* [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters
@ 2026-10-23 9:20 Vincent Donnefort
2026-10-23 12:29 ` Fuad Tabba
2026-10-23 12:37 ` Marc Zyngier
0 siblings, 2 replies; 6+ messages in thread
From: Vincent Donnefort @ 2026-10-23 9:20 UTC (permalink / raw)
To: maz, oupton, kvmarm, linux-arm-kernel
Cc: joey.gouly, seiden, suzuki.poulose, yuzenghui, catalin.marinas,
will, kernel-team, fuad.tabba, Vincent Donnefort
nr_event_counters is only used by the host to populate mdcr_el2.
However, host_data_ptr() for pKVM switches between the hypervisor copy
of kvm_host_data before it is deprivileged to the host copy after.
So kvm_init_host_debug_data() initialises nr_event_counters in the
hypervisor's kvm_host_data, while kvm_arm_setup_mdcr_el2() reads the
host copy which is 0. With a HPMN of 0, this stops programmable event
counters from counting a non-protected guest.
Always use the host copy for nr_event_counters.
Fixes: 2417218f2f23 ("KVM: arm64: Get rid of __kvm_get_mdcr_el2() and related warts")
Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
---
arch/arm64/kvm/debug.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/arch/arm64/kvm/debug.c b/arch/arm64/kvm/debug.c
index f4d7b12045e8..a00771339c62 100644
--- a/arch/arm64/kvm/debug.c
+++ b/arch/arm64/kvm/debug.c
@@ -44,7 +44,7 @@ static void kvm_arm_setup_mdcr_el2(struct kvm_vcpu *vcpu)
* to disable guest access to the profiling and trace buffers
*/
vcpu->arch.mdcr_el2 = FIELD_PREP(MDCR_EL2_HPMN,
- *host_data_ptr(nr_event_counters));
+ __this_cpu_read(kvm_host_data.nr_event_counters));
vcpu->arch.mdcr_el2 |= (MDCR_EL2_TPM |
MDCR_EL2_TPMS |
MDCR_EL2_TTRF |
@@ -80,8 +80,8 @@ void kvm_init_host_debug_data(void)
ID_AA64DFR0_EL1_PMUVer_SHIFT);
if (pmuv3_implemented(pmuver))
- *host_data_ptr(nr_event_counters) = FIELD_GET(ARMV8_PMU_PMCR_N,
- read_sysreg(pmcr_el0));
+ __this_cpu_write(kvm_host_data.nr_event_counters,
+ FIELD_GET(ARMV8_PMU_PMCR_N, read_sysreg(pmcr_el0)));
*host_data_ptr(debug_brps) = SYS_FIELD_GET(ID_AA64DFR0_EL1, BRPs, dfr0);
*host_data_ptr(debug_wrps) = SYS_FIELD_GET(ID_AA64DFR0_EL1, WRPs, dfr0);
base-commit: c03c1a14507dfaf8b856c2608f9ac216021fc222
--
2.56.0.257.gf7497c7d27-goog
^ permalink raw reply related [flat|nested] 6+ messages in thread
* Re: [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters
2026-10-23 9:20 [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters Vincent Donnefort
@ 2026-10-23 12:29 ` Fuad Tabba
2026-10-23 12:43 ` Marc Zyngier
2026-10-23 12:37 ` Marc Zyngier
1 sibling, 1 reply; 6+ messages in thread
From: Fuad Tabba @ 2026-10-23 12:29 UTC (permalink / raw)
To: Vincent Donnefort
Cc: maz, oupton, kvmarm, linux-arm-kernel, joey.gouly, seiden,
suzuki.poulose, yuzenghui, catalin.marinas, will, kernel-team
Hi Vincent,
On Fri, 23 Oct 2026 10:20:07 +0100, Vincent Donnefort
<vdonnefort@google.com> wrote:
[...]
> So kvm_init_host_debug_data() initialises nr_event_counters in the
> hypervisor's kvm_host_data, while kvm_arm_setup_mdcr_el2() reads the
> host copy which is 0. With a HPMN of 0, this stops programmable event
> counters from counting a non-protected guest.
