Skip to contentSkip to conversation

Intel TPEBS min/max/mean/last support

PATCH Superseded Ian Rogers 1 participant 17 messages Apr 7, 2025 perf events

v3 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

v3 cover Intel TPEBS min/max/mean/last support

The patches add support to computing the min, max, mean or last retirement latency and then using that value as the basis for metrics. When values aren't available, support is added to use the retirement latency as recorded for an event in the perf json.

Support is added for reading the retirement latency from the forked perf command more than once. To avoid killing the process commands are sent through the control fd. Some name handling is changed to make it more robust.

Rather than retirement latency events having issues with perf record, make it so that the retirement latency modifier enables sample weights.

v3: Two fixes from Kan Liang. Ensure min/max statistics don't vary
when real samples are encountered.

v2: Addition of header cleanup patch originally posted:
https://lore.kernel.org/lkml/[email protected]/
as there are no arch specific reasons not to build this code. Fix bug in "perf pmu-events: Add retirement latency to JSON events inside of perf" where "t->stats.n != 0" should have been "t->stats.n == 0".
Add patch so that perf record of a retirement latency event doesn't crash but instead enables sample weights for the event.

Ian Rogers (16):
perf intel-tpebs: Cleanup header
perf intel-tpebs: Simplify tpebs_cmd
perf intel-tpebs: Rename tpebs_start to evsel__tpebs_open perf intel-tpebs: Separate evsel__tpebs_prepare out of
evsel__tpebs_open
perf intel-tpebs: Move cpumap_buf out of evsel__tpebs_open perf intel-tpebs: Reduce scope of tpebs_events_size
perf intel-tpebs: Inline get_perf_record_args
perf intel-tpebs: Ensure events are opened, factor out finding perf intel-tpebs: Refactor tpebs_results list
perf intel-tpebs: Add support for updating counts in evsel__tpebs_read perf intel-tpebs: Add mutex for tpebs_results
perf intel-tpebs: Don't close record on read
perf intel-tpebs: Use stats for retirement latency statistics perf stat: Add mean, min, max and last --tpebs-mode options perf pmu-events: Add retirement latency to JSON events inside of perf perf record: Retirement latency cleanup in evsel__config

 tools/perf/Documentation/perf-stat.txt   |   7 +
 tools/perf/builtin-stat.c                |  29 +-
 tools/perf/pmu-events/empty-pmu-events.c | 216 +++----
 tools/perf/pmu-events/jevents.py         |   6 +
 tools/perf/pmu-events/pmu-events.h       |   3 +
 tools/perf/util/Build                    |   2 +-
 tools/perf/util/evlist.c                 |   1 -
 tools/perf/util/evsel.c                  |  22 +-
 tools/perf/util/evsel.h                  |   6 +
 tools/perf/util/intel-tpebs.c            | 682 ++++++++++++++---------
 tools/perf/util/intel-tpebs.h            |  40 +-
 tools/perf/util/parse-events.c           |   4 +
 tools/perf/util/pmu.c                    |  52 +-
 tools/perf/util/pmu.h                    |   3 +
 14 files changed, 666 insertions(+), 407 deletions(-)
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#1 lore raw

PATCH v3 1/16 perf intel-tpebs: Cleanup header

Remove arch conditional compilation. Arch conditional compilation belongs in the arch/ directory.

Tidy header guards to match other files. Remove unneeded includes and switch to forward declarations when necesary.

---
 tools/perf/util/Build         |  2 +-
 tools/perf/util/intel-tpebs.c |  1 +
 tools/perf/util/intel-tpebs.h | 30 ++++++------------------------
 3 files changed, 8 insertions(+), 25 deletions(-)
tools/perf/util/Build +1 -1
index 946bce6628f3..815274b199fd 100644
@@ -161,7 +161,7 @@ perf-util-y += clockid.o
 perf-util-y += list_sort.o
 perf-util-y += mutex.o
 perf-util-y += sharded_mutex.o
-perf-util-$(CONFIG_X86_64) += intel-tpebs.o
+perf-util-y += intel-tpebs.o
 
 perf-util-$(CONFIG_LIBBPF) += bpf_map.o
 perf-util-$(CONFIG_PERF_BPF_SKEL) += bpf_counter.o
tools/perf/util/intel-tpebs.c +1 -0
index 2c421b475b3b..3503da28a12f 100644
@@ -19,6 +19,7 @@
 #include "tool.h"
 #include "cpumap.h"
 #include "metricgroup.h"
+#include "stat.h"
 #include <sys/stat.h>
 #include <sys/file.h>
 #include <poll.h>
tools/perf/util/intel-tpebs.h +6 -24
index 766b3fbd79f1..63c16e759a71 100644
@@ -2,34 +2,16 @@
 /*
  * intel_tpebs.h: Intel TEPBS support
  */
-#ifndef INCLUDE__PERF_INTEL_TPEBS_H__
-#define INCLUDE__PERF_INTEL_TPEBS_H__
+#ifndef __INTEL_TPEBS_H
+#define __INTEL_TPEBS_H
 
-#include "stat.h"
-#include "evsel.h"
-
-#ifdef HAVE_ARCH_X86_64_SUPPORT
+struct evlist;
+struct evsel;
 
 extern bool tpebs_recording;
+
 int tpebs_start(struct evlist *evsel_list);
 void tpebs_delete(void);
 int tpebs_set_evsel(struct evsel *evsel, int cpu_map_idx, int thread);
 
-#else
-
-static inline int tpebs_start(struct evlist *evsel_list __maybe_unused)
-{
-	return 0;
-}
-
-static inline void tpebs_delete(void) {};
-
-static inline int tpebs_set_evsel(struct evsel *evsel  __maybe_unused,
-				int cpu_map_idx  __maybe_unused,
-				int thread  __maybe_unused)
-{
-	return 0;
-}
-
-#endif
-#endif
+#endif /* __INTEL_TPEBS_H */
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#2 lore raw

PATCH v3 2/16 perf intel-tpebs: Simplify tpebs_cmd

In reply to Ian Rogers

No need to dynamically allocate when there is 1. tpebs_pid duplicates tpebs_cmd.pid, so remove. Use 0 as the uninitialized value (PID == 0 is reserved for the kernel) rather than -1.

---
 tools/perf/util/intel-tpebs.c | 55 ++++++++++++-----------------------
 1 file changed, 18 insertions(+), 37 deletions(-)
tools/perf/util/intel-tpebs.c +18 -37
index 3503da28a12f..74b43faab986 100644
@@ -28,11 +28,10 @@
 #define PERF_DATA		"-"
 
 bool tpebs_recording;
-static pid_t tpebs_pid = -1;
 static size_t tpebs_event_size;
 static LIST_HEAD(tpebs_results);
 static pthread_t tpebs_reader_thread;
-static struct child_process *tpebs_cmd;
+static struct child_process tpebs_cmd;
 
 struct tpebs_retire_lat {
 	struct list_head nd;
@@ -83,16 +82,6 @@ static int get_perf_record_args(const char **record_argv, char buf[],
 	return 0;
 }
 
-static int prepare_run_command(const char **argv)
-{
-	tpebs_cmd = zalloc(sizeof(struct child_process));
-	if (!tpebs_cmd)
-		return -ENOMEM;
-	tpebs_cmd->argv = argv;
-	tpebs_cmd->out = -1;
-	return 0;
-}
-
 static int start_perf_record(int control_fd[], int ack_fd[],
 				const char *cpumap_buf)
 {
@@ -110,10 +99,10 @@ static int start_perf_record(int control_fd[], int ack_fd[],
 	if (ret)
 		goto out;
 
-	ret = prepare_run_command(record_argv);
-	if (ret)
-		goto out;
-	ret = start_command(tpebs_cmd);
+	assert(tpebs_cmd.pid == 0);
+	tpebs_cmd.argv = record_argv;
+	tpebs_cmd.out = -1;
+	ret = start_command(&tpebs_cmd);
 out:
 	free(record_argv);
 	return ret;
@@ -156,14 +145,13 @@ static int process_feature_event(struct perf_session *session,
 	return 0;
 }
 
-static void *__sample_reader(void *arg)
+static void *__sample_reader(void *arg __maybe_unused)
 {
-	struct child_process *child = arg;
 	struct perf_session *session;
 	struct perf_data data = {
 		.mode = PERF_DATA_MODE_READ,
 		.path = PERF_DATA,
-		.file.fd = child->out,
+		.file.fd = tpebs_cmd.out,
 	};
 	struct perf_tool tool;
 
@@ -189,12 +177,12 @@ static int tpebs_stop(void)
 	int ret = 0;
 
 	/* Like tpebs_start, we should only run tpebs_end once. */
-	if (tpebs_pid != -1) {
-		kill(tpebs_cmd->pid, SIGTERM);
-		tpebs_pid = -1;
+	if (tpebs_cmd.pid != 0) {
+		kill(tpebs_cmd.pid, SIGTERM);
 		pthread_join(tpebs_reader_thread, NULL);
-		close(tpebs_cmd->out);
-		ret = finish_command(tpebs_cmd);
+		close(tpebs_cmd.out);
+		ret = finish_command(&tpebs_cmd);
+		tpebs_cmd.pid = 0;
 		if (ret == -ERR_RUN_COMMAND_WAITPID_SIGNAL)
 			ret = 0;
 	}
@@ -219,7 +207,7 @@ int tpebs_start(struct evlist *evsel_list)
 	 * We should only run tpebs_start when tpebs_recording is enabled.
 	 * And we should only run it once with all the required events.
 	 */
-	if (tpebs_pid != -1 || !tpebs_recording)
+	if (tpebs_cmd.pid != 0 || !tpebs_recording)
 		return 0;
 
 	cpu_map__snprint(evsel_list->core.user_requested_cpus, cpumap_buf, sizeof(cpumap_buf));
@@ -284,10 +272,11 @@ int tpebs_start(struct evlist *evsel_list)
 		ret = start_perf_record(control_fd, ack_fd, cpumap_buf);
 		if (ret)
 			goto out;
-		tpebs_pid = tpebs_cmd->pid;
-		if (pthread_create(&tpebs_reader_thread, NULL, __sample_reader, tpebs_cmd)) {
-			kill(tpebs_cmd->pid, SIGTERM);
-			close(tpebs_cmd->out);
+
+		if (pthread_create(&tpebs_reader_thread, /*attr=*/NULL, __sample_reader,
+				   /*arg=*/NULL)) {
+			kill(tpebs_cmd.pid, SIGTERM);
+			close(tpebs_cmd.out);
 			pr_err("Could not create thread to process sample data.\n");
 			ret = -1;
 			goto out;
@@ -416,18 +405,10 @@ void tpebs_delete(void)
 {
 	struct tpebs_retire_lat *r, *rtmp;
 
-	if (tpebs_pid == -1)
-		return;
-
 	tpebs_stop();
 
 	list_for_each_entry_safe(r, rtmp, &tpebs_results, nd) {
 		list_del_init(&r->nd);
 		tpebs_retire_lat__delete(r);
 	}
-
-	if (tpebs_cmd) {
-		free(tpebs_cmd);
-		tpebs_cmd = NULL;
-	}
 }
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#3 lore raw

PATCH v3 3/16 perf intel-tpebs: Rename tpebs_start to evsel__tpebs_open

In reply to Ian Rogers

Try to add more consistency to evsel by having tpebs_start renamed to evsel__tpebs_open, passing the evsel that is being opened. The unusual behavior of evsel__tpebs_open opening all events on the evlist is kept and will be cleaned up further in later patches. The comments are cleaned up as tpebs_start isn't called from evlist.

