Skip to contentSkip to conversation

perf evsel: Fallback to task-clock when not system wide

RFC PATCH In review Ian Rogers 4 participants 6 messages Nov 21, 2023 – Dec 6, 2023 perf events

lore
Share & tools
lore

Ian Rogers

author reviewer
#0 lore raw

RFC PATCH v2 1/1 perf evsel: Fallback to task-clock when not system wide

When the cycles event isn't available evsel will fallback to the cpu-clock software event. task-clock is similar to cpu-clock but only runs when the process is running. Falling back to cpu-clock when not system wide leads to confusion, by falling back to task-clock it is hoped the confusion is less.

Pass the target to determine if task-clock is more appropriate. Update a nearby comment and debug string for the change.

---
v2. Use target__has_cpu as suggested by Namhyung.
https://lpc.events/event/17/contributions/1556/

---
 tools/perf/builtin-record.c |  2 +-
 tools/perf/builtin-stat.c   |  2 +-
 tools/perf/builtin-top.c    |  2 +-
 tools/perf/util/evsel.c     | 18 ++++++++++--------
 tools/perf/util/evsel.h     |  3 ++-
 5 files changed, 15 insertions(+), 12 deletions(-)
tools/perf/builtin-record.c +1 -1
index 8ec818568662..d8bb59511fdd 100644
@@ -1350,7 +1350,7 @@ static int record__open(struct record *rec)
 	evlist__for_each_entry(evlist, pos) {
 try_again:
 		if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
-			if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
+			if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {
 				if (verbose > 0)
 					ui__warning("%s\n", msg);
 				goto try_again;
tools/perf/builtin-stat.c +1 -1
index a3af805a1d57..d8e5d6f7a87a 100644
@@ -653,7 +653,7 @@ static enum counter_recovery stat_handle_error(struct evsel *counter)
 		if ((evsel__leader(counter) != counter) ||
 		    !(counter->core.leader->nr_members > 1))
 			return COUNTER_SKIP;
-	} else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+	} else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
 		if (verbose > 0)
 			ui__warning("%s\n", msg);
 		return COUNTER_RETRY;
tools/perf/builtin-top.c +1 -1
index ea8c7eca5eee..1e42bd1c7d5a 100644
@@ -1044,7 +1044,7 @@ static int perf_top__start_counters(struct perf_top *top)
 			    perf_top_overwrite_fallback(top, counter))
 				goto try_again;
 
-			if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+			if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
 				if (verbose > 0)
 					ui__warning("%s\n", msg);
 				goto try_again;
tools/perf/util/evsel.c +10 -8
index a5da74e3a517..532f34d9fcb5 100644
@@ -2853,7 +2853,8 @@ u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const
 
 #endif
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+		     char *msg, size_t msgsize)
 {
 	int paranoid;
 
@@ -2861,18 +2862,19 @@ bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
 	    evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
 	    evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
 		/*
-		 * If it's cycles then fall back to hrtimer based
-		 * cpu-clock-tick sw counter, which is always available even if
-		 * no PMU support.
+		 * If it's cycles then fall back to hrtimer based cpu-clock sw
+		 * counter, which is always available even if no PMU support.
 		 *
 		 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
 		 * b0a873e).
 		 */
-		scnprintf(msg, msgsize, "%s",
-"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
-
 		evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
-		evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
+		evsel->core.attr.config = target__has_cpu(target)
+			? PERF_COUNT_SW_CPU_CLOCK
+			: PERF_COUNT_SW_TASK_CLOCK;
+		scnprintf(msg, msgsize,
+			"The cycles event is not supported, trying to fall back to %s",
+			target__has_cpu(target) ? "cpu-clock" : "task-clock");
 
 		zfree(&evsel->name);
 		return true;
tools/perf/util/evsel.h +2 -1
index f19ac9f027ef..efbb6e848287 100644
@@ -460,7 +460,8 @@ static inline bool evsel__is_clock(const struct evsel *evsel)
 	       evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK);
 }
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize);
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+		     char *msg, size_t msgsize);
 int evsel__open_strerror(struct evsel *evsel, struct target *target,
 			 int err, char *msg, size_t size);
 
Signature · 1 line
2.43.0.rc1.413.gea7ed67945-goog
author reviewer
#1 lore raw
Quoted 7 lines · Ian Rogers

When the cycles event isn't available evsel will fallback to the cpu-clock software event. task-clock is similar to cpu-clock but only runs when the process is running. Falling back to cpu-clock when not system wide leads to confusion, by falling back to task-clock it is hoped the confusion is less.

