diff mbox series

[1/4] drivers/perf: Add support for ARMv8.3-SPE

Message ID 20200724091607.41903-2-liwei391@huawei.com (mailing list archive)
State New, archived
Headers show
Series Add support for ARMv8.3-SPE | expand

Commit Message

Wei Li July 24, 2020, 9:16 a.m. UTC
Armv8.3 extends the SPE by adding:
- Alignment field in the Events packet, and filtering on this event
  using PMSEVFR_EL1.
- Support for the Scalable Vector Extension (SVE).

The main additions for SVE are:
- Recording the vector length for SVE operations in the Operation Type
  packet. It is not possible to filter on vector length.
- Incomplete predicate and empty predicate fields in the Events packet,
  and filtering on these events using PMSEVFR_EL1.

Update the check of pmsevfr for empty/partial predicated SVE and
alignment event in kernel driver.

Signed-off-by: Wei Li <liwei391@huawei.com>
---
 arch/arm64/include/asm/sysreg.h |  4 +++-
 drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
 2 files changed, 17 insertions(+), 5 deletions(-)

Comments

Leo Yan July 28, 2020, 12:27 p.m. UTC | #1
Hi Wei,

On Fri, Jul 24, 2020 at 05:16:04PM +0800, Wei Li wrote:
> Armv8.3 extends the SPE by adding:
> - Alignment field in the Events packet, and filtering on this event
>   using PMSEVFR_EL1.
> - Support for the Scalable Vector Extension (SVE).
> 
> The main additions for SVE are:
> - Recording the vector length for SVE operations in the Operation Type
>   packet. It is not possible to filter on vector length.
> - Incomplete predicate and empty predicate fields in the Events packet,
>   and filtering on these events using PMSEVFR_EL1.
> 
> Update the check of pmsevfr for empty/partial predicated SVE and
> alignment event in kernel driver.
> 
> Signed-off-by: Wei Li <liwei391@huawei.com>
> ---
>  arch/arm64/include/asm/sysreg.h |  4 +++-
>  drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
>  2 files changed, 17 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
> index 463175f80341..be4c44ccdb56 100644
> --- a/arch/arm64/include/asm/sysreg.h
> +++ b/arch/arm64/include/asm/sysreg.h
> @@ -281,7 +281,6 @@
>  #define SYS_PMSFCR_EL1_ST_SHIFT		18
>  
>  #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
>  
>  #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>  #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
> @@ -769,6 +768,9 @@
>  #define ID_AA64DFR0_PMUVER_8_5		0x6
>  #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>  
> +#define ID_AA64DFR0_PMSVER_8_2		0x1
> +#define ID_AA64DFR0_PMSVER_8_3		0x2
> +
>  #define ID_DFR0_PERFMON_SHIFT		24
>  
>  #define ID_DFR0_PERFMON_8_1		0x4
> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> index e51ddb6d63ed..5ec7ee0c8fa1 100644
> --- a/drivers/perf/arm_spe_pmu.c
> +++ b/drivers/perf/arm_spe_pmu.c
> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>  	struct hlist_node			hotplug_node;
>  
>  	int					irq; /* PPI */
> -
> +	int					pmuver;

Since the version number is only 4 bits width, 'u16' would be enough
to record SPE version number.

>  	u16					min_period;
>  	u16					counter_sz;
>  
> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>  /* Keep track of our dynamic hotplug state */
>  static enum cpuhp_state arm_spe_pmu_online;
>  
> +static u64 sys_pmsevfr_el1_mask[] = {
> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
> +		BIT_ULL(1),
> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
> +};

Seems to me, the definitions for Aarch64 system registers should be
placed into the file 'arch/arm64/include/asm/sysreg.h'.  Like below
two macros:

  #define SYS_PMSEVFR_EL1_RES0_8_2		0x0000ffff00ff0f55UL
  #define SYS_PMSEVFR_EL1_RES0_8_3		...

Let's wait for Will or Mark Rutland's comments for this, in case I
mislead for this.

> +
>  enum arm_spe_pmu_buf_fault_action {
>  	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
>  	SPE_PMU_BUF_FAULT_ACT_FATAL,
> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
>  	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
>  		return -ENOENT;
>  
> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
>  		return -EOPNOTSUPP;
>  
>  	if (attr->exclude_idle)
> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
>  			fld, smp_processor_id());
>  		return;
>  	}
> +	spe_pmu->pmuver = fld;
>  
>  	/* Read PMBIDR first to determine whether or not we have access */
>  	reg = read_sysreg_s(SYS_PMBIDR_EL1);
> @@ -1027,8 +1037,8 @@ static void __arm_spe_pmu_dev_probe(void *info)
>  	}
>  
>  	dev_info(dev,
> -		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
> -		 cpumask_pr_args(&spe_pmu->supported_cpus),
> +		 "v%d probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",

Let's output explict info, like:

  "probed for CPUs %*pbl [pmuver %d, max_record_sz %u, align %u, features 0x%llx]\n",

Thanks,
Leo

> +		 spe_pmu->pmuver, cpumask_pr_args(&spe_pmu->supported_cpus),
>  		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
>  
>  	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
> -- 
> 2.17.1
>
Wei Li July 28, 2020, 1:24 p.m. UTC | #2
Hi Leo,

