Skip to contentSkip to conversation

perf test stat_all_pmu.sh: Parse return value of perf stat

PATCH Superseded [email protected] 4 participants 18 messages Apr 9, 2024 – Aug 14, 2024 perf events

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf test stat_all_pmu.sh: Parse return value of perf stat

From: Veronika Molnarova <[email protected]>

With the MR a381bd3615de6 ('powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test, however, exits immediately if any command exits with a non-zero value due to 'set -e' option.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with appriopriate action.

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +25 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#1 lore raw
Quoted 3 lines · On Tue, Apr 09, 2024 at 03:02:40PM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the MR a381bd3615de6 ('powerpc/hv-gpci: Fix the

Is this MR publicly accessible? If so, please provide an URL, otherwise it will be interesting to state if this problem takes place upstream and how to reproduce it, do we need a specific environmnt (I assume so, from looking at "powerpc" in the mentioned MR, etc.

- Arnaldo

Quoted 68 lines

H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test, however, exits immediately if any command exits with a non-zero value due to 'set -e' option.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with appriopriate action.

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +25 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#2 lore raw

PATCH v2 1/1 perf test stat_all_pmu.sh: Parse return value of perf stat

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit a381bd3615de6 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit a381bd3615de6:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +25 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#3 lore raw
Quoted 3 lines · On Mon, Apr 15, 2024 at 11:42:20AM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit a381bd3615de6 ('powerpc/hv-gpci:

I'm still not finding this commit upstream:

⬢[acme@toolbox perf-tools-next]$ git remote update torvalds Fetching torvalds
remote: Enumerating objects: 298, done.
remote: Counting objects: 100% (173/173), done.
remote: Compressing objects: 100% (18/18), done.
remote: Total 298 (delta 156), reused 168 (delta 155), pack-reused 125 Receiving objects: 100% (298/298), 208.17 KiB | 859.00 KiB/s, done. Resolving deltas: 100% (213/213), completed with 71 local objects. From git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux

   586b5dfb51b962c1..0bbac3facb5d6cc0  master     -> torvalds/master
 * [new tag]                           v6.9-rc4   -> v6.9-rc4

⬢[acme@toolbox perf-tools-next]$ git show a381bd3615de6
fatal: ambiguous argument 'a381bd3615de6': unknown revision or path not in the working tree. Use '--' to separate paths from revisions, like this:
'git <command> [<revision>...] -- [<file>...]'
⬢[acme@toolbox perf-tools-next]$

⬢[acme@toolbox perf-tools-next]$

IS this one:

⬢[acme@toolbox perf-tools-next]$ git log | grep "Fix the H_GET_PERF_COUNTER_INFO hcall return value checks"
powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks
powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks
^C
⬢[acme@toolbox perf-tools-next]$

⬢[acme@toolbox perf-tools-next]$ git log | grep -B4 "Fix the H_GET_PERF_COUNTER_INFO hcall return value checks"
powerpc/fsl: Fix mfpmr build errors with newer binutils
powerpc/64s: Use .machine power4 around dcbt
powerpc/64s: Move dcbt/dcbtst sequence into a macro
powerpc/mm: Code cleanup for __hash_page_thp
powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks
--
commit ad86d7ee43b22aa2ed60fb982ae94b285c1be671

Author: Kajol Jain <[email protected]>
Date:   Thu Feb 29 17:58:47 2024 +0530

powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks
^C
⬢[acme@toolbox perf-tools-next]$

I think it is, now looking at
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916
I see the reference to ad86d7ee43b22aa2ed60fb982ae94b285c1be671

I'm replacing this:

Quoted 1 line

With the upstream MR !3916 of commit a381bd3615de6 ('powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks')

with:

Quoted 1 line

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ("powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks")

Ok?

- Arnaldo

Quoted 82 lines

Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit a381bd3615de6:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +25 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#4 lore raw
In reply to [email protected]
Quoted 4 lines · On Mon, Apr 15, 2024 at 11:42:20AM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit a381bd3615de6 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf

Athira/Kajol,

I don't have access to such machines, since you did work on this
area, could you please take a look and maybe test it?

