Skip to contentSkip to conversation

perf jevents: Don't stop at the first matched pmu when searching a events table

PATCH Applied [email protected] 4 participants 7 messages Jul 16, 2024 – Sep 30, 2024 perf events

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 perf jevents: Don't stop at the first matched pmu when searching a events table

From: Kan Liang <[email protected]>

The "perf all PMU test" fails on a Coffee Lake machine.

The failure is caused by the below change in the commit e2641db83f18 ("perf vendor events: Add/update skylake events/metrics").

+ {

+        "BriefDescription": "This 48-bit fixed counter counts the UCLK cycles",
+        "Counter": "FIXED",
+        "EventCode": "0xff",
+        "EventName": "UNC_CLOCK.SOCKET",
+        "PerPkg": "1",
+        "PublicDescription": "This 48-bit fixed counter counts the UCLK cycles.",
+        "Unit": "cbox_0"
     }

The other cbox events have the unit name "CBOX", while the fixed counter has a unit name "cbox_0". So the events_table will maintain separate entries for cbox and cbox_0.

The perf_pmus__print_pmu_events() calculates the total number of events, allocate an aliases buffer, store all the events into the buffer, sort, and print all the aliases one by one.

The problem is that the calculated total number of events doesn't match the stored events in the aliases buffer.

The perf_pmu__num_events() is used to calculate the number of events. It invokes the pmu_events_table__num_events() to go through the entire events_table to find all events. Because of the pmu_uncore_alias_match(), the suffix of uncore PMU will be ignored. So the events for cbox and cbox_0 are all counted.

When storing events into the aliases buffer, the perf_pmu__for_each_event() only process the events for cbox.

Since a bigger buffer was allocated, the last entry are all 0. When printing all the aliases, null will be outputted, and trigger the failure.

The mismatch was introduced from the commit e3edd6cf6399 ("perf
pmu-events: Reduce processed events by passing PMU"). The pmu_events_table__for_each_event() stops immediately once a pmu is set. But for uncore, especially this case, the method is wrong and mismatch what perf does in the perf_pmu__num_events().

With the patch,

$ perf list pmu | grep -A 1 clock.socket
   unc_clock.socket
        [This 48-bit fixed counter counts the UCLK cycles. Unit: uncore_cbox_0
$ perf test "perf all PMU test"
  107: perf all PMU test                                               : Ok
Fixes e3edd6cf6399 ("perf pmu-events: Reduce processed events by passing PMU")
---
 tools/perf/pmu-events/jevents.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/pmu-events/jevents.py +1 -1
index ac9b7ca41856..97a3b018f865 100755
@@ -923,7 +923,7 @@ int pmu_events_table__for_each_event(const struct pmu_events_table *table,
                         continue;
 
                 ret = pmu_events_table__for_each_event_pmu(table, table_pmu, fn, data);
-                if (pmu || ret)
+                if (ret)
                         return ret;
         }
         return 0;
Signature · 1 line
2.38.1
reviewer Reviewed-by on v1 1/1
#1 lore raw
Quoted 55 lines · On Tue, Jul 16, 2024 at 6:29 AM <[email protected]> wrote:

From: Kan Liang <[email protected]>

The "perf all PMU test" fails on a Coffee Lake machine.

The failure is caused by the below change in the commit e2641db83f18 ("perf vendor events: Add/update skylake events/metrics").

(quoted diff) +8 -0
+    {
+        "BriefDescription": "This 48-bit fixed counter counts the UCLK cycles",
+        "Counter": "FIXED",
+        "EventCode": "0xff",
+        "EventName": "UNC_CLOCK.SOCKET",
+        "PerPkg": "1",
+        "PublicDescription": "This 48-bit fixed counter counts the UCLK cycles.",
+        "Unit": "cbox_0"
     }

The other cbox events have the unit name "CBOX", while the fixed counter has a unit name "cbox_0". So the events_table will maintain separate entries for cbox and cbox_0.

