Skip to contentSkip to conversation

perf x86/topdown: Fix topdown leader sampling test error on hybrid

PATCH In review Dapeng Mi 3 participants 7 messages Feb 24, 2025 – Mar 5, 2025 perf events

lore
Share & tools
lore

Dapeng Mi

author
#0 lore raw

PATCH 1/2 perf x86/topdown: Fix topdown leader sampling test error on hybrid

When running topdown leader smapling test on Intel hybrid platforms, such as LNL/ARL, we see the below error.

Topdown leader sampling test
Topdown leader sampling [Failed topdown events not reordered correctly]

It indciates the below command fails.

perf record -o "${perfdata}" -e "{instructions,slots,topdown-retiring}:S" true

The root cause is that perf tool creats a perf event for each PMU type if it can create.

As for this command, there would be 5 perf events created, cpu_atom/instructions/,cpu_atom/topdown_retiring/, cpu_core/slots/,cpu_core/instructions/,cpu_core/topdown-retiring/

For these 5 events, the 2 cpu_atom events are in a group and the other 3 cpu_core events are in another group.

When arch_topdown_sample_read() traverses all these 5 events, events cpu_atom/instructions/ and cpu_core/slots/ don't have a same group leade, and then return false directly and lead to cpu_core/slots/ event is used to sample and this is not allowed by PMU driver.

It's a overkill to return false directly if "evsel->core.leader !=
leader->core.leader" since there could be multiple groups in the event
list.

Just "continue" instead of "return false" to fix this issue.

Fixes 1e53e9d1787b ("perf x86/topdown: Correct leader selection with sample_read enabled")
---
 tools/perf/arch/x86/util/topdown.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/arch/x86/util/topdown.c +1 -1
index f63747d0abdf..d1c654839049 100644
@@ -81,7 +81,7 @@ bool arch_topdown_sample_read(struct evsel *leader)
 	 */
 	evlist__for_each_entry(leader->evlist, evsel) {
 		if (evsel->core.leader != leader->core.leader)
-			return false;
+			continue;
 		if (evsel != leader && arch_is_topdown_metrics(evsel))
 			return true;
 	}

base-commit: 4bac7fb5862740087825eda3ed6168e91da8b7e6

Signature · 1 line
2.40.1
author
#1 lore raw

PATCH 2/2 perf tools/tests: Fix topdown groups test on hybrid platforms

When running topdown groups test on hybrid platforms like LNL/ARL, we see the following 2 commands fail.

perf stat $cputype -e '{instructions,slots},topdown-retiring' true perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true

Take the 1st command as an example, 5 events are created on hybrid platform. They are cpu_atom/instructions/, cpu_core/instructions/, cpu_core/slots/, cpu_atom/topdown-retiring/ and cpu_core/topdown-retiring/ events. The former 3 events are in a group and the latter 2 topdown-retiring events are independent events.

As the limitation of current implementation, the cpu_core/topdown-retiring/ event can't be moved into previous group as long as there are other events before it. That's the reason why we see the failure.

Thus add "--cputype core" option to limit only P-core events are tested.

