Skip to contentSkip to conversation

perf stat fixes and improvements

PATCH Superseded Ian Rogers 2 participants 11 messages Nov 6, 2025 perf events

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

v3 cover perf stat fixes and improvements

A collection of fixes aiming to stabilize and make more reasonable measurements/metrics such as memory bandwidth.

Tool events are changed from getting a PMU cpu map of all online CPUs to either CPU 0 or all online CPUs. This avoids iterating over useless CPUs for events in particular duration_time. Fix a bug where duration_time didn't correct use the previous raw counts and would skip values in interval mode.

Change how json metrics handle tool events. Use the counter value rather than using shared state with perf stat. A later patch changes it so that tool events are read last, so that if reading say memory bandwidth counters you don't divide by an earlier read time and exceed the theoretical maximum memory bandwidth.

Do some clean up around the shared state in stat-shadow that's no longer used. It can be fully removed when the legacy json metrics patch series lands:
https://lore.kernel.org/lkml/[email protected]/

Change how affinities work with evlist__for_each_cpu. Move the affinity code into the iterator to simplify setting it up. Detect when affinities will and won't be profitable, for example a tool event and a regular perf event (or read group) may face less delay from a single IPI for the event read than from a call to sched_setaffinity. Add a
--no-affinity flag to perf stat to allow affinities to be disabled.

v3: Add affinity clean ups and read tool events last.

v2: Fixed an aggregation index issue:
https://lore.kernel.org/lkml/[email protected]/

v1:
https://lore.kernel.org/lkml/[email protected]/

Ian Rogers (9):
libperf cpumap: Reduce allocations and sorting in intersect perf pmu: perf_cpu_map__new_int to avoid parsing a string perf tool_pmu: Use old_count when computing count values for time
events
perf stat-shadow: Read tool events directly
perf stat: Reduce scope of ru_stats
perf tool_pmu: More accurately set the cpus for tool events perf evlist: Reduce affinity use and move into iterator, fix no
affinity
perf stat: Read tool events last
perf stat: Add no-affinity flag

 tools/lib/perf/cpumap.c                |  29 ++--
 tools/perf/Documentation/perf-stat.txt |   4 +
 tools/perf/builtin-stat.c              | 189 +++++++++++++++----------
 tools/perf/util/config.c               |   1 -
 tools/perf/util/drm_pmu.c              |   2 +-
 tools/perf/util/evlist.c               | 156 ++++++++++++--------
 tools/perf/util/evlist.h               |  27 +++-
 tools/perf/util/hwmon_pmu.c            |   2 +-
 tools/perf/util/parse-events.c         |   9 +-
 tools/perf/util/pmu.c                  |  12 ++
 tools/perf/util/pmu.h                  |   1 +
 tools/perf/util/stat-shadow.c          | 149 +++++++++----------
 tools/perf/util/stat.h                 |  16 ---
 tools/perf/util/tool_pmu.c             |  78 ++++++----
 tools/perf/util/tool_pmu.h             |   1 +
 15 files changed, 395 insertions(+), 281 deletions(-)
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#1 lore raw

PATCH v3 1/9 libperf cpumap: Reduce allocations and sorting in intersect

On hybrid platforms the CPU maps are often disjoint. Rather than copy CPUs and trim, compute the number of common CPUs, if none early exit, otherwise copy in an sorted order. This avoids memory allocation in the disjoint case and avoids a second malloc and useless sort in the previous trim cases.

---
 tools/lib/perf/cpumap.c | 29 +++++++++++++++++++----------
 1 file changed, 19 insertions(+), 10 deletions(-)
tools/lib/perf/cpumap.c +19 -10
index b20a5280f2b3..7e88417ba84d 100644
@@ -453,21 +453,33 @@ int perf_cpu_map__merge(struct perf_cpu_map **orig, struct perf_cpu_map *other)
 struct perf_cpu_map *perf_cpu_map__intersect(struct perf_cpu_map *orig,
 					     struct perf_cpu_map *other)
 {
-	struct perf_cpu *tmp_cpus;
-	int tmp_len;
 	int i, j, k;
-	struct perf_cpu_map *merged = NULL;
+	struct perf_cpu_map *merged;
 
 	if (perf_cpu_map__is_subset(other, orig))
 		return perf_cpu_map__get(orig);
 	if (perf_cpu_map__is_subset(orig, other))
 		return perf_cpu_map__get(other);
 
-	tmp_len = max(__perf_cpu_map__nr(orig), __perf_cpu_map__nr(other));
-	tmp_cpus = malloc(tmp_len * sizeof(struct perf_cpu));
-	if (!tmp_cpus)
+	i = j = k = 0;
+	while (i < __perf_cpu_map__nr(orig) && j < __perf_cpu_map__nr(other)) {
+		if (__perf_cpu_map__cpu(orig, i).cpu < __perf_cpu_map__cpu(other, j).cpu)
+			i++;
+		else if (__perf_cpu_map__cpu(orig, i).cpu > __perf_cpu_map__cpu(other, j).cpu)
+			j++;
+		else { /* CPUs match. */
+			i++;
+			j++;
+			k++;
+		}
+	}
+	if (k == 0) /* Maps are completely disjoint. */
 		return NULL;
 
+	merged = perf_cpu_map__alloc(k);
+	if (!merged)
+		return NULL;
+	/* Entries are added to merged in sorted order, so no need to sort again. */
 	i = j = k = 0;
 	while (i < __perf_cpu_map__nr(orig) && j < __perf_cpu_map__nr(other)) {
 		if (__perf_cpu_map__cpu(orig, i).cpu < __perf_cpu_map__cpu(other, j).cpu)
@@ -476,11 +488,8 @@ struct perf_cpu_map *perf_cpu_map__intersect(struct perf_cpu_map *orig,
 			j++;
 		else {
 			j++;
-			tmp_cpus[k++] = __perf_cpu_map__cpu(orig, i++);
+			RC_CHK_ACCESS(merged)->map[k++] = __perf_cpu_map__cpu(orig, i++);
 		}
 	}
-	if (k)
-		merged = cpu_map__trim_new(k, tmp_cpus);
-	free(tmp_cpus);
 	return merged;
 }
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#2 lore raw

PATCH v3 2/9 perf pmu: perf_cpu_map__new_int to avoid parsing a string

In reply to Ian Rogers

Prefer perf_cpu_map__new_int(0) to perf_cpu_map__new("0") as it avoids strings parsing.