The perf_pmus__print_pmu_events() calculates the total number of events, allocate an aliases buffer, store all the events into the buffer, sort, and print all the aliases one by one.

The problem is that the calculated total number of events doesn't match the stored events in the aliases buffer.

The perf_pmu__num_events() is used to calculate the number of events. It invokes the pmu_events_table__num_events() to go through the entire events_table to find all events. Because of the pmu_uncore_alias_match(), the suffix of uncore PMU will be ignored. So the events for cbox and cbox_0 are all counted.

When storing events into the aliases buffer, the perf_pmu__for_each_event() only process the events for cbox.

Since a bigger buffer was allocated, the last entry are all 0. When printing all the aliases, null will be outputted, and trigger the failure.

The mismatch was introduced from the commit e3edd6cf6399 ("perf
pmu-events: Reduce processed events by passing PMU"). The pmu_events_table__for_each_event() stops immediately once a pmu is set. But for uncore, especially this case, the method is wrong and mismatch what perf does in the perf_pmu__num_events().

With the patch,

$ perf list pmu | grep -A 1 clock.socket
   unc_clock.socket
        [This 48-bit fixed counter counts the UCLK cycles. Unit: uncore_cbox_0
$ perf test "perf all PMU test"
  107: perf all PMU test                                               : Ok
Quoted 4 lines
Fixes e3edd6cf6399 ("perf pmu-events: Reduce processed events by passing PMU")

Awesome sauce, thanks!

Thanks,
Ian

Quoted 20 lines
---
 tools/perf/pmu-events/jevents.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/pmu-events/jevents.py +1 -2
index ac9b7ca41856..97a3b018f865 100755
@@ -923,7 +923,7 @@ int pmu_events_table__for_each_event(const struct pmu_events_table *table,
                         continue;
 
                 ret = pmu_events_table__for_each_event_pmu(table, table_pmu, fn, data);
-                if (pmu || ret)
+                if (ret)
                         return ret;
         }
         return 0;
--

2.38.1

maintainer Applied
#2 lore raw
Quoted 56 lines · Ian Rogers
Quoted 55 lines · On Tue, Jul 16, 2024 at 6:29 AM <[email protected]> wrote:

From: Kan Liang <[email protected]>

The "perf all PMU test" fails on a Coffee Lake machine.

The failure is caused by the below change in the commit e2641db83f18 ("perf vendor events: Add/update skylake events/metrics").

(quoted diff) +8 -0
+    {
+        "BriefDescription": "This 48-bit fixed counter counts the UCLK cycles",
+        "Counter": "FIXED",
+        "EventCode": "0xff",
+        "EventName": "UNC_CLOCK.SOCKET",
+        "PerPkg": "1",
+        "PublicDescription": "This 48-bit fixed counter counts the UCLK cycles.",
+        "Unit": "cbox_0"
     }

The other cbox events have the unit name "CBOX", while the fixed counter has a unit name "cbox_0". So the events_table will maintain separate entries for cbox and cbox_0.

The perf_pmus__print_pmu_events() calculates the total number of events, allocate an aliases buffer, store all the events into the buffer, sort, and print all the aliases one by one.

The problem is that the calculated total number of events doesn't match the stored events in the aliases buffer.

The perf_pmu__num_events() is used to calculate the number of events. It invokes the pmu_events_table__num_events() to go through the entire events_table to find all events. Because of the pmu_uncore_alias_match(), the suffix of uncore PMU will be ignored. So the events for cbox and cbox_0 are all counted.

When storing events into the aliases buffer, the perf_pmu__for_each_event() only process the events for cbox.

Since a bigger buffer was allocated, the last entry are all 0. When printing all the aliases, null will be outputted, and trigger the failure.

The mismatch was introduced from the commit e3edd6cf6399 ("perf
pmu-events: Reduce processed events by passing PMU"). The pmu_events_table__for_each_event() stops immediately once a pmu is set. But for uncore, especially this case, the method is wrong and mismatch what perf does in the perf_pmu__num_events().