On 2020/7/28 20:27, Leo Yan wrote:
> Hi Wei,
> 
> On Fri, Jul 24, 2020 at 05:16:04PM +0800, Wei Li wrote:
>> Armv8.3 extends the SPE by adding:
>> - Alignment field in the Events packet, and filtering on this event
>>   using PMSEVFR_EL1.
>> - Support for the Scalable Vector Extension (SVE).
>>
>> The main additions for SVE are:
>> - Recording the vector length for SVE operations in the Operation Type
>>   packet. It is not possible to filter on vector length.
>> - Incomplete predicate and empty predicate fields in the Events packet,
>>   and filtering on these events using PMSEVFR_EL1.
>>
>> Update the check of pmsevfr for empty/partial predicated SVE and
>> alignment event in kernel driver.
>>
>> Signed-off-by: Wei Li <liwei391@huawei.com>
>> ---
>>  arch/arm64/include/asm/sysreg.h |  4 +++-
>>  drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
>>  2 files changed, 17 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
>> index 463175f80341..be4c44ccdb56 100644
>> --- a/arch/arm64/include/asm/sysreg.h
>> +++ b/arch/arm64/include/asm/sysreg.h
>> @@ -281,7 +281,6 @@
>>  #define SYS_PMSFCR_EL1_ST_SHIFT		18
>>  
>>  #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
>> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
>>  
>>  #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>>  #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
>> @@ -769,6 +768,9 @@
>>  #define ID_AA64DFR0_PMUVER_8_5		0x6
>>  #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>>  
>> +#define ID_AA64DFR0_PMSVER_8_2		0x1
>> +#define ID_AA64DFR0_PMSVER_8_3		0x2
>> +
>>  #define ID_DFR0_PERFMON_SHIFT		24
>>  
>>  #define ID_DFR0_PERFMON_8_1		0x4
>> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
>> index e51ddb6d63ed..5ec7ee0c8fa1 100644
>> --- a/drivers/perf/arm_spe_pmu.c
>> +++ b/drivers/perf/arm_spe_pmu.c
>> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>>  	struct hlist_node			hotplug_node;
>>  
>>  	int					irq; /* PPI */
>> -
>> +	int					pmuver;
> 
> Since the version number is only 4 bits width, 'u16' would be enough
> to record SPE version number.

Sounds reasonable, i can change it to 'u16' if you insist.

>>  	u16					min_period;
>>  	u16					counter_sz;
>>  
>> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>>  /* Keep track of our dynamic hotplug state */
>>  static enum cpuhp_state arm_spe_pmu_online;
>>  
>> +static u64 sys_pmsevfr_el1_mask[] = {
>> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
>> +		BIT_ULL(1),
>> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
>> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
>> +};
> 
> Seems to me, the definitions for Aarch64 system registers should be
> placed into the file 'arch/arm64/include/asm/sysreg.h'.  Like below
> two macros:
> 
>   #define SYS_PMSEVFR_EL1_RES0_8_2		0x0000ffff00ff0f55UL
>   #define SYS_PMSEVFR_EL1_RES0_8_3		...

I really think using GENMASK_ULL() to generate the mask is better than a definition
with magic number. It is beneficial to be reviewed and extended later.

> Let's wait for Will or Mark Rutland's comments for this, in case I
> mislead for this.
>> +
>>  enum arm_spe_pmu_buf_fault_action {
>>  	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
>>  	SPE_PMU_BUF_FAULT_ACT_FATAL,
>> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
>>  	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
>>  		return -ENOENT;
>>  
>> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
>> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
>>  		return -EOPNOTSUPP;
>>  
>>  	if (attr->exclude_idle)
>> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
>>  			fld, smp_processor_id());
>>  		return;
>>  	}
>> +	spe_pmu->pmuver = fld;
>>  
>>  	/* Read PMBIDR first to determine whether or not we have access */
>>  	reg = read_sysreg_s(SYS_PMBIDR_EL1);
>> @@ -1027,8 +1037,8 @@ static void __arm_spe_pmu_dev_probe(void *info)
>>  	}
>>  
>>  	dev_info(dev,
>> -		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
>> -		 cpumask_pr_args(&spe_pmu->supported_cpus),
>> +		 "v%d probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
> 
> Let's output explict info, like:
> 
>   "probed for CPUs %*pbl [pmuver %d, max_record_sz %u, align %u, features 0x%llx]\n",
> 

Agree, and i have a little question here:
Currently, the of_compatible of SPE PMU is "arm,statistical-profiling-extension-v1", and
the platform_device name is "arm,spe-v1". So this message looks weird when supporting
ARMv8.3-SPE because the pmuver is 2.

As the version of SPE can be probed by reading 'ID_AA64DFR0_EL1.PMSVer', can we remove
the version hint in of_compatible and platform_device name?

> 
>> +		 spe_pmu->pmuver, cpumask_pr_args(&spe_pmu->supported_cpus),
>>  		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
>>  
>>  	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
>> -- 
>> 2.17.1
>>

Thanks,
Wei
Leo Yan July 29, 2020, 7:08 a.m. UTC | #3
On Tue, Jul 28, 2020 at 09:24:42PM +0800, liwei (GF) wrote:

[...]

> >> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> >> index e51ddb6d63ed..5ec7ee0c8fa1 100644
> >> --- a/drivers/perf/arm_spe_pmu.c
> >> +++ b/drivers/perf/arm_spe_pmu.c
> >> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
> >>  	struct hlist_node			hotplug_node;
> >>  
> >>  	int					irq; /* PPI */
> >> -
> >> +	int					pmuver;
> > 
> > Since the version number is only 4 bits width, 'u16' would be enough
> > to record SPE version number.
> 
> Sounds reasonable, i can change it to 'u16' if you insist.
> 
> >>  	u16					min_period;
> >>  	u16					counter_sz;
> >>  
> >> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
> >>  /* Keep track of our dynamic hotplug state */
> >>  static enum cpuhp_state arm_spe_pmu_online;
> >>  
> >> +static u64 sys_pmsevfr_el1_mask[] = {
> >> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> >> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
> >> +		BIT_ULL(1),
> >> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> >> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
> >> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
> >> +};
> > 
> > Seems to me, the definitions for Aarch64 system registers should be
> > placed into the file 'arch/arm64/include/asm/sysreg.h'.  Like below
> > two macros:
> > 
> >   #define SYS_PMSEVFR_EL1_RES0_8_2		0x0000ffff00ff0f55UL
> >   #define SYS_PMSEVFR_EL1_RES0_8_3		...
> 
> I really think using GENMASK_ULL() to generate the mask is better than a definition
> with magic number. It is beneficial to be reviewed and extended later.