The upstream patch mentioned is in fact:

ad86d7ee43b22aa2 ("powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks")

Thanks,

- Arnaldo

Quoted 81 lines

stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit a381bd3615de6:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +25 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#5 lore raw
Quoted 11 lines · Arnaldo Carvalho de Melo
Quoted 4 lines · On Mon, Apr 15, 2024 at 11:42:20AM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit a381bd3615de6 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf

Athira/Kajol,

I don't have access to such machines, since you did work on this
area, could you please take a look and maybe test it?

Quoted 7 lines

The upstream patch mentioned is in fact:

ad86d7ee43b22aa2 ("powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks")

Thanks,

- Arnaldo

Hi Arnaldo,
Sure we will review and test the patch.

Thanks,
Kajol Jain

Quoted 83 lines
Quoted 82 lines

stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit a381bd3615de6:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +25 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#6 lore raw
In reply to [email protected]

Hi Veronika,
I was trying to test this patch. But I am not able to apply this
patch cleanly on both upstream and acme's tmp.perf-tools-next branch.

error: patch failed: tools/perf/tests/shell/stat_all_pmu.sh:2
error: tools/perf/tests/shell/stat_all_pmu.sh: patch does not apply

Can you check that part, but since its simple change I added it manually and could test the patch.

Patch looks fine to me. I tested it by disabling hv-gpci events in HMC

Without patch changes:

localhost:/dev/shm/linux/tools/perf # ./perf test 95 -vv
95: perf all PMU test:
--- start ---
test child forked, pid 2234
---
Testing cpu/stalled-cycles-frontend/
Testing hv_gpci/system_hypervisor_times_time_spent_managing_partitions_over_entitlement/ ---- end(-1) ----
95: perf all PMU test :
FAILED!

with patch changes:

Testing event 'hv_gpci/system_hypervisor_times_time_spent_managing_partitions_over_entitlement/' -- no permission to enable
Testing event 'hv_gpci/system_hypervisor_times_time_spent_on_system_management/' -- no permission to enable
Testing event 'hv_gpci/system_hypervisor_times_time_spent_processing_virtual_processor_timers/' -- no permission to enable
Testing event 'hv_gpci/system_hypervisor_times_time_spent_to_dispatch_virtual_processors/' -- no permission to enable
Testing event 'hv_gpci/system_tlbie_count_and_time_time_spent_issuing_tlbies/' -- no permission to enable
Testing event 'hv_gpci/system_tlbie_count_and_time_tlbie_instructions_issued/' -- no permission to enable
Testing event 'pm_br_mpred_cmpl' -- supported

Quoted 36 lines · On 4/15/24 15:12, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit a381bd3615de6 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit a381bd3615de6:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +0 -0
index c77955419173..d9f0d2100baa 100755

tools/perf/tests/shell/stat_all_pmu.sh line 5

Quoted 4 lines
tools/perf/tests/shell/stat_all_pmu.sh +0 -1
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e

It will create extra blank line can you remove that. Also make sure to test the patch with checkpatch script to avoid any format issue.

Thanks,
Kajol Jain

Quoted 42 lines
tools/perf/tests/shell/stat_all_pmu.sh +25 -10
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
#7 lore raw

Hi Jain,

Thanks for letting me know, there was just some typo fixed in the comment upstream, which caused it to fail. Will fix the patch, also add Ian's suggestion and resend it shortly.

Thanks,
Veronika

Quoted 139 lines · kajoljain

Hi Veronika,
I was trying to test this patch. But I am not able to apply this
patch cleanly on both upstream and acme's tmp.perf-tools-next branch.

error: patch failed: tools/perf/tests/shell/stat_all_pmu.sh:2
error: tools/perf/tests/shell/stat_all_pmu.sh: patch does not apply

Can you check that part, but since its simple change I added it manually and could test the patch.