---
 tools/perf/util/evsel.c       |  2 +-
 tools/perf/util/intel-tpebs.c | 33 ++++++++++++++++-----------------
 tools/perf/util/intel-tpebs.h |  2 +-
 3 files changed, 18 insertions(+), 19 deletions(-)
tools/perf/util/evsel.c +1 -1
index 1974395492d7..121283f2f382 100644
@@ -2595,7 +2595,7 @@ static int evsel__open_cpu(struct evsel *evsel, struct perf_cpu_map *cpus,
 	struct perf_cpu cpu;
 
 	if (evsel__is_retire_lat(evsel))
-		return tpebs_start(evsel->evlist);
+		return evsel__tpebs_open(evsel);
 
 	err = __evsel__prepare_open(evsel, cpus, threads);
 	if (err)
tools/perf/util/intel-tpebs.c +16 -17
index 74b43faab986..566e0ddcad88 100644
@@ -12,6 +12,7 @@
 #include <linux/zalloc.h>
 #include <linux/err.h>
 #include "sample.h"
+#include "counts.h"
 #include "debug.h"
 #include "evlist.h"
 #include "evsel.h"
@@ -189,18 +190,16 @@ static int tpebs_stop(void)
 	return ret;
 }
 
-/*
- * tpebs_start - start tpebs execution.
- * @evsel_list: retire_latency evsels in this list will be selected and sampled
- * to get the average retire_latency value.
- *
- * This function will be called from evlist level later when evlist__open() is
- * called consistently.
+/**
+ * evsel__tpebs_open - starts tpebs execution.
+ * @evsel: retire_latency evsel, all evsels on its list will be selected. Each
+ *         evsel is sampled to get the average retire_latency value.
  */
-int tpebs_start(struct evlist *evsel_list)
+int evsel__tpebs_open(struct evsel *evsel)
 {
 	int ret = 0;
-	struct evsel *evsel;
+	struct evsel *pos;
+	struct evlist *evsel_list = evsel->evlist;
 	char cpumap_buf[50];
 
 	/*
@@ -215,25 +214,25 @@ int tpebs_start(struct evlist *evsel_list)
 	 * Prepare perf record for sampling event retire_latency before fork and
 	 * prepare workload
 	 */
-	evlist__for_each_entry(evsel_list, evsel) {
+	evlist__for_each_entry(evsel_list, pos) {
 		int i;
 		char *name;
 		struct tpebs_retire_lat *new;
 
-		if (!evsel->retire_lat)
+		if (!pos->retire_lat)
 			continue;
 
-		pr_debug("tpebs: Retire_latency of event %s is required\n", evsel->name);
-		for (i = strlen(evsel->name) - 1; i > 0; i--) {
-			if (evsel->name[i] == 'R')
+		pr_debug("tpebs: Retire_latency of event %s is required\n", pos->name);
+		for (i = strlen(pos->name) - 1; i > 0; i--) {
+			if (pos->name[i] == 'R')
 				break;
 		}
-		if (i <= 0 || evsel->name[i] != 'R') {
+		if (i <= 0 || pos->name[i] != 'R') {
 			ret = -1;
 			goto err;
 		}
 
-		name = strdup(evsel->name);
+		name = strdup(pos->name);
 		if (!name) {
 			ret = -ENOMEM;
 			goto err;
@@ -247,7 +246,7 @@ int tpebs_start(struct evlist *evsel_list)
 			goto err;
 		}
 		new->name = name;
-		new->tpebs_name = evsel->name;
+		new->tpebs_name = pos->name;
 		list_add_tail(&new->nd, &tpebs_results);
 		tpebs_event_size += 1;
 	}
tools/perf/util/intel-tpebs.h +1 -1
index 63c16e759a71..cc98203719c8 100644
@@ -10,7 +10,7 @@ struct evsel;
 
 extern bool tpebs_recording;
 
-int tpebs_start(struct evlist *evsel_list);
+int evsel__tpebs_open(struct evsel *evsel);
 void tpebs_delete(void);
 int tpebs_set_evsel(struct evsel *evsel, int cpu_map_idx, int thread);
 
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#4 lore raw

PATCH v3 4/16 perf intel-tpebs: Separate evsel__tpebs_prepare out of evsel__tpebs_open

In reply to Ian Rogers

Separate the creation of the tpebs_retire_lat result out of the opening step. This is in preparation for adding a prepare operation for evlists.

---
 tools/perf/util/intel-tpebs.c | 133 ++++++++++++++++++++++------------
 1 file changed, 86 insertions(+), 47 deletions(-)
tools/perf/util/intel-tpebs.c +86 -47
index 566e0ddcad88..2186818b2c9b 100644
@@ -37,7 +37,7 @@ static struct child_process tpebs_cmd;
 struct tpebs_retire_lat {
 	struct list_head nd;
 	/* Event name */
-	const char *name;
+	char *name;
 	/* Event name with the TPEBS modifier R */
 	const char *tpebs_name;
 	/* Count of retire_latency values found in sample data */
@@ -190,6 +190,82 @@ static int tpebs_stop(void)
 	return ret;
 }
 
+static char *evsel__tpebs_name(struct evsel *evsel)
+{
+	char *name, *modifier;
+
+	name = strdup(evsel->name);
+	if (!name)
+		return NULL;
+
+	modifier = strrchr(name, 'R');
+	if (!modifier) {
+		pr_err("Tpebs event missing modifier '%s'\n", name);
+		free(name);
+		return NULL;
+	}
+
+	*modifier = 'p';
+	return name;
+}
+
+static struct tpebs_retire_lat *tpebs_retire_lat__new(struct evsel *evsel)
+{
+	struct tpebs_retire_lat *result = zalloc(sizeof(*result));
+
+	if (!result)
+		return NULL;
+
+	result->tpebs_name = evsel->name;
+	result->name = evsel__tpebs_name(evsel);
+	if (!result->name) {
+		free(result);
+		return NULL;
+	}
+	list_add_tail(&result->nd, &tpebs_results);
+	tpebs_event_size++;
+	return result;
+}
+
+/**
+ * evsel__tpebs_prepare - create tpebs data structures ready for opening.
+ * @evsel: retire_latency evsel, all evsels on its list will be prepared.
+ */
+static int evsel__tpebs_prepare(struct evsel *evsel)
+{
+	struct evsel *pos;
+	struct tpebs_retire_lat *tpebs_event;
+
+	list_for_each_entry(tpebs_event, &tpebs_results, nd) {
+		if (!strcmp(tpebs_event->tpebs_name, evsel->name)) {
+			/*
+			 * evsel, or an identically named one, was already
+			 * prepared.
+			 */
+			return 0;
+		}
+	}
+	tpebs_event = tpebs_retire_lat__new(evsel);
+	if (!tpebs_event)
+		return -ENOMEM;
+
+	/*
+	 * Eagerly prepare all other evsels on the list to try to ensure that by
+	 * open they are all known.
+	 */
+	evlist__for_each_entry(evsel->evlist, pos) {
+		int ret;
+
+		if (pos == evsel || !pos->retire_lat)
+			continue;
+
+		ret = evsel__tpebs_prepare(pos);
+		if (ret)
+			return ret;
+	}
+	return 0;
+}
+
 /**
  * evsel__tpebs_open - starts tpebs execution.
  * @evsel: retire_latency evsel, all evsels on its list will be selected. Each
@@ -197,10 +273,7 @@ static int tpebs_stop(void)
  */
 int evsel__tpebs_open(struct evsel *evsel)
 {
-	int ret = 0;
-	struct evsel *pos;
-	struct evlist *evsel_list = evsel->evlist;
-	char cpumap_buf[50];
+	int ret;
 
 	/*
 	 * We should only run tpebs_start when tpebs_recording is enabled.
@@ -209,49 +282,13 @@ int evsel__tpebs_open(struct evsel *evsel)
 	if (tpebs_cmd.pid != 0 || !tpebs_recording)
 		return 0;
 
-	cpu_map__snprint(evsel_list->core.user_requested_cpus, cpumap_buf, sizeof(cpumap_buf));
-	/*
-	 * Prepare perf record for sampling event retire_latency before fork and
-	 * prepare workload
-	 */
-	evlist__for_each_entry(evsel_list, pos) {
-		int i;
-		char *name;
-		struct tpebs_retire_lat *new;
-
-		if (!pos->retire_lat)
-			continue;
-
-		pr_debug("tpebs: Retire_latency of event %s is required\n", pos->name);
-		for (i = strlen(pos->name) - 1; i > 0; i--) {
-			if (pos->name[i] == 'R')
-				break;
-		}
-		if (i <= 0 || pos->name[i] != 'R') {
-			ret = -1;
-			goto err;
-		}
-
-		name = strdup(pos->name);
-		if (!name) {
-			ret = -ENOMEM;
-			goto err;
-		}
-		name[i] = 'p';
-
-		new = zalloc(sizeof(*new));
-		if (!new) {
-			ret = -1;
-			zfree(&name);
-			goto err;
-		}
-		new->name = name;
-		new->tpebs_name = pos->name;
-		list_add_tail(&new->nd, &tpebs_results);
-		tpebs_event_size += 1;
-	}
+	ret = evsel__tpebs_prepare(evsel);
+	if (ret)
+		return ret;
 
 	if (tpebs_event_size > 0) {
+		struct evlist *evsel_list = evsel->evlist;
+		char cpumap_buf[50];
 		struct pollfd pollfd = { .events = POLLIN, };
 		int control_fd[2], ack_fd[2], len;
 		char ack_buf[8];
@@ -268,6 +305,9 @@ int evsel__tpebs_open(struct evsel *evsel)
 			goto out;
 		}
 
+		cpu_map__snprint(evsel_list->core.user_requested_cpus, cpumap_buf,
+				 sizeof(cpumap_buf));
+
 		ret = start_perf_record(control_fd, ack_fd, cpumap_buf);
 		if (ret)
 			goto out;
@@ -321,7 +361,6 @@ int evsel__tpebs_open(struct evsel *evsel)
 		close(ack_fd[0]);
 		close(ack_fd[1]);
 	}
-err:
 	if (ret)
 		tpebs_delete();
 	return ret;
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#5 lore raw

PATCH v3 5/16 perf intel-tpebs: Move cpumap_buf out of evsel__tpebs_open

In reply to Ian Rogers

The buffer holds the cpumap to pass to the perf record command, so move it down to the perf record function. Make this function an evsel function given the need for the evsel for the cpumap.

---
 tools/perf/util/intel-tpebs.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)
tools/perf/util/intel-tpebs.c +6 -8
index 2186818b2c9b..2b04deaf66ff 100644
@@ -83,12 +83,15 @@ static int get_perf_record_args(const char **record_argv, char buf[],
 	return 0;
 }
 
-static int start_perf_record(int control_fd[], int ack_fd[],
-				const char *cpumap_buf)
+static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[], int ack_fd[])
 {
 	const char **record_argv;
 	int ret;
 	char buf[32];
+	char cpumap_buf[50];
+
+	cpu_map__snprint(evsel->evlist->core.user_requested_cpus, cpumap_buf,
+			 sizeof(cpumap_buf));
 
 	scnprintf(buf, sizeof(buf), "--control=fd:%d,%d", control_fd[0], ack_fd[1]);
 
@@ -287,8 +290,6 @@ int evsel__tpebs_open(struct evsel *evsel)
 		return ret;
 
 	if (tpebs_event_size > 0) {
-		struct evlist *evsel_list = evsel->evlist;
-		char cpumap_buf[50];
 		struct pollfd pollfd = { .events = POLLIN, };
 		int control_fd[2], ack_fd[2], len;
 		char ack_buf[8];
@@ -305,10 +306,7 @@ int evsel__tpebs_open(struct evsel *evsel)
 			goto out;
 		}
 
-		cpu_map__snprint(evsel_list->core.user_requested_cpus, cpumap_buf,
-				 sizeof(cpumap_buf));
-
-		ret = start_perf_record(control_fd, ack_fd, cpumap_buf);
+		ret = evsel__tpebs_start_perf_record(evsel, control_fd, ack_fd);
 		if (ret)
 			goto out;
 
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#6 lore raw

PATCH v3 6/16 perf intel-tpebs: Reduce scope of tpebs_events_size

In reply to Ian Rogers

Moved to record argument computation rather than being global.