---
 tools/perf/arch/x86/util/evlist.c | 26 +++++++++++++++++++++++---
 tools/perf/tests/shell/stat.sh    | 20 ++++++++++++++++++--
 2 files changed, 41 insertions(+), 5 deletions(-)
tools/perf/arch/x86/util/evlist.c +23 -3
index 447a734e591c..0a71ba975871 100644
@@ -9,7 +9,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
 {
 	/*
 	 * Currently the following topdown events sequence are supported to
-	 * move and regroup correctly.
+	 * move and regroup correctly on non-hybrid platforms.
 	 *
 	 * a. all events in a group
 	 *    perf stat -e "{instructions,topdown-retiring,slots}" -C0 sleep 1
@@ -44,7 +44,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
 	 * topdown metrics events must be first event after the slots event group,
 	 * otherwise topdown metrics events can't be regrouped correctly, e.g.
 	 *
-	 * a. perf stat -e "{instructions,slots},cycles,topdown-retiring" -C0 sleep 1
+	 * e. perf stat -e "{instructions,slots},cycles,topdown-retiring" -C0 sleep 1
 	 *    WARNING: events were regrouped to match PMUs
 	 *     Performance counter stats for 'CPU(s) 0':
 	 *         17,923,134      slots
@@ -56,11 +56,31 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
 	 * has topdown metrics events must contain only the topdown metrics event,
 	 * otherwise topdown metrics event can't be regrouped correctly as well, e.g.
 	 *
-	 * a. perf stat -e "{instructions,slots},{topdown-retiring,cycles}" -C0 sleep 1
+	 * f. perf stat -e "{instructions,slots},{topdown-retiring,cycles}" -C0 sleep 1
 	 *    WARNING: events were regrouped to match PMUs
 	 *    Error:
 	 *    The sys_perf_event_open() syscall returned with 22 (Invalid argument) for
 	 *    event (topdown-retiring)
+	 *
+	 * For hybrid platforms, the sequences 'c' and 'd' are not supported as well
+	 * besides above sequences 'e' and 'f'.
+	 *
+	 *    perf stat -e "{instructions,slots},topdown-retiring" -C0 sleep 1
+	 *    perf stat -e "{instructions,slots},{topdown-retiring}" -C0 sleep 1
+	 *
+	 * On hybrid platforms each event would create an instance on all types of PMU
+	 * if the event can be supported by the PMU, i.e., the "topdown-retiring" event
+	 * would create two instances on hybrid platforms with P-cores and E-cores,
+	 * "cpu_core/topdown-retiring/" and "cpu_atom/topdown_retiring".
+	 *
+	 * Take the first command as an example, the events list would be converted to
+	 * below list in fact.
+	 *
+	 * "{cpu_atom/instructions/,cpu_core/instructions/,cpu_core/slots/},
+	 *  cpu_atom/topdown-retiring/,cpu_core/topdown-retiring/"
+	 *
+	 * This is actually same with event list in case 'e', "cpu_core/topdown-retiring/"
+	 * event can't be moved into previous events group.
 	 */
 	if (topdown_sys_has_perf_metrics() &&
 	    (arch_evsel__must_be_in_group(lhs) || arch_evsel__must_be_in_group(rhs))) {
tools/perf/tests/shell/stat.sh +18 -2
index 68323d636fb7..cdfe27c25528 100755
@@ -5,6 +5,16 @@
 set -e
 
 err=0
+is_hybrid=false
+
+check_hybrid_platform() {
+  pmus=$(ls /sys/bus/event_source/devices/*/cpus 2>/dev/null | wc -l)
+  if [ "$pmus" -gt 1 ]
+  then
+    is_hybrid=true
+  fi
+}
+
 test_default_stat() {
   echo "Basic stat command test"
   if ! perf stat true 2>&1 | grep -E -q "Performance counter stats for 'true':"
@@ -62,6 +72,11 @@ test_topdown_groups() {
   # Topdown events must be grouped with the slots event first. Test that
   # parse-events reorders this.
   echo "Topdown event group test"
+  cputype=""
+  if $is_hybrid
+  then
+    cputype="--cputype core"
+  fi
   if ! perf stat -e '{slots,topdown-retiring}' true > /dev/null 2>&1
   then
     echo "Topdown event group test [Skipped event parsing failed]"
@@ -85,13 +100,13 @@ test_topdown_groups() {
     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"
   then
     echo "Topdown event group test [Failed topdown metrics event not move into slots group]"
     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"
   then
     echo "Topdown event group test [Failed topdown metrics group not merge into slots group]"
     err=1
@@ -200,6 +215,7 @@ test_hybrid() {
   echo "hybrid test [Success]"
 }
 
+check_hybrid_platform
 test_default_stat
 test_stat_record_report
 test_stat_record_script
Signature · 1 line
2.40.1
Tested-by on v1 2/2
#2 lore raw

Re: [PATCH 2/2] perf tools/tests: Fix topdown groups test on hybrid platforms

Quoted 19 lines · Dapeng Mi

When running topdown groups test on hybrid platforms like LNL/ARL, we see the following 2 commands fail.

perf stat $cputype -e '{instructions,slots},topdown-retiring' true perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true

Take the 1st command as an example, 5 events are created on hybrid platform. They are cpu_atom/instructions/, cpu_core/instructions/, cpu_core/slots/, cpu_atom/topdown-retiring/ and cpu_core/topdown-retiring/ events. The former 3 events are in a group and the latter 2 topdown-retiring events are independent events.

