arm64: perf: Don't disgard upper 32 bits from PMCEID0/1 registers

[PATCH] arm64: perf: Don't disgard upper 32 bits from PMCEID0/1 registers

Posted by Ilkka Koskinen 2 years, 1 month ago

The upper 32 bits of PMCEID[n] registers are used to describe whether
architectural and microarchitectural events in range 0x4000-0x401f
exist. Due to disgarding the bits, the driver made the events invisible,
even if they existed.

Fixes: df29ddf4f04b ("arm64: perf: Abstract system register accesses away")
Reported-by: Carl Worth <carl@os.amperecomputing.com>
Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
---
 arch/arm64/include/asm/arm_pmuv3.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/arm64/include/asm/arm_pmuv3.h b/arch/arm64/include/asm/arm_pmuv3.h
index 18dc2fb3d7b7..3e92b7cb57a4 100644
--- a/arch/arm64/include/asm/arm_pmuv3.h
+++ b/arch/arm64/include/asm/arm_pmuv3.h
@@ -126,12 +126,12 @@ static inline void write_pmuserenr(u32 val)
 	write_sysreg(val, pmuserenr_el0);
 }
 
-static inline u32 read_pmceid0(void)
+static inline u64 read_pmceid0(void)
 {
 	return read_sysreg(pmceid0_el0);
 }
 
-static inline u32 read_pmceid1(void)
+static inline u64 read_pmceid1(void)
 {
 	return read_sysreg(pmceid1_el0);
 }
-- 
2.41.0

Re: [PATCH] arm64: perf: Don't disgard upper 32 bits from PMCEID0/1 registers

Posted by Marc Zyngier 2 years, 1 month ago

On Wed, 25 Oct 2023 21:08:15 +0100,
Ilkka Koskinen <ilkka@os.amperecomputing.com> wrote:
> 
> The upper 32 bits of PMCEID[n] registers are used to describe whether
> architectural and microarchitectural events in range 0x4000-0x401f
> exist. Due to disgarding the bits, the driver made the events invisible,
> even if they existed.
> 
> Fixes: df29ddf4f04b ("arm64: perf: Abstract system register accesses away")
> Reported-by: Carl Worth <carl@os.amperecomputing.com>
> Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
> ---
>  arch/arm64/include/asm/arm_pmuv3.h | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/arm_pmuv3.h b/arch/arm64/include/asm/arm_pmuv3.h
> index 18dc2fb3d7b7..3e92b7cb57a4 100644
> --- a/arch/arm64/include/asm/arm_pmuv3.h
> +++ b/arch/arm64/include/asm/arm_pmuv3.h
> @@ -126,12 +126,12 @@ static inline void write_pmuserenr(u32 val)
>  	write_sysreg(val, pmuserenr_el0);
>  }
>  
> -static inline u32 read_pmceid0(void)
> +static inline u64 read_pmceid0(void)
>  {
>  	return read_sysreg(pmceid0_el0);
>  }
>  
> -static inline u32 read_pmceid1(void)
> +static inline u64 read_pmceid1(void)
>  {
>  	return read_sysreg(pmceid1_el0);
>  }

This is necessary, but not sufficient. You also need to update the
corresponding 32bit accessors to match this behaviour. Something along
the lines of the patch below

	M.

diff --git a/arch/arm/include/asm/arm_pmuv3.h b/arch/arm/include/asm/arm_pmuv3.h
index 72529f5e2bed..90841cb7ce43 100644
--- a/arch/arm/include/asm/arm_pmuv3.h
+++ b/arch/arm/include/asm/arm_pmuv3.h
@@ -23,6 +23,8 @@
 #define PMUSERENR		__ACCESS_CP15(c9,  0, c14, 0)
 #define PMINTENSET		__ACCESS_CP15(c9,  0, c14, 1)
 #define PMINTENCLR		__ACCESS_CP15(c9,  0, c14, 2)
+#define PMCEID2			__ACCESS_CP15(c9,  0, c14, 4)
+#define PMCEID3			__ACCESS_CP15(c9,  0, c14, 5)
 #define PMMIR			__ACCESS_CP15(c9,  0, c14, 6)
 #define PMCCFILTR		__ACCESS_CP15(c14, 0, c15, 7)
 
@@ -205,16 +207,6 @@ static inline void write_pmuserenr(u32 val)
 	write_sysreg(val, PMUSERENR);
 }
 
-static inline u32 read_pmceid0(void)
-{
-	return read_sysreg(PMCEID0);
-}
-
-static inline u32 read_pmceid1(void)
-{
-	return read_sysreg(PMCEID1);
-}
-
 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_pmu_counter_deferred(struct perf_event_attr *attr)
@@ -231,6 +223,7 @@ static inline void kvm_vcpu_pmu_resync_el0(void) {}
 
 /* PMU Version in DFR Register */
 #define ARMV8_PMU_DFR_VER_NI        0
+#define ARMV8_PMU_DFR_VER_V3P1      0x4
 #define ARMV8_PMU_DFR_VER_V3P4      0x5
 #define ARMV8_PMU_DFR_VER_V3P5      0x6
 #define ARMV8_PMU_DFR_VER_IMP_DEF   0xF
@@ -251,4 +244,24 @@ static inline bool is_pmuv3p5(int pmuver)
 	return pmuver >= ARMV8_PMU_DFR_VER_V3P5;
 }
 
+static inline u64 read_pmceid0(void)
+{
+	u64 val = read_sysreg(PMCEID0);
+
+	if (read_pmuver() >= ARMV8_PMU_DFR_VER_V3P1)
+		val |= (u64)read_sysreg(PMCEID2) << 32;
+
+	return val;
+}
+
+static inline u64 read_pmceid1(void)
+{
+	u64 val = read_sysreg(PMCEID1);
+
+	if (read_pmuver() >= ARMV8_PMU_DFR_VER_V3P1)
+		val |= (u64)read_sysreg(PMCEID3) << 32;
+
+	return val;
+}
+
 #endif

-- 
Without deviation from the norm, progress is not possible.

