Skip to contentSkip to conversation

perf pmu: Assume sysfs events are always lowercase

RFC PATCH In review Ian Rogers 2 participants 3 messages Apr 13, 2024 – Apr 18, 2024 perf events

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

RFC PATCH 1/1 perf pmu: Assume sysfs events are always lowercase

Perf event names aren't case sensitive. For sysfs events the entire directory of events is read then iterated comparing names in a case insensitive way, most often to see if an event is present.

Consider:

$ perf stat -e inst_retired.any true

The event inst_retired.any may be present in any PMU, so every PMU's sysfs events are loaded and then searched with strcasecmp to see if any match. This event is only present on the cpu PMU as a json event so a lot of events were loaded from sysfs unnecessarily just to prove an event didn't exist there.

This change avoids loading all the events by assuming sysfs event names are always lowercase. It then uses file exists and only loads the events when the desired event is present.

For the example above, the number of openat calls measured by perf trace on a tigerlake laptop goes from 325 down to 255. The reduction will be larger for machines with many PMUs, particularly replicated uncore PMUs.

Make pmu_aliases_parse early return when aliases are loaded, ensure the function is called before all uses of the aliases list.

---
 tools/perf/util/pmu.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)
tools/perf/util/pmu.c +15 -5
index ab30f22eaf10..ce72c99e4f61 100644
@@ -425,9 +425,16 @@ static struct perf_pmu_alias *perf_pmu__find_alias(struct perf_pmu *pmu,
 {
 	struct perf_pmu_alias *alias;
 
-	if (load && !pmu->sysfs_aliases_loaded)
-		pmu_aliases_parse(pmu);
+	if (load && !pmu->sysfs_aliases_loaded) {
+		char event_file_name[FILENAME_MAX + 8];
 
+		scnprintf(event_file_name, sizeof(event_file_name), "events/%s", name);
+		for (size_t i = 7, n = 7 + strlen(name); i < n; i++)
+			event_file_name[i] = tolower(event_file_name[i]);
+
+		if (perf_pmu__file_exists(pmu, event_file_name))
+			pmu_aliases_parse(pmu);
+	}
 	list_for_each_entry(alias, &pmu->aliases, list) {
 		if (!strcasecmp(alias->name, name))
 			return alias;
@@ -605,6 +612,9 @@ static int pmu_aliases_parse(struct perf_pmu *pmu)
 	size_t len;
 	int fd, dir_fd;
 
+	if (pmu->sysfs_aliases_loaded)
+		return 0;
+
 	len = perf_pmu__event_source_devices_scnprintf(path, sizeof(path));
 	if (!len)
 		return 0;
@@ -1689,9 +1699,7 @@ size_t perf_pmu__num_events(struct perf_pmu *pmu)
 {
 	size_t nr;
 
-	if (!pmu->sysfs_aliases_loaded)
-		pmu_aliases_parse(pmu);
-
+	pmu_aliases_parse(pmu);
 	nr = pmu->sysfs_aliases;
 
 	if (pmu->cpu_aliases_added)
@@ -1750,6 +1758,7 @@ int perf_pmu__for_each_event(struct perf_pmu *pmu, bool skip_duplicate_pmus,
 	struct strbuf sb;
 
 	strbuf_init(&sb, /*hint=*/ 0);
+	pmu_aliases_parse(pmu);
 	pmu_add_cpu_aliases(pmu);
 	list_for_each_entry(event, &pmu->aliases, list) {
 		size_t buf_used;
@@ -2154,6 +2163,7 @@ const char *perf_pmu__name_from_config(struct perf_pmu *pmu, u64 config)
 	if (!pmu)
 		return NULL;
 
+	pmu_aliases_parse(pmu);
 	pmu_add_cpu_aliases(pmu);
 	list_for_each_entry(event, &pmu->aliases, list) {
 		struct perf_event_attr attr = {.config = 0,};
Signature · 1 line
2.44.0.683.g7961c838ac-goog
#1 lore raw
Quoted 7 lines · Ian Rogers

Perf event names aren't case sensitive. For sysfs events the entire directory of events is read then iterated comparing names in a case insensitive way, most often to see if an event is present.

