Skip to contentSkip to conversation

perf metrics: Support parsing metrics if platforms only have json table of system PMU

PATCH In review Junhao He 2 participants 2 messages Oct 10, 2024 – Oct 22, 2024 perf events

lore
Share & tools
lore

Junhao He

author
#0 lore raw

PATCH v2 1/1 perf metrics: Support parsing metrics if platforms only have json table of system PMU

The system PMUs don't depend on the certain CPUs and we don't need a CPUID to metric table mapping to match the json event for generating the metric table. For example HiSilicon HIP09 only have json events table of system PMUs in the "sys/" subdirectory.

Currently for this case the struct of system metric table "pmu_metrics__hisilicon_hip09_sys" generating works fine and metrics display as expected by using perf list. But perf stat doesn't work for such metrics.

  $ perf list metric
  Metrics:
    cpa_p0_avg_bw
         [Average bandwidth of CPA Port 0]
    cpa_p1_avg_bw
         [Average bandwidth of CPA Port 1]
  $ perf stat -M cpa_p0_avg_bw --timeout 1000 --> No error messages output
  $ echo $?
  234

The metricgroup__parse_groups() expects to find an cpu metric table, but the hisilicon/hip09 doesn't uses CPUID to map json events and metrics, so pmu_metrics_table__find() will return NULL, than the cmd run failed.

But in metricgroup__add_metric(), the function parse for each sys metric and add it to metric_list, which also will get an valid sys metric table. So, we can ignore the NULL result of pmu_metrics_table__find() and to use the sys metric table.

metricgroup__parse_groups
-> parse_groups

     -> metricgroup__add_metric_list
         -> metricgroup__add_metric
	     -> pmu_for_each_sys_metric   --> parse for each sys metric

Testing:

  $ perf stat -M cpa_p0_avg_bw --timeout 1000

Performance counter stats for 'system wide':

     4,004,863,602      cpa_cycles                #     0.00 cpa_p0_avg_bw
                 0      cpa_p0_wr_dat
                 0      cpa_p0_rd_dat_64b
                 0      cpa_p0_rd_dat_32b
       1.001306160 seconds time elapsed

---
v1 --> v2:
-Add some comments to explain the table is null.
-Modify the patch commit.
-Add Yicong Test-by.
v1:https://lore.kernel.org/all/[email protected]/

---
 tools/perf/util/metricgroup.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
tools/perf/util/metricgroup.c +8 -4
index 69f6a46402c3..cb428eabd485 100644
@@ -1123,7 +1123,7 @@ static int metricgroup__add_metric_sys_event_iter(const struct pmu_metric *pm,
 
 	ret = add_metric(d->metric_list, pm, d->modifier, d->metric_no_group,
 			 d->metric_no_threshold, d->user_requested_cpu_list,
-			 d->system_wide, d->root_metric, d->visited, d->table);
+			 d->system_wide, d->root_metric, d->visited, d->table ?: table);
 	if (ret)
 		goto out;
 
@@ -1226,7 +1226,8 @@ static int metricgroup__add_metric_callback(const struct pmu_metric *pm,
  * @system_wide: Are events for all processes recorded.
  * @metric_list: The list that the metric or metric group are added to.
  * @table: The table that is searched for metrics, most commonly the table for the
- *       architecture perf is running upon.
+ *       architecture perf is running upon. This value could be NULL if no core
+ *       metrics matches the architecture and we'll try to use the table of system PMUs.
  */
 static int metricgroup__add_metric(const char *pmu, const char *metric_name, const char *modifier,
 				   bool metric_no_group, bool metric_no_threshold,
@@ -1239,7 +1240,8 @@ static int metricgroup__add_metric(const char *pmu, const char *metric_name, con
 	int ret;
 	bool has_match = false;
 
-	{
+	/* Add core metrics to the metric list */
+	if (table) {
 		struct metricgroup__add_metric_data data = {
 			.list = &list,
 			.pmu = pmu,
@@ -1263,6 +1265,7 @@ static int metricgroup__add_metric(const char *pmu, const char *metric_name, con
 		has_match = data.has_match;
 	}
 	{
+		/* Parse metrics table of system PMUs */
 		struct metricgroup_iter_data data = {
 			.fn = metricgroup__add_metric_sys_event_iter,
 			.data = (void *) &(struct metricgroup_add_iter_data) {
@@ -1697,7 +1700,8 @@ int metricgroup__parse_groups(struct evlist *perf_evlist,
 	const struct pmu_metrics_table *table = pmu_metrics_table__find();
 
 	if (!table)
-		return -EINVAL;
+		pr_debug("The core metric table not found, continue to parse system metric table\n");
+
 	if (hardware_aware_grouping)
 		pr_debug("Use hardware aware grouping instead of traditional metric grouping method\n");
 
Signature · 1 line
2.33.0
maintainer
#1 lore raw

Hello,

Quoted 38 lines · Junhao He

The system PMUs don't depend on the certain CPUs and we don't need a CPUID to metric table mapping to match the json event for generating the metric table. For example HiSilicon HIP09 only have json events table of system PMUs in the "sys/" subdirectory.