---
 tools/perf/util/intel-tpebs.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)
tools/perf/util/intel-tpebs.c +6 -3
index 2b04deaf66ff..e3bed86145b9 100644
@@ -29,7 +29,6 @@
 #define PERF_DATA		"-"
 
 bool tpebs_recording;
-static size_t tpebs_event_size;
 static LIST_HEAD(tpebs_results);
 static pthread_t tpebs_reader_thread;
 static struct child_process tpebs_cmd;
@@ -86,15 +85,20 @@ static int get_perf_record_args(const char **record_argv, char buf[],
 static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[], int ack_fd[])
 {
 	const char **record_argv;
+	size_t tpebs_event_size = 0;
 	int ret;
 	char buf[32];
 	char cpumap_buf[50];
+	struct tpebs_retire_lat *t;
 
 	cpu_map__snprint(evsel->evlist->core.user_requested_cpus, cpumap_buf,
 			 sizeof(cpumap_buf));
 
 	scnprintf(buf, sizeof(buf), "--control=fd:%d,%d", control_fd[0], ack_fd[1]);
 
+	list_for_each_entry(t, &tpebs_results, nd)
+		tpebs_event_size++;
+
 	record_argv = calloc(12 + 2 * tpebs_event_size, sizeof(char *));
 	if (!record_argv)
 		return -ENOMEM;
@@ -226,7 +230,6 @@ static struct tpebs_retire_lat *tpebs_retire_lat__new(struct evsel *evsel)
 		return NULL;
 	}
 	list_add_tail(&result->nd, &tpebs_results);
-	tpebs_event_size++;
 	return result;
 }
 
@@ -289,7 +292,7 @@ int evsel__tpebs_open(struct evsel *evsel)
 	if (ret)
 		return ret;
 
-	if (tpebs_event_size > 0) {
+	if (!list_empty(&tpebs_results)) {
 		struct pollfd pollfd = { .events = POLLIN, };
 		int control_fd[2], ack_fd[2], len;
 		char ack_buf[8];
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#7 lore raw

PATCH v3 7/16 perf intel-tpebs: Inline get_perf_record_args

In reply to Ian Rogers

Code is short enough to be inlined and there are no error cases when made inline. Make the implicit NULL pointer at the end of the argv explicit. Move the fixed number of arguments before the variable number of arguments. Correctly size the argv allocation and zero when feeing to avoid a dangling pointer.

---
 tools/perf/util/intel-tpebs.c | 75 +++++++++++++----------------------
 1 file changed, 28 insertions(+), 47 deletions(-)
tools/perf/util/intel-tpebs.c +28 -47
index e3bed86145b9..c4c818f32239 100644
@@ -47,72 +47,53 @@ struct tpebs_retire_lat {
 	double val;
 };
 
-static int get_perf_record_args(const char **record_argv, char buf[],
-				const char *cpumap_buf)
+static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[], int ack_fd[])
 {
-	struct tpebs_retire_lat *e;
-	int i = 0;
+	const char **record_argv;
+	int tpebs_event_size = 0, i = 0, ret;
+	char control_fd_buf[32];
+	char cpumap_buf[50];
+	struct tpebs_retire_lat *t;
+
+	list_for_each_entry(t, &tpebs_results, nd)
+		tpebs_event_size++;
 
-	pr_debug("tpebs: Prepare perf record for retire_latency\n");
+	record_argv = malloc((10 + 2 * tpebs_event_size) * sizeof(*record_argv));
+	if (!record_argv)
+		return -ENOMEM;
 
 	record_argv[i++] = "perf";
 	record_argv[i++] = "record";
 	record_argv[i++] = "-W";
 	record_argv[i++] = "--synth=no";
-	record_argv[i++] = buf;
 
-	if (!cpumap_buf) {
-		pr_err("tpebs: Require cpumap list to run sampling\n");
-		return -ECANCELED;
-	}
-	/* Use -C when cpumap_buf is not "-1" */
-	if (strcmp(cpumap_buf, "-1")) {
+	scnprintf(control_fd_buf, sizeof(control_fd_buf), "--control=fd:%d,%d",
+		  control_fd[0], ack_fd[1]);
+	record_argv[i++] = control_fd_buf;
+
+	record_argv[i++] = "-o";
+	record_argv[i++] = PERF_DATA;
+
+	if (!perf_cpu_map__is_any_cpu_or_is_empty(evsel->evlist->core.user_requested_cpus)) {
+		cpu_map__snprint(evsel->evlist->core.user_requested_cpus, cpumap_buf,
+				 sizeof(cpumap_buf));
 		record_argv[i++] = "-C";
 		record_argv[i++] = cpumap_buf;
 	}
 
-	list_for_each_entry(e, &tpebs_results, nd) {
+	list_for_each_entry(t, &tpebs_results, nd) {
 		record_argv[i++] = "-e";
-		record_argv[i++] = e->name;
+		record_argv[i++] = t->name;
 	}
-
-	record_argv[i++] = "-o";
-	record_argv[i++] = PERF_DATA;
-
-	return 0;
-}
-
-static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[], int ack_fd[])
-{
-	const char **record_argv;
-	size_t tpebs_event_size = 0;
-	int ret;
-	char buf[32];
-	char cpumap_buf[50];
-	struct tpebs_retire_lat *t;
-
-	cpu_map__snprint(evsel->evlist->core.user_requested_cpus, cpumap_buf,
-			 sizeof(cpumap_buf));
-
-	scnprintf(buf, sizeof(buf), "--control=fd:%d,%d", control_fd[0], ack_fd[1]);
-
-	list_for_each_entry(t, &tpebs_results, nd)
-		tpebs_event_size++;
-
-	record_argv = calloc(12 + 2 * tpebs_event_size, sizeof(char *));
-	if (!record_argv)
-		return -ENOMEM;
-
-	ret = get_perf_record_args(record_argv, buf, cpumap_buf);
-	if (ret)
-		goto out;
+	record_argv[i++] = NULL;
+	assert(i == 10 + 2 * tpebs_event_size || i == 8 + 2 * tpebs_event_size);
+	/* Note, no workload given so system wide is implied. */
 
 	assert(tpebs_cmd.pid == 0);
 	tpebs_cmd.argv = record_argv;
 	tpebs_cmd.out = -1;
 	ret = start_command(&tpebs_cmd);
-out:
-	free(record_argv);
+	zfree(&tpebs_cmd.argv);
 	return ret;
 }
 
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#8 lore raw

PATCH v3 8/16 perf intel-tpebs: Ensure events are opened, factor out finding

In reply to Ian Rogers

Factor out finding an tpebs_retire_lat from an evsel. Don't blindly return when ignoring an open request, which happens after the first open request, ensure the event was started on a fork of perf record.

---
 tools/perf/util/intel-tpebs.c | 61 +++++++++++++++++++++--------------
 1 file changed, 37 insertions(+), 24 deletions(-)
tools/perf/util/intel-tpebs.c +37 -24
index c4c818f32239..e42f3ec39a64 100644
@@ -45,6 +45,8 @@ struct tpebs_retire_lat {
 	int sum;
 	/* Average of retire_latency, val = sum / count */
 	double val;
+	/* Has the event been sent to perf record? */
+	bool started;
 };
 
 static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[], int ack_fd[])
@@ -94,6 +96,9 @@ static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[],
 	tpebs_cmd.out = -1;
 	ret = start_command(&tpebs_cmd);
 	zfree(&tpebs_cmd.argv);
+	list_for_each_entry(t, &tpebs_results, nd)
+		t->started = true;
+
 	return ret;
 }
 
@@ -214,6 +219,19 @@ static struct tpebs_retire_lat *tpebs_retire_lat__new(struct evsel *evsel)
 	return result;
 }
 
+static struct tpebs_retire_lat *tpebs_retire_lat__find(struct evsel *evsel)
+{
+	struct tpebs_retire_lat *t;
+
+	list_for_each_entry(t, &tpebs_results, nd) {
+		if (t->tpebs_name == evsel->name ||
+		    !strcmp(t->tpebs_name, evsel->name) ||
+		    (evsel->metric_id && !strcmp(t->tpebs_name, evsel->metric_id)))
+			return t;
+	}
+	return NULL;
+}
+
 /**
  * evsel__tpebs_prepare - create tpebs data structures ready for opening.
  * @evsel: retire_latency evsel, all evsels on its list will be prepared.
@@ -221,16 +239,11 @@ static struct tpebs_retire_lat *tpebs_retire_lat__new(struct evsel *evsel)
 static int evsel__tpebs_prepare(struct evsel *evsel)
 {
 	struct evsel *pos;
-	struct tpebs_retire_lat *tpebs_event;
-
-	list_for_each_entry(tpebs_event, &tpebs_results, nd) {
-		if (!strcmp(tpebs_event->tpebs_name, evsel->name)) {
-			/*
-			 * evsel, or an identically named one, was already
-			 * prepared.
-			 */
-			return 0;
-		}
+	struct tpebs_retire_lat *tpebs_event = tpebs_retire_lat__find(evsel);
+
+	if (tpebs_event) {
+		/* evsel, or an identically named one, was already prepared. */
+		return 0;
 	}
 	tpebs_event = tpebs_retire_lat__new(evsel);
 	if (!tpebs_event)
@@ -262,12 +275,19 @@ int evsel__tpebs_open(struct evsel *evsel)
 {
 	int ret;
 
-	/*
-	 * We should only run tpebs_start when tpebs_recording is enabled.
-	 * And we should only run it once with all the required events.
-	 */
-	if (tpebs_cmd.pid != 0 || !tpebs_recording)
+	/* We should only run tpebs_start when tpebs_recording is enabled. */
+	if (!tpebs_recording)
 		return 0;
+	/* Only start the events once. */
+	if (tpebs_cmd.pid != 0) {
+		struct tpebs_retire_lat *t = tpebs_retire_lat__find(evsel);
+
+		if (!t || !t->started) {
+			/* Fail, as the event wasn't started. */
+			return -EBUSY;
+		}
+		return 0;
+	}
 
 	ret = evsel__tpebs_prepare(evsel);
 	if (ret)
@@ -352,7 +372,6 @@ int evsel__tpebs_open(struct evsel *evsel)
 int tpebs_set_evsel(struct evsel *evsel, int cpu_map_idx, int thread)
 {
 	__u64 val;
-	bool found = false;
 	struct tpebs_retire_lat *t;
 	struct perf_counts_values *count;
 
@@ -367,19 +386,13 @@ int tpebs_set_evsel(struct evsel *evsel, int cpu_map_idx, int thread)
 	tpebs_stop();
 	count = perf_counts(evsel->counts, cpu_map_idx, thread);
 
-	list_for_each_entry(t, &tpebs_results, nd) {
-		if (t->tpebs_name == evsel->name ||
-		    (evsel->metric_id && !strcmp(t->tpebs_name, evsel->metric_id))) {
-			found = true;
-			break;
-		}
-	}
+	t = tpebs_retire_lat__find(evsel);
 
 	/* Set ena and run to non-zero */
 	count->ena = count->run = 1;
 	count->lost = 0;
 
-	if (!found) {
+	if (!t) {
 		/*
 		 * Set default value or 0 when retire_latency for this event is
 		 * not found from sampling data (record_tpebs not set or 0
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#9 lore raw

PATCH v3 9/16 perf intel-tpebs: Refactor tpebs_results list

In reply to Ian Rogers

evsel names and metric-ids are used for matching but this can be problematic, for example, multiple occurrences of the same retirement latency event become a single event for the record. Change the name of the record events so they are unique and reflect the evsel of the retirement latency event that opens them (the retirement latency event's evsel address is embedded within them). This allows an evsel based close to close the event when the retirement latency event is closed. This is important as perf stat has an evlist and the session listen to the record events has an evlist, knowing which event should remove the tpebs_retire_lat can't be tied to an evlist list as there is more than 1, so closing which evlist should cause the tpebs to stop? Using the evsel and the last one out doing the tpebs_stop is cleaner.