Understand.  Here I just want to remind, you could see the ARMv8's
system registers definition usually are placed into the global header
sysreg.h rather than define them in separate source files.

You could define the bit mask with GENMASK_ULL() for the two macros
in sysreg.h.

> > Let's wait for Will or Mark Rutland's comments for this, in case I
> > mislead for this.
> >> +
> >>  enum arm_spe_pmu_buf_fault_action {
> >>  	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
> >>  	SPE_PMU_BUF_FAULT_ACT_FATAL,
> >> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
> >>  	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
> >>  		return -ENOENT;
> >>  
> >> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
> >> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
> >>  		return -EOPNOTSUPP;
> >>  
> >>  	if (attr->exclude_idle)
> >> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
> >>  			fld, smp_processor_id());
> >>  		return;
> >>  	}
> >> +	spe_pmu->pmuver = fld;
> >>  
> >>  	/* Read PMBIDR first to determine whether or not we have access */
> >>  	reg = read_sysreg_s(SYS_PMBIDR_EL1);
> >> @@ -1027,8 +1037,8 @@ static void __arm_spe_pmu_dev_probe(void *info)
> >>  	}
> >>  
> >>  	dev_info(dev,
> >> -		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
> >> -		 cpumask_pr_args(&spe_pmu->supported_cpus),
> >> +		 "v%d probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
> > 
> > Let's output explict info, like:
> > 
> >   "probed for CPUs %*pbl [pmuver %d, max_record_sz %u, align %u, features 0x%llx]\n",
> > 
> 
> Agree, and i have a little question here:
> Currently, the of_compatible of SPE PMU is "arm,statistical-profiling-extension-v1", and
> the platform_device name is "arm,spe-v1". So this message looks weird when supporting
> ARMv8.3-SPE because the pmuver is 2.

I think here we need to distinguish two things: SPE (as an IP) and
ARMv8.2/ARMv8.3 (as CPU architectures).  From my understanding, now we
are working on SPE-v1, but it needs to support ARMv8 variants, e.g.
ARMv8.2 and ARMv8.3 with SVE extension.

I am not the best person to clarify the version number for SPE, if Arm
colleagues disagree with this, very welcome to correct me.

Also loop in Al for this.

> As the version of SPE can be probed by reading 'ID_AA64DFR0_EL1.PMSVer', can we remove
> the version hint in of_compatible and platform_device name?

No, for device tree, usually we need to keep back compability for the
DT binding, so we cannot remove compatible string.

Thanks,
Leo
Suzuki K Poulose July 29, 2020, 9:12 a.m. UTC | #4
On 07/24/2020 10:16 AM, Wei Li wrote:
> Armv8.3 extends the SPE by adding:
> - Alignment field in the Events packet, and filtering on this event
>    using PMSEVFR_EL1.
> - Support for the Scalable Vector Extension (SVE).
> 
> The main additions for SVE are:
> - Recording the vector length for SVE operations in the Operation Type
>    packet. It is not possible to filter on vector length.
> - Incomplete predicate and empty predicate fields in the Events packet,
>    and filtering on these events using PMSEVFR_EL1.
> 
> Update the check of pmsevfr for empty/partial predicated SVE and
> alignment event in kernel driver.
> 
> Signed-off-by: Wei Li <liwei391@huawei.com>
> ---
>   arch/arm64/include/asm/sysreg.h |  4 +++-
>   drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
>   2 files changed, 17 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
> index 463175f80341..be4c44ccdb56 100644
> --- a/arch/arm64/include/asm/sysreg.h
> +++ b/arch/arm64/include/asm/sysreg.h
> @@ -281,7 +281,6 @@
>   #define SYS_PMSFCR_EL1_ST_SHIFT		18
>   
>   #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
>   
>   #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>   #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
> @@ -769,6 +768,9 @@
>   #define ID_AA64DFR0_PMUVER_8_5		0x6
>   #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>   
> +#define ID_AA64DFR0_PMSVER_8_2		0x1
> +#define ID_AA64DFR0_PMSVER_8_3		0x2
> +
>   #define ID_DFR0_PERFMON_SHIFT		24
>   
>   #define ID_DFR0_PERFMON_8_1		0x4
> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> index e51ddb6d63ed..5ec7ee0c8fa1 100644
> --- a/drivers/perf/arm_spe_pmu.c
> +++ b/drivers/perf/arm_spe_pmu.c
> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>   	struct hlist_node			hotplug_node;
>   
>   	int					irq; /* PPI */
> -
> +	int					pmuver;
>   	u16					min_period;
>   	u16					counter_sz;
>   
> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>   /* Keep track of our dynamic hotplug state */
>   static enum cpuhp_state arm_spe_pmu_online;
>   
> +static u64 sys_pmsevfr_el1_mask[] = {
> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
> +		BIT_ULL(1),
> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
> +};
> +
>   enum arm_spe_pmu_buf_fault_action {
>   	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
>   	SPE_PMU_BUF_FAULT_ACT_FATAL,
> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
>   	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
>   		return -ENOENT;
>   
> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
>   		return -EOPNOTSUPP;
>   
>   	if (attr->exclude_idle)
> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
>   			fld, smp_processor_id());
>   		return;
>   	}
> +	spe_pmu->pmuver = fld;

How do we deal with cases where we have big.LITTLE system with differing
SPE versions ?

