perf tools: Remove needless affinity setup/cleanup for pid targets
Arnaldo Carvalho de Melo
maintainercover perf tools: Remove needless affinity setup/cleanup for pid targets
From: Arnaldo Carvalho de Melo <[email protected]>
Hi,
Continuing with the removal of needless affinity setup/cleanup
for pid targets, please ack.
- Arnaldo
Arnaldo Carvalho de Melo (4):
perf affinity: Allow passing a NULL arg to affinity__cleanup() perf stat: No need to setup affinities when starting a workload perf evlist: No need to setup affinities when enabling events for pid
targets
perf evlist: No need to setup affinities when disabling events for pid
targets
tools/perf/builtin-stat.c | 17 ++++++++++------- tools/perf/util/affinity.c | 8 +++++++- tools/perf/util/evlist.c | 28 ++++++++++++++++++---------- 3 files changed, 35 insertions(+), 18 deletions(-)
Signature · 1 line
2.34.1
Arnaldo Carvalho de Melo
maintainerPATCH 1/4 perf affinity: Allow passing a NULL arg to affinity__cleanup()
From: Arnaldo Carvalho de Melo <[email protected]>
Just like with free(), NULL is checked to avoid having all callers do it.
Its convenient for when not using affinity setup/cleanup for dummy CPU maps, i.e. CPU maps for pid targets.
--- tools/perf/util/affinity.c | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-)
tools/perf/util/affinity.c +7 -1
@@ -62,7 +62,7 @@ void affinity__set(struct affinity *a, int cpu)clear_bit (cpu, a->sched_cpus); } -void affinity__cleanup (struct affinity *a) +static void __affinity__cleanup (struct affinity *a) {int cpu_set_size =get_cpu_set_size ();
@@ -71,3 +71,9 @@ void affinity__cleanup(struct affinity *a)zfree (&a->sched_cpus);zfree (&a->orig_cpus); } + +void affinity__cleanup (struct affinity *a) +{ +if (a !=NULL ) +__affinity__cleanup (a); +}
Signature · 1 line
2.34.1
Arnaldo Carvalho de Melo
maintainerPATCH 2/4 perf stat: No need to setup affinities when starting a workload
In reply to Arnaldo Carvalho de MeloFrom: Arnaldo Carvalho de Melo <[email protected]>
I.e. the simple:
$ perf stat sleep 1
Uses a dummy CPU map and thus there is no need to setup/cleanup affinities to avoid IPIs, etc.
With this we're down to a sched_getaffinity() call, in the libnuma initialization, that probably can be removed in a followup patch.
--- tools/perf/builtin-stat.c | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-)
tools/perf/builtin-stat.c +10 -7
@@ -788,7 +788,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)const bool forks = (argc >0 );bool is_pipe =STAT_RECORD ? perf_stat.data.is_pipe : false;struct evlist_cpu_iterator evlist_cpu_itr; -struct affinity affinity; +struct affinity saved_affinity, *affinity =NULL ;int err;bool second_pass = false;
@@ -803,8 +803,11 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)if (group)evlist__set_leader (evsel_list); -if (affinity__setup (&affinity) <0 ) -return -1 ; +if (!cpu_map__is_dummy (evsel_list->core.cpus)) { +if (affinity__setup (&saved_affinity) <0 ) +return -1 ; + affinity = &saved_affinity; + }evlist__for_each_entry (evsel_list, counter) {if (bpf_counter__load (counter, &target))
@@ -813,7 +816,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx) all_counters_use_bpf = false; } -evlist__for_each_cpu (evlist_cpu_itr, evsel_list, &affinity) { +evlist__for_each_cpu (evlist_cpu_itr, evsel_list, affinity) { counter = evlist_cpu_itr.evsel;/*
@@ -869,7 +872,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)*/ /* First close errored or weak retry */ -evlist__for_each_cpu (evlist_cpu_itr, evsel_list, &affinity) { +evlist__for_each_cpu (evlist_cpu_itr, evsel_list, affinity) { counter = evlist_cpu_itr.evsel;if (!counter->reset_group && !counter->errored)
@@ -878,7 +881,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)perf_evsel__close_cpu (&counter->core, evlist_cpu_itr.cpu_map_idx); }/* Now reopen weak */ -evlist__for_each_cpu (evlist_cpu_itr, evsel_list, &affinity) { +evlist__for_each_cpu (evlist_cpu_itr, evsel_list, affinity) { counter = evlist_cpu_itr.evsel;if (!counter->reset_group && !counter->errored)
@@ -904,7 +907,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx) counter->supported = true; } } -affinity__cleanup (&affinity); +affinity__cleanup (affinity);evlist__for_each_entry (evsel_list, counter) {if (!counter->supported) {
Signature · 1 line
2.34.1
Arnaldo Carvalho de Melo
maintainerPATCH 3/4 perf evlist: No need to setup affinities when enabling events for pid targets
In reply to Arnaldo Carvalho de MeloFrom: Arnaldo Carvalho de Melo <[email protected]>
When the target is a pid, not started by 'perf stat' we need to enable the events, and in that case there is no need to setup affinities as we use a dummy CPU map, with just one entry set to -1.