---
 tools/perf/util/drm_pmu.c   | 2 +-
 tools/perf/util/hwmon_pmu.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
tools/perf/util/drm_pmu.c +1 -1
index 98d4d2b556d4..fc864bde4cb2 100644
@@ -119,7 +119,7 @@ static struct drm_pmu *add_drm_pmu(struct list_head *pmus, char *line, size_t li
 		return NULL;
 	}
 
-	drm->pmu.cpus = perf_cpu_map__new("0");
+	drm->pmu.cpus = perf_cpu_map__new_int(0);
 	if (!drm->pmu.cpus) {
 		perf_pmu__delete(&drm->pmu);
 		return NULL;
tools/perf/util/hwmon_pmu.c +1 -1
index 5c27256a220a..279d6b1a47f0 100644
@@ -376,7 +376,7 @@ struct perf_pmu *hwmon_pmu__new(struct list_head *pmus, const char *hwmon_dir,
 		perf_pmu__delete(&hwm->pmu);
 		return NULL;
 	}
-	hwm->pmu.cpus = perf_cpu_map__new("0");
+	hwm->pmu.cpus = perf_cpu_map__new_int(0);
 	if (!hwm->pmu.cpus) {
 		perf_pmu__delete(&hwm->pmu);
 		return NULL;
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#3 lore raw

PATCH v3 3/9 perf tool_pmu: Use old_count when computing count values for time events

In reply to Ian Rogers

When running in interval mode every third count of a time event isn't showing properly:

```
$ perf stat -e duration_time -a -I 1000
     1.001082862      1,002,290,425      duration_time
     2.004264262      1,003,183,516      duration_time
     3.007381401      <not counted>      duration_time
     4.011160141      1,003,705,631      duration_time
     5.014515385      1,003,290,110      duration_time
     6.018539680      <not counted>      duration_time
     7.022065321      1,003,591,720      duration_time
```

The regression came in with a different fix, found through bisection, commit 68cb1567439f ("perf tool_pmu: Fix aggregation on duration_time"). The issue is caused by the enabled and running time of the event matching the old_count's and creating a delta of 0, which is indicative of an error.

Fixes 68cb1567439f ("perf tool_pmu: Fix aggregation on duration_time")
---
 tools/perf/builtin-stat.c  | 16 +++++++++--
 tools/perf/util/tool_pmu.c | 59 +++++++++++++++++++++-----------------
 2 files changed, 46 insertions(+), 29 deletions(-)
tools/perf/builtin-stat.c +13 -3
index 886727ae8529..d89fa9468f89 100644
@@ -282,17 +282,27 @@ static int read_single_counter(struct evsel *counter, int cpu_map_idx, int threa
 	if (err && cpu_map_idx == 0 &&
 	    (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME ||
 	     evsel__tool_event(counter) == TOOL_PMU__EVENT_SYSTEM_TIME)) {
-		u64 val, *start_time;
 		struct perf_counts_values *count =
 			perf_counts(counter->counts, cpu_map_idx, thread);
+		struct perf_counts_values *old_count = NULL;
+		u64 val;
+
+		if (counter->prev_raw_counts)
+			old_count = perf_counts(counter->prev_raw_counts, cpu_map_idx, thread);
 
-		start_time = xyarray__entry(counter->start_times, cpu_map_idx, thread);
 		if (evsel__tool_event(counter) == TOOL_PMU__EVENT_USER_TIME)
 			val = ru_stats.ru_utime_usec_stat.mean;
 		else
 			val = ru_stats.ru_stime_usec_stat.mean;
-		count->ena = count->run = *start_time + val;
+
 		count->val = val;
+		if (old_count) {
+			count->run = old_count->run + 1;
+			count->ena = old_count->ena + 1;
+		} else {
+			count->run++;
+			count->ena++;
+		}
 		return 0;
 	}
 	return err;
tools/perf/util/tool_pmu.c +33 -26

101 lines not on this page · raw message

Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#4 lore raw

PATCH v3 4/9 perf stat-shadow: Read tool events directly

In reply to Ian Rogers

When reading time values for metrics don't use the globals updated in builtin-stat, just read the events as regular events. The only exception is for time events where nanoseconds need converting to seconds as metrics assume time metrics are in seconds.

---
 tools/perf/util/stat-shadow.c | 147 +++++++++++++++-------------------
 1 file changed, 66 insertions(+), 81 deletions(-)
tools/perf/util/stat-shadow.c +66 -81
index abaf6b579bfc..939ac3269a44 100644
@@ -371,7 +371,32 @@ static void print_nsecs(struct perf_stat_config *config,
 	}
 }
 
-static int prepare_metric(const struct metric_expr *mexp,
+static bool tool_pmu__is_time_event(const struct perf_stat_config *config,
+				    const struct evsel *evsel, int *tool_aggr_idx)
+{
+	enum tool_pmu_event event = evsel__tool_event(evsel);
+	int aggr_idx;
+
+	if (event != TOOL_PMU__EVENT_DURATION_TIME &&
+	    event != TOOL_PMU__EVENT_USER_TIME &&
+	    event != TOOL_PMU__EVENT_SYSTEM_TIME)
+		return false;
+
+	if (config) {
+		cpu_aggr_map__for_each_idx(aggr_idx, config->aggr_map) {
+			if (config->aggr_map->map[aggr_idx].cpu.cpu == 0) {
+				*tool_aggr_idx = aggr_idx;
+				return true;
+			}
+		}
+		pr_debug("Unexpected CPU0 missing in aggregation for tool event.\n");
+	}
+	*tool_aggr_idx = 0; /* Assume the first aggregation index works. */
+	return true;
+}
+
+static int prepare_metric(const struct perf_stat_config *config,
+			  const struct metric_expr *mexp,
 			  const struct evsel *evsel,
 			  struct expr_parse_ctx *pctx,
 			  int aggr_idx)
@@ -381,91 +406,51 @@ static int prepare_metric(const struct metric_expr *mexp,
 	int i;
 
 	for (i = 0; metric_events[i]; i++) {
+		int source_count = 0, tool_aggr_idx;
+		bool is_tool_time =
+			tool_pmu__is_time_event(config, metric_events[i], &tool_aggr_idx);
+		struct perf_stat_evsel *ps = metric_events[i]->stats;
+		struct perf_stat_aggr *aggr;
 		char *n;
 		double val;
-		int source_count = 0;
 
