[PATCH 0/3] perf intel-pt: Add PEBS-via-PT side-band

[PATCH 0/3] perf intel-pt: Add PEBS-via-PT side-band

[Adrian Hunter]

Hi

PEBS output to Intel Processor Trace was introduced with Atom
processors based on Tremont.  Currently there is software support
only for a single PEBS-via-PT event.

Here is support for multiple PEBS-via-PT events.

The first patch is the kernel change which adds a new event,
namely PERF_RECORD_AUX_OUTPUT_HW_ID which contains the counter
index.  That is output when the PEBS-via-PT event is enabled in hardware.
There is an optimization, to report the index only when it changes
for the event.  That will work only so long as all PEBS-via-PT
events are scheduled together, which they are for a recording
session because they are in a single group.

Also no attribute bit is used to select the new event, so a new
kernel is not compatible with older perf tools.  The assumption
being that PEBS-via-PT is sufficiently esoteric that users will not
be troubled by this.

The second patch adds the usual boiler plate to perf tools for
a new event.

The third patch adds support for processing the new event by
perf tool's Intel PT decoder.


Adrian Hunter (3):
      perf/x86: Add new event for AUX output counter index
      perf tools: Add support for PERF_RECORD_AUX_OUTPUT_HW_ID
      perf intel-pt: Add support for PERF_RECORD_AUX_OUTPUT_HW_ID

 arch/x86/events/core.c                     |  1 +
 arch/x86/events/intel/ds.c                 | 16 ++++++
 include/linux/perf_event.h                 |  2 +
 include/uapi/linux/perf_event.h            | 15 ++++++
 kernel/events/core.c                       | 30 +++++++++++
 tools/include/uapi/linux/perf_event.h      | 15 ++++++
 tools/lib/perf/include/perf/event.h        |  6 +++
 tools/perf/Documentation/perf-intel-pt.txt |  7 ++-
 tools/perf/builtin-inject.c                |  4 +-
 tools/perf/builtin-record.c                |  2 +-
 tools/perf/util/event.c                    | 18 +++++++
 tools/perf/util/event.h                    |  5 ++
 tools/perf/util/intel-pt.c                 | 85 ++++++++++++++++++++++++++++--
 tools/perf/util/machine.c                  | 10 ++++
 tools/perf/util/machine.h                  |  2 +
 tools/perf/util/session.c                  |  5 ++
 tools/perf/util/tool.h                     |  1 +
 17 files changed, 217 insertions(+), 7 deletions(-)


Regards
Adrian

[PATCH 1/3] perf/x86: Add new event for AUX output counter index

[Adrian Hunter]

PEBS-via-PT records contain a mask of applicable counters. To identify
which event belongs to which counter, a side-band event is needed. Until
now, there has been no side-band event, and consequently users were limited
to using a single event.

Add such a side-band event. Note the event is optimised to output only
when the counter index changes for an event. That works only so long as
all PEBS-via-PT events are scheduled together, which they are for a
recording session because they are in a single group.

Also no attribute bit is used to select the new event, so a new
kernel is not compatible with older perf tools.  The assumption
being that PEBS-via-PT is sufficiently esoteric that users will not
be troubled by this.

Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Reviewed-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
---
 arch/x86/events/core.c          |  1 +
 arch/x86/events/intel/ds.c      | 16 ++++++++++++++++
 include/linux/perf_event.h      |  2 ++
 include/uapi/linux/perf_event.h | 15 +++++++++++++++
 kernel/events/core.c            | 30 ++++++++++++++++++++++++++++++
 5 files changed, 64 insertions(+)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 8f71dd72ef95..30d0d18eb4b6 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -660,6 +660,7 @@ static int __x86_pmu_event_init(struct perf_event *event)
 	event->destroy = hw_perf_event_destroy;
 
 	event->hw.idx = -1;
+	event->hw.idx_reported = -1;
 	event->hw.last_cpu = -1;
 	event->hw.last_tag = ~0ULL;
 
diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
index 1ec8fd311f38..d4c149b6e82b 100644
--- a/arch/x86/events/intel/ds.c
+++ b/arch/x86/events/intel/ds.c
@@ -1182,6 +1182,20 @@ static void intel_pmu_pebs_via_pt_disable(struct perf_event *event)
 		cpuc->pebs_enabled &= ~PEBS_VIA_PT_MASK;
 }
 
+static void intel_pmu_report_aux_output_id(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+
+	/*
+	 * So long as all PEBS-via-PT events for a recording session are
+	 * scheduled together, then only changes to hwc->idx need be reported.
+	 */
+	if (hwc->idx != hwc->idx_reported) {
+		hwc->idx_reported = hwc->idx;
+		perf_report_aux_output_id(event, hwc->idx);
+	}
+}
+
 static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
 {
 	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
@@ -1191,6 +1205,8 @@ static void intel_pmu_pebs_via_pt_enable(struct perf_event *event)
 	if (!is_pebs_pt(event))
 		return;
 
