perf,tools: get correct cpu id for print_aggr
PATCH 1/2 perf,tools: get correct cpu id for print_aggr
From: Kan Liang <[email protected]>
print_aggr fails to print per-core/per-socket statistics after commit b7f0c203586b ("perf evlist: Propagate cpu maps to evsels in an evlist"), if events have differnt cpus. Because in print_aggr, aggr_get_id needs index (not cpu id) to find core/pkg id.
This patch introduced perf_evsel__get_cpumap_index to get the index by cpu id for a given event. The index can be used to find correct cpu id for print_aggr.
Here is an example.
Counting events cycles,uncore_imc_0/cas_count_read/. (Uncore event has cpumask 0,18)
"perf stat -e cycles,uncore_imc_0/cas_count_read/ -C0,18 --per-core sleep 2"
Without this patch, it failes to get CPU 18 result.
Performance counter stats for 'CPU(s) 0,18':
S0-C0 1 7526851 cycles S0-C0 1 1.05 MiB uncore_imc_0/cas_count_read/ S1-C0 0 <not counted> cycles S1-C0 0 <not counted> MiB uncore_imc_0/cas_count_read/
With this patch, it can get both CPU0 and CPU18 result.
Performance counter stats for 'CPU(s) 0,18':
S0-C0 1 6327768 cycles S0-C0 1 0.47 MiB uncore_imc_0/cas_count_read/ S1-C0 1 330228 cycles S1-C0 1 0.29 MiB uncore_imc_0/cas_count_read/
--- tools/perf/builtin-stat.c | 1 + tools/perf/util/cpumap.c | 4 ++-- tools/perf/util/evsel.c | 14 ++++++++++++++ tools/perf/util/evsel.h | 2 ++ 4 files changed, 19 insertions(+), 2 deletions(-)
tools/perf/builtin-stat.c +1 -0
@@ -708,6 +708,7 @@ static void print_aggr(char *prefix) nr =0 ;for (cpu =0 ; cpu <perf_evsel__nr_cpus (counter); cpu++) { cpu2 =perf_evsel__cpus (counter)->map[cpu]; + cpu2 =perf_evsel__get_cpumap_index (cpu2, evsel_list->cpus); s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue ;
tools/perf/util/cpumap.c +2 -2
@@ -232,7 +232,7 @@ int cpu_map__get_socket(struct cpu_map *map, int idx)char path[PATH_MAX ];int cpu, ret; -if (idx > map->nr) +if ((idx > map->nr) || (idx <0 ))return -1 ; cpu = map->map[idx];
@@ -296,7 +296,7 @@ int cpu_map__get_core(struct cpu_map *map, int idx)char path[PATH_MAX ];int cpu, ret, s; -if (idx > map->nr) +if ((idx > map->nr) || (idx <0 ))return -1 ; cpu = map->map[idx];
tools/perf/util/evsel.c +14 -0
@@ -974,6 +974,20 @@ int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,return 0 ; } +int perf_evsel__get_cpumap_index (int cpu,struct cpu_map *evsel_cpus) +{ +int i; + +if (evsel_cpus ==NULL ) +return -1 ; + +for (i =0 ; i < evsel_cpus->nr; i++) { +if (cpu == evsel_cpus->map[i]) +return i; + } +return -1 ; +} +static int get_group_fd (struct perf_evsel *evsel,int cpu,int thread) {struct perf_evsel *leader = evsel->leader;
tools/perf/util/evsel.h +2 -0
@@ -355,4 +355,6 @@ typedef int (*attr__fprintf_f)(FILE *, const char *, const char *, void *);int perf_event_attr__fprintf (FILE *fp,struct perf_event_attr *attr,attr__fprintf_f attr__fprintf,void *priv); +int perf_evsel__get_cpumap_index (int cpu,struct cpu_map *evsel_cpus); +#endif /* __PERF_EVSEL_H */
Signature · 1 line
1.8.3.1
PATCH 2/2 perf,tools: check and re-organize evsel cpu maps
From: Kan Liang <[email protected]>
Some PMU events have cpumask, e.g uncore events. The cpu list set by user may be incompatible with event's cpumask.
