perf stat: fix per-pkg event reporting bug
PATCH 1/1 perf stat: fix per-pkg event reporting bug
Per-pkg events need to be captured once per processor socket. The code in check_per_pkg() ensures only one value per processor package is used. However there is a problem with this function in case the first CPU of the package does not measure anything for the per-pkg event, but other CPUs do.
Consider the following
$ create cgroup FOO; echo $$ >FOO/tasks; taskset -c 1 noploop & $ perf stat -a -I 1000 -e intel_cqm/llc_occupancy/ -G FOO sleep 100 1.00000 <not counted> Bytes intel_cqm/llc_occupancy/ FOO
The reason for this is that CPU0 in the cgrop has nothing running on it. Yet check_per_plg() will mark socket0 as processed and no other event value will be considered for the socket.
This patch fixes the problem by having check_per_pkg() only consider events which actually ran.
Patch is relative to tip.git.
--- tools/perf/util/stat.c | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)
tools/perf/util/stat.c +14 -2
@@ -196,7 +196,8 @@ static void zero_per_pkg(struct perf_evsel *counter)memset (counter->per_pkg_mask,0 ,MAX_NR_CPUS ); } -static int check_per_pkg (struct perf_evsel *counter,int cpu,bool *skip) +static int check_per_pkg (struct perf_evsel *counter, +struct perf_counts_values *vals,int cpu,bool *skip) {unsigned long *mask = counter->per_pkg_mask;struct cpu_map *cpus =perf_evsel__cpus (counter);
@@ -218,6 +219,17 @@ static int check_per_pkg(struct perf_evsel *counter, int cpu, bool *skip) counter->per_pkg_mask = mask; } +/* +* we do not consider an event that has not run as a good +* instance to mark a package as used (skip=1). Otherwise +* we may run into a situation where the first CPU in a package +* is not running anything, yet the second is, and this function +* would mark the package as used after the first CPU and would +* not read the values from the second CPU. +*/ +if (!(vals->run && vals->ena)) +return 0 ; + s =cpu_map__get_socket (cpus, cpu);if (s <0 )return -1 ;
@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false; -if (check_per_pkg (evsel, cpu, &skip)) { +if (check_per_pkg (evsel, aggr, cpu, &skip)) {pr_err ("failed to read per-pkg counter\n" );return -1 ; }
Signature · 1 line
1.9.1
Quoted 6 lines · Stephane Eranian
Per-pkg events need to be captured once per processor socket. The code in
check_per_pkg()ensures only one value per processor package is used. However there is a problem with this function in case the first CPU of the package does not measure anything for the per-pkg event, but other CPUs do.
I've seen a similar(?) bug with -C and --per-core combined. Some logical threads are not correctly accounted there either. Do you think that's another instance of such a bug?
-Andi
Quoted 11 lines · Andi Kleen
Quoted 6 lines · Stephane Eranian
Per-pkg events need to be captured once per processor socket. The code in
check_per_pkg()ensures only one value per processor package is used. However there is a problem with this function in case the first CPU of the package does not measure anything for the per-pkg event, but other CPUs do.I've seen a similar(?) bug with -C and --per-core combined. Some logical threads are not correctly accounted there either. Do you think that's another instance of such a bug?
AFAICS this is related strictly to per-pkg events, is this what you were doing?
jirka
On Wed, Sep 02, 2015 at 03:17:51PM +0200, Stephane Eranian wrote:
SNIP
Quoted 15 lines
tools/perf/util/stat.c +11
-0+/* +* we do not consider an event that has not run as a good +* instance to mark a package as used (skip=1). Otherwise +* we may run into a situation where the first CPU in a package +* is not running anything, yet the second is, and this function +* would mark the package as used after the first CPU and would +* not read the values from the second CPU. +*/ +if (!(vals->run && vals->ena)) +return 0 ; + s =cpu_map__get_socket (cpus, cpu);if (s <0 )return -1 ;
tools/perf/util/stat.c process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel · line 250
Quoted 5 lines
tools/perf/util/stat.c +1
-1@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false; -if (check_per_pkg (evsel, cpu, &skip)) { +if (check_per_pkg (evsel, aggr, cpu, &skip)) {
should we pass 'count' instead o 'aggr' ?