I tested this in pKVM with a non-protected guest counting with
perf_event_open(). On its own the patch doesn't change what the guest
sees: every counter, the cycle counter included, still reads 0. That's
a separate bug: flush_hyp_vcpu() never copies vcpu->arch.pmu.events to
the hyp vCPU, so the guest's events are never enabled on entry. I'll
send a fix for that shortly.
With that fix applied, the cycle counter counts without your patch but
a programmable counter still reads 0, and with your patch both count.
Reviewed-by: Fuad Tabba <fuad.tabba@linux.dev>
Tested-by: Fuad Tabba <fuad.tabba@linux.dev>
Cheers,
/fuad
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters
2026-10-23 9:20 [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters Vincent Donnefort
2026-10-23 12:29 ` Fuad Tabba
@ 2026-10-23 12:37 ` Marc Zyngier
2026-10-26 13:08 ` Vincent Donnefort
1 sibling, 1 reply; 6+ messages in thread
From: Marc Zyngier @ 2026-10-23 12:37 UTC (permalink / raw)
To: Vincent Donnefort
Cc: oupton, kvmarm, linux-arm-kernel, joey.gouly, seiden,
suzuki.poulose, yuzenghui, catalin.marinas, will, kernel-team,
fuad.tabba
On Fri, 23 Oct 2026 10:20:07 +0100,
Vincent Donnefort <vdonnefort@google.com> wrote:
>
> nr_event_counters is only used by the host to populate mdcr_el2.
> However, host_data_ptr() for pKVM switches between the hypervisor copy
> of kvm_host_data before it is deprivileged to the host copy after.
>
> So kvm_init_host_debug_data() initialises nr_event_counters in the
> hypervisor's kvm_host_data, while kvm_arm_setup_mdcr_el2() reads the
> host copy which is 0. With a HPMN of 0, this stops programmable event
> counters from counting a non-protected guest.
>
> Always use the host copy for nr_event_counters.
>
> Fixes: 2417218f2f23 ("KVM: arm64: Get rid of __kvm_get_mdcr_el2() and related warts")
> Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
> ---
> arch/arm64/kvm/debug.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/arch/arm64/kvm/debug.c b/arch/arm64/kvm/debug.c
> index f4d7b12045e8..a00771339c62 100644
> --- a/arch/arm64/kvm/debug.c
> +++ b/arch/arm64/kvm/debug.c
> @@ -44,7 +44,7 @@ static void kvm_arm_setup_mdcr_el2(struct kvm_vcpu *vcpu)
> * to disable guest access to the profiling and trace buffers
> */
> vcpu->arch.mdcr_el2 = FIELD_PREP(MDCR_EL2_HPMN,
> - *host_data_ptr(nr_event_counters));
> + __this_cpu_read(kvm_host_data.nr_event_counters));
> vcpu->arch.mdcr_el2 |= (MDCR_EL2_TPM |
> MDCR_EL2_TPMS |
> MDCR_EL2_TTRF |
> @@ -80,8 +80,8 @@ void kvm_init_host_debug_data(void)
> ID_AA64DFR0_EL1_PMUVer_SHIFT);
>
> if (pmuv3_implemented(pmuver))
> - *host_data_ptr(nr_event_counters) = FIELD_GET(ARMV8_PMU_PMCR_N,
> - read_sysreg(pmcr_el0));
> + __this_cpu_write(kvm_host_data.nr_event_counters,
> + FIELD_GET(ARMV8_PMU_PMCR_N, read_sysreg(pmcr_el0)));
>
> *host_data_ptr(debug_brps) = SYS_FIELD_GET(ID_AA64DFR0_EL1, BRPs, dfr0);
> *host_data_ptr(debug_wrps) = SYS_FIELD_GET(ID_AA64DFR0_EL1, WRPs, dfr0);
>
This really looks like the wrong approach.
host_data_ptr() should be reliable, and always point at the correct
data, irrespective of the backend. If host_data_ptr() returns a
pointer to stale data on pKVM, then it is purely a pKVM bug, and
that's what needs fixing by synchronising the two copies by whatever
means necessary.