Re: [PATCH] arm64: perf: Don't disgard upper 32 bits from PMCEID0/1 registers

Posted by Ilkka Koskinen 2 years, 1 month ago

Hi Marc,

On Thu, 26 Oct 2023, Marc Zyngier wrote:
> On Wed, 25 Oct 2023 21:08:15 +0100,
> Ilkka Koskinen <ilkka@os.amperecomputing.com> wrote:
>>
>> The upper 32 bits of PMCEID[n] registers are used to describe whether
>> architectural and microarchitectural events in range 0x4000-0x401f
>> exist. Due to disgarding the bits, the driver made the events invisible,
>> even if they existed.
>>
>> Fixes: df29ddf4f04b ("arm64: perf: Abstract system register accesses away")
>> Reported-by: Carl Worth <carl@os.amperecomputing.com>
>> Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
>> ---
>>  arch/arm64/include/asm/arm_pmuv3.h | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/arm_pmuv3.h b/arch/arm64/include/asm/arm_pmuv3.h
>> index 18dc2fb3d7b7..3e92b7cb57a4 100644
>> --- a/arch/arm64/include/asm/arm_pmuv3.h
>> +++ b/arch/arm64/include/asm/arm_pmuv3.h
>> @@ -126,12 +126,12 @@ static inline void write_pmuserenr(u32 val)
>>  	write_sysreg(val, pmuserenr_el0);
>>  }
>>
>> -static inline u32 read_pmceid0(void)
>> +static inline u64 read_pmceid0(void)
>>  {
>>  	return read_sysreg(pmceid0_el0);
>>  }
>>
>> -static inline u32 read_pmceid1(void)
>> +static inline u64 read_pmceid1(void)
>>  {
>>  	return read_sysreg(pmceid1_el0);
>>  }
>
> This is necessary, but not sufficient. You also need to update the
> corresponding 32bit accessors to match this behaviour. Something along
> the lines of the patch below
>
> 	M.

Thanks! I'll add your code to the second version.

Cheers, Ilkka


>
> diff --git a/arch/arm/include/asm/arm_pmuv3.h b/arch/arm/include/asm/arm_pmuv3.h
> index 72529f5e2bed..90841cb7ce43 100644
> --- a/arch/arm/include/asm/arm_pmuv3.h
> +++ b/arch/arm/include/asm/arm_pmuv3.h
> @@ -23,6 +23,8 @@
> #define PMUSERENR		__ACCESS_CP15(c9,  0, c14, 0)
> #define PMINTENSET		__ACCESS_CP15(c9,  0, c14, 1)
> #define PMINTENCLR		__ACCESS_CP15(c9,  0, c14, 2)
> +#define PMCEID2			__ACCESS_CP15(c9,  0, c14, 4)
> +#define PMCEID3			__ACCESS_CP15(c9,  0, c14, 5)
> #define PMMIR			__ACCESS_CP15(c9,  0, c14, 6)
> #define PMCCFILTR		__ACCESS_CP15(c14, 0, c15, 7)
>
> @@ -205,16 +207,6 @@ static inline void write_pmuserenr(u32 val)
> 	write_sysreg(val, PMUSERENR);
> }
>
> -static inline u32 read_pmceid0(void)
> -{
> -	return read_sysreg(PMCEID0);
> -}
> -
> -static inline u32 read_pmceid1(void)
> -{
> -	return read_sysreg(PMCEID1);
> -}
> -
> 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_pmu_counter_deferred(struct perf_event_attr *attr)
> @@ -231,6 +223,7 @@ static inline void kvm_vcpu_pmu_resync_el0(void) {}
>
> /* PMU Version in DFR Register */
> #define ARMV8_PMU_DFR_VER_NI        0
> +#define ARMV8_PMU_DFR_VER_V3P1      0x4
> #define ARMV8_PMU_DFR_VER_V3P4      0x5
> #define ARMV8_PMU_DFR_VER_V3P5      0x6
> #define ARMV8_PMU_DFR_VER_IMP_DEF   0xF
> @@ -251,4 +244,24 @@ static inline bool is_pmuv3p5(int pmuver)
> 	return pmuver >= ARMV8_PMU_DFR_VER_V3P5;
> }
>
> +static inline u64 read_pmceid0(void)
> +{
> +	u64 val = read_sysreg(PMCEID0);
> +
> +	if (read_pmuver() >= ARMV8_PMU_DFR_VER_V3P1)
> +		val |= (u64)read_sysreg(PMCEID2) << 32;
> +
> +	return val;
> +}
> +
> +static inline u64 read_pmceid1(void)
> +{
> +	u64 val = read_sysreg(PMCEID1);
> +
> +	if (read_pmuver() >= ARMV8_PMU_DFR_VER_V3P1)
> +		val |= (u64)read_sysreg(PMCEID3) << 32;
> +
> +	return val;
> +}
> +
> #endif
>
> -- 
> Without deviation from the norm, progress is not possible.
>