This patch will check the user defined cpu list. If the incompatible cpu is found, it will warn the user and discard the incompatible cpu. Only available cpu can be stored in evsel->cpus->map. If there is no cpu from cpu list compatible with event's cpumask. It will error out.
Here is an example.
According to cpumask, uncore should only available on CPU0 and CPU18. So the S0-C1 for uncore should not count.
Without this patch
$ sudo ./perf stat -e cycles,uncore_imc_0/cas_count_read/ -C0,1,18 --per-core sleep 2
Performance counter stats for 'CPU(s) 0,1,18':
S0-C0 1 6749638 cycles S0-C0 1 0.83 MiB uncore_imc_0/cas_count_read/ (100.00%) S0-C1 1 232421 cycles S0-C1 1 0.83 MiB uncore_imc_0/cas_count_read/ S1-C0 1 236997 cycles S1-C0 1 0.35 MiB uncore_imc_0/cas_count_read/
2.001094019 seconds time elapsed
With this patch
$ perf stat -e cycles,uncore_imc_0/cas_count_read/ -C0,1,18 --per-core sleep 2 event uncore_imc_0/cas_count_read/ can only be monitored on CPU 0 18. Other CPUs will be discard.
Performance counter stats for 'CPU(s) 0,1,18':
S0-C0 1 5557406 cycles S0-C0 1 0.21 MiB uncore_imc_0/cas_count_read/ S0-C1 1 1012534 cycles S0-C1 0 <not counted> MiB uncore_imc_0/cas_count_read/ S1-C0 1 916130 cycles S1-C0 1 0.08 MiB uncore_imc_0/cas_count_read/
2.001110843 seconds time elapsed
--- tools/perf/util/evlist.c | 77 ++++++++++++++++++++++++++++++++++++++++++++---- 1 file changed, 72 insertions(+), 5 deletions(-)
tools/perf/util/evlist.c +72 -5
@@ -1101,6 +1101,71 @@ int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,return perf_evlist__mmap_ex (evlist, pages, overwrite,0 , false); } +static int cmp_ids (const void *a,const void *b) +{ +return *(int *)a - *(int *)b; +} + +static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist,struct perf_evsel *evsel) +{ +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +int j =0 , cpu_nr =0 , tmp =0 ; +int i; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids); + +/* find the common cpus between evsel and evlist. */ +for (i =0 ; i <cpu_map__nr (evsel->cpus);) { + +if (j >= ncpus) { + evsel->cpus->map[i++] =-1 ; +continue ; + } +for (; j < ncpus; j++) { +if (cpus->map[j] < evsel->cpus->map[i]) +continue ; +if (cpus->map[j] == evsel->cpus->map[i]) { + cpu_nr++; + j++; + i++; + }else + evsel->cpus->map[i++] =-1 ; +break ; + } + } + +if (cpu_nr ==0 ) { +pr_warning ("event %s cannot be monitored on the given cpus." +"Please check cpumask\n" , evsel->name); +return -1 ; + } + +if (ncpus > cpu_nr) +pr_warning ("event %s can only be monitored on CPU" , evsel->name); + +/* order evsel cpus */ +for (i =0 ; i < cpu_nr; i++) { +if (evsel->cpus->map[i] ==-1 ) { +while (evsel->cpus->map[tmp] ==-1 ) { + tmp++; +BUG_ON (tmp >=cpu_map__nr (evsel->cpus)); + } + evsel->cpus->map[i] = evsel->cpus->map[tmp]; + evsel->cpus->map[tmp] =-1 ; + } +if (ncpus > cpu_nr) +pr_warning (" %d" , evsel->cpus->map[i]); + tmp++; + } + evsel->cpus->nr = cpu_nr; +if (ncpus > cpu_nr) +pr_warning (". Other CPUs will be discard.\n" ); + +return 0 ; +} +static int perf_evlist__propagate_maps (struct perf_evlist *evlist,struct target *target) {
@@ -1108,13 +1173,15 @@ static int perf_evlist__propagate_maps(struct perf_evlist *evlist,evlist__for_each (evlist, evsel) {/* -* We already have cpus for evsel (via PMU sysfs) so -* keep it, if there's no target cpu list defined. +* We already have cpus for evsel (via PMU sysfs) +* and target cpu list defined, check if they are +* compatible. If not, discard incompatible cpus. */ -if (evsel->cpus && target->cpu_list) -cpu_map__put (evsel->cpus); +if (evsel->cpus && target->cpu_list && +perf_evlist__check_evsel_cpus (evlist, evsel)) +return -EINVAL ; -if (!evsel->cpus || target->cpu_list) +if (!evsel->cpus) evsel->cpus =cpu_map__get (evlist->cpus); evsel->threads =thread_map__get (evlist->threads);
