From: Yang Jihong <yangjihong1@huawei.com>
Date: Sat, 12 Aug 2023 08:49:12 +0000
Subject: perf kwork top: Add statistics on hardirq event support
Git-commit: 2f21f5e4b48d910cbbe234264527b23b7cfc49f2
Patch-mainline: v6.7-rc1
References: perf-v6.7 (jsc#PED-6012 jsc#PED-6121)
Calculate the runtime of the hardirq events and subtract it from
the corresponding task runtime to improve the precision.
Example usage:
# perf kwork -k sched,irq record -- perf record -o perf_record.data -a sleep 10
[ perf record: Woken up 2 times to write data ]
[ perf record: Captured and wrote 1.054 MB perf_record.data (18019 samples) ]
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.798 MB perf.data (16334 samples) ]
#
# perf kwork top
Total : 139240.869 ms, 8 cpus
%Cpu(s): 94.91% id, 0.05% hi
%Cpu0 [ 0.05%]
%Cpu1 [| 5.00%]
%Cpu2 [ 0.43%]
%Cpu3 [ 0.57%]
%Cpu4 [ 1.19%]
%Cpu5 [|||||| 20.46%]
%Cpu6 [ 0.48%]
%Cpu7 [||| 12.10%]
PID %CPU RUNTIME COMMMAND
----------------------------------------------------
0 99.54 17325.622 ms swapper/2
0 99.54 17327.527 ms swapper/0
0 99.51 17319.909 ms swapper/6
0 99.42 17304.934 ms swapper/3
0 98.80 17197.385 ms swapper/4
0 94.99 16534.991 ms swapper/1
0 87.89 15295.264 ms swapper/7
0 79.53 13843.182 ms swapper/5
4252 36.50 6361.768 ms perf
4256 1.17 205.215 ms bash
151 0.53 93.298 ms systemd-resolve
4254 0.39 69.468 ms perf
423 0.34 59.368 ms bash
412 0.29 51.204 ms sshd
249 0.20 35.288 ms sd-resolve
16 0.17 30.287 ms rcu_preempt
153 0.09 17.266 ms systemd-timesyn
1 0.09 17.078 ms systemd
4253 0.07 12.457 ms perf
4255 0.06 11.559 ms perf
4234 0.03 6.105 ms kworker/u16:1
69 0.03 6.259 ms kworker/1:1H
4251 0.02 4.615 ms perf
4095 0.02 4.890 ms kworker/7:1
61 0.02 4.005 ms kcompactd0
75 0.02 3.546 ms kworker/2:1
97 0.01 3.106 ms kworker/7:1H
98 0.01 1.995 ms jbd2/sda-8
4088 0.01 1.779 ms kworker/u16:3
2909 0.01 1.795 ms kworker/0:2
4246 0.00 1.117 ms kworker/7:2
51 0.00 0.327 ms ksoftirqd/7
50 0.00 0.369 ms migration/7
102 0.00 0.160 ms kworker/6:1H
76 0.00 0.609 ms kworker/6:1
45 0.00 0.779 ms migration/6
87 0.00 0.504 ms kworker/5:1H
73 0.00 1.130 ms kworker/5:1
41 0.00 0.152 ms ksoftirqd/5
40 0.00 0.702 ms migration/5
64 0.00 0.316 ms kworker/4:1
35 0.00 0.791 ms migration/4
353 0.00 2.211 ms sshd
74 0.00 0.272 ms kworker/3:1
30 0.00 0.819 ms migration/3
25 0.00 0.784 ms migration/2
397 0.00 0.539 ms kworker/1:1
21 0.00 1.600 ms ksoftirqd/1
20 0.00 0.773 ms migration/1
17 0.00 1.682 ms migration/0
15 0.00 0.076 ms ksoftirqd/0
Reviewed-by: Ian Rogers <irogers@google.com>
Signed-off-by: Yang Jihong <yangjihong1@huawei.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Kan Liang <kan.liang@linux.intel.com>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ravi Bangoria <ravi.bangoria@amd.com>
Cc: Sandipan Das <sandipan.das@amd.com>
Link: https://lore.kernel.org/r/20230812084917.169338-12-yangjihong1@huawei.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Tony Jones <tonyj@suse.de>
---
tools/perf/builtin-kwork.c | 93 +++++++++++++++++++++++++++++++++++++++++++---
tools/perf/util/kwork.h | 1 +
2 files changed, 88 insertions(+), 6 deletions(-)
diff --git a/tools/perf/builtin-kwork.c b/tools/perf/builtin-kwork.c
index 1b213af59135..66ad8c2e8020 100644
--- a/tools/perf/builtin-kwork.c
+++ b/tools/perf/builtin-kwork.c
@@ -866,6 +866,36 @@ static int top_entry_event(struct perf_kwork *kwork,
machine, NULL, true);
}
+static int top_exit_event(struct perf_kwork *kwork,
+ struct kwork_class *class,
+ struct evsel *evsel,
+ struct perf_sample *sample,
+ struct machine *machine)
+{
+ struct kwork_work *work, *sched_work;
+ struct kwork_class *sched_class;
+ struct kwork_atom *atom;
+
+ atom = work_pop_atom(kwork, class, KWORK_TRACE_EXIT,
+ KWORK_TRACE_ENTRY, evsel, sample,
+ machine, &work);
+ if (!work)
+ return -1;
+
+ if (atom) {
+ sched_class = get_kwork_class(kwork, KWORK_CLASS_SCHED);
+ if (sched_class) {
+ sched_work = find_work_by_id(&sched_class->work_root,
+ work->id, work->cpu);
+ if (sched_work)
+ top_update_runtime(work, atom, sample);
+ }
+ atom_del(atom);
+ }
+
+ return 0;
+}
+
static int top_sched_switch_event(struct perf_kwork *kwork,