+	intel_pmu_report_aux_output_id(event);
+
 	if (!(event->hw.flags & PERF_X86_EVENT_LARGE_PEBS))
 		cpuc->pebs_enabled |= PEBS_PMI_AFTER_EACH_RECORD;
 
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index f5a6a2f069ed..18bfa05537ab 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -141,6 +141,7 @@ struct hw_perf_event {
 			unsigned long	event_base;
 			int		event_base_rdpmc;
 			int		idx;
+			int		idx_reported;
 			int		last_cpu;
 			int		flags;
 
@@ -1397,6 +1398,7 @@ perf_event_addr_filters(struct perf_event *event)
 }
 
 extern void perf_event_addr_filters_sync(struct perf_event *event);
+extern void perf_report_aux_output_id(struct perf_event *event, u64 hw_id);
 
 extern int perf_output_begin(struct perf_output_handle *handle,
 			     struct perf_sample_data *data,
diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
index f92880a15645..c89535de1ec8 100644
--- a/include/uapi/linux/perf_event.h
+++ b/include/uapi/linux/perf_event.h
@@ -1141,6 +1141,21 @@ enum perf_event_type {
 	 */
 	PERF_RECORD_TEXT_POKE			= 20,
 
+	/*
+	 * Data written to the AUX area by hardware due to aux_output, may need
+	 * to be matched to the event by an architecture-specific hardware ID.
+	 * This records the hardware ID, but requires sample_id to provide the
+	 * event ID. e.g. Intel PT uses this record to disambiguate PEBS-via-PT
+	 * records from multiple events.
+	 *
+	 * struct {
+	 *	struct perf_event_header	header;
+	 *	u64				hw_id;
+	 *	struct sample_id		sample_id;
+	 * };
+	 */
+	PERF_RECORD_AUX_OUTPUT_HW_ID		= 21,
+
 	PERF_RECORD_MAX,			/* non-ABI */
 };
 
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 6fee4a7e88d7..bf98c8932ae2 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -9055,6 +9055,36 @@ static void perf_log_itrace_start(struct perf_event *event)
 	perf_output_end(&handle);
 }
 
+void perf_report_aux_output_id(struct perf_event *event, u64 hw_id)
+{
+	struct perf_output_handle handle;
+	struct perf_sample_data sample;
+	struct perf_aux_event {
+		struct perf_event_header        header;
+		u64				hw_id;
+	} rec;
+	int ret;
+
+	if (event->parent)
+		event = event->parent;
+
+	rec.header.type	= PERF_RECORD_AUX_OUTPUT_HW_ID;
+	rec.header.misc	= 0;
+	rec.header.size	= sizeof(rec);
+	rec.hw_id	= hw_id;
+
+	perf_event_header__init_id(&rec.header, &sample, event);
+	ret = perf_output_begin(&handle, &sample, event, rec.header.size);
+
+	if (ret)
+		return;
+
+	perf_output_put(&handle, rec);
+	perf_event__output_id_sample(event, &handle, &sample);
+
+	perf_output_end(&handle);
+}
+
 static int
 __perf_event_account_interrupt(struct perf_event *event, int throttle)
 {
-- 
2.17.1

Re: [PATCH 1/3] perf/x86: Add new event for AUX output counter index

[Peter Zijlstra]

On Wed, Jun 09, 2021 at 05:20:53PM +0300, Adrian Hunter wrote:
> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> index f5a6a2f069ed..18bfa05537ab 100644
> --- a/include/linux/perf_event.h
> +++ b/include/linux/perf_event.h
> @@ -141,6 +141,7 @@ struct hw_perf_event {
>  			unsigned long	event_base;
>  			int		event_base_rdpmc;
>  			int		idx;
> +			int		idx_reported;
>  			int		last_cpu;
>  			int		flags;
>  

This is sad, it blows up the largest element in that union from 96 to
104 bytes.

Is there really no better place for this?

Re: [PATCH 1/3] perf/x86: Add new event for AUX output counter index

[Peter Zijlstra]

On Wed, Jun 09, 2021 at 05:20:53PM +0300, Adrian Hunter wrote:

> +static void intel_pmu_report_aux_output_id(struct perf_event *event)
> +{
> +	struct hw_perf_event *hwc = &event->hw;
> +
> +	/*
> +	 * So long as all PEBS-via-PT events for a recording session are
> +	 * scheduled together, then only changes to hwc->idx need be reported.
> +	 */
> +	if (hwc->idx != hwc->idx_reported) {
> +		hwc->idx_reported = hwc->idx;
> +		perf_report_aux_output_id(event, hwc->idx);
> +	}
> +}

AFAICT you want a callback in x86_assign_hw_event(), is that so?