---
 tools/perf/builtin-stat.c     |   2 -
 tools/perf/util/evlist.c      |   1 -
 tools/perf/util/evsel.c       |   2 +-
 tools/perf/util/intel-tpebs.c | 150 +++++++++++++++++++++-------------
 tools/perf/util/intel-tpebs.h |   2 +-
 5 files changed, 93 insertions(+), 64 deletions(-)
tools/perf/builtin-stat.c +0 -2
index 68ea7589c143..80e491bd775b 100644
@@ -681,8 +681,6 @@ static enum counter_recovery stat_handle_error(struct evsel *counter)
 	if (child_pid != -1)
 		kill(child_pid, SIGTERM);
 
-	tpebs_delete();
-
 	return COUNTER_FATAL;
 }
 
tools/perf/util/evlist.c +0 -1
index c1a04141aed0..0a21da4f990f 100644
@@ -183,7 +183,6 @@ void evlist__delete(struct evlist *evlist)
 	if (evlist == NULL)
 		return;
 
-	tpebs_delete();
 	evlist__free_stats(evlist);
 	evlist__munmap(evlist);
 	evlist__close(evlist);
tools/perf/util/evsel.c +1 -1
index 121283f2f382..554252ed1aab 100644
@@ -2759,7 +2759,7 @@ int evsel__open(struct evsel *evsel, struct perf_cpu_map *cpus,
 void evsel__close(struct evsel *evsel)
 {
 	if (evsel__is_retire_lat(evsel))
-		tpebs_delete();
+		evsel__tpebs_close(evsel);
 	perf_evsel__close(&evsel->core);
 	perf_evsel__free_id(&evsel->core);
 }
tools/perf/util/intel-tpebs.c +91 -59

233 lines not on this page · raw message

tools/perf/util/intel-tpebs.h +1 -1
index cc98203719c8..5c671181ec60 100644
@@ -11,7 +11,7 @@ struct evsel;
 extern bool tpebs_recording;
 
 int evsel__tpebs_open(struct evsel *evsel);
-void tpebs_delete(void);
+void evsel__tpebs_close(struct evsel *evsel);
 int tpebs_set_evsel(struct evsel *evsel, int cpu_map_idx, int thread);
 
 #endif /* __INTEL_TPEBS_H */
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#10 lore raw

PATCH v3 10/16 perf intel-tpebs: Add support for updating counts in evsel__tpebs_read

In reply to Ian Rogers

Rename to reflect evsel argument and for consistency with other tpebs functions. Update count from prev_raw_counts when available. Eventually this will allow inteval mode support.

---
 tools/perf/util/evsel.c       | 11 ++------
 tools/perf/util/intel-tpebs.c | 52 ++++++++++++++---------------------
 tools/perf/util/intel-tpebs.h |  2 +-
 3 files changed, 25 insertions(+), 40 deletions(-)
tools/perf/util/evsel.c +3 -8
index 554252ed1aab..1d343f51225b 100644
@@ -1718,11 +1718,6 @@ static int evsel__read_one(struct evsel *evsel, int cpu_map_idx, int thread)
 	return perf_evsel__read(&evsel->core, cpu_map_idx, thread, count);
 }
 
-static int evsel__read_retire_lat(struct evsel *evsel, int cpu_map_idx, int thread)
-{
-	return tpebs_set_evsel(evsel, cpu_map_idx, thread);
-}
-
 static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
 			     u64 val, u64 ena, u64 run, u64 lost)
 {
@@ -1730,8 +1725,8 @@ static void evsel__set_count(struct evsel *counter, int cpu_map_idx, int thread,
 
 	count = perf_counts(counter->counts, cpu_map_idx, thread);
 
-	if (counter->retire_lat) {
-		evsel__read_retire_lat(counter, cpu_map_idx, thread);
+	if (evsel__is_retire_lat(counter)) {
+		evsel__tpebs_read(counter, cpu_map_idx, thread);
 		perf_counts__set_loaded(counter->counts, cpu_map_idx, thread, true);
 		return;
 	}
@@ -1889,7 +1884,7 @@ int evsel__read_counter(struct evsel *evsel, int cpu_map_idx, int thread)
 		return evsel__hwmon_pmu_read(evsel, cpu_map_idx, thread);
 
 	if (evsel__is_retire_lat(evsel))
-		return evsel__read_retire_lat(evsel, cpu_map_idx, thread);
+		return evsel__tpebs_read(evsel, cpu_map_idx, thread);
 
 	if (evsel->core.attr.read_format & PERF_FORMAT_GROUP)
 		return evsel__read_group(evsel, cpu_map_idx, thread);
tools/perf/util/intel-tpebs.c +21 -31

70 lines not on this page · raw message

tools/perf/util/intel-tpebs.h +1 -1
index 5c671181ec60..218a82866cee 100644
@@ -12,6 +12,6 @@ extern bool tpebs_recording;
 
 int evsel__tpebs_open(struct evsel *evsel);
 void evsel__tpebs_close(struct evsel *evsel);
-int tpebs_set_evsel(struct evsel *evsel, int cpu_map_idx, int thread);
+int evsel__tpebs_read(struct evsel *evsel, int cpu_map_idx, int thread);
 
 #endif /* __INTEL_TPEBS_H */
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#11 lore raw

PATCH v3 11/16 perf intel-tpebs: Add mutex for tpebs_results

In reply to Ian Rogers

Ensure sample reader isn't racing with events being added/removed.

---
 tools/perf/util/intel-tpebs.c | 51 ++++++++++++++++++++++++++++++-----
 1 file changed, 44 insertions(+), 7 deletions(-)
tools/perf/util/intel-tpebs.c +44 -7
index f648fca17556..c5ccdbc42dc6 100644
@@ -16,6 +16,7 @@
 #include "debug.h"
 #include "evlist.h"
 #include "evsel.h"
+#include "mutex.h"
 #include "session.h"
 #include "tool.h"
 #include "cpumap.h"
@@ -32,6 +33,7 @@ bool tpebs_recording;
 static LIST_HEAD(tpebs_results);
 static pthread_t tpebs_reader_thread;
 static struct child_process tpebs_cmd;
+static struct mutex tpebs_mtx;
 
 struct tpebs_retire_lat {
 	struct list_head nd;
@@ -51,6 +53,19 @@ struct tpebs_retire_lat {
 
 static struct tpebs_retire_lat *tpebs_retire_lat__find(struct evsel *evsel);
 
+static void tpebs_mtx_init(void)
+{
+	mutex_init(&tpebs_mtx);
+}
+
+static struct mutex *tpebs_mtx_get(void)
+{
+	static pthread_once_t tpebs_mtx_once = PTHREAD_ONCE_INIT;
+
+	pthread_once(&tpebs_mtx_once, tpebs_mtx_init);
+	return &tpebs_mtx;
+}
+
 static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[], int ack_fd[])
 {
 	const char **record_argv;
@@ -59,13 +74,15 @@ static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[],
 	char cpumap_buf[50];
 	struct tpebs_retire_lat *t;
 
+	mutex_lock(tpebs_mtx_get());
 	list_for_each_entry(t, &tpebs_results, nd)
 		tpebs_event_size++;
 
 	record_argv = malloc((10 + 2 * tpebs_event_size) * sizeof(*record_argv));
-	if (!record_argv)
+	if (!record_argv) {
+		mutex_unlock(tpebs_mtx_get());
 		return -ENOMEM;
-
+	}
 	record_argv[i++] = "perf";
 	record_argv[i++] = "record";
 	record_argv[i++] = "-W";
@@ -101,6 +118,7 @@ static int evsel__tpebs_start_perf_record(struct evsel *evsel, int control_fd[],
 	list_for_each_entry(t, &tpebs_results, nd)
 		t->started = true;
 
+	mutex_unlock(tpebs_mtx_get());
 	return ret;
 }
 
@@ -112,9 +130,12 @@ static int process_sample_event(const struct perf_tool *tool __maybe_unused,
 {
 	struct tpebs_retire_lat *t;
 
+	mutex_lock(tpebs_mtx_get());
 	t = tpebs_retire_lat__find(evsel);
-	if (!t)
+	if (!t) {
+		mutex_unlock(tpebs_mtx_get());
 		return -EINVAL;
+	}
 	/*
 	 * Need to handle per core results? We are assuming average retire
 	 * latency value will be used. Save the number of samples and the sum of
@@ -123,6 +144,7 @@ static int process_sample_event(const struct perf_tool *tool __maybe_unused,
 	t->count += 1;
 	t->sum += sample->retire_lat;
 	t->val = (double) t->sum / t->count;
+	mutex_unlock(tpebs_mtx_get());
 	return 0;
 }
 
@@ -229,7 +251,6 @@ static struct tpebs_retire_lat *tpebs_retire_lat__new(struct evsel *evsel)
 		return NULL;
 	}
 	result->evsel = evsel;
-	list_add_tail(&result->nd, &tpebs_results);
 	return result;
 }
 
@@ -282,16 +303,22 @@ static struct tpebs_retire_lat *tpebs_retire_lat__find(struct evsel *evsel)
 static int evsel__tpebs_prepare(struct evsel *evsel)
 {
 	struct evsel *pos;
-	struct tpebs_retire_lat *tpebs_event = tpebs_retire_lat__find(evsel);
+	struct tpebs_retire_lat *tpebs_event;
 
+	mutex_lock(tpebs_mtx_get());
+	tpebs_event = tpebs_retire_lat__find(evsel);
 	if (tpebs_event) {
 		/* evsel, or an identically named one, was already prepared. */
+		mutex_unlock(tpebs_mtx_get());
 		return 0;
 	}
 	tpebs_event = tpebs_retire_lat__new(evsel);
 	if (!tpebs_event)
 		return -ENOMEM;
 
+	list_add_tail(&tpebs_event->nd, &tpebs_results);
+	mutex_unlock(tpebs_mtx_get());
+
 	/*
 	 * Eagerly prepare all other evsels on the list to try to ensure that by
 	 * open they are all known.
@@ -317,6 +344,7 @@ static int evsel__tpebs_prepare(struct evsel *evsel)
 int evsel__tpebs_open(struct evsel *evsel)
 {
 	int ret;
+	bool tpebs_empty;
 
 	/* We should only run tpebs_start when tpebs_recording is enabled. */
 	if (!tpebs_recording)
@@ -336,7 +364,10 @@ int evsel__tpebs_open(struct evsel *evsel)
 	if (ret)
 		return ret;
 
-	if (!list_empty(&tpebs_results)) {
+	mutex_lock(tpebs_mtx_get());
+	tpebs_empty = list_empty(&tpebs_results);
+	mutex_unlock(tpebs_mtx_get());
+	if (!tpebs_empty) {
 		struct pollfd pollfd = { .events = POLLIN, };
 		int control_fd[2], ack_fd[2], len;
 		char ack_buf[8];
@@ -436,8 +467,10 @@ int evsel__tpebs_read(struct evsel *evsel, int cpu_map_idx, int thread)
 	 */
 	tpebs_stop();
 
+	mutex_lock(tpebs_mtx_get());
 	t = tpebs_retire_lat__find(evsel);
 	val = rint(t->val);
+	mutex_unlock(tpebs_mtx_get());
 
 	if (old_count) {
 		count->val = old_count->val + val;
@@ -460,9 +493,13 @@ int evsel__tpebs_read(struct evsel *evsel, int cpu_map_idx, int thread)
  */
 void evsel__tpebs_close(struct evsel *evsel)
 {
-	struct tpebs_retire_lat *t = tpebs_retire_lat__find(evsel);
+	struct tpebs_retire_lat *t;
 
+	mutex_lock(tpebs_mtx_get());
+	t = tpebs_retire_lat__find(evsel);
+	list_del_init(&t->nd);
 	tpebs_retire_lat__delete(t);
+	mutex_unlock(tpebs_mtx_get());
 
 	if (list_empty(&tpebs_results))
 		tpebs_stop();
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#12 lore raw

PATCH v3 12/16 perf intel-tpebs: Don't close record on read

In reply to Ian Rogers

Factor sending record control fd code into its own function. Rather than killing the record process send it a ping when reading. Timeouts were witnessed if done too frequently, so only ping for the first tpebs events. Don't kill the record command send it a stop command. As close isn't reliably called also close on evsel__exit. Add extra checks on the record being terminated to avoid warnings. Adjust the locking as needed and incorporate extra -Wthread-safety checks. Check to do six 500ms poll timeouts when sending commands, rather than the larger 3000ms, to allow the record process terminating to be better witnessed.