So stop doing it to avoid this needless call to sched_getaffinity():
# strace -ke sched_getaffinity perf stat -e cycles -p 241957 sleep 1 <SNIP> sched_getaffinity(0, 512, [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]) = 8
Quoted 9 lines
/usr/lib64/libc-2.33.so(sched_getaffinity@@GLIBC_2.3.4+0x1a) [
0xe6eea] /var/home/acme/bin/perf(affinity__setup+0x6a) [0x5329ca] /var/home/acme/bin/perf(__evlist__enable.constprop.0+0x23) [0x4b9693] /var/home/acme/bin/perf(enable_counters+0x14d) [0x42de5d] /var/home/acme/bin/perf(cmd_stat+0x2358) [0x4310c8] /var/home/acme/bin/perf(run_builtin+0x6a) [0x4a2cfa] /var/home/acme/bin/perf(main+0x612) [0x40f8c2] /usr/lib64/libc-2.33.so(__libc_start_main+0xd4) [0x27b74] /var/home/acme/bin/perf(_start+0x2d) [0x40fadd]
<SNIP>
--- tools/perf/util/evlist.c | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-)
tools/perf/util/evlist.c +9 -5
@@ -487,12 +487,16 @@ static void __evlist__enable(struct evlist *evlist, char *evsel_name) {struct evsel *pos;struct evlist_cpu_iterator evlist_cpu_itr; -struct affinity affinity; +struct affinity saved_affinity, *affinity =NULL ; -if (affinity__setup (&affinity) <0 ) -return ; +// See explanation in evlist__close() +if (!cpu_map__is_dummy (evlist->core.cpus)) { +if (affinity__setup (&saved_affinity) <0 ) +return ; + affinity = &saved_affinity; + } -evlist__for_each_cpu (evlist_cpu_itr, evlist, &affinity) { +evlist__for_each_cpu (evlist_cpu_itr, evlist, affinity) { pos = evlist_cpu_itr.evsel;if (evsel__strcmp (pos, evsel_name))continue ;
@@ -500,7 +504,7 @@ static void __evlist__enable(struct evlist *evlist, char *evsel_name)continue ;evsel__enable_cpu (pos, evlist_cpu_itr.cpu_map_idx); } -affinity__cleanup (&affinity); +affinity__cleanup (affinity);evlist__for_each_entry (evlist, pos) {if (evsel__strcmp (pos, evsel_name))continue ;
Signature · 1 line
2.34.1
Arnaldo Carvalho de Melo
maintainerPATCH 4/4 perf evlist: No need to setup affinities when disabling events for pid targets
In reply to Arnaldo Carvalho de MeloFrom: Arnaldo Carvalho de Melo <[email protected]>
When the target is a pid, not started by 'perf stat' we need to disable the events, and in that case there is no need to setup affinities as we use a dummy CPU map, with just one entry set to -1.
So stop doing it to avoid this needless call to sched_getaffinity():
# strace -ke sched_getaffinity perf stat -e cycles -p 241957 sleep 1 <SNIP> sched_getaffinity(0, 512, [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]) = 8
Quoted 8 lines
/usr/lib64/libc-2.33.so(sched_getaffinity@@GLIBC_2.3.4+0x1a) [
0xe6eea] /var/home/acme/bin/perf(affinity__setup+0x6a) [0x532a2a] /var/home/acme/bin/perf(__evlist__disable.constprop.0+0x27) [0x4b9827] /var/home/acme/bin/perf(cmd_stat+0x29b5) [0x431725] /var/home/acme/bin/perf(run_builtin+0x6a) [0x4a2cfa] /var/home/acme/bin/perf(main+0x612) [0x40f8c2] /usr/lib64/libc-2.33.so(__libc_start_main+0xd4) [0x27b74] /var/home/acme/bin/perf(_start+0x2d) [0x40fadd]
<SNIP>
--- tools/perf/util/evlist.c | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-)
tools/perf/util/evlist.c +9 -5
@@ -430,15 +430,19 @@ static void __evlist__disable(struct evlist *evlist, char *evsel_name) {struct evsel *pos;struct evlist_cpu_iterator evlist_cpu_itr; -struct affinity affinity; +struct affinity saved_affinity, *affinity =NULL ;bool has_imm = false; -if (affinity__setup (&affinity) <0 ) -return ; +// See explanation in evlist__close() +if (!cpu_map__is_dummy (evlist->core.cpus)) { +if (affinity__setup (&saved_affinity) <0 ) +return ; + affinity = &saved_affinity; + }/* Disable 'immediate' events last */ for (int imm =0 ; imm <=1 ; imm++) { -evlist__for_each_cpu (evlist_cpu_itr, evlist, &affinity) { +evlist__for_each_cpu (evlist_cpu_itr, evlist, affinity) { pos = evlist_cpu_itr.evsel;if (evsel__strcmp (pos, evsel_name))continue ;
@@ -454,7 +458,7 @@ static void __evlist__disable(struct evlist *evlist, char *evsel_name)break ; } -affinity__cleanup (&affinity); +affinity__cleanup (affinity);evlist__for_each_entry (evlist, pos) {if (evsel__strcmp (pos, evsel_name))continue ;
Signature · 1 line
2.34.1
Quoted 7 lines · Arnaldo Carvalho de Melo
From: Arnaldo Carvalho de Melo <[email protected]>
Hi,
Continuing with the removal of needless affinity setup/cleanup
for pid targets, please ack.
The whole set:
Thanks!
Ian
Quoted 18 lines
- Arnaldo
Arnaldo Carvalho de Melo (4):
perf affinity: Allow passing a NULL arg toaffinity__cleanup()perf stat: No need to setup affinities when starting a workload perf evlist: No need to setup affinities when enabling events for pid
targets
perf evlist: No need to setup affinities when disabling events for pid
targetstools/perf/builtin-stat.c | 17 ++++++++++------- tools/perf/util/affinity.c | 8 +++++++- tools/perf/util/evlist.c | 28 ++++++++++++++++++---------- 3 files changed, 35 insertions(+), 18 deletions(-)--
2.34.1