With the patch,

$ perf list pmu | grep -A 1 clock.socket
   unc_clock.socket
        [This 48-bit fixed counter counts the UCLK cycles. Unit: uncore_cbox_0
$ perf test "perf all PMU test"
  107: perf all PMU test                                               : Ok
Quoted 8 lines
Quoted 4 lines
Fixes e3edd6cf6399 ("perf pmu-events: Reduce processed events by passing PMU")

Awesome sauce, thanks!

Thanks, applied to tmp.perf-tools-next,

- Arnaldo

reviewer
#3 lore raw
Quoted 63 lines · Arnaldo Carvalho de Melo
Quoted 61 lines · Ian Rogers
Quoted 59 lines · On Tue, Jul 16, 2024 at 6:29 AM <[email protected]> wrote:

From: Kan Liang <[email protected]>

The "perf all PMU test" fails on a Coffee Lake machine.

The failure is caused by the below change in the commit e2641db83f18 ("perf vendor events: Add/update skylake events/metrics").

(quoted diff) +8 -0
+    {
+        "BriefDescription": "This 48-bit fixed counter counts the UCLK cycles",
+        "Counter": "FIXED",
+        "EventCode": "0xff",
+        "EventName": "UNC_CLOCK.SOCKET",
+        "PerPkg": "1",
+        "PublicDescription": "This 48-bit fixed counter counts the UCLK cycles.",
+        "Unit": "cbox_0"
     }

The other cbox events have the unit name "CBOX", while the fixed counter has a unit name "cbox_0". So the events_table will maintain separate entries for cbox and cbox_0.

The perf_pmus__print_pmu_events() calculates the total number of events, allocate an aliases buffer, store all the events into the buffer, sort, and print all the aliases one by one.

The problem is that the calculated total number of events doesn't match the stored events in the aliases buffer.

The perf_pmu__num_events() is used to calculate the number of events. It invokes the pmu_events_table__num_events() to go through the entire events_table to find all events. Because of the pmu_uncore_alias_match(), the suffix of uncore PMU will be ignored. So the events for cbox and cbox_0 are all counted.

When storing events into the aliases buffer, the perf_pmu__for_each_event() only process the events for cbox.

Since a bigger buffer was allocated, the last entry are all 0. When printing all the aliases, null will be outputted, and trigger the failure.

The mismatch was introduced from the commit e3edd6cf6399 ("perf
pmu-events: Reduce processed events by passing PMU"). The pmu_events_table__for_each_event() stops immediately once a pmu is set. But for uncore, especially this case, the method is wrong and mismatch what perf does in the perf_pmu__num_events().

With the patch,

$ perf list pmu | grep -A 1 clock.socket
   unc_clock.socket
        [This 48-bit fixed counter counts the UCLK cycles. Unit: uncore_cbox_0
$ perf test "perf all PMU test"
  107: perf all PMU test                                               : Ok
Fixes e3edd6cf6399 ("perf pmu-events: Reduce processed events by passing PMU")
Quoted 5 lines
Quoted 3 lines

Awesome sauce, thanks!

Thanks, applied to tmp.perf-tools-next,

Did this get applied? I'm not seeing it in perf-tools-next:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/log/tools/perf/pmu-events/jevents.py?h=perf-tools-next

Thanks,
Ian

maintainer
#4 lore raw
Quoted 65 lines · Ian Rogers
Quoted 63 lines · Arnaldo Carvalho de Melo
Quoted 61 lines · Ian Rogers
Quoted 59 lines · On Tue, Jul 16, 2024 at 6:29 AM <[email protected]> wrote:

From: Kan Liang <[email protected]>

The "perf all PMU test" fails on a Coffee Lake machine.

The failure is caused by the below change in the commit e2641db83f18 ("perf vendor events: Add/update skylake events/metrics").

