powerpc/perf: Record counter overflow always if SAMPLE_IP is unset

While sampling for marked events, currently we record the sample only
if the SIAR valid bit of Sampled Instruction Event Register (SIER) is
set. SIAR_VALID bit is used for fetching the instruction address from
Sampled Instruction Address Register(SIAR). But there are some usecases,
where the user is interested only in the PMU stats at each counter
overflow and the exact IP of the overflow event is not required.
Dropping SIAR invalid samples will fail to record some of the counter
overflows in such cases.

Example of such usecase is dumping the PMU stats (event counts)
after some regular amount of instructions/events from the userspace
(ex: via ptrace). Here counter overflow is indicated to userspace via
signal handler, and captured by monitoring and enabling I/O
signaling on the event file descriptor. In these cases, we expect to
get sample/overflow indication after each specified sample_period.

Perf event attribute will not have PERF_SAMPLE_IP set in the
sample_type if exact IP of the overflow event is not requested. So
while profiling if SAMPLE_IP is not set, just record the counter overflow
irrespective of SIAR_VALID check.

Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
---
 arch/powerpc/perf/core-book3s.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

Athira Rajeev <atrajeev@linux.vnet.ibm.com> writes:
> While sampling for marked events, currently we record the sample only
> if the SIAR valid bit of Sampled Instruction Event Register (SIER) is
> set. SIAR_VALID bit is used for fetching the instruction address from
> Sampled Instruction Address Register(SIAR). But there are some usecases,
> where the user is interested only in the PMU stats at each counter
> overflow and the exact IP of the overflow event is not required.
> Dropping SIAR invalid samples will fail to record some of the counter
> overflows in such cases.
>
> Example of such usecase is dumping the PMU stats (event counts)
> after some regular amount of instructions/events from the userspace
> (ex: via ptrace). Here counter overflow is indicated to userspace via
> signal handler, and captured by monitoring and enabling I/O
> signaling on the event file descriptor. In these cases, we expect to
> get sample/overflow indication after each specified sample_period.
>
> Perf event attribute will not have PERF_SAMPLE_IP set in the
> sample_type if exact IP of the overflow event is not requested. So
> while profiling if SAMPLE_IP is not set, just record the counter overflow
> irrespective of SIAR_VALID check.
>
> Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
> ---
>  arch/powerpc/perf/core-book3s.c | 10 ++++++++--
>  1 file changed, 8 insertions(+), 2 deletions(-)
>
> diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
> index 28206b1fe172..bb4828a05e4d 100644
> --- a/arch/powerpc/perf/core-book3s.c
> +++ b/arch/powerpc/perf/core-book3s.c
> @@ -2166,10 +2166,16 @@ static void record_and_restart(struct perf_event *event, unsigned long val,
>  	 * address even when freeze on supervisor state (kernel) is set in
>  	 * MMCR2. Check attr.exclude_kernel and address to drop the sample in
>  	 * these cases.
> +	 *
> +	 * If address is not requested in the sample
> +	 * via PERF_SAMPLE_IP, just record that sample
> +	 * irrespective of SIAR valid check.
>  	 */
> -	if (event->attr.exclude_kernel && record)
> -		if (is_kernel_addr(mfspr(SPRN_SIAR)))
> +	if (event->attr.exclude_kernel && record) {
> +		if (is_kernel_addr(mfspr(SPRN_SIAR)) && (event->attr.sample_type & PERF_SAMPLE_IP))
>  			record = 0;
> +	} else if (!record && !(event->attr.sample_type & PERF_SAMPLE_IP))
> +		record = 1;

This seems wrong, you're assuming that record was set previously to
= siar_valid(), but it may be that record is still 0 from the
initialisation and we weren't going to record.

Don't we need something more like this?

diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
index 9fd06010e8b6..e4e8a017d339 100644
--- a/arch/powerpc/perf/core-book3s.c
+++ b/arch/powerpc/perf/core-book3s.c
@@ -2136,7 +2136,12 @@ static void record_and_restart(struct perf_event *event, unsigned long val,
 			left += period;
 			if (left <= 0)
 				left = period;
-			record = siar_valid(regs);
+
+			if (event->attr.sample_type & PERF_SAMPLE_IP)
+				record = siar_valid(regs);
+			else
+				record = 1;
+
 			event->hw.last_period = event->hw.sample_period;
 		}
 		if (left < 0x80000000LL)
@@ -2154,9 +2159,10 @@ static void record_and_restart(struct perf_event *event, unsigned long val,
 	 * MMCR2. Check attr.exclude_kernel and address to drop the sample in
 	 * these cases.
 	 */
-	if (event->attr.exclude_kernel && record)
-		if (is_kernel_addr(mfspr(SPRN_SIAR)))
-			record = 0;
+	if (event->attr.exclude_kernel &&
+	    (event->attr.sample_type & PERF_SAMPLE_IP) &&
+	    is_kernel_addr(mfspr(SPRN_SIAR)))
+		record = 0;
 