Patch looks fine to me. I tested it by disabling hv-gpci events in HMC

Without patch changes:

localhost:/dev/shm/linux/tools/perf # ./perf test 95 -vv
95: perf all PMU test:
--- start ---
test child forked, pid 2234
---
Testing cpu/stalled-cycles-frontend/
Testing hv_gpci/system_hypervisor_times_time_spent_managing_partitions_over_entitlement/ ---- end(-1) ----
95: perf all PMU test :
FAILED!

with patch changes:

Testing event 'hv_gpci/system_hypervisor_times_time_spent_managing_partitions_over_entitlement/' -- no permission to enable
Testing event 'hv_gpci/system_hypervisor_times_time_spent_on_system_management/' -- no permission to enable
Testing event 'hv_gpci/system_hypervisor_times_time_spent_processing_virtual_processor_timers/' -- no permission to enable
Testing event 'hv_gpci/system_hypervisor_times_time_spent_to_dispatch_virtual_processors/' -- no permission to enable
Testing event 'hv_gpci/system_tlbie_count_and_time_time_spent_issuing_tlbies/' -- no permission to enable
Testing event 'hv_gpci/system_tlbie_count_and_time_tlbie_instructions_issued/' -- no permission to enable
Testing event 'pm_br_mpred_cmpl' -- supported

Quoted 36 lines · On 4/15/24 15:12, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit a381bd3615de6 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit a381bd3615de6:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +0 -0
index c77955419173..d9f0d2100baa 100755

tools/perf/tests/shell/stat_all_pmu.sh line 5

Quoted 4 lines
tools/perf/tests/shell/stat_all_pmu.sh +0 -1
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e

It will create extra blank line can you remove that. Also make sure to test the patch with checkpatch script to avoid any format issue.

Thanks,
Kajol Jain

Quoted 42 lines
tools/perf/tests/shell/stat_all_pmu.sh +25 -10
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
reviewer
#8 lore raw
In reply to [email protected]
Quoted 10 lines · On Tue, Apr 9, 2024 at 6:03 AM <[email protected]> wrote:

From: Veronika Molnarova <[email protected]>

With the MR a381bd3615de6 ('powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test, however, exits immediately if any command exits with a non-zero value due to 'set -e' option.

Quoted 4 lines

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with appriopriate action.

nit: s/appriopriate/appropriate/

There was a similar issue with metric groups that should be on its way to landing:
https://lore.kernel.org/lkml/[email protected]/

Quoted 26 lines
---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +0 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1

tools/perf/tests/shell/stat_all_pmu.sh line 17

Quoted 10 lines
tools/perf/tests/shell/stat_all_pmu.sh +10 -0
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set

nit: s/pemission/permission/

tools/perf/tests/shell/stat_all_pmu.sh line 24

Quoted 7 lines
tools/perf/tests/shell/stat_all_pmu.sh +7 -0
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1

Something we've been seeing with parallel testing is getting busy PMUs. The error looks like:

```
$ sudo bash -c "perf stat -e intel_bts// -a sleep 1; echo $?" & sudo
bash -c "perf stat -e intel_bts// -a sleep 1; echo $?"
[2] 953102
Error:
The sys_perf_event_open() syscall returned with 16 (Device or resource
busy) for event (intel_bts//).
/bin/dmesg | grep -i perf may provide additional information.

255

Performance counter stats for 'system wide':

    <not counted>      intel_bts//
                        (0.00%)

      1.002334792 seconds time elapsed

0
[2]+  Done                    sudo bash -c "perf stat -e intel_bts//
-a sleep 1; echo $?"
```

To avoid failing for this we probably need something like:

```
   elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q
"Device or resource busy"; then
       echo "device busy"
```

It'd be great if you could add this.

Thanks,
Ian

Quoted 16 lines
tools/perf/tests/shell/stat_all_pmu.sh +8 -1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
--

2.43.0

#9 lore raw
Quoted 16 lines · Ian Rogers
Quoted 10 lines · On Tue, Apr 9, 2024 at 6:03 AM <[email protected]> wrote:

From: Veronika Molnarova <[email protected]>

With the MR a381bd3615de6 ('powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test, however, exits immediately if any command exits with a non-zero value due to 'set -e' option.

Quoted 4 lines

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with appriopriate action.

Quoted 5 lines

nit: s/appriopriate/appropriate/

There was a similar issue with metric groups that should be on its way to landing:
https://lore.kernel.org/lkml/[email protected]/

Cool, looks similar indeed, Veronika, can you consider reviewing Ian's patch and providing a Reviewed-by: for it?