As the limitation of current implementation, the cpu_core/topdown-retiring/ event can't be moved into previous group as
long as there are other events before it. That's the reason why we see
the failure.

Quoted 4 lines

Thus add "--cputype core" option to limit only P-core events are tested.

Quoted 153 lines

---
 tools/perf/arch/x86/util/evlist.c | 26 +++++++++++++++++++++++--- tools/perf/tests/shell/stat.sh    | 20 ++++++++++++++++++--

 2 files changed, 41 insertions(+), 5 deletions(-)
tools/perf/arch/x86/util/evlist.c +0 -0
index 447a734e591c..0a71ba975871 100644
@@ -9,7 +9,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const
 struct evsel *rhs)
  {
  	/*
  	 * Currently the following topdown events sequence are

supported to

tools/perf/arch/x86/util/evlist.c +1 -1
-	 * move and regroup correctly.
+	 * move and regroup correctly on non-hybrid platforms.

  *
  * a. all events in a group
  *    perf stat -e "{instructions,topdown-retiring,slots}" - C0 sleep 1

tools/perf/arch/x86/util/evlist.c +0 -0
@@ -44,7 +44,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const
 struct evsel *rhs)
  	 * topdown metrics events must be first event after the
 slots event group,
  	 * otherwise topdown metrics events can't be regrouped

correctly, e.g.
  *

tools/perf/arch/x86/util/evlist.c +0 -1
-	 * a. perf stat -e "{instructions,slots},cycles,topdown-

retiring" -C0 sleep 1

tools/perf/arch/x86/util/evlist.c +1 -0
+	 * e. perf stat -e "{instructions,slots},cycles,topdown-

retiring" -C0 sleep 1
  *    WARNING: events were regrouped to match PMUs *     Performance counter stats for 'CPU(s) 0':
  *         17,923,134      slots

tools/perf/arch/x86/util/evlist.c +0 -0
@@ -56,11 +56,31 @@ int arch_evlist__cmp(const struct evsel *lhs,
 const struct evsel *rhs)
  	 * has topdown metrics events must contain only the topdown
 metrics event,
  	 * otherwise topdown metrics event can't be regrouped

correctly as well, e.g.
  *

tools/perf/arch/x86/util/evlist.c +0 -1
-	 * a. perf stat -e "{instructions,slots},{topdown-

retiring,cycles}" -C0 sleep 1

tools/perf/arch/x86/util/evlist.c +1 -0
+	 * f. perf stat -e "{instructions,slots},{topdown-

retiring,cycles}" -C0 sleep 1
  *    WARNING: events were regrouped to match PMUs *    Error:
  *    The sys_perf_event_open() syscall returned with 22 (Invalid argument) for
  *    event (topdown-retiring)

tools/perf/arch/x86/util/evlist.c +2 -0
+	 *
+	 * For hybrid platforms, the sequences 'c' and 'd' are not

supported as well

tools/perf/arch/x86/util/evlist.c +3 -0
+	 * besides above sequences 'e' and 'f'.
+	 *
+	 *    perf stat -e "{instructions,slots},topdown-retiring" -

C0 sleep 1

tools/perf/arch/x86/util/evlist.c +3 -1
+	 *    perf stat -e "{instructions,slots},{topdown-retiring}"
-C0 sleep 1
+	 *
+	 * On hybrid platforms each event would create an instance

on all types of PMU

tools/perf/arch/x86/util/evlist.c +1 -0
+	 * if the event can be supported by the PMU, i.e., the

"topdown-retiring" event

tools/perf/arch/x86/util/evlist.c +1 -0
+	 * would create two instances on hybrid platforms with P-

cores and E-cores,

tools/perf/arch/x86/util/evlist.c +1 -0
+	 * "cpu_core/topdown-retiring/" and

"cpu_atom/topdown_retiring".

tools/perf/arch/x86/util/evlist.c +2 -0
+	 *
+	 * Take the first command as an example, the events list

would be converted to

tools/perf/arch/x86/util/evlist.c +3 -0
+	 * below list in fact.
+	 *
+	 *

"{cpu_atom/instructions/,cpu_core/instructions/,cpu_core/slots/},

tools/perf/arch/x86/util/evlist.c +3 -0
+	 *  cpu_atom/topdown-retiring/,cpu_core/topdown-retiring/"
+	 *
+	 * This is actually same with event list in case 'e',

"cpu_core/topdown-retiring/"

tools/perf/arch/x86/util/evlist.c +1 -0
+	 * event can't be moved into previous events group.