(quoted diff) +8 -0
+    {
+        "BriefDescription": "This 48-bit fixed counter counts the UCLK cycles",
+        "Counter": "FIXED",
+        "EventCode": "0xff",
+        "EventName": "UNC_CLOCK.SOCKET",
+        "PerPkg": "1",
+        "PublicDescription": "This 48-bit fixed counter counts the UCLK cycles.",
+        "Unit": "cbox_0"
     }

The other cbox events have the unit name "CBOX", while the fixed counter has a unit name "cbox_0". So the events_table will maintain separate entries for cbox and cbox_0.

The perf_pmus__print_pmu_events() calculates the total number of events, allocate an aliases buffer, store all the events into the buffer, sort, and print all the aliases one by one.

The problem is that the calculated total number of events doesn't match the stored events in the aliases buffer.

The perf_pmu__num_events() is used to calculate the number of events. It invokes the pmu_events_table__num_events() to go through the entire events_table to find all events. Because of the pmu_uncore_alias_match(), the suffix of uncore PMU will be ignored. So the events for cbox and cbox_0 are all counted.

When storing events into the aliases buffer, the perf_pmu__for_each_event() only process the events for cbox.

Since a bigger buffer was allocated, the last entry are all 0. When printing all the aliases, null will be outputted, and trigger the failure.

The mismatch was introduced from the commit e3edd6cf6399 ("perf
pmu-events: Reduce processed events by passing PMU"). The pmu_events_table__for_each_event() stops immediately once a pmu is set. But for uncore, especially this case, the method is wrong and mismatch what perf does in the perf_pmu__num_events().

With the patch,

$ perf list pmu | grep -A 1 clock.socket
   unc_clock.socket
        [This 48-bit fixed counter counts the UCLK cycles. Unit: uncore_cbox_0
$ perf test "perf all PMU test"
  107: perf all PMU test                                               : Ok
Fixes e3edd6cf6399 ("perf pmu-events: Reduce processed events by passing PMU")
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

Awesome sauce, thanks!

Thanks, applied to tmp.perf-tools-next,

Did this get applied? I'm not seeing it in perf-tools-next:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/log/tools/perf/pmu-events/jevents.py?h=perf-tools-next

I'm seeing this build failure after applying this.

TEST pmu-events/empty-pmu-events.log
--- pmu-events/empty-pmu-events.c 2024-09-26 10:27:15.600055460 -0700
+++ pmu-events/test-empty-pmu-events.c 2024-09-30 13:20:34.631357001 -0700
@@ -380,7 +380,7 @@

                           continue;
   
                   ret = pmu_events_table__for_each_event_pmu(table, table_pmu, fn, data);
  -                if (pmu || ret)
  +                if (ret)
                           return ret;
           }
           return 0;

make[3]: *** [pmu-events/Build:42: pmu-events/empty-pmu-events.log] Error 1 make[3]: *** Deleting file 'pmu-events/empty-pmu-events.log' make[2]: *** [Makefile.perf:765: pmu-events/pmu-events-in.o] Error 2 make[2]: *** Waiting for unfinished jobs....

Thanks,
Namhyung

reviewer
#5 lore raw
Quoted 90 lines · Namhyung Kim
Quoted 65 lines · Ian Rogers
Quoted 63 lines · Arnaldo Carvalho de Melo
Quoted 61 lines · Ian Rogers
Quoted 59 lines · On Tue, Jul 16, 2024 at 6:29 AM <[email protected]> wrote:

From: Kan Liang <[email protected]>

The "perf all PMU test" fails on a Coffee Lake machine.

The failure is caused by the below change in the commit e2641db83f18 ("perf vendor events: Add/update skylake events/metrics").

(quoted diff) +8 -0
+    {
+        "BriefDescription": "This 48-bit fixed counter counts the UCLK cycles",
+        "Counter": "FIXED",
+        "EventCode": "0xff",
+        "EventName": "UNC_CLOCK.SOCKET",
+        "PerPkg": "1",
+        "PublicDescription": "This 48-bit fixed counter counts the UCLK cycles.",
+        "Unit": "cbox_0"
     }

The other cbox events have the unit name "CBOX", while the fixed counter has a unit name "cbox_0". So the events_table will maintain separate entries for cbox and cbox_0.