Cheers
Suzuki
Leo Yan July 30, 2020, 8:14 a.m. UTC | #5
Hi Suzuki,

On Wed, Jul 29, 2020 at 10:12:50AM +0100, Suzuki Kuruppassery Poulose wrote:
> On 07/24/2020 10:16 AM, Wei Li wrote:
> > Armv8.3 extends the SPE by adding:
> > - Alignment field in the Events packet, and filtering on this event
> >    using PMSEVFR_EL1.
> > - Support for the Scalable Vector Extension (SVE).
> > 
> > The main additions for SVE are:
> > - Recording the vector length for SVE operations in the Operation Type
> >    packet. It is not possible to filter on vector length.
> > - Incomplete predicate and empty predicate fields in the Events packet,
> >    and filtering on these events using PMSEVFR_EL1.
> > 
> > Update the check of pmsevfr for empty/partial predicated SVE and
> > alignment event in kernel driver.
> > 
> > Signed-off-by: Wei Li <liwei391@huawei.com>
> > ---
> >   arch/arm64/include/asm/sysreg.h |  4 +++-
> >   drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
> >   2 files changed, 17 insertions(+), 5 deletions(-)
> > 
> > diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
> > index 463175f80341..be4c44ccdb56 100644
> > --- a/arch/arm64/include/asm/sysreg.h
> > +++ b/arch/arm64/include/asm/sysreg.h
> > @@ -281,7 +281,6 @@
> >   #define SYS_PMSFCR_EL1_ST_SHIFT		18
> >   #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
> > -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
> >   #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
> >   #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
> > @@ -769,6 +768,9 @@
> >   #define ID_AA64DFR0_PMUVER_8_5		0x6
> >   #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
> > +#define ID_AA64DFR0_PMSVER_8_2		0x1
> > +#define ID_AA64DFR0_PMSVER_8_3		0x2
> > +
> >   #define ID_DFR0_PERFMON_SHIFT		24
> >   #define ID_DFR0_PERFMON_8_1		0x4
> > diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> > index e51ddb6d63ed..5ec7ee0c8fa1 100644
> > --- a/drivers/perf/arm_spe_pmu.c
> > +++ b/drivers/perf/arm_spe_pmu.c
> > @@ -54,7 +54,7 @@ struct arm_spe_pmu {
> >   	struct hlist_node			hotplug_node;
> >   	int					irq; /* PPI */
> > -
> > +	int					pmuver;
> >   	u16					min_period;
> >   	u16					counter_sz;
> > @@ -80,6 +80,15 @@ struct arm_spe_pmu {
> >   /* Keep track of our dynamic hotplug state */
> >   static enum cpuhp_state arm_spe_pmu_online;
> > +static u64 sys_pmsevfr_el1_mask[] = {
> > +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> > +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
> > +		BIT_ULL(1),
> > +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> > +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
> > +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
> > +};
> > +
> >   enum arm_spe_pmu_buf_fault_action {
> >   	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
> >   	SPE_PMU_BUF_FAULT_ACT_FATAL,
> > @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
> >   	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
> >   		return -ENOENT;
> > -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
> > +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
> >   		return -EOPNOTSUPP;
> >   	if (attr->exclude_idle)
> > @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
> >   			fld, smp_processor_id());
> >   		return;
> >   	}
> > +	spe_pmu->pmuver = fld;
> 
> How do we deal with cases where we have big.LITTLE system with differing
> SPE versions ?

Good point.

The first question we need to answer is: how to define SPE version?
From my understanding, if SPE uses the same sample specification and
the same packet format, then we should consider the SPE is the same
version cross CPUs.  So even some CPUs are ARMv8.2 and other CPUs are
ARMv8.3 variants, we still should take the SPE as the same version.

And when read the SPE driver in the file drivers/perf/arm_spe_pmu.c and
I concluded that so far the SPE perf driver is to only support SPE-v1
with single instance, it cannot support a complex usage case like
below:

  CPU0-3: ARMv8.2 architecture with SPE
  CPU4-7: ARMv8.3 architecture with SPE

For this case, if we take SPE as two different versions, let's say
SPE-8.2 and SPE-8.3, then should the SPE driver need to create multi
perf PMU events?  For example, we should create a perf PMU event
'arm_spe_8.2' and another PMU event 'arm_spe_8.3'.

Another option is we always take this as SPE-v1 and only create single
PMU event, just keep what's we are doing with the perf event
'arm_spe_0', but the driver needs to dynamically detect SPE PMU version
number in the function arm_spe_pmu_event_init(), and then based on
version number to select corresponding mask for PMSEVFR.