-		if (evsel__is_tool(metric_events[i])) {
-			struct stats *stats;
-			double scale;
-
-			switch (evsel__tool_event(metric_events[i])) {
-			case TOOL_PMU__EVENT_DURATION_TIME:
-				stats = &walltime_nsecs_stats;
-				scale = 1e-9;
-				break;
-			case TOOL_PMU__EVENT_USER_TIME:
-				stats = &ru_stats.ru_utime_usec_stat;
-				scale = 1e-6;
-				break;
-			case TOOL_PMU__EVENT_SYSTEM_TIME:
-				stats = &ru_stats.ru_stime_usec_stat;
-				scale = 1e-6;
+		/*
+		 * If there are multiple uncore PMUs and we're not reading the
+		 * leader's stats, determine the stats for the appropriate
+		 * uncore PMU.
+		 */
+		if (evsel && evsel->metric_leader &&
+		    evsel->pmu != evsel->metric_leader->pmu &&
+		    mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
+			struct evsel *pos;
+
+			evlist__for_each_entry(evsel->evlist, pos) {
+				if (pos->pmu != evsel->pmu)
+					continue;
+				if (pos->metric_leader != mexp->metric_events[i])
+					continue;
+				ps = pos->stats;
+				source_count = 1;
 				break;
-			case TOOL_PMU__EVENT_NONE:
-				pr_err("Invalid tool event 'none'");
-				abort();
-			case TOOL_PMU__EVENT_MAX:
-				pr_err("Invalid tool event 'max'");
-				abort();
-			case TOOL_PMU__EVENT_HAS_PMEM:
-			case TOOL_PMU__EVENT_NUM_CORES:
-			case TOOL_PMU__EVENT_NUM_CPUS:
-			case TOOL_PMU__EVENT_NUM_CPUS_ONLINE:
-			case TOOL_PMU__EVENT_NUM_DIES:
-			case TOOL_PMU__EVENT_NUM_PACKAGES:
-			case TOOL_PMU__EVENT_SLOTS:
-			case TOOL_PMU__EVENT_SMT_ON:
-			case TOOL_PMU__EVENT_SYSTEM_TSC_FREQ:
-			default:
-				pr_err("Unexpected tool event '%s'", evsel__name(metric_events[i]));
-				abort();
 			}
-			val = avg_stats(stats) * scale;
-			source_count = 1;
-		} else {
-			struct perf_stat_evsel *ps = metric_events[i]->stats;
-			struct perf_stat_aggr *aggr;
-
+		}
+		/* Time events are always on CPU0, the first aggregation index. */
+		aggr = &ps->aggr[is_tool_time ? tool_aggr_idx : aggr_idx];
+		if (!aggr || !metric_events[i]->supported) {
 			/*
-			 * If there are multiple uncore PMUs and we're not
-			 * reading the leader's stats, determine the stats for
-			 * the appropriate uncore PMU.
+			 * Not supported events will have a count of 0, which
+			 * can be confusing in a metric. Explicitly set the
+			 * value to NAN. Not counted events (enable time of 0)
+			 * are read as 0.
 			 */
-			if (evsel && evsel->metric_leader &&
-			    evsel->pmu != evsel->metric_leader->pmu &&
-			    mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
-				struct evsel *pos;
-
-				evlist__for_each_entry(evsel->evlist, pos) {
-					if (pos->pmu != evsel->pmu)
-						continue;
-					if (pos->metric_leader != mexp->metric_events[i])
-						continue;
-					ps = pos->stats;
-					source_count = 1;
-					break;
-				}
-			}
-			aggr = &ps->aggr[aggr_idx];
-			if (!aggr)
-				break;
-
-			if (!metric_events[i]->supported) {
-				/*
-				 * Not supported events will have a count of 0,
-				 * which can be confusing in a
-				 * metric. Explicitly set the value to NAN. Not
-				 * counted events (enable time of 0) are read as
-				 * 0.
-				 */
-				val = NAN;
-				source_count = 0;
-			} else {
-				val = aggr->counts.val;
-				if (!source_count)
-					source_count = evsel__source_count(metric_events[i]);
-			}
+			val = NAN;
+			source_count = 0;
+		} else {
+			val = aggr->counts.val;
+			if (is_tool_time)
+				val *= 1e-9; /* Convert time event nanoseconds to seconds. */
+			if (!source_count)
+				source_count = evsel__source_count(metric_events[i]);
 		}
 		n = strdup(evsel__metric_id(metric_events[i]));
 		if (!n)
@@ -511,7 +496,7 @@ static void generic_metric(struct perf_stat_config *config,
 		pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
 	pctx->sctx.runtime = runtime;
 	pctx->sctx.system_wide = config->system_wide;
-	i = prepare_metric(mexp, evsel, pctx, aggr_idx);
+	i = prepare_metric(config, mexp, evsel, pctx, aggr_idx);
 	if (i < 0) {
 		expr__ctx_free(pctx);
 		return;
@@ -572,7 +557,7 @@ double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
 	if (!pctx)
 		return NAN;
 
-	if (prepare_metric(mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
+	if (prepare_metric(/*config=*/NULL, mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
 		goto out;
 
 	if (expr__parse(&ratio, pctx, mexp->metric_expr))
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#5 lore raw

PATCH v3 5/9 perf stat: Reduce scope of ru_stats

In reply to Ian Rogers

The ru_stats are used to capture user and system time stats when a process exits. These are then applied to user and system time tool events if their reads fail due to the process terminating. Reduce the scope now the metric code no longer reads these values.