  */
  if (topdown_sys_has_perf_metrics() &&
      (arch_evsel__must_be_in_group(lhs) ||
arch_evsel__must_be_in_group(rhs))) {

tools/perf/tests/shell/stat.sh +0 -0
index 68323d636fb7..cdfe27c25528 100755
@@ -5,6 +5,16 @@

 set -e

 err=0

tools/perf/tests/shell/stat.sh +4 -0
+is_hybrid=false
+
+check_hybrid_platform() {
+  pmus=$(ls /sys/bus/event_source/devices/*/cpus 2>/dev/null | wc -

l)
+  if [ "$pmus" -gt 1 ]
+  then
+    is_hybrid=true
+  fi

tools/perf/tests/shell/stat.sh +2 -0
+}
+

 test_default_stat() {
   echo "Basic stat command test"
   if ! perf stat true 2>&1 | grep -E -q "Performance counter stats for 'true':"

tools/perf/tests/shell/stat.sh +0 -0
@@ -62,6 +72,11 @@ test_topdown_groups() {
    # Topdown events must be grouped with the slots event first. Test

that

   # parse-events reorders this.
   echo "Topdown event group test"
+  cputype=""
+  if $is_hybrid
+  then
+    cputype="--cputype core"
+  fi
   if ! perf stat -e '{slots,topdown-retiring}' true > /dev/null 2>&1
   then
     echo "Topdown event group test [Skipped event parsing failed]"
tools/perf/tests/shell/stat.sh +0 -0
@@ -85,13 +100,13 @@ test_topdown_groups() {

     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"
   then
     echo "Topdown event group test [Failed topdown metrics event not move into slots group]"
     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"
   then
     echo "Topdown event group test [Failed topdown metrics group not merge into slots group]"
     err=1

tools/perf/tests/shell/stat.sh +0 -0
@@ -200,6 +215,7 @@ test_hybrid() {

   echo "hybrid test [Success]"
 }

+check_hybrid_platform
 test_default_stat
 test_stat_record_report
 test_stat_record_script

reviewer
#3 lore raw

Re: [PATCH 2/2] perf tools/tests: Fix topdown groups test on hybrid platforms

In reply to Dapeng Mi
Quoted 115 lines · Dapeng Mi

When running topdown groups test on hybrid platforms like LNL/ARL, we see the following 2 commands fail.

perf stat $cputype -e '{instructions,slots},topdown-retiring' true perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true

Take the 1st command as an example, 5 events are created on hybrid platform. They are cpu_atom/instructions/, cpu_core/instructions/, cpu_core/slots/, cpu_atom/topdown-retiring/ and cpu_core/topdown-retiring/ events. The former 3 events are in a group and the latter 2 topdown-retiring events are independent events.

As the limitation of current implementation, the cpu_core/topdown-retiring/ event can't be moved into previous group as long as there are other events before it. That's the reason why we see the failure.

Thus add "--cputype core" option to limit only P-core events are tested.