Thanks,
Leo

[1] https://lore.kernel.org/linux-arm-kernel/20200724071111.35593-1-liwei391@huawei.com/
Wei Li July 31, 2020, 12:18 p.m. UTC | #6
On 2020/7/30 16:14, Leo Yan wrote:
> Hi Suzuki,
> 
> On Wed, Jul 29, 2020 at 10:12:50AM +0100, Suzuki Kuruppassery Poulose wrote:
>> On 07/24/2020 10:16 AM, Wei Li wrote:
>>> Armv8.3 extends the SPE by adding:
>>> - Alignment field in the Events packet, and filtering on this event
>>>    using PMSEVFR_EL1.
>>> - Support for the Scalable Vector Extension (SVE).
>>>
>>> The main additions for SVE are:
>>> - Recording the vector length for SVE operations in the Operation Type
>>>    packet. It is not possible to filter on vector length.
>>> - Incomplete predicate and empty predicate fields in the Events packet,
>>>    and filtering on these events using PMSEVFR_EL1.
>>>
>>> Update the check of pmsevfr for empty/partial predicated SVE and
>>> alignment event in kernel driver.
>>>
>>> Signed-off-by: Wei Li <liwei391@huawei.com>
>>> ---
>>>   arch/arm64/include/asm/sysreg.h |  4 +++-
>>>   drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
>>>   2 files changed, 17 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
>>> index 463175f80341..be4c44ccdb56 100644
>>> --- a/arch/arm64/include/asm/sysreg.h
>>> +++ b/arch/arm64/include/asm/sysreg.h
>>> @@ -281,7 +281,6 @@
>>>   #define SYS_PMSFCR_EL1_ST_SHIFT		18
>>>   #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
>>> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
>>>   #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>>>   #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
>>> @@ -769,6 +768,9 @@
>>>   #define ID_AA64DFR0_PMUVER_8_5		0x6
>>>   #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>>> +#define ID_AA64DFR0_PMSVER_8_2		0x1
>>> +#define ID_AA64DFR0_PMSVER_8_3		0x2
>>> +
>>>   #define ID_DFR0_PERFMON_SHIFT		24
>>>   #define ID_DFR0_PERFMON_8_1		0x4
>>> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
>>> index e51ddb6d63ed..5ec7ee0c8fa1 100644
>>> --- a/drivers/perf/arm_spe_pmu.c
>>> +++ b/drivers/perf/arm_spe_pmu.c
>>> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>>>   	struct hlist_node			hotplug_node;
>>>   	int					irq; /* PPI */
>>> -
>>> +	int					pmuver;
>>>   	u16					min_period;
>>>   	u16					counter_sz;
>>> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>>>   /* Keep track of our dynamic hotplug state */
>>>   static enum cpuhp_state arm_spe_pmu_online;
>>> +static u64 sys_pmsevfr_el1_mask[] = {
>>> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>>> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
>>> +		BIT_ULL(1),
>>> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>>> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
>>> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
>>> +};
>>> +
>>>   enum arm_spe_pmu_buf_fault_action {
>>>   	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
>>>   	SPE_PMU_BUF_FAULT_ACT_FATAL,
>>> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
>>>   	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
>>>   		return -ENOENT;
>>> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
>>> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
>>>   		return -EOPNOTSUPP;
>>>   	if (attr->exclude_idle)
>>> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
>>>   			fld, smp_processor_id());
>>>   		return;
>>>   	}
>>> +	spe_pmu->pmuver = fld;
>>
>> How do we deal with cases where we have big.LITTLE system with differing
>> SPE versions ?
> 
> Good point.
> 
> The first question we need to answer is: how to define SPE version?
> From my understanding, if SPE uses the same sample specification and
> the same packet format, then we should consider the SPE is the same
> version cross CPUs.  So even some CPUs are ARMv8.2 and other CPUs are
> ARMv8.3 variants, we still should take the SPE as the same version.
> 
> And when read the SPE driver in the file drivers/perf/arm_spe_pmu.c and
> I concluded that so far the SPE perf driver is to only support SPE-v1
> with single instance, it cannot support a complex usage case like
> below:
> 
>   CPU0-3: ARMv8.2 architecture with SPE
>   CPU4-7: ARMv8.3 architecture with SPE
> 
> For this case, if we take SPE as two different versions, let's say
> SPE-8.2 and SPE-8.3, then should the SPE driver need to create multi
> perf PMU events?  For example, we should create a perf PMU event
> 'arm_spe_8.2' and another PMU event 'arm_spe_8.3'.

As we have supported SPE v2 (ARMv8.3-SPE) now, if we add the new
of_device_id: "arm,statistical-profiling-extension-v2" and the new
platform_device_id: "arm,spe-v2", we may really support two instance now.
Even two different versions of SPE pmus which work on different range of
cores with the same PPI. Their functional scopes are the same as the PPI partitions.

> Another option is we always take this as SPE-v1 and only create single
> PMU event, just keep what's we are doing with the perf event
> 'arm_spe_0', but the driver needs to dynamically detect SPE PMU version
> number in the function arm_spe_pmu_event_init(), and then based on
> version number to select corresponding mask for PMSEVFR.

Thus, the driver will service two devices, and will also register two PMUs 'arm_spe_0' and
'arm_spe_1', so i think there is no conflict here.

Back to Suzuki's question, refer to the ACPI parsing code for SPE device, function
arm_spe_acpi_register_device(), there is a check of hetero_id for all cores.
It seems that we only support homogeneous ACPI/SPE machines, but i can't find the similar
check in OF/SPE parsing code.

> Thanks,
> Leo
> 
> [1] https://lore.kernel.org/linux-arm-kernel/20200724071111.35593-1-liwei391@huawei.com/
> 