Quoted 8 lines

Pass the target to determine if task-clock is more appropriate. Update a nearby comment and debug string for the change.

---
v2. Use target__has_cpu as suggested by Namhyung.
https://lpc.events/event/17/contributions/1556/

Ping.

Thanks,
Ian

Quoted 105 lines
---
 tools/perf/builtin-record.c |  2 +-
 tools/perf/builtin-stat.c   |  2 +-
 tools/perf/builtin-top.c    |  2 +-
 tools/perf/util/evsel.c     | 18 ++++++++++--------
 tools/perf/util/evsel.h     |  3 ++-
 5 files changed, 15 insertions(+), 12 deletions(-)
tools/perf/builtin-record.c +1 -1
index 8ec818568662..d8bb59511fdd 100644
@@ -1350,7 +1350,7 @@ static int record__open(struct record *rec)
        evlist__for_each_entry(evlist, pos) {
 try_again:
                if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
-                       if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
+                       if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {
                                if (verbose > 0)
                                        ui__warning("%s\n", msg);
                                goto try_again;
tools/perf/builtin-stat.c +1 -1
index a3af805a1d57..d8e5d6f7a87a 100644
@@ -653,7 +653,7 @@ static enum counter_recovery stat_handle_error(struct evsel *counter)
                if ((evsel__leader(counter) != counter) ||
                    !(counter->core.leader->nr_members > 1))
                        return COUNTER_SKIP;
-       } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+       } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
                if (verbose > 0)
                        ui__warning("%s\n", msg);
                return COUNTER_RETRY;
tools/perf/builtin-top.c +1 -1
index ea8c7eca5eee..1e42bd1c7d5a 100644
@@ -1044,7 +1044,7 @@ static int perf_top__start_counters(struct perf_top *top)
                            perf_top_overwrite_fallback(top, counter))
                                goto try_again;
 
-                       if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+                       if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
                                if (verbose > 0)
                                        ui__warning("%s\n", msg);
                                goto try_again;
tools/perf/util/evsel.c +10 -8
index a5da74e3a517..532f34d9fcb5 100644
@@ -2853,7 +2853,8 @@ u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const
 
 #endif
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+                    char *msg, size_t msgsize)
 {
        int paranoid;
 
@@ -2861,18 +2862,19 @@ bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
            evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
            evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
                /*
-                * If it's cycles then fall back to hrtimer based
-                * cpu-clock-tick sw counter, which is always available even if
-                * no PMU support.
+                * If it's cycles then fall back to hrtimer based cpu-clock sw
+                * counter, which is always available even if no PMU support.
                 *
                 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
                 * b0a873e).
                 */
-               scnprintf(msg, msgsize, "%s",
-"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
-
                evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
-               evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
+               evsel->core.attr.config = target__has_cpu(target)
+                       ? PERF_COUNT_SW_CPU_CLOCK
+                       : PERF_COUNT_SW_TASK_CLOCK;
+               scnprintf(msg, msgsize,
+                       "The cycles event is not supported, trying to fall back to %s",
+                       target__has_cpu(target) ? "cpu-clock" : "task-clock");
 
                zfree(&evsel->name);
                return true;
tools/perf/util/evsel.h +2 -2
index f19ac9f027ef..efbb6e848287 100644
@@ -460,7 +460,8 @@ static inline bool evsel__is_clock(const struct evsel *evsel)
               evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK);
 }
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize);
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+                    char *msg, size_t msgsize);
 int evsel__open_strerror(struct evsel *evsel, struct target *target,
                         int err, char *msg, size_t size);
 
--

2.43.0.rc1.413.gea7ed67945-goog

maintainer Acked-by
#2 lore raw
Quoted 9 lines · Ian Rogers
Quoted 7 lines · Ian Rogers

When the cycles event isn't available evsel will fallback to the cpu-clock software event. task-clock is similar to cpu-clock but only runs when the process is running. Falling back to cpu-clock when not system wide leads to confusion, by falling back to task-clock it is hoped the confusion is less.

