commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
is returning error for perf syscall with mixed attribute set for exclude_kernel
and exlude_hv.
This change is breaking some applications (observed with hhvm) when
ran with VHE enabled. Adding change to enable EL2 event counting,
if either of or both of exclude_kernel and exlude_hv are not set.
Signed-off-by: Ganapatrao Kulkarni <[email protected]>
---
arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
1 file changed, 12 insertions(+), 7 deletions(-)
diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
index daf95919..ea5848a 100644
--- a/arch/arm64/kernel/perf_event.c
+++ b/arch/arm64/kernel/perf_event.c
@@ -871,15 +871,20 @@ static int armv8pmu_set_event_filter(struct hw_perf_event *event,
if (attr->exclude_idle)
return -EPERM;
- if (is_kernel_in_hyp_mode() &&
- attr->exclude_kernel != attr->exclude_hv)
- return -EINVAL;
+ if (is_kernel_in_hyp_mode()) {
+ /* include EL2 events, if either of not excluded */
+ if ((attr->exclude_kernel != attr->exclude_hv) ||
+ !attr->exclude_kernel ||
+ !attr->exclude_hv)
+ config_base |= ARMV8_PMU_INCLUDE_EL2;
+ } else {
+ if (attr->exclude_kernel)
+ config_base |= ARMV8_PMU_EXCLUDE_EL1;
+ if (!attr->exclude_hv)
+ config_base |= ARMV8_PMU_INCLUDE_EL2;
+ }
if (attr->exclude_user)
config_base |= ARMV8_PMU_EXCLUDE_EL0;
- if (!is_kernel_in_hyp_mode() && attr->exclude_kernel)
- config_base |= ARMV8_PMU_EXCLUDE_EL1;
- if (!attr->exclude_hv)
- config_base |= ARMV8_PMU_INCLUDE_EL2;
/*
* Install the filter into config_base as this is used to
--
1.8.1.4
On Tue, Apr 04, 2017 at 04:10:55PM +0530, Ganapatrao Kulkarni wrote:
> commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
> is returning error for perf syscall with mixed attribute set for exclude_kernel
> and exlude_hv.
>
> This change is breaking some applications (observed with hhvm) when
> ran with VHE enabled. Adding change to enable EL2 event counting,
> if either of or both of exclude_kernel and exlude_hv are not set.
>
> Signed-off-by: Ganapatrao Kulkarni <[email protected]>
> ---
> arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
> 1 file changed, 12 insertions(+), 7 deletions(-)
Hmm. When we have VHE, we can't distinguish between hypervisor and kernel,
so this patch doesn't seem right to me. The code currently requires
both exclude_kernel and exclude_hv to be clear before we enable profiling
EL2, otherwise we're failing to exclude samples that were asked to be
excluded.
Will
> diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
> index daf95919..ea5848a 100644
> --- a/arch/arm64/kernel/perf_event.c
> +++ b/arch/arm64/kernel/perf_event.c
> @@ -871,15 +871,20 @@ static int armv8pmu_set_event_filter(struct hw_perf_event *event,
>
> if (attr->exclude_idle)
> return -EPERM;
> - if (is_kernel_in_hyp_mode() &&
> - attr->exclude_kernel != attr->exclude_hv)
> - return -EINVAL;
> + if (is_kernel_in_hyp_mode()) {
> + /* include EL2 events, if either of not excluded */
> + if ((attr->exclude_kernel != attr->exclude_hv) ||
> + !attr->exclude_kernel ||
> + !attr->exclude_hv)
> + config_base |= ARMV8_PMU_INCLUDE_EL2;
> + } else {
> + if (attr->exclude_kernel)
> + config_base |= ARMV8_PMU_EXCLUDE_EL1;
> + if (!attr->exclude_hv)
> + config_base |= ARMV8_PMU_INCLUDE_EL2;
> + }
> if (attr->exclude_user)
> config_base |= ARMV8_PMU_EXCLUDE_EL0;
> - if (!is_kernel_in_hyp_mode() && attr->exclude_kernel)
> - config_base |= ARMV8_PMU_EXCLUDE_EL1;
> - if (!attr->exclude_hv)
> - config_base |= ARMV8_PMU_INCLUDE_EL2;
>
> /*
> * Install the filter into config_base as this is used to
> --
> 1.8.1.4
>
Hi Will,
On Tue, Apr 4, 2017 at 4:48 PM, Will Deacon <[email protected]> wrote:
> On Tue, Apr 04, 2017 at 04:10:55PM +0530, Ganapatrao Kulkarni wrote:
>> commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
>> is returning error for perf syscall with mixed attribute set for exclude_kernel
>> and exlude_hv.