---
 tools/perf/util/evsel.c       |   2 +
 tools/perf/util/intel-tpebs.c | 205 +++++++++++++++++++++-------------
 2 files changed, 132 insertions(+), 75 deletions(-)
tools/perf/util/evsel.c +2 -0
index 1d343f51225b..661a07cbdb25 100644
@@ -1656,6 +1656,8 @@ void evsel__exit(struct evsel *evsel)
 {
 	assert(list_empty(&evsel->core.node));
 	assert(evsel->evlist == NULL);
+	if (evsel__is_retire_lat(evsel))
+		evsel__tpebs_close(evsel);
 	bpf_counter__destroy(evsel);
 	perf_bpf_filter__destroy(evsel);
 	evsel__free_counts(evsel);
tools/perf/util/intel-tpebs.c +130 -75

320 lines not on this page · raw message

Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#13 lore raw

PATCH v3 13/16 perf intel-tpebs: Use stats for retirement latency statistics

In reply to Ian Rogers

struct stats provides access to mean, min and max. It also provides uniformity with statistics code used elsewhere in perf.

---
 tools/perf/util/intel-tpebs.c | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)
tools/perf/util/intel-tpebs.c +5 -10
index a9446e7a1c5c..de9fea601964 100644
@@ -18,6 +18,7 @@
 #include "evsel.h"
 #include "mutex.h"
 #include "session.h"
+#include "stat.h"
 #include "tool.h"
 #include "cpumap.h"
 #include "metricgroup.h"
@@ -42,12 +43,8 @@ struct tpebs_retire_lat {
 	struct evsel *evsel;
 	/** @event: Event passed to perf record. */
 	char *event;
-	/* Count of retire_latency values found in sample data */
-	size_t count;
-	/* Sum of all the retire_latency values in sample data */
-	int sum;
-	/* Average of retire_latency, val = sum / count */
-	double val;
+	/** @stats: Recorded retirement latency stats. */
+	struct stats stats;
 	/* Has the event been sent to perf record? */
 	bool started;
 };
@@ -145,9 +142,7 @@ static int process_sample_event(const struct perf_tool *tool __maybe_unused,
 	 * latency value will be used. Save the number of samples and the sum of
 	 * retire latency value for each event.
 	 */
-	t->count += 1;
-	t->sum += sample->retire_lat;
-	t->val = (double) t->sum / t->count;
+	update_stats(&t->stats, sample->retire_lat);
 	mutex_unlock(tpebs_mtx_get());
 	return 0;
 }
@@ -522,7 +517,7 @@ int evsel__tpebs_read(struct evsel *evsel, int cpu_map_idx, int thread)
 			return ret;
 		mutex_lock(tpebs_mtx_get());
 	}
-	val = rint(t->val);
+	val = rint(t->stats.mean);
 	mutex_unlock(tpebs_mtx_get());
 
 	if (old_count) {
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#14 lore raw

PATCH v3 14/16 perf stat: Add mean, min, max and last --tpebs-mode options

In reply to Ian Rogers

Add command line configuration option for how retirement latency events are combined. The default "mean" gives the average of retirement latency. "min" or "max" give the smallest or largest retirment latency times respectively. "last" uses the last retirment latency sample's time.

---
 tools/perf/Documentation/perf-stat.txt |  7 +++++++
 tools/perf/builtin-stat.c              | 27 ++++++++++++++++++++++++++
 tools/perf/util/intel-tpebs.c          | 20 ++++++++++++++++++-
 tools/perf/util/intel-tpebs.h          |  8 ++++++++
 4 files changed, 61 insertions(+), 1 deletion(-)
tools/perf/Documentation/perf-stat.txt +7 -0
index 2bc063672486..61d091670dee 100644
@@ -506,6 +506,13 @@ this option is not set. The TPEBS hardware feature starts from Intel Granite
 Rapids microarchitecture. This option only exists in X86_64 and is meaningful on
 Intel platforms with TPEBS feature.
 
+--tpebs-mode=[mean|min|max|last]::
+Set how retirement latency events have their sample times
+combined. The default "mean" gives the average of retirement
+latency. "min" or "max" give the smallest or largest retirment latency
+times respectively. "last" uses the last retirment latency sample's
+time.
+
 --td-level::
 Print the top-down statistics that equal the input level. It allows
 users to print the interested top-down metrics level instead of the
tools/perf/builtin-stat.c +27 -0
index 80e491bd775b..4adf2ae53b11 100644
@@ -2327,6 +2327,30 @@ static void setup_system_wide(int forks)
 	}
 }
 
+static int parse_tpebs_mode(const struct option *opt, const char *str,
+			    int unset __maybe_unused)
+{
+	enum tpebs_mode *mode = opt->value;
+
+	if (!strcasecmp("mean", str)) {
+		*mode = TPEBS_MODE__MEAN;
+		return 0;
+	}
+	if (!strcasecmp("min", str)) {
+		*mode = TPEBS_MODE__MIN;
+		return 0;
+	}
+	if (!strcasecmp("max", str)) {
+		*mode = TPEBS_MODE__MAX;
+		return 0;
+	}
+	if (!strcasecmp("last", str)) {
+		*mode = TPEBS_MODE__LAST;
+		return 0;
+	}
+	return -1;
+}
+
 int cmd_stat(int argc, const char **argv)
 {
 	struct opt_aggr_mode opt_mode = {};
@@ -2431,6 +2455,9 @@ int cmd_stat(int argc, const char **argv)
 #ifdef HAVE_ARCH_X86_64_SUPPORT
 		OPT_BOOLEAN(0, "record-tpebs", &tpebs_recording,
 			"enable recording for tpebs when retire_latency required"),
+		OPT_CALLBACK(0, "tpebs-mode", &tpebs_mode, "tpebs-mode",
+			"Mode of TPEBS recording: mean, min or max",
+			parse_tpebs_mode),
 #endif
 		OPT_UINTEGER(0, "td-level", &stat_config.topdown_level,
 			"Set the metrics level for the top-down statistics (0: max level)"),
tools/perf/util/intel-tpebs.c +19 -1
index de9fea601964..6b00bd5b0af1 100644
@@ -31,6 +31,7 @@
 #define PERF_DATA		"-"
 
 bool tpebs_recording;
+enum tpebs_mode tpebs_mode;
 static LIST_HEAD(tpebs_results);
 static pthread_t tpebs_reader_thread;
 static struct child_process tpebs_cmd;
@@ -45,6 +46,8 @@ struct tpebs_retire_lat {
 	char *event;
 	/** @stats: Recorded retirement latency stats. */
 	struct stats stats;
+	/** @last: Last retirement latency read. */
+	uint64_t last;
 	/* Has the event been sent to perf record? */
 	bool started;
 };
@@ -142,6 +145,7 @@ static int process_sample_event(const struct perf_tool *tool __maybe_unused,
 	 * latency value will be used. Save the number of samples and the sum of
 	 * retire latency value for each event.
 	 */
+	t->last = sample->retire_lat;
 	update_stats(&t->stats, sample->retire_lat);
 	mutex_unlock(tpebs_mtx_get());
 	return 0;
@@ -517,7 +521,21 @@ int evsel__tpebs_read(struct evsel *evsel, int cpu_map_idx, int thread)
 			return ret;
 		mutex_lock(tpebs_mtx_get());
 	}
-	val = rint(t->stats.mean);
+	switch (tpebs_mode) {
+	case TPEBS_MODE__MIN:
+		val = rint(t->stats.min);
+		break;
+	case TPEBS_MODE__MAX:
+		val = rint(t->stats.max);
+		break;
+	case TPEBS_MODE__LAST:
+		val = t->last;
+		break;
+	default:
+	case TPEBS_MODE__MEAN:
+		val = rint(t->stats.mean);
+		break;
+	}
 	mutex_unlock(tpebs_mtx_get());
 
 	if (old_count) {
tools/perf/util/intel-tpebs.h +8 -0
index 218a82866cee..9475e2e6ea74 100644
@@ -8,7 +8,15 @@
 struct evlist;
 struct evsel;
 
+enum tpebs_mode {
+	TPEBS_MODE__MEAN,
+	TPEBS_MODE__MIN,
+	TPEBS_MODE__MAX,
+	TPEBS_MODE__LAST,
+};
+
 extern bool tpebs_recording;
+extern enum tpebs_mode tpebs_mode;
 
 int evsel__tpebs_open(struct evsel *evsel);
 void evsel__tpebs_close(struct evsel *evsel);
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#15 lore raw

PATCH v3 15/16 perf pmu-events: Add retirement latency to JSON events inside of perf

In reply to Ian Rogers

The updated Intel vendor events add retirement latency for graniterapids:
https://lore.kernel.org/lkml/[email protected]/
This change makes those values available within an alias/event within a PMU and saves them into the evsel at event parse time. When no TPEBS data is available the default values are substituted in for TMA metrics that are using retirement latency events - currently just those on graniterapids. If the default values are used the min or max in the sampled statistics are also updated so that the max never decreases or the min increases.

---
 tools/perf/pmu-events/empty-pmu-events.c | 216 ++++++++++++-----------
 tools/perf/pmu-events/jevents.py         |   6 +
 tools/perf/pmu-events/pmu-events.h       |   3 +
 tools/perf/util/evsel.h                  |   6 +
 tools/perf/util/intel-tpebs.c            |  52 ++++--
 tools/perf/util/parse-events.c           |   4 +
 tools/perf/util/pmu.c                    |  52 +++++-
 tools/perf/util/pmu.h                    |   3 +
 8 files changed, 220 insertions(+), 122 deletions(-)
tools/perf/pmu-events/empty-pmu-events.c +111 -105
index 0cb7ba7912e8..0361bcc1eb58 100644
@@ -20,73 +20,73 @@ struct pmu_table_entry {
 
