[PATCH v2] KVM: arm64: Allow to limit number of PMU counters
Robin Murphy
robin.murphy at arm.com
Thu Sep 10 13:41:44 EDT 2020
On 2020-09-10 17:46, Alexander Graf wrote:
>
>
> On 10.09.20 17:52, Robin Murphy wrote:
>>
>> On 2020-09-10 11:18, Alexander Graf wrote:
>>>
>>>
>>> On 10.09.20 12:06, Marc Zyngier wrote:
>>>>
>>>> On 2020-09-08 21:57, Alexander Graf wrote:
>>>>> We currently pass through the number of PMU counters that we have
>>>>> available
>>>>> in hardware to guests. So if my host supports 10 concurrently active
>>>>> PMU
>>>>> counters, my guest will be able to spawn 10 counters as well.
>>>>>
>>>>> This is undesireable if we also want to use the PMU on the host for
>>>>> monitoring. In that case, we want to split the PMU between guest and
>>>>> host.
>>>>>
>>>>> To help that case, let's add a PMU attr that allows us to limit the
>>>>> number
>>>>> of PMU counters that we expose. With this patch in place, user space
>>>>> can
>>>>> keep some counters free for host use.
>>>>>
>>>>> Signed-off-by: Alexander Graf <graf at amazon.com>
>>>>>
>>>>> ---
>>>>>
>>>>> Because this patch touches the same code paths as the vPMU filtering
>>>>> one
>>>>> and the vPMU filtering generalized a few conditions in the attr path,
>>>>> I've based it on top. Please let me know if you want it independent
>>>>> instead.
>>>>>
>>>>> v1 -> v2:
>>>>>
>>>>> - Add documentation
>>>>> - Add read support
>>>>> ---
>>>>> Documentation/virt/kvm/devices/vcpu.rst | 25
>>>>> +++++++++++++++++++++++++
>>>>> arch/arm64/include/uapi/asm/kvm.h | 7 ++++---
>>>>> arch/arm64/kvm/pmu-emul.c | 32
>>>>> ++++++++++++++++++++++++++++++++
>>>>> arch/arm64/kvm/sys_regs.c | 5 +++++
>>>>> include/kvm/arm_pmu.h | 1 +
>>>>> 5 files changed, 67 insertions(+), 3 deletions(-)
>>>>>
>>>>> diff --git a/Documentation/virt/kvm/devices/vcpu.rst
>>>>> b/Documentation/virt/kvm/devices/vcpu.rst
>>>>> index 203b91e93151..1a1c8d8c8b1d 100644
>>>>> --- a/Documentation/virt/kvm/devices/vcpu.rst
>>>>> +++ b/Documentation/virt/kvm/devices/vcpu.rst
>>>>> @@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering the
>>>>> cycle counter is possible
>>>>> using event 0x11 (CPU_CYCLES).
>>>>>
>>>>>
>>>>> +1.4 ATTRIBUTE: KVM_ARM_VCPU_PMU_V3_NUM_EVENTS
>>>>> +---------------------------------------------
>>>>> +
>>>>> +:Parameters: in kvm_device_attr.addr the address for the limit of
>>>>> concurrent
>>>>> + events is a pointer to an int
>>>>> +
>>>>> +:Returns:
>>>>> +
>>>>> + ======= ======================================================
>>>>> + -ENODEV: PMUv3 not supported
>>>>> + -EBUSY: PMUv3 already initialized
>>>>> + -EINVAL: Too large number of events
>>>>> + ======= ======================================================
>>>>> +
>>>>> +Reconfigure the limit of concurrent PMU events that the guest can
>>>>> monitor.
>>>>> +This number is directly exposed as part of the PMCR_EL0 register.
>>>>> +
>>>>> +On vcpu creation, this attribute is set to the hardware limit of the
>>>>> current
>>>>> +platform. If you need to determine the hardware limit, you can read
>>>>> this
>>>>> +attribute before setting it.
>>>>> +
>>>>> +Restrictions: The default value for this property is the number of
>>>>> hardware
>>>>> +supported events. Only values that are smaller than the hardware
>>>>> limit
>>>>> can
>>>>> +be set.