---
 tools/perf/builtin-stat.c     | 14 +++++++++++++-
 tools/perf/util/config.c      |  1 -
 tools/perf/util/stat-shadow.c |  2 --
 tools/perf/util/stat.h        | 16 ----------------
 4 files changed, 13 insertions(+), 20 deletions(-)
tools/perf/builtin-stat.c +13 -1
index d89fa9468f89..3cd663b3b357 100644
@@ -132,6 +132,7 @@ static bool			interval_count;
 static const char		*output_name;
 static int			output_fd;
 static char			*metrics;
+static struct rusage_stats	ru_stats;
 
 struct perf_stat {
 	bool			 record;
@@ -729,6 +730,17 @@ static int create_perf_stat_counter(struct evsel *evsel,
 					      evsel->core.threads);
 }
 
+static void update_rusage_stats(const struct rusage *rusage)
+{
+	const u64 us_to_ns = 1000;
+	const u64 s_to_ns = 1000000000;
+
+	update_stats(&ru_stats.ru_utime_usec_stat,
+		(rusage->ru_utime.tv_usec * us_to_ns + rusage->ru_utime.tv_sec * s_to_ns));
+	update_stats(&ru_stats.ru_stime_usec_stat,
+		(rusage->ru_stime.tv_usec * us_to_ns + rusage->ru_stime.tv_sec * s_to_ns));
+}
+
 static int __run_perf_stat(int argc, const char **argv, int run_idx)
 {
 	int interval = stat_config.interval;
@@ -978,7 +990,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 		evlist__reset_aggr_stats(evsel_list);
 	} else {
 		update_stats(&walltime_nsecs_stats, t1 - t0);
-		update_rusage_stats(&ru_stats, &stat_config.ru_data);
+		update_rusage_stats(&stat_config.ru_data);
 	}
 
 	/*
tools/perf/util/config.c +0 -1
index 6f914620c6ff..cc0746f494f4 100644
@@ -45,7 +45,6 @@ struct perf_stat_config stat_config = {
 	.run_count		= 1,
 	.metric_only_len	= METRIC_ONLY_LEN,
 	.walltime_nsecs_stats	= &walltime_nsecs_stats,
-	.ru_stats		= &ru_stats,
 	.big_num		= true,
 	.ctl_fd			= -1,
 	.ctl_fd_ack		= -1,
tools/perf/util/stat-shadow.c +0 -2
index 939ac3269a44..c59578886b4f 100644
@@ -18,7 +18,6 @@
 #include "tool_pmu.h"
 
 struct stats walltime_nsecs_stats;
-struct rusage_stats ru_stats;
 
 enum {
 	CTX_BIT_USER	= 1 << 0,
@@ -74,7 +73,6 @@ static int evsel_context(const struct evsel *evsel)
 void perf_stat__reset_shadow_stats(void)
 {
 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
-	memset(&ru_stats, 0, sizeof(ru_stats));
 }
 
 static enum stat_type evsel__stat_type(struct evsel *evsel)
tools/perf/util/stat.h +0 -16
index 34f30a295f89..5b8b4675883c 100644
@@ -102,7 +102,6 @@ struct perf_stat_config {
 	const char		*csv_sep;
 	struct stats		*walltime_nsecs_stats;
 	struct rusage		 ru_data;
-	struct rusage_stats		 *ru_stats;
 	struct cpu_aggr_map	*aggr_map;
 	aggr_get_id_t		 aggr_get_id;
 	struct cpu_aggr_map	*cpus_aggr_map;
@@ -132,25 +131,10 @@ static inline void init_stats(struct stats *stats)
 	stats->max  = 0;
 }
 
-static inline void init_rusage_stats(struct rusage_stats *ru_stats) {
-	init_stats(&ru_stats->ru_utime_usec_stat);
-	init_stats(&ru_stats->ru_stime_usec_stat);
-}
-
-static inline void update_rusage_stats(struct rusage_stats *ru_stats, struct rusage* rusage) {
-	const u64 us_to_ns = 1000;
-	const u64 s_to_ns = 1000000000;
-	update_stats(&ru_stats->ru_utime_usec_stat,
-	             (rusage->ru_utime.tv_usec * us_to_ns + rusage->ru_utime.tv_sec * s_to_ns));
-	update_stats(&ru_stats->ru_stime_usec_stat,
-	             (rusage->ru_stime.tv_usec * us_to_ns + rusage->ru_stime.tv_sec * s_to_ns));
-}
-
 struct evsel;
 struct evlist;
 
 extern struct stats walltime_nsecs_stats;
-extern struct rusage_stats ru_stats;
 
 enum metric_threshold_classify {
 	METRIC_THRESHOLD_UNKNOWN,
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#6 lore raw

PATCH v3 6/9 perf tool_pmu: More accurately set the cpus for tool events

In reply to Ian Rogers

The user and system time events can record on different CPUs, but for all other events a single CPU map of just CPU 0 makes sense. In parse-events detect a tool PMU and then pass the perf_event_attr so that the tool_pmu can return CPUs specific for the event. This avoids a CPU map of all online CPUs being used for events like duration_time. Avoiding this avoids the evlist CPUs containing CPUs for which duration_time just gives 0. Minimizing the evlist CPUs can remove unnecessary sched_setaffinity syscalls that delay metric calculations.