Quoted 8 lines
Quoted 8 lines

Pass the target to determine if task-clock is more appropriate. Update a nearby comment and debug string for the change.

---
v2. Use target__has_cpu as suggested by Namhyung.
https://lpc.events/event/17/contributions/1556/

Thanks,
Namhyung

#3 lore raw
In reply to Ian Rogers
Quoted 32 lines · Ian Rogers

When the cycles event isn't available evsel will fallback to the cpu-clock software event. task-clock is similar to cpu-clock but only runs when the process is running. Falling back to cpu-clock when not system wide leads to confusion, by falling back to task-clock it is hoped the confusion is less.

Pass the target to determine if task-clock is more appropriate. Update a nearby comment and debug string for the change.

---
v2. Use target__has_cpu as suggested by Namhyung.
https://lpc.events/event/17/contributions/1556/

---

tools/perf/builtin-record.c |  2 +-
tools/perf/builtin-stat.c   |  2 +-
tools/perf/builtin-top.c    |  2 +-
tools/perf/util/evsel.c     | 18 ++++++++++--------
tools/perf/util/evsel.h     |  3 ++-
5 files changed, 15 insertions(+), 12 deletions(-)
tools/perf/builtin-record.c +0 -0
index 8ec818568662..d8bb59511fdd 100644
@@ -1350,7 +1350,7 @@ static int record__open(struct record *rec)
 evlist__for_each_entry(evlist, pos) {
Quoted 4 lines

try_again:
if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
- if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
+ if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {

Hi Ian

Tested this with perf record and I could find the code fallback to using task-clock

./perf record -v ls
Warning:
The cycles event is not supported, trying to fall back to task-clock

But in case of “perf stat”, in my environment, found that the code path won't invoke “evsel__fallback”.

Snippet for builtin-stat.c

            if (errno == EINVAL || errno == ENOSYS ||
            errno == ENOENT || errno == EOPNOTSUPP ||
            errno == ENXIO) {
                if (verbose > 0)
                        ui__warning("%s event is not supported by the kernel.\n",
                                    evsel__name(counter));
                counter->supported = false;
                /*
                 * errored is a sticky flag that means one of the counter's
                 * cpu event had a problem and needs to be reexamined.
                 */
                counter->errored = true;

                if ((evsel__leader(counter) != counter) ||
                    !(counter->core.leader->nr_members > 1))
                        return COUNTER_SKIP;
        } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
                if (verbose > 0)
                        ui__warning("%s\n", msg);
                return COUNTER_RETRY;

So if the perf_event_open returns ENOENT, we won’t do a fallback in builtin-stat.c Should we address cycles differently here ? Any comments ?

Thanks
Athira

Quoted 88 lines

if (verbose > 0)
ui__warning("%s\n", msg);
goto try_again;

tools/perf/builtin-stat.c +0 -0
index a3af805a1d57..d8e5d6f7a87a 100644
@@ -653,7 +653,7 @@ static enum counter_recovery stat_handle_error(struct evsel *counter)

if ((evsel__leader(counter) != counter) ||
!(counter->core.leader->nr_members > 1))
return COUNTER_SKIP;
- } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+ } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) { if (verbose > 0)
ui__warning("%s\n", msg);
return COUNTER_RETRY;

tools/perf/builtin-top.c +1 -1
index ea8c7eca5eee..1e42bd1c7d5a 100644
@@ -1044,7 +1044,7 @@ static int perf_top__start_counters(struct perf_top *top)
   perf_top_overwrite_fallback(top, counter))
 goto try_again;
 
- if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+ if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
 if (verbose > 0)
 ui__warning("%s\n", msg);
 goto try_again;
tools/perf/util/evsel.c +10 -8
index a5da74e3a517..532f34d9fcb5 100644
@@ -2853,7 +2853,8 @@ u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const
 
 #endif
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+     char *msg, size_t msgsize)
 {
 int paranoid;
 
