Re: [PATCH 3/3] hw/arm/virt: Implement kvm-steal-time

[Peter Maydell <peter.maydell@linaro.org>]

On Sat, 11 Jul 2020 at 11:10, Andrew Jones <drjones@redhat.com> wrote:
> We add the kvm-steal-time CPU property and implement it for machvirt.
> A tiny bit of refactoring was also done to allow pmu and pvtime to
> use the same vcpu device helper functions.
>
> Signed-off-by: Andrew Jones <drjones@redhat.com>

Hi; I'm forwarding a couple of comments here from Beata,
(whose secondment with Linaro is coming to an end so she won't
have access to her Linaro email address to continue the conversation):


>  static void virt_cpu_post_init(VirtMachineState *vms)
>  {
> -    bool aarch64, pmu;
> +    bool aarch64, pmu, steal_time;
>      CPUState *cpu;
>
>      aarch64 = object_property_get_bool(OBJECT(first_cpu), "aarch64", NULL);
>      pmu = object_property_get_bool(OBJECT(first_cpu), "pmu", NULL);
> +    steal_time = object_property_get_bool(OBJECT(first_cpu),
> +                                          "kvm-steal-time", NULL);
>
>      if (kvm_enabled()) {
> +        hwaddr pvtime_base = vms->memmap[VIRT_PVTIME].base;
> +        hwaddr pvtime_size = vms->memmap[VIRT_PVTIME].size;
> +
> +        if (steal_time) {
> +            MemoryRegion *pvtime = g_new(MemoryRegion, 1);
> +
> +            memory_region_init_ram(pvtime, NULL, "pvtime", pvtime_size, NULL);
> +            memory_region_add_subregion(get_system_memory(), pvtime_base,
> +                                        pvtime);
> +        }

B: I'm not sure whether it wouldn't be useful to have the area
allocated with size determined by number of VCPUs instead of having
pre-defined size.

> +        if (vmc->kvm_no_steal_time &&
> +            object_property_find(cpuobj, "kvm-steal-time", NULL)) {
> +            object_property_set_bool(cpuobj, false, "kvm-steal-time", NULL);
> +        }
> +
>          if (vmc->no_pmu && object_property_find(cpuobj, "pmu", NULL)) {
>              object_property_set_bool(cpuobj, "pmu", false, NULL);
>          }
> @@ -2528,6 +2558,7 @@ static void virt_machine_5_0_options(MachineClass *mc)
>      mc->numa_mem_supported = true;
>      vmc->acpi_expose_flash = true;
>      mc->auto_enable_numa_with_memdev = false;
> +    vmc->kvm_no_steal_time = true;
>  }
>  DEFINE_VIRT_MACHINE(5, 0)
>
> diff --git a/include/hw/arm/virt.h b/include/hw/arm/virt.h
> index 54bcf17afd35..b5153afedcdf 100644
> --- a/include/hw/arm/virt.h
> +++ b/include/hw/arm/virt.h
> @@ -80,6 +80,7 @@ enum {
>      VIRT_PCDIMM_ACPI,
>      VIRT_ACPI_GED,
>      VIRT_NVDIMM_ACPI,
> +    VIRT_PVTIME,
>      VIRT_LOWMEMMAP_LAST,
>  };
>
> @@ -126,6 +127,7 @@ typedef struct {
>      bool no_ged;   /* Machines < 4.2 has no support for ACPI GED device */
>      bool kvm_no_adjvtime;
>      bool acpi_expose_flash;
> +    bool kvm_no_steal_time;

B: It is slightly confusing : using kvm_no_steal_time vs kvm_steal_time

P: I have to admit I get confused about which sense this flag
should have. I think the sense of the flags in this struct is
"the false case is the one that the older virt boards had",
so original virt didn't have an ITS or a PMU and so we have
no_its and no_pmu. Similarly here old virt didn't have steal-time
and so we want a no_ flag (ie the patch is correct). Why
kvm_no_steal_time rather than no_kvm_steal_time, though ?

>  } VirtMachineClass;

> +void kvm_arm_pvtime_init(CPUState *cs, uint64_t ipa)
> +{
> +    struct kvm_device_attr attr = {
> +        .group = KVM_ARM_VCPU_PVTIME_CTRL,
> +        .attr = KVM_ARM_VCPU_PVTIME_IPA,
> +        .addr = (uint64_t)&ipa,
> +    };
> +
> +    if (!ARM_CPU(cs)->kvm_steal_time) {
> +        return;
> +    }
> +    if (!kvm_arm_set_device_attr(cs, &attr, "PVTIME IPA")) {
> +        error_report("failed to init PVTIME IPA");
> +        abort();
> +    }
> +}

B: I am probably missing smth but .. there is a trigger missing to
update the stats
and write them back to pre-allocated guest memory.
Looking at the kernel code the stats are updated upon pending
VCPU request :
in arch/arm64/kvm/arm.c:
static void check_vcpu_requests(struct kvm_vcpu *vcpu) {
        ...
         if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
                kvm_update_stolen_time(vcpu);
}


thanks
-- PMM