jirka
Quoted 22 lines · Jiri Olsa
On Wed, Sep 02, 2015 at 03:17:51PM +0200, Stephane Eranian wrote:
SNIP
Quoted 18 lines
tools/perf/util/stat.c +11
-0+/* +* we do not consider an event that has not run as a good +* instance to mark a package as used (skip=1). Otherwise +* we may run into a situation where the first CPU in a package +* is not running anything, yet the second is, and this function +* would mark the package as used after the first CPU and would +* not read the values from the second CPU. +*/ +if (!(vals->run && vals->ena)) +return 0 ; + s =cpu_map__get_socket (cpus, cpu);if (s <0 )return -1 ;@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false;
Quoted 5 lines
tools/perf/util/stat.c process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel · line 250
Quoted 2 lines
tools/perf/util/stat.c +1
-1-if (check_per_pkg (evsel, cpu, &skip)) { +if (check_per_pkg (evsel, aggr, cpu, &skip)) {should we pass 'count' instead o 'aggr' ?
the reason I passed counts_values is in case this function needs to be called from other places which do
not use aggr mode.
Quoted 23 lines · Stephane Eranian
Quoted 22 lines · Jiri Olsa
On Wed, Sep 02, 2015 at 03:17:51PM +0200, Stephane Eranian wrote:
SNIP
Quoted 18 lines
tools/perf/util/stat.c +11
-0+/* +* we do not consider an event that has not run as a good +* instance to mark a package as used (skip=1). Otherwise +* we may run into a situation where the first CPU in a package +* is not running anything, yet the second is, and this function +* would mark the package as used after the first CPU and would +* not read the values from the second CPU. +*/ +if (!(vals->run && vals->ena)) +return 0 ; + s =cpu_map__get_socket (cpus, cpu);if (s <0 )return -1 ;@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false;
Quoted 8 lines
Quoted 5 lines
tools/perf/util/stat.c process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel · line 250
Quoted 2 lines
tools/perf/util/stat.c +1
-1-if (check_per_pkg (evsel, cpu, &skip)) { +if (check_per_pkg (evsel, aggr, cpu, &skip)) {should we pass 'count' instead o 'aggr' ?
the reason I passed
counts_valuesis in case this function needs to be called from other places which do
not use aggr mode.
sure, but 'aggr' is being computed within process_counter_values
process_counter_values gets 'count' argument with values read for given cpu/thread for further processing, and it seems to me that 'count' values should be passed to check_per_pkg
jirka
Quoted 33 lines · Jiri Olsa
Quoted 23 lines · Stephane Eranian
Quoted 22 lines · Jiri Olsa
On Wed, Sep 02, 2015 at 03:17:51PM +0200, Stephane Eranian wrote:
SNIP
Quoted 18 lines
tools/perf/util/stat.c +11
-0+/* +* we do not consider an event that has not run as a good +* instance to mark a package as used (skip=1). Otherwise +* we may run into a situation where the first CPU in a package +* is not running anything, yet the second is, and this function +* would mark the package as used after the first CPU and would +* not read the values from the second CPU. +*/ +if (!(vals->run && vals->ena)) +return 0 ; + s =cpu_map__get_socket (cpus, cpu);if (s <0 )return -1 ;@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false;Quoted 8 lines
Quoted 5 lines
tools/perf/util/stat.c process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel · line 250
Quoted 2 lines
tools/perf/util/stat.c +1
-1-if (check_per_pkg (evsel, cpu, &skip)) { +if (check_per_pkg (evsel, aggr, cpu, &skip)) {should we pass 'count' instead o 'aggr' ?
the reason I passed
counts_valuesis in case this function needs to be called from other places which do
not use aggr mode.
Quoted 6 lines
sure, but 'aggr' is being computed within
process_counter_values
process_counter_valuesgets 'count' argument with values read for given cpu/thread for further processing, and it seems to me that 'count' values should be passed tocheck_per_pkg
You do not want to aggregate values, you want to look at the individual events for each CPU because you need to look at their run/ena fields.
Quoted 34 lines · Stephane Eranian
Quoted 33 lines · Jiri Olsa
Quoted 31 lines · Stephane Eranian
Quoted 22 lines · Jiri Olsa
On Wed, Sep 02, 2015 at 03:17:51PM +0200, Stephane Eranian wrote:
SNIP
Quoted 18 lines
(quoted diff) +11
-0+ /* + * we do not consider an event that has not run as a good + * instance to mark a package as used (skip=1). Otherwise + * we may run into a situation where the first CPU in a package + * is not running anything, yet the second is, and this function + * would mark the package as used after the first CPU and would + * not read the values from the second CPU. + */ + if (!(vals->run && vals->ena)) + return 0; + s = cpu_map__get_socket(cpus, cpu); if (s < 0) return -1;@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel static struct perf_counts_values zero; bool skip = false;Quoted 5 lines
Review process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel · line 250
Quoted 2 lines
(quoted diff) +1
-1- if (check_per_pkg(evsel, cpu, &skip)) { + if (check_per_pkg(evsel, aggr, cpu, &skip)) {should we pass 'count' instead o 'aggr' ?
the reason I passed
counts_valuesis in case this function needs to be called from other places which do
not use aggr mode.
