[PATCH v6 0/4] perf: arm_cspmu: ampere: Add support for Ampere SoC PMUs

[PATCH v6 0/4] perf: arm_cspmu: ampere: Add support for Ampere SoC PMUs

[Ilkka Koskinen]

Changes since v5:
    * Implemented the needed parts for vendor registration API
    * Rebased on top of Besar's patch 

      [PATCH v5] perf: arm_cspmu: Separate Arm and vendor module
      https://lore.kernel.org/all/20230705104745.52255-1-bwicaksono@nvidia.com/

    * v5: https://lore.kernel.org/all/20230714010141.824226-1-ilkka@os.amperecomputing.com/

Changes since v4:
    * "Support implementation specific filters" patch:
        - Added comment about filter and impdef registers and reference
          to the Coresight PMU specification to the commit message

    * "Add support for Ampere SoC PMU" patch:
	- Fixed the documentation and added more comments
        - Changed the incrementing PMU index number to idr_alloc()
	  (Needs a impdef release hook patch to release unused index)
	- Fixed style in init_ops() to more reasonable
	- Moved bank parameter to config1

Changes since v3:
    * use_64b_counter_reg => has_atomic_dword (patch 1/4)
    * Removed the unnecessary hook for group validation (patch 3/4)
    * Added group config validation to ampere_cspmu_validate_event() (patch 4/4)
    * Rebased the patchset

Changes since v2:
    * Changed to use supports_64bits_atomics() and replaced the split writes
      with lo_hi_writeq()
    * Added implementation specific group validation to patch 3
    * Dropped shared interrupt patch
    * Removed unnecessary filter_enable parameter from ampere module
    * Added group validation to ampere module

Changes since v1:
    * Rather than creating a completely new driver, implemented as a submodule
      of Arm CoreSight PMU driver
      * Fixed shared filter handling

Ilkka Koskinen (4):
  perf: arm_cspmu: Split 64-bit write to 32-bit writes
  perf: arm_cspmu: Support implementation specific filters
  perf: arm_cspmu: Support implementation specific validation
  perf: arm_cspmu: ampere_cspmu: Add support for Ampere SoC PMU

 .../admin-guide/perf/ampere_cspmu.rst         |  29 ++
 drivers/perf/arm_cspmu/Kconfig                |  10 +
 drivers/perf/arm_cspmu/Makefile               |   2 +
 drivers/perf/arm_cspmu/ampere_cspmu.c         | 271 ++++++++++++++++++
 drivers/perf/arm_cspmu/arm_cspmu.c            |  33 ++-
 drivers/perf/arm_cspmu/arm_cspmu.h            |   7 +
 6 files changed, 346 insertions(+), 6 deletions(-)
 create mode 100644 Documentation/admin-guide/perf/ampere_cspmu.rst
 create mode 100644 drivers/perf/arm_cspmu/ampere_cspmu.c

-- 
2.41.0

[PATCH v6 1/4] perf: arm_cspmu: Split 64-bit write to 32-bit writes

[Ilkka Koskinen]

Split the 64-bit register accesses if 64-bit access is not supported
by the PMU.

Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Reviewed-by: Besar Wicaksono <bwicaksono@nvidia.com>
---
 drivers/perf/arm_cspmu/arm_cspmu.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
index 04be94b4aa48..6387cbad7a7d 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.c
+++ b/drivers/perf/arm_cspmu/arm_cspmu.c
@@ -715,7 +715,10 @@ static void arm_cspmu_write_counter(struct perf_event *event, u64 val)
 	if (use_64b_counter_reg(cspmu)) {
 		offset = counter_offset(sizeof(u64), event->hw.idx);
 
-		writeq(val, cspmu->base1 + offset);
+		if (cspmu->has_atomic_dword)
+			writeq(val, cspmu->base1 + offset);
+		else
+			lo_hi_writeq(val, cspmu->base1 + offset);
 	} else {
 		offset = counter_offset(sizeof(u32), event->hw.idx);
 
-- 
2.41.0

Re: [PATCH v6 1/4] perf: arm_cspmu: Split 64-bit write to 32-bit writes

[Suzuki K Poulose]

On 15/08/2023 07:35, Ilkka Koskinen wrote:
> Split the 64-bit register accesses if 64-bit access is not supported
> by the PMU.
> 
> Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
> Reviewed-by: Besar Wicaksono <bwicaksono@nvidia.com>

Do we need a Fixes tag ?