---
 tools/perf/util/parse-events.c |  9 +++++++--
 tools/perf/util/tool_pmu.c     | 19 +++++++++++++++++++
 tools/perf/util/tool_pmu.h     |  1 +
 3 files changed, 27 insertions(+), 2 deletions(-)
tools/perf/util/parse-events.c +7 -2
index 0c0dc20b1c13..7b2422ccb554 100644
@@ -30,6 +30,7 @@
 #include "util/event.h"
 #include "util/bpf-filter.h"
 #include "util/stat.h"
+#include "util/tool_pmu.h"
 #include "util/util.h"
 #include "tracepoint.h"
 #include <api/fs/tracing_path.h>
@@ -227,8 +228,12 @@ __add_event(struct list_head *list, int *idx,
 	if (pmu) {
 		is_pmu_core = pmu->is_core;
 		pmu_cpus = perf_cpu_map__get(pmu->cpus);
-		if (perf_cpu_map__is_empty(pmu_cpus))
-			pmu_cpus = cpu_map__online();
+		if (perf_cpu_map__is_empty(pmu_cpus)) {
+			if (perf_pmu__is_tool(pmu))
+				pmu_cpus = tool_pmu__cpus(attr);
+			else
+				pmu_cpus = cpu_map__online();
+		}
 	} else {
 		is_pmu_core = (attr->type == PERF_TYPE_HARDWARE ||
 			       attr->type == PERF_TYPE_HW_CACHE);
tools/perf/util/tool_pmu.c +19 -0
index b895e88ff740..03864a8f5e91 100644
@@ -2,6 +2,7 @@
 #include "cgroup.h"
 #include "counts.h"
 #include "cputopo.h"
+#include "debug.h"
 #include "evsel.h"
 #include "pmu.h"
 #include "print-events.h"
@@ -12,6 +13,7 @@
 #include <api/fs/fs.h>
 #include <api/io.h>
 #include <internal/threadmap.h>
+#include <perf/cpumap.h>
 #include <perf/threadmap.h>
 #include <fcntl.h>
 #include <strings.h>
@@ -106,6 +108,23 @@ const char *evsel__tool_pmu_event_name(const struct evsel *evsel)
 	return tool_pmu__event_to_str(evsel->core.attr.config);
 }
 
+struct perf_cpu_map *tool_pmu__cpus(struct perf_event_attr *attr)
+{
+	static struct perf_cpu_map *cpu0_map;
+	enum tool_pmu_event event = (enum tool_pmu_event)attr->config;
+
+	if (event <= TOOL_PMU__EVENT_NONE || event >= TOOL_PMU__EVENT_MAX) {
+		pr_err("Invalid tool PMU event config %llx\n", attr->config);
+		return NULL;
+	}
+	if (event == TOOL_PMU__EVENT_USER_TIME || event == TOOL_PMU__EVENT_SYSTEM_TIME)
+		return cpu_map__online();
+
+	if (!cpu0_map)
+		cpu0_map = perf_cpu_map__new_int(0);
+	return perf_cpu_map__get(cpu0_map);
+}
+
 static bool read_until_char(struct io *io, char e)
 {
 	int c;
tools/perf/util/tool_pmu.h +1 -0
index d642e7d73910..e95fcbd55384 100644
@@ -39,6 +39,7 @@ bool tool_pmu__read_event(enum tool_pmu_event ev, struct evsel *evsel, u64 *resu
 u64 tool_pmu__cpu_slots_per_cycle(void);
 
 bool perf_pmu__is_tool(const struct perf_pmu *pmu);
+struct perf_cpu_map *tool_pmu__cpus(struct perf_event_attr *attr);
 
 bool evsel__is_tool(const struct evsel *evsel);
 enum tool_pmu_event evsel__tool_event(const struct evsel *evsel);
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#7 lore raw

PATCH v3 7/9 perf evlist: Reduce affinity use and move into iterator, fix no affinity

In reply to Ian Rogers

The evlist__for_each_cpu iterator will call sched_setaffitinity when moving between CPUs to avoid IPIs. If only 1 IPI is saved then this may be unprofitable as the delay to get scheduled may be considerable. This may be particularly true if reading an event group in perf stat in interval mode.

Move the affinity handling completely into the iterator so that a single evlist__use_affinity can determine whether CPU affinities will be used. For perf record the change is minimal as the dummy event and the real event will always make the use of affinities the thing to do. In perf stat, tool events are ignored and affinities only used if >1 event on the same CPU occur. Determining if affinities are useful is done by per-event in a new PMU benefits from affinity function.

Fix a bug where when there are no affinities that the CPU map iterator may reference a CPU not present in the initial evsel. Fix by making the iterator and non-iterator code common.

Fix a bug where closing events on an evlist wasn't closing TPEBS events.

---
 tools/perf/builtin-stat.c | 128 ++++++++++++++----------------
 tools/perf/util/evlist.c  | 160 ++++++++++++++++++++++++--------------
 tools/perf/util/evlist.h  |  26 +++++--
 tools/perf/util/pmu.c     |  12 +++
 tools/perf/util/pmu.h     |   1 +
 5 files changed, 189 insertions(+), 138 deletions(-)
tools/perf/builtin-stat.c +57 -71
index 3cd663b3b357..5f31cd5bb03b 100644
@@ -360,22 +360,14 @@ static int read_counter_cpu(struct evsel *counter, int cpu_map_idx)
 	return 0;
 }
 
-static int read_affinity_counters(void)
+static int read_counters_with_affinity(void)
 {
 	struct evlist_cpu_iterator evlist_cpu_itr;
-	struct affinity saved_affinity, *affinity;
 
 	if (all_counters_use_bpf)
 		return 0;
 