@@ -2861,18 +2862,19 @@ bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
   evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
   evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
 /*
- * If it's cycles then fall back to hrtimer based
- * cpu-clock-tick sw counter, which is always available even if
- * no PMU support.
+ * If it's cycles then fall back to hrtimer based cpu-clock sw
+ * counter, which is always available even if no PMU support.
 *
 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
 * b0a873e).
 */
- scnprintf(msg, msgsize, "%s",
-"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
-
 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
- evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
+ evsel->core.attr.config = target__has_cpu(target)
+ ? PERF_COUNT_SW_CPU_CLOCK
+ : PERF_COUNT_SW_TASK_CLOCK;
+ scnprintf(msg, msgsize,
+ "The cycles event is not supported, trying to fall back to %s",
+ target__has_cpu(target) ? "cpu-clock" : "task-clock");
 
 zfree(&evsel->name);
 return true;
tools/perf/util/evsel.h +2 -1
index f19ac9f027ef..efbb6e848287 100644
@@ -460,7 +460,8 @@ static inline bool evsel__is_clock(const struct evsel *evsel)
      evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK);
 }
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize);
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+     char *msg, size_t msgsize);
 int evsel__open_strerror(struct evsel *evsel, struct target *target,
 int err, char *msg, size_t size);
 
Signature · 1 line
2.43.0.rc1.413.gea7ed67945-goog
#4 lore raw

Em Tue, Dec 05, 2023 at 04:51:01PM +0530, Athira Rajeev escreveu:

Quoted 39 lines
Quoted 32 lines · Ian Rogers

When the cycles event isn't available evsel will fallback to the cpu-clock software event. task-clock is similar to cpu-clock but only runs when the process is running. Falling back to cpu-clock when not system wide leads to confusion, by falling back to task-clock it is hoped the confusion is less.

Pass the target to determine if task-clock is more appropriate. Update a nearby comment and debug string for the change.

---
v2. Use target__has_cpu as suggested by Namhyung.
https://lpc.events/event/17/contributions/1556/

---

tools/perf/builtin-record.c |  2 +-
tools/perf/builtin-stat.c   |  2 +-
tools/perf/builtin-top.c    |  2 +-
tools/perf/util/evsel.c     | 18 ++++++++++--------
tools/perf/util/evsel.h     |  3 ++-
5 files changed, 15 insertions(+), 12 deletions(-)
tools/perf/builtin-record.c +0 -0
index 8ec818568662..d8bb59511fdd 100644
@@ -1350,7 +1350,7 @@ static int record__open(struct record *rec)
 evlist__for_each_entry(evlist, pos) {
Quoted 4 lines

try_again:
if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
- if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
+ if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {

Hi Ian

Quoted 5 lines

Tested this with perf record and I could find the code fallback to using task-clock

./perf record -v ls
Warning:
The cycles event is not supported, trying to fall back to task-clock

Ok, so I'll take that as a Tested-by: you, ok?

The "perf stat" part can be addressed in a follow up patch, when that error handling is researched to remember why we have that ->supported, ->errored thing.

- Arnaldo

Quoted 118 lines

But in case of “perf stat”, in my environment, found that the code path won't invoke “evsel__fallback”.

Snippet for builtin-stat.c

            if (errno == EINVAL || errno == ENOSYS ||
            errno == ENOENT || errno == EOPNOTSUPP ||
            errno == ENXIO) {
                if (verbose > 0)
                        ui__warning("%s event is not supported by the kernel.\n",
                                    evsel__name(counter));
                counter->supported = false;
                /*
                 * errored is a sticky flag that means one of the counter's
                 * cpu event had a problem and needs to be reexamined.
                 */
                counter->errored = true;

                if ((evsel__leader(counter) != counter) ||
                    !(counter->core.leader->nr_members > 1))
                        return COUNTER_SKIP;
        } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
                if (verbose > 0)
                        ui__warning("%s\n", msg);
                return COUNTER_RETRY;

So if the perf_event_open returns ENOENT, we won’t do a fallback in builtin-stat.c Should we address cycles differently here ? Any comments ?

