Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
perf stat: Support new per thread TopDown metrics
Icelake has support for reporting per thread TopDown metrics.
These are reported differently than the previous TopDown support,
each metric is standalone, but scaled to pipeline "slots".
We don't need to do anything special for HyperThreading anymore.
Teach perf stat --topdown to handle these new metrics and
print them in the same way as the previous TopDown metrics.
The restrictions of only being able to report information per core is
gone.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Co-developed-by: Kan Liang <kan.liang@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Acked-by: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lore.kernel.org/lkml/20200911144808.27603-4-kan.liang@linux.intel.com
Signed-off-by: Kan Liang <kan.liang@linux.intel.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
authored by
Andi Kleen
and committed by
Arnaldo Carvalho de Melo
55c36a9f
acb65150
+134
-4
+5
-2
tools/perf/Documentation/perf-stat.txt
+5
-2
tools/perf/Documentation/perf-stat.txt
···
363
363
For best results it is usually a good idea to use it with interval
364
364
mode like -I 1000, as the bottleneck of workloads can change often.
365
365
366
+
This enables --metric-only, unless overridden with --no-metric-only.
367
+
368
+
The following restrictions only apply to older Intel CPUs and Atom,
369
+
on newer CPUs (IceLake and later) TopDown can be collected for any thread:
370
+
366
371
The top down metrics are collected per core instead of per
367
372
CPU thread. Per core mode is automatically enabled
368
373
and -a (global monitoring) is needed, requiring root rights or
···
378
373
echo 0 > /proc/sys/kernel/nmi_watchdog
379
374
for best results. Otherwise the bottlenecks may be inconsistent
380
375
on workload with changing phases.
381
-
382
-
This enables --metric-only, unless overridden with --no-metric-only.
383
376
384
377
To interpret the results it is usually needed to know on which
385
378
CPUs the workload runs on. If needed the CPUs can be forced using
+28
-2
tools/perf/builtin-stat.c
+28
-2
tools/perf/builtin-stat.c
···
128
128
NULL,
129
129
};
130
130
131
+
static const char *topdown_metric_attrs[] = {
132
+
"slots",
133
+
"topdown-retiring",
134
+
"topdown-bad-spec",
135
+
"topdown-fe-bound",
136
+
"topdown-be-bound",
137
+
NULL,
138
+
};
139
+
131
140
static const char *smi_cost_attrs = {
132
141
"{"
133
142
"msr/aperf/,"
···
1686
1677
char *str = NULL;
1687
1678
bool warn = false;
1688
1679
1680
+
if (!force_metric_only)
1681
+
stat_config.metric_only = true;
1682
+
1683
+
if (topdown_filter_events(topdown_metric_attrs, &str, 1) < 0) {
1684
+
pr_err("Out of memory\n");
1685
+
return -1;
1686
+
}
1687
+
if (topdown_metric_attrs[0] && str) {
1688
+
if (!stat_config.interval && !stat_config.metric_only) {
1689
+
fprintf(stat_config.output,
1690
+
"Topdown accuracy may decrease when measuring long periods.\n"
1691
+
"Please print the result regularly, e.g. -I1000\n");
1692
+
}
1693
+
goto setup_metrics;
1694
+
}
1695
+
1696
+
zfree(&str);
1697
+
1689
1698
if (stat_config.aggr_mode != AGGR_GLOBAL &&
1690
1699
stat_config.aggr_mode != AGGR_CORE) {
1691
1700
pr_err("top down event configuration requires --per-core mode\n");
···
1715
1688
return -1;
1716
1689
}
1717
1690
1718
-
if (!force_metric_only)
1719
-
stat_config.metric_only = true;
1720
1691
if (topdown_filter_events(topdown_attrs, &str,
1721
1692
arch_topdown_check_group(&warn)) < 0) {
1722
1693
pr_err("Out of memory\n");
···
1723
1698
if (topdown_attrs[0] && str) {
1724
1699
if (warn)
1725
1700
arch_topdown_group_warn();
1701
+
setup_metrics:
1726
1702
err = parse_events(evsel_list, str, &errinfo);
1727
1703
if (err) {
1728
1704
fprintf(stderr,
+89
tools/perf/util/stat-shadow.c
+89
tools/perf/util/stat-shadow.c
···
241
241
else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
242
242
update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
243
243
ctx, cpu, count);
244
+
else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
245
+
update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
246
+
ctx, cpu, count);
247
+
else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
248
+
update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
249
+
ctx, cpu, count);
250
+
else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
251
+
update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
252
+
ctx, cpu, count);
253
+
else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
254
+
update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
255
+
ctx, cpu, count);
244
256
else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
245
257
update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
246
258
ctx, cpu, count);
···
717
705
return sanitize_val(1.0 - sum);
718
706
}
719
707
708
+
/*
709
+
* Kernel reports metrics multiplied with slots. To get back
710
+
* the ratios we need to recreate the sum.
