I noticed "perf list" reported the error as follows on an Armv8 Neoverse
N2 server:
Unexpected event smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
The root cause is due to the incorrect calculation in
perf_pmu__num_events().
Jia He (2):
perf pmu: Allow finishing loading json events when !events_table
perf pmu: Fix num_events calculation
tools/perf/util/pmu.c | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
--
2.34.1
When pe is NULL in the function perf_pmu__new_alias(), the total number
of events is added to loaded_json_aliases. However, if pmu->events_table
is NULL and cpu_aliases_added is false, the calculation for the events
number in perf_pmu__num_events() is incorrect.
Then cause the error report after "perf list":
Unexpected event smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
Fix it by adding loaded_json_aliases in the calculation under the
mentioned conditions.
Test it also with "perf bench internals pmu-scan" and there is no
regression.
Fixes: e6ff1eed3584 ("perf pmu: Lazily add JSON events")
Signed-off-by: Jia He <[email protected]>
---
tools/perf/util/pmu.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
index a1eef7b2e389..a53224e2ce7e 100644
--- a/tools/perf/util/pmu.c
+++ b/tools/perf/util/pmu.c
@@ -1639,6 +1639,8 @@ size_t perf_pmu__num_events(struct perf_pmu *pmu)
nr += pmu->loaded_json_aliases;
else if (pmu->events_table)
nr += pmu_events_table__num_events(pmu->events_table, pmu) - pmu->loaded_json_aliases;
+ else
+ nr += pmu->loaded_json_aliases;
return pmu->selectable ? nr + 1 : nr;
}
--
2.34.1
Otherwise, cpu_aliases_added is never set to true on an Arm v8a
Neoverse N2 server.
Signed-off-by: Jia He <[email protected]>
---
tools/perf/util/pmu.c | 6 ++----
1 file changed, 2 insertions(+), 4 deletions(-)
diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
index f39cbbc1a7ec..a1eef7b2e389 100644
--- a/tools/perf/util/pmu.c
+++ b/tools/perf/util/pmu.c
@@ -915,13 +915,11 @@ void pmu_add_cpu_aliases_table(struct perf_pmu *pmu, const struct pmu_events_tab
static void pmu_add_cpu_aliases(struct perf_pmu *pmu)
{
- if (!pmu->events_table)
- return;
-
if (pmu->cpu_aliases_added)
return;
- pmu_add_cpu_aliases_table(pmu, pmu->events_table);
+ if (pmu->events_table)
+ pmu_add_cpu_aliases_table(pmu, pmu->events_table);
pmu->cpu_aliases_added = true;
}
--
2.34.1
On Thu, May 9, 2024 at 7:47 PM Jia He <[email protected]> wrote:
>
> When pe is NULL in the function perf_pmu__new_alias(), the total number
> of events is added to loaded_json_aliases. However, if pmu->events_table
> is NULL and cpu_aliases_added is false, the calculation for the events
> number in perf_pmu__num_events() is incorrect.
>
> Then cause the error report after "perf list":
> Unexpected event smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
>
> Fix it by adding loaded_json_aliases in the calculation under the
> mentioned conditions.
>
> Test it also with "perf bench internals pmu-scan" and there is no
> regression.
>
> Fixes: e6ff1eed3584 ("perf pmu: Lazily add JSON events")
> Signed-off-by: Jia He <[email protected]>
> ---
> tools/perf/util/pmu.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
> index a1eef7b2e389..a53224e2ce7e 100644
> --- a/tools/perf/util/pmu.c
> +++ b/tools/perf/util/pmu.c
> @@ -1639,6 +1639,8 @@ size_t perf_pmu__num_events(struct perf_pmu *pmu)
> nr += pmu->loaded_json_aliases;
> else if (pmu->events_table)
> nr += pmu_events_table__num_events(pmu->events_table, pmu) - pmu->loaded_json_aliases;
> + else
> + nr += pmu->loaded_json_aliases;
Thanks for working on this! The "struct pmu_event *pe" in new_alias is
an entry from the json data, and "pmu->events_table" should NULL if
there is no json data. I believe the code is assuming that these lines
aren't necessary as it shouldn't be possible to load json data if the
json events table doesn't exist for the PMU - if there is no json data
then loaded_json_aliases should be 0 and no addition is necessary. I'm
wondering why this case isn't true for you.