 static const char *const big_c_string =
 /* offset=0 */ "tool\000"
-/* offset=5 */ "duration_time\000tool\000Wall clock interval time in nanoseconds\000config=1\000\00000\000\000"
-/* offset=78 */ "user_time\000tool\000User (non-kernel) time in nanoseconds\000config=2\000\00000\000\000"
-/* offset=145 */ "system_time\000tool\000System/kernel time in nanoseconds\000config=3\000\00000\000\000"
-/* offset=210 */ "has_pmem\000tool\0001 if persistent memory installed otherwise 0\000config=4\000\00000\000\000"
-/* offset=283 */ "num_cores\000tool\000Number of cores. A core consists of 1 or more thread, with each thread being associated with a logical Linux CPU\000config=5\000\00000\000\000"
-/* offset=425 */ "num_cpus\000tool\000Number of logical Linux CPUs. There may be multiple such CPUs on a core\000config=6\000\00000\000\000"
-/* offset=525 */ "num_cpus_online\000tool\000Number of online logical Linux CPUs. There may be multiple such CPUs on a core\000config=7\000\00000\000\000"
-/* offset=639 */ "num_dies\000tool\000Number of dies. Each die has 1 or more cores\000config=8\000\00000\000\000"
-/* offset=712 */ "num_packages\000tool\000Number of packages. Each package has 1 or more die\000config=9\000\00000\000\000"
-/* offset=795 */ "slots\000tool\000Number of functional units that in parallel can execute parts of an instruction\000config=0xa\000\00000\000\000"
-/* offset=902 */ "smt_on\000tool\0001 if simultaneous multithreading (aka hyperthreading) is enable otherwise 0\000config=0xb\000\00000\000\000"
-/* offset=1006 */ "system_tsc_freq\000tool\000The amount a Time Stamp Counter (TSC) increases per second\000config=0xc\000\00000\000\000"
-/* offset=1102 */ "default_core\000"
-/* offset=1115 */ "bp_l1_btb_correct\000branch\000L1 BTB Correction\000event=0x8a\000\00000\000\000"
-/* offset=1174 */ "bp_l2_btb_correct\000branch\000L2 BTB Correction\000event=0x8b\000\00000\000\000"
-/* offset=1233 */ "l3_cache_rd\000cache\000L3 cache access, read\000event=0x40\000\00000\000Attributable Level 3 cache access, read\000"
-/* offset=1328 */ "segment_reg_loads.any\000other\000Number of segment register loads\000event=6,period=200000,umask=0x80\000\00000\000\000"
-/* offset=1427 */ "dispatch_blocked.any\000other\000Memory cluster signals to block micro-op dispatch for any reason\000event=9,period=200000,umask=0x20\000\00000\000\000"
-/* offset=1557 */ "eist_trans\000other\000Number of Enhanced Intel SpeedStep(R) Technology (EIST) transitions\000event=0x3a,period=200000\000\00000\000\000"
-/* offset=1672 */ "hisi_sccl,ddrc\000"
-/* offset=1687 */ "uncore_hisi_ddrc.flux_wcmd\000uncore\000DDRC write commands\000event=2\000\00000\000DDRC write commands\000"
-/* offset=1773 */ "uncore_cbox\000"
-/* offset=1785 */ "unc_cbo_xsnp_response.miss_eviction\000uncore\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000event=0x22,umask=0x81\000\00000\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000"
-/* offset=2016 */ "event-hyphen\000uncore\000UNC_CBO_HYPHEN\000event=0xe0\000\00000\000UNC_CBO_HYPHEN\000"
-/* offset=2081 */ "event-two-hyph\000uncore\000UNC_CBO_TWO_HYPH\000event=0xc0\000\00000\000UNC_CBO_TWO_HYPH\000"
-/* offset=2152 */ "hisi_sccl,l3c\000"
-/* offset=2166 */ "uncore_hisi_l3c.rd_hit_cpipe\000uncore\000Total read hits\000event=7\000\00000\000Total read hits\000"
-/* offset=2246 */ "uncore_imc_free_running\000"
-/* offset=2270 */ "uncore_imc_free_running.cache_miss\000uncore\000Total cache misses\000event=0x12\000\00000\000Total cache misses\000"
-/* offset=2365 */ "uncore_imc\000"
-/* offset=2376 */ "uncore_imc.cache_hits\000uncore\000Total cache hits\000event=0x34\000\00000\000Total cache hits\000"
-/* offset=2454 */ "uncore_sys_ddr_pmu\000"
-/* offset=2473 */ "sys_ddr_pmu.write_cycles\000uncore\000ddr write-cycles event\000event=0x2b\000v8\00000\000\000"
-/* offset=2546 */ "uncore_sys_ccn_pmu\000"
-/* offset=2565 */ "sys_ccn_pmu.read_cycles\000uncore\000ccn read-cycles event\000config=0x2c\0000x01\00000\000\000"
-/* offset=2639 */ "uncore_sys_cmn_pmu\000"
-/* offset=2658 */ "sys_cmn_pmu.hnf_cache_miss\000uncore\000Counts total cache misses in first lookup result (high priority)\000eventid=1,type=5\000(434|436|43c|43a).*\00000\000\000"
-/* offset=2798 */ "CPI\000\0001 / IPC\000\000\000\000\000\000\000\00000"
-/* offset=2820 */ "IPC\000group1\000inst_retired.any / cpu_clk_unhalted.thread\000\000\000\000\000\000\000\00000"
-/* offset=2883 */ "Frontend_Bound_SMT\000\000idq_uops_not_delivered.core / (4 * (cpu_clk_unhalted.thread / 2 * (1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk)))\000\000\000\000\000\000\000\00000"
-/* offset=3049 */ "dcache_miss_cpi\000\000l1d\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000"
-/* offset=3113 */ "icache_miss_cycles\000\000l1i\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000"
-/* offset=3180 */ "cache_miss_cycles\000group1\000dcache_miss_cpi + icache_miss_cycles\000\000\000\000\000\000\000\00000"
-/* offset=3251 */ "DCache_L2_All_Hits\000\000l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit\000\000\000\000\000\000\000\00000"
-/* offset=3345 */ "DCache_L2_All_Miss\000\000max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + l2_rqsts.pf_miss + l2_rqsts.rfo_miss\000\000\000\000\000\000\000\00000"
-/* offset=3479 */ "DCache_L2_All\000\000DCache_L2_All_Hits + DCache_L2_All_Miss\000\000\000\000\000\000\000\00000"
-/* offset=3543 */ "DCache_L2_Hits\000\000d_ratio(DCache_L2_All_Hits, DCache_L2_All)\000\000\000\000\000\000\000\00000"
-/* offset=3611 */ "DCache_L2_Misses\000\000d_ratio(DCache_L2_All_Miss, DCache_L2_All)\000\000\000\000\000\000\000\00000"
-/* offset=3681 */ "M1\000\000ipc + M2\000\000\000\000\000\000\000\00000"
-/* offset=3703 */ "M2\000\000ipc + M1\000\000\000\000\000\000\000\00000"
-/* offset=3725 */ "M3\000\0001 / M3\000\000\000\000\000\000\000\00000"
-/* offset=3745 */ "L1D_Cache_Fill_BW\000\00064 * l1d.replacement / 1e9 / duration_time\000\000\000\000\000\000\000\00000"
+/* offset=5 */ "duration_time\000tool\000Wall clock interval time in nanoseconds\000config=1\000\00000\000\000\000\000\000"
+/* offset=81 */ "user_time\000tool\000User (non-kernel) time in nanoseconds\000config=2\000\00000\000\000\000\000\000"
+/* offset=151 */ "system_time\000tool\000System/kernel time in nanoseconds\000config=3\000\00000\000\000\000\000\000"
+/* offset=219 */ "has_pmem\000tool\0001 if persistent memory installed otherwise 0\000config=4\000\00000\000\000\000\000\000"
+/* offset=295 */ "num_cores\000tool\000Number of cores. A core consists of 1 or more thread, with each thread being associated with a logical Linux CPU\000config=5\000\00000\000\000\000\000\000"
+/* offset=440 */ "num_cpus\000tool\000Number of logical Linux CPUs. There may be multiple such CPUs on a core\000config=6\000\00000\000\000\000\000\000"
+/* offset=543 */ "num_cpus_online\000tool\000Number of online logical Linux CPUs. There may be multiple such CPUs on a core\000config=7\000\00000\000\000\000\000\000"
+/* offset=660 */ "num_dies\000tool\000Number of dies. Each die has 1 or more cores\000config=8\000\00000\000\000\000\000\000"
+/* offset=736 */ "num_packages\000tool\000Number of packages. Each package has 1 or more die\000config=9\000\00000\000\000\000\000\000"
+/* offset=822 */ "slots\000tool\000Number of functional units that in parallel can execute parts of an instruction\000config=0xa\000\00000\000\000\000\000\000"
+/* offset=932 */ "smt_on\000tool\0001 if simultaneous multithreading (aka hyperthreading) is enable otherwise 0\000config=0xb\000\00000\000\000\000\000\000"
+/* offset=1039 */ "system_tsc_freq\000tool\000The amount a Time Stamp Counter (TSC) increases per second\000config=0xc\000\00000\000\000\000\000\000"
+/* offset=1138 */ "default_core\000"
+/* offset=1151 */ "bp_l1_btb_correct\000branch\000L1 BTB Correction\000event=0x8a\000\00000\000\000\000\000\000"
+/* offset=1213 */ "bp_l2_btb_correct\000branch\000L2 BTB Correction\000event=0x8b\000\00000\000\000\000\000\000"
+/* offset=1275 */ "l3_cache_rd\000cache\000L3 cache access, read\000event=0x40\000\00000\000\000\000\000Attributable Level 3 cache access, read\000"
+/* offset=1373 */ "segment_reg_loads.any\000other\000Number of segment register loads\000event=6,period=200000,umask=0x80\000\00000\000\000\000\000\000"
+/* offset=1475 */ "dispatch_blocked.any\000other\000Memory cluster signals to block micro-op dispatch for any reason\000event=9,period=200000,umask=0x20\000\00000\000\000\000\000\000"
+/* offset=1608 */ "eist_trans\000other\000Number of Enhanced Intel SpeedStep(R) Technology (EIST) transitions\000event=0x3a,period=200000\000\00000\000\000\000\000\000"
+/* offset=1726 */ "hisi_sccl,ddrc\000"
+/* offset=1741 */ "uncore_hisi_ddrc.flux_wcmd\000uncore\000DDRC write commands\000event=2\000\00000\000\000\000\000DDRC write commands\000"
+/* offset=1830 */ "uncore_cbox\000"
+/* offset=1842 */ "unc_cbo_xsnp_response.miss_eviction\000uncore\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000event=0x22,umask=0x81\000\00000\000\000\000\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000"
+/* offset=2076 */ "event-hyphen\000uncore\000UNC_CBO_HYPHEN\000event=0xe0\000\00000\000\000\000\000UNC_CBO_HYPHEN\000"
+/* offset=2144 */ "event-two-hyph\000uncore\000UNC_CBO_TWO_HYPH\000event=0xc0\000\00000\000\000\000\000UNC_CBO_TWO_HYPH\000"
+/* offset=2218 */ "hisi_sccl,l3c\000"
+/* offset=2232 */ "uncore_hisi_l3c.rd_hit_cpipe\000uncore\000Total read hits\000event=7\000\00000\000\000\000\000Total read hits\000"
+/* offset=2315 */ "uncore_imc_free_running\000"
+/* offset=2339 */ "uncore_imc_free_running.cache_miss\000uncore\000Total cache misses\000event=0x12\000\00000\000\000\000\000Total cache misses\000"
+/* offset=2437 */ "uncore_imc\000"
+/* offset=2448 */ "uncore_imc.cache_hits\000uncore\000Total cache hits\000event=0x34\000\00000\000\000\000\000Total cache hits\000"
+/* offset=2529 */ "uncore_sys_ddr_pmu\000"
+/* offset=2548 */ "sys_ddr_pmu.write_cycles\000uncore\000ddr write-cycles event\000event=0x2b\000v8\00000\000\000\000\000\000"
+/* offset=2624 */ "uncore_sys_ccn_pmu\000"
+/* offset=2643 */ "sys_ccn_pmu.read_cycles\000uncore\000ccn read-cycles event\000config=0x2c\0000x01\00000\000\000\000\000\000"
+/* offset=2720 */ "uncore_sys_cmn_pmu\000"
+/* offset=2739 */ "sys_cmn_pmu.hnf_cache_miss\000uncore\000Counts total cache misses in first lookup result (high priority)\000eventid=1,type=5\000(434|436|43c|43a).*\00000\000\000\000\000\000"
+/* offset=2882 */ "CPI\000\0001 / IPC\000\000\000\000\000\000\000\00000"
+/* offset=2904 */ "IPC\000group1\000inst_retired.any / cpu_clk_unhalted.thread\000\000\000\000\000\000\000\00000"
+/* offset=2967 */ "Frontend_Bound_SMT\000\000idq_uops_not_delivered.core / (4 * (cpu_clk_unhalted.thread / 2 * (1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk)))\000\000\000\000\000\000\000\00000"
+/* offset=3133 */ "dcache_miss_cpi\000\000l1d\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000"
+/* offset=3197 */ "icache_miss_cycles\000\000l1i\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000"
+/* offset=3264 */ "cache_miss_cycles\000group1\000dcache_miss_cpi + icache_miss_cycles\000\000\000\000\000\000\000\00000"
+/* offset=3335 */ "DCache_L2_All_Hits\000\000l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit\000\000\000\000\000\000\000\00000"
+/* offset=3429 */ "DCache_L2_All_Miss\000\000max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + l2_rqsts.pf_miss + l2_rqsts.rfo_miss\000\000\000\000\000\000\000\00000"
+/* offset=3563 */ "DCache_L2_All\000\000DCache_L2_All_Hits + DCache_L2_All_Miss\000\000\000\000\000\000\000\00000"
+/* offset=3627 */ "DCache_L2_Hits\000\000d_ratio(DCache_L2_All_Hits, DCache_L2_All)\000\000\000\000\000\000\000\00000"
+/* offset=3695 */ "DCache_L2_Misses\000\000d_ratio(DCache_L2_All_Miss, DCache_L2_All)\000\000\000\000\000\000\000\00000"
+/* offset=3765 */ "M1\000\000ipc + M2\000\000\000\000\000\000\000\00000"
+/* offset=3787 */ "M2\000\000ipc + M1\000\000\000\000\000\000\000\00000"
+/* offset=3809 */ "M3\000\0001 / M3\000\000\000\000\000\000\000\00000"
+/* offset=3829 */ "L1D_Cache_Fill_BW\000\00064 * l1d.replacement / 1e9 / duration_time\000\000\000\000\000\000\000\00000"
 ;
 