Signature · 1 line
1.8.3.1
Re: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
On Mon, Jun 29, 2015 at 03:55:35PM -0400, [email protected] wrote:
SNIP
Quoted 12 lines
tools/perf/util/evlist.c +4
-0@@ -1101,6 +1101,71 @@ int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,return perf_evlist__mmap_ex (evlist, pages, overwrite,0 , false); } +static int cmp_ids (const void *a,const void *b) +{ +return *(int *)a - *(int *)b; +}
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1117
Quoted 10 lines
tools/perf/util/evlist.c +10
-0+ +static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist,struct perf_evsel *evsel) +{ +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +int j =0 , cpu_nr =0 , tmp =0 ; +int i; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids);
wouldn't sorting maps affect some other code?
Quoted 4 lines
tools/perf/util/evlist.c +4
-0+ +/* find the common cpus between evsel and evlist. */ +for (i =0 ; i <cpu_map__nr (evsel->cpus);) { +
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1131
Quoted 10 lines
tools/perf/util/evlist.c +10
-0+if (j >= ncpus) { + evsel->cpus->map[i++] =-1 ; +continue ; + } +for (; j < ncpus; j++) { +if (cpus->map[j] < evsel->cpus->map[i]) +continue ; +if (cpus->map[j] == evsel->cpus->map[i]) { + cpu_nr++; + j++;
hum, do you skip 1 item in cpus by j++ here and in for loop header?
also some easy it'd be easier to read if you identify different cpu maps somehow.. like evlist_cpus and evsel_cpus or such
I have no way of testing this.. could you please add automated test for this one?
thanks,
jirka
Quoted 7 lines
tools/perf/util/evlist.c +7
-0+ i++; + }else + evsel->cpus->map[i++] =-1 ; +break ; + } + } +
SNIP
RE: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
Quoted 10 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist
Quoted 7 lines
tools/perf/util/evlist.c index renamed from tools/perf/util/evlist.c +0
-06cfdee6..f179379 100644
tools/perf/util/evlist.c +0
-0@@ -1101,6 +1101,71 @@ int perf_evlist__mmap(struct perf_evlist*evlist, unsigned int pages,
Quoted 2 lines
tools/perf/util/evlist.c +0
-0return perf_evlist__mmap_ex (evlist, pages, overwrite,0 , false); }
Quoted 16 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1117
Quoted 13 lines
tools/perf/util/evlist.c +13
-0+static int cmp_ids (const void *a,const void *b) { +return *(int *)a - *(int *)b; +} + +static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist, +struct perf_evsel *evsel) { +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +int j =0 , cpu_nr =0 , tmp =0 ; +int i; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids);wouldn't sorting maps affect some other code?
I didn't find any bad effect after sorting the maps. Any codes I need to check?
Quoted 8 lines
Quoted 8 lines
tools/perf/util/evlist.c +8
-0+ +/* find the common cpus between evsel and evlist. */ +for (i =0 ; i <cpu_map__nr (evsel->cpus);) { + +if (j >= ncpus) { + evsel->cpus->map[i++] =-1 ; +continue ; + }
Quoted 9 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1131
Quoted 6 lines
tools/perf/util/evlist.c +6
-0+for (; j < ncpus; j++) { +if (cpus->map[j] < evsel->cpus->map[i]) +continue ; +if (cpus->map[j] == evsel->cpus->map[i]) { + cpu_nr++; + j++;hum, do you skip 1 item in cpus by j++ here and in for loop header?