>>
>> This change is breaking some applications (observed with hhvm) when
>> ran with VHE enabled. Adding change to enable EL2 event counting,
>> if either of or both of exclude_kernel and exlude_hv are not set.
>>
>> Signed-off-by: Ganapatrao Kulkarni <[email protected]>
>> ---
>> arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
>> 1 file changed, 12 insertions(+), 7 deletions(-)
>
> Hmm. When we have VHE, we can't distinguish between hypervisor and kernel,
> so this patch doesn't seem right to me. The code currently requires
> both exclude_kernel and exclude_hv to be clear before we enable profiling
> EL2, otherwise we're failing to exclude samples that were asked to be
> excluded.
The application cant differentiate that kernel is running in EL2/VHE or in EL1
when VHE=1, is it makes sense to enable EL2 event counting when there
is request from application to either include kernel or hypervisor
event count, since both are same.
IMO, it is not appropriate to have different application behaviour
when kernel booted with VHE=0/1
>
> Will
>
>> diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
>> index daf95919..ea5848a 100644
>> --- a/arch/arm64/kernel/perf_event.c
>> +++ b/arch/arm64/kernel/perf_event.c
>> @@ -871,15 +871,20 @@ static int armv8pmu_set_event_filter(struct hw_perf_event *event,
>>
>> if (attr->exclude_idle)
>> return -EPERM;
>> - if (is_kernel_in_hyp_mode() &&
>> - attr->exclude_kernel != attr->exclude_hv)
>> - return -EINVAL;
>> + if (is_kernel_in_hyp_mode()) {
>> + /* include EL2 events, if either of not excluded */
>> + if ((attr->exclude_kernel != attr->exclude_hv) ||
>> + !attr->exclude_kernel ||
>> + !attr->exclude_hv)
>> + config_base |= ARMV8_PMU_INCLUDE_EL2;
>> + } else {
>> + if (attr->exclude_kernel)
>> + config_base |= ARMV8_PMU_EXCLUDE_EL1;
>> + if (!attr->exclude_hv)
>> + config_base |= ARMV8_PMU_INCLUDE_EL2;
>> + }
>> if (attr->exclude_user)
>> config_base |= ARMV8_PMU_EXCLUDE_EL0;
>> - if (!is_kernel_in_hyp_mode() && attr->exclude_kernel)
>> - config_base |= ARMV8_PMU_EXCLUDE_EL1;
>> - if (!attr->exclude_hv)
>> - config_base |= ARMV8_PMU_INCLUDE_EL2;
>>
>> /*
>> * Install the filter into config_base as this is used to
>> --
>> 1.8.1.4
>>
thanks
Ganapat
On Tue, Apr 04, 2017 at 05:37:10PM +0530, Ganapatrao Kulkarni wrote:
> On Tue, Apr 4, 2017 at 4:48 PM, Will Deacon <[email protected]> wrote:
> > On Tue, Apr 04, 2017 at 04:10:55PM +0530, Ganapatrao Kulkarni wrote:
> >> commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
> >> is returning error for perf syscall with mixed attribute set for exclude_kernel
> >> and exlude_hv.
> >>
> >> This change is breaking some applications (observed with hhvm) when
> >> ran with VHE enabled. Adding change to enable EL2 event counting,
> >> if either of or both of exclude_kernel and exlude_hv are not set.