 	/*
 	 * Finally record data if requested.



cheers

> On 04-Feb-2021, at 8:25 AM, Michael Ellerman <mpe@ellerman.id.au> wrote:
> 
> Athira Rajeev <atrajeev@linux.vnet.ibm.com> writes:
>> While sampling for marked events, currently we record the sample only
>> if the SIAR valid bit of Sampled Instruction Event Register (SIER) is
>> set. SIAR_VALID bit is used for fetching the instruction address from
>> Sampled Instruction Address Register(SIAR). But there are some usecases,
>> where the user is interested only in the PMU stats at each counter
>> overflow and the exact IP of the overflow event is not required.
>> Dropping SIAR invalid samples will fail to record some of the counter
>> overflows in such cases.
>> 
>> Example of such usecase is dumping the PMU stats (event counts)
>> after some regular amount of instructions/events from the userspace
>> (ex: via ptrace). Here counter overflow is indicated to userspace via
>> signal handler, and captured by monitoring and enabling I/O
>> signaling on the event file descriptor. In these cases, we expect to
>> get sample/overflow indication after each specified sample_period.
>> 
>> Perf event attribute will not have PERF_SAMPLE_IP set in the
>> sample_type if exact IP of the overflow event is not requested. So
>> while profiling if SAMPLE_IP is not set, just record the counter overflow
>> irrespective of SIAR_VALID check.
>> 
>> Signed-off-by: Athira Rajeev <atrajeev@linux.vnet.ibm.com>
>> ---
>> arch/powerpc/perf/core-book3s.c | 10 ++++++++--
>> 1 file changed, 8 insertions(+), 2 deletions(-)
>> 
>> diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
>> index 28206b1fe172..bb4828a05e4d 100644
>> --- a/arch/powerpc/perf/core-book3s.c
>> +++ b/arch/powerpc/perf/core-book3s.c
>> @@ -2166,10 +2166,16 @@ static void record_and_restart(struct perf_event *event, unsigned long val,
>> 	 * address even when freeze on supervisor state (kernel) is set in
>> 	 * MMCR2. Check attr.exclude_kernel and address to drop the sample in
>> 	 * these cases.
>> +	 *
>> +	 * If address is not requested in the sample
>> +	 * via PERF_SAMPLE_IP, just record that sample
>> +	 * irrespective of SIAR valid check.
>> 	 */
>> -	if (event->attr.exclude_kernel && record)
>> -		if (is_kernel_addr(mfspr(SPRN_SIAR)))
>> +	if (event->attr.exclude_kernel && record) {
>> +		if (is_kernel_addr(mfspr(SPRN_SIAR)) && (event->attr.sample_type & PERF_SAMPLE_IP))
>> 			record = 0;
>> +	} else if (!record && !(event->attr.sample_type & PERF_SAMPLE_IP))
>> +		record = 1;
> 
> This seems wrong, you're assuming that record was set previously to
> = siar_valid(), but it may be that record is still 0 from the
> initialisation and we weren't going to record.
> 
> Don't we need something more like this?

Hi Michael,

Thanks for checking the patch and sharing the suggestion.

Yes, the below change looks good and tested with my scenario. 
I will send a V2 with new change.

Thanks
Athira
> 
> diff --git a/arch/powerpc/perf/core-book3s.c b/arch/powerpc/perf/core-book3s.c
> index 9fd06010e8b6..e4e8a017d339 100644
> --- a/arch/powerpc/perf/core-book3s.c
> +++ b/arch/powerpc/perf/core-book3s.c
> @@ -2136,7 +2136,12 @@ static void record_and_restart(struct perf_event *event, unsigned long val,
> 			left += period;
> 			if (left <= 0)
> 				left = period;
> -			record = siar_valid(regs);
> +
> +			if (event->attr.sample_type & PERF_SAMPLE_IP)
> +				record = siar_valid(regs);
> +			else
> +				record = 1;
> +
> 			event->hw.last_period = event->hw.sample_period;
> 		}
> 		if (left < 0x80000000LL)
> @@ -2154,9 +2159,10 @@ static void record_and_restart(struct perf_event *event, unsigned long val,
> 	 * MMCR2. Check attr.exclude_kernel and address to drop the sample in
> 	 * these cases.
> 	 */
> -	if (event->attr.exclude_kernel && record)
> -		if (is_kernel_addr(mfspr(SPRN_SIAR)))
> -			record = 0;
> +	if (event->attr.exclude_kernel &&
> +	    (event->attr.sample_type & PERF_SAMPLE_IP) &&
> +	    is_kernel_addr(mfspr(SPRN_SIAR)))
> +		record = 0;
> 
> 	/*
> 	 * Finally record data if requested.
> 
> 
> 
> cheers