Re: [PATCH 1/3] perf/x86: Add new event for AUX output counter index

[Adrian Hunter]

On 25/06/21 4:15 pm, Peter Zijlstra wrote:
> On Wed, Jun 09, 2021 at 05:20:53PM +0300, Adrian Hunter wrote:
> 
>> +static void intel_pmu_report_aux_output_id(struct perf_event *event)
>> +{
>> +	struct hw_perf_event *hwc = &event->hw;
>> +
>> +	/*
>> +	 * So long as all PEBS-via-PT events for a recording session are
>> +	 * scheduled together, then only changes to hwc->idx need be reported.
>> +	 */
>> +	if (hwc->idx != hwc->idx_reported) {
>> +		hwc->idx_reported = hwc->idx;
>> +		perf_report_aux_output_id(event, hwc->idx);
>> +	}
>> +}
> 
> AFAICT you want a callback in x86_assign_hw_event(), is that so?
> 

Yes, or open-coded e.g.

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 8f71dd72ef95..46dac45298d1 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -1207,6 +1207,11 @@ static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader,
 	return n;
 }
 
+static inline bool report_aux_output_id(struct perf_event *event)
+{
+	return is_pebs_pt(event);
+}
+
 static inline void x86_assign_hw_event(struct perf_event *event,
 				struct cpu_hw_events *cpuc, int i)
 {
@@ -1217,6 +1222,9 @@ static inline void x86_assign_hw_event(struct perf_event *event,
 	hwc->last_cpu = smp_processor_id();
 	hwc->last_tag = ++cpuc->tags[i];
 
+	if (report_aux_output_id(event))
+		perf_report_aux_output_id(event, idx);
+
 	switch (hwc->idx) {
 	case INTEL_PMC_IDX_FIXED_BTS:
 	case INTEL_PMC_IDX_FIXED_VLBR:

Re: [PATCH 1/3] perf/x86: Add new event for AUX output counter index

[Peter Zijlstra]

On Wed, Jun 30, 2021 at 12:21:07PM +0300, Adrian Hunter wrote:
> On 25/06/21 4:15 pm, Peter Zijlstra wrote:
> > On Wed, Jun 09, 2021 at 05:20:53PM +0300, Adrian Hunter wrote:
> > 
> >> +static void intel_pmu_report_aux_output_id(struct perf_event *event)
> >> +{
> >> +	struct hw_perf_event *hwc = &event->hw;
> >> +
> >> +	/*
> >> +	 * So long as all PEBS-via-PT events for a recording session are
> >> +	 * scheduled together, then only changes to hwc->idx need be reported.
> >> +	 */
> >> +	if (hwc->idx != hwc->idx_reported) {
> >> +		hwc->idx_reported = hwc->idx;
> >> +		perf_report_aux_output_id(event, hwc->idx);
> >> +	}
> >> +}
> > 
> > AFAICT you want a callback in x86_assign_hw_event(), is that so?
> > 
> 
> Yes, or open-coded e.g.
> 
> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
> index 8f71dd72ef95..46dac45298d1 100644
> --- a/arch/x86/events/core.c
> +++ b/arch/x86/events/core.c
> @@ -1207,6 +1207,11 @@ static int collect_events(struct cpu_hw_events *cpuc, struct perf_event *leader,
>  	return n;
>  }
>  
> +static inline bool report_aux_output_id(struct perf_event *event)
> +{
> +	return is_pebs_pt(event);
> +}
> +
>  static inline void x86_assign_hw_event(struct perf_event *event,
>  				struct cpu_hw_events *cpuc, int i)
>  {
> @@ -1217,6 +1222,9 @@ static inline void x86_assign_hw_event(struct perf_event *event,
>  	hwc->last_cpu = smp_processor_id();
>  	hwc->last_tag = ++cpuc->tags[i];
>  
> +	if (report_aux_output_id(event))
> +		perf_report_aux_output_id(event, idx);
> +
>  	switch (hwc->idx) {
>  	case INTEL_PMC_IDX_FIXED_BTS:
>  	case INTEL_PMC_IDX_FIXED_VLBR:

Right, bit yuck, but I suppose it works. The alternative is something
like:

	static_call_cond(x86_pmu_assign)(event, idx);

but I'm not sure that's worth it, but it avoids stuffing even more intel
specific bits into the core.

[PATCH 2/3] perf tools: Add support for PERF_RECORD_AUX_OUTPUT_HW_ID

[Adrian Hunter]

The PERF_RECORD_AUX_OUTPUT_HW_ID event provides a way to match AUX output
data like Intel PT PEBS-via-PT back to the event that it came from, by
providing a hardware ID that is present in the AUX output.

Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Reviewed-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
---
 tools/include/uapi/linux/perf_event.h | 15 +++++++++++++++
 tools/lib/perf/include/perf/event.h   |  6 ++++++
 tools/perf/builtin-inject.c           |  4 +++-
 tools/perf/builtin-record.c           |  2 +-
 tools/perf/util/event.c               | 18 ++++++++++++++++++
 tools/perf/util/event.h               |  5 +++++
 tools/perf/util/machine.c             | 10 ++++++++++
 tools/perf/util/machine.h             |  2 ++
 tools/perf/util/session.c             |  5 +++++
 tools/perf/util/tool.h                |  1 +
 10 files changed, 66 insertions(+), 2 deletions(-)

diff --git a/tools/include/uapi/linux/perf_event.h b/tools/include/uapi/linux/perf_event.h
index f92880a15645..c89535de1ec8 100644
--- a/tools/include/uapi/linux/perf_event.h
+++ b/tools/include/uapi/linux/perf_event.h
@@ -1141,6 +1141,21 @@ enum perf_event_type {
 	 */
 	PERF_RECORD_TEXT_POKE			= 20,
 
+	/*
+	 * Data written to the AUX area by hardware due to aux_output, may need
+	 * to be matched to the event by an architecture-specific hardware ID.
+	 * This records the hardware ID, but requires sample_id to provide the
+	 * event ID. e.g. Intel PT uses this record to disambiguate PEBS-via-PT
+	 * records from multiple events.
+	 *
+	 * struct {
+	 *	struct perf_event_header	header;
+	 *	u64				hw_id;
+	 *	struct sample_id		sample_id;
+	 * };
+	 */
+	PERF_RECORD_AUX_OUTPUT_HW_ID		= 21,
+
 	PERF_RECORD_MAX,			/* non-ABI */
 };
 
diff --git a/tools/lib/perf/include/perf/event.h b/tools/lib/perf/include/perf/event.h
index 4d0c02ba3f7d..75ee385fb078 100644
--- a/tools/lib/perf/include/perf/event.h
+++ b/tools/lib/perf/include/perf/event.h
@@ -289,6 +289,11 @@ struct perf_record_itrace_start {
 	__u32			 tid;
 };
 
+struct perf_record_aux_output_hw_id {
+	struct perf_event_header header;
+	__u64			hw_id;
+};
+
 struct perf_record_thread_map_entry {
 	__u64			 pid;
 	char			 comm[16];
@@ -414,6 +419,7 @@ union perf_event {
 	struct perf_record_auxtrace_error	auxtrace_error;
 	struct perf_record_aux			aux;
 	struct perf_record_itrace_start		itrace_start;
+	struct perf_record_aux_output_hw_id	aux_output_hw_id;
 	struct perf_record_switch		context_switch;
 	struct perf_record_thread_map		thread_map;
 	struct perf_record_cpu_map		cpu_map;
diff --git a/tools/perf/builtin-inject.c b/tools/perf/builtin-inject.c
index 5d6f583e2cd3..8d7a0a0a4cb2 100644
--- a/tools/perf/builtin-inject.c
+++ b/tools/perf/builtin-inject.c
@@ -813,7 +813,8 @@ static int __cmd_inject(struct perf_inject *inject)
 		inject->tool.auxtrace_info  = perf_event__process_auxtrace_info;
 		inject->tool.auxtrace	    = perf_event__process_auxtrace;
 		inject->tool.aux	    = perf_event__drop_aux;
-		inject->tool.itrace_start   = perf_event__drop_aux,
+		inject->tool.itrace_start   = perf_event__drop_aux;
+		inject->tool.aux_output_hw_id = perf_event__drop_aux;
 		inject->tool.ordered_events = true;
 		inject->tool.ordering_requires_timestamps = true;
 		/* Allow space in the header for new attributes */
@@ -880,6 +881,7 @@ int cmd_inject(int argc, const char **argv)
 			.lost_samples	= perf_event__repipe,
 			.aux		= perf_event__repipe,
 			.itrace_start	= perf_event__repipe,
+			.aux_output_hw_id = perf_event__repipe,
 			.context_switch	= perf_event__repipe,
 			.throttle	= perf_event__repipe,
 			.unthrottle	= perf_event__repipe,
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index bc3dd379eb67..95f93e0799d3 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -1437,7 +1437,7 @@ static int record__synthesize(struct record *rec, bool tail)
 		goto out;
 