-	if (!target__has_cpu(&target) || target__has_per_thread(&target))
-		affinity = NULL;
-	else if (affinity__setup(&saved_affinity) < 0)
-		return -1;
-	else
-		affinity = &saved_affinity;
-
-	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
+	evlist__for_each_cpu(evlist_cpu_itr, evsel_list) {
 		struct evsel *counter = evlist_cpu_itr.evsel;
 
 		if (evsel__is_bpf(counter))
@@ -384,8 +376,6 @@ static int read_affinity_counters(void)
 		if (!counter->err)
 			counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx);
 	}
-	if (affinity)
-		affinity__cleanup(&saved_affinity);
 
 	return 0;
 }
@@ -408,12 +398,18 @@ static int read_bpf_map_counters(void)
 
 static int read_counters(void)
 {
-	if (!stat_config.stop_read_counter) {
-		if (read_bpf_map_counters() ||
-		    read_affinity_counters())
-			return -1;
-	}
-	return 0;
+	int ret;
+
+	if (stat_config.stop_read_counter)
+		return 0;
+
+	// Read all BPF counters first.
+	ret = read_bpf_map_counters();
+	if (ret)
+		return ret;
+
+	// Read non-BPF and non-tool counters next.
+	return read_counters_with_affinity();
 }
 
 static void process_counters(void)
@@ -754,7 +750,6 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	const bool forks = (argc > 0);
 	bool is_pipe = STAT_RECORD ? perf_stat.data.is_pipe : false;
 	struct evlist_cpu_iterator evlist_cpu_itr;
-	struct affinity saved_affinity, *affinity = NULL;
 	int err, open_err = 0;
 	bool second_pass = false, has_supported_counters;
 
@@ -766,14 +761,6 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 		child_pid = evsel_list->workload.pid;
 	}
 
-	if (!cpu_map__is_dummy(evsel_list->core.user_requested_cpus)) {
-		if (affinity__setup(&saved_affinity) < 0) {
-			err = -1;
-			goto err_out;
-		}
-		affinity = &saved_affinity;
-	}
-
 	evlist__for_each_entry(evsel_list, counter) {
 		counter->reset_group = false;
 		if (bpf_counter__load(counter, &target)) {
@@ -786,49 +773,48 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 
 	evlist__reset_aggr_stats(evsel_list);
 
-	evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
-		counter = evlist_cpu_itr.evsel;
+	/*
+	 * bperf calls evsel__open_per_cpu() in bperf__load(), so
+	 * no need to call it again here.
+	 */
+	if (!target.use_bpf) {
+		evlist__for_each_cpu(evlist_cpu_itr, evsel_list) {
+			counter = evlist_cpu_itr.evsel;
 
-		/*
-		 * bperf calls evsel__open_per_cpu() in bperf__load(), so
-		 * no need to call it again here.
-		 */
-		if (target.use_bpf)
-			break;
+			if (counter->reset_group || !counter->supported)
+				continue;
+			if (evsel__is_bperf(counter))
+				continue;
 
-		if (counter->reset_group || !counter->supported)
-			continue;
-		if (evsel__is_bperf(counter))
-			continue;
+			while (true) {
+				if (create_perf_stat_counter(counter, &stat_config,
+							      evlist_cpu_itr.cpu_map_idx) == 0)
+					break;
 
-		while (true) {
-			if (create_perf_stat_counter(counter, &stat_config,
-						     evlist_cpu_itr.cpu_map_idx) == 0)
-				break;
+				open_err = errno;
+				/*
+				 * Weak group failed. We cannot just undo this
+				 * here because earlier CPUs might be in group
+				 * mode, and the kernel doesn't support mixing
+				 * group and non group reads. Defer it to later.
+				 * Don't close here because we're in the wrong
+				 * affinity.
+				 */
+				if ((open_err == EINVAL || open_err == EBADF) &&
+					evsel__leader(counter) != counter &&
+					counter->weak_group) {
+					evlist__reset_weak_group(evsel_list, counter, false);
+					assert(counter->reset_group);
+					counter->supported = true;
+					second_pass = true;
+					break;
+				}
 
-			open_err = errno;
-			/*
-			 * Weak group failed. We cannot just undo this here
-			 * because earlier CPUs might be in group mode, and the kernel
-			 * doesn't support mixing group and non group reads. Defer
-			 * it to later.
-			 * Don't close here because we're in the wrong affinity.
-			 */
-			if ((open_err == EINVAL || open_err == EBADF) &&
-				evsel__leader(counter) != counter &&
-				counter->weak_group) {
-				evlist__reset_weak_group(evsel_list, counter, false);
-				assert(counter->reset_group);
-				counter->supported = true;
-				second_pass = true;
-				break;
+				if (stat_handle_error(counter, open_err) != COUNTER_RETRY)
+					break;
 			}
-
-			if (stat_handle_error(counter, open_err) != COUNTER_RETRY)
-				break;
 		}
 	}
-
 	if (second_pass) {
 		/*
 		 * Now redo all the weak group after closing them,
@@ -836,7 +822,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 		 */
 
 		/* First close errored or weak retry */
-		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
+		evlist__for_each_cpu(evlist_cpu_itr, evsel_list) {
 			counter = evlist_cpu_itr.evsel;
 
 			if (!counter->reset_group && counter->supported)
@@ -845,7 +831,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			perf_evsel__close_cpu(&counter->core, evlist_cpu_itr.cpu_map_idx);
 		}
 		/* Now reopen weak */
-		evlist__for_each_cpu(evlist_cpu_itr, evsel_list, affinity) {
+		evlist__for_each_cpu(evlist_cpu_itr, evsel_list) {
 			counter = evlist_cpu_itr.evsel;
 
 			if (!counter->reset_group)
@@ -854,17 +840,18 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			while (true) {
 				pr_debug2("reopening weak %s\n", evsel__name(counter));
 				if (create_perf_stat_counter(counter, &stat_config,
-							     evlist_cpu_itr.cpu_map_idx) == 0)
+							     evlist_cpu_itr.cpu_map_idx) == 0) {
+					evlist_cpu_iterator__exit(&evlist_cpu_itr);
 					break;
-
+				}
 				open_err = errno;
-				if (stat_handle_error(counter, open_err) != COUNTER_RETRY)
+				if (stat_handle_error(counter, open_err) != COUNTER_RETRY) {
+					evlist_cpu_iterator__exit(&evlist_cpu_itr);
 					break;
+				}
 			}
 		}
 	}
-	affinity__cleanup(affinity);
-	affinity = NULL;
 
 	has_supported_counters = false;
 	evlist__for_each_entry(evsel_list, counter) {
@@ -1015,7 +1002,6 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 	if (forks)
 		evlist__cancel_workload(evsel_list);
 