 static const struct compact_pmu_event pmu_events__common_tool[] = {
-{ 5 }, /* duration_time\000tool\000Wall clock interval time in nanoseconds\000config=1\000\00000\000\000 */
-{ 210 }, /* has_pmem\000tool\0001 if persistent memory installed otherwise 0\000config=4\000\00000\000\000 */
-{ 283 }, /* num_cores\000tool\000Number of cores. A core consists of 1 or more thread, with each thread being associated with a logical Linux CPU\000config=5\000\00000\000\000 */
-{ 425 }, /* num_cpus\000tool\000Number of logical Linux CPUs. There may be multiple such CPUs on a core\000config=6\000\00000\000\000 */
-{ 525 }, /* num_cpus_online\000tool\000Number of online logical Linux CPUs. There may be multiple such CPUs on a core\000config=7\000\00000\000\000 */
-{ 639 }, /* num_dies\000tool\000Number of dies. Each die has 1 or more cores\000config=8\000\00000\000\000 */
-{ 712 }, /* num_packages\000tool\000Number of packages. Each package has 1 or more die\000config=9\000\00000\000\000 */
-{ 795 }, /* slots\000tool\000Number of functional units that in parallel can execute parts of an instruction\000config=0xa\000\00000\000\000 */
-{ 902 }, /* smt_on\000tool\0001 if simultaneous multithreading (aka hyperthreading) is enable otherwise 0\000config=0xb\000\00000\000\000 */
-{ 145 }, /* system_time\000tool\000System/kernel time in nanoseconds\000config=3\000\00000\000\000 */
-{ 1006 }, /* system_tsc_freq\000tool\000The amount a Time Stamp Counter (TSC) increases per second\000config=0xc\000\00000\000\000 */
-{ 78 }, /* user_time\000tool\000User (non-kernel) time in nanoseconds\000config=2\000\00000\000\000 */
+{ 5 }, /* duration_time\000tool\000Wall clock interval time in nanoseconds\000config=1\000\00000\000\000\000\000\000 */
+{ 219 }, /* has_pmem\000tool\0001 if persistent memory installed otherwise 0\000config=4\000\00000\000\000\000\000\000 */
+{ 295 }, /* num_cores\000tool\000Number of cores. A core consists of 1 or more thread, with each thread being associated with a logical Linux CPU\000config=5\000\00000\000\000\000\000\000 */
+{ 440 }, /* num_cpus\000tool\000Number of logical Linux CPUs. There may be multiple such CPUs on a core\000config=6\000\00000\000\000\000\000\000 */
+{ 543 }, /* num_cpus_online\000tool\000Number of online logical Linux CPUs. There may be multiple such CPUs on a core\000config=7\000\00000\000\000\000\000\000 */
+{ 660 }, /* num_dies\000tool\000Number of dies. Each die has 1 or more cores\000config=8\000\00000\000\000\000\000\000 */
+{ 736 }, /* num_packages\000tool\000Number of packages. Each package has 1 or more die\000config=9\000\00000\000\000\000\000\000 */
+{ 822 }, /* slots\000tool\000Number of functional units that in parallel can execute parts of an instruction\000config=0xa\000\00000\000\000\000\000\000 */
+{ 932 }, /* smt_on\000tool\0001 if simultaneous multithreading (aka hyperthreading) is enable otherwise 0\000config=0xb\000\00000\000\000\000\000\000 */
+{ 151 }, /* system_time\000tool\000System/kernel time in nanoseconds\000config=3\000\00000\000\000\000\000\000 */
+{ 1039 }, /* system_tsc_freq\000tool\000The amount a Time Stamp Counter (TSC) increases per second\000config=0xc\000\00000\000\000\000\000\000 */
+{ 81 }, /* user_time\000tool\000User (non-kernel) time in nanoseconds\000config=2\000\00000\000\000\000\000\000 */
 
 };
 
@@ -99,29 +99,29 @@ const struct pmu_table_entry pmu_events__common[] = {
 };
 
 static const struct compact_pmu_event pmu_events__test_soc_cpu_default_core[] = {
-{ 1115 }, /* bp_l1_btb_correct\000branch\000L1 BTB Correction\000event=0x8a\000\00000\000\000 */
-{ 1174 }, /* bp_l2_btb_correct\000branch\000L2 BTB Correction\000event=0x8b\000\00000\000\000 */
-{ 1427 }, /* dispatch_blocked.any\000other\000Memory cluster signals to block micro-op dispatch for any reason\000event=9,period=200000,umask=0x20\000\00000\000\000 */
-{ 1557 }, /* eist_trans\000other\000Number of Enhanced Intel SpeedStep(R) Technology (EIST) transitions\000event=0x3a,period=200000\000\00000\000\000 */
-{ 1233 }, /* l3_cache_rd\000cache\000L3 cache access, read\000event=0x40\000\00000\000Attributable Level 3 cache access, read\000 */
-{ 1328 }, /* segment_reg_loads.any\000other\000Number of segment register loads\000event=6,period=200000,umask=0x80\000\00000\000\000 */
+{ 1151 }, /* bp_l1_btb_correct\000branch\000L1 BTB Correction\000event=0x8a\000\00000\000\000\000\000\000 */
+{ 1213 }, /* bp_l2_btb_correct\000branch\000L2 BTB Correction\000event=0x8b\000\00000\000\000\000\000\000 */
+{ 1475 }, /* dispatch_blocked.any\000other\000Memory cluster signals to block micro-op dispatch for any reason\000event=9,period=200000,umask=0x20\000\00000\000\000\000\000\000 */
+{ 1608 }, /* eist_trans\000other\000Number of Enhanced Intel SpeedStep(R) Technology (EIST) transitions\000event=0x3a,period=200000\000\00000\000\000\000\000\000 */
+{ 1275 }, /* l3_cache_rd\000cache\000L3 cache access, read\000event=0x40\000\00000\000\000\000\000Attributable Level 3 cache access, read\000 */
+{ 1373 }, /* segment_reg_loads.any\000other\000Number of segment register loads\000event=6,period=200000,umask=0x80\000\00000\000\000\000\000\000 */
 };
 static const struct compact_pmu_event pmu_events__test_soc_cpu_hisi_sccl_ddrc[] = {
-{ 1687 }, /* uncore_hisi_ddrc.flux_wcmd\000uncore\000DDRC write commands\000event=2\000\00000\000DDRC write commands\000 */
+{ 1741 }, /* uncore_hisi_ddrc.flux_wcmd\000uncore\000DDRC write commands\000event=2\000\00000\000\000\000\000DDRC write commands\000 */
 };
 static const struct compact_pmu_event pmu_events__test_soc_cpu_hisi_sccl_l3c[] = {
-{ 2166 }, /* uncore_hisi_l3c.rd_hit_cpipe\000uncore\000Total read hits\000event=7\000\00000\000Total read hits\000 */
+{ 2232 }, /* uncore_hisi_l3c.rd_hit_cpipe\000uncore\000Total read hits\000event=7\000\00000\000\000\000\000Total read hits\000 */
 };
 static const struct compact_pmu_event pmu_events__test_soc_cpu_uncore_cbox[] = {
-{ 2016 }, /* event-hyphen\000uncore\000UNC_CBO_HYPHEN\000event=0xe0\000\00000\000UNC_CBO_HYPHEN\000 */
-{ 2081 }, /* event-two-hyph\000uncore\000UNC_CBO_TWO_HYPH\000event=0xc0\000\00000\000UNC_CBO_TWO_HYPH\000 */
-{ 1785 }, /* unc_cbo_xsnp_response.miss_eviction\000uncore\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000event=0x22,umask=0x81\000\00000\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000 */
+{ 2076 }, /* event-hyphen\000uncore\000UNC_CBO_HYPHEN\000event=0xe0\000\00000\000\000\000\000UNC_CBO_HYPHEN\000 */
+{ 2144 }, /* event-two-hyph\000uncore\000UNC_CBO_TWO_HYPH\000event=0xc0\000\00000\000\000\000\000UNC_CBO_TWO_HYPH\000 */
+{ 1842 }, /* unc_cbo_xsnp_response.miss_eviction\000uncore\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000event=0x22,umask=0x81\000\00000\000\000\000\000A cross-core snoop resulted from L3 Eviction which misses in some processor core\000 */
 };
 static const struct compact_pmu_event pmu_events__test_soc_cpu_uncore_imc[] = {
-{ 2376 }, /* uncore_imc.cache_hits\000uncore\000Total cache hits\000event=0x34\000\00000\000Total cache hits\000 */
+{ 2448 }, /* uncore_imc.cache_hits\000uncore\000Total cache hits\000event=0x34\000\00000\000\000\000\000Total cache hits\000 */
 };
 static const struct compact_pmu_event pmu_events__test_soc_cpu_uncore_imc_free_running[] = {
-{ 2270 }, /* uncore_imc_free_running.cache_miss\000uncore\000Total cache misses\000event=0x12\000\00000\000Total cache misses\000 */
+{ 2339 }, /* uncore_imc_free_running.cache_miss\000uncore\000Total cache misses\000event=0x12\000\00000\000\000\000\000Total cache misses\000 */
 
 };
 
@@ -129,51 +129,51 @@ const struct pmu_table_entry pmu_events__test_soc_cpu[] = {
 {
      .entries = pmu_events__test_soc_cpu_default_core,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_cpu_default_core),
-     .pmu_name = { 1102 /* default_core\000 */ },
+     .pmu_name = { 1138 /* default_core\000 */ },
 },
 {
      .entries = pmu_events__test_soc_cpu_hisi_sccl_ddrc,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_cpu_hisi_sccl_ddrc),
-     .pmu_name = { 1672 /* hisi_sccl,ddrc\000 */ },
+     .pmu_name = { 1726 /* hisi_sccl,ddrc\000 */ },
 },
 {
      .entries = pmu_events__test_soc_cpu_hisi_sccl_l3c,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_cpu_hisi_sccl_l3c),
-     .pmu_name = { 2152 /* hisi_sccl,l3c\000 */ },
+     .pmu_name = { 2218 /* hisi_sccl,l3c\000 */ },
 },
 {
      .entries = pmu_events__test_soc_cpu_uncore_cbox,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_cpu_uncore_cbox),
-     .pmu_name = { 1773 /* uncore_cbox\000 */ },
+     .pmu_name = { 1830 /* uncore_cbox\000 */ },
 },
 {
      .entries = pmu_events__test_soc_cpu_uncore_imc,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_cpu_uncore_imc),
-     .pmu_name = { 2365 /* uncore_imc\000 */ },
+     .pmu_name = { 2437 /* uncore_imc\000 */ },
 },
 {
      .entries = pmu_events__test_soc_cpu_uncore_imc_free_running,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_cpu_uncore_imc_free_running),
-     .pmu_name = { 2246 /* uncore_imc_free_running\000 */ },
+     .pmu_name = { 2315 /* uncore_imc_free_running\000 */ },
 },
 };
 