If matching, it will break the j loop then. There is no chance to j++ in the loop header. So I do j++ here, and it will not skip any items.
Quoted 3 lines
also some easy it'd be easier to read if you identify different cpu maps somehow.. like
evlist_cpusandevsel_cpusor such
OK. I will change the name.
Quoted 2 lines
I have no way of testing this.. could you please add automated test for this one?
OK
Thanks,
Kan
Quoted 13 lines
thanks,
jirkaQuoted 7 lines
tools/perf/util/evlist.c +7
-0+ i++; + }else + evsel->cpus->map[i++] =-1 ; +break ; + } + } +SNIP
Re: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
Quoted 27 lines · Kan
Quoted 10 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist
Quoted 7 lines
tools/perf/util/evlist.c index renamed from tools/perf/util/evlist.c +0
-06cfdee6..f179379 100644
tools/perf/util/evlist.c +0
-0@@ -1101,6 +1101,71 @@ int perf_evlist__mmap(struct perf_evlist*evlist, unsigned int pages,
Quoted 2 lines
tools/perf/util/evlist.c +0
-0return perf_evlist__mmap_ex (evlist, pages, overwrite,0 , false); }Quoted 16 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1117
Quoted 13 lines
tools/perf/util/evlist.c +13
-0+static int cmp_ids (const void *a,const void *b) { +return *(int *)a - *(int *)b; +} + +static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist, +struct perf_evsel *evsel) { +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +int j =0 , cpu_nr =0 , tmp =0 ; +int i; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids);wouldn't sorting maps affect some other code?
Quoted 2 lines
I didn't find any bad effect after sorting the maps. Any codes I need to check?
I dont think so, but I'm not sure either.. thats why I asked ;-)
I guess any code dealing with cpu maps.. it might affect perf stat output.. but it looks sorted now anyway ;-) I dont think it's an issue
jirka
[email protected]
maintainerRe: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
Em Tue, Jun 30, 2015 at 03:54:49PM +0200, Jiri Olsa escreveu:
Quoted 14 lines · Kan
Quoted 13 lines
Quoted 12 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1117
Quoted 9 lines
tools/perf/util/evlist.c +9
-0+static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist, +struct perf_evsel *evsel) { +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +int j =0 , cpu_nr =0 , tmp =0 ; +int i; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids);wouldn't sorting maps affect some other code?
Quoted 8 lines
Quoted 2 lines
I didn't find any bad effect after sorting the maps. Any codes I need to check?
I dont think so, but I'm not sure either.. thats why I asked ;-)
I guess any code dealing with cpu maps.. it might affect perf stat output.. but it looks sorted now anyway ;-) I dont think it's an issue
Is this being done at cpu_map__new time, i.e. when we first parse it?
That way any assumption about repositioning gets out of the way.
- Arnaldo
Re: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
Quoted 18 lines · On Tue, Jun 30, 2015 at 11:42:25AM -0300, [email protected] wrote:
Em Tue, Jun 30, 2015 at 03:54:49PM +0200, Jiri Olsa escreveu:
Quoted 17 lines · Kan
Quoted 15 lines
Quoted 12 lines
tools/perf/util/evlist.c int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages, · line 1117
Quoted 9 lines
tools/perf/util/evlist.c +9
-0+static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist, +struct perf_evsel *evsel) { +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +int j =0 , cpu_nr =0 , tmp =0 ; +int i; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids);wouldn't sorting maps affect some other code?
I didn't find any bad effect after sorting the maps. Any codes I need to check?
Quoted 7 lines
Quoted 5 lines
I dont think so, but I'm not sure either.. thats why I asked ;-)
I guess any code dealing with cpu maps.. it might affect perf stat output.. but it looks sorted now anyway ;-) I dont think it's an issue
Is this being done at
cpu_map__newtime, i.e. when we first parse it?
it's done within perf_evlist__create_maps.. checking the v2
jirka
Quoted 3 lines
That way any assumption about repositioning gets out of the way.