 	/* Synthesize id_index before auxtrace_info */
-	if (rec->opts.auxtrace_sample_mode) {
+	if (rec->opts.auxtrace_sample_mode || rec->opts.full_auxtrace) {
 		err = perf_event__synthesize_id_index(tool,
 						      process_synthesized_event,
 						      session->evlist, machine);
diff --git a/tools/perf/util/event.c b/tools/perf/util/event.c
index ac706304afe9..fe24801f8e9f 100644
--- a/tools/perf/util/event.c
+++ b/tools/perf/util/event.c
@@ -57,6 +57,7 @@ static const char *perf_event__names[] = {
 	[PERF_RECORD_BPF_EVENT]			= "BPF_EVENT",
 	[PERF_RECORD_CGROUP]			= "CGROUP",
 	[PERF_RECORD_TEXT_POKE]			= "TEXT_POKE",
+	[PERF_RECORD_AUX_OUTPUT_HW_ID]		= "AUX_OUTPUT_HW_ID",
 	[PERF_RECORD_HEADER_ATTR]		= "ATTR",
 	[PERF_RECORD_HEADER_EVENT_TYPE]		= "EVENT_TYPE",
 	[PERF_RECORD_HEADER_TRACING_DATA]	= "TRACING_DATA",
@@ -237,6 +238,14 @@ int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused,
 	return machine__process_itrace_start_event(machine, event);
 }
 
+int perf_event__process_aux_output_hw_id(struct perf_tool *tool __maybe_unused,
+					 union perf_event *event,
+					 struct perf_sample *sample __maybe_unused,
+					 struct machine *machine)
+{
+	return machine__process_aux_output_hw_id_event(machine, event);
+}
+
 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused,
 				     union perf_event *event,
 				     struct perf_sample *sample,
@@ -407,6 +416,12 @@ size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp)
 		       event->itrace_start.pid, event->itrace_start.tid);
 }
 
+size_t perf_event__fprintf_aux_output_hw_id(union perf_event *event, FILE *fp)
+{
+	return fprintf(fp, " hw_id: %#"PRI_lx64"\n",
+		       event->aux_output_hw_id.hw_id);
+}
+
 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp)
 {
 	bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT;
@@ -534,6 +549,9 @@ size_t perf_event__fprintf(union perf_event *event, struct machine *machine, FIL
 	case PERF_RECORD_TEXT_POKE:
 		ret += perf_event__fprintf_text_poke(event, machine, fp);
 		break;
+	case PERF_RECORD_AUX_OUTPUT_HW_ID:
+		ret += perf_event__fprintf_aux_output_hw_id(event, fp);
+		break;
 	default:
 		ret += fprintf(fp, "\n");
 	}
diff --git a/tools/perf/util/event.h b/tools/perf/util/event.h
index 19ad64f2bd83..95ffed66369c 100644
--- a/tools/perf/util/event.h
+++ b/tools/perf/util/event.h
@@ -330,6 +330,10 @@ int perf_event__process_itrace_start(struct perf_tool *tool,
 				     union perf_event *event,
 				     struct perf_sample *sample,
 				     struct machine *machine);
+int perf_event__process_aux_output_hw_id(struct perf_tool *tool,
+					 union perf_event *event,
+					 struct perf_sample *sample,
+					 struct machine *machine);
 int perf_event__process_switch(struct perf_tool *tool,
 			       union perf_event *event,
 			       struct perf_sample *sample,
@@ -397,6 +401,7 @@ size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp);
 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp);
 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp);
 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp);
+size_t perf_event__fprintf_aux_output_hw_id(union perf_event *event, FILE *fp);
 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp);
 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp);
 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp);
diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
index 3ff4936a15a4..dad2e89095d7 100644
--- a/tools/perf/util/machine.c
+++ b/tools/perf/util/machine.c
@@ -755,6 +755,14 @@ int machine__process_itrace_start_event(struct machine *machine __maybe_unused,
 	return 0;
 }
 