And also please check Ian's suggestions for your patch, below:

Quoted 99 lines
Quoted 26 lines
---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +0 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1

tools/perf/tests/shell/stat_all_pmu.sh line 17

Quoted 10 lines
tools/perf/tests/shell/stat_all_pmu.sh +10 -0
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set

nit: s/pemission/permission/

tools/perf/tests/shell/stat_all_pmu.sh line 24

Quoted 7 lines
tools/perf/tests/shell/stat_all_pmu.sh +7 -0
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1

Something we've been seeing with parallel testing is getting busy PMUs. The error looks like:

```
$ sudo bash -c "perf stat -e intel_bts// -a sleep 1; echo $?" & sudo
bash -c "perf stat -e intel_bts// -a sleep 1; echo $?"
[2] 953102
Error:
The sys_perf_event_open() syscall returned with 16 (Device or resource
busy) for event (intel_bts//).
/bin/dmesg | grep -i perf may provide additional information.

255

Performance counter stats for 'system wide':

    <not counted>      intel_bts//
                        (0.00%)

      1.002334792 seconds time elapsed

0
[2]+  Done                    sudo bash -c "perf stat -e intel_bts//
-a sleep 1; echo $?"
```

To avoid failing for this we probably need something like:

```
   elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q
"Device or resource busy"; then
       echo "device busy"
```

It'd be great if you could add this.

Thanks,
Ian

Quoted 16 lines
tools/perf/tests/shell/stat_all_pmu.sh +8 -1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
--

2.43.0

#10 lore raw

PATCH v2 1/1 perf test stat_all_pmu.sh: Parse return value of perf stat

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit ad86d7ee43b22aa2:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
Fixed the issue with applying due to a fixed comment typo upstream and added Ian's suggestion for the 'device busy' issue during parallel testing.

 tools/perf/tests/shell/stat_all_pmu.sh | 41 ++++++++++++++++++--------
 1 file changed, 29 insertions(+), 12 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +28 -11
index c77955419173..a75beddda4db 100755
@@ -2,21 +2,38 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parameterized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters get printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required permission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 255 in case of resource busy during parallel testing of events
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "Device or resource busy"; then
+        echo "resource busy"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#11 lore raw
Quoted 9 lines · On Mon, Apr 29, 2024 at 12:56:24PM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Is this v2 or v3? Kajol, can I have your tested-by?

Thanks,

- Arnaldo

Quoted 82 lines

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit ad86d7ee43b22aa2:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
Fixed the issue with applying due to a fixed comment typo upstream and added Ian's suggestion for the 'device busy' issue during parallel testing.

 tools/perf/tests/shell/stat_all_pmu.sh | 41 ++++++++++++++++++--------
 1 file changed, 29 insertions(+), 12 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +28 -11
index c77955419173..a75beddda4db 100755
@@ -2,21 +2,38 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parameterized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters get printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required permission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 255 in case of resource busy during parallel testing of events
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "Device or resource busy"; then
+        echo "resource busy"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#12 lore raw
Quoted 11 lines · Arnaldo Carvalho de Melo
Quoted 9 lines · On Mon, Apr 29, 2024 at 12:56:24PM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Quoted 6 lines

Is this v2 or v3? Kajol, can I have your tested-by?

Thanks,

- Arnaldo

From my point of view it was v2 of the patch, but now I noticed that the a v2 was created when the commit hash was firstly fixed, so maybe its then a v3?

Sorry for confusion,
Veronika

Quoted 83 lines
Quoted 82 lines

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit ad86d7ee43b22aa2:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
Fixed the issue with applying due to a fixed comment typo upstream and added Ian's suggestion for the 'device busy' issue during parallel testing.

 tools/perf/tests/shell/stat_all_pmu.sh | 41 ++++++++++++++++++--------
 1 file changed, 29 insertions(+), 12 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +28 -11
index c77955419173..a75beddda4db 100755
@@ -2,21 +2,38 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parameterized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters get printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required permission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 255 in case of resource busy during parallel testing of events
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "Device or resource busy"; then
+        echo "resource busy"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#13 lore raw
In reply to Arnaldo Carvalho de Melo
Quoted 11 lines · Arnaldo Carvalho de Melo
Quoted 9 lines · On Mon, Apr 29, 2024 at 12:56:24PM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Quoted 1 line

Is this v2 or v3? Kajol, can I have your tested-by?

