diff mbox series

[v2,1/5] perf: arm_cspmu: Support 32-bit accesses to 64-bit registers

Message ID 20230601030144.3458136-2-ilkka@os.amperecomputing.com (mailing list archive)
State New, archived
Headers show
Series perf: ampere: Add support for Ampere SoC PMUs | expand

Commit Message

Ilkka Koskinen June 1, 2023, 3:01 a.m. UTC
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>
---
 drivers/perf/arm_cspmu/arm_cspmu.c | 8 ++++++--
 drivers/perf/arm_cspmu/arm_cspmu.h | 1 +
 2 files changed, 7 insertions(+), 2 deletions(-)

Comments

Robin Murphy June 1, 2023, 2:49 p.m. UTC | #1
On 2023-06-01 04:01, 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>
> ---
>   drivers/perf/arm_cspmu/arm_cspmu.c | 8 ++++++--
>   drivers/perf/arm_cspmu/arm_cspmu.h | 1 +
>   2 files changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
> index a3f1c410b417..88547a2b73e6 100644
> --- a/drivers/perf/arm_cspmu/arm_cspmu.c
> +++ b/drivers/perf/arm_cspmu/arm_cspmu.c
> @@ -701,8 +701,12 @@ 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->impl.split_64bit_access) {

Could we not just hang this off the 64-bit atomicity property to match 
the read path? It doesn't seem like there's much benefit in 
micro-optimising for whether the interconnect splits 64-bit accesses 
into 32-bit bursts vs. just not accepting them at all.

> +			writeq(val, cspmu->base1 + offset);
> +		} else {
> +			writel(lower_32_bits(val), cspmu->base1 + offset);
> +			writel(upper_32_bits(val), cspmu->base1 + offset + 4);

lo_hi_writeq() - the header's already included for 32-bit build 
coverage, so we may as well put it to use :)

Thanks,
Robin.

> +		}
>   	} else {
>   		offset = counter_offset(sizeof(u32), event->hw.idx);
>   
> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h b/drivers/perf/arm_cspmu/arm_cspmu.h
> index 51323b175a4a..c0412cf2bd97 100644
> --- a/drivers/perf/arm_cspmu/arm_cspmu.h
> +++ b/drivers/perf/arm_cspmu/arm_cspmu.h
> @@ -110,6 +110,7 @@ struct arm_cspmu_impl_ops {
>   /* Vendor/implementer descriptor. */
>   struct arm_cspmu_impl {
>   	u32 pmiidr;
> +	bool split_64bit_access;
>   	struct arm_cspmu_impl_ops ops;
>   	void *ctx;
>   };
Ilkka Koskinen June 2, 2023, 6:47 a.m. UTC | #2
Hi Robin,

On Thu, 1 Jun 2023, Robin Murphy wrote:
> On 2023-06-01 04:01, 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>
>> ---
>>   drivers/perf/arm_cspmu/arm_cspmu.c | 8 ++++++--
>>   drivers/perf/arm_cspmu/arm_cspmu.h | 1 +
>>   2 files changed, 7 insertions(+), 2 deletions(-)
>> 
>> diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c 
>> b/drivers/perf/arm_cspmu/arm_cspmu.c
>> index a3f1c410b417..88547a2b73e6 100644
>> --- a/drivers/perf/arm_cspmu/arm_cspmu.c
>> +++ b/drivers/perf/arm_cspmu/arm_cspmu.c
>> @@ -701,8 +701,12 @@ 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->impl.split_64bit_access) {
>
> Could we not just hang this off the 64-bit atomicity property to match the 
> read path? It doesn't seem like there's much benefit in micro-optimising for 
> whether the interconnect splits 64-bit accesses into 32-bit bursts vs. just 
> not accepting them at all.

Sure, I was actually wondering if I could use it or if there could be 
a really weird hw implementation. I'll change it.

>
>> +			writeq(val, cspmu->base1 + offset);
>> +		} else {
>> +			writel(lower_32_bits(val), cspmu->base1 + offset);
>> +			writel(upper_32_bits(val), cspmu->base1 + offset + 
>> 4);
>
> lo_hi_writeq() - the header's already included for 32-bit build coverage, so 
> we may as well put it to use :)

Oh, I have missed that function completely. I fix this too.

Cheers, Ilkka


> Thanks,
> Robin.
>
>> +		}
>>   	} else {
>>   		offset = counter_offset(sizeof(u32), event->hw.idx);
>>   diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h 
>> b/drivers/perf/arm_cspmu/arm_cspmu.h
>> index 51323b175a4a..c0412cf2bd97 100644
>> --- a/drivers/perf/arm_cspmu/arm_cspmu.h
>> +++ b/drivers/perf/arm_cspmu/arm_cspmu.h
>> @@ -110,6 +110,7 @@ struct arm_cspmu_impl_ops {
>>   /* Vendor/implementer descriptor. */
>>   struct arm_cspmu_impl {
>>   	u32 pmiidr;
>> +	bool split_64bit_access;
>>   	struct arm_cspmu_impl_ops ops;
>>   	void *ctx;
>>   };
>
diff mbox series

Patch

diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c b/drivers/perf/arm_cspmu/arm_cspmu.c
index a3f1c410b417..88547a2b73e6 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.c
+++ b/drivers/perf/arm_cspmu/arm_cspmu.c
@@ -701,8 +701,12 @@  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->impl.split_64bit_access) {
+			writeq(val, cspmu->base1 + offset);
+		} else {
+			writel(lower_32_bits(val), cspmu->base1 + offset);
+			writel(upper_32_bits(val), cspmu->base1 + offset + 4);
+		}
 	} else {
 		offset = counter_offset(sizeof(u32), event->hw.idx);
 
diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h b/drivers/perf/arm_cspmu/arm_cspmu.h
index 51323b175a4a..c0412cf2bd97 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.h
+++ b/drivers/perf/arm_cspmu/arm_cspmu.h
@@ -110,6 +110,7 @@  struct arm_cspmu_impl_ops {
 /* Vendor/implementer descriptor. */
 struct arm_cspmu_impl {
 	u32 pmiidr;
+	bool split_64bit_access;
 	struct arm_cspmu_impl_ops ops;
 	void *ctx;
 };