struct kwork_class *class,
struct evsel *evsel,
@@ -933,7 +963,7 @@ static int irq_class_init(struct kwork_class *class,
return 0;
}
-static void irq_work_init(struct perf_kwork *kwork __maybe_unused,
+static void irq_work_init(struct perf_kwork *kwork,
struct kwork_class *class,
struct kwork_work *work,
enum kwork_trace_type src_type __maybe_unused,
@@ -943,8 +973,14 @@ static void irq_work_init(struct perf_kwork *kwork __maybe_unused,
{
work->class = class;
work->cpu = sample->cpu;
- work->id = evsel__intval(evsel, sample, "irq");
- work->name = evsel__strval(evsel, sample, "name");
+
+ if (kwork->report == KWORK_REPORT_TOP) {
+ work->id = evsel__intval_common(evsel, sample, "common_pid");
+ work->name = NULL;
+ } else {
+ work->id = evsel__intval(evsel, sample, "irq");
+ work->name = evsel__strval(evsel, sample, "name");
+ }
}
static void irq_work_name(struct kwork_work *work, char *buf, int len)
@@ -1510,6 +1546,7 @@ static void top_print_cpu_usage(struct perf_kwork *kwork)
{
struct kwork_top_stat *stat = &kwork->top_stat;
u64 idle_time = stat->cpus_runtime[MAX_NR_CPUS].idle;
+ u64 hardirq_time = stat->cpus_runtime[MAX_NR_CPUS].irq;
int cpus_nr = bitmap_weight(stat->all_cpus_bitmap, MAX_NR_CPUS);
u64 cpus_total_time = stat->cpus_runtime[MAX_NR_CPUS].total;
@@ -1518,9 +1555,12 @@ static void top_print_cpu_usage(struct perf_kwork *kwork)
(double)cpus_total_time / NSEC_PER_MSEC,
cpus_nr);
- printf("%%Cpu(s): %*.*f%% id\n",
+ printf("%%Cpu(s): %*.*f%% id, %*.*f%% hi\n",
+ PRINT_CPU_USAGE_WIDTH, PRINT_CPU_USAGE_DECIMAL_WIDTH,
+ cpus_total_time ? (double)idle_time * 100 / cpus_total_time : 0,
+
PRINT_CPU_USAGE_WIDTH, PRINT_CPU_USAGE_DECIMAL_WIDTH,
- cpus_total_time ? (double)idle_time * 100 / cpus_total_time : 0);
+ cpus_total_time ? (double)hardirq_time * 100 / cpus_total_time : 0);
top_print_per_cpu_load(kwork);
}
@@ -1621,6 +1661,8 @@ static int perf_kwork__check_config(struct perf_kwork *kwork,
.exit_event = timehist_exit_event,
};
static struct trace_kwork_handler top_ops = {
+ .entry_event = timehist_entry_event,
+ .exit_event = top_exit_event,
.sched_switch_event = top_sched_switch_event,
};
@@ -1915,6 +1957,43 @@ static void top_calc_idle_time(struct perf_kwork *kwork,
}
}
+static void top_calc_irq_runtime(struct perf_kwork *kwork,
+ enum kwork_class_type type,
+ struct kwork_work *work)
+{
+ struct kwork_top_stat *stat = &kwork->top_stat;
+
+ if (type == KWORK_CLASS_IRQ) {
+ stat->cpus_runtime[work->cpu].irq += work->total_runtime;
+ stat->cpus_runtime[MAX_NR_CPUS].irq += work->total_runtime;
+ }
+}
+
+static void top_subtract_irq_runtime(struct perf_kwork *kwork,
+ struct kwork_work *work)
+{
+ struct kwork_class *class;
+ struct kwork_work *data;
+ unsigned int i;
+ int irq_class_list[] = {KWORK_CLASS_IRQ};
+
+ for (i = 0; i < ARRAY_SIZE(irq_class_list); i++) {
+ class = get_kwork_class(kwork, irq_class_list[i]);
+ if (!class)
+ continue;
+
+ data = find_work_by_id(&class->work_root,
+ work->id, work->cpu);
+ if (!data)
+ continue;
+
+ if (work->total_runtime > data->total_runtime) {
+ work->total_runtime -= data->total_runtime;
+ top_calc_irq_runtime(kwork, irq_class_list[i], data);
+ }
+ }
+}
+
static void top_calc_cpu_usage(struct perf_kwork *kwork)
{
struct kwork_class *class;
@@ -1935,6 +2014,8 @@ static void top_calc_cpu_usage(struct perf_kwork *kwork)
__set_bit(work->cpu, stat->all_cpus_bitmap);
+ top_subtract_irq_runtime(kwork, work);
+
work->cpu_usage = work->total_runtime * 10000 /
stat->cpus_runtime[work->cpu].total;
@@ -2311,7 +2392,7 @@ int cmd_kwork(int argc, const char **argv)
}
kwork.report = KWORK_REPORT_TOP;
if (!kwork.event_list_str)
- kwork.event_list_str = "sched";
+ kwork.event_list_str = "sched, irq";
setup_event_list(&kwork, kwork_options, kwork_usage);
setup_sorting(&kwork, top_options, top_usage);
return perf_kwork__top(&kwork);
diff --git a/tools/perf/util/kwork.h b/tools/perf/util/kwork.h
index 41ed193d5d8b..b3e410d02043 100644
--- a/tools/perf/util/kwork.h
+++ b/tools/perf/util/kwork.h
@@ -183,6 +183,7 @@ struct trace_kwork_handler {
struct __top_cpus_runtime {
u64 load;
u64 idle;
+ u64 irq;
u64 total;
};