[PATCH] ARM: test for PMU feature on v7.

[PATCH] ARM: test for PMU feature on v7.

[Rusty Russell]

The v7 ARM (B4.1.82) specifies that bits 27-24 of the ID_DFR0 reg show
what performance monitoring (if any) is available.  We should test this
before assuming (useful for inside virtualized environments, for example).

Signed-off-by: Rusty Russell <rusty.russell@linaro.org>

diff --git a/arch/arm/kernel/perf_event_v7.c b/arch/arm/kernel/perf_event_v7.c
index 6933244..303d1e1 100644
--- a/arch/arm/kernel/perf_event_v7.c
+++ b/arch/arm/kernel/perf_event_v7.c
@@ -1108,6 +1108,12 @@ static struct arm_pmu armv7pmu = {
 	.max_period		= (1LLU << 32) - 1,
 };
 
+static bool armv7_pmu_avail(void)
+{
+	/* 0 means unknown (maybe v1), 1 means v2, 2 means v2, 15 means none */
+	return ((read_cpuid_ext(CPUID_EXT_DFR0) >> 24) & 0xF) != 0xF;
+}
+
 static u32 __init armv7_read_num_pmnc_events(void)
 {
 	u32 nb_cnt;
@@ -1121,6 +1127,8 @@ static u32 __init armv7_read_num_pmnc_events(void)
 
 static struct arm_pmu *__init armv7_a8_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA8;
 	armv7pmu.name		= "ARMv7 Cortex-A8";
 	armv7pmu.map_event	= armv7_a8_map_event;
@@ -1130,6 +1138,8 @@ static struct arm_pmu *__init armv7_a8_pmu_init(void)
 
 static struct arm_pmu *__init armv7_a9_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA9;
 	armv7pmu.name		= "ARMv7 Cortex-A9";
 	armv7pmu.map_event	= armv7_a9_map_event;
@@ -1139,6 +1149,8 @@ static struct arm_pmu *__init armv7_a9_pmu_init(void)
 
 static struct arm_pmu *__init armv7_a5_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA5;
 	armv7pmu.name		= "ARMv7 Cortex-A5";
 	armv7pmu.map_event	= armv7_a5_map_event;
@@ -1148,6 +1160,8 @@ static struct arm_pmu *__init armv7_a5_pmu_init(void)
 
 static struct arm_pmu *__init armv7_a15_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA15;
 	armv7pmu.name		= "ARMv7 Cortex-A15";
 	armv7pmu.map_event	= armv7_a15_map_event;

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Rusty Russell]

The v7 ARM (B4.1.82) specifies that bits 27-24 of the ID_DFR0 reg show
what performance monitoring (if any) is available.  We should test this
before assuming (useful for inside virtualized environments, for example).

Signed-off-by: Rusty Russell <rusty.russell@linaro.org>

diff --git a/arch/arm/kernel/perf_event_v7.c b/arch/arm/kernel/perf_event_v7.c
index 6933244..303d1e1 100644
--- a/arch/arm/kernel/perf_event_v7.c
+++ b/arch/arm/kernel/perf_event_v7.c
@@ -1108,6 +1108,12 @@ static struct arm_pmu armv7pmu = {
 	.max_period		= (1LLU << 32) - 1,
 };
 
+static bool armv7_pmu_avail(void)
+{
+	/* 0 means unknown (maybe v1), 1 means v1, 2 means v2, 15 means none */
+	return ((read_cpuid_ext(CPUID_EXT_DFR0) >> 24) & 0xF) != 0xF;
+}
+
 static u32 __init armv7_read_num_pmnc_events(void)
 {
 	u32 nb_cnt;
@@ -1121,6 +1127,8 @@ static u32 __init armv7_read_num_pmnc_events(void)
 
 static struct arm_pmu *__init armv7_a8_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA8;
 	armv7pmu.name		= "ARMv7 Cortex-A8";
 	armv7pmu.map_event	= armv7_a8_map_event;
@@ -1130,6 +1138,8 @@ static struct arm_pmu *__init armv7_a8_pmu_init(void)
 