711
+
*/
712
+
713
+
static double td_metric_ratio(int ctx, int cpu,
714
+
enum stat_type type,
715
+
struct runtime_stat *stat)
716
+
{
717
+
double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, ctx, cpu) +
718
+
runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, ctx, cpu) +
719
+
runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, ctx, cpu) +
720
+
runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, ctx, cpu);
721
+
double d = runtime_stat_avg(stat, type, ctx, cpu);
722
+
723
+
if (sum)
724
+
return d / sum;
725
+
return 0;
726
+
}
727
+
728
+
/*
729
+
* ... but only if most of the values are actually available.
730
+
* We allow two missing.
731
+
*/
732
+
733
+
static bool full_td(int ctx, int cpu,
734
+
struct runtime_stat *stat)
735
+
{
736
+
int c = 0;
737
+
738
+
if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, ctx, cpu) > 0)
739
+
c++;
740
+
if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, ctx, cpu) > 0)
741
+
c++;
742
+
if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, ctx, cpu) > 0)
743
+
c++;
744
+
if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, ctx, cpu) > 0)
745
+
c++;
746
+
return c >= 2;
747
+
}
748
+
720
749
static void print_smi_cost(struct perf_stat_config *config,
721
750
int cpu, struct evsel *evsel,
722
751
struct perf_stat_output_ctx *out,
···
1126
1073
be_bound * 100.);
1127
1074
else
1128
1075
print_metric(config, ctxp, NULL, NULL, name, 0);
1076
+
} else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1077
+
full_td(ctx, cpu, st)) {
1078
+
double retiring = td_metric_ratio(ctx, cpu,
1079
+
STAT_TOPDOWN_RETIRING, st);
1080
+
1081
+
if (retiring > 0.7)
1082
+
color = PERF_COLOR_GREEN;
1083
+
print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1084
+
retiring * 100.);
1085
+
} else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1086
+
full_td(ctx, cpu, st)) {
1087
+
double fe_bound = td_metric_ratio(ctx, cpu,
1088
+
STAT_TOPDOWN_FE_BOUND, st);
1089
+
1090
+
if (fe_bound > 0.2)
1091
+
color = PERF_COLOR_RED;
1092
+
print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1093
+
fe_bound * 100.);
1094
+
} else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1095
+
full_td(ctx, cpu, st)) {
1096
+
double be_bound = td_metric_ratio(ctx, cpu,
1097
+
STAT_TOPDOWN_BE_BOUND, st);
1098
+
1099
+
if (be_bound > 0.2)
1100
+
color = PERF_COLOR_RED;
1101
+
print_metric(config, ctxp, color, "%8.1f%%", "backend bound",
1102
+
be_bound * 100.);
1103
+
} else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1104
+
full_td(ctx, cpu, st)) {
1105
+
double bad_spec = td_metric_ratio(ctx, cpu,
1106
+
STAT_TOPDOWN_BAD_SPEC, st);
1107
+
1108
+
if (bad_spec > 0.1)
1109
+
color = PERF_COLOR_RED;
1110
+
print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1111
+
bad_spec * 100.);
1129
1112
} else if (evsel->metric_expr) {
1130
1113
generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1131
1114
evsel->name, evsel->metric_name, NULL, 1, cpu, out, st);
+4
tools/perf/util/stat.c
+4
tools/perf/util/stat.c
···
95
95
ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
96
96
ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
97
97
ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
98
+
ID(TOPDOWN_RETIRING, topdown-retiring),
99
+
ID(TOPDOWN_BAD_SPEC, topdown-bad-spec),
100
+
ID(TOPDOWN_FE_BOUND, topdown-fe-bound),
101
+
ID(TOPDOWN_BE_BOUND, topdown-be-bound),
98
102
ID(SMI_NUM, msr/smi/),
99
103
ID(APERF, msr/aperf/),
100
104
};
+8
tools/perf/util/stat.h
+8
tools/perf/util/stat.h
···
28
28
PERF_STAT_EVSEL_ID__TOPDOWN_SLOTS_RETIRED,
29
29
PERF_STAT_EVSEL_ID__TOPDOWN_FETCH_BUBBLES,
30
30
PERF_STAT_EVSEL_ID__TOPDOWN_RECOVERY_BUBBLES,
31
+
PERF_STAT_EVSEL_ID__TOPDOWN_RETIRING,
32
+
PERF_STAT_EVSEL_ID__TOPDOWN_BAD_SPEC,
33
+
PERF_STAT_EVSEL_ID__TOPDOWN_FE_BOUND,
34
+
PERF_STAT_EVSEL_ID__TOPDOWN_BE_BOUND,
31
35
PERF_STAT_EVSEL_ID__SMI_NUM,
32
36
PERF_STAT_EVSEL_ID__APERF,
33
37
PERF_STAT_EVSEL_ID__MAX,
···
86
82
STAT_TOPDOWN_SLOTS_RETIRED,
87
83
STAT_TOPDOWN_FETCH_BUBBLES,
88
84
STAT_TOPDOWN_RECOVERY_BUBBLES,
85
+
STAT_TOPDOWN_RETIRING,
86
+
STAT_TOPDOWN_BAD_SPEC,
87
+
STAT_TOPDOWN_FE_BOUND,
88
+
STAT_TOPDOWN_BE_BOUND,
89
89
STAT_SMI_NUM,
90
90
STAT_APERF,
91
91
STAT_MAX