Thanks
Athira

Quoted 88 lines

if (verbose > 0)
ui__warning("%s\n", msg);
goto try_again;

tools/perf/builtin-stat.c +0 -0
index a3af805a1d57..d8e5d6f7a87a 100644
@@ -653,7 +653,7 @@ static enum counter_recovery stat_handle_error(struct evsel *counter)

if ((evsel__leader(counter) != counter) ||
!(counter->core.leader->nr_members > 1))
return COUNTER_SKIP;
- } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+ } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) { if (verbose > 0)
ui__warning("%s\n", msg);
return COUNTER_RETRY;

tools/perf/builtin-top.c +1 -1
index ea8c7eca5eee..1e42bd1c7d5a 100644
@@ -1044,7 +1044,7 @@ static int perf_top__start_counters(struct perf_top *top)
   perf_top_overwrite_fallback(top, counter))
 goto try_again;
 
- if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+ if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
 if (verbose > 0)
 ui__warning("%s\n", msg);
 goto try_again;
tools/perf/util/evsel.c +10 -8
index a5da74e3a517..532f34d9fcb5 100644
@@ -2853,7 +2853,8 @@ u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const
 
 #endif
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+     char *msg, size_t msgsize)
 {
 int paranoid;
 
@@ -2861,18 +2862,19 @@ bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
   evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
   evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
 /*
- * If it's cycles then fall back to hrtimer based
- * cpu-clock-tick sw counter, which is always available even if
- * no PMU support.
+ * If it's cycles then fall back to hrtimer based cpu-clock sw
+ * counter, which is always available even if no PMU support.
 *
 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
 * b0a873e).
 */
- scnprintf(msg, msgsize, "%s",
-"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
-
 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
- evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
+ evsel->core.attr.config = target__has_cpu(target)
+ ? PERF_COUNT_SW_CPU_CLOCK
+ : PERF_COUNT_SW_TASK_CLOCK;
+ scnprintf(msg, msgsize,
+ "The cycles event is not supported, trying to fall back to %s",
+ target__has_cpu(target) ? "cpu-clock" : "task-clock");
 
 zfree(&evsel->name);
 return true;
tools/perf/util/evsel.h +2 -1
index f19ac9f027ef..efbb6e848287 100644
@@ -460,7 +460,8 @@ static inline bool evsel__is_clock(const struct evsel *evsel)
      evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK);
 }
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize);
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+     char *msg, size_t msgsize);
 int evsel__open_strerror(struct evsel *evsel, struct target *target,
 int err, char *msg, size_t size);
 
Signature · 1 line
2.43.0.rc1.413.gea7ed67945-goog
Signature · 1 line
- Arnaldo
Tested-by
#5 lore raw
Quoted 44 lines · Arnaldo Carvalho de Melo

Em Tue, Dec 05, 2023 at 04:51:01PM +0530, Athira Rajeev escreveu:

Quoted 39 lines
Quoted 32 lines · Ian Rogers

When the cycles event isn't available evsel will fallback to the cpu-clock software event. task-clock is similar to cpu-clock but only runs when the process is running. Falling back to cpu-clock when not system wide leads to confusion, by falling back to task-clock it is hoped the confusion is less.

Pass the target to determine if task-clock is more appropriate. Update a nearby comment and debug string for the change.

---
v2. Use target__has_cpu as suggested by Namhyung.
https://lpc.events/event/17/contributions/1556/

---

tools/perf/builtin-record.c |  2 +-
tools/perf/builtin-stat.c   |  2 +-
tools/perf/builtin-top.c    |  2 +-
tools/perf/util/evsel.c     | 18 ++++++++++--------
tools/perf/util/evsel.h     |  3 ++-
5 files changed, 15 insertions(+), 12 deletions(-)
tools/perf/builtin-record.c +0 -0
index 8ec818568662..d8bb59511fdd 100644
@@ -1350,7 +1350,7 @@ static int record__open(struct record *rec)
 evlist__for_each_entry(evlist, pos) {
Quoted 4 lines

try_again:
if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
- if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
+ if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {

Hi Ian

Quoted 7 lines
Quoted 5 lines

Tested this with perf record and I could find the code fallback to using task-clock

./perf record -v ls
Warning:
The cycles event is not supported, trying to fall back to task-clock

Ok, so I'll take that as a Tested-by: you, ok?