Consider:

$ perf stat -e inst_retired.any true
Quoted 8 lines

The event inst_retired.any may be present in any PMU, so every PMU's sysfs events are loaded and then searched with strcasecmp to see if any match. This event is only present on the cpu PMU as a json event so a lot of events were loaded from sysfs unnecessarily just to prove an event didn't exist there.

This change avoids loading all the events by assuming sysfs event names are always lowercase.

From what I searched in the kernel, it looks like all the sysfs event names (from different ARCHs and devices) are lowercase. It should not break the existing usages.
I don't see a reason why a event name must be uppercase. However, I think we need to add something to guarantee the assumption.

Could you please update the doc to describe the assumption? Documentation/ABI/testing/sysfs-bus-event_source-devices-events So everybody can follow the rule.

I think a perf test is required as well.
Maybe we can extend the test__pmu_events() to do the name check.

Thanks,
Kan

Quoted 76 lines

It then uses file exists and only loads
the events when the desired event is present.

For the example above, the number of openat calls measured by perf trace on a tigerlake laptop goes from 325 down to 255. The reduction will be larger for machines with many PMUs, particularly replicated uncore PMUs.

Make pmu_aliases_parse early return when aliases are loaded, ensure the function is called before all uses of the aliases list.

---
 tools/perf/util/pmu.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)
tools/perf/util/pmu.c +15 -5
index ab30f22eaf10..ce72c99e4f61 100644
@@ -425,9 +425,16 @@ static struct perf_pmu_alias *perf_pmu__find_alias(struct perf_pmu *pmu,
 {
 	struct perf_pmu_alias *alias;
 
-	if (load && !pmu->sysfs_aliases_loaded)
-		pmu_aliases_parse(pmu);
+	if (load && !pmu->sysfs_aliases_loaded) {
+		char event_file_name[FILENAME_MAX + 8];
 
+		scnprintf(event_file_name, sizeof(event_file_name), "events/%s", name);
+		for (size_t i = 7, n = 7 + strlen(name); i < n; i++)
+			event_file_name[i] = tolower(event_file_name[i]);
+
+		if (perf_pmu__file_exists(pmu, event_file_name))
+			pmu_aliases_parse(pmu);
+	}
 	list_for_each_entry(alias, &pmu->aliases, list) {
 		if (!strcasecmp(alias->name, name))
 			return alias;
@@ -605,6 +612,9 @@ static int pmu_aliases_parse(struct perf_pmu *pmu)
 	size_t len;
 	int fd, dir_fd;
 
+	if (pmu->sysfs_aliases_loaded)
+		return 0;
+
 	len = perf_pmu__event_source_devices_scnprintf(path, sizeof(path));
 	if (!len)
 		return 0;
@@ -1689,9 +1699,7 @@ size_t perf_pmu__num_events(struct perf_pmu *pmu)
 {
 	size_t nr;
 
-	if (!pmu->sysfs_aliases_loaded)
-		pmu_aliases_parse(pmu);
-
+	pmu_aliases_parse(pmu);
 	nr = pmu->sysfs_aliases;
 
 	if (pmu->cpu_aliases_added)
@@ -1750,6 +1758,7 @@ int perf_pmu__for_each_event(struct perf_pmu *pmu, bool skip_duplicate_pmus,
 	struct strbuf sb;
 
 	strbuf_init(&sb, /*hint=*/ 0);
+	pmu_aliases_parse(pmu);
 	pmu_add_cpu_aliases(pmu);
 	list_for_each_entry(event, &pmu->aliases, list) {
 		size_t buf_used;
@@ -2154,6 +2163,7 @@ const char *perf_pmu__name_from_config(struct perf_pmu *pmu, u64 config)
 	if (!pmu)
 		return NULL;
 
+	pmu_aliases_parse(pmu);
 	pmu_add_cpu_aliases(pmu);
 	list_for_each_entry(event, &pmu->aliases, list) {
 		struct perf_event_attr attr = {.config = 0,};
author reviewer
#2 lore raw
Quoted 26 lines · Kan
Quoted 7 lines · Ian Rogers

Perf event names aren't case sensitive. For sysfs events the entire directory of events is read then iterated comparing names in a case insensitive way, most often to see if an event is present.