Oops, Kajol wasn't on the CC list, nor Ian, I'm adding them, maybe Kajol can test it on those machines and Ian, who wrote the test can take a look, also adding all the perf tools reviewers, as listed in MAINTAINERS.

Thanks,

- Arnaldo

Quoted 82 lines
Quoted 82 lines

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit ad86d7ee43b22aa2:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
Fixed the issue with applying due to a fixed comment typo upstream and added Ian's suggestion for the 'device busy' issue during parallel testing.

 tools/perf/tests/shell/stat_all_pmu.sh | 41 ++++++++++++++++++--------
 1 file changed, 29 insertions(+), 12 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +28 -11
index c77955419173..a75beddda4db 100755
@@ -2,21 +2,38 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parameterized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters get printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required permission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 255 in case of resource busy during parallel testing of events
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "Device or resource busy"; then
+        echo "resource busy"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
Reviewed-by Tested-by
#14 lore raw
Quoted 14 lines · Arnaldo Carvalho de Melo
Quoted 11 lines · Arnaldo Carvalho de Melo
Quoted 9 lines · On Mon, Apr 29, 2024 at 12:56:24PM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Quoted 1 line

Is this v2 or v3? Kajol, can I have your tested-by?

Quoted 9 lines

Oops, Kajol wasn't on the CC list, nor Ian, I'm adding them, maybe Kajol can test it on those machines and Ian, who wrote the test can take a look, also adding all the perf tools reviewers, as listed in MAINTAINERS.

Thanks,

- Arnaldo

Hi Veronika,
I checked the patch, code wise patch looks fine to me. I also tested it on powerpc system.

Just have concern on your commit message, you added below text as part of your commit message:

It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

The hv-gpci events are supported in pseries system.
And for different power processors, we already have code to include only supported event as part of perf list.

If you are taking about parametrized events which are skipped in this testcase, its not that those parametrized events are not supported. Since testing proper values of the parameters is out of scope of this script we are skipping those events.

Can you remove that part from commit message? Rest looks fine to me.

Thanks,
Kajol Jain

Quoted 83 lines
Quoted 82 lines
Quoted 82 lines

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit ad86d7ee43b22aa2:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
Fixed the issue with applying due to a fixed comment typo upstream and added Ian's suggestion for the 'device busy' issue during parallel testing.

 tools/perf/tests/shell/stat_all_pmu.sh | 41 ++++++++++++++++++--------
 1 file changed, 29 insertions(+), 12 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +28 -11
index c77955419173..a75beddda4db 100755
@@ -2,21 +2,38 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parameterized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters get printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required permission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 255 in case of resource busy during parallel testing of events
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "Device or resource busy"; then
+        echo "resource busy"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#15 lore raw
Quoted 41 lines · kajoljain
Quoted 14 lines · Arnaldo Carvalho de Melo
Quoted 11 lines · Arnaldo Carvalho de Melo
Quoted 9 lines · On Mon, Apr 29, 2024 at 12:56:24PM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Quoted 1 line

Is this v2 or v3? Kajol, can I have your tested-by?

Quoted 9 lines

Oops, Kajol wasn't on the CC list, nor Ian, I'm adding them, maybe Kajol can test it on those machines and Ian, who wrote the test can take a look, also adding all the perf tools reviewers, as listed in MAINTAINERS.

Thanks,

- Arnaldo

Hi Veronika,
I checked the patch, code wise patch looks fine to me. I also tested it on powerpc system.

Just have concern on your commit message, you added below text as part of your commit message:

It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

The hv-gpci events are supported in pseries system.
And for different power processors, we already have code to include only supported event as part of perf list.

Quoted 6 lines

If you are taking about parametrized events which are skipped in this testcase, its not that those parametrized events are not supported. Since testing proper values of the parameters is out of scope of this script we are skipping those events.

Can you remove that part from commit message? Rest looks fine to me.

Veronika, can you please address the reviewer's comment and send a v4 patch, collecting the tags he provided?