Currently for this case the struct of system metric table "pmu_metrics__hisilicon_hip09_sys" generating works fine and metrics display as expected by using perf list. But perf stat doesn't work for such metrics.

  $ perf list metric
  Metrics:
    cpa_p0_avg_bw
         [Average bandwidth of CPA Port 0]
    cpa_p1_avg_bw
         [Average bandwidth of CPA Port 1]
  $ perf stat -M cpa_p0_avg_bw --timeout 1000 --> No error messages output
  $ echo $?
  234

The metricgroup__parse_groups() expects to find an cpu metric table, but the hisilicon/hip09 doesn't uses CPUID to map json events and metrics, so pmu_metrics_table__find() will return NULL, than the cmd run failed.

But in metricgroup__add_metric(), the function parse for each sys metric and add it to metric_list, which also will get an valid sys metric table. So, we can ignore the NULL result of pmu_metrics_table__find() and to use the sys metric table.

metricgroup__parse_groups
-> parse_groups

     -> metricgroup__add_metric_list
         -> metricgroup__add_metric
	     -> pmu_for_each_sys_metric   --> parse for each sys metric

Testing:

  $ perf stat -M cpa_p0_avg_bw --timeout 1000
Quoted 8 lines

Performance counter stats for 'system wide':

     4,004,863,602      cpa_cycles                #     0.00 cpa_p0_avg_bw
                 0      cpa_p0_wr_dat
                 0      cpa_p0_rd_dat_64b
                 0      cpa_p0_rd_dat_32b
       1.001306160 seconds time elapsed

Ian, can you please review this?

Thanks,
Namhyung

Quoted 67 lines

---
v1 --> v2:
-Add some comments to explain the table is null.
-Modify the patch commit.
-Add Yicong Test-by.
v1:https://lore.kernel.org/all/[email protected]/

---
 tools/perf/util/metricgroup.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)
tools/perf/util/metricgroup.c +8 -4
index 69f6a46402c3..cb428eabd485 100644
@@ -1123,7 +1123,7 @@ static int metricgroup__add_metric_sys_event_iter(const struct pmu_metric *pm,
 
 	ret = add_metric(d->metric_list, pm, d->modifier, d->metric_no_group,
 			 d->metric_no_threshold, d->user_requested_cpu_list,
-			 d->system_wide, d->root_metric, d->visited, d->table);
+			 d->system_wide, d->root_metric, d->visited, d->table ?: table);
 	if (ret)
 		goto out;
 
@@ -1226,7 +1226,8 @@ static int metricgroup__add_metric_callback(const struct pmu_metric *pm,
  * @system_wide: Are events for all processes recorded.
  * @metric_list: The list that the metric or metric group are added to.
  * @table: The table that is searched for metrics, most commonly the table for the
- *       architecture perf is running upon.
+ *       architecture perf is running upon. This value could be NULL if no core
+ *       metrics matches the architecture and we'll try to use the table of system PMUs.
  */
 static int metricgroup__add_metric(const char *pmu, const char *metric_name, const char *modifier,
 				   bool metric_no_group, bool metric_no_threshold,
@@ -1239,7 +1240,8 @@ static int metricgroup__add_metric(const char *pmu, const char *metric_name, con
 	int ret;
 	bool has_match = false;
 
-	{
+	/* Add core metrics to the metric list */
+	if (table) {
 		struct metricgroup__add_metric_data data = {
 			.list = &list,
 			.pmu = pmu,
@@ -1263,6 +1265,7 @@ static int metricgroup__add_metric(const char *pmu, const char *metric_name, con
 		has_match = data.has_match;
 	}
 	{
+		/* Parse metrics table of system PMUs */
 		struct metricgroup_iter_data data = {
 			.fn = metricgroup__add_metric_sys_event_iter,
 			.data = (void *) &(struct metricgroup_add_iter_data) {
@@ -1697,7 +1700,8 @@ int metricgroup__parse_groups(struct evlist *perf_evlist,
 	const struct pmu_metrics_table *table = pmu_metrics_table__find();
 
 	if (!table)
-		return -EINVAL;
+		pr_debug("The core metric table not found, continue to parse system metric table\n");
+
 	if (hardware_aware_grouping)
 		pr_debug("Use hardware aware grouping instead of traditional metric grouping method\n");
 
Signature · 1 line
2.33.0