 static struct arm_pmu *__init armv7_a9_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA9;
 	armv7pmu.name		= "ARMv7 Cortex-A9";
 	armv7pmu.map_event	= armv7_a9_map_event;
@@ -1139,6 +1149,8 @@ static struct arm_pmu *__init armv7_a9_pmu_init(void)
 
 static struct arm_pmu *__init armv7_a5_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA5;
 	armv7pmu.name		= "ARMv7 Cortex-A5";
 	armv7pmu.map_event	= armv7_a5_map_event;
@@ -1148,6 +1160,8 @@ static struct arm_pmu *__init armv7_a5_pmu_init(void)
 
 static struct arm_pmu *__init armv7_a15_pmu_init(void)
 {
+	if (!armv7_pmu_avail())
+		return NULL;
 	armv7pmu.id		= ARM_PERF_PMU_ID_CA15;
 	armv7pmu.name		= "ARMv7 Cortex-A15";
 	armv7pmu.map_event	= armv7_a15_map_event;

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Will Deacon]

Hi Rusty,

On Fri, Mar 23, 2012 at 12:34:57AM +0000, Rusty Russell wrote:
> The v7 ARM (B4.1.82) specifies that bits 27-24 of the ID_DFR0 reg show
> what performance monitoring (if any) is available.  We should test this
> before assuming (useful for inside virtualized environments, for example).

Ok, I can see why you don't want to emulate a PMU but I don't agree with the
patch :)

> diff --git a/arch/arm/kernel/perf_event_v7.c b/arch/arm/kernel/perf_event_v7.c
> index 6933244..303d1e1 100644
> --- a/arch/arm/kernel/perf_event_v7.c
> +++ b/arch/arm/kernel/perf_event_v7.c
> @@ -1108,6 +1108,12 @@ static struct arm_pmu armv7pmu = {
>  	.max_period		= (1LLU << 32) - 1,
>  };
>  
> +static bool armv7_pmu_avail(void)
> +{
> +	/* 0 means unknown (maybe v1), 1 means v1, 2 means v2, 15 means none */
> +	return ((read_cpuid_ext(CPUID_EXT_DFR0) >> 24) & 0xF) != 0xF;
> +}
> +
>  static u32 __init armv7_read_num_pmnc_events(void)
>  {
>  	u32 nb_cnt;
> @@ -1121,6 +1127,8 @@ static u32 __init armv7_read_num_pmnc_events(void)
>  
>  static struct arm_pmu *__init armv7_a8_pmu_init(void)
>  {
> +	if (!armv7_pmu_avail())
> +		return NULL;

This code is only executed if we know for sure that we are running on a
Cortex-A8. Cortex-A8 implementations have a PMU, so I don't like having the
probe here in case we are running on a virtualised CPU that doesn't quite
match the hardware.

Would you be able to advertise 0 event counters instead and treat the PMU
largely as RAZ/WI? The only tricky bit I can see is the cycle counter, which
is mandated by the presence of a PMU, however this could also just RAZ
initially.

>  	armv7pmu.id		= ARM_PERF_PMU_ID_CA8;
>  	armv7pmu.name		= "ARMv7 Cortex-A8";
>  	armv7pmu.map_event	= armv7_a8_map_event;
> @@ -1130,6 +1138,8 @@ static struct arm_pmu *__init armv7_a8_pmu_init(void)
>  
>  static struct arm_pmu *__init armv7_a9_pmu_init(void)
>  {
> +	if (!armv7_pmu_avail())
> +		return NULL;

Similarly for A5, A9, A7 and A15...

Cheers,

Will

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Rusty Russell]

On Fri, 23 Mar 2012 09:53:14 +0000, Will Deacon <will.deacon@arm.com> wrote:
> Hi Rusty,
> 
> On Fri, Mar 23, 2012 at 12:34:57AM +0000, Rusty Russell wrote:
> > The v7 ARM (B4.1.82) specifies that bits 27-24 of the ID_DFR0 reg show
> > what performance monitoring (if any) is available.  We should test this
> > before assuming (useful for inside virtualized environments, for example).
> 
> Ok, I can see why you don't want to emulate a PMU but I don't agree with the
> patch :)
...
> >  static struct arm_pmu *__init armv7_a8_pmu_init(void)
> >  {
> > +	if (!armv7_pmu_avail())
> > +		return NULL;
> 
> This code is only executed if we know for sure that we are running on a
> Cortex-A8. Cortex-A8 implementations have a PMU, so I don't like having the
> probe here in case we are running on a virtualised CPU that doesn't quite
> match the hardware.