> >>
> >> Signed-off-by: Ganapatrao Kulkarni <[email protected]>
> >> ---
> >> arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
> >> 1 file changed, 12 insertions(+), 7 deletions(-)
> >
> > Hmm. When we have VHE, we can't distinguish between hypervisor and kernel,
> > so this patch doesn't seem right to me. The code currently requires
> > both exclude_kernel and exclude_hv to be clear before we enable profiling
> > EL2, otherwise we're failing to exclude samples that were asked to be
> > excluded.
>
> The application cant differentiate that kernel is running in EL2/VHE or in EL1
> when VHE=1, is it makes sense to enable EL2 event counting when there
> is request from application to either include kernel or hypervisor
> event count, since both are same.
You can make exactly the same argument against your proposal by saying that
it makes sense to disable EL2 event counting when there is a request from
an application to either exclude kernel or hypervisor event counting.
> IMO, it is not appropriate to have different application behaviour
> when kernel booted with VHE=0/1
Then find another solution to that. How about a mechanism to advertise
that exclude_hv is effectively always set if the kernel is running at EL2?
That would mean that you would use exclude_kernel to determine the profiling
controls for the host.
Will
On Tue, Apr 4, 2017 at 5:56 PM, Will Deacon <[email protected]> wrote:
> On Tue, Apr 04, 2017 at 05:37:10PM +0530, Ganapatrao Kulkarni wrote:
>> On Tue, Apr 4, 2017 at 4:48 PM, Will Deacon <[email protected]> wrote:
>> > On Tue, Apr 04, 2017 at 04:10:55PM +0530, Ganapatrao Kulkarni wrote:
>> >> commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
>> >> is returning error for perf syscall with mixed attribute set for exclude_kernel
>> >> and exlude_hv.
>> >>
>> >> This change is breaking some applications (observed with hhvm) when
>> >> ran with VHE enabled. Adding change to enable EL2 event counting,
>> >> if either of or both of exclude_kernel and exlude_hv are not set.
>> >>
>> >> Signed-off-by: Ganapatrao Kulkarni <[email protected]>
>> >> ---
>> >> arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
>> >> 1 file changed, 12 insertions(+), 7 deletions(-)
>> >
>> > Hmm. When we have VHE, we can't distinguish between hypervisor and kernel,
>> > so this patch doesn't seem right to me. The code currently requires
>> > both exclude_kernel and exclude_hv to be clear before we enable profiling
>> > EL2, otherwise we're failing to exclude samples that were asked to be
>> > excluded.
>>
>> The application cant differentiate that kernel is running in EL2/VHE or in EL1
>> when VHE=1, is it makes sense to enable EL2 event counting when there
>> is request from application to either include kernel or hypervisor
>> event count, since both are same.
>
> You can make exactly the same argument against your proposal by saying that
> it makes sense to disable EL2 event counting when there is a request from
> an application to either exclude kernel or hypervisor event counting.
yes, the argument is equally valid on either side.
>
>> IMO, it is not appropriate to have different application behaviour
>> when kernel booted with VHE=0/1
>
> Then find another solution to that. How about a mechanism to advertise
> that exclude_hv is effectively always set if the kernel is running at EL2?
I am not sure, how we can advertise to user that kernel is running at EL2.
we may add a note to man page of perf_event_open?
"exclude_hv is always set, if host kernel and hypervisor are running
at same privilege level",
>
> That would mean that you would use exclude_kernel to determine the profiling
> controls for the host.
yes, this seems to be more appropriate.
>
> Will
thanks
Ganapat
On Wed, Apr 05, 2017 at 09:29:32AM +0530, Ganapatrao Kulkarni wrote:
> On Tue, Apr 4, 2017 at 5:56 PM, Will Deacon <[email protected]> wrote:
> > On Tue, Apr 04, 2017 at 05:37:10PM +0530, Ganapatrao Kulkarni wrote:
> >> On Tue, Apr 4, 2017 at 4:48 PM, Will Deacon <[email protected]> wrote:
> >> > On Tue, Apr 04, 2017 at 04:10:55PM +0530, Ganapatrao Kulkarni wrote:
> >> >> commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
> >> >> is returning error for perf syscall with mixed attribute set for exclude_kernel
> >> >> and exlude_hv.