Hi Arnaldo,

Please add my Tested-by.

Quoted 4 lines

The "perf stat" part can be addressed in a follow up patch, when that error handling is researched to remember why we have that ->supported, ->errored thing.

Sure, I will take a look at why we have that difference in “perf stat”

Thanks
Athira

Quoted 125 lines

- Arnaldo

Quoted 118 lines

But in case of “perf stat”, in my environment, found that the code path won't invoke “evsel__fallback”.

Snippet for builtin-stat.c

           if (errno == EINVAL || errno == ENOSYS ||
           errno == ENOENT || errno == EOPNOTSUPP ||
           errno == ENXIO) {
               if (verbose > 0)
                       ui__warning("%s event is not supported by the kernel.\n",
                                   evsel__name(counter));
               counter->supported = false;
               /*
                * errored is a sticky flag that means one of the counter's
                * cpu event had a problem and needs to be reexamined.
                */
               counter->errored = true;

               if ((evsel__leader(counter) != counter) ||
                   !(counter->core.leader->nr_members > 1))
                       return COUNTER_SKIP;
       } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
               if (verbose > 0)
                       ui__warning("%s\n", msg);
               return COUNTER_RETRY;

So if the perf_event_open returns ENOENT, we won’t do a fallback in builtin-stat.c Should we address cycles differently here ? Any comments ?

Thanks
Athira

Quoted 88 lines

if (verbose > 0)
ui__warning("%s\n", msg);
goto try_again;

tools/perf/builtin-stat.c +0 -0
index a3af805a1d57..d8e5d6f7a87a 100644
@@ -653,7 +653,7 @@ static enum counter_recovery stat_handle_error(struct evsel *counter)

if ((evsel__leader(counter) != counter) ||
!(counter->core.leader->nr_members > 1))
return COUNTER_SKIP;
- } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+ } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) { if (verbose > 0)
ui__warning("%s\n", msg);
return COUNTER_RETRY;

tools/perf/builtin-top.c +1 -1
index ea8c7eca5eee..1e42bd1c7d5a 100644
@@ -1044,7 +1044,7 @@ static int perf_top__start_counters(struct perf_top *top)
  perf_top_overwrite_fallback(top, counter))
 goto try_again;
 
- if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
+ if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
 if (verbose > 0)
 ui__warning("%s\n", msg);
 goto try_again;
tools/perf/util/evsel.c +10 -8
index a5da74e3a517..532f34d9fcb5 100644
@@ -2853,7 +2853,8 @@ u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const
 
 #endif
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+     char *msg, size_t msgsize)
 {
 int paranoid;
 
@@ -2861,18 +2862,19 @@ bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
  evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
  evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
 /*
- * If it's cycles then fall back to hrtimer based
- * cpu-clock-tick sw counter, which is always available even if
- * no PMU support.
+ * If it's cycles then fall back to hrtimer based cpu-clock sw
+ * counter, which is always available even if no PMU support.
 *
 * PPC returns ENXIO until 2.6.37 (behavior changed with commit
 * b0a873e).
 */
- scnprintf(msg, msgsize, "%s",
-"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
-
 evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
- evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
+ evsel->core.attr.config = target__has_cpu(target)
+ ? PERF_COUNT_SW_CPU_CLOCK
+ : PERF_COUNT_SW_TASK_CLOCK;
+ scnprintf(msg, msgsize,
+ "The cycles event is not supported, trying to fall back to %s",
+ target__has_cpu(target) ? "cpu-clock" : "task-clock");
 
 zfree(&evsel->name);
 return true;
tools/perf/util/evsel.h +2 -1
index f19ac9f027ef..efbb6e848287 100644
@@ -460,7 +460,8 @@ static inline bool evsel__is_clock(const struct evsel *evsel)
     evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK);
 }
 
-bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize);
+bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
+     char *msg, size_t msgsize);
 int evsel__open_strerror(struct evsel *evsel, struct target *target,
 int err, char *msg, size_t size);
 
Signature · 1 line
2.43.0.rc1.413.gea7ed67945-goog
Signature · 1 line
- Arnaldo