Randomly picking a different accessor is not an acceptable outcome.
M.
--
Without deviation from the norm, progress is not possible.
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters
2026-10-23 12:29 ` Fuad Tabba
@ 2026-10-23 12:43 ` Marc Zyngier
2026-10-23 12:45 ` Fuad Tabba
0 siblings, 1 reply; 6+ messages in thread
From: Marc Zyngier @ 2026-10-23 12:43 UTC (permalink / raw)
To: Fuad Tabba
Cc: Vincent Donnefort, oupton, kvmarm, linux-arm-kernel, joey.gouly,
seiden, suzuki.poulose, yuzenghui, catalin.marinas, will,
kernel-team
On Fri, 23 Oct 2026 13:29:04 +0100,
Fuad Tabba <fuad.tabba@linux.dev> wrote:
>
> Hi Vincent,
>
> On Fri, 23 Oct 2026 10:20:07 +0100, Vincent Donnefort
> <vdonnefort@google.com> wrote:
> [...]
> > So kvm_init_host_debug_data() initialises nr_event_counters in the
> > hypervisor's kvm_host_data, while kvm_arm_setup_mdcr_el2() reads the
> > host copy which is 0. With a HPMN of 0, this stops programmable event
> > counters from counting a non-protected guest.
>
> I tested this in pKVM with a non-protected guest counting with
> perf_event_open(). On its own the patch doesn't change what the guest
> sees: every counter, the cycle counter included, still reads 0. That's
> a separate bug: flush_hyp_vcpu() never copies vcpu->arch.pmu.events to
> the hyp vCPU, so the guest's events are never enabled on entry. I'll
> send a fix for that shortly.
Please synchronise with Vincent and sent this as a series fixing PMU
handling on pKVM, not as discrete patches.
The approach taken for this patch is wrong anyway, so you have plenty
of opportunity to work together.
Thanks,
M.
--
Without deviation from the norm, progress is not possible.
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters
2026-10-23 12:43 ` Marc Zyngier
@ 2026-10-23 12:45 ` Fuad Tabba
0 siblings, 0 replies; 6+ messages in thread
From: Fuad Tabba @ 2026-10-23 12:45 UTC (permalink / raw)
To: Marc Zyngier
Cc: Vincent Donnefort, oupton, kvmarm, linux-arm-kernel, joey.gouly,
seiden, suzuki.poulose, yuzenghui, catalin.marinas, will,
kernel-team
On Fri, 23 Oct 2026 at 13:43, Marc Zyngier <maz@kernel.org> wrote:
[...]
> > I tested this in pKVM with a non-protected guest counting with
> > perf_event_open(). On its own the patch doesn't change what the guest
> > sees: every counter, the cycle counter included, still reads 0. That's
> > a separate bug: flush_hyp_vcpu() never copies vcpu->arch.pmu.events to
> > the hyp vCPU, so the guest's events are never enabled on entry. I'll
> > send a fix for that shortly.
>
> Please synchronise with Vincent and sent this as a series fixing PMU
> handling on pKVM, not as discrete patches.
>
> The approach taken for this patch is wrong anyway, so you have plenty
> of opportunity to work together.
Will do.
Cheers,
/fuad
>
> Thanks,
>
> M.
>
> --
> Without deviation from the norm, progress is not possible.
>
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters
2026-10-23 12:37 ` Marc Zyngier
@ 2026-10-26 13:08 ` Vincent Donnefort
0 siblings, 0 replies; 6+ messages in thread
From: Vincent Donnefort @ 2026-10-26 13:08 UTC (permalink / raw)
To: Marc Zyngier
Cc: oupton, kvmarm, linux-arm-kernel, joey.gouly, seiden,
suzuki.poulose, yuzenghui, catalin.marinas, will, kernel-team,
fuad.tabba
On Fri, Oct 23, 2026 at 01:37:45PM +0100, Marc Zyngier wrote:
> On Fri, 23 Oct 2026 10:20:07 +0100,
> Vincent Donnefort <vdonnefort@google.com> wrote:
> >
> > nr_event_counters is only used by the host to populate mdcr_el2.