Thanks,
Wei
Suzuki K Poulose July 31, 2020, 2:01 p.m. UTC | #7
On 07/31/2020 01:18 PM, liwei (GF) wrote:
> 
> 
> On 2020/7/30 16:14, Leo Yan wrote:
>> Hi Suzuki,
>>
>> On Wed, Jul 29, 2020 at 10:12:50AM +0100, Suzuki Kuruppassery Poulose wrote:
>>> On 07/24/2020 10:16 AM, Wei Li wrote:
>>>> Armv8.3 extends the SPE by adding:
>>>> - Alignment field in the Events packet, and filtering on this event
>>>>     using PMSEVFR_EL1.
>>>> - Support for the Scalable Vector Extension (SVE).
>>>>
>>>> The main additions for SVE are:
>>>> - Recording the vector length for SVE operations in the Operation Type
>>>>     packet. It is not possible to filter on vector length.
>>>> - Incomplete predicate and empty predicate fields in the Events packet,
>>>>     and filtering on these events using PMSEVFR_EL1.
>>>>
>>>> Update the check of pmsevfr for empty/partial predicated SVE and
>>>> alignment event in kernel driver.
>>>>
>>>> Signed-off-by: Wei Li <liwei391@huawei.com>
>>>> ---
>>>>    arch/arm64/include/asm/sysreg.h |  4 +++-
>>>>    drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
>>>>    2 files changed, 17 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
>>>> index 463175f80341..be4c44ccdb56 100644
>>>> --- a/arch/arm64/include/asm/sysreg.h
>>>> +++ b/arch/arm64/include/asm/sysreg.h
>>>> @@ -281,7 +281,6 @@
>>>>    #define SYS_PMSFCR_EL1_ST_SHIFT		18
>>>>    #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
>>>> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
>>>>    #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>>>>    #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
>>>> @@ -769,6 +768,9 @@
>>>>    #define ID_AA64DFR0_PMUVER_8_5		0x6
>>>>    #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>>>> +#define ID_AA64DFR0_PMSVER_8_2		0x1
>>>> +#define ID_AA64DFR0_PMSVER_8_3		0x2
>>>> +
>>>>    #define ID_DFR0_PERFMON_SHIFT		24
>>>>    #define ID_DFR0_PERFMON_8_1		0x4
>>>> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
>>>> index e51ddb6d63ed..5ec7ee0c8fa1 100644
>>>> --- a/drivers/perf/arm_spe_pmu.c
>>>> +++ b/drivers/perf/arm_spe_pmu.c
>>>> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>>>>    	struct hlist_node			hotplug_node;
>>>>    	int					irq; /* PPI */
>>>> -
>>>> +	int					pmuver;
>>>>    	u16					min_period;
>>>>    	u16					counter_sz;
>>>> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>>>>    /* Keep track of our dynamic hotplug state */
>>>>    static enum cpuhp_state arm_spe_pmu_online;
>>>> +static u64 sys_pmsevfr_el1_mask[] = {
>>>> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>>>> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
>>>> +		BIT_ULL(1),
>>>> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>>>> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
>>>> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
>>>> +};
>>>> +
>>>>    enum arm_spe_pmu_buf_fault_action {
>>>>    	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
>>>>    	SPE_PMU_BUF_FAULT_ACT_FATAL,
>>>> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
>>>>    	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
>>>>    		return -ENOENT;
>>>> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
>>>> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
>>>>    		return -EOPNOTSUPP;
>>>>    	if (attr->exclude_idle)
>>>> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
>>>>    			fld, smp_processor_id());
>>>>    		return;
>>>>    	}
>>>> +	spe_pmu->pmuver = fld;
>>>
>>> How do we deal with cases where we have big.LITTLE system with differing
>>> SPE versions ?
>>
>> Good point.
>>
>> The first question we need to answer is: how to define SPE version?
>>  From my understanding, if SPE uses the same sample specification and
>> the same packet format, then we should consider the SPE is the same
>> version cross CPUs.  So even some CPUs are ARMv8.2 and other CPUs are
>> ARMv8.3 variants, we still should take the SPE as the same version.
>>
>> And when read the SPE driver in the file drivers/perf/arm_spe_pmu.c and
>> I concluded that so far the SPE perf driver is to only support SPE-v1
>> with single instance, it cannot support a complex usage case like
>> below:
>>
>>    CPU0-3: ARMv8.2 architecture with SPE
>>    CPU4-7: ARMv8.3 architecture with SPE
>>
>> For this case, if we take SPE as two different versions, let's say
>> SPE-8.2 and SPE-8.3, then should the SPE driver need to create multi
>> perf PMU events?  For example, we should create a perf PMU event
>> 'arm_spe_8.2' and another PMU event 'arm_spe_8.3'.
> 
> As we have supported SPE v2 (ARMv8.3-SPE) now, if we add the new
> of_device_id: "arm,statistical-profiling-extension-v2" and the new
> platform_device_id: "arm,spe-v2", we may really support two instance now.
> Even two different versions of SPE pmus which work on different range of
> cores with the same PPI. Their functional scopes are the same as the PPI partitions.
> 
>> Another option is we always take this as SPE-v1 and only create single
>> PMU event, just keep what's we are doing with the perf event
>> 'arm_spe_0', but the driver needs to dynamically detect SPE PMU version
>> number in the function arm_spe_pmu_event_init(), and then based on
>> version number to select corresponding mask for PMSEVFR.
> 
> Thus, the driver will service two devices, and will also register two PMUs 'arm_spe_0' and
> 'arm_spe_1', so i think there is no conflict here.

Or the other option is to strictly support only one version - the lowest
version supported by the system, which is compatible with all the
others. This could be achieved by using the sanitised feature value of
the SPE version.

Will,

What do you prefer ?

> 
> Back to Suzuki's question, refer to the ACPI parsing code for SPE device, function
> arm_spe_acpi_register_device(), there is a check of hetero_id for all cores.
> It seems that we only support homogeneous ACPI/SPE machines, but i can't find the similar
> check in OF/SPE parsing code.

I think the driver assumes that the system supports uniform version
of the SPE (which was valid when the driver was written).

Cheers
Suzuki
Will Deacon Sept. 7, 2020, 12:51 p.m. UTC | #8
On Fri, Jul 24, 2020 at 05:16:04PM +0800, Wei Li wrote:
> Armv8.3 extends the SPE by adding:
> - Alignment field in the Events packet, and filtering on this event
>   using PMSEVFR_EL1.
> - Support for the Scalable Vector Extension (SVE).
> 
> The main additions for SVE are:
> - Recording the vector length for SVE operations in the Operation Type
>   packet. It is not possible to filter on vector length.
> - Incomplete predicate and empty predicate fields in the Events packet,
>   and filtering on these events using PMSEVFR_EL1.
> 
> Update the check of pmsevfr for empty/partial predicated SVE and
> alignment event in kernel driver.
> 
> Signed-off-by: Wei Li <liwei391@huawei.com>
> ---
>  arch/arm64/include/asm/sysreg.h |  4 +++-
>  drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
>  2 files changed, 17 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
> index 463175f80341..be4c44ccdb56 100644
> --- a/arch/arm64/include/asm/sysreg.h
> +++ b/arch/arm64/include/asm/sysreg.h
> @@ -281,7 +281,6 @@
>  #define SYS_PMSFCR_EL1_ST_SHIFT		18
>  
>  #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL

I think we can just update this mask unconditionally to allow the new bits.