---
 tools/perf/arch/x86/util/evlist.c | 26 +++++++++++++++++++++++---
 tools/perf/tests/shell/stat.sh    | 20 ++++++++++++++++++--
 2 files changed, 41 insertions(+), 5 deletions(-)
tools/perf/arch/x86/util/evlist.c +23 -3
index 447a734e591c..0a71ba975871 100644
@@ -9,7 +9,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
 {
        /*
         * Currently the following topdown events sequence are supported to
-        * move and regroup correctly.
+        * move and regroup correctly on non-hybrid platforms.
         *
         * a. all events in a group
         *    perf stat -e "{instructions,topdown-retiring,slots}" -C0 sleep 1
@@ -44,7 +44,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
         * topdown metrics events must be first event after the slots event group,
         * otherwise topdown metrics events can't be regrouped correctly, e.g.
         *
-        * a. perf stat -e "{instructions,slots},cycles,topdown-retiring" -C0 sleep 1
+        * e. perf stat -e "{instructions,slots},cycles,topdown-retiring" -C0 sleep 1
         *    WARNING: events were regrouped to match PMUs
         *     Performance counter stats for 'CPU(s) 0':
         *         17,923,134      slots
@@ -56,11 +56,31 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
         * has topdown metrics events must contain only the topdown metrics event,
         * otherwise topdown metrics event can't be regrouped correctly as well, e.g.
         *
-        * a. perf stat -e "{instructions,slots},{topdown-retiring,cycles}" -C0 sleep 1
+        * f. perf stat -e "{instructions,slots},{topdown-retiring,cycles}" -C0 sleep 1
         *    WARNING: events were regrouped to match PMUs
         *    Error:
         *    The sys_perf_event_open() syscall returned with 22 (Invalid argument) for
         *    event (topdown-retiring)
+        *
+        * For hybrid platforms, the sequences 'c' and 'd' are not supported as well
+        * besides above sequences 'e' and 'f'.
+        *
+        *    perf stat -e "{instructions,slots},topdown-retiring" -C0 sleep 1
+        *    perf stat -e "{instructions,slots},{topdown-retiring}" -C0 sleep 1
+        *
+        * On hybrid platforms each event would create an instance on all types of PMU
+        * if the event can be supported by the PMU, i.e., the "topdown-retiring" event
+        * would create two instances on hybrid platforms with P-cores and E-cores,
+        * "cpu_core/topdown-retiring/" and "cpu_atom/topdown_retiring".
+        *
+        * Take the first command as an example, the events list would be converted to
+        * below list in fact.
+        *
+        * "{cpu_atom/instructions/,cpu_core/instructions/,cpu_core/slots/},
+        *  cpu_atom/topdown-retiring/,cpu_core/topdown-retiring/"
+        *
+        * This is actually same with event list in case 'e', "cpu_core/topdown-retiring/"
+        * event can't be moved into previous events group.
         */
        if (topdown_sys_has_perf_metrics() &&
            (arch_evsel__must_be_in_group(lhs) || arch_evsel__must_be_in_group(rhs))) {
tools/perf/tests/shell/stat.sh +15 -0
index 68323d636fb7..cdfe27c25528 100755
@@ -5,6 +5,16 @@
 set -e
 
 err=0
+is_hybrid=false
+
+check_hybrid_platform() {
+  pmus=$(ls /sys/bus/event_source/devices/*/cpus 2>/dev/null | wc -l)
+  if [ "$pmus" -gt 1 ]
+  then
+    is_hybrid=true
+  fi
+}
+
 test_default_stat() {
   echo "Basic stat command test"
   if ! perf stat true 2>&1 | grep -E -q "Performance counter stats for 'true':"
@@ -62,6 +72,11 @@ test_topdown_groups() {
   # Topdown events must be grouped with the slots event first. Test that
   # parse-events reorders this.
   echo "Topdown event group test"
+  cputype=""
+  if $is_hybrid
+  then
+    cputype="--cputype core"
+  fi
   if ! perf stat -e '{slots,topdown-retiring}' true > /dev/null 2>&1
   then
     echo "Topdown event group test [Skipped event parsing failed]"

tools/perf/tests/shell/stat.sh test_topdown_groups() { · line 103

Quoted 5 lines
tools/perf/tests/shell/stat.sh +1 -1
@@ -85,13 +100,13 @@ test_topdown_groups() {
     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"

So I can make this test fail on non-hybrid with:

```
$ perf stat -e '{instructions,slots},faults,topdown-retiring' true
WARNING: events were regrouped to match PMUs

Performance counter stats for 'true':

        5,312,770      slots
        1,078,401      instructions
               56      faults
  <not supported>      topdown-retiring
```
The issue is that the slots isn't a group leader, so the "force group
index" we try to insert must be grouped events into is miscalculated
to be topdown-retiring. The following fixes this:
```
diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index 35e48fe56dfa..68ddc335cde4 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -2056,8 +2056,7 @@ static int
parse_events__sort_events_and_fix_groups(struct list_head *list)
                pos->core.idx = idx++;

                /* Remember an index to sort all forced grouped events
together to. */
-               if (force_grouped_idx == -1 && pos == pos_leader &&
pos->core.nr_members < 2 &&
-                   arch_evsel__must_be_in_group(pos))
+               if (force_grouped_idx == -1 &&
arch_evsel__must_be_in_group(pos))
                        force_grouped_idx = pos->core.idx;
        }
```

tools/perf/tests/shell/stat.sh test_topdown_groups() { · line 109

Quoted 7 lines
tools/perf/tests/shell/stat.sh +1 -1
   then
     echo "Topdown event group test [Failed topdown metrics event not move into slots group]"
     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"

So this is trickier as '{topdown-retiring}' ends up looking as a 2 event group. I'll post a proposed fix.