Thanks,

- Arnaldo

Quoted 89 lines

Thanks,
Kajol Jain

Quoted 83 lines
Quoted 82 lines
Quoted 82 lines

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit ad86d7ee43b22aa2:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
Fixed the issue with applying due to a fixed comment typo upstream and added Ian's suggestion for the 'device busy' issue during parallel testing.

 tools/perf/tests/shell/stat_all_pmu.sh | 41 ++++++++++++++++++--------
 1 file changed, 29 insertions(+), 12 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +28 -11
index c77955419173..a75beddda4db 100755
@@ -2,21 +2,38 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parameterized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters get printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required permission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 255 in case of resource busy during parallel testing of events
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "Device or resource busy"; then
+        echo "resource busy"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0
#16 lore raw
Quoted 47 lines · Arnaldo Carvalho de Melo
Quoted 46 lines · kajoljain
Quoted 14 lines · Arnaldo Carvalho de Melo
Quoted 11 lines · Arnaldo Carvalho de Melo
Quoted 9 lines · On Mon, Apr 29, 2024 at 12:56:24PM +0200, [email protected] wrote:

From: Veronika Molnarova <[email protected]>

With the upstream MR !3916 of commit ad86d7ee43b22aa2 ('powerpc/hv-gpci:
Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test was designed in a way, that if any command exits with a non-zero value the test exits with an error value without any information provided due to the 'set -e' option.

Quoted 1 line

Is this v2 or v3? Kajol, can I have your tested-by?

Quoted 9 lines

Oops, Kajol wasn't on the CC list, nor Ian, I'm adding them, maybe Kajol can test it on those machines and Ian, who wrote the test can take a look, also adding all the perf tools reviewers, as listed in MAINTAINERS.

Thanks,

- Arnaldo

Hi Veronika,
I checked the patch, code wise patch looks fine to me. I also tested it on powerpc system.

Just have concern on your commit message, you added below text as part of your commit message:

It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

The hv-gpci events are supported in pseries system.
And for different power processors, we already have code to include only supported event as part of perf list.

If you are taking about parametrized events which are skipped in this testcase, its not that those parametrized events are not supported. Since testing proper values of the parameters is out of scope of this script we are skipping those events.

Quoted 99 lines
Quoted 1 line

Can you remove that part from commit message? Rest looks fine to me.

Veronika, can you please address the reviewer's comment and send a v4 patch, collecting the tags he provided?

Thanks,

- Arnaldo

Quoted 89 lines

Thanks,
Kajol Jain

Quoted 83 lines
Quoted 82 lines
Quoted 82 lines

Running stat_all_pmu test on powerpc machine with unsupported hv_gpci event causes failure after the MR as the zero return value was required. The issue propagated upstream as the list of the files that affected perf did not cover the changed files and was updated after the issue was discovered. It was caught by CKI testing where it was triaged to stop blocking further MRs, as most of the powerpc machines do not support some of the hv_gpci events.

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with the appropriate action.

Link to the MR !3916 of commit ad86d7ee43b22aa2:
https://gitlab.com/redhat/centos-stream/src/kernel/centos-stream-9/-/merge_requests/3916

---
Fixed the issue with applying due to a fixed comment typo upstream and added Ian's suggestion for the 'device busy' issue during parallel testing.

 tools/perf/tests/shell/stat_all_pmu.sh | 41 ++++++++++++++++++--------
 1 file changed, 29 insertions(+), 12 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +28 -11
index c77955419173..a75beddda4db 100755
@@ -2,21 +2,38 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parameterized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters get printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required permission for the event is not set
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 255 in case of resource busy during parallel testing of events
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "Device or resource busy"; then
+        echo "resource busy"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
Signature · 1 line
2.43.0

Looking at it, the issue wasn't with the unsupported hv-gpci events, but the ones without permissions to retrieve information from their partitions as mentioned in the first part of the commit message.

Will rework the commit message and re-send the patch.

Thanks,
-Veronika

#17 lore raw
In reply to Ian Rogers

Hi Ian,

Thanks for looking at this, will check the problem with busy PMUs and add your proposition.

Best,
Veronika

Quoted 122 lines · Ian Rogers
Quoted 10 lines · On Tue, Apr 9, 2024 at 6:03 AM <[email protected]> wrote:

From: Veronika Molnarova <[email protected]>

With the MR a381bd3615de6 ('powerpc/hv-gpci: Fix the H_GET_PERF_COUNTER_INFO hcall return value checks') the perf stat for hv_gpci events without required permission set returns an error value of -1 to differentiate the output from the unsupported events. The stat_all_pmu test, however, exits immediately if any command exits with a non-zero value due to 'set -e' option.