- Arnaldo
Re: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
In reply to [email protected]Quoted 10 lines · On Mon, Jun 29, 2015 at 12:55 PM, <[email protected]> wrote:
From: Kan Liang <[email protected]>
Some PMU events have cpumask, e.g uncore events. The cpu list set by user may be incompatible with event's cpumask.
This patch will check the user defined cpu list. If the incompatible cpu is found, it will warn the user and discard the incompatible cpu. Only available cpu can be stored inevsel->cpus->map. If there is no cpu from cpu list compatible with event's cpumask. It will error out.
Quoted 4 lines
Here is an example.
According to cpumask, uncore should only available on CPU0 and CPU18. So the S0-C1 for uncore should not count.
I don't think this is correct. The cpumask is a default set of CPUs to be used by
perf. The cpumask does not indicate the ONLY CPUs on which to monitor. It is just a default. You can monitor an uncore event using a CPU not listed in the cpumask, unless the kernel code has changed recently. If you are not on the default CPUs, the kernel will IPI.
Quoted 139 lines
Without this patch
$ sudo ./perf stat -e cycles,uncore_imc_0/cas_count_read/ -C0,1,18 --per-core sleep 2Performance counter stats for 'CPU(s) 0,1,18':
S0-C0 1 6749638 cycles S0-C0 1 0.83 MiB uncore_imc_0/cas_count_read/ (100.00%) S0-C1 1 232421 cycles S0-C1 1 0.83 MiB uncore_imc_0/cas_count_read/ S1-C0 1 236997 cycles S1-C0 1 0.35 MiB uncore_imc_0/cas_count_read/2.001094019 seconds time elapsedWith this patch
$ perf stat -e cycles,uncore_imc_0/cas_count_read/ -C0,1,18 --per-core sleep 2 event uncore_imc_0/cas_count_read/ can only be monitored on CPU 0 18. Other CPUs will be discard.Performance counter stats for 'CPU(s) 0,1,18':
S0-C0 1 5557406 cycles S0-C0 1 0.21 MiB uncore_imc_0/cas_count_read/ S0-C1 1 1012534 cycles S0-C1 0 <not counted> MiB uncore_imc_0/cas_count_read/ S1-C0 1 916130 cycles S1-C0 1 0.08 MiB uncore_imc_0/cas_count_read/2.001110843 seconds time elapsedSigned-off-by Kan Liang <[email protected]>--- tools/perf/util/evlist.c | 77 ++++++++++++++++++++++++++++++++++++++++++++---- 1 file changed, 72 insertions(+), 5 deletions(-)tools/perf/util/evlist.c +72
-6@@ -1101,6 +1101,71 @@ int perf_evlist__mmap(struct perf_evlist *evlist, unsigned int pages,return perf_evlist__mmap_ex (evlist, pages, overwrite,0 , false); } +static int cmp_ids (const void *a,const void *b) +{ +return *(int *)a - *(int *)b; +} + +static int perf_evlist__check_evsel_cpus (struct perf_evlist *evlist,struct perf_evsel *evsel) +{ +const struct cpu_map *cpus = evlist->cpus; +const int ncpus =cpu_map__nr (evlist->cpus); +int j =0 , cpu_nr =0 , tmp =0 ; +int i; + +/* ensure we process id in increasing order */ +qsort (evlist->cpus->map, evlist->cpus->nr,sizeof (int ), cmp_ids); + +/* find the common cpus between evsel and evlist. */ +for (i =0 ; i <cpu_map__nr (evsel->cpus);) { + +if (j >= ncpus) { + evsel->cpus->map[i++] =-1 ; +continue ; + } +for (; j < ncpus; j++) { +if (cpus->map[j] < evsel->cpus->map[i]) +continue ; +if (cpus->map[j] == evsel->cpus->map[i]) { + cpu_nr++; + j++; + i++; + }else + evsel->cpus->map[i++] =-1 ; +break ; + } + } + +if (cpu_nr ==0 ) { +pr_warning ("event %s cannot be monitored on the given cpus." +"Please check cpumask\n" , evsel->name); +return -1 ; + } + +if (ncpus > cpu_nr) +pr_warning ("event %s can only be monitored on CPU" , evsel->name); + +/* order evsel cpus */ +for (i =0 ; i < cpu_nr; i++) { +if (evsel->cpus->map[i] ==-1 ) { +while (evsel->cpus->map[tmp] ==-1 ) { + tmp++; +BUG_ON (tmp >=cpu_map__nr (evsel->cpus)); + } + evsel->cpus->map[i] = evsel->cpus->map[tmp]; + evsel->cpus->map[tmp] =-1 ; + } +if (ncpus > cpu_nr) +pr_warning (" %d" , evsel->cpus->map[i]); + tmp++; + } + evsel->cpus->nr = cpu_nr; +if (ncpus > cpu_nr) +pr_warning (". Other CPUs will be discard.\n" ); + +return 0 ; +} +static int perf_evlist__propagate_maps (struct perf_evlist *evlist,struct target *target) {@@ -1108,13 +1173,15 @@ static int perf_evlist__propagate_maps(struct perf_evlist *evlist,evlist__for_each (evlist, evsel) {/* -* We already have cpus for evsel (via PMU sysfs) so -* keep it, if there's no target cpu list defined. +* We already have cpus for evsel (via PMU sysfs) +* and target cpu list defined, check if they are +* compatible. If not, discard incompatible cpus. */ -if (evsel->cpus && target->cpu_list) -cpu_map__put (evsel->cpus); +if (evsel->cpus && target->cpu_list && +perf_evlist__check_evsel_cpus (evlist, evsel)) +return -EINVAL ; -if (!evsel->cpus || target->cpu_list) +if (!evsel->cpus) evsel->cpus =cpu_map__get (evlist->cpus); evsel->threads =thread_map__get (evlist->threads); --1.8.3.1
RE: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
Quoted 17 lines · On Mon, Jun 29, 2015 at 12:55 PM, <[email protected]> wrote:
Quoted 10 lines
From: Kan Liang <[email protected]>
Some PMU events have cpumask, e.g uncore events. The cpu list set by user may be incompatible with event's cpumask.
This patch will check the user defined cpu list. If the incompatible cpu is found, it will warn the user and discard the incompatible cpu. Only available cpu can be stored inevsel->cpus->map. If there is no cpu from cpu list compatible with event's cpumask. It will error out.Quoted 4 lines
Here is an example.
According to cpumask, uncore should only available on CPU0 and CPU18. So the S0-C1 for uncore should not count.
Quoted 6 lines
I don't think this is correct. The cpumask is a default set of CPUs to be used by perf. The cpumask does not indicate the ONLY CPUs on which to monitor. It is just a default. You can monitor an uncore event using a CPU not listed in the cpumask, unless the kernel code has changed recently. If you are not on the default CPUs, the kernel will IPI.
Here I mean that the S0-C1 for uncore should show "<not counted>", as we showed the same thing on "perf stat -a --per-core".
Yes, in theory, user can use a CPU not listed in the cpumask. Because uncore is per-socket event.
However, it brings error and confusion.
- As the below example, if we run -C0,1, we get two results for socket 0. I think it's incorrect. Per-socket event should only have one result per socket.
- Since the cpumask has already been exported to user space, I think users should follow it. Otherwise, why we export cpumask to user space? Implicitly changing the monitored CPU in kernel is not a good way I think.
