perf pmus: Allow just core PMU scanning

Scanning all PMUs is expensive as all PMUs sysfs entries are loaded,
benchmarking shows more than 4x the cost:

```
$ perf bench internals pmu-scan -i 1000
Computing performance of sysfs PMU event scan for 1000 times
Average core PMU scanning took: 989.231 usec (+- 1.535 usec)
Average PMU scanning took: 4309.425 usec (+- 74.322 usec)
```

Add new perf_pmus__scan_core routine that scans just core
PMUs. Replace perf_pmus__scan calls with perf_pmus__scan_core when
non-core PMUs are being ignored.

Reviewed-by: Kan Liang <[e-mail]>
Signed-off-by: Ian Rogers <[e-mail]>
Cc: Adrian Hunter <[e-mail]>
Cc: Alexander Shishkin <[e-mail]>
Cc: Ali Saidi <[e-mail]>
Cc: Athira Rajeev <[e-mail]>
Cc: Dmitrii Dolgov <[e-mail]>
Cc: Huacai Chen <[e-mail]>
Cc: Ingo Molnar <[e-mail]>
Cc: James Clark <[e-mail]>
Cc: Jing Zhang <[e-mail]>
Cc: Jiri Olsa <[e-mail]>
Cc: John Garry <[e-mail]>
Cc: Kajol Jain <[e-mail]>
Cc: Kang Minchul <[e-mail]>
Cc: Leo Yan <[e-mail]>
Cc: Madhavan Srinivasan <[e-mail]>
Cc: Mark Rutland <[e-mail]>
Cc: Mike Leach <[e-mail]>
Cc: Ming Wang <[e-mail]>
Cc: Namhyung Kim <[e-mail]>
Cc: Peter Zijlstra <[e-mail]>
Cc: Ravi Bangoria <[e-mail]>
Cc: Rob Herring <[e-mail]>
Cc: Sandipan Das <[e-mail]>
Cc: Sean Christopherson <[e-mail]>
Cc: Suzuki Poulouse <[e-mail]>
Cc: Thomas Richter <[e-mail]>
Cc: Will Deacon <[e-mail]>
Cc: Xing Zhengjun <[e-mail]>
Cc: [e-mail]
Cc: [e-mail]
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Arnaldo Carvalho de Melo <[e-mail]>

tools/perf/arch/arm64/util/pmu.c

···
 {
 	struct perf_pmu *pmu = NULL;
 
-	while ((pmu = perf_pmus__scan(pmu))) {
-		if (!is_pmu_core(pmu->name))
-			continue;
-
+	while ((pmu = perf_pmus__scan_core(pmu))) {
 		/*
 		 * The cpumap should cover all CPUs. Otherwise, some CPUs may
 		 * not support some events or have different event IDs.

tools/perf/arch/x86/util/evlist.c

···
 			continue;
 		}
 
-		while ((pmu = perf_pmus__scan(pmu)) != NULL) {
+		while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 			struct perf_cpu_map *cpus;
 			struct evsel *evsel;
-
-			if (!pmu->is_core)
-				continue;
 
 			evsel = evsel__new(attrs + i);
 			if (evsel == NULL)

tools/perf/arch/x86/util/perf_regs.c

···
 		 * The same register set is supported among different hybrid PMUs.
 		 * Only check the first available one.
 		 */
-		while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-			if (pmu->is_core) {
-				type = pmu->type;
-				break;
-			}
+		while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
+			type = pmu->type;
+			break;
 		}
 		attr.config |= type << PERF_PMU_TYPE_SHIFT;
 	}

tools/perf/bench/pmu-scan.c

···
 	int nr_aliases;
 	int nr_formats;
 	int nr_caps;
+	bool is_core;
 };
 
 static const struct option options[] = {
···
 		r = results + nr_pmus;
 
 		r->name = strdup(pmu->name);
+		r->is_core = pmu->is_core;
 		r->nr_caps = pmu->nr_caps;
 
 		r->nr_aliases = 0;
···
 	return 0;
 }
 
-static int check_result(void)
+static int check_result(bool core_only)
 {
 	struct pmu_scan_result *r;
 	struct perf_pmu *pmu;
···
 
 	for (int i = 0; i < nr_pmus; i++) {
 		r = &results[i];
+		if (core_only && !r->is_core)
+			continue;
+
 		pmu = perf_pmus__find(r->name);
 		if (pmu == NULL) {
 			pr_err("Cannot find PMU %s\n", r->name);
···
 	struct timeval start, end, diff;
 	double time_average, time_stddev;
 	u64 runtime_us;
-	unsigned int i;
 	int ret;
 
 	init_stats(&stats);
···
 		return -1;
 	}
 
-	for (i = 0; i < iterations; i++) {
-		gettimeofday(&start, NULL);
-		perf_pmus__scan(NULL);
-		gettimeofday(&end, NULL);
+	for (int j = 0; j < 2; j++) {
+		bool core_only = (j == 0);
 