Quoted 4 lines

Remove the 'set -e' option from the test and rework the test case to log the status of the event for better maintainability. Instead of exiting immediately after 'perf stat' ends with a non-zero value, check the return value and output of the 'perf stat' command with appriopriate action.

nit: s/appriopriate/appropriate/

There was a similar issue with metric groups that should be on its way to landing:
https://lore.kernel.org/lkml/[email protected]/

Quoted 26 lines
---
 tools/perf/tests/shell/stat_all_pmu.sh | 36 ++++++++++++++++++--------
 1 file changed, 25 insertions(+), 11 deletions(-)
tools/perf/tests/shell/stat_all_pmu.sh +0 -11
index c77955419173..d9f0d2100baa 100755
@@ -2,21 +2,35 @@
 # perf all PMU test
 # SPDX-License-Identifier: GPL-2.0
 
-set -e
 
 # Test all PMU events; however exclude parametrized ones (name contains '?')
 for p in $(perf list --raw-dump pmu | sed 's/[[:graph:]]\+?[[:graph:]]\+[[:space:]]//g'); do
-  echo "Testing $p"
-  result=$(perf stat -e "$p" true 2>&1)
-  if ! echo "$result" | grep -q "$p" && ! echo "$result" | grep -q "<not supported>" ; then
-    # We failed to see the event and it is supported. Possibly the workload was
-    # too small so retry with something longer.
-    result=$(perf stat -e "$p" perf bench internals synthesize 2>&1)
-    if ! echo "$result" | grep -q "$p" ; then
-      echo "Event '$p' not printed in:"
-      echo "$result"
-      exit 1

tools/perf/tests/shell/stat_all_pmu.sh line 17

Quoted 10 lines
tools/perf/tests/shell/stat_all_pmu.sh +10 -0
+  echo -n "Testing event '$p' -- "
+  stat_output=$(perf stat -e "$p" true 2>&1)
+  stat_result=$?
+  if echo "$stat_output" | grep -q "$p"; then
+    # return value 0 if counters gets printed either if the event is supported or not
+    if [ $stat_result -eq 0 ] && ! echo "$stat_output" | grep -q "<not supported>"; then
+        echo "supported"
+    elif [ $stat_result -eq 0 ]; then
+        echo "not supported"
+    # return value 255 when the required pemission for the event is not set

nit: s/pemission/permission/

tools/perf/tests/shell/stat_all_pmu.sh line 24

Quoted 7 lines
tools/perf/tests/shell/stat_all_pmu.sh +7 -0
+    elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q "No permission"; then
+        echo "no permission to enable"
+    # return value 129 when trying to run 'perf stat' with a non-existent event
+    elif [ $stat_result -eq 129 ] && echo "$stat_output" | grep -q "Bad event name"; then
+        echo "Fail: Bad event name"
+        echo "$stat_output"
+        exit 1

Something we've been seeing with parallel testing is getting busy PMUs. The error looks like:

```
$ sudo bash -c "perf stat -e intel_bts// -a sleep 1; echo $?" & sudo
bash -c "perf stat -e intel_bts// -a sleep 1; echo $?"
[2] 953102
Error:
The sys_perf_event_open() syscall returned with 16 (Device or resource
busy) for event (intel_bts//).
/bin/dmesg | grep -i perf may provide additional information.

255

Performance counter stats for 'system wide':

    <not counted>      intel_bts//
                        (0.00%)

      1.002334792 seconds time elapsed

0
[2]+  Done                    sudo bash -c "perf stat -e intel_bts//
-a sleep 1; echo $?"
```

To avoid failing for this we probably need something like:

```
   elif [ $stat_result -eq 255 ] && echo "$stat_output" | grep -q
"Device or resource busy"; then
       echo "device busy"
```

It'd be great if you could add this.

Thanks,
Ian

Quoted 16 lines
tools/perf/tests/shell/stat_all_pmu.sh +8 -1
+    else
+        echo "Fail: Unknown return value $stat_result"
+        echo "$stat_output"
+        exit 1
     fi
+  else
+      echo "Fail: Event '$p' not printed in:"
+      echo "$stat_output"
+      exit 1
   fi
 done
 
--

2.43.0