Re: [PATCH] perf tools: Enable on a list of CPUs for hybrid
From: Jin, Yao
Date: Thu Jun 17 2021 - 22:38:15 EST
Hi Arnaldo,
On 6/17/2021 10:29 PM, Arnaldo Carvalho de Melo wrote:
Em Thu, Jun 17, 2021 at 03:00:26PM +0800, Jin Yao escreveu:
The perf-record and perf-stat have supported the option '-C/--cpus'
to count or collect only on the list of CPUs provided. This option
needs to be supported for hybrid as well.
For hybrid support, it needs to check that the CPUs are available on
hybrid PMU. On AlderLake, for example, 'cpu_core' has CPU0-CPU15,
and 'cpu_atom' has CPU16-CPU23.
Before:
# perf stat -e cpu_core/cycles/ -C16 true
Performance counter stats for 'CPU(s) 16':
<not supported> cpu_core/cycles/
The perf-stat silently returned "<not supported>" without any helpful
information. It should error out that CPU16 was not available on
'cpu_core'.
After:
# perf stat -e cpu_core/cycles/ -C16 true
'cpu_core' doesn't have cpu 16
failed to use cpu list 16
The above is sensible, as the user specified a specific CPU type
(cpu_core) and asked for a CPU list that is of another type (cpu_atom).
Perhaps the message can be improved a bit by checking if the CPU list
specified is valid for the other cpu type (cpu_atom) and print:
# perf stat -e cpu_core/cycles/ -C16 true
CPU 16 isn't a 'cpu_core', please use 'cpu_atom/cycles/' or use a CPU list in the 'cpu_core' range (0-15).
failed to use cpu list 16
Can we just print the message such as:
"CPU 16 isn't a 'cpu_core', please use a CPU list in the 'cpu_core' range (0-15)".
The user specified the 'cpu_core', so he might only want to enable the event on 'cpu_core'. Maybe we
directly return the cpulist in 'cpu_core' range is more straightfoward.
It also supports to count only on a given CPU.
# perf stat -e cpu_core/cycles/ -C15 -vv true
Humm, what if I want to count events on some cpu_core CPUs plus on some
cpu_atom ones?
I.e. if I do:
perf stat -e cycles -C16
If only with this patch, it would be failed.
# perf stat -e cycles -C 2 -- sleep 1
'cpu_atom' doesn't have cpu 2
failed to use cpu list 2
CPU2 is 'cpu_core', so errors out.
probably it should map to cpu_atom/cycles/, right? And if I do:
Yes, if we can map to cpu_atom/cycles/, that's better.
perf stat -e cycles -C15,16
This should map to cpu_core/cycles/ on CPU 15 and to cpu_atom/cycles/ on
CPU 16?
Ditto, if we can map, that's better.
Or perhaps we should make the CPU list a evsel term? I.e.:
perf stat -e cpu_core/cycles,cpus=15/,cpu_atom/cycles,cpus=16/
Introducing a new term 'cpus' is a more flexible solution. I like it! :)
Or perhaps both?
Let me improve this patch to v2 for soltuion #1 and a follow up patch for solution #2. Is that OK?
Thanks
Jin Yao
- Arnaldo
------------------------------------------------------------
perf_event_attr:
size 128
config 0x400000000
sample_type IDENTIFIER
read_format TOTAL_TIME_ENABLED|TOTAL_TIME_RUNNING
disabled 1
inherit 1
exclude_guest 1
------------------------------------------------------------
sys_perf_event_open: pid -1 cpu 15 group_fd -1 flags 0x8 = 3
cycles: 0: 103287 569776 569776
cycles: 103287 569776 569776
Performance counter stats for 'CPU(s) 15':
103,287 cpu_core/cycles/
0.000566813 seconds time elapsed
Collect the counts of 'cycles' on CPU15 (CPU15 is in 'cpu_core').