With that:

Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>

Suzuki

> ---
>   drivers/perf/arm_cspmu/arm_cspmu.c | 5 ++++-
>   1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
> index 04be94b4aa48..6387cbad7a7d 100644
> --- a/drivers/perf/arm_cspmu/arm_cspmu.c
> +++ b/drivers/perf/arm_cspmu/arm_cspmu.c
> @@ -715,7 +715,10 @@ static void arm_cspmu_write_counter(struct perf_event *event, u64 val)
>   	if (use_64b_counter_reg(cspmu)) {
>   		offset = counter_offset(sizeof(u64), event->hw.idx);
>   
> -		writeq(val, cspmu->base1 + offset);
> +		if (cspmu->has_atomic_dword)
> +			writeq(val, cspmu->base1 + offset);
> +		else
> +			lo_hi_writeq(val, cspmu->base1 + offset);


>   	} else {
>   		offset = counter_offset(sizeof(u32), event->hw.idx);
>   

Re: [PATCH v6 1/4] perf: arm_cspmu: Split 64-bit write to 32-bit writes

[Ilkka Koskinen]

Hi Suzuki,

On Tue, 15 Aug 2023, Suzuki K Poulose wrote:
> On 15/08/2023 07:35, Ilkka Koskinen wrote:
>> Split the 64-bit register accesses if 64-bit access is not supported
>> by the PMU.
>> 
>> Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
>> Reviewed-by: Besar Wicaksono <bwicaksono@nvidia.com>
>
> Do we need a Fixes tag ?

I believe, NVIDIA's PMU supports 64-bit access while Ampere's one doesn't 
and since this patchset enables support for the latter one, it doesn't 
seem like we need a Fixes tag here.

Cheers, Ilkka

>
> With that:
>
> Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
>
> Suzuki
>
>> ---
>>   drivers/perf/arm_cspmu/arm_cspmu.c | 5 ++++-
>>   1 file changed, 4 insertions(+), 1 deletion(-)
>> 
>> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c 
>> b/drivers/perf/arm_cspmu/arm_cspmu.c
>> index 04be94b4aa48..6387cbad7a7d 100644
>> --- a/drivers/perf/arm_cspmu/arm_cspmu.c
>> +++ b/drivers/perf/arm_cspmu/arm_cspmu.c
>> @@ -715,7 +715,10 @@ static void arm_cspmu_write_counter(struct perf_event 
>> *event, u64 val)
>>   	if (use_64b_counter_reg(cspmu)) {
>>   		offset = counter_offset(sizeof(u64), event->hw.idx);
>>   -		writeq(val, cspmu->base1 + offset);
>> +		if (cspmu->has_atomic_dword)
>> +			writeq(val, cspmu->base1 + offset);
>> +		else
>> +			lo_hi_writeq(val, cspmu->base1 + offset);
>
>
>>   	} else {
>>   		offset = counter_offset(sizeof(u32), event->hw.idx);
>> 
>
>

Re: [PATCH v6 1/4] perf: arm_cspmu: Split 64-bit write to 32-bit writes

[Suzuki K Poulose]

On 15/08/2023 21:46, Ilkka Koskinen wrote:
> 
> Hi Suzuki,
> 
> On Tue, 15 Aug 2023, Suzuki K Poulose wrote:
>> On 15/08/2023 07:35, Ilkka Koskinen wrote:
>>> Split the 64-bit register accesses if 64-bit access is not supported
>>> by the PMU.
>>>
>>> Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
>>> Reviewed-by: Besar Wicaksono <bwicaksono@nvidia.com>
>>
>> Do we need a Fixes tag ?
> 
> I believe, NVIDIA's PMU supports 64-bit access while Ampere's one 
> doesn't and since this patchset enables support for the latter one, it 
> doesn't seem like we need a Fixes tag here.

Ok, makes sense.

Suzuki


> 
> Cheers, Ilkka
> 
>>
>> With that:
>>
>> Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
>>
>> Suzuki
>>
>>> ---
>>>   drivers/perf/arm_cspmu/arm_cspmu.c | 5 ++++-
>>>   1 file changed, 4 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c 
>>> b/drivers/perf/arm_cspmu/arm_cspmu.c
>>> index 04be94b4aa48..6387cbad7a7d 100644
>>> --- a/drivers/perf/arm_cspmu/arm_cspmu.c
>>> +++ b/drivers/perf/arm_cspmu/arm_cspmu.c
>>> @@ -715,7 +715,10 @@ static void arm_cspmu_write_counter(struct 
>>> perf_event *event, u64 val)
>>>       if (use_64b_counter_reg(cspmu)) {
>>>           offset = counter_offset(sizeof(u64), event->hw.idx);
>>>   -        writeq(val, cspmu->base1 + offset);
>>> +        if (cspmu->has_atomic_dword)
>>> +            writeq(val, cspmu->base1 + offset);
>>> +        else
>>> +            lo_hi_writeq(val, cspmu->base1 + offset);
>>
>>
>>>       } else {
>>>           offset = counter_offset(sizeof(u32), event->hw.idx);
>>>
>>
>>

[PATCH v6 2/4] perf: arm_cspmu: Support implementation specific filters

[Ilkka Koskinen]

ARM Coresight PMU architecture specification [1] defines PMEVTYPER and
PMEVFILT* registers as optional in Chapter 2.1. Moreover, implementers may
choose to use PMIMPDEF* registers (offset: 0xD80-> 0xDFF) to filter the
events. Add support for those by adding implementation specific filter
callback function.