Thanks,
Ian
>
> return pmu->selectable ? nr + 1 : nr;
> }
> --
> 2.34.1
>
Hi, Ian
> -----Original Message-----
> From: Ian Rogers <[email protected]>
> Sent: Friday, May 10, 2024 2:17 PM
> To: Justin He <[email protected]>
> Cc: Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>;
> Arnaldo Carvalho de Melo <[email protected]>; Namhyung Kim
> <[email protected]>; Mark Rutland <[email protected]>; Alexander
> Shishkin <[email protected]>; Jiri Olsa <[email protected]>;
> Adrian Hunter <[email protected]>; Kan Liang
> <[email protected]>; James Clark <[email protected]>;
> [email protected]; [email protected]
> Subject: Re: [PATCH 2/2] perf pmu: Fix num_events calculation
>
> On Thu, May 9, 2024 at 7:47 PM Jia He <[email protected]> wrote:
> >
> > When pe is NULL in the function perf_pmu__new_alias(), the total
> > number of events is added to loaded_json_aliases. However, if
> > pmu->events_table is NULL and cpu_aliases_added is false, the
> > calculation for the events number in perf_pmu__num_events() is incorrect.
> >
> > Then cause the error report after "perf list":
> > Unexpected event
> smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
> >
> > Fix it by adding loaded_json_aliases in the calculation under the
> > mentioned conditions.
> >
> > Test it also with "perf bench internals pmu-scan" and there is no
> > regression.
> >
> > Fixes: e6ff1eed3584 ("perf pmu: Lazily add JSON events")
> > Signed-off-by: Jia He <[email protected]>
> > ---
> > tools/perf/util/pmu.c | 2 ++
> > 1 file changed, 2 insertions(+)
> >
> > diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index
> > a1eef7b2e389..a53224e2ce7e 100644
> > --- a/tools/perf/util/pmu.c
> > +++ b/tools/perf/util/pmu.c
> > @@ -1639,6 +1639,8 @@ size_t perf_pmu__num_events(struct perf_pmu
> *pmu)
> > nr += pmu->loaded_json_aliases;
> > else if (pmu->events_table)
> > nr +=
> pmu_events_table__num_events(pmu->events_table,
> > pmu) - pmu->loaded_json_aliases;
> > + else
> > + nr += pmu->loaded_json_aliases;
>
> Thanks for working on this! The "struct pmu_event *pe" in new_alias is an entry
> from the json data, and "pmu->events_table" should NULL if there is no json
> data. I believe the code is assuming that these lines aren't necessary as it
> shouldn't be possible to load json data if the json events table doesn't exist for
> the PMU - if there is no json data then loaded_json_aliases should be 0 and no
> addition is necessary. I'm wondering why this case isn't true for you.
On my Armv8a N2 server, "pmu->events_table" is NULL because perf_pmu__find_events_table()
return NULL.
I also noticed that pmu->loaded_json_aliases is *not* 0. The missing adding calculation will cause
perf_pmu__num_events() less than normal case and will trigger latter check failure in
perf_pmus__print_pmu_events__callback().