+int machine__process_aux_output_hw_id_event(struct machine *machine __maybe_unused,
+					    union perf_event *event)
+{
+	if (dump_trace)
+		perf_event__fprintf_aux_output_hw_id(event, stdout);
+	return 0;
+}
+
 int machine__process_switch_event(struct machine *machine __maybe_unused,
 				  union perf_event *event)
 {
@@ -2027,6 +2035,8 @@ int machine__process_event(struct machine *machine, union perf_event *event,
 		ret = machine__process_bpf(machine, event, sample); break;
 	case PERF_RECORD_TEXT_POKE:
 		ret = machine__process_text_poke(machine, event, sample); break;
+	case PERF_RECORD_AUX_OUTPUT_HW_ID:
+		ret = machine__process_aux_output_hw_id_event(machine, event); break;
 	default:
 		ret = -1;
 		break;
diff --git a/tools/perf/util/machine.h b/tools/perf/util/machine.h
index 7377ed6efdf1..a143087eeb47 100644
--- a/tools/perf/util/machine.h
+++ b/tools/perf/util/machine.h
@@ -124,6 +124,8 @@ int machine__process_aux_event(struct machine *machine,
 			       union perf_event *event);
 int machine__process_itrace_start_event(struct machine *machine,
 					union perf_event *event);
+int machine__process_aux_output_hw_id_event(struct machine *machine,
+					    union perf_event *event);
 int machine__process_switch_event(struct machine *machine,
 				  union perf_event *event);
 int machine__process_namespaces_event(struct machine *machine,
diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
index 0dbb4f2628f3..9ec78dea118d 100644
--- a/tools/perf/util/session.c
+++ b/tools/perf/util/session.c
@@ -504,6 +504,8 @@ void perf_tool__fill_defaults(struct perf_tool *tool)
 		tool->bpf = perf_event__process_bpf;
 	if (tool->text_poke == NULL)
 		tool->text_poke = perf_event__process_text_poke;
+	if (tool->aux_output_hw_id == NULL)
+		tool->aux_output_hw_id = perf_event__process_aux_output_hw_id;
 	if (tool->read == NULL)
 		tool->read = process_event_sample_stub;
 	if (tool->throttle == NULL)
@@ -995,6 +997,7 @@ static perf_event__swap_op perf_event__swap_ops[] = {
 	[PERF_RECORD_NAMESPACES]	  = perf_event__namespaces_swap,
 	[PERF_RECORD_CGROUP]		  = perf_event__cgroup_swap,
 	[PERF_RECORD_TEXT_POKE]		  = perf_event__text_poke_swap,
+	[PERF_RECORD_AUX_OUTPUT_HW_ID]	  = perf_event__all64_swap,
 	[PERF_RECORD_HEADER_ATTR]	  = perf_event__hdr_attr_swap,
 	[PERF_RECORD_HEADER_EVENT_TYPE]	  = perf_event__event_type_swap,
 	[PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
@@ -1549,6 +1552,8 @@ static int machines__deliver_event(struct machines *machines,
 		return tool->bpf(tool, event, sample, machine);
 	case PERF_RECORD_TEXT_POKE:
 		return tool->text_poke(tool, event, sample, machine);
+	case PERF_RECORD_AUX_OUTPUT_HW_ID:
+		return tool->aux_output_hw_id(tool, event, sample, machine);
 	default:
 		++evlist->stats.nr_unknown_events;
 		return -1;
diff --git a/tools/perf/util/tool.h b/tools/perf/util/tool.h
index bbbc0dcd461f..ef873f2cc38f 100644
--- a/tools/perf/util/tool.h
+++ b/tools/perf/util/tool.h
@@ -53,6 +53,7 @@ struct perf_tool {
 			lost_samples,
 			aux,
 			itrace_start,
+			aux_output_hw_id,
 			context_switch,
 			throttle,
 			unthrottle,
-- 
2.17.1

[PATCH 3/3] perf intel-pt: Add support for PERF_RECORD_AUX_OUTPUT_HW_ID

[Adrian Hunter]

Originally, software only supported redirecting at most one PEBS event to
Intel PT (PEBS-via-PT) because it was not able to differentiate one event
from another. To overcome that, add support for the
PERF_RECORD_AUX_OUTPUT_HW_ID side-band event.

Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
Reviewed-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reviewed-by: Andi Kleen <ak@linux.intel.com>
---
 tools/perf/Documentation/perf-intel-pt.txt |  7 +-
 tools/perf/util/intel-pt.c                 | 85 +++++++++++++++++++++-
 2 files changed, 87 insertions(+), 5 deletions(-)

diff --git a/tools/perf/Documentation/perf-intel-pt.txt b/tools/perf/Documentation/perf-intel-pt.txt
index 184ba62420f0..19f792876085 100644
--- a/tools/perf/Documentation/perf-intel-pt.txt
+++ b/tools/perf/Documentation/perf-intel-pt.txt
@@ -1144,7 +1144,12 @@ Recording is selected by using the aux-output config term e.g.
 
 	perf record -c 10000 -e '{intel_pt/branch=0/,cycles/aux-output/ppp}' uname
 
-Note that currently, software only supports redirecting at most one PEBS event.
+Originally, software only supported redirecting at most one PEBS event because it
+was not able to differentiate one event from another. To overcome that, more recent
+kernels and perf tools add support for the PERF_RECORD_AUX_OUTPUT_HW_ID side-band event.
+To check for the presence of that event in a PEBS-via-PT trace:
+
+	perf script -D --no-itrace | grep PERF_RECORD_AUX_OUTPUT_HW_ID
 