> >> >>
> >> >> This change is breaking some applications (observed with hhvm) when
> >> >> ran with VHE enabled. Adding change to enable EL2 event counting,
> >> >> if either of or both of exclude_kernel and exlude_hv are not set.
> >> >>
> >> >> Signed-off-by: Ganapatrao Kulkarni <[email protected]>
> >> >> ---
> >> >> arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
> >> >> 1 file changed, 12 insertions(+), 7 deletions(-)
> >> >
> >> > Hmm. When we have VHE, we can't distinguish between hypervisor and kernel,
> >> > so this patch doesn't seem right to me. The code currently requires
> >> > both exclude_kernel and exclude_hv to be clear before we enable profiling
> >> > EL2, otherwise we're failing to exclude samples that were asked to be
> >> > excluded.
> >>
> >> The application cant differentiate that kernel is running in EL2/VHE or in EL1
> >> when VHE=1, is it makes sense to enable EL2 event counting when there
> >> is request from application to either include kernel or hypervisor
> >> event count, since both are same.
> >
> > You can make exactly the same argument against your proposal by saying that
> > it makes sense to disable EL2 event counting when there is a request from
> > an application to either exclude kernel or hypervisor event counting.
>
> yes, the argument is equally valid on either side.
>
> >
> >> IMO, it is not appropriate to have different application behaviour
> >> when kernel booted with VHE=0/1
> >
> > Then find another solution to that. How about a mechanism to advertise
> > that exclude_hv is effectively always set if the kernel is running at EL2?
>
> I am not sure, how we can advertise to user that kernel is running at EL2.
> we may add a note to man page of perf_event_open?
> "exclude_hv is always set, if host kernel and hypervisor are running
> at same privilege level",
I was thinking of putting something into sysfs, alongside the other things
we have in there. For example, a file that describes whether any of the
perf_event_attr behave as though they are fixed to a certain value. We
should involve the perf maintainers (and perf tool developers) in this,
but perhaps something like an attr directory, where we could have a file
called exclude_hv that contains the value 1.
Will
Hi Perf Maintainers,
your suggestion on below discussion is much appreciated!
On Wed, Apr 5, 2017 at 2:01 PM, Will Deacon <[email protected]> wrote:
> On Wed, Apr 05, 2017 at 09:29:32AM +0530, Ganapatrao Kulkarni wrote:
>> On Tue, Apr 4, 2017 at 5:56 PM, Will Deacon <[email protected]> wrote:
>> > On Tue, Apr 04, 2017 at 05:37:10PM +0530, Ganapatrao Kulkarni wrote:
>> >> On Tue, Apr 4, 2017 at 4:48 PM, Will Deacon <[email protected]> wrote:
>> >> > On Tue, Apr 04, 2017 at 04:10:55PM +0530, Ganapatrao Kulkarni wrote:
>> >> >> commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
>> >> >> is returning error for perf syscall with mixed attribute set for exclude_kernel
>> >> >> and exlude_hv.
>> >> >>
>> >> >> This change is breaking some applications (observed with hhvm) when
>> >> >> ran with VHE enabled. Adding change to enable EL2 event counting,
>> >> >> if either of or both of exclude_kernel and exlude_hv are not set.
>> >> >>
>> >> >> Signed-off-by: Ganapatrao Kulkarni <[email protected]>
>> >> >> ---
>> >> >> arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
>> >> >> 1 file changed, 12 insertions(+), 7 deletions(-)
>> >> >
>> >> > Hmm. When we have VHE, we can't distinguish between hypervisor and kernel,
>> >> > so this patch doesn't seem right to me. The code currently requires
>> >> > both exclude_kernel and exclude_hv to be clear before we enable profiling
>> >> > EL2, otherwise we're failing to exclude samples that were asked to be
>> >> > excluded.
>> >>
>> >> The application cant differentiate that kernel is running in EL2/VHE or in EL1
>> >> when VHE=1, is it makes sense to enable EL2 event counting when there
>> >> is request from application to either include kernel or hypervisor
>> >> event count, since both are same.
>> >
>> > You can make exactly the same argument against your proposal by saying that
>> > it makes sense to disable EL2 event counting when there is a request from
>> > an application to either exclude kernel or hypervisor event counting.