At last, perf list will report many lines similar as:
Unexpected event smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
--
Cheers,
Justin (Jia He)
On Thu, May 9, 2024 at 11:52 PM Justin He <[email protected]> wrote:
>
> Hi, Ian
>
> > -----Original Message-----
> > From: Ian Rogers <[email protected]>
> > Sent: Friday, May 10, 2024 2:17 PM
> > To: Justin He <[email protected]>
> > Cc: Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>;
> > Arnaldo Carvalho de Melo <[email protected]>; Namhyung Kim
> > <[email protected]>; Mark Rutland <[email protected]>; Alexander
> > Shishkin <[email protected]>; Jiri Olsa <[email protected]>;
> > Adrian Hunter <[email protected]>; Kan Liang
> > <[email protected]>; James Clark <[email protected]>;
> > [email protected]; [email protected]
> > Subject: Re: [PATCH 2/2] perf pmu: Fix num_events calculation
> >
> > On Thu, May 9, 2024 at 7:47 PM Jia He <[email protected]> wrote:
> > >
> > > When pe is NULL in the function perf_pmu__new_alias(), the total
> > > number of events is added to loaded_json_aliases. However, if
> > > pmu->events_table is NULL and cpu_aliases_added is false, the
> > > calculation for the events number in perf_pmu__num_events() is incorrect.
> > >
> > > Then cause the error report after "perf list":
> > > Unexpected event
> > smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
> > >
> > > Fix it by adding loaded_json_aliases in the calculation under the
> > > mentioned conditions.
> > >
> > > Test it also with "perf bench internals pmu-scan" and there is no
> > > regression.
> > >
> > > Fixes: e6ff1eed3584 ("perf pmu: Lazily add JSON events")
> > > Signed-off-by: Jia He <[email protected]>
> > > ---
> > > tools/perf/util/pmu.c | 2 ++
> > > 1 file changed, 2 insertions(+)
> > >
> > > diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index
> > > a1eef7b2e389..a53224e2ce7e 100644
> > > --- a/tools/perf/util/pmu.c
> > > +++ b/tools/perf/util/pmu.c
> > > @@ -1639,6 +1639,8 @@ size_t perf_pmu__num_events(struct perf_pmu
> > *pmu)
> > > nr += pmu->loaded_json_aliases;
> > > else if (pmu->events_table)
> > > nr +=
> > pmu_events_table__num_events(pmu->events_table,
> > > pmu) - pmu->loaded_json_aliases;
> > > + else
> > > + nr += pmu->loaded_json_aliases;
> >
> > Thanks for working on this! The "struct pmu_event *pe" in new_alias is an entry
> > from the json data, and "pmu->events_table" should NULL if there is no json
> > data. I believe the code is assuming that these lines aren't necessary as it
> > shouldn't be possible to load json data if the json events table doesn't exist for
> > the PMU - if there is no json data then loaded_json_aliases should be 0 and no
> > addition is necessary. I'm wondering why this case isn't true for you.
> On my Armv8a N2 server, "pmu->events_table" is NULL because perf_pmu__find_events_table()
> return NULL.
>
> I also noticed that pmu->loaded_json_aliases is *not* 0. The missing adding calculation will cause
> perf_pmu__num_events() less than normal case and will trigger latter check failure in
> perf_pmus__print_pmu_events__callback().