 static const struct compact_pmu_event pmu_metrics__test_soc_cpu_default_core[] = {
-{ 2798 }, /* CPI\000\0001 / IPC\000\000\000\000\000\000\000\00000 */
-{ 3479 }, /* DCache_L2_All\000\000DCache_L2_All_Hits + DCache_L2_All_Miss\000\000\000\000\000\000\000\00000 */
-{ 3251 }, /* DCache_L2_All_Hits\000\000l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit\000\000\000\000\000\000\000\00000 */
-{ 3345 }, /* DCache_L2_All_Miss\000\000max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + l2_rqsts.pf_miss + l2_rqsts.rfo_miss\000\000\000\000\000\000\000\00000 */
-{ 3543 }, /* DCache_L2_Hits\000\000d_ratio(DCache_L2_All_Hits, DCache_L2_All)\000\000\000\000\000\000\000\00000 */
-{ 3611 }, /* DCache_L2_Misses\000\000d_ratio(DCache_L2_All_Miss, DCache_L2_All)\000\000\000\000\000\000\000\00000 */
-{ 2883 }, /* Frontend_Bound_SMT\000\000idq_uops_not_delivered.core / (4 * (cpu_clk_unhalted.thread / 2 * (1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk)))\000\000\000\000\000\000\000\00000 */
-{ 2820 }, /* IPC\000group1\000inst_retired.any / cpu_clk_unhalted.thread\000\000\000\000\000\000\000\00000 */
-{ 3745 }, /* L1D_Cache_Fill_BW\000\00064 * l1d.replacement / 1e9 / duration_time\000\000\000\000\000\000\000\00000 */
-{ 3681 }, /* M1\000\000ipc + M2\000\000\000\000\000\000\000\00000 */
-{ 3703 }, /* M2\000\000ipc + M1\000\000\000\000\000\000\000\00000 */
-{ 3725 }, /* M3\000\0001 / M3\000\000\000\000\000\000\000\00000 */
-{ 3180 }, /* cache_miss_cycles\000group1\000dcache_miss_cpi + icache_miss_cycles\000\000\000\000\000\000\000\00000 */
-{ 3049 }, /* dcache_miss_cpi\000\000l1d\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000 */
-{ 3113 }, /* icache_miss_cycles\000\000l1i\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000 */
+{ 2882 }, /* CPI\000\0001 / IPC\000\000\000\000\000\000\000\00000 */
+{ 3563 }, /* DCache_L2_All\000\000DCache_L2_All_Hits + DCache_L2_All_Miss\000\000\000\000\000\000\000\00000 */
+{ 3335 }, /* DCache_L2_All_Hits\000\000l2_rqsts.demand_data_rd_hit + l2_rqsts.pf_hit + l2_rqsts.rfo_hit\000\000\000\000\000\000\000\00000 */
+{ 3429 }, /* DCache_L2_All_Miss\000\000max(l2_rqsts.all_demand_data_rd - l2_rqsts.demand_data_rd_hit, 0) + l2_rqsts.pf_miss + l2_rqsts.rfo_miss\000\000\000\000\000\000\000\00000 */
+{ 3627 }, /* DCache_L2_Hits\000\000d_ratio(DCache_L2_All_Hits, DCache_L2_All)\000\000\000\000\000\000\000\00000 */
+{ 3695 }, /* DCache_L2_Misses\000\000d_ratio(DCache_L2_All_Miss, DCache_L2_All)\000\000\000\000\000\000\000\00000 */
+{ 2967 }, /* Frontend_Bound_SMT\000\000idq_uops_not_delivered.core / (4 * (cpu_clk_unhalted.thread / 2 * (1 + cpu_clk_unhalted.one_thread_active / cpu_clk_unhalted.ref_xclk)))\000\000\000\000\000\000\000\00000 */
+{ 2904 }, /* IPC\000group1\000inst_retired.any / cpu_clk_unhalted.thread\000\000\000\000\000\000\000\00000 */
+{ 3829 }, /* L1D_Cache_Fill_BW\000\00064 * l1d.replacement / 1e9 / duration_time\000\000\000\000\000\000\000\00000 */
+{ 3765 }, /* M1\000\000ipc + M2\000\000\000\000\000\000\000\00000 */
+{ 3787 }, /* M2\000\000ipc + M1\000\000\000\000\000\000\000\00000 */
+{ 3809 }, /* M3\000\0001 / M3\000\000\000\000\000\000\000\00000 */
+{ 3264 }, /* cache_miss_cycles\000group1\000dcache_miss_cpi + icache_miss_cycles\000\000\000\000\000\000\000\00000 */
+{ 3133 }, /* dcache_miss_cpi\000\000l1d\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000 */
+{ 3197 }, /* icache_miss_cycles\000\000l1i\\-loads\\-misses / inst_retired.any\000\000\000\000\000\000\000\00000 */
 
 };
 
@@ -181,18 +181,18 @@ const struct pmu_table_entry pmu_metrics__test_soc_cpu[] = {
 {
      .entries = pmu_metrics__test_soc_cpu_default_core,
      .num_entries = ARRAY_SIZE(pmu_metrics__test_soc_cpu_default_core),
-     .pmu_name = { 1102 /* default_core\000 */ },
+     .pmu_name = { 1138 /* default_core\000 */ },
 },
 };
 
 static const struct compact_pmu_event pmu_events__test_soc_sys_uncore_sys_ccn_pmu[] = {
-{ 2565 }, /* sys_ccn_pmu.read_cycles\000uncore\000ccn read-cycles event\000config=0x2c\0000x01\00000\000\000 */
+{ 2643 }, /* sys_ccn_pmu.read_cycles\000uncore\000ccn read-cycles event\000config=0x2c\0000x01\00000\000\000\000\000\000 */
 };
 static const struct compact_pmu_event pmu_events__test_soc_sys_uncore_sys_cmn_pmu[] = {
-{ 2658 }, /* sys_cmn_pmu.hnf_cache_miss\000uncore\000Counts total cache misses in first lookup result (high priority)\000eventid=1,type=5\000(434|436|43c|43a).*\00000\000\000 */
+{ 2739 }, /* sys_cmn_pmu.hnf_cache_miss\000uncore\000Counts total cache misses in first lookup result (high priority)\000eventid=1,type=5\000(434|436|43c|43a).*\00000\000\000\000\000\000 */
 };
 static const struct compact_pmu_event pmu_events__test_soc_sys_uncore_sys_ddr_pmu[] = {
-{ 2473 }, /* sys_ddr_pmu.write_cycles\000uncore\000ddr write-cycles event\000event=0x2b\000v8\00000\000\000 */
+{ 2548 }, /* sys_ddr_pmu.write_cycles\000uncore\000ddr write-cycles event\000event=0x2b\000v8\00000\000\000\000\000\000 */
 
 };
 
@@ -200,17 +200,17 @@ const struct pmu_table_entry pmu_events__test_soc_sys[] = {
 {
      .entries = pmu_events__test_soc_sys_uncore_sys_ccn_pmu,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_sys_uncore_sys_ccn_pmu),
-     .pmu_name = { 2546 /* uncore_sys_ccn_pmu\000 */ },
+     .pmu_name = { 2624 /* uncore_sys_ccn_pmu\000 */ },
 },
 {
      .entries = pmu_events__test_soc_sys_uncore_sys_cmn_pmu,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_sys_uncore_sys_cmn_pmu),
-     .pmu_name = { 2639 /* uncore_sys_cmn_pmu\000 */ },
+     .pmu_name = { 2720 /* uncore_sys_cmn_pmu\000 */ },
 },
 {
      .entries = pmu_events__test_soc_sys_uncore_sys_ddr_pmu,
      .num_entries = ARRAY_SIZE(pmu_events__test_soc_sys_uncore_sys_ddr_pmu),
-     .pmu_name = { 2454 /* uncore_sys_ddr_pmu\000 */ },
+     .pmu_name = { 2529 /* uncore_sys_ddr_pmu\000 */ },
 },
 };
 
@@ -316,6 +316,12 @@ static void decompress_event(int offset, struct pmu_event *pe)
 	p++;
 	pe->unit = (*p == '\0' ? NULL : p);
 	while (*p++);
+	pe->retirement_latency_mean = (*p == '\0' ? NULL : p);
+	while (*p++);
+	pe->retirement_latency_min = (*p == '\0' ? NULL : p);
+	while (*p++);
+	pe->retirement_latency_max = (*p == '\0' ? NULL : p);
+	while (*p++);
 	pe->long_desc = (*p == '\0' ? NULL : p);
 }
 
tools/perf/pmu-events/jevents.py +6 -0
index 7499a35bfadd..e3a55486c08e 100755
@@ -47,6 +47,9 @@ _json_event_attributes = [
     'event',
     # Short things in alphabetical order.
     'compat', 'deprecated', 'perpkg', 'unit',
+    # Retirement latency specific to Intel granite rapids currently.
+    'retirement_latency_mean', 'retirement_latency_min',
+    'retirement_latency_max',
     # Longer things (the last won't be iterated over during decompress).
     'long_desc'
 ]
@@ -341,6 +344,9 @@ class JsonEvent:
     self.perpkg = jd.get('PerPkg')
     self.aggr_mode = convert_aggr_mode(jd.get('AggregationMode'))
     self.deprecated = jd.get('Deprecated')
+    self.retirement_latency_mean = jd.get('RetirementLatencyMean')
+    self.retirement_latency_min = jd.get('RetirementLatencyMin')
+    self.retirement_latency_max = jd.get('RetirementLatencyMax')
     self.metric_name = jd.get('MetricName')
     self.metric_group = jd.get('MetricGroup')
     self.metricgroup_no_group = jd.get('MetricgroupNoGroup')
tools/perf/pmu-events/pmu-events.h +3 -0
index 675562e6f770..a95fee561622 100644
@@ -47,6 +47,9 @@ struct pmu_event {
 	const char *long_desc;
 	const char *pmu;
 	const char *unit;
+	const char *retirement_latency_mean;
+	const char *retirement_latency_min;
+	const char *retirement_latency_max;
 	bool perpkg;
 	bool deprecated;
 };
tools/perf/util/evsel.h +6 -0
index aae431d63d64..42dcadfef8ce 100644
@@ -177,6 +177,12 @@ struct evsel {
 	/* For tool events */
 	/* Beginning time subtracted when the counter is read. */
 	union {
+		/* Defaults for retirement latency events. */
+		struct _retirement_latency {
+			double mean;
+			double min;
+			double max;
+		} retirement_latency;
 		/* duration_time is a single global time. */
 		__u64 start_time;
 		/*
tools/perf/util/intel-tpebs.c +37 -15

64 lines not on this page · raw message

tools/perf/util/parse-events.c +4 -0
index 5152fd5a6ead..89708d1769c8 100644
@@ -1539,6 +1539,10 @@ static int parse_events_add_pmu(struct parse_events_state *parse_state,
 	evsel->scale = info.scale;
 	evsel->per_pkg = info.per_pkg;
 	evsel->snapshot = info.snapshot;
+	evsel->retirement_latency.mean = info.retirement_latency_mean;
+	evsel->retirement_latency.min = info.retirement_latency_min;
+	evsel->retirement_latency.max = info.retirement_latency_max;
+
 	return 0;
 }
 
tools/perf/util/pmu.c +50 -2

100 lines not on this page · raw message

tools/perf/util/pmu.h +3 -0
index b93014cc3670..13dd3511f504 100644
@@ -194,6 +194,9 @@ struct perf_pmu {
 struct perf_pmu_info {
 	const char *unit;
 	double scale;
+	double retirement_latency_mean;
+	double retirement_latency_min;
+	double retirement_latency_max;
 	bool per_pkg;
 	bool snapshot;
 };
Signature · 1 line
2.49.0.504.g3bcea36a83-goog
author reviewer
#16 lore raw

PATCH v3 16/16 perf record: Retirement latency cleanup in evsel__config

In reply to Ian Rogers

Perf record will fail with retirement latency events as the open doesn't do a perf_event_open system call. Use evsel__config to set up such events for recording by removing the flag and enabling sample weights - the sample weights containing the retirement latency.

---
 tools/perf/util/evsel.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
tools/perf/util/evsel.c +3 -2
index 661a07cbdb25..6a84893e3c58 100644
@@ -1440,9 +1440,10 @@ void evsel__config(struct evsel *evsel, struct record_opts *opts,
 		attr->branch_sample_type = opts->branch_stack;
 	}
 
-	if (opts->sample_weight)
+	if (opts->sample_weight || evsel->retire_lat) {
 		arch_evsel__set_sample_weight(evsel);
-
+		evsel->retire_lat = false;
+	}
 	attr->task     = track;
 	attr->mmap     = track;
 	attr->mmap2    = track && !perf_missing_features.mmap2;
Signature · 1 line
2.49.0.504.g3bcea36a83-goog