-	affinity__cleanup(affinity);
 	return err;
 }
 
tools/perf/util/evlist.c +100 -60

227 lines not on this page · raw message

tools/perf/util/evlist.h +19 -7
index 5e71e3dc6042..b4604c3f03d6 100644
@@ -10,6 +10,7 @@
 #include <internal/evlist.h>
 #include <internal/evsel.h>
 #include <perf/evlist.h>
+#include "affinity.h"
 #include "events_stats.h"
 #include "evsel.h"
 #include "rblist.h"
@@ -361,6 +362,8 @@ struct evlist_cpu_iterator {
 	struct perf_cpu cpu;
 	/** If present, used to set the affinity when switching between CPUs. */
 	struct affinity *affinity;
+	/** Maybe be used to hold affinity state prior to iterating. */
+	struct affinity saved_affinity;
 };
 
 /**
@@ -368,22 +371,31 @@ struct evlist_cpu_iterator {
  *                        affinity, iterate over all CPUs and then the evlist
  *                        for each evsel on that CPU. When switching between
  *                        CPUs the affinity is set to the CPU to avoid IPIs
- *                        during syscalls.
+ *                        during syscalls. The affinity is set up and removed
+ *                        automatically, if the loop is broken a call to
+ *                        evlist_cpu_iterator__exit is necessary.
  * @evlist_cpu_itr: the iterator instance.
  * @evlist: evlist instance to iterate.
- * @affinity: NULL or used to set the affinity to the current CPU.
  */
-#define evlist__for_each_cpu(evlist_cpu_itr, evlist, affinity)		\
-	for ((evlist_cpu_itr) = evlist__cpu_begin(evlist, affinity);	\
+#define evlist__for_each_cpu(evlist_cpu_itr, evlist)			\
+	for (evlist_cpu_iterator__init(&(evlist_cpu_itr), evlist);	\
 	     !evlist_cpu_iterator__end(&evlist_cpu_itr);		\
 	     evlist_cpu_iterator__next(&evlist_cpu_itr))
 
-/** Returns an iterator set to the first CPU/evsel of evlist. */
-struct evlist_cpu_iterator evlist__cpu_begin(struct evlist *evlist, struct affinity *affinity);
+/** Setup an iterator set to the first CPU/evsel of evlist. */
+void evlist_cpu_iterator__init(struct evlist_cpu_iterator *itr, struct evlist *evlist);
+/**
+ * Cleans up the iterator, automatically done by evlist_cpu_iterator__next when
+ * the end of the list is reached. Multiple calls are safe.
+ */
+void evlist_cpu_iterator__exit(struct evlist_cpu_iterator *itr);
 /** Move to next element in iterator, updating CPU, evsel and the affinity. */
 void evlist_cpu_iterator__next(struct evlist_cpu_iterator *evlist_cpu_itr);
 /** Returns true when iterator is at the end of the CPUs and evlist. */
-bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr);
+static inline bool evlist_cpu_iterator__end(const struct evlist_cpu_iterator *evlist_cpu_itr)
+{
+	return evlist_cpu_itr->evlist_cpu_map_idx >= evlist_cpu_itr->evlist_cpu_map_nr;
+}
 
 struct evsel *evlist__get_tracking_event(struct evlist *evlist);
 void evlist__set_tracking_event(struct evlist *evlist, struct evsel *tracking_evsel);
tools/perf/util/pmu.c +12 -0
index d597263fab4f..caa3bd5e0956 100644
@@ -2402,6 +2402,18 @@ bool perf_pmu__is_software(const struct perf_pmu *pmu)
 	return false;
 }
 
+bool perf_pmu__benefits_from_affinity(struct perf_pmu *pmu)
+{
+	if (!pmu)
+		return true; /* Assume is core. */
+
+	/*
+	 * All perf event PMUs should benefit from accessing the perf event
+	 * contexts on the local CPU.
+	 */
+	return pmu->type <= PERF_PMU_TYPE_PE_END;
+}
+
 FILE *perf_pmu__open_file(const struct perf_pmu *pmu, const char *name)
 {
 	char path[PATH_MAX];
tools/perf/util/pmu.h +1 -0
index 1ebcf0242af8..87e12a9a0e67 100644
@@ -259,6 +259,7 @@ bool perf_pmu__name_no_suffix_match(const struct perf_pmu *pmu, const char *to_m
  *                        perf_sw_context in the kernel?
  */
 bool perf_pmu__is_software(const struct perf_pmu *pmu);
+bool perf_pmu__benefits_from_affinity(struct perf_pmu *pmu);
 
 FILE *perf_pmu__open_file(const struct perf_pmu *pmu, const char *name);
 FILE *perf_pmu__open_file_at(const struct perf_pmu *pmu, int dirfd, const char *name);
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#8 lore raw

PATCH v3 8/9 perf stat: Read tool events last

In reply to Ian Rogers

When reading a metric like memory bandwidth on multiple sockets, the additional sockets will be on CPUS > 0. Because of the affinity reading, the counters are read on CPU 0 along with the time, then the later sockets are read. This can lead to the later sockets having a bandwidth larger than is possible for the period of time. To avoid this moving the reading of tool events to occur after all other events are read.