> At last, perf list will report many lines similar as:
> Unexpected event smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
The issue is the sys events which currently are ARM only. Here is an
lldb stack trace:
```
frame #6: 0x0000aaaaabc9b49c
perf`__assert_fail(assertion="pmu->events_table",
file="tools/perf/util/pmu.c", line=522, function="int
perf_pmu__new_alias(struct perf_pmu *, const char *, const char *,
const char *, FILE *, const struct pmu_event *)")
frame #7: 0x0000aaaaab41e018
perf`perf_pmu__new_alias(pmu=0x00007377e7e09b40,
name="hnf_brd_snoops_sent", desc="Counts number of multicast snoops
sent (not including SF back invalidation)", val="eventid=9,type=5",
val_fd=0x0000000000000000, pe=0x0000ffffffffde88) at pmu.c:522:3
frame #8: 0x0000aaaaab4175cc
perf`pmu_add_sys_aliases_iter_fn(pe=0x0000ffffffffde88,
table=0x0000aaaaac299bb0, vdata=0x00007377e7e09b40) at pmu.c:939:3
frame #9: 0x0000aaaaaaf6d000
perf`pmu_events_table__for_each_event_pmu(table=0x0000aaaaac299bb0,
pmu=0x0000aaaaac299238, fn=(perf`pmu_add_sys_aliases_iter_fn at
pmu.c:931), data=0x00007377e7e09b40) at pmu-events.c:5994:23
frame #10: 0x0000aaaaaaf6cd78
perf`pmu_events_table__for_each_event(table=0x0000aaaaac299bb0,
pmu=0x0000000000000000, fn=(perf`pmu_add_sys_aliases_iter_fn at
pmu.c:931), data=0x00007377e7e09b40) at pmu-events.c:6057:23
frame #11: 0x0000aaaaaaf6ec44
perf`pmu_for_each_sys_event(fn=(perf`pmu_add_sys_aliases_iter_fn at
pmu.c:931), data=0x00007377e7e09b40) at pmu-events.c:6295:27
frame #12: 0x0000aaaaab41738c
perf`pmu_add_sys_aliases(pmu=0x00007377e7e09b40) at pmu.c:955:2
frame #13: 0x0000aaaaab4179fc
perf`perf_pmu__lookup(pmus=0x0000aaaaac500970, dirfd=34,
name="arm_cmn_0") at pmu.c:1037:2
frame #14: 0x0000aaaaab4242d0 perf`perf_pmu__find2(dirfd=34,
name="arm_cmn_0") at pmus.c:161:9
frame #15: 0x0000aaaaab421614 perf`pmu_read_sysfs(core_only=false)
at pmus.c:209:3
frame #16: 0x0000aaaaab42278c
perf`perf_pmus__scan_skip_duplicates(pmu=0x0000000000000000) at
pmus.c:297:3
frame #17: 0x0000aaaaab422078
perf`perf_pmus__print_pmu_events(print_cb=0x0000ffffffffec68,
print_state=0x0000ffffffffecd0) at pmus.c:462:16
frame #18: 0x0000aaaaab425d9c
perf`print_events(print_cb=0x0000ffffffffec68,
print_state=0x0000ffffffffecd0) at print-events.c:409:2
frame #19: 0x0000aaaaab093ef0 perf`cmd_list(argc=0,
argv=0x0000fffffffff340) at builtin-list.c:592:3
frame #20: 0x0000aaaaaaf5b490
perf`run_builtin(p=0x0000aaaaac4e8220, argc=1,
argv=0x0000fffffffff340) at perf.c:349:11
frame #21: 0x0000aaaaaaf5a3e0 perf`handle_internal_command(argc=1,
argv=0x0000fffffffff340) at perf.c:402:8
frame #22: 0x0000aaaaaaf5b1a0
perf`run_argv(argcp=0x0000fffffffff1c8, argv=0x0000fffffffff1c0) at
perf.c:446:2
frame #23: 0x0000aaaaaaf59f44 perf`main(argc=1,
argv=0x0000fffffffff340) at perf.c:562:3
```
As such the fix here is incomplete. There may be both sys json events
(detected by PMU/id name) and CPU json events (detected by CPUID). I'm
looking into a fix.
Thanks,
Ian
> --
> Cheers,
> Justin (Jia He)
On Fri, May 10, 2024 at 1:41 PM Ian Rogers <[email protected]> wrote:
>
> On Thu, May 9, 2024 at 11:52 PM Justin He <[email protected]> wrote:
> >
> > Hi, Ian
> >
> > > -----Original Message-----
> > > From: Ian Rogers <[email protected]>
> > > Sent: Friday, May 10, 2024 2:17 PM
> > > To: Justin He <[email protected]>
> > > Cc: Peter Zijlstra <[email protected]>; Ingo Molnar <[email protected]>;
> > > Arnaldo Carvalho de Melo <[email protected]>; Namhyung Kim
> > > <[email protected]>; Mark Rutland <[email protected]>; Alexander
> > > Shishkin <[email protected]>; Jiri Olsa <[email protected]>;
> > > Adrian Hunter <[email protected]>; Kan Liang
> > > <[email protected]>; James Clark <[email protected]>;
> > > [email protected]; [email protected]
> > > Subject: Re: [PATCH 2/2] perf pmu: Fix num_events calculation
> > >
> > > On Thu, May 9, 2024 at 7:47 PM Jia He <[email protected]> wrote:
> > > >
> > > > When pe is NULL in the function perf_pmu__new_alias(), the total
> > > > number of events is added to loaded_json_aliases. However, if
> > > > pmu->events_table is NULL and cpu_aliases_added is false, the
> > > > calculation for the events number in perf_pmu__num_events() is incorrect.