[1] https://developer.arm.com/documentation/ihi0091/latest

Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Reviewed-by: Besar Wicaksono <bwicaksono@nvidia.com>
---
 drivers/perf/arm_cspmu/arm_cspmu.c | 12 ++++++++----
 drivers/perf/arm_cspmu/arm_cspmu.h |  3 +++
 2 files changed, 11 insertions(+), 4 deletions(-)

diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
index 6387cbad7a7d..94f6856ec786 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.c
+++ b/drivers/perf/arm_cspmu/arm_cspmu.c
@@ -116,6 +116,9 @@ static unsigned long arm_cspmu_cpuhp_state;
 
 static DEFINE_MUTEX(arm_cspmu_lock);
 
+static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
+				    struct hw_perf_event *hwc, u32 filter);
+
 static struct acpi_apmt_node *arm_cspmu_apmt_node(struct device *dev)
 {
 	return *(struct acpi_apmt_node **)dev_get_platdata(dev);
@@ -450,6 +453,7 @@ static int arm_cspmu_init_impl_ops(struct arm_cspmu *cspmu)
 	CHECK_DEFAULT_IMPL_OPS(impl_ops, event_type);
 	CHECK_DEFAULT_IMPL_OPS(impl_ops, event_filter);
 	CHECK_DEFAULT_IMPL_OPS(impl_ops, event_attr_is_visible);
+	CHECK_DEFAULT_IMPL_OPS(impl_ops, set_ev_filter);
 
 	return 0;
 }
@@ -811,9 +815,9 @@ static inline void arm_cspmu_set_event(struct arm_cspmu *cspmu,
 	writel(hwc->config, cspmu->base0 + offset);
 }
 
-static inline void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
-					   struct hw_perf_event *hwc,
-					   u32 filter)
+static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
+					struct hw_perf_event *hwc,
+					u32 filter)
 {
 	u32 offset = PMEVFILTR + (4 * hwc->idx);
 
@@ -845,7 +849,7 @@ static void arm_cspmu_start(struct perf_event *event, int pmu_flags)
 		arm_cspmu_set_cc_filter(cspmu, filter);
 	} else {
 		arm_cspmu_set_event(cspmu, hwc);
-		arm_cspmu_set_ev_filter(cspmu, hwc, filter);
+		cspmu->impl.ops.set_ev_filter(cspmu, hwc, filter);
 	}
 
 	hwc->state = 0;
diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h b/drivers/perf/arm_cspmu/arm_cspmu.h
index e5c6dff2ce7f..274ca3d10578 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.h
+++ b/drivers/perf/arm_cspmu/arm_cspmu.h
@@ -104,6 +104,9 @@ struct arm_cspmu_impl_ops {
 	u32 (*event_type)(const struct perf_event *event);
 	/* Decode filter value from configs */
 	u32 (*event_filter)(const struct perf_event *event);
+	/* Set event filter */
+	void (*set_ev_filter)(struct arm_cspmu *cspmu,
+			      struct hw_perf_event *hwc, u32 filter);
 	/* Hide/show unsupported events */
 	umode_t (*event_attr_is_visible)(struct kobject *kobj,
 					 struct attribute *attr, int unused);
-- 
2.41.0

Re: [PATCH v6 2/4] perf: arm_cspmu: Support implementation specific filters

[Suzuki K Poulose]

On 15/08/2023 07:35, Ilkka Koskinen wrote:
> ARM Coresight PMU architecture specification [1] defines PMEVTYPER and
> PMEVFILT* registers as optional in Chapter 2.1. Moreover, implementers may
> choose to use PMIMPDEF* registers (offset: 0xD80-> 0xDFF) to filter the
> events. Add support for those by adding implementation specific filter
> callback function.
> 
> [1] https://developer.arm.com/documentation/ihi0091/latest
> 
> Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
> Reviewed-by: Besar Wicaksono <bwicaksono@nvidia.com>

Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>

> ---
>   drivers/perf/arm_cspmu/arm_cspmu.c | 12 ++++++++----
>   drivers/perf/arm_cspmu/arm_cspmu.h |  3 +++
>   2 files changed, 11 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
> index 6387cbad7a7d..94f6856ec786 100644
> --- a/drivers/perf/arm_cspmu/arm_cspmu.c
> +++ b/drivers/perf/arm_cspmu/arm_cspmu.c
> @@ -116,6 +116,9 @@ static unsigned long arm_cspmu_cpuhp_state;
>   
>   static DEFINE_MUTEX(arm_cspmu_lock);
>   
> +static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
> +				    struct hw_perf_event *hwc, u32 filter);
> +
>   static struct acpi_apmt_node *arm_cspmu_apmt_node(struct device *dev)
>   {
>   	return *(struct acpi_apmt_node **)dev_get_platdata(dev);
> @@ -450,6 +453,7 @@ static int arm_cspmu_init_impl_ops(struct arm_cspmu *cspmu)
>   	CHECK_DEFAULT_IMPL_OPS(impl_ops, event_type);
>   	CHECK_DEFAULT_IMPL_OPS(impl_ops, event_filter);
>   	CHECK_DEFAULT_IMPL_OPS(impl_ops, event_attr_is_visible);
> +	CHECK_DEFAULT_IMPL_OPS(impl_ops, set_ev_filter);
>   
>   	return 0;
>   }
> @@ -811,9 +815,9 @@ static inline void arm_cspmu_set_event(struct arm_cspmu *cspmu,
>   	writel(hwc->config, cspmu->base0 + offset);
>   }
>   
> -static inline void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
> -					   struct hw_perf_event *hwc,
> -					   u32 filter)
> +static void arm_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
> +					struct hw_perf_event *hwc,
> +					u32 filter)
>   {
>   	u32 offset = PMEVFILTR + (4 * hwc->idx);
>   
> @@ -845,7 +849,7 @@ static void arm_cspmu_start(struct perf_event *event, int pmu_flags)
>   		arm_cspmu_set_cc_filter(cspmu, filter);
>   	} else {
>   		arm_cspmu_set_event(cspmu, hwc);
> -		arm_cspmu_set_ev_filter(cspmu, hwc, filter);
> +		cspmu->impl.ops.set_ev_filter(cspmu, hwc, filter);
>   	}
>   
>   	hwc->state = 0;
> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h b/drivers/perf/arm_cspmu/arm_cspmu.h
> index e5c6dff2ce7f..274ca3d10578 100644
> --- a/drivers/perf/arm_cspmu/arm_cspmu.h
> +++ b/drivers/perf/arm_cspmu/arm_cspmu.h
> @@ -104,6 +104,9 @@ struct arm_cspmu_impl_ops {
>   	u32 (*event_type)(const struct perf_event *event);
>   	/* Decode filter value from configs */
>   	u32 (*event_filter)(const struct perf_event *event);
> +	/* Set event filter */
> +	void (*set_ev_filter)(struct arm_cspmu *cspmu,
> +			      struct hw_perf_event *hwc, u32 filter);
>   	/* Hide/show unsupported events */
>   	umode_t (*event_attr_is_visible)(struct kobject *kobj,
>   					 struct attribute *attr, int unused);

[PATCH v6 3/4] perf: arm_cspmu: Support implementation specific validation

[Ilkka Koskinen]

Some platforms may use e.g. different filtering mechanism and, thus,
may need different way to validate the events and group.

Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
---
 drivers/perf/arm_cspmu/arm_cspmu.c | 8 +++++++-
 drivers/perf/arm_cspmu/arm_cspmu.h | 3 +++
 2 files changed, 10 insertions(+), 1 deletion(-)

diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
index 94f6856ec786..585ce96ac03f 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.c
+++ b/drivers/perf/arm_cspmu/arm_cspmu.c
@@ -572,7 +572,7 @@ static void arm_cspmu_disable(struct pmu *pmu)
 static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events,
 				struct perf_event *event)
 {
-	int idx;
+	int idx, ret;
 	struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
 
 	if (supports_cycle_counter(cspmu)) {
@@ -606,6 +606,12 @@ static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events,
 	if (idx >= cspmu->num_logical_ctrs)
 		return -EAGAIN;
 
+	if (cspmu->impl.ops.validate_event) {
+		ret = cspmu->impl.ops.validate_event(cspmu, event);
+		if (ret)
+			return ret;
+	}
+
 	set_bit(idx, hw_events->used_ctrs);
 
 	return idx;
diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h b/drivers/perf/arm_cspmu/arm_cspmu.h
index 274ca3d10578..05577f74b8a0 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.h
+++ b/drivers/perf/arm_cspmu/arm_cspmu.h
@@ -107,6 +107,9 @@ struct arm_cspmu_impl_ops {
 	/* Set event filter */
 	void (*set_ev_filter)(struct arm_cspmu *cspmu,
 			      struct hw_perf_event *hwc, u32 filter);
+	/* Implementation specific event validation */
+	int (*validate_event)(struct arm_cspmu *cspmu,
+			      struct perf_event *event);
 	/* Hide/show unsupported events */
 	umode_t (*event_attr_is_visible)(struct kobject *kobj,
 					 struct attribute *attr, int unused);
-- 
2.41.0

Re: [PATCH v6 3/4] perf: arm_cspmu: Support implementation specific validation

[Suzuki K Poulose]

On 15/08/2023 07:35, Ilkka Koskinen wrote:
> Some platforms may use e.g. different filtering mechanism and, thus,
> may need different way to validate the events and group.
> 
> Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
> Reviewed-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>

Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>


> ---
>   drivers/perf/arm_cspmu/arm_cspmu.c | 8 +++++++-
>   drivers/perf/arm_cspmu/arm_cspmu.h | 3 +++
>   2 files changed, 10 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
> index 94f6856ec786..585ce96ac03f 100644
> --- a/drivers/perf/arm_cspmu/arm_cspmu.c
> +++ b/drivers/perf/arm_cspmu/arm_cspmu.c
> @@ -572,7 +572,7 @@ static void arm_cspmu_disable(struct pmu *pmu)
>   static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events,
>   				struct perf_event *event)
>   {
> -	int idx;
> +	int idx, ret;
>   	struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
>   
>   	if (supports_cycle_counter(cspmu)) {
> @@ -606,6 +606,12 @@ static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events,
>   	if (idx >= cspmu->num_logical_ctrs)
>   		return -EAGAIN;
>   
> +	if (cspmu->impl.ops.validate_event) {
> +		ret = cspmu->impl.ops.validate_event(cspmu, event);
> +		if (ret)
> +			return ret;
> +	}
> +
>   	set_bit(idx, hw_events->used_ctrs);
>   
>   	return idx;
> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h b/drivers/perf/arm_cspmu/arm_cspmu.h
> index 274ca3d10578..05577f74b8a0 100644
> --- a/drivers/perf/arm_cspmu/arm_cspmu.h
> +++ b/drivers/perf/arm_cspmu/arm_cspmu.h
> @@ -107,6 +107,9 @@ struct arm_cspmu_impl_ops {
>   	/* Set event filter */
>   	void (*set_ev_filter)(struct arm_cspmu *cspmu,
>   			      struct hw_perf_event *hwc, u32 filter);
> +	/* Implementation specific event validation */
> +	int (*validate_event)(struct arm_cspmu *cspmu,
> +			      struct perf_event *event);
>   	/* Hide/show unsupported events */
>   	umode_t (*event_attr_is_visible)(struct kobject *kobj,
>   					 struct attribute *attr, int unused);

[PATCH v6 4/4] perf: arm_cspmu: ampere_cspmu: Add support for Ampere SoC PMU

[Ilkka Koskinen]

Ampere SoC PMU follows CoreSight PMU architecture. It uses implementation
specific registers to filter events rather than PMEVFILTnR registers.

Signed-off-by: Ilkka Koskinen <ilkka@os.amperecomputing.com>
---
 .../admin-guide/perf/ampere_cspmu.rst         |  29 ++
 drivers/perf/arm_cspmu/Kconfig                |  10 +
 drivers/perf/arm_cspmu/Makefile               |   2 +
 drivers/perf/arm_cspmu/ampere_cspmu.c         | 271 ++++++++++++++++++
 drivers/perf/arm_cspmu/arm_cspmu.c            |   8 +
 drivers/perf/arm_cspmu/arm_cspmu.h            |   1 +
 6 files changed, 321 insertions(+)
 create mode 100644 Documentation/admin-guide/perf/ampere_cspmu.rst
 create mode 100644 drivers/perf/arm_cspmu/ampere_cspmu.c

diff --git a/Documentation/admin-guide/perf/ampere_cspmu.rst b/Documentation/admin-guide/perf/ampere_cspmu.rst
new file mode 100644
index 000000000000..94f93f5aee6c
--- /dev/null
+++ b/Documentation/admin-guide/perf/ampere_cspmu.rst
@@ -0,0 +1,29 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+============================================
+Ampere SoC Performance Monitoring Unit (PMU)
+============================================
+
+Ampere SoC PMU is a generic PMU IP that follows Arm CoreSight PMU architecture.
+Therefore, the driver is implemented as a submodule of arm_cspmu driver. At the
+first phase it's used for counting MCU events on AmpereOne.
+
+
+MCU PMU events
+--------------
+
+The PMU driver supports setting filters for "rank", "bank", and "threshold".
+Note, that the filters are per PMU instance rather than per event.
+
+
+Example for perf tool use::
+
+  / # perf list ampere
+
+    ampere_mcu_pmu_0/act_sent/                         [Kernel PMU event]
+    <...>
+    ampere_mcu_pmu_1/rd_sent/                          [Kernel PMU event]
+    <...>
+
+  / # perf stat -a -e ampere_mcu_pmu_0/act_sent,bank=5,rank=3,threshold=2/,ampere_mcu_pmu_1/rd_sent/ \
+        sleep 1
diff --git a/drivers/perf/arm_cspmu/Kconfig b/drivers/perf/arm_cspmu/Kconfig
index d5f787d22234..6f4e28fc84a2 100644
--- a/drivers/perf/arm_cspmu/Kconfig
+++ b/drivers/perf/arm_cspmu/Kconfig
@@ -17,3 +17,13 @@ config NVIDIA_CORESIGHT_PMU_ARCH_SYSTEM_PMU
 	help
 	  Provides NVIDIA specific attributes for performance monitoring unit
 	  (PMU) devices based on ARM CoreSight PMU architecture.
+
+config AMPERE_CORESIGHT_PMU_ARCH_SYSTEM_PMU
+       tristate "Ampere Coresight Architecture PMU"
+       depends on  ARM_CORESIGHT_PMU_ARCH_SYSTEM_PMU
+	help
+	  Provides Ampere specific attributes for performance monitoring unit
+	  (PMU) devices based on ARM CoreSight PMU architecture.
+
+	  In the first phase, the driver enables support on MCU PMU used in
+	  AmpereOne SoC family.
diff --git a/drivers/perf/arm_cspmu/Makefile b/drivers/perf/arm_cspmu/Makefile
index 0309d2ff264a..220a734efd54 100644
--- a/drivers/perf/arm_cspmu/Makefile
+++ b/drivers/perf/arm_cspmu/Makefile
@@ -3,6 +3,8 @@
 # SPDX-License-Identifier: GPL-2.0
 
 obj-$(CONFIG_ARM_CORESIGHT_PMU_ARCH_SYSTEM_PMU) += arm_cspmu_module.o
+
 arm_cspmu_module-y := arm_cspmu.o
 