Thanks,
Kan
Quoted 32 lines
Quoted 32 lines
Without this patch
$ sudo ./perf stat -e cycles,uncore_imc_0/cas_count_read/ -C0,1,18 --per-core sleep 2Performance counter stats for 'CPU(s) 0,1,18':
S0-C0 1 6749638 cycles S0-C0 1 0.83 MiB uncore_imc_0/cas_count_read/ (100.00%) S0-C1 1 232421 cycles S0-C1 1 0.83 MiB uncore_imc_0/cas_count_read/ S1-C0 1 236997 cycles S1-C0 1 0.35 MiB uncore_imc_0/cas_count_read/2.001094019 seconds time elapsedWith this patch
$ perf stat -e cycles,uncore_imc_0/cas_count_read/ -C0,1,18 --per-core sleep 2 event uncore_imc_0/cas_count_read/ can only be monitored on CPU 0 18. Other CPUs will be discard.Performance counter stats for 'CPU(s) 0,1,18':
S0-C0 1 5557406 cycles S0-C0 1 0.21 MiB uncore_imc_0/cas_count_read/ S0-C1 1 1012534 cycles S0-C1 0 <not counted> MiB uncore_imc_0/cas_count_read/ S1-C0 1 916130 cycles S1-C0 1 0.08 MiB uncore_imc_0/cas_count_read/
Quoted 3 lines
Quoted 3 lines
2.001110843 seconds time elapsedSigned-off-by Kan Liang <[email protected]>
���{.n�+�������+%�����ݶ��w��{.n�+����{��G�����{ay�ʇڙ�,j��f���h��������z_�(�階�ݢj"���m������G������?����&���~��iO��z��v�^�m����������?�I�
Re: [PATCH 2/2] perf,tools: check and re-organize evsel cpu maps
Quoted 27 lines · Kan
Quoted 17 lines · On Mon, Jun 29, 2015 at 12:55 PM, <[email protected]> wrote:
Quoted 10 lines
From: Kan Liang <[email protected]>
Some PMU events have cpumask, e.g uncore events. The cpu list set by user may be incompatible with event's cpumask.
This patch will check the user defined cpu list. If the incompatible cpu is found, it will warn the user and discard the incompatible cpu. Only available cpu can be stored inevsel->cpus->map. If there is no cpu from cpu list compatible with event's cpumask. It will error out.Quoted 4 lines
Here is an example.
According to cpumask, uncore should only available on CPU0 and CPU18. So the S0-C1 for uncore should not count.Quoted 6 lines
I don't think this is correct. The cpumask is a default set of CPUs to be used by perf. The cpumask does not indicate the ONLY CPUs on which to monitor. It is just a default. You can monitor an uncore event using a CPU not listed in the cpumask, unless the kernel code has changed recently. If you are not on the default CPUs, the kernel will IPI.
Here I mean that the S0-C1 for uncore should show "<not counted>", as we showed the same thing on "perf stat -a --per-core".
Quoted 9 lines
Yes, in theory, user can use a CPU not listed in the cpumask. Because uncore is per-socket event.
However, it brings error and confusion.
- As the below example, if we run -C0,1, we get two results for socket 0. I think it's incorrect. Per-socket event should only have one result per socket.
- Since the cpumask has already been exported to user space, I think users should follow it. Otherwise, why we export cpumask to user space? Implicitly changing the monitored CPU in kernel is not a good way I think.
I dont follow uncore stuff much :-\ Stephane, does this make sense to you ^^^ ? please check v2
thanks,
jirka
Quoted 10 lines · On Mon, Jun 29, 2015 at 03:55:34PM -0400, [email protected] wrote:
From: Kan Liang <[email protected]>
print_aggrfails to print per-core/per-socket statistics after commitb7f0c203586b("perf evlist: Propagate cpu maps to evsels in an evlist"), if events have differnt cpus. Because inprint_aggr,aggr_get_idneeds index (not cpu id) to find core/pkg id.
This patch introducedperf_evsel__get_cpumap_indexto get the index by cpu id for a given event. The index can be used to find correct cpu id forprint_aggr.