The perf_pmus__print_pmu_events() calculates the total number of events, allocate an aliases buffer, store all the events into the buffer, sort, and print all the aliases one by one.

The problem is that the calculated total number of events doesn't match the stored events in the aliases buffer.

The perf_pmu__num_events() is used to calculate the number of events. It invokes the pmu_events_table__num_events() to go through the entire events_table to find all events. Because of the pmu_uncore_alias_match(), the suffix of uncore PMU will be ignored. So the events for cbox and cbox_0 are all counted.

When storing events into the aliases buffer, the perf_pmu__for_each_event() only process the events for cbox.

Since a bigger buffer was allocated, the last entry are all 0. When printing all the aliases, null will be outputted, and trigger the failure.

The mismatch was introduced from the commit e3edd6cf6399 ("perf
pmu-events: Reduce processed events by passing PMU"). The pmu_events_table__for_each_event() stops immediately once a pmu is set. But for uncore, especially this case, the method is wrong and mismatch what perf does in the perf_pmu__num_events().

With the patch,

$ perf list pmu | grep -A 1 clock.socket
   unc_clock.socket
        [This 48-bit fixed counter counts the UCLK cycles. Unit: uncore_cbox_0
$ perf test "perf all PMU test"
  107: perf all PMU test                                               : Ok
Fixes e3edd6cf6399 ("perf pmu-events: Reduce processed events by passing PMU")
Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

Awesome sauce, thanks!

Thanks, applied to tmp.perf-tools-next,

Did this get applied? I'm not seeing it in perf-tools-next:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/log/tools/perf/pmu-events/jevents.py?h=perf-tools-next

I'm seeing this build failure after applying this.

TEST pmu-events/empty-pmu-events.log
--- pmu-events/empty-pmu-events.c 2024-09-26 10:27:15.600055460 -0700
+++ pmu-events/test-empty-pmu-events.c 2024-09-30 13:20:34.631357001 -0700
@@ -380,7 +380,7 @@

                           continue;

                   ret = pmu_events_table__for_each_event_pmu(table, table_pmu, fn, data);
  -                if (pmu || ret)
  +                if (ret)
                           return ret;
           }
           return 0;
Quoted 4 lines

make[3]: *** [pmu-events/Build:42: pmu-events/empty-pmu-events.log] Error 1 make[3]: *** Deleting file 'pmu-events/empty-pmu-events.log' make[2]: *** [Makefile.perf:765: pmu-events/pmu-events-in.o] Error 2 make[2]: *** Waiting for unfinished jobs....

Need to copy pmu-events/test-empty-pmu-events.c to pmu-events/empty-pmu-events.c, do you need the patch resent with this done?

Thanks,
Ian

maintainer
#6 lore raw
Quoted 20 lines · Ian Rogers
Quoted 14 lines · Namhyung Kim

I'm seeing this build failure after applying this.

TEST pmu-events/empty-pmu-events.log
--- pmu-events/empty-pmu-events.c 2024-09-26 10:27:15.600055460 -0700
+++ pmu-events/test-empty-pmu-events.c 2024-09-30 13:20:34.631357001 -0700
@@ -380,7 +380,7 @@

                           continue;

                   ret = pmu_events_table__for_each_event_pmu(table, table_pmu, fn, data);
  -                if (pmu || ret)
  +                if (ret)
                           return ret;
           }
           return 0;
Quoted 4 lines

make[3]: *** [pmu-events/Build:42: pmu-events/empty-pmu-events.log] Error 1 make[3]: *** Deleting file 'pmu-events/empty-pmu-events.log' make[2]: *** [Makefile.perf:765: pmu-events/pmu-events-in.o] Error 2 make[2]: *** Waiting for unfinished jobs....

Quoted 3 lines

Need to copy pmu-events/test-empty-pmu-events.c to pmu-events/empty-pmu-events.c, do you need the patch resent with this done?

Yes please!

Thanks
Namhyung