[PATCH v2 2/2] KVM: arm64: PMU: Don't overwrite PMUSERENR with vcpu loaded
Reiji Watanabe
reijiw at google.com
Tue Apr 11 22:14:10 PDT 2023
On Tue, Apr 11, 2023 at 10:33:50AM +0100, Mark Rutland wrote:
> On Fri, Apr 07, 2023 at 08:47:59PM -0700, Reiji Watanabe wrote:
> > Currently, with VHE, KVM sets ER, CR, SW and EN bits of
> > PMUSERENR_EL0 to 1 on vcpu_load(), and saves and restores
> > the register value for the host on vcpu_load() and vcpu_put().
> > If the value of those bits are cleared on a pCPU with a vCPU
> > loaded (armv8pmu_start() would do that when PMU counters are
> > programmed for the guest), PMU access from the guest EL0 might
> > be trapped to the guest EL1 directly regardless of the current
> > PMUSERENR_EL0 value of the vCPU.
> >
> > Fix this by not letting armv8pmu_start() overwrite PMUSERENR on
> > the pCPU on which a vCPU is loaded, and instead updating the
> > saved shadow register value for the host, so that the value can
> > be restored on vcpu_put() later.
>
> I'm happy with the hook in the PMU code, but I think there's still a race
> between an IPI and vcpu_{load,put}() where we can lose an update to
> PMUSERERNR_EL0. I tried to point that out in my final question in:
>
> https://lore.kernel.org/all/ZCwzV7ACl21VbLru@FVFF77S0Q05N.cambridge.arm.com/
>
> ... but I looks like that wasn't all that clear.
>
> Consider vcpu_load():
>
> void vcpu_load(struct kvm_vcpu *vcpu)
> {
> int cpu = get_cpu();
>
> __this_cpu_write(kvm_running_vcpu, vcpu);
> preempt_notifier_register(&vcpu->preempt_notifier);
> kvm_arch_vcpu_load(vcpu, cpu);
> put_cpu();
> }
>
> AFAICT that's called with IRQs enabled, and the {get,put}_cpu() calls will only
> disable migration/preemption. After the write to kvm_running_vcpu, the code in
> kvm_set_pmuserenr() will see that there is a running vcpu, and write to the
> host context without updating the real PMUSERENR_EL0 register.
>
> If we take an IPI and call kvm_set_pmuserenr() after the write to
> kvm_running_vcpu but before kvm_running_vcpu() completes, the call to
> kvm_set_pmuserenr() could update the host context (without updating the real
> PMUSERENR_EL0 value) before __activate_traps_common() saves the host value
> with:
>
> ctxt_sys_reg(hctxt, PMUSERENR_EL0) = read_sysreg(pmuserenr_el0);
>
> ... which would discard the write made by kvm_set_pmuserenr().
>
> Similar can happen in vcpu_put() where an IPI after __deactivate_traps_common()
> but before kvm_running_vcpu is cleared would result in kvm_set_pmuserenr()
> writing to the host context, but this value would never be written into HW.
>
> Unless I'm missing something (e.g. if interrupts are actually masked during
> those windows), I don't think this is a complete fix as-is.
>
> I'm not sure if there is a smart fix for that.
Thank you for the comment.
Uh, right, interrupts are not masked during those windows...
What I am currently considering on this would be disabling
IRQs while manipulating the register, and introducing a new flag
to indicate whether the PMUSERENR for the guest EL0 is loaded,
and having kvm_set_pmuserenr() check the new flag.
The code would be something like below (local_irq_save/local_irq_restore
needs to be excluded for NVHE though).
What do you think ?