> > > >
> > > > Then cause the error report after "perf list":
> > > > Unexpected event
> > > smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
> > > >
> > > > Fix it by adding loaded_json_aliases in the calculation under the
> > > > mentioned conditions.
> > > >
> > > > Test it also with "perf bench internals pmu-scan" and there is no
> > > > regression.
> > > >
> > > > Fixes: e6ff1eed3584 ("perf pmu: Lazily add JSON events")
> > > > Signed-off-by: Jia He <[email protected]>
> > > > ---
> > > > tools/perf/util/pmu.c | 2 ++
> > > > 1 file changed, 2 insertions(+)
> > > >
> > > > diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c index
> > > > a1eef7b2e389..a53224e2ce7e 100644
> > > > --- a/tools/perf/util/pmu.c
> > > > +++ b/tools/perf/util/pmu.c
> > > > @@ -1639,6 +1639,8 @@ size_t perf_pmu__num_events(struct perf_pmu
> > > *pmu)
> > > > nr += pmu->loaded_json_aliases;
> > > > else if (pmu->events_table)
> > > > nr +=
> > > pmu_events_table__num_events(pmu->events_table,
> > > > pmu) - pmu->loaded_json_aliases;
> > > > + else
> > > > + nr += pmu->loaded_json_aliases;
> > >
> > > Thanks for working on this! The "struct pmu_event *pe" in new_alias is an entry
> > > from the json data, and "pmu->events_table" should NULL if there is no json
> > > data. I believe the code is assuming that these lines aren't necessary as it
> > > shouldn't be possible to load json data if the json events table doesn't exist for
> > > the PMU - if there is no json data then loaded_json_aliases should be 0 and no
> > > addition is necessary. I'm wondering why this case isn't true for you.
> > On my Armv8a N2 server, "pmu->events_table" is NULL because perf_pmu__find_events_table()
> > return NULL.
> >
> > I also noticed that pmu->loaded_json_aliases is *not* 0. The missing adding calculation will cause
> > perf_pmu__num_events() less than normal case and will trigger latter check failure in
> > perf_pmus__print_pmu_events__callback().
> > At last, perf list will report many lines similar as:
> > Unexpected event smmuv3_pmcg_3f062/smmuv3_pmcg_3f062/transaction//
>
> The issue is the sys events which currently are ARM only. Here is an
> lldb stack trace:
> ```
> frame #6: 0x0000aaaaabc9b49c
> perf`__assert_fail(assertion="pmu->events_table",
> file="tools/perf/util/pmu.c", line=522, function="int
> perf_pmu__new_alias(struct perf_pmu *, const char *, const char *,
> const char *, FILE *, const struct pmu_event *)")
> frame #7: 0x0000aaaaab41e018
> perf`perf_pmu__new_alias(pmu=0x00007377e7e09b40,
> name="hnf_brd_snoops_sent", desc="Counts number of multicast snoops
> sent (not including SF back invalidation)", val="eventid=9,type=5",
> val_fd=0x0000000000000000, pe=0x0000ffffffffde88) at pmu.c:522:3
> frame #8: 0x0000aaaaab4175cc
> perf`pmu_add_sys_aliases_iter_fn(pe=0x0000ffffffffde88,