 obj-$(CONFIG_NVIDIA_CORESIGHT_PMU_ARCH_SYSTEM_PMU) += nvidia_cspmu.o
+obj-$(CONFIG_AMPERE_CORESIGHT_PMU_ARCH_SYSTEM_PMU) += ampere_cspmu.o
diff --git a/drivers/perf/arm_cspmu/ampere_cspmu.c b/drivers/perf/arm_cspmu/ampere_cspmu.c
new file mode 100644
index 000000000000..a365f59fbfe7
--- /dev/null
+++ b/drivers/perf/arm_cspmu/ampere_cspmu.c
@@ -0,0 +1,271 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Ampere SoC PMU (Performance Monitor Unit)
+ *
+ * Copyright (c) 2023, Ampere Computing LLC
+ */
+#include <linux/module.h>
+#include <linux/topology.h>
+
+#include "arm_cspmu.h"
+
+#define PMAUXR0		0xD80
+#define PMAUXR1		0xD84
+#define PMAUXR2		0xD88
+#define PMAUXR3		0xD8C
+
+#define to_ampere_cspmu_ctx(cspmu)	((struct ampere_cspmu_ctx *)(cspmu->impl.ctx))
+
+struct ampere_cspmu_ctx {
+	const char *name;
+	struct attribute **event_attr;
+	struct attribute **format_attr;
+};
+
+static DEFINE_IDR(mcu_pmu_idr);
+
+#define SOC_PMU_EVENT_ATTR_EXTRACTOR(_name, _config, _start, _end)        \
+	static inline u32 get_##_name(const struct perf_event *event)     \
+	{                                                                 \
+		return FIELD_GET(GENMASK_ULL(_end, _start),               \
+				 event->attr._config);                    \
+	}                                                                 \
+
+SOC_PMU_EVENT_ATTR_EXTRACTOR(event, config, 0, 8);
+SOC_PMU_EVENT_ATTR_EXTRACTOR(threshold, config1, 0, 7);
+SOC_PMU_EVENT_ATTR_EXTRACTOR(rank, config1, 8, 23);
+SOC_PMU_EVENT_ATTR_EXTRACTOR(bank, config1, 24, 55);
+
+static struct attribute *ampereone_mcu_pmu_event_attrs[] = {
+	ARM_CSPMU_EVENT_ATTR(cycle_count,		0x00),
+	ARM_CSPMU_EVENT_ATTR(act_sent,			0x01),
+	ARM_CSPMU_EVENT_ATTR(pre_sent,			0x02),
+	ARM_CSPMU_EVENT_ATTR(rd_sent,			0x03),
+	ARM_CSPMU_EVENT_ATTR(rda_sent,			0x04),
+	ARM_CSPMU_EVENT_ATTR(wr_sent,			0x05),
+	ARM_CSPMU_EVENT_ATTR(wra_sent,			0x06),
+	ARM_CSPMU_EVENT_ATTR(pd_entry_vld,		0x07),
+	ARM_CSPMU_EVENT_ATTR(sref_entry_vld,		0x08),
+	ARM_CSPMU_EVENT_ATTR(prea_sent,			0x09),
+	ARM_CSPMU_EVENT_ATTR(pre_sb_sent,		0x0a),
+	ARM_CSPMU_EVENT_ATTR(ref_sent,			0x0b),
+	ARM_CSPMU_EVENT_ATTR(rfm_sent,			0x0c),
+	ARM_CSPMU_EVENT_ATTR(ref_sb_sent,		0x0d),
+	ARM_CSPMU_EVENT_ATTR(rfm_sb_sent,		0x0e),
+	ARM_CSPMU_EVENT_ATTR(rd_rda_sent,		0x0f),
+	ARM_CSPMU_EVENT_ATTR(wr_wra_sent,		0x10),
+	ARM_CSPMU_EVENT_ATTR(raw_hazard,		0x11),
+	ARM_CSPMU_EVENT_ATTR(war_hazard,		0x12),
+	ARM_CSPMU_EVENT_ATTR(waw_hazard,		0x13),
+	ARM_CSPMU_EVENT_ATTR(rar_hazard,		0x14),
+	ARM_CSPMU_EVENT_ATTR(raw_war_waw_hazard,	0x15),
+	ARM_CSPMU_EVENT_ATTR(hprd_lprd_wr_req_vld,	0x16),
+	ARM_CSPMU_EVENT_ATTR(lprd_req_vld,		0x17),
+	ARM_CSPMU_EVENT_ATTR(hprd_req_vld,		0x18),
+	ARM_CSPMU_EVENT_ATTR(hprd_lprd_req_vld,		0x19),
+	ARM_CSPMU_EVENT_ATTR(prefetch_tgt,		0x1a),
+	ARM_CSPMU_EVENT_ATTR(wr_req_vld,		0x1b),
+	ARM_CSPMU_EVENT_ATTR(partial_wr_req_vld,	0x1c),
+	ARM_CSPMU_EVENT_ATTR(rd_retry,			0x1d),
+	ARM_CSPMU_EVENT_ATTR(wr_retry,			0x1e),
+	ARM_CSPMU_EVENT_ATTR(retry_gnt,			0x1f),
+	ARM_CSPMU_EVENT_ATTR(rank_change,		0x20),
+	ARM_CSPMU_EVENT_ATTR(dir_change,		0x21),
+	ARM_CSPMU_EVENT_ATTR(rank_dir_change,		0x22),
+	ARM_CSPMU_EVENT_ATTR(rank_active,		0x23),
+	ARM_CSPMU_EVENT_ATTR(rank_idle,			0x24),