--- a/arch/arm64/include/asm/kvm_host.h
+++ b/arch/arm64/include/asm/kvm_host.h
@@ -668,6 +668,8 @@ struct kvm_vcpu_arch {
/* Software step state is Active-pending */
#define DBG_SS_ACTIVE_PENDING __vcpu_single_flag(sflags, BIT(5))
+/* PMUSERENR for the guest EL0 is on physical CPU */
+#define PMUSERENR_ON_CPU __vcpu_single_flag(sflags, BIT(6))
/* Pointer to the vcpu's SVE FFR for sve_{save,load}_state() */
#define vcpu_sve_pffr(vcpu) (kern_hyp_va((vcpu)->arch.sve_state) + \
diff --git a/arch/arm64/kvm/hyp/include/hyp/switch.h b/arch/arm64/kvm/hyp/include/hyp/switch.h
index 6718731729fd..57e4f480874a 100644
--- a/arch/arm64/kvm/hyp/include/hyp/switch.h
+++ b/arch/arm64/kvm/hyp/include/hyp/switch.h
@@ -82,12 +82,19 @@ static inline void __activate_traps_common(struct kvm_vcpu *vcpu)
*/
if (kvm_arm_support_pmu_v3()) {
struct kvm_cpu_context *hctxt;
+ unsigned long flags;
write_sysreg(0, pmselr_el0);
hctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
+
+ local_irq_save(flags);
+
ctxt_sys_reg(hctxt, PMUSERENR_EL0) = read_sysreg(pmuserenr_el0);
write_sysreg(ARMV8_PMU_USERENR_MASK, pmuserenr_el0);
+ vcpu_set_flag(vcpu, PMUSERENR_ON_CPU);
+
+ local_irq_restore(flags);
}
vcpu->arch.mdcr_el2_host = read_sysreg(mdcr_el2);
@@ -112,9 +119,16 @@ static inline void __deactivate_traps_common(struct kvm_vcpu *vcpu)
write_sysreg(0, hstr_el2);
if (kvm_arm_support_pmu_v3()) {
struct kvm_cpu_context *hctxt;
+ unsigned long flags;
hctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
+
+ local_irq_save(flags);
+
write_sysreg(ctxt_sys_reg(hctxt, PMUSERENR_EL0), pmuserenr_el0);
+ vcpu_clear_flag(vcpu, PMUSERENR_ON_CPU);
+
+ local_irq_restore(flags);
}
if (cpus_have_final_cap(ARM64_SME)) {
diff --git a/arch/arm64/kvm/pmu.c b/arch/arm64/kvm/pmu.c
index 40bb2cb13317..33cd8e1ecbd6 100644
--- a/arch/arm64/kvm/pmu.c
+++ b/arch/arm64/kvm/pmu.c
@@ -221,8 +221,13 @@ void kvm_vcpu_pmu_restore_host(struct kvm_vcpu *vcpu)
bool kvm_set_pmuserenr(u64 val)
{
struct kvm_cpu_context *hctxt;
+ struct kvm_vcpu *vcpu;
- if (!kvm_arm_support_pmu_v3() || !has_vhe() || !kvm_get_running_vcpu())
+ if (!kvm_arm_support_pmu_v3() || !has_vhe())
+ return false;
+
+ vcpu = kvm_get_running_vcpu();
+ if (!vcpu || !vcpu_get_flag(vcpu, PMUSERENR_ON_CPU))
return false;
hctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
--
Thank you,
Reiji
>
> Thanks,
> Mark.