Consider:

$ perf stat -e inst_retired.any true
Quoted 8 lines

The event inst_retired.any may be present in any PMU, so every PMU's sysfs events are loaded and then searched with strcasecmp to see if any match. This event is only present on the cpu PMU as a json event so a lot of events were loaded from sysfs unnecessarily just to prove an event didn't exist there.

This change avoids loading all the events by assuming sysfs event names are always lowercase.

From what I searched in the kernel, it looks like all the sysfs event names (from different ARCHs and devices) are lowercase. It should not break the existing usages.
I don't see a reason why a event name must be uppercase. However, I think we need to add something to guarantee the assumption.

Quoted 6 lines

Could you please update the doc to describe the assumption? Documentation/ABI/testing/sysfs-bus-event_source-devices-events So everybody can follow the rule.

I think a perf test is required as well.
Maybe we can extend the test__pmu_events() to do the name check.

Great suggestions, I'll do a v2 with this. I've been thinking for a while we need a PMU driver test. A particular problem is for weak groups, where broken drivers don't return perf event open failures for groups that exceed the number of counters. There's the enum value MetricNoGroupEvents to workaround this problem in metrics, but it would still be nice to get the drivers fixed. So rather than keep growing the event parsing tests, I think adding it to the pmu test and making the pmu test a fuller suite checking properties of PMUs would be good.

Thanks,
Ian

Quoted 79 lines

Thanks,
Kan

Quoted 76 lines

It then uses file exists and only loads
the events when the desired event is present.

For the example above, the number of openat calls measured by perf trace on a tigerlake laptop goes from 325 down to 255. The reduction will be larger for machines with many PMUs, particularly replicated uncore PMUs.

Make pmu_aliases_parse early return when aliases are loaded, ensure the function is called before all uses of the aliases list.

---
 tools/perf/util/pmu.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)
tools/perf/util/pmu.c +15 -5
index ab30f22eaf10..ce72c99e4f61 100644
@@ -425,9 +425,16 @@ static struct perf_pmu_alias *perf_pmu__find_alias(struct perf_pmu *pmu,
 {
      struct perf_pmu_alias *alias;
 
-     if (load && !pmu->sysfs_aliases_loaded)
-             pmu_aliases_parse(pmu);
+     if (load && !pmu->sysfs_aliases_loaded) {
+             char event_file_name[FILENAME_MAX + 8];
 
+             scnprintf(event_file_name, sizeof(event_file_name), "events/%s", name);
+             for (size_t i = 7, n = 7 + strlen(name); i < n; i++)
+                     event_file_name[i] = tolower(event_file_name[i]);
+
+             if (perf_pmu__file_exists(pmu, event_file_name))
+                     pmu_aliases_parse(pmu);
+     }
      list_for_each_entry(alias, &pmu->aliases, list) {
              if (!strcasecmp(alias->name, name))
                      return alias;
@@ -605,6 +612,9 @@ static int pmu_aliases_parse(struct perf_pmu *pmu)
      size_t len;
      int fd, dir_fd;
 
+     if (pmu->sysfs_aliases_loaded)
+             return 0;
+
      len = perf_pmu__event_source_devices_scnprintf(path, sizeof(path));
      if (!len)
              return 0;
@@ -1689,9 +1699,7 @@ size_t perf_pmu__num_events(struct perf_pmu *pmu)
 {
      size_t nr;
 
-     if (!pmu->sysfs_aliases_loaded)
-             pmu_aliases_parse(pmu);
-
+     pmu_aliases_parse(pmu);
      nr = pmu->sysfs_aliases;
 
      if (pmu->cpu_aliases_added)
@@ -1750,6 +1758,7 @@ int perf_pmu__for_each_event(struct perf_pmu *pmu, bool skip_duplicate_pmus,
      struct strbuf sb;
 
      strbuf_init(&sb, /*hint=*/ 0);
+     pmu_aliases_parse(pmu);
      pmu_add_cpu_aliases(pmu);
      list_for_each_entry(event, &pmu->aliases, list) {
              size_t buf_used;
@@ -2154,6 +2163,7 @@ const char *perf_pmu__name_from_config(struct perf_pmu *pmu, u64 config)
      if (!pmu)
              return NULL;
 
+     pmu_aliases_parse(pmu);
      pmu_add_cpu_aliases(pmu);
      list_for_each_entry(event, &pmu->aliases, list) {
              struct perf_event_attr attr = {.config = 0,};