> table=0x0000aaaaac299bb0, vdata=0x00007377e7e09b40) at pmu.c:939:3
> frame #9: 0x0000aaaaaaf6d000
> perf`pmu_events_table__for_each_event_pmu(table=0x0000aaaaac299bb0,
> pmu=0x0000aaaaac299238, fn=(perf`pmu_add_sys_aliases_iter_fn at
> pmu.c:931), data=0x00007377e7e09b40) at pmu-events.c:5994:23
> frame #10: 0x0000aaaaaaf6cd78
> perf`pmu_events_table__for_each_event(table=0x0000aaaaac299bb0,
> pmu=0x0000000000000000, fn=(perf`pmu_add_sys_aliases_iter_fn at
> pmu.c:931), data=0x00007377e7e09b40) at pmu-events.c:6057:23
> frame #11: 0x0000aaaaaaf6ec44
> perf`pmu_for_each_sys_event(fn=(perf`pmu_add_sys_aliases_iter_fn at
> pmu.c:931), data=0x00007377e7e09b40) at pmu-events.c:6295:27
> frame #12: 0x0000aaaaab41738c
> perf`pmu_add_sys_aliases(pmu=0x00007377e7e09b40) at pmu.c:955:2
> frame #13: 0x0000aaaaab4179fc
> perf`perf_pmu__lookup(pmus=0x0000aaaaac500970, dirfd=34,
> name="arm_cmn_0") at pmu.c:1037:2
> frame #14: 0x0000aaaaab4242d0 perf`perf_pmu__find2(dirfd=34,
> name="arm_cmn_0") at pmus.c:161:9
> frame #15: 0x0000aaaaab421614 perf`pmu_read_sysfs(core_only=false)
> at pmus.c:209:3
> frame #16: 0x0000aaaaab42278c
> perf`perf_pmus__scan_skip_duplicates(pmu=0x0000000000000000) at
> pmus.c:297:3
> frame #17: 0x0000aaaaab422078
> perf`perf_pmus__print_pmu_events(print_cb=0x0000ffffffffec68,
> print_state=0x0000ffffffffecd0) at pmus.c:462:16
> frame #18: 0x0000aaaaab425d9c
> perf`print_events(print_cb=0x0000ffffffffec68,
> print_state=0x0000ffffffffecd0) at print-events.c:409:2
> frame #19: 0x0000aaaaab093ef0 perf`cmd_list(argc=0,
> argv=0x0000fffffffff340) at builtin-list.c:592:3
> frame #20: 0x0000aaaaaaf5b490
> perf`run_builtin(p=0x0000aaaaac4e8220, argc=1,
> argv=0x0000fffffffff340) at perf.c:349:11
> frame #21: 0x0000aaaaaaf5a3e0 perf`handle_internal_command(argc=1,
> argv=0x0000fffffffff340) at perf.c:402:8
> frame #22: 0x0000aaaaaaf5b1a0
> perf`run_argv(argcp=0x0000fffffffff1c8, argv=0x0000fffffffff1c0) at
> perf.c:446:2
> frame #23: 0x0000aaaaaaf59f44 perf`main(argc=1,
> argv=0x0000fffffffff340) at perf.c:562:3
> ```
>
> As such the fix here is incomplete. There may be both sys json events
> (detected by PMU/id name) and CPU json events (detected by CPUID). I'm
> looking into a fix.
Patch sent:
https://lore.kernel.org/lkml/[email protected]/
Thanks,
Ian
> Thanks,
> Ian
>
> > --
> > Cheers,
> > Justin (Jia He)
On Fri, May 10, 2024 at 05:50:29PM -0700, Ian Rogers wrote:
> On Fri, May 10, 2024 at 1:41 PM Ian Rogers <[email protected]> wrote:
> > As such the fix here is incomplete. There may be both sys json events
> > (detected by PMU/id name) and CPU json events (detected by CPUID). I'm
> > looking into a fix.
>
> Patch sent:
> https://lore.kernel.org/lkml/[email protected]/
I applied the patch to tmp.perf-tools-next, will go to perf-tools-next
soon, would be good for Justin to test and provide a Tested-by.
- Arnaldo