+	ARM_CSPMU_EVENT_ATTR(rank_pd,			0x25),
+	ARM_CSPMU_EVENT_ATTR(rank_sref,			0x26),
+	ARM_CSPMU_EVENT_ATTR(queue_fill_gt_thresh,	0x27),
+	ARM_CSPMU_EVENT_ATTR(queue_rds_gt_thresh,	0x28),
+	ARM_CSPMU_EVENT_ATTR(queue_wrs_gt_thresh,	0x29),
+	ARM_CSPMU_EVENT_ATTR(phy_updt_complt,		0x2a),
+	ARM_CSPMU_EVENT_ATTR(tz_fail,			0x2b),
+	ARM_CSPMU_EVENT_ATTR(dram_errc,			0x2c),
+	ARM_CSPMU_EVENT_ATTR(dram_errd,			0x2d),
+	ARM_CSPMU_EVENT_ATTR(read_data_return,		0x32),
+	ARM_CSPMU_EVENT_ATTR(chi_wr_data_delta,		0x33),
+	ARM_CSPMU_EVENT_ATTR(zq_start,			0x34),
+	ARM_CSPMU_EVENT_ATTR(zq_latch,			0x35),
+	ARM_CSPMU_EVENT_ATTR(wr_fifo_full,		0x36),
+	ARM_CSPMU_EVENT_ATTR(info_fifo_full,		0x37),
+	ARM_CSPMU_EVENT_ATTR(cmd_fifo_full,		0x38),
+	ARM_CSPMU_EVENT_ATTR(dfi_nop,			0x39),
+	ARM_CSPMU_EVENT_ATTR(dfi_cmd,			0x3a),
+	ARM_CSPMU_EVENT_ATTR(rd_run_len,		0x3b),
+	ARM_CSPMU_EVENT_ATTR(wr_run_len,		0x3c),
+
+	ARM_CSPMU_EVENT_ATTR(cycles, ARM_CSPMU_EVT_CYCLES_DEFAULT),
+	NULL,
+};
+
+static struct attribute *ampereone_mcu_format_attrs[] = {
+	ARM_CSPMU_FORMAT_EVENT_ATTR,
+	ARM_CSPMU_FORMAT_ATTR(threshold, "config1:0-7"),
+	ARM_CSPMU_FORMAT_ATTR(rank, "config1:8-23"),
+	ARM_CSPMU_FORMAT_ATTR(bank, "config1:24-55"),
+	NULL,
+};
+
+static struct attribute **
+ampere_cspmu_get_event_attrs(const struct arm_cspmu *cspmu)
+{
+	const struct ampere_cspmu_ctx *ctx = to_ampere_cspmu_ctx(cspmu);
+
+	return ctx->event_attr;
+}
+
+static struct attribute **
+ampere_cspmu_get_format_attrs(const struct arm_cspmu *cspmu)
+{
+	const struct ampere_cspmu_ctx *ctx = to_ampere_cspmu_ctx(cspmu);
+
+	return ctx->format_attr;
+}
+
+static const char *
+ampere_cspmu_get_name(const struct arm_cspmu *cspmu)
+{
+	const struct ampere_cspmu_ctx *ctx = to_ampere_cspmu_ctx(cspmu);
+
+	return ctx->name;
+}
+
+static u32 ampere_cspmu_event_filter(const struct perf_event *event)
+{
+	/*
+	 * PMEVFILTR or PMCCFILTR aren't used in Ampere SoC PMU but are marked
+	 * as RES0. Make sure, PMCCFILTR is written zero.
+	 */
+	return 0;
+}
+
+static void ampere_cspmu_set_ev_filter(struct arm_cspmu *cspmu,
+				       struct hw_perf_event *hwc,
+				       u32 filter)
+{
+	struct perf_event *event;
+	unsigned int idx;
+	u32 threshold, rank, bank;
+
+	/*
+	 * At this point, all the events have the same filter settings.
+	 * Therefore, take the first event and use its configuration.
+	 */
+	idx = find_first_bit(cspmu->hw_events.used_ctrs,
+			     cspmu->cycle_counter_logical_idx);
+
+	event = cspmu->hw_events.events[idx];
+
+	threshold	= get_threshold(event);
+	rank		= get_rank(event);
+	bank		= get_bank(event);
+
+	writel(threshold, cspmu->base0 + PMAUXR0);
+	writel(rank, cspmu->base0 + PMAUXR1);
+	writel(bank, cspmu->base0 + PMAUXR2);
+}
+
+static int ampere_cspmu_validate_configs(struct perf_event *event,
+					 struct perf_event *event2)
+{
+	if (get_threshold(event) != get_threshold(event2) ||
+	    get_rank(event) != get_rank(event2) ||
+	    get_bank(event) != get_bank(event2))
+		return -EINVAL;
+
+	return 0;
+}
+
+static int ampere_cspmu_validate_event(struct arm_cspmu *cspmu,
+				       struct perf_event *new)
+{
+	struct perf_event *curr, *leader = new->group_leader;
+	unsigned int idx;
+	int ret;
+