OK, I see that argument, but I don't like it.  eg. the Cortex A-15 is
defined to have a PMUv2 with 6 counters, and below you're arguing we
should implement zero of them.

I had assumed that as we trend towards a common cross-platform kernel,
we should be relying on feature registers where available, not the
model.

> Would you be able to advertise 0 event counters instead and treat the PMU
> largely as RAZ/WI? The only tricky bit I can see is the cycle counter, which
> is mandated by the presence of a PMU, however this could also just RAZ
> initially.

Yes, that's the effect of the current emulation hack.  Far better not to
lie to the guest, however, than get weird results from profiling.

Cheers,
Rusty.
-- 
  How could I marry someone with more hair than me?  http://baldalex.org

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Will Deacon]

[removed some dead email addresses from CC]

On Fri, Mar 23, 2012 at 11:17:07PM +0000, Rusty Russell wrote:
> On Fri, 23 Mar 2012 09:53:14 +0000, Will Deacon <will.deacon@arm.com> wrote:
> > On Fri, Mar 23, 2012 at 12:34:57AM +0000, Rusty Russell wrote:
> ...
> > >  static struct arm_pmu *__init armv7_a8_pmu_init(void)
> > >  {
> > > +	if (!armv7_pmu_avail())
> > > +		return NULL;
> > 
> > This code is only executed if we know for sure that we are running on a
> > Cortex-A8. Cortex-A8 implementations have a PMU, so I don't like having the
> > probe here in case we are running on a virtualised CPU that doesn't quite
> > match the hardware.
> 
> OK, I see that argument, but I don't like it.  eg. the Cortex A-15 is
> defined to have a PMUv2 with 6 counters, and below you're arguing we
> should implement zero of them.

Well, I think you should actually implement all 6 of them if that's the
target which you're emulating. However, there are bigger peripherals to fry
than the PMU so if you want to get something working quickly then a dummy
PMU makes sense as a stopgap.

> I had assumed that as we trend towards a common cross-platform kernel,
> we should be relying on feature registers where available, not the
> model.

The PMU is very tightly coupled with the CPU and it is necessary to know
exactly which CPU you have so that you can map microarchitectural events to
their identifiers. This is not discoverable in the architecture [PMUv2
provides some support for discovering the status of architected events, but
that is still not enough and is only present on A7 and A15 so far] so that's
why we probe the CPU instead.

Now, if everything was device-tree based then we could simply use a
different binding for each CPU but since we support perf on non-DT
platforms, probing the CPU type is the best solution. I would like to avoid
the probing code if we are initialised from DT, but I've not got round to it
yet (this would be useful for big.LITTLE).

> > Would you be able to advertise 0 event counters instead and treat the PMU
> > largely as RAZ/WI? The only tricky bit I can see is the cycle counter, which
> > is mandated by the presence of a PMU, however this could also just RAZ
> > initially.
> 
> Yes, that's the effect of the current emulation hack.  Far better not to
> lie to the guest, however, than get weird results from profiling.

I'd rather lie to the guest than modify it just to avoid emulating the
device correctly. Even then, the profiling results wouldn't be especially
weird -- everything would be 0, which is a lot better than random numbers.