Thanks,
Ian

Quoted 14 lines
tools/perf/tests/shell/stat.sh +1 -1
   then
     echo "Topdown event group test [Failed topdown metrics group not merge into slots group]"
     err=1
@@ -200,6 +215,7 @@ test_hybrid() {
   echo "hybrid test [Success]"
 }
 
+check_hybrid_platform
 test_default_stat
 test_stat_record_report
 test_stat_record_script
--

2.40.1

reviewer
#4 lore raw

Re: [PATCH 2/2] perf tools/tests: Fix topdown groups test on hybrid platforms

Quoted 166 lines · Ian Rogers
Quoted 115 lines · Dapeng Mi

When running topdown groups test on hybrid platforms like LNL/ARL, we see the following 2 commands fail.

perf stat $cputype -e '{instructions,slots},topdown-retiring' true perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true

Take the 1st command as an example, 5 events are created on hybrid platform. They are cpu_atom/instructions/, cpu_core/instructions/, cpu_core/slots/, cpu_atom/topdown-retiring/ and cpu_core/topdown-retiring/ events. The former 3 events are in a group and the latter 2 topdown-retiring events are independent events.

As the limitation of current implementation, the cpu_core/topdown-retiring/ event can't be moved into previous group as long as there are other events before it. That's the reason why we see the failure.

Thus add "--cputype core" option to limit only P-core events are tested.

---
 tools/perf/arch/x86/util/evlist.c | 26 +++++++++++++++++++++++---
 tools/perf/tests/shell/stat.sh    | 20 ++++++++++++++++++--
 2 files changed, 41 insertions(+), 5 deletions(-)
tools/perf/arch/x86/util/evlist.c +23 -3
index 447a734e591c..0a71ba975871 100644
@@ -9,7 +9,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
 {
        /*
         * Currently the following topdown events sequence are supported to
-        * move and regroup correctly.
+        * move and regroup correctly on non-hybrid platforms.
         *
         * a. all events in a group
         *    perf stat -e "{instructions,topdown-retiring,slots}" -C0 sleep 1
@@ -44,7 +44,7 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
         * topdown metrics events must be first event after the slots event group,
         * otherwise topdown metrics events can't be regrouped correctly, e.g.
         *
-        * a. perf stat -e "{instructions,slots},cycles,topdown-retiring" -C0 sleep 1
+        * e. perf stat -e "{instructions,slots},cycles,topdown-retiring" -C0 sleep 1
         *    WARNING: events were regrouped to match PMUs
         *     Performance counter stats for 'CPU(s) 0':
         *         17,923,134      slots
@@ -56,11 +56,31 @@ int arch_evlist__cmp(const struct evsel *lhs, const struct evsel *rhs)
         * has topdown metrics events must contain only the topdown metrics event,
         * otherwise topdown metrics event can't be regrouped correctly as well, e.g.
         *
-        * a. perf stat -e "{instructions,slots},{topdown-retiring,cycles}" -C0 sleep 1
+        * f. perf stat -e "{instructions,slots},{topdown-retiring,cycles}" -C0 sleep 1
         *    WARNING: events were regrouped to match PMUs
         *    Error:
         *    The sys_perf_event_open() syscall returned with 22 (Invalid argument) for
         *    event (topdown-retiring)
+        *
+        * For hybrid platforms, the sequences 'c' and 'd' are not supported as well
+        * besides above sequences 'e' and 'f'.
+        *
+        *    perf stat -e "{instructions,slots},topdown-retiring" -C0 sleep 1
+        *    perf stat -e "{instructions,slots},{topdown-retiring}" -C0 sleep 1
+        *
+        * On hybrid platforms each event would create an instance on all types of PMU
+        * if the event can be supported by the PMU, i.e., the "topdown-retiring" event
+        * would create two instances on hybrid platforms with P-cores and E-cores,
+        * "cpu_core/topdown-retiring/" and "cpu_atom/topdown_retiring".
+        *
+        * Take the first command as an example, the events list would be converted to
+        * below list in fact.
+        *
+        * "{cpu_atom/instructions/,cpu_core/instructions/,cpu_core/slots/},
+        *  cpu_atom/topdown-retiring/,cpu_core/topdown-retiring/"
+        *
+        * This is actually same with event list in case 'e', "cpu_core/topdown-retiring/"
+        * event can't be moved into previous events group.
         */
        if (topdown_sys_has_perf_metrics() &&
            (arch_evsel__must_be_in_group(lhs) || arch_evsel__must_be_in_group(rhs))) {
tools/perf/tests/shell/stat.sh +15 -0
index 68323d636fb7..cdfe27c25528 100755
@@ -5,6 +5,16 @@
 set -e
 