SNIP
tools/perf/builtin-stat.c static void print_aggr(char *prefix) · line 714
Quoted 13 lines
tools/perf/builtin-stat.c +1
-0@@ -708,6 +708,7 @@ static void print_aggr(char *prefix) nr =0 ;for (cpu =0 ; cpu <perf_evsel__nr_cpus (counter); cpu++) { cpu2 =perf_evsel__cpus (counter)->map[cpu]; + cpu2 =perf_evsel__get_cpumap_index (cpu2, evsel_list->cpus); s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue ;
hum, looks like passing the actual cpu number was introduced in:
582ec0829b3d perf stat: Fix per-socket output bug for uncore events
also, we already have the index into counter's cpus, why not use those for getting aggregated id (core,socket).. what do I miss?
thanks,
jirka
---
tools/perf/builtin-stat.c +2 -3
@@ -694,7 +694,7 @@ static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)static void print_aggr (char *prefix) {struct perf_evsel *counter; -int cpu, cpu2, s, s2, id, nr; +int cpu, s, s2, id, nr;double uval;u64 ena, run, val;
@@ -707,8 +707,7 @@ static void print_aggr(char *prefix) val = ena = run =0 ; nr =0 ;for (cpu =0 ; cpu <perf_evsel__nr_cpus (counter); cpu++) { - cpu2 =perf_evsel__cpus (counter)->map[cpu]; - s2 =aggr_get_id (evsel_list->cpus, cpu2); + s2 =aggr_get_id (perf_evsel__cpus (counter), cpu);if (s2 != id)continue ; val +=perf_counts (counter->counts, cpu,0 )->val;
Quoted 30 lines · On Mon, Jun 29, 2015 at 03:55:34PM -0400, [email protected] wrote:
Quoted 7 lines
From: Kan Liang <[email protected]>
print_aggrfails to print per-core/per-socket statistics after commitb7f0c203586b("perf evlist: Propagate cpu maps to evsels in an evlist"), if events have differnt cpus. Because inprint_aggr,aggr_get_idneeds index (not cpu id) to find core/pkg id. This patch introducedperf_evsel__get_cpumap_indexto get the indexby
Quoted 3 lines
cpu id for a given event. The index can be used to find correct cpu id for
print_aggr.SNIP
tools/perf/builtin-stat.c static void print_aggr(char *prefix) · line 709
Quoted 8 lines
tools/perf/builtin-stat.c +0
-0@@ -708,6 +708,7 @@ static void print_aggr(char *prefix) nr =0 ; for (cpu = 0; cpu < perf_evsel__nr_cpus(counter);cpu++) {
tools/perf/builtin-stat.c static void print_aggr(char *prefix) · line 712
Quoted 3 lines
tools/perf/builtin-stat.c +1
-0cpu2 =perf_evsel__cpus (counter)- map[cpu]; + cpu2 =
perf_evsel__get_cpumap_index(cpu2, evsel_list->cpus);
Quoted 7 lines
Quoted 3 lines
tools/perf/builtin-stat.c +0
-0s2 =aggr_get_id (evsel_list->cpus, cpu2);if (s2 != id)continue;
hum, looks like passing the actual cpu number was introduced in:
582ec0829b3dperf stat: Fix per-socket output bug for uncore events
Right, I will change it.
Quoted 2 lines
also, we already have the index into counter's cpus, why not use those for getting aggregated id (core,socket).. what do I miss?
Yes, the index is already there. But we still pass evlist's cpu map to aggr_get_id. So once evsel and evlist have different cpu maps (e.g. -a with uncore event), the wrong aggregated id will be return.
I think the patch as below will fix all these issues.
Thanks,
Kan
---
tools/perf/builtin-stat.c +2 -3
@@ -694,7 +694,7 @@ static void abs_printout(int id, int nr, struct perf_evsel *evsel, double avg)static void print_aggr (char *prefix) {struct perf_evsel *counter; -int cpu, cpu2, s, s2, id, nr; +int cpu, s, s2, id, nr;double uval;u64 ena, run, val;
@@ -707,8 +707,7 @@ static void print_aggr(char *prefix) val = ena = run =0 ; nr =0 ;for (cpu =0 ; cpu <perf_evsel__nr_cpus (counter); cpu++) { - cpu2 =perf_evsel__cpus (counter)->map[cpu]; - s2 =aggr_get_id (evsel_list->cpus, cpu2); + s2 =aggr_get_id (perf_evsel__cpus (counter), cpu);if (s2 != id)continue ; val +=perf_counts (counter->counts, cpu,0 )->val;