Quoted 8 lines
Quoted 6 lines
sure, but 'aggr' is being computed within
process_counter_values
process_counter_valuesgets 'count' argument with values read for given cpu/thread for further processing, and it seems to me that 'count' values should be passed tocheck_per_pkgYou do not want to aggregate values, you want to look at the individual events for each CPU because you need to look at their run/ena fields.
yes, but for 'count' not 'aggr'
jirka
---
tools/perf/util/stat.c +1 -1
@@ -247,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false; -if (check_per_pkg (evsel, aggr, cpu, &skip)) { +if (check_per_pkg (evsel, count, cpu, &skip)) {pr_err ("failed to read per-pkg counter\n" );return -1 ; }
Quoted 37 lines · Jiri Olsa
Quoted 36 lines · Stephane Eranian
Quoted 35 lines · Jiri Olsa
Quoted 23 lines · Stephane Eranian
Quoted 22 lines · Jiri Olsa
On Wed, Sep 02, 2015 at 03:17:51PM +0200, Stephane Eranian wrote:
SNIP
Quoted 18 lines
tools/perf/util/stat.c +11
-0+/* +* we do not consider an event that has not run as a good +* instance to mark a package as used (skip=1). Otherwise +* we may run into a situation where the first CPU in a package +* is not running anything, yet the second is, and this function +* would mark the package as used after the first CPU and would +* not read the values from the second CPU. +*/ +if (!(vals->run && vals->ena)) +return 0 ; + s =cpu_map__get_socket (cpus, cpu);if (s <0 )return -1 ;@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false;Quoted 8 lines
Quoted 5 lines
Review process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel · line 250
Quoted 2 lines
(quoted diff) +1
-1- if (check_per_pkg(evsel, cpu, &skip)) { + if (check_per_pkg(evsel, aggr, cpu, &skip)) {should we pass 'count' instead o 'aggr' ?
the reason I passed
counts_valuesis in case this function needs to be called from other places which do
not use aggr mode.sure, but 'aggr' is being computed within
process_counter_values
Quoted 9 lines
Quoted 6 lines
Quoted 4 lines
process_counter_valuesgets 'count' argument with values read for given cpu/thread for further processing, and it seems to me that 'count' values should be passed tocheck_per_pkgYou do not want to aggregate values, you want to look at the individual events for each CPU because you need to look at their run/ena fields.
yes, but for 'count' not 'aggr'
Ah, yes, sorry, has to be count and not aggr. Sent the wrong version. Can you fix it? Or do you want me to resubmit?
Quoted 17 lines
jirka
---
tools/perf/util/stat.c +1
-1@@ -247,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false; -if (check_per_pkg (evsel, aggr, cpu, &skip)) { +if (check_per_pkg (evsel, count, cpu, &skip)) {pr_err ("failed to read per-pkg counter\n" );return -1 ; }
Quoted 42 lines · Stephane Eranian
Quoted 41 lines · Jiri Olsa
Quoted 40 lines · Stephane Eranian
Quoted 39 lines · Jiri Olsa
Quoted 23 lines · Stephane Eranian
Quoted 22 lines · Jiri Olsa
On Wed, Sep 02, 2015 at 03:17:51PM +0200, Stephane Eranian wrote:
SNIP
Quoted 18 lines
tools/perf/util/stat.c +11
-0+/* +* we do not consider an event that has not run as a good +* instance to mark a package as used (skip=1). Otherwise +* we may run into a situation where the first CPU in a package +* is not running anything, yet the second is, and this function +* would mark the package as used after the first CPU and would +* not read the values from the second CPU. +*/ +if (!(vals->run && vals->ena)) +return 0 ; + s =cpu_map__get_socket (cpus, cpu);if (s <0 )return -1 ;@@ -235,7 +247,7 @@ process_counter_values(struct perf_stat_config *config, struct perf_evsel *evselstatic struct perf_counts_values zero;bool skip = false;Quoted 8 lines
Quoted 5 lines
Review process_counter_values(struct perf_stat_config *config, struct perf_evsel *evsel · line 250
Quoted 2 lines
(quoted diff) +1
-1- if (check_per_pkg(evsel, cpu, &skip)) { + if (check_per_pkg(evsel, aggr, cpu, &skip)) {should we pass 'count' instead o 'aggr' ?