Will

[Android-virt] [PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Peter Maydell]

On 26 March 2012 10:40, Will Deacon <will.deacon@arm.com> wrote:
> [removed some dead email addresses from CC]
>
> On Fri, Mar 23, 2012 at 11:17:07PM +0000, Rusty Russell wrote:
>> On Fri, 23 Mar 2012 09:53:14 +0000, Will Deacon <will.deacon@arm.com> wrote:
>> > This code is only executed if we know for sure that we are running on a
>> > Cortex-A8. Cortex-A8 implementations have a PMU, so I don't like having the
>> > probe here in case we are running on a virtualised CPU that doesn't quite
>> > match the hardware.
>>
>> OK, I see that argument, but I don't like it. ?eg. the Cortex A-15 is
>> defined to have a PMUv2 with 6 counters, and below you're arguing we
>> should implement zero of them.
>
> Well, I think you should actually implement all 6 of them if that's the
> target which you're emulating.

The trouble with this line of thought is that it leads to the requirement
to support nested virtualization and virtualization of TrustZone, neither
of which are actually particularly useful or simple to do. At some point
we're going to end up saying that the "A15" we expose to the guest is
not an exact match to the real hardware, that it's missing certain
features (and hopefully that the lack of those features is advertised
via the feature registers). So the question is where you draw the line...

-- PMM

[Android-virt] [PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Will Deacon]

On Mon, Mar 26, 2012 at 10:50:35AM +0100, Peter Maydell wrote:
> On 26 March 2012 10:40, Will Deacon <will.deacon@arm.com> wrote:
> > On Fri, Mar 23, 2012 at 11:17:07PM +0000, Rusty Russell wrote:
> >>
> >> OK, I see that argument, but I don't like it. ?eg. the Cortex A-15 is
> >> defined to have a PMUv2 with 6 counters, and below you're arguing we
> >> should implement zero of them.
> >
> > Well, I think you should actually implement all 6 of them if that's the
> > target which you're emulating.
> 
> The trouble with this line of thought is that it leads to the requirement
> to support nested virtualization and virtualization of TrustZone, neither
> of which are actually particularly useful or simple to do. At some point
> we're going to end up saying that the "A15" we expose to the guest is
> not an exact match to the real hardware, that it's missing certain
> features (and hopefully that the lack of those features is advertised
> via the feature registers). So the question is where you draw the line...

I'd be inclined to say that, at least for kvm, you should draw the line at
the point where the guest doesn't need modifications to run in the virtual
environment. That is, everything used by Linux is emulated enough for the
kernel to run happily. Clearly you can improve on that in the areas which
are considered to be most critical.

Aside: given that guests have to run non-secure, how would you emulate a
secure platform without emulating the corresponding TrustZone API? Or is
that one of the reasons why you're currently targetting Versatile Express?

Will

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Rusty Russell]

On Mon, 26 Mar 2012 10:40:54 +0100, Will Deacon <will.deacon@arm.com> wrote:
> [removed some dead email addresses from CC]

Thanks!

> On Fri, Mar 23, 2012 at 11:17:07PM +0000, Rusty Russell wrote:
> > I had assumed that as we trend towards a common cross-platform kernel,
> > we should be relying on feature registers where available, not the
> > model.
> 
> The PMU is very tightly coupled with the CPU and it is necessary to know
> exactly which CPU you have so that you can map microarchitectural events to
> their identifiers. This is not discoverable in the architecture [PMUv2
> provides some support for discovering the status of architected events, but
> that is still not enough and is only present on A7 and A15 so far] so that's
> why we probe the CPU instead.

OK.

> Now, if everything was device-tree based then we could simply use a
> different binding for each CPU but since we support perf on non-DT
> platforms, probing the CPU type is the best solution. I would like to avoid
> the probing code if we are initialised from DT, but I've not got round to it
> yet (this would be useful for big.LITTLE).

I'll be interested to see how you encode the mapping of
microarchitectural events to their identifiers here.

> > > Would you be able to advertise 0 event counters instead and treat the PMU
> > > largely as RAZ/WI? The only tricky bit I can see is the cycle counter, which
> > > is mandated by the presence of a PMU, however this could also just RAZ
> > > initially.
> > 
> > Yes, that's the effect of the current emulation hack.  Far better not to
> > lie to the guest, however, than get weird results from profiling.
> 
> I'd rather lie to the guest than modify it just to avoid emulating the
> device correctly. Even then, the profiling results wouldn't be especially
> weird -- everything would be 0, which is a lot better than random
> numbers.

