Re: [PATCH] KVM: arm64: Properly restore PMU state during live-migration

[Marc Zyngier]

Hi Jinank,

On Thu, 03 Jun 2021 12:05:54 +0100,
Jinank Jain <jinankj@xxxxxxxxx> wrote:
> 
> Currently if a guest is live-migrated while it is actively using perf
> counters, then after live-migrate it will notice that all counters would
> suddenly start reporting 0s. This is due to the fact we are not
> re-creating the relevant perf events inside the kernel.
> 
> Usually on live-migration guest state is restored using KVM_SET_ONE_REG
> ioctl interface, which simply restores the value of PMU registers
> values but does not re-program the perf events so that the guest can seamlessly
> use these counters even after live-migration like it was doing before
> live-migration.
> 
> Instead there are two completely different code path between guest
> accessing PMU registers and VMM restoring counters on
> live-migration.
> 
> In case of KVM_SET_ONE_REG:
> 
> kvm_arm_set_reg()
> ...... kvm_arm_sys_reg_set_reg()
> ........... reg_from_user()
> 
> but in case when guest tries to access these counters:
> 
> handle_exit()
> ..... kvm_handle_sys_reg()
> ..........perform_access()
> ...............access_pmu_evcntr()
> ...................kvm_pmu_set_counter_value()
> .......................kvm_pmu_create_perf_event()
> 
> The drawback of using the KVM_SET_ONE_REG interface is that the host pmu
> events which were registered for the source instance and not present for
> the destination instance.

I can't parse this sentence. Do you mean "are not present"?

> Thus passively restoring PMCR_EL0 using
> KVM_SET_ONE_REG interface would not create the necessary host pmu events
> which are crucial for seamless guest experience across live migration.
> 
> In ordet to fix the situation, on first vcpu load we should restore
> PMCR_EL0 in the same exact way like the guest was trying to access
> these counters. And then we will also recreate the relevant host pmu
> events.
> 
> Signed-off-by: Jinank Jain <jinankj@xxxxxxxxx>
> Cc: Alexander Graf (AWS) <graf@xxxxxxxxx>
> Cc: Marc Zyngier <maz@xxxxxxxxxx>
> Cc: James Morse <james.morse@xxxxxxx>
> Cc: Alexandru Elisei <alexandru.elisei@xxxxxxx>
> Cc: Suzuki K Poulose <suzuki.poulose@xxxxxxx>
> Cc: Catalin Marinas <catalin.marinas@xxxxxxx>
> Cc: Will Deacon <will@xxxxxxxxxx>
> ---
>  arch/arm64/include/asm/kvm_host.h |  1 +
>  arch/arm64/kvm/arm.c              |  1 +
>  arch/arm64/kvm/pmu-emul.c         | 10 ++++++++--
>  arch/arm64/kvm/pmu.c              | 15 +++++++++++++++
>  include/kvm/arm_pmu.h             |  3 +++
>  5 files changed, 28 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> index 7cd7d5c8c4bc..2376ad3c2fc2 100644
> --- a/arch/arm64/include/asm/kvm_host.h
> +++ b/arch/arm64/include/asm/kvm_host.h
> @@ -745,6 +745,7 @@ static inline int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
>  void kvm_set_pmu_events(u32 set, struct perf_event_attr *attr);
>  void kvm_clr_pmu_events(u32 clr);
>  
> +void kvm_vcpu_pmu_restore(struct kvm_vcpu *vcpu);
>  void kvm_vcpu_pmu_restore_guest(struct kvm_vcpu *vcpu);
>  void kvm_vcpu_pmu_restore_host(struct kvm_vcpu *vcpu);
>  #else
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index e720148232a0..c66f6d16ec06 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -408,6 +408,7 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
>  	if (has_vhe())
>  		kvm_vcpu_load_sysregs_vhe(vcpu);
>  	kvm_arch_vcpu_load_fp(vcpu);
> +	kvm_vcpu_pmu_restore(vcpu);

If this only needs to be run once per vcpu, why not trigger it from
kvm_arm_pmu_v3_enable(), which is also called once per vcpu?

This can done on the back of a request, saving most of the overhead
and not requiring any extra field. Essentially, something like the
(untested) patch below.

>  	kvm_vcpu_pmu_restore_guest(vcpu);
>  	if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
>  		kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
> diff --git a/arch/arm64/kvm/pmu-emul.c b/arch/arm64/kvm/pmu-emul.c
> index fd167d4f4215..12a40f4b5f0d 100644
> --- a/arch/arm64/kvm/pmu-emul.c
> +++ b/arch/arm64/kvm/pmu-emul.c
> @@ -574,10 +574,16 @@ void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val)
>  		kvm_pmu_disable_counter_mask(vcpu, mask);
>  	}
>  
> -	if (val & ARMV8_PMU_PMCR_C)
> +	/*
> +	 * Cycle counter needs to reset in case of first vcpu load.
> +	 */
> +	if (val & ARMV8_PMU_PMCR_C || !kvm_arm_pmu_v3_restored(vcpu))

Why? There is no architectural guarantee that a counter resets to 0
without writing PMCR_EL0.C. And if you want the guest to continue
counting where it left off, resetting the counter is at best
counter-productive.

So I must be missing something...

>  		kvm_pmu_set_counter_value(vcpu, ARMV8_PMU_CYCLE_IDX, 0);
>  
> -	if (val & ARMV8_PMU_PMCR_P) {
> +	/*
> +	 * All the counters needs to reset in case of first vcpu load.
> +	 */
> +	if (val & ARMV8_PMU_PMCR_P || !kvm_arm_pmu_v3_restored(vcpu)) {

Same thing here.

>  		for_each_set_bit(i, &mask, 32)
>  			kvm_pmu_set_counter_value(vcpu, i, 0);
>  	}

The rest of the changes should be unnecessary with the patch below.

Thanks,

	M.