>>>>> +
>>>>> 2. GROUP: KVM_ARM_VCPU_TIMER_CTRL
>>>>> =================================
>>>>>
>>>>> diff --git a/arch/arm64/include/uapi/asm/kvm.h
>>>>> b/arch/arm64/include/uapi/asm/kvm.h
>>>>> index 7b1511d6ce44..db025c0b5a40 100644
>>>>> --- a/arch/arm64/include/uapi/asm/kvm.h
>>>>> +++ b/arch/arm64/include/uapi/asm/kvm.h
>>>>> @@ -342,9 +342,10 @@ struct kvm_vcpu_events {
>>>>>
>>>>> /* Device Control API on vcpu fd */
>>>>> #define KVM_ARM_VCPU_PMU_V3_CTRL 0
>>>>> -#define KVM_ARM_VCPU_PMU_V3_IRQ 0
>>>>> -#define KVM_ARM_VCPU_PMU_V3_INIT 1
>>>>> -#define KVM_ARM_VCPU_PMU_V3_FILTER 2
>>>>> +#define KVM_ARM_VCPU_PMU_V3_IRQ 0
>>>>> +#define KVM_ARM_VCPU_PMU_V3_INIT 1
>>>>> +#define KVM_ARM_VCPU_PMU_V3_FILTER 2
>>>>> +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3
>>>>> #define KVM_ARM_VCPU_TIMER_CTRL 1
>>>>> #define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0
>>>>> #define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1
>>>>> diff --git a/arch/arm64/kvm/pmu-emul.c b/arch/arm64/kvm/pmu-emul.c
>>>>> index 0458860bade2..c7915b95fec0 100644
>>>>> --- a/arch/arm64/kvm/pmu-emul.c
>>>>> +++ b/arch/arm64/kvm/pmu-emul.c
>>>>> @@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)
>>>>>
>>>>> for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++)
>>>>> pmu->pmc[i].idx = i;
>>>>> +
>>>>> + pmu->num_events = perf_num_counters() - 1;
>>>>> }
>>>>>
>>>>> /**
>>>>> @@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu
>>>>> *vcpu, struct kvm_device_attr *attr)
>>>>>
>>>>> return 0;
>>>>> }
>>>>> + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: {
>>>>> + u64 mask = ARMV8_PMU_PMCR_N_MASK <<
>>>>> ARMV8_PMU_PMCR_N_SHIFT;
>>>>> + int __user *uaddr = (int __user *)(long)attr->addr;
>>>>> + u32 num_events;
>>>>> +
>>>>> + if (get_user(num_events, uaddr))
>>>>> + return -EFAULT;
>>>>> +
>>>>> + if (num_events >= perf_num_counters())
>>>>> + return -EINVAL;
>>>>> +
>>>>> + vcpu->arch.pmu.num_events = num_events;
>>>>> +
>>>>> + num_events <<= ARMV8_PMU_PMCR_N_SHIFT;
>>>>> + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) &= ~mask;
>>>>> + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) |= num_events;
>>>>> +
>>>>> + return 0;
>>>>> + }
>>>>> case KVM_ARM_VCPU_PMU_V3_INIT:
>>>>> return kvm_arm_pmu_v3_init(vcpu);
>>>>> }
>>>>> @@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu
>>>>> *vcpu, struct kvm_device_attr *attr)
>>>>> irq = vcpu->arch.pmu.irq_num;
>>>>> return put_user(irq, uaddr);
>>>>> }
>>>>> + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: {
>>>>> + int __user *uaddr = (int __user *)(long)attr->addr;
>>>>> + u32 num_events;
>>>>> +
>>>>> + if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
>>>>> + return -ENODEV;
>>>>> +
>>>>> + num_events = vcpu->arch.pmu.num_events;
>>>>> + return put_user(num_events, uaddr);
>>>>> + }
>>>>> }
>>>>>
>>>>> return -ENXIO;
>>>>> @@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu
>>>>> *vcpu, struct kvm_device_attr *attr)
>>>>> case KVM_ARM_VCPU_PMU_V3_IRQ:
>>>>> case KVM_ARM_VCPU_PMU_V3_INIT:
>>>>> case KVM_ARM_VCPU_PMU_V3_FILTER:
>>>>> + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS:
>>>>> if (kvm_arm_support_pmu_v3() &&
>>>>> test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
>>>>> return 0;
>>>>> diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
>>>>> index 20ab2a7d37ca..d51e39600bbd 100644
>>>>> --- a/arch/arm64/kvm/sys_regs.c
>>>>> +++ b/arch/arm64/kvm/sys_regs.c
>>>>> @@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu,
>>>>> const struct sys_reg_desc *r)
>>>>> | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) &
>>>>> (~ARMV8_PMU_PMCR_E);
>>>>> if (!system_supports_32bit_el0())
>>>>> val |= ARMV8_PMU_PMCR_LC;
>>>>> +
>>>>> + /* Override number of event selectors */
>>>>> + val &= ~(ARMV8_PMU_PMCR_N_MASK << ARMV8_PMU_PMCR_N_SHIFT);