-		timersub(&end, &start, &diff);
-		runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
-		update_stats(&stats, runtime_us);
+		for (unsigned int i = 0; i < iterations; i++) {
+			gettimeofday(&start, NULL);
+			if (core_only)
+				perf_pmus__scan_core(NULL);
+			else
+				perf_pmus__scan(NULL);
+			gettimeofday(&end, NULL);
+			timersub(&end, &start, &diff);
+			runtime_us = diff.tv_sec * USEC_PER_SEC + diff.tv_usec;
+			update_stats(&stats, runtime_us);
 
-		ret = check_result();
-		perf_pmus__destroy();
-		if (ret < 0)
-			break;
+			ret = check_result(core_only);
+			perf_pmus__destroy();
+			if (ret < 0)
+				break;
+		}
+		time_average = avg_stats(&stats);
+		time_stddev = stddev_stats(&stats);
+		pr_info("  Average%s PMU scanning took: %.3f usec (+- %.3f usec)\n",
+			core_only ? " core" : "", time_average, time_stddev);
 	}
-
-	time_average = avg_stats(&stats);
-	time_stddev = stddev_stats(&stats);
-	pr_info("  Average PMU scanning took: %.3f usec (+- %.3f usec)\n",
-		time_average, time_stddev);
-
 	delete_result();
 	return 0;
 }

tools/perf/tests/pmu-events.c

···
 	struct perf_pmu *pmu = NULL;
 	unsigned long i;
 
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 		int count = 0;
-
-		if (!is_pmu_core(pmu->name))
-			continue;
 
 		if (list_empty(&pmu->format)) {
 			pr_debug2("skipping testing core PMU %s\n", pmu->name);

tools/perf/util/cputopo.c

···
 	if (!perf_pmus__has_hybrid())
 		return NULL;
 
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-		if (pmu->is_core)
-			nr++;
-	}
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL)
+		nr++;
+
 	if (nr == 0)
 		return NULL;
 
···
 		return NULL;
 
 	tp->nr = nr;
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-		if (!pmu->is_core)
-			continue;
-
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 		if (load_hybrid_node(&tp->nodes[i], pmu)) {
 			hybrid_topology__delete(tp);
 			return NULL;

tools/perf/util/header.c

···
 	 */
 	if (perf_pmus__has_hybrid()) {
 		pmu = NULL;
-		while ((pmu = perf_pmus__scan(pmu))) {
-			if (!pmu->is_core)
-				continue;
-
+		while ((pmu = perf_pmus__scan_core(pmu))) {
 			ret = __write_pmu_caps(ff, pmu, true);
 			if (ret < 0)
 				return ret;

tools/perf/util/mem-events.c

···
 		} else {
 			struct perf_pmu *pmu = NULL;
 
-			while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-				if (!pmu->is_core)
-					continue;
-
+			while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 				scnprintf(sysfs_name, sizeof(sysfs_name),
 					  e->sysfs_name, pmu->name);
 				e->supported |= perf_mem_event__supported(mnt, sysfs_name);
···
 	char sysfs_name[100];
 	struct perf_pmu *pmu = NULL;
 
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-		if (!pmu->is_core)
-			continue;
-
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 		scnprintf(sysfs_name, sizeof(sysfs_name), e->sysfs_name,
 			  pmu->name);
 		if (!perf_mem_event__supported(mnt, sysfs_name)) {
···
 				return -1;
 			}
 
-			while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-				if (!pmu->is_core)
-					continue;
+			while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 				rec_argv[i++] = "-e";
 				s = perf_mem_events__name(j, pmu->name);
 				if (s) {

tools/perf/util/parse-events.c

···
 	const char *config_name = get_config_name(head_config);
 	const char *metric_id = get_config_metric_id(head_config);
 
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
+	/* Legacy cache events are only supported by core PMUs. */
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 		LIST_HEAD(config_terms);
 		struct perf_event_attr attr;
 		int ret;
-
-		/* Skip unsupported PMUs. */
-		if (!perf_pmu__supports_legacy_cache(pmu))
-			continue;
 
 		if (parse_events__filter_pmu(parse_state, pmu))
 			continue;
···
 		return __parse_events_add_numeric(parse_state, list, /*pmu=*/NULL,
 						  type, config, head_config);
 