> > However, host_data_ptr() for pKVM switches between the hypervisor copy
> > of kvm_host_data before it is deprivileged to the host copy after.
> >
> > So kvm_init_host_debug_data() initialises nr_event_counters in the
> > hypervisor's kvm_host_data, while kvm_arm_setup_mdcr_el2() reads the
> > host copy which is 0. With a HPMN of 0, this stops programmable event
> > counters from counting a non-protected guest.
> >
> > Always use the host copy for nr_event_counters.
> >
> > Fixes: 2417218f2f23 ("KVM: arm64: Get rid of __kvm_get_mdcr_el2() and related warts")
> > Signed-off-by: Vincent Donnefort <vdonnefort@google.com>
> > ---
> > arch/arm64/kvm/debug.c | 6 +++---
> > 1 file changed, 3 insertions(+), 3 deletions(-)
> >
> > diff --git a/arch/arm64/kvm/debug.c b/arch/arm64/kvm/debug.c
> > index f4d7b12045e8..a00771339c62 100644
> > --- a/arch/arm64/kvm/debug.c
> > +++ b/arch/arm64/kvm/debug.c
> > @@ -44,7 +44,7 @@ static void kvm_arm_setup_mdcr_el2(struct kvm_vcpu *vcpu)
> > * to disable guest access to the profiling and trace buffers
> > */
> > vcpu->arch.mdcr_el2 = FIELD_PREP(MDCR_EL2_HPMN,
> > - *host_data_ptr(nr_event_counters));
> > + __this_cpu_read(kvm_host_data.nr_event_counters));
> > vcpu->arch.mdcr_el2 |= (MDCR_EL2_TPM |
> > MDCR_EL2_TPMS |
> > MDCR_EL2_TTRF |
> > @@ -80,8 +80,8 @@ void kvm_init_host_debug_data(void)
> > ID_AA64DFR0_EL1_PMUVer_SHIFT);
> >
> > if (pmuv3_implemented(pmuver))
> > - *host_data_ptr(nr_event_counters) = FIELD_GET(ARMV8_PMU_PMCR_N,
> > - read_sysreg(pmcr_el0));
> > + __this_cpu_write(kvm_host_data.nr_event_counters,
> > + FIELD_GET(ARMV8_PMU_PMCR_N, read_sysreg(pmcr_el0)));
> >
> > *host_data_ptr(debug_brps) = SYS_FIELD_GET(ID_AA64DFR0_EL1, BRPs, dfr0);
> > *host_data_ptr(debug_wrps) = SYS_FIELD_GET(ID_AA64DFR0_EL1, WRPs, dfr0);
> >
>
> This really looks like the wrong approach.
>
> host_data_ptr() should be reliable, and always point at the correct
> data, irrespective of the backend. If host_data_ptr() returns a
> pointer to stale data on pKVM, then it is purely a pKVM bug, and
> that's what needs fixing by synchronising the two copies by whatever
> means necessary.
I'll update both copies and make a common series Fuad.
Thanks
--
Vincent
>
> Randomly picking a different accessor is not an acceptable outcome.
>
> M.
>
> --
> Without deviation from the norm, progress is not possible.
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2026-10-26 13:09 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-10-23 9:20 [PATCH v1] KVM: arm64: Use the host kvm_host_data for nr_event_counters Vincent Donnefort
2026-10-23 12:29 ` Fuad Tabba
2026-10-23 12:43 ` Marc Zyngier
2026-10-23 12:45 ` Fuad Tabba
2026-10-23 12:37 ` Marc Zyngier
2026-10-26 13:08 ` Vincent Donnefort
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.