 err=0
+is_hybrid=false
+
+check_hybrid_platform() {
+  pmus=$(ls /sys/bus/event_source/devices/*/cpus 2>/dev/null | wc -l)
+  if [ "$pmus" -gt 1 ]
+  then
+    is_hybrid=true
+  fi
+}
+
 test_default_stat() {
   echo "Basic stat command test"
   if ! perf stat true 2>&1 | grep -E -q "Performance counter stats for 'true':"
@@ -62,6 +72,11 @@ test_topdown_groups() {
   # Topdown events must be grouped with the slots event first. Test that
   # parse-events reorders this.
   echo "Topdown event group test"
+  cputype=""
+  if $is_hybrid
+  then
+    cputype="--cputype core"
+  fi
   if ! perf stat -e '{slots,topdown-retiring}' true > /dev/null 2>&1
   then
     echo "Topdown event group test [Skipped event parsing failed]"

tools/perf/tests/shell/stat.sh test_topdown_groups() { · line 103

Quoted 5 lines
tools/perf/tests/shell/stat.sh +1 -1
@@ -85,13 +100,13 @@ test_topdown_groups() {
     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},topdown-retiring' true 2>&1 | grep -E -q "<not supported>"

So I can make this test fail on non-hybrid with:

```
$ perf stat -e '{instructions,slots},faults,topdown-retiring' true
WARNING: events were regrouped to match PMUs

Performance counter stats for 'true':

        5,312,770      slots
        1,078,401      instructions
               56      faults
  <not supported>      topdown-retiring
```

The issue is that the slots isn't a group leader, so the "force group index" we try to insert must be grouped events into is miscalculated to be topdown-retiring. The following fixes this:
```

tools/perf/util/parse-events.c +0 -0
index 35e48fe56dfa..68ddc335cde4 100644
@@ -2056,8 +2056,7 @@ static int
 parse_events__sort_events_and_fix_groups(struct list_head *list)
                pos->core.idx = idx++;
 
                /* Remember an index to sort all forced grouped events

together to. */
- if (force_grouped_idx == -1 && pos == pos_leader &&
pos->core.nr_members < 2 &&

tools/perf/util/parse-events.c +1 -1
-                   arch_evsel__must_be_in_group(pos))
+               if (force_grouped_idx == -1 &&

arch_evsel__must_be_in_group(pos))

                        force_grouped_idx = pos->core.idx;
        }

```

tools/perf/tests/shell/stat.sh test_topdown_groups() { · line 109

Quoted 7 lines
tools/perf/tests/shell/stat.sh +1 -1
   then
     echo "Topdown event group test [Failed topdown metrics event not move into slots group]"
     err=1
     return
   fi
-  if perf stat -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"
+  if perf stat $cputype -e '{instructions,slots},{topdown-retiring}' true 2>&1 | grep -E -q "<not supported>"
Quoted 2 lines

So this is trickier as '{topdown-retiring}' ends up looking as a 2 event group. I'll post a proposed fix.