 To display PEBS events from the Intel PT trace, use the itrace 'o' option e.g.
 
diff --git a/tools/perf/util/intel-pt.c b/tools/perf/util/intel-pt.c
index 154a1077f22e..9e0539d3a79c 100644
--- a/tools/perf/util/intel-pt.c
+++ b/tools/perf/util/intel-pt.c
@@ -111,6 +111,7 @@ struct intel_pt {
 	u64 cbr_id;
 	u64 psb_id;
 
+	bool single_pebs;
 	bool sample_pebs;
 	struct evsel *pebs_evsel;
 
@@ -147,6 +148,14 @@ enum switch_state {
 	INTEL_PT_SS_EXPECTING_SWITCH_IP,
 };
 
+/* applicable_counters is 64-bits */
+#define INTEL_PT_MAX_PEBS 64
+
+struct intel_pt_pebs_event {
+	struct evsel *evsel;
+	u64 id;
+};
+
 struct intel_pt_queue {
 	struct intel_pt *pt;
 	unsigned int queue_nr;
@@ -188,6 +197,7 @@ struct intel_pt_queue {
 	u64 last_br_cyc_cnt;
 	unsigned int cbr_seen;
 	char insn[INTEL_PT_INSN_BUF_SZ];
+	struct intel_pt_pebs_event pebs[INTEL_PT_MAX_PEBS];
 };
 
 static void intel_pt_dump(struct intel_pt *pt __maybe_unused,
@@ -1976,15 +1986,13 @@ static void intel_pt_add_lbrs(struct branch_stack *br_stack,
 	}
 }
 
-static int intel_pt_synth_pebs_sample(struct intel_pt_queue *ptq)
+static int intel_pt_do_synth_pebs_sample(struct intel_pt_queue *ptq, struct evsel *evsel, u64 id)
 {
 	const struct intel_pt_blk_items *items = &ptq->state->items;
 	struct perf_sample sample = { .ip = 0, };
 	union perf_event *event = ptq->event_buf;
 	struct intel_pt *pt = ptq->pt;
-	struct evsel *evsel = pt->pebs_evsel;
 	u64 sample_type = evsel->core.attr.sample_type;
-	u64 id = evsel->core.id[0];
 	u8 cpumode;
 	u64 regs[8 * sizeof(sample.intr_regs.mask)];
 
@@ -2110,6 +2118,45 @@ static int intel_pt_synth_pebs_sample(struct intel_pt_queue *ptq)
 	return intel_pt_deliver_synth_event(pt, event, &sample, sample_type);
 }
 
+static int intel_pt_synth_single_pebs_sample(struct intel_pt_queue *ptq)
+{
+	struct intel_pt *pt = ptq->pt;
+	struct evsel *evsel = pt->pebs_evsel;
+	u64 id = evsel->core.id[0];
+
+	return intel_pt_do_synth_pebs_sample(ptq, evsel, id);
+}
+
+static int intel_pt_synth_pebs_sample(struct intel_pt_queue *ptq)
+{
+	const struct intel_pt_blk_items *items = &ptq->state->items;
+	struct intel_pt_pebs_event *pe;
+	struct intel_pt *pt = ptq->pt;
+	int err = -EINVAL;
+	int hw_id;
+
+	if (!items->has_applicable_counters || !items->applicable_counters) {
+		if (!pt->single_pebs)
+			pr_err("PEBS-via-PT record with no applicable_counters\n");
+		return intel_pt_synth_single_pebs_sample(ptq);
+	}
+
+	for_each_set_bit(hw_id, &items->applicable_counters, INTEL_PT_MAX_PEBS) {
+		pe = &ptq->pebs[hw_id];
+		if (!pe->evsel) {
+			if (!pt->single_pebs)
+				pr_err("PEBS-via-PT record with no matching event, hw_id %d\n",
+				       hw_id);
+			return intel_pt_synth_single_pebs_sample(ptq);
+		}
+		err = intel_pt_do_synth_pebs_sample(ptq, pe->evsel, pe->id);
+		if (err)
+			return err;
+	}
+
+	return err;
+}
+
 static int intel_pt_synth_error(struct intel_pt *pt, int code, int cpu,
 				pid_t pid, pid_t tid, u64 ip, u64 timestamp)
 {
@@ -2880,6 +2927,30 @@ static int intel_pt_process_itrace_start(struct intel_pt *pt,
 					event->itrace_start.tid);
 }
 