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
+	/* Wildcards on numeric values are only supported by core PMUs. */
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 		int ret;
-
-		if (!perf_pmu__supports_wildcard_numeric(pmu))
-			continue;
 
 		if (parse_events__filter_pmu(parse_state, pmu))
 			continue;

tools/perf/util/pmu.c

···
 	return pmu->is_core;
 }
 
-bool perf_pmu__supports_wildcard_numeric(const struct perf_pmu *pmu)
-{
-	return pmu->is_core;
-}
-
 bool perf_pmu__auto_merge_stats(const struct perf_pmu *pmu)
 {
 	return !is_pmu_hybrid(pmu->name);
-}
-
-bool perf_pmu__is_mem_pmu(const struct perf_pmu *pmu)
-{
-	return pmu->is_core;
 }
 
 bool perf_pmu__have_event(const struct perf_pmu *pmu, const char *name)

tools/perf/util/pmu.h

···
 bool is_pmu_core(const char *name);
 bool is_pmu_hybrid(const char *name);
 bool perf_pmu__supports_legacy_cache(const struct perf_pmu *pmu);
-bool perf_pmu__supports_wildcard_numeric(const struct perf_pmu *pmu);
 bool perf_pmu__auto_merge_stats(const struct perf_pmu *pmu);
-bool perf_pmu__is_mem_pmu(const struct perf_pmu *pmu);
 bool perf_pmu__have_event(const struct perf_pmu *pmu, const char *name);
 
 FILE *perf_pmu__open_file(struct perf_pmu *pmu, const char *name);

tools/perf/util/pmus.c

···
 }
 
 /* Add all pmus in sysfs to pmu list: */
-static void pmu_read_sysfs(void)
+static void pmu_read_sysfs(bool core_only)
 {
 	int fd;
 	DIR *dir;
···
 
 	while ((dent = readdir(dir))) {
 		if (!strcmp(dent->d_name, ".") || !strcmp(dent->d_name, ".."))
+			continue;
+		if (core_only && !is_pmu_core(dent->d_name))
 			continue;
 		/* add to static LIST_HEAD(core_pmus) or LIST_HEAD(other_pmus): */
 		perf_pmu__find2(fd, dent->d_name);
···
 	bool use_core_pmus = !pmu || pmu->is_core;
 
 	if (!pmu) {
-		pmu_read_sysfs();
+		pmu_read_sysfs(/*core_only=*/false);
 		pmu = list_prepare_entry(pmu, &core_pmus, list);
 	}
 	if (use_core_pmus) {
···
 	}
 	list_for_each_entry_continue(pmu, &other_pmus, list)
 		return pmu;
+	return NULL;
+}
+
+struct perf_pmu *perf_pmus__scan_core(struct perf_pmu *pmu)
+{
+	if (!pmu) {
+		pmu_read_sysfs(/*core_only=*/true);
+		pmu = list_prepare_entry(pmu, &core_pmus, list);
+	}
+	list_for_each_entry_continue(pmu, &core_pmus, list)
+		return pmu;
+
 	return NULL;
 }
 
···
 	struct perf_pmu *pmu = NULL;
 	int count = 0;
 
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-		if (perf_pmu__is_mem_pmu(pmu))
-			count++;
-	}
+	/* All core PMUs are for mem events. */
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL)
+		count++;
+
 	return count;
 }
 
···
 	if (!hybrid_scanned) {
 		struct perf_pmu *pmu = NULL;
 
-		while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-			if (pmu->is_core && is_pmu_hybrid(pmu->name)) {
+		while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
+			if (is_pmu_hybrid(pmu->name)) {
 				has_hybrid = true;
 				break;
 			}

tools/perf/util/pmus.h

···
 struct perf_pmu *perf_pmus__find_by_type(unsigned int type);
 
 struct perf_pmu *perf_pmus__scan(struct perf_pmu *pmu);
+struct perf_pmu *perf_pmus__scan_core(struct perf_pmu *pmu);
 
 const struct perf_pmu *perf_pmus__pmu_for_pmu_filter(const char *str);
 

tools/perf/util/print-events.c

···
 	struct perf_pmu *pmu = NULL;
 	const char *event_type_descriptor = event_type_descriptors[PERF_TYPE_HW_CACHE];
 
-	while ((pmu = perf_pmus__scan(pmu)) != NULL) {
-		/*
-		 * Skip uncore PMUs for performance. PERF_TYPE_HW_CACHE type
-		 * attributes can accept software PMUs in the extended type, so
-		 * also skip.
-		 */
+	/*
+	 * Only print core PMUs, skipping uncore for performance and
+	 * PERF_TYPE_SOFTWARE that can succeed in opening legacy cache evenst.
+	 */
+	while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
 		if (pmu->is_uncore || pmu->type == PERF_TYPE_SOFTWARE)
 			continue;