Signed-off-by: Jin Yao <yao.jin@xxxxxxxxxxxxxxx>
---
tools/perf/builtin-record.c | 6 +++++
tools/perf/builtin-stat.c | 5 ++++
tools/perf/util/evlist-hybrid.c | 43 +++++++++++++++++++++++++++++++++
tools/perf/util/evlist-hybrid.h | 1 +
tools/perf/util/evlist.c | 1 +
tools/perf/util/pmu.c | 23 ++++++++++++++++++
tools/perf/util/pmu.h | 3 +++
7 files changed, 82 insertions(+)
diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index bc3dd379eb67..bd39d4260549 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -2877,6 +2877,12 @@ int cmd_record(int argc, const char **argv)
/* Enable ignoring missing threads when -u/-p option is defined. */
rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
+ if (evlist__use_cpu_list(rec->evlist, rec->opts.target.cpu_list)) {
+ pr_err("failed to use cpu list %s\n",
+ rec->opts.target.cpu_list);
+ goto out;
+ }
+
err = -ENOMEM;
if (evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
usage_with_options(record_usage, record_options);
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index f9f74a514315..9c27c90d069f 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -2427,6 +2427,11 @@ int cmd_stat(int argc, const char **argv)
if ((stat_config.aggr_mode == AGGR_THREAD) && (target.system_wide))
target.per_thread = true;
+ if (evlist__use_cpu_list(evsel_list, target.cpu_list)) {
+ pr_err("failed to use cpu list %s\n", target.cpu_list);
+ goto out;
+ }
+
if (evlist__create_maps(evsel_list, &target) < 0) {
if (target__has_task(&target)) {
pr_err("Problems finding threads of monitor\n");
diff --git a/tools/perf/util/evlist-hybrid.c b/tools/perf/util/evlist-hybrid.c
index db3f5fbdebe1..e8fdd98aed3f 100644
--- a/tools/perf/util/evlist-hybrid.c
+++ b/tools/perf/util/evlist-hybrid.c
@@ -86,3 +86,46 @@ bool evlist__has_hybrid(struct evlist *evlist)
return false;
}
+
+int evlist__use_cpu_list(struct evlist *evlist, const char *cpu_list)
+{
+ struct perf_cpu_map *cpus;
+ struct evsel *evsel;
+ struct perf_pmu *pmu;
+ int ret;
+
+ if (!perf_pmu__has_hybrid() || !cpu_list)
+ return 0;
+
+ cpus = perf_cpu_map__new(cpu_list);
+ if (!cpus)
+ return -1;
+
+ evlist__for_each_entry(evlist, evsel) {
+ bool exact_match;
+
+ pmu = perf_pmu__find_hybrid_pmu(evsel->pmu_name);
+ if (!pmu)
+ continue;
+
+ if (!perf_pmu__cpus_matched(pmu, cpus, &exact_match)) {
+ ret = -1;
+ goto out;
+ }
+
+ if (!exact_match) {
+ /*
+ * Use the cpus in cpu_list.
+ */
+ perf_cpu_map__put(evsel->core.cpus);
+ perf_cpu_map__put(evsel->core.own_cpus);
+ evsel->core.cpus = perf_cpu_map__get(cpus);
+ evsel->core.own_cpus = perf_cpu_map__get(cpus);
+ }
+ }
+
+ ret = 0;
+out:
+ perf_cpu_map__put(cpus);
+ return ret;
+}
diff --git a/tools/perf/util/evlist-hybrid.h b/tools/perf/util/evlist-hybrid.h
index 19f74b4c340a..f33a4e8443a1 100644
--- a/tools/perf/util/evlist-hybrid.h
+++ b/tools/perf/util/evlist-hybrid.h
@@ -10,5 +10,6 @@
int evlist__add_default_hybrid(struct evlist *evlist, bool precise);
void evlist__warn_hybrid_group(struct evlist *evlist);
bool evlist__has_hybrid(struct evlist *evlist);
+int evlist__use_cpu_list(struct evlist *evlist, const char *cpu_list);
#endif /* __PERF_EVLIST_HYBRID_H */
diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index 6ba9664089bd..e8a0f95f7f47 100644
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -27,6 +27,7 @@
#include "util/perf_api_probe.h"
#include "util/evsel_fprintf.h"
#include "util/evlist-hybrid.h"
+#include "util/pmu.h"
#include <signal.h>
#include <unistd.h>
#include <sched.h>
diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
index 88c8ecdc60b0..0e3a19a6736d 100644
--- a/tools/perf/util/pmu.c
+++ b/tools/perf/util/pmu.c
@@ -1872,3 +1872,26 @@ bool perf_pmu__has_hybrid(void)
return !list_empty(&perf_pmu__hybrid_pmus);
}
+
+bool perf_pmu__cpus_matched(struct perf_pmu *pmu, struct perf_cpu_map *cpus,
+ bool *exact_match)
+{
+ struct perf_cpu_map *pmu_cpus = pmu->cpus;
+ int cpu;
+
+ *exact_match = false;
+
+ for (int i = 0; i < cpus->nr; i++) {
+ cpu = perf_cpu_map__idx(pmu_cpus, cpus->map[i]);
+ if (cpu == -1) {
+ pr_err("'%s' doesn't have cpu %d\n",
+ pmu->name, cpus->map[i]);
+ return false;
+ }
+ }
+
+ if (cpus->nr == pmu_cpus->nr)
+ *exact_match = true;
+
+ return true;
+}
diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h
index a790ef758171..1cca48e02b5d 100644
--- a/tools/perf/util/pmu.h
+++ b/tools/perf/util/pmu.h
@@ -11,6 +11,7 @@
#include "pmu-events/pmu-events.h"
struct evsel_config_term;
+struct perf_cpu_map;
enum {
PERF_PMU_FORMAT_VALUE_CONFIG,
@@ -133,5 +134,7 @@ void perf_pmu__warn_invalid_config(struct perf_pmu *pmu, __u64 config,
char *name);
bool perf_pmu__has_hybrid(void);
+bool perf_pmu__cpus_matched(struct perf_pmu *pmu, struct perf_cpu_map *cpus,
+ bool *exact_match);
#endif /* __PMU_H */
--
2.17.1