>
> > Suggested-by: Mark Rutland <mark.rutland at arm.com>
> > Suggested-by: Marc Zyngier <maz at kernel.org>
> > Fixes: 83a7a4d643d3 ("arm64: perf: Enable PMU counter userspace access for perf event")
> > Signed-off-by: Reiji Watanabe <reijiw at google.com>
> > ---
> > arch/arm64/include/asm/kvm_host.h | 5 +++++
> > arch/arm64/kernel/perf_event.c | 21 ++++++++++++++++++---
> > arch/arm64/kvm/pmu.c | 20 ++++++++++++++++++++
> > 3 files changed, 43 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> > index bcd774d74f34..22db2f885c17 100644
> > --- a/arch/arm64/include/asm/kvm_host.h
> > +++ b/arch/arm64/include/asm/kvm_host.h
> > @@ -1028,9 +1028,14 @@ void kvm_arch_vcpu_put_debug_state_flags(struct kvm_vcpu *vcpu);
> > #ifdef CONFIG_KVM
> > void kvm_set_pmu_events(u32 set, struct perf_event_attr *attr);
> > void kvm_clr_pmu_events(u32 clr);
> > +bool kvm_set_pmuserenr(u64 val);
> > #else
> > static inline void kvm_set_pmu_events(u32 set, struct perf_event_attr *attr) {}
> > static inline void kvm_clr_pmu_events(u32 clr) {}
> > +static inline bool kvm_set_pmuserenr(u64 val)
> > +{
> > + return false;
> > +}
> > #endif
> >
> > void kvm_vcpu_load_sysregs_vhe(struct kvm_vcpu *vcpu);
> > diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
> > index dde06c0f97f3..0fffe4c56c28 100644
> > --- a/arch/arm64/kernel/perf_event.c
> > +++ b/arch/arm64/kernel/perf_event.c
> > @@ -741,9 +741,25 @@ static inline u32 armv8pmu_getreset_flags(void)
> > return value;
> > }
> >
> > +static void update_pmuserenr(u64 val)
> > +{
> > + lockdep_assert_irqs_disabled();
> > +
> > + /*
> > + * The current pmuserenr value might be the value for the guest.
> > + * If that's the case, have KVM keep tracking of the register value
> > + * for the host EL0 so that KVM can restore it before returning to
> > + * the host EL0. Otherwise, update the register now.
> > + */
> > + if (kvm_set_pmuserenr(val))
> > + return;
> > +
> > + write_sysreg(val, pmuserenr_el0);
> > +}
> > +
> > static void armv8pmu_disable_user_access(void)
> > {
> > - write_sysreg(0, pmuserenr_el0);
> > + update_pmuserenr(0);
> > }
> >
> > static void armv8pmu_enable_user_access(struct arm_pmu *cpu_pmu)
> > @@ -759,8 +775,7 @@ static void armv8pmu_enable_user_access(struct arm_pmu *cpu_pmu)
> > armv8pmu_write_evcntr(i, 0);
> > }
> >
> > - write_sysreg(0, pmuserenr_el0);
> > - write_sysreg(ARMV8_PMU_USERENR_ER | ARMV8_PMU_USERENR_CR, pmuserenr_el0);
> > + update_pmuserenr(ARMV8_PMU_USERENR_ER | ARMV8_PMU_USERENR_CR);
> > }
> >
> > static void armv8pmu_enable_event(struct perf_event *event)
> > diff --git a/arch/arm64/kvm/pmu.c b/arch/arm64/kvm/pmu.c
> > index 7887133d15f0..40bb2cb13317 100644
> > --- a/arch/arm64/kvm/pmu.c
> > +++ b/arch/arm64/kvm/pmu.c
> > @@ -209,3 +209,23 @@ void kvm_vcpu_pmu_restore_host(struct kvm_vcpu *vcpu)
> > kvm_vcpu_pmu_enable_el0(events_host);
> > kvm_vcpu_pmu_disable_el0(events_guest);
> > }
> > +
> > +/*
> > + * With VHE, keep track of the PMUSERENR_EL0 value for the host EL0 on
> > + * the pCPU where vCPU is loaded, since PMUSERENR_EL0 is switched to
> > + * the value for the guest on vcpu_load(). The value for the host EL0
> > + * will be restored on vcpu_put(), before returning to the EL0.
> > + *
> > + * Return true if KVM takes care of the register. Otherwise return false.
> > + */
> > +bool kvm_set_pmuserenr(u64 val)
> > +{
> > + struct kvm_cpu_context *hctxt;
> > +
> > + if (!kvm_arm_support_pmu_v3() || !has_vhe() || !kvm_get_running_vcpu())
> > + return false;
> > +
> > + hctxt = &this_cpu_ptr(&kvm_host_data)->host_ctxt;
> > + ctxt_sys_reg(hctxt, PMUSERENR_EL0) = val;
> > + return true;
> > +}
> > --
> > 2.40.0.577.gac1e443424-goog
> >
> >
More information about the linux-arm-kernel
mailing list