+static int intel_pt_process_aux_output_hw_id(struct intel_pt *pt,
+					     union perf_event *event,
+					     struct perf_sample *sample)
+{
+	u64 hw_id = event->aux_output_hw_id.hw_id;
+	struct auxtrace_queue *queue;
+	struct intel_pt_queue *ptq;
+	struct evsel *evsel;
+
+	queue = auxtrace_queues__sample_queue(&pt->queues, sample, pt->session);
+	evsel = evlist__id2evsel_strict(pt->session->evlist, sample->id);
+	if (!queue || !queue->priv || !evsel || hw_id > INTEL_PT_MAX_PEBS) {
+		pr_err("Bad AUX output hardware ID\n");
+		return -EINVAL;
+	}
+
+	ptq = queue->priv;
+
+	ptq->pebs[hw_id].evsel = evsel;
+	ptq->pebs[hw_id].id = sample->id;
+
+	return 0;
+}
+
 static int intel_pt_find_map(struct thread *thread, u8 cpumode, u64 addr,
 			     struct addr_location *al)
 {
@@ -3007,6 +3078,8 @@ static int intel_pt_process_event(struct perf_session *session,
 		err = intel_pt_process_switch(pt, sample);
 	else if (event->header.type == PERF_RECORD_ITRACE_START)
 		err = intel_pt_process_itrace_start(pt, event, sample);
+	else if (event->header.type == PERF_RECORD_AUX_OUTPUT_HW_ID)
+		err = intel_pt_process_aux_output_hw_id(pt, event, sample);
 	else if (event->header.type == PERF_RECORD_SWITCH ||
 		 event->header.type == PERF_RECORD_SWITCH_CPU_WIDE)
 		err = intel_pt_context_switch(pt, event, sample);
@@ -3391,9 +3464,13 @@ static void intel_pt_setup_pebs_events(struct intel_pt *pt)
 
 	evlist__for_each_entry(pt->session->evlist, evsel) {
 		if (evsel->core.attr.aux_output && evsel->core.id) {
+			if (pt->single_pebs) {
+				pt->single_pebs = false;
+				return;
+			}
+			pt->single_pebs = true;
 			pt->sample_pebs = true;
 			pt->pebs_evsel = evsel;
-			return;
 		}
 	}
 }
-- 
2.17.1

Re: [PATCH 0/3] perf intel-pt: Add PEBS-via-PT side-band

[Adrian Hunter]

On 9/06/21 5:20 pm, Adrian Hunter wrote:
> Hi
> 
> PEBS output to Intel Processor Trace was introduced with Atom
> processors based on Tremont.  Currently there is software support
> only for a single PEBS-via-PT event.
> 
> Here is support for multiple PEBS-via-PT events.

Any comments on this patch set?

> 
> The first patch is the kernel change which adds a new event,
> namely PERF_RECORD_AUX_OUTPUT_HW_ID which contains the counter
> index.  That is output when the PEBS-via-PT event is enabled in hardware.
> There is an optimization, to report the index only when it changes
> for the event.  That will work only so long as all PEBS-via-PT
> events are scheduled together, which they are for a recording
> session because they are in a single group.
> 
> Also no attribute bit is used to select the new event, so a new
> kernel is not compatible with older perf tools.  The assumption
> being that PEBS-via-PT is sufficiently esoteric that users will not
> be troubled by this.
> 
> The second patch adds the usual boiler plate to perf tools for
> a new event.
> 
> The third patch adds support for processing the new event by
> perf tool's Intel PT decoder.
> 
> 
> Adrian Hunter (3):
>       perf/x86: Add new event for AUX output counter index
>       perf tools: Add support for PERF_RECORD_AUX_OUTPUT_HW_ID
>       perf intel-pt: Add support for PERF_RECORD_AUX_OUTPUT_HW_ID
> 
>  arch/x86/events/core.c                     |  1 +
>  arch/x86/events/intel/ds.c                 | 16 ++++++
>  include/linux/perf_event.h                 |  2 +
>  include/uapi/linux/perf_event.h            | 15 ++++++
>  kernel/events/core.c                       | 30 +++++++++++
>  tools/include/uapi/linux/perf_event.h      | 15 ++++++
>  tools/lib/perf/include/perf/event.h        |  6 +++
>  tools/perf/Documentation/perf-intel-pt.txt |  7 ++-
>  tools/perf/builtin-inject.c                |  4 +-
>  tools/perf/builtin-record.c                |  2 +-
>  tools/perf/util/event.c                    | 18 +++++++
>  tools/perf/util/event.h                    |  5 ++
>  tools/perf/util/intel-pt.c                 | 85 ++++++++++++++++++++++++++++--
>  tools/perf/util/machine.c                  | 10 ++++
>  tools/perf/util/machine.h                  |  2 +
>  tools/perf/util/session.c                  |  5 ++
>  tools/perf/util/tool.h                     |  1 +
>  17 files changed, 217 insertions(+), 7 deletions(-)
> 
> 
> Regards
> Adrian
>