +
> + if (pmu->irq_pending && (pmu->irq_num != -1)) {
How likely is that pmu->irq_num could be -1? I don't think the interrupt
can be made optional.
> + kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id, pmu->irq_num, 1);
> + pmu->irq_pending = false;
> + }
So you're signalling the interrupt as an edge, not a level. Is that an
accurate modelling of the PMU interrupt?
My hunch is that the interrupt should still be pending until the guest
has cleared the overflow condition (by writing to PMOVSCLR_EL0). You can
probably lift most of that logic from Christoffer's rework of the timer
state.
> +}
> +
> +/**
> + * When perf event overflows, set irq_pending and call kvm_vcpu_kick() to inject
> + * the interrupt.
> + */
> +static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
> + struct perf_sample_data *data,
> + struct pt_regs *regs)
> +{
> + struct kvm_pmc *pmc = perf_event->overflow_handler_context;
> + struct kvm_vcpu *vcpu = pmc->vcpu;
> + struct kvm_pmu *pmu = &vcpu->arch.pmu;
> + int idx = pmc->idx;
> +
> + if (!vcpu_mode_is_32bit(vcpu)) {
> + if ((vcpu_sys_reg(vcpu, PMINTENSET_EL1) >> idx) & 0x1) {
> + __set_bit(idx,
> + (unsigned long *)&vcpu_sys_reg(vcpu, PMOVSSET_EL0));
> + __set_bit(idx,
> + (unsigned long *)&vcpu_sys_reg(vcpu, PMOVSCLR_EL0));
> + pmu->irq_pending = true;
> + kvm_vcpu_kick(vcpu);
> + }
> + } else {
> + if ((vcpu_cp15(vcpu, c9_PMINTENSET) >> idx) & 0x1) {
> + __set_bit(idx,
> + (unsigned long *)&vcpu_cp15(vcpu, c9_PMOVSSET));
> + __set_bit(idx,
> + (unsigned long *)&vcpu_cp15(vcpu, c9_PMOVSCLR));
> + pmu->irq_pending = true;
> + kvm_vcpu_kick(vcpu);
> + }
> + }
> +}
> +
> +/**
> * kvm_pmu_get_counter_value - get PMU counter value
> * @vcpu: The vcpu pointer
> * @select_idx: The counter index
> @@ -225,7 +276,8 @@ void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u32 data,
> /* The initial sample period (overflow count) of an event. */
> attr.sample_period = (-counter) & (((u64)1 << overflow_bit) - 1);
>
> - event = perf_event_create_kernel_counter(&attr, -1, current, NULL, pmc);
> + event = perf_event_create_kernel_counter(&attr, -1, current,
> + kvm_pmu_perf_overflow, pmc);
> if (IS_ERR(event)) {
> printk_once("kvm: pmu event creation failed %ld\n",
> PTR_ERR(event));
>
Thanks,
M.
--
Jazz is not dead. It just smells funny...