>>>>> + val |= (u32)vcpu->arch.pmu.num_events << ARMV8_PMU_PMCR_N_SHIFT;
>>>>> +
>>>>> __vcpu_sys_reg(vcpu, r->reg) = val;
>>>>> }
>>>>>
>>>>> diff --git a/include/kvm/arm_pmu.h b/include/kvm/arm_pmu.h
>>>>> index 98cbfe885a53..ea3fc96a37d9 100644
>>>>> --- a/include/kvm/arm_pmu.h
>>>>> +++ b/include/kvm/arm_pmu.h
>>>>> @@ -27,6 +27,7 @@ struct kvm_pmu {
>>>>> bool ready;
>>>>> bool created;
>>>>> bool irq_level;
>>>>> + u8 num_events;
>>>>> };
>>>>>
>>>>> #define kvm_arm_pmu_v3_ready(v) ((v)->arch.pmu.ready)
>>>>
>>>> I see several problems with this approach:
>>>>
>>>> - userspace doesn't really have a good way to retrieve the number of
>>>> counters.
>>> It does with v2, because it can then just read the register ;). I agree
>>> that it's clunky though.
>>>
>>>>
>>>> - Limiting the number of counters for the guest doesn't mean anything
>>>> when it comes to the actual use of the HW counters, given that we
>>>> don't allocate them ourselves (it's all perf doing the actual work).
>>>
>>> We do cap the number of actively requestable counters via perf by the
>>> PMCR.N limit. So in a way, it does mean something.
>>>
>>>> - If you want to "pin" counters for the host, why don't you just do
>>>> that before starting the guest?
>>>
>>> You can do that. Imagine I have 10 counters. I pin 4 of them to the
>>> host. I still tell my guest that it can use 6. That means perf will then
>>> time slice and juggle 10 guest event counters on those remaining 6
>>> hardware counters. That juggling heavily reduces accuracy.
>>>
>>>> I think you need to look at the bigger picture: how to limit the use
>>>> of physical counter usage for a given userspace task. This needs
>>>> to happen in perf itself, and not in KVM.
>>>
>>> That's definitely another way to look at it that I agree with.
>>>
>>> What we really want is to expose the number of counters the guest has
>>> available, not the number of counters hardware can support at maximum.
>>>
>>> So in theory it would be enough to ask perf how many counters it does
>>> have free for me to consume without overcommitting. But that would
>>> potentially change between multiple invocations of KVM and thus break
>>> things like live migration, no?
>>>
>>> Maybe what we really want is an interface to perf from user space to say
>>> "how many counters can you dedicate to me?" and "reserve them for me".
>>> Then user space could reserve them as dedicated counters and KVM would
>>> just need to either probe for the reservation or get told by user space
>>> what to expose via ONE_REG as Drew suggested. It'd be up to user space
>>> to ensure that the reservation matches the number of exposed counters
>>> then.
>>
>> Note that if the aim is to avoid the guest seeing unexpectedly weird
>> behaviour, then it's not just the *number* of counters that matters, but
>> the underlying physical allocation too, thanks to the possibility of
>> chained events.
>
> Wouldn't ideally guest chaining propagate into host chaining as well?
> I'd have to double check if it does, but in my naive thinking if I
> reserve 4 hardware counters for the guest and the guest ends up using 4
> hardware counters regardless of their chaining attributes, I'd still be
> able to fit them all?
It depends what you mean by "reserve" - if you merely tell the guest
that the vPMU only has 4 counters, and at any given time your (n-4) host
events happen to have ended up scheduled such that no physical even/odd
pair of counters is free, then even a single chained event from the
guest might trigger context rotation despite there being 4 or more
counters free in total.
If you want truly deterministic results then you'll probably need some
brain surgery on the PMU driver to actually partition the physical
counters rather than simply relying on perf's scheduling.
Robin.
More information about the linux-arm-kernel
mailing list