>>
>> yes, the argument is equally valid on either side.
>>
>> >
>> >> IMO, it is not appropriate to have different application behaviour
>> >> when kernel booted with VHE=0/1
>> >
>> > Then find another solution to that. How about a mechanism to advertise
>> > that exclude_hv is effectively always set if the kernel is running at EL2?
>>
>> I am not sure, how we can advertise to user that kernel is running at EL2.
>> we may add a note to man page of perf_event_open?
>> "exclude_hv is always set, if host kernel and hypervisor are running
>> at same privilege level",
>
> I was thinking of putting something into sysfs, alongside the other things
> we have in there. For example, a file that describes whether any of the
> perf_event_attr behave as though they are fixed to a certain value. We
> should involve the perf maintainers (and perf tool developers) in this,
> but perhaps something like an attr directory, where we could have a file
> called exclude_hv that contains the value 1.
>
> Will
thanks
Ganapat
[Adding maintainers]
On Wed, Apr 5, 2017 at 2:01 PM, Will Deacon <[email protected]> wrote:
> On Wed, Apr 05, 2017 at 09:29:32AM +0530, Ganapatrao Kulkarni wrote:
>> On Tue, Apr 4, 2017 at 5:56 PM, Will Deacon <[email protected]> wrote:
>> > On Tue, Apr 04, 2017 at 05:37:10PM +0530, Ganapatrao Kulkarni wrote:
>> >> On Tue, Apr 4, 2017 at 4:48 PM, Will Deacon <[email protected]> wrote:
>> >> > On Tue, Apr 04, 2017 at 04:10:55PM +0530, Ganapatrao Kulkarni wrote:
>> >> >> commit d98ecda (arm64: perf: Count EL2 events if the kernel is running in HYP)
>> >> >> is returning error for perf syscall with mixed attribute set for exclude_kernel
>> >> >> and exlude_hv.
>> >> >>
>> >> >> This change is breaking some applications (observed with hhvm) when
>> >> >> ran with VHE enabled. Adding change to enable EL2 event counting,
>> >> >> if either of or both of exclude_kernel and exlude_hv are not set.
>> >> >>
>> >> >> Signed-off-by: Ganapatrao Kulkarni <[email protected]>
>> >> >> ---
>> >> >> arch/arm64/kernel/perf_event.c | 19 ++++++++++++-------
>> >> >> 1 file changed, 12 insertions(+), 7 deletions(-)
>> >> >
>> >> > Hmm. When we have VHE, we can't distinguish between hypervisor and kernel,
>> >> > so this patch doesn't seem right to me. The code currently requires
>> >> > both exclude_kernel and exclude_hv to be clear before we enable profiling
>> >> > EL2, otherwise we're failing to exclude samples that were asked to be
>> >> > excluded.
>> >>
>> >> The application cant differentiate that kernel is running in EL2/VHE or in EL1
>> >> when VHE=1, is it makes sense to enable EL2 event counting when there
>> >> is request from application to either include kernel or hypervisor
>> >> event count, since both are same.
>> >
>> > You can make exactly the same argument against your proposal by saying that
>> > it makes sense to disable EL2 event counting when there is a request from
>> > an application to either exclude kernel or hypervisor event counting.
>>
>> yes, the argument is equally valid on either side.
>>
>> >
>> >> IMO, it is not appropriate to have different application behaviour
>> >> when kernel booted with VHE=0/1
>> >
>> > Then find another solution to that. How about a mechanism to advertise
>> > that exclude_hv is effectively always set if the kernel is running at EL2?
>>
>> I am not sure, how we can advertise to user that kernel is running at EL2.
>> we may add a note to man page of perf_event_open?
>> "exclude_hv is always set, if host kernel and hypervisor are running
>> at same privilege level",
>
> I was thinking of putting something into sysfs, alongside the other things
> we have in there. For example, a file that describes whether any of the
> perf_event_attr behave as though they are fixed to a certain value. We
> should involve the perf maintainers (and perf tool developers) in this,
> but perhaps something like an attr directory, where we could have a file
> called exclude_hv that contains the value 1.
>
> Will