Sure, but nowhere as neat an experience as having "no hardware support
available" printed, and nicely defined failure mode :(

When I get back in May, I'll look at how hard it will to be implement
this properly, if noone else does.

Thanks,
Rusty.
-- 
  How could I marry someone with more hair than me?  http://baldalex.org

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Nicolas Pitre]

On Mon, 26 Mar 2012, Will Deacon wrote:

> Now, if everything was device-tree based then we could simply use a
> different binding for each CPU but since we support perf on non-DT
> platforms, probing the CPU type is the best solution. I would like to avoid
> the probing code if we are initialised from DT, but I've not got round to it
> yet (this would be useful for big.LITTLE).

Still... my opinion is that we should try to autodetect as much as 
possible and avoid overstuffing the DT with content that can otherwise 
be run-time probed.  OK to use DT to override the probe for corner 
cases, but IMHO the probe should be the default method of 
initialization.  The rational is that we want to spread knowledge about 
part of the system and have it confined into respective drivers and 
subsystems for easier maintenance.  If the guy who has to maintain the 
dts has to know all the details for everything then that won't scale and 
the risk for discrepancies is increased.


Nicolas

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Will Deacon]

Hi Nicolas,

On Wed, Mar 28, 2012 at 03:17:29PM +0100, Nicolas Pitre wrote:
> On Mon, 26 Mar 2012, Will Deacon wrote:
> 
> > Now, if everything was device-tree based then we could simply use a
> > different binding for each CPU but since we support perf on non-DT
> > platforms, probing the CPU type is the best solution. I would like to avoid
> > the probing code if we are initialised from DT, but I've not got round to it
> > yet (this would be useful for big.LITTLE).
> 
> Still... my opinion is that we should try to autodetect as much as 
> possible and avoid overstuffing the DT with content that can otherwise 
> be run-time probed.  OK to use DT to override the probe for corner 
> cases, but IMHO the probe should be the default method of 
> initialization.  The rational is that we want to spread knowledge about 
> part of the system and have it confined into respective drivers and 
> subsystems for easier maintenance.  If the guy who has to maintain the 
> dts has to know all the details for everything then that won't scale and 
> the risk for discrepancies is increased.

I agree that probing is preferable where possible but, since the PMUs are
banked, we cannot reliably probe them on big.LITTLE platforms. I guess if
there were some infrastructure for probing on the remote cluster, I could
use that, but it seems like DT would be easier.

Will

[PATCH] ARM: test for PMU feature on v7 (v2 with typo fix)

[Nicolas Pitre]

On Fri, 30 Mar 2012, Will Deacon wrote:

> Hi Nicolas,
> 
> On Wed, Mar 28, 2012 at 03:17:29PM +0100, Nicolas Pitre wrote:
> > On Mon, 26 Mar 2012, Will Deacon wrote:
> > 
> > > Now, if everything was device-tree based then we could simply use a
> > > different binding for each CPU but since we support perf on non-DT
> > > platforms, probing the CPU type is the best solution. I would like to avoid
> > > the probing code if we are initialised from DT, but I've not got round to it
> > > yet (this would be useful for big.LITTLE).
> > 
> > Still... my opinion is that we should try to autodetect as much as 
> > possible and avoid overstuffing the DT with content that can otherwise 
> > be run-time probed.  OK to use DT to override the probe for corner 
> > cases, but IMHO the probe should be the default method of 
> > initialization.  The rational is that we want to spread knowledge about 
> > part of the system and have it confined into respective drivers and 
> > subsystems for easier maintenance.  If the guy who has to maintain the 
> > dts has to know all the details for everything then that won't scale and 
> > the risk for discrepancies is increased.
> 
> I agree that probing is preferable where possible but, since the PMUs are
> banked, we cannot reliably probe them on big.LITTLE platforms.

Are they tied to each cluster power domains?  If so you'll have to 
preserve/restore their state when clusters are suspended even if the 
interface tries to present a consistent view to users.  The probe can be 
done there, and differences in capabilities adjusted dynamically.


Nicolas