the reason I passed
counts_valuesis in case this function needs to be called from other places which do
not use aggr mode.sure, but 'aggr' is being computed within
process_counter_values
process_counter_valuesgets 'count' argument with values read for given cpu/thread for further processing, and it seems to me that 'count' values should be passed tocheck_per_pkg
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines
You do not want to aggregate values, you want to look at the individual events for each CPU because you need to look at their run/ena fields.
yes, but for 'count' not 'aggr'
Ah, yes, sorry, has to be count and not aggr. Sent the wrong version. Can you fix it? Or do you want me to resubmit?
well, Arnaldo will queue it.. leaving up to him ;-)
jirka
Em Thu, Sep 03, 2015 at 02:25:44PM +0200, Jiri Olsa escreveu:
Quoted 8 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa
Quoted 1 line
yes, but for 'count' not 'aggr'
Ah, yes, sorry, has to be count and not aggr. Sent the wrong version. Can you fix it? Or do you want me to resubmit?
well, Arnaldo will queue it.. leaving up to him ;-)
Please resubmit, with a [PATCH v2 ...] in it, and with a v2 right before your Signed-off-by: stating what you changed, that helps when I see multiple patches, i.e. you document what was changed and I don't have to follow that many threads :-)
- Arnaldo
Quoted 8 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 02:25:44PM +0200, Jiri Olsa escreveu:
Quoted 7 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa
Quoted 1 line
yes, but for 'count' not 'aggr'
Ah, yes, sorry, has to be count and not aggr. Sent the wrong version. Can you fix it? Or do you want me to resubmit?
Quoted 7 lines
Quoted 1 line
well, Arnaldo will queue it.. leaving up to him ;-)
Please resubmit, with a [PATCH v2 ...] in it, and with a v2 right before your Signed-off-by: stating what you changed, that helps when I see multiple patches, i.e. you document what was changed and I don't have to follow that many threads :-)
I already sent the V2, but I forgot to state the small change.
Do you want a V3? Or you add: fix perf_counter_value argument to check_per_pkg. Thanks.
Em Thu, Sep 03, 2015 at 10:12:12AM -0700, Stephane Eranian escreveu:
Quoted 12 lines
Quoted 10 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 02:25:44PM +0200, Jiri Olsa escreveu:
Quoted 8 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa
Quoted 1 line
yes, but for 'count' not 'aggr'
Ah, yes, sorry, has to be count and not aggr. Sent the wrong version. Can you fix it? Or do you want me to resubmit?
well, Arnaldo will queue it.. leaving up to him ;-)
Quoted 8 lines
Quoted 5 lines
Please resubmit, with a [PATCH v2 ...] in it, and with a v2 right before your Signed-off-by: stating what you changed, that helps when I see multiple patches, i.e. you document what was changed and I don't have to follow that many threads :-)
I already sent the V2, but I forgot to state the small change.
Do you want a V3? Or you add: fixperf_counter_valueargument tocheck_per_pkg. Thanks.
Ok, I can do it...
- Arnaldo
Quoted 18 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 10:12:12AM -0700, Stephane Eranian escreveu:
Quoted 17 lines
Quoted 15 lines · Arnaldo Carvalho de Melo
Em Thu, Sep 03, 2015 at 02:25:44PM +0200, Jiri Olsa escreveu:
Quoted 8 lines · Stephane Eranian
Quoted 5 lines · Jiri Olsa
Quoted 1 line
yes, but for 'count' not 'aggr'
Ah, yes, sorry, has to be count and not aggr. Sent the wrong version. Can you fix it? Or do you want me to resubmit?
well, Arnaldo will queue it.. leaving up to him ;-)
Please resubmit, with a [PATCH v2 ...] in it, and with a v2 right before your Signed-off-by: stating what you changed, that helps when I see multiple patches, i.e. you document what was changed and I don't have to follow that many threads :-)
Quoted 6 lines
Quoted 3 lines
I already sent the V2, but I forgot to state the small change.
Do you want a V3? Or you add: fixperf_counter_valueargument tocheck_per_pkg. Thanks.Ok, I can do it...
Thanks.
Quoted 1 line
- Arnaldo