>  #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>  #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
> @@ -769,6 +768,9 @@
>  #define ID_AA64DFR0_PMUVER_8_5		0x6
>  #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>  
> +#define ID_AA64DFR0_PMSVER_8_2		0x1
> +#define ID_AA64DFR0_PMSVER_8_3		0x2
> +
>  #define ID_DFR0_PERFMON_SHIFT		24
>  
>  #define ID_DFR0_PERFMON_8_1		0x4
> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
> index e51ddb6d63ed..5ec7ee0c8fa1 100644
> --- a/drivers/perf/arm_spe_pmu.c
> +++ b/drivers/perf/arm_spe_pmu.c
> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>  	struct hlist_node			hotplug_node;
>  
>  	int					irq; /* PPI */
> -
> +	int					pmuver;

nit: please call this "pmsver" to align with the architecture (where
"pmuver" means something else).

>  	u16					min_period;
>  	u16					counter_sz;
>  
> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>  /* Keep track of our dynamic hotplug state */
>  static enum cpuhp_state arm_spe_pmu_online;
>  
> +static u64 sys_pmsevfr_el1_mask[] = {
> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
> +		BIT_ULL(1),
> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
> +};

As I said above, you can drop this and just update the #define.

> +
>  enum arm_spe_pmu_buf_fault_action {
>  	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
>  	SPE_PMU_BUF_FAULT_ACT_FATAL,
> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
>  	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
>  		return -ENOENT;
>  
> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
>  		return -EOPNOTSUPP;

Same here.

>  
>  	if (attr->exclude_idle)
> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
>  			fld, smp_processor_id());
>  		return;
>  	}
> +	spe_pmu->pmuver = fld;
>  
>  	/* Read PMBIDR first to determine whether or not we have access */
>  	reg = read_sysreg_s(SYS_PMBIDR_EL1);
> @@ -1027,8 +1037,8 @@ static void __arm_spe_pmu_dev_probe(void *info)
>  	}
>  
>  	dev_info(dev,
> -		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
> -		 cpumask_pr_args(&spe_pmu->supported_cpus),
> +		 "v%d probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
> +		 spe_pmu->pmuver, cpumask_pr_args(&spe_pmu->supported_cpus),

There's no need for this. If userspace finds this information useful, then
we should expose it in sysfs, like we do for other PMU paramaters. If
userspace doesn't find it useful, then there's no need to expose it at all.

So I would suggest adding something like SPE_PMU_CAP_PMSVER and exposing the
field on a per-SPE-PMU basis in sysfs.

big.LITTLE should work as before, where we expose a completely separate PMU
instance for each CPU type.

Will
>
Wei Li Sept. 29, 2020, 8:17 a.m. UTC | #9
Hi Will,

On 2020/9/7 20:51, Will Deacon wrote:
> On Fri, Jul 24, 2020 at 05:16:04PM +0800, Wei Li wrote:
>> Armv8.3 extends the SPE by adding:
>> - Alignment field in the Events packet, and filtering on this event
>>   using PMSEVFR_EL1.
>> - Support for the Scalable Vector Extension (SVE).
>>
>> The main additions for SVE are:
>> - Recording the vector length for SVE operations in the Operation Type
>>   packet. It is not possible to filter on vector length.
>> - Incomplete predicate and empty predicate fields in the Events packet,
>>   and filtering on these events using PMSEVFR_EL1.
>>
>> Update the check of pmsevfr for empty/partial predicated SVE and
>> alignment event in kernel driver.
>>
>> Signed-off-by: Wei Li <liwei391@huawei.com>
>> ---
>>  arch/arm64/include/asm/sysreg.h |  4 +++-
>>  drivers/perf/arm_spe_pmu.c      | 18 ++++++++++++++----
>>  2 files changed, 17 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
>> index 463175f80341..be4c44ccdb56 100644
>> --- a/arch/arm64/include/asm/sysreg.h
>> +++ b/arch/arm64/include/asm/sysreg.h
>> @@ -281,7 +281,6 @@
>>  #define SYS_PMSFCR_EL1_ST_SHIFT		18
>>  
>>  #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
>> -#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
> 
> I think we can just update this mask unconditionally to allow the new bits.
> 
>>  #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
>>  #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
>> @@ -769,6 +768,9 @@
>>  #define ID_AA64DFR0_PMUVER_8_5		0x6
>>  #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
>>  
>> +#define ID_AA64DFR0_PMSVER_8_2		0x1
>> +#define ID_AA64DFR0_PMSVER_8_3		0x2
>> +
>>  #define ID_DFR0_PERFMON_SHIFT		24
>>  
>>  #define ID_DFR0_PERFMON_8_1		0x4
>> diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
>> index e51ddb6d63ed..5ec7ee0c8fa1 100644
>> --- a/drivers/perf/arm_spe_pmu.c
>> +++ b/drivers/perf/arm_spe_pmu.c
>> @@ -54,7 +54,7 @@ struct arm_spe_pmu {
>>  	struct hlist_node			hotplug_node;
>>  
>>  	int					irq; /* PPI */
>> -
>> +	int					pmuver;
> 
> nit: please call this "pmsver" to align with the architecture (where
> "pmuver" means something else).

OK, i will rename it in v2.

>>  	u16					min_period;
>>  	u16					counter_sz;
>>  
>> @@ -80,6 +80,15 @@ struct arm_spe_pmu {
>>  /* Keep track of our dynamic hotplug state */
>>  static enum cpuhp_state arm_spe_pmu_online;
>>  
>> +static u64 sys_pmsevfr_el1_mask[] = {
>> +	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>> +		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
>> +		BIT_ULL(1),
>> +	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
>> +		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
>> +		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
>> +};
> 
> As I said above, you can drop this and just update the #define.
> 
>> +
>>  enum arm_spe_pmu_buf_fault_action {
>>  	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
>>  	SPE_PMU_BUF_FAULT_ACT_FATAL,
>> @@ -670,7 +679,7 @@ static int arm_spe_pmu_event_init(struct perf_event *event)
>>  	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
>>  		return -ENOENT;
>>  
>> -	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
>> +	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
>>  		return -EOPNOTSUPP;
> 
> Same here.