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

Thanks,
Ian

Quoted 15 lines
Quoted 15 lines
tools/perf/tests/shell/stat.sh +1 -1
   then
     echo "Topdown event group test [Failed topdown metrics group not merge into slots group]"
     err=1
@@ -200,6 +215,7 @@ test_hybrid() {
   echo "hybrid test [Success]"
 }
 
+check_hybrid_platform
 test_default_stat
 test_stat_record_report
 test_stat_record_script
--

2.40.1

Tested-by on v1 1/2
#5 lore raw
In reply to Dapeng Mi
Quoted 35 lines · Dapeng Mi

When running topdown leader smapling test on Intel hybrid platforms, such as LNL/ARL, we see the below error.

Topdown leader sampling test
Topdown leader sampling [Failed topdown events not reordered correctly]

It indciates the below command fails.

perf record -o "${perfdata}" -e "{instructions,slots,topdown- retiring}:S" true

The root cause is that perf tool creats a perf event for each PMU type
if it can create.

As for this command, there would be 5 perf events created, cpu_atom/instructions/,cpu_atom/topdown_retiring/, cpu_core/slots/,cpu_core/instructions/,cpu_core/topdown-retiring/

For these 5 events, the 2 cpu_atom events are in a group and the other 3
cpu_core events are in another group.

When arch_topdown_sample_read() traverses all these 5 events, events cpu_atom/instructions/ and cpu_core/slots/ don't have a same group leade, and then return false directly and lead to cpu_core/slots/ event
is used to sample and this is not allowed by PMU driver.

It's a overkill to return false directly if "evsel->core.leader != leader->core.leader" since there could be multiple groups in the event
list.

Quoted 5 lines

Just "continue" instead of "return false" to fix this issue.

Fixes 1e53e9d1787b ("perf x86/topdown: Correct leader selection with

sample_read enabled")

Quoted 21 lines

---
 tools/perf/arch/x86/util/topdown.c | 2 +-

 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/arch/x86/util/topdown.c +1 -1
index f63747d0abdf..d1c654839049 100644
@@ -81,7 +81,7 @@ bool arch_topdown_sample_read(struct evsel *leader)
  	 */
  	evlist__for_each_entry(leader->evlist, evsel) {
  		if (evsel->core.leader != leader->core.leader)
-			return false;
+			continue;

  if (evsel != leader &&
arch_is_topdown_metrics(evsel))
  return true;
  }

base-commit: 4bac7fb5862740087825eda3ed6168e91da8b7e6

reviewer Tested-by
#6 lore raw
Quoted 37 lines · Thomas
Quoted 35 lines · Dapeng Mi

When running topdown leader smapling test on Intel hybrid platforms, such as LNL/ARL, we see the below error.

Topdown leader sampling test
Topdown leader sampling [Failed topdown events not reordered correctly]

It indciates the below command fails.

perf record -o "${perfdata}" -e "{instructions,slots,topdown- retiring}:S" true

The root cause is that perf tool creats a perf event for each PMU type
if it can create.

As for this command, there would be 5 perf events created, cpu_atom/instructions/,cpu_atom/topdown_retiring/, cpu_core/slots/,cpu_core/instructions/,cpu_core/topdown-retiring/

For these 5 events, the 2 cpu_atom events are in a group and the other 3
cpu_core events are in another group.

When arch_topdown_sample_read() traverses all these 5 events, events cpu_atom/instructions/ and cpu_core/slots/ don't have a same group leade, and then return false directly and lead to cpu_core/slots/ event
is used to sample and this is not allowed by PMU driver.

It's a overkill to return false directly if "evsel->core.leader !=
leader->core.leader" since there could be multiple groups in the
event
list.

Quoted 7 lines
Quoted 5 lines

Just "continue" instead of "return false" to fix this issue.

Fixes 1e53e9d1787b ("perf x86/topdown: Correct leader selection with

sample_read enabled")

On an Alderlake:

Thanks,
Ian

Quoted 22 lines
Quoted 21 lines
---
 tools/perf/arch/x86/util/topdown.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/arch/x86/util/topdown.c +1 -1
index f63747d0abdf..d1c654839049 100644
@@ -81,7 +81,7 @@ bool arch_topdown_sample_read(struct evsel *leader)
       */
      evlist__for_each_entry(leader->evlist, evsel) {
              if (evsel->core.leader != leader->core.leader)
-                     return false;
+                     continue;
              if (evsel != leader &&
 arch_is_topdown_metrics(evsel))
                      return true;
      }

base-commit: 4bac7fb5862740087825eda3ed6168e91da8b7e6