+	ret = ampere_cspmu_validate_configs(new, leader);
+	if (ret)
+		return ret;
+
+	/* We compare the global filter settings to the existing events */
+	idx = find_first_bit(cspmu->hw_events.used_ctrs,
+			     cspmu->cycle_counter_logical_idx);
+
+	/* This is the first event, thus any configuration is fine */
+	if (idx == cspmu->cycle_counter_logical_idx)
+		return 0;
+
+	curr = cspmu->hw_events.events[idx];
+
+	return ampere_cspmu_validate_configs(curr, new);
+}
+
+static char *ampere_cspmu_format_name(const struct arm_cspmu *cspmu,
+				      const char *name_pattern)
+{
+	struct device *dev = cspmu->dev;
+	int id;
+
+	id = idr_alloc(&mcu_pmu_idr, NULL, 0, 0, GFP_KERNEL);
+	if (id < 0)
+		return ERR_PTR(id);
+
+	return devm_kasprintf(dev, GFP_KERNEL, name_pattern, id);
+}
+
+static int ampere_cspmu_init_ops(struct arm_cspmu *cspmu)
+{
+	struct device *dev = cspmu->dev;
+	struct ampere_cspmu_ctx *ctx;
+	struct arm_cspmu_impl_ops *impl_ops = &cspmu->impl.ops;
+
+	ctx = devm_kzalloc(dev, sizeof(struct ampere_cspmu_ctx), GFP_KERNEL);
+	if (!ctx)
+		return -ENOMEM;
+
+	ctx->event_attr	= ampereone_mcu_pmu_event_attrs;
+	ctx->format_attr = ampereone_mcu_format_attrs;
+	ctx->name = ampere_cspmu_format_name(cspmu, "ampere_mcu_pmu_%d");
+	if (IS_ERR_OR_NULL(ctx->name))
+		return ctx->name ? PTR_ERR(ctx->name) : -ENOMEM;
+
+	cspmu->impl.ctx = ctx;
+
+	impl_ops->event_filter		= ampere_cspmu_event_filter;
+	impl_ops->set_ev_filter		= ampere_cspmu_set_ev_filter;
+	impl_ops->validate_event	= ampere_cspmu_validate_event;
+	impl_ops->get_name		= ampere_cspmu_get_name;
+	impl_ops->get_event_attrs	= ampere_cspmu_get_event_attrs;
+	impl_ops->get_format_attrs	= ampere_cspmu_get_format_attrs;
+
+	return 0;
+}
+
+/* Match all Ampere Coresight PMU devices */
+static const struct arm_cspmu_impl_match ampere_cspmu_param = {
+	.pmiidr_val	= ARM_CSPMU_IMPL_ID_AMPERE,
+	.module		= THIS_MODULE,
+	.impl_init_ops	= ampere_cspmu_init_ops
+};
+
+static int __init ampere_cspmu_init(void)
+{
+	int ret;
+
+	ret = arm_cspmu_impl_register(&ampere_cspmu_param);
+	if (ret)
+		pr_err("ampere_cspmu backend registration error: %d\n", ret);
+
+	return ret;
+}
+
+static void __exit ampere_cspmu_exit(void)
+{
+	arm_cspmu_impl_unregister(&ampere_cspmu_param);
+}
+
+module_init(ampere_cspmu_init);
+module_exit(ampere_cspmu_exit);
+
+MODULE_LICENSE("GPL");
diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
index 585ce96ac03f..d6c50e06c5bc 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.c
+++ b/drivers/perf/arm_cspmu/arm_cspmu.c
@@ -383,6 +383,14 @@ static struct arm_cspmu_impl_match impl_match[] = {
 		.module		= NULL,
 		.impl_init_ops	= NULL,
 	},
+	{
+		.module_name	= "ampere_cspmu",
+		.pmiidr_val	= ARM_CSPMU_IMPL_ID_AMPERE,
+		.pmiidr_mask	= ARM_CSPMU_PMIIDR_IMPLEMENTER,
+		.module		= NULL,
+		.impl_init_ops	= NULL,
+	},
+
 	{0}
 };
 
diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h b/drivers/perf/arm_cspmu/arm_cspmu.h
index 05577f74b8a0..586441d4dba1 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.h
+++ b/drivers/perf/arm_cspmu/arm_cspmu.h
@@ -71,6 +71,7 @@
 
 /* JEDEC-assigned JEP106 identification code */
 #define ARM_CSPMU_IMPL_ID_NVIDIA	0x36B
+#define ARM_CSPMU_IMPL_ID_AMPERE	0xA16
 
 struct arm_cspmu;
 
-- 
2.41.0