What if we use these new bits on the system which just support ARMv8.2-SPE?
It will return success and never work, i don't think that is suitable.

>>  
>>  	if (attr->exclude_idle)
>> @@ -937,6 +946,7 @@ static void __arm_spe_pmu_dev_probe(void *info)
>>  			fld, smp_processor_id());
>>  		return;
>>  	}
>> +	spe_pmu->pmuver = fld;
>>  
>>  	/* Read PMBIDR first to determine whether or not we have access */
>>  	reg = read_sysreg_s(SYS_PMBIDR_EL1);
>> @@ -1027,8 +1037,8 @@ static void __arm_spe_pmu_dev_probe(void *info)
>>  	}
>>  
>>  	dev_info(dev,
>> -		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
>> -		 cpumask_pr_args(&spe_pmu->supported_cpus),
>> +		 "v%d probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
>> +		 spe_pmu->pmuver, cpumask_pr_args(&spe_pmu->supported_cpus),
> 
> There's no need for this. If userspace finds this information useful, then
> we should expose it in sysfs, like we do for other PMU paramaters. If
> userspace doesn't find it useful, then there's no need to expose it at all.
> 
> So I would suggest adding something like SPE_PMU_CAP_PMSVER and exposing the
> field on a per-SPE-PMU basis in sysfs.

We may need this as then we can know which events are supported. It is meaningful to testcases.
So i will expose it as cap attribute in v2.

> big.LITTLE should work as before, where we expose a completely separate PMU
> instance for each CPU type.
> 

The of_compatible of SPE PMU is "arm,statistical-profiling-extension-v1", and the platform_device
name is "arm,spe-v1". Should we add a "v2" entry for ARMv8.3-SPE or not?

Thanks for your time.

Best regards,
Wei
diff mbox series

Patch

diff --git a/arch/arm64/include/asm/sysreg.h b/arch/arm64/include/asm/sysreg.h
index 463175f80341..be4c44ccdb56 100644
--- a/arch/arm64/include/asm/sysreg.h
+++ b/arch/arm64/include/asm/sysreg.h
@@ -281,7 +281,6 @@ 
 #define SYS_PMSFCR_EL1_ST_SHIFT		18
 
 #define SYS_PMSEVFR_EL1			sys_reg(3, 0, 9, 9, 5)
-#define SYS_PMSEVFR_EL1_RES0		0x0000ffff00ff0f55UL
 
 #define SYS_PMSLATFR_EL1		sys_reg(3, 0, 9, 9, 6)
 #define SYS_PMSLATFR_EL1_MINLAT_SHIFT	0
@@ -769,6 +768,9 @@ 
 #define ID_AA64DFR0_PMUVER_8_5		0x6
 #define ID_AA64DFR0_PMUVER_IMP_DEF	0xf
 
+#define ID_AA64DFR0_PMSVER_8_2		0x1
+#define ID_AA64DFR0_PMSVER_8_3		0x2
+
 #define ID_DFR0_PERFMON_SHIFT		24
 
 #define ID_DFR0_PERFMON_8_1		0x4
diff --git a/drivers/perf/arm_spe_pmu.c b/drivers/perf/arm_spe_pmu.c
index e51ddb6d63ed..5ec7ee0c8fa1 100644
--- a/drivers/perf/arm_spe_pmu.c
+++ b/drivers/perf/arm_spe_pmu.c
@@ -54,7 +54,7 @@  struct arm_spe_pmu {
 	struct hlist_node			hotplug_node;
 
 	int					irq; /* PPI */
-
+	int					pmuver;
 	u16					min_period;
 	u16					counter_sz;
 
@@ -80,6 +80,15 @@  struct arm_spe_pmu {
 /* Keep track of our dynamic hotplug state */
 static enum cpuhp_state arm_spe_pmu_online;
 
+static u64 sys_pmsevfr_el1_mask[] = {
+	[ID_AA64DFR0_PMSVER_8_2] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
+		GENMASK_ULL(15, 12) | BIT_ULL(7) | BIT_ULL(5) | BIT_ULL(3) |
+		BIT_ULL(1),
+	[ID_AA64DFR0_PMSVER_8_3] = GENMASK_ULL(63, 48) | GENMASK_ULL(31, 24) |
+		GENMASK_ULL(18, 17) | GENMASK_ULL(15, 11) | BIT_ULL(7) |
+		BIT_ULL(5) | BIT_ULL(3) | BIT_ULL(1),
+};
+
 enum arm_spe_pmu_buf_fault_action {
 	SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
 	SPE_PMU_BUF_FAULT_ACT_FATAL,
@@ -670,7 +679,7 @@  static int arm_spe_pmu_event_init(struct perf_event *event)
 	    !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
 		return -ENOENT;
 
-	if (arm_spe_event_to_pmsevfr(event) & SYS_PMSEVFR_EL1_RES0)
+	if (arm_spe_event_to_pmsevfr(event) & ~sys_pmsevfr_el1_mask[spe_pmu->pmuver])
 		return -EOPNOTSUPP;
 
 	if (attr->exclude_idle)
@@ -937,6 +946,7 @@  static void __arm_spe_pmu_dev_probe(void *info)
 			fld, smp_processor_id());
 		return;
 	}
+	spe_pmu->pmuver = fld;
 
 	/* Read PMBIDR first to determine whether or not we have access */
 	reg = read_sysreg_s(SYS_PMBIDR_EL1);
@@ -1027,8 +1037,8 @@  static void __arm_spe_pmu_dev_probe(void *info)
 	}
 
 	dev_info(dev,
-		 "probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
-		 cpumask_pr_args(&spe_pmu->supported_cpus),
+		 "v%d probed for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
+		 spe_pmu->pmuver, cpumask_pr_args(&spe_pmu->supported_cpus),
 		 spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
 
 	spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;