---
 tools/perf/builtin-stat.c | 29 ++++++++++++++++++++++++++++-
 tools/perf/util/evlist.c  |  4 ----
 2 files changed, 28 insertions(+), 5 deletions(-)
tools/perf/builtin-stat.c +28 -1
index 5f31cd5bb03b..75ff08c5f0ed 100644
@@ -373,6 +373,9 @@ static int read_counters_with_affinity(void)
 		if (evsel__is_bpf(counter))
 			continue;
 
+		if (evsel__is_tool(counter))
+			continue;
+
 		if (!counter->err)
 			counter->err = read_counter_cpu(counter, evlist_cpu_itr.cpu_map_idx);
 	}
@@ -396,6 +399,24 @@ static int read_bpf_map_counters(void)
 	return 0;
 }
 
+static int read_tool_counters(void)
+{
+	struct evsel *counter;
+
+	evlist__for_each_entry(evsel_list, counter) {
+		int idx;
+
+		if (!evsel__is_tool(counter))
+			continue;
+
+		perf_cpu_map__for_each_idx(idx, counter->core.cpus) {
+			if (!counter->err)
+				counter->err = read_counter_cpu(counter, idx);
+		}
+	}
+	return 0;
+}
+
 static int read_counters(void)
 {
 	int ret;
@@ -409,7 +430,13 @@ static int read_counters(void)
 		return ret;
 
 	// Read non-BPF and non-tool counters next.
-	return read_counters_with_affinity();
+	ret = read_counters_with_affinity();
+	if (ret)
+		return ret;
+
+	// Read the tool counters last. This way the duration_time counter
+	// should always be greater than any other counter's enabled time.
+	return read_tool_counters();
 }
 
 static void process_counters(void)
tools/perf/util/evlist.c +0 -4
index b6df81b8a236..fc3dae7cdfca 100644
@@ -368,10 +368,6 @@ static bool evlist__use_affinity(struct evlist *evlist)
 	struct perf_cpu_map *used_cpus = NULL;
 	bool ret = false;
 
-	/*
-	 * With perf record core.user_requested_cpus is usually NULL.
-	 * Use the old method to handle this for now.
-	 */
 	if (!evlist->core.user_requested_cpus ||
 	    cpu_map__is_dummy(evlist->core.user_requested_cpus))
 		return false;
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
author reviewer
#9 lore raw

PATCH v3 9/9 perf stat: Add no-affinity flag

In reply to Ian Rogers

Add flag that disables affinity behavior. Using sched_setaffinity to place a perf thread on a CPU can avoid certain interprocessor interrupts but may introduce a delay due to the scheduling, particularly on loaded machines. Add a command line option to disable the behavior. This behavior is less present in other tools like perf record, as it uses a ring buffer and doesn't make repeated system calls.

---
 tools/perf/Documentation/perf-stat.txt | 4 ++++
 tools/perf/builtin-stat.c              | 6 ++++++
 tools/perf/util/evlist.c               | 2 +-
 tools/perf/util/evlist.h               | 1 +
 4 files changed, 12 insertions(+), 1 deletion(-)
tools/perf/Documentation/perf-stat.txt +4 -0
index 1a766d4a2233..1ffb510606af 100644
@@ -382,6 +382,10 @@ color the metric's computed value.
 Don't print output, warnings or messages. This is useful with perf stat
 record below to only write data to the perf.data file.
 
+--no-affinity::
+Don't change scheduler affinities when iterating over CPUs. Disables
+an optimization aimed at minimizing interprocessor interrupts.
+
 STAT RECORD
 -----------
 Stores stat data into perf data file.
tools/perf/builtin-stat.c +6 -0
index 75ff08c5f0ed..f2f149f0c2cf 100644
@@ -2479,6 +2479,7 @@ static int parse_tpebs_mode(const struct option *opt, const char *str,
 int cmd_stat(int argc, const char **argv)
 {
 	struct opt_aggr_mode opt_mode = {};
+	bool no_affinity = false;
 	struct option stat_options[] = {
 		OPT_BOOLEAN('T', "transaction", &transaction_run,
 			"hardware transaction statistics"),
@@ -2607,6 +2608,8 @@ int cmd_stat(int argc, const char **argv)
 			"don't print 'summary' for CSV summary output"),
 		OPT_BOOLEAN(0, "quiet", &quiet,
 			"don't print any output, messages or warnings (useful with record)"),
+		OPT_BOOLEAN(0, "no-affinity", &no_affinity,
+			"don't allow affinity optimizations aimed at reducing IPIs"),
 		OPT_CALLBACK(0, "cputype", &evsel_list, "hybrid cpu type",
 			"Only enable events on applying cpu with this type "
 			"for hybrid platform (e.g. core or atom)",
@@ -2664,6 +2667,9 @@ int cmd_stat(int argc, const char **argv)
 	} else
 		stat_config.csv_sep = DEFAULT_SEPARATOR;
 
+	if (no_affinity)
+		evsel_list->no_affinity = true;
+
 	if (argc && strlen(argv[0]) > 2 && strstarts("record", argv[0])) {
 		argc = __cmd_record(stat_options, &opt_mode, argc, argv);
 		if (argc < 0)
tools/perf/util/evlist.c +1 -1
index fc3dae7cdfca..53c8e974de8b 100644
@@ -368,7 +368,7 @@ static bool evlist__use_affinity(struct evlist *evlist)
 	struct perf_cpu_map *used_cpus = NULL;
 	bool ret = false;
 
-	if (!evlist->core.user_requested_cpus ||
+	if (evlist->no_affinity || !evlist->core.user_requested_cpus ||
 	    cpu_map__is_dummy(evlist->core.user_requested_cpus))
 		return false;
 
tools/perf/util/evlist.h +1 -0
index b4604c3f03d6..c7ba0e0b2219 100644
@@ -59,6 +59,7 @@ struct event_enable_timer;
 struct evlist {
 	struct perf_evlist core;
 	bool		 enabled;
+	bool		 no_affinity;
 	int		 id_pos;
 	int		 is_pos;
 	int		 nr_br_cntr;
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
#10 lore raw

Re: [PATCH v3 9/9] perf stat: Add no-affinity flag

Quoted 7 lines · Ian Rogers

Add flag that disables affinity behavior. Using sched_setaffinity to place a perf thread on a CPU can avoid certain interprocessor interrupts but may introduce a delay due to the scheduling, particularly on loaded machines. Add a command line option to disable the behavior. This behavior is less present in other tools like perf record, as it uses a ring buffer and doesn't make repeated system calls.

I would rather use a real time scheduling policy instead if you see starvation. That would still get the batching efficiency.

If it's done it would be better in .perfconfig instead of adding ever more obscure command line options.

-Andi