Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include <math.h>
3#include <stdio.h>
4#include "evsel.h"
5#include "stat.h"
6#include "color.h"
7#include "debug.h"
8#include "pmu.h"
9#include "rblist.h"
10#include "evlist.h"
11#include "expr.h"
12#include "metricgroup.h"
13#include "cgroup.h"
14#include "units.h"
15#include <linux/zalloc.h>
16#include "iostat.h"
17#include "util/hashmap.h"
18
19struct stats walltime_nsecs_stats;
20struct rusage_stats ru_stats;
21
22enum {
23 CTX_BIT_USER = 1 << 0,
24 CTX_BIT_KERNEL = 1 << 1,
25 CTX_BIT_HV = 1 << 2,
26 CTX_BIT_HOST = 1 << 3,
27 CTX_BIT_IDLE = 1 << 4,
28 CTX_BIT_MAX = 1 << 5,
29};
30
31enum stat_type {
32 STAT_NONE = 0,
33 STAT_NSECS,
34 STAT_CYCLES,
35 STAT_INSTRUCTIONS,
36 STAT_STALLED_CYCLES_FRONT,
37 STAT_STALLED_CYCLES_BACK,
38 STAT_BRANCHES,
39 STAT_BRANCH_MISS,
40 STAT_CACHE_REFS,
41 STAT_CACHE_MISSES,
42 STAT_L1_DCACHE,
43 STAT_L1_ICACHE,
44 STAT_LL_CACHE,
45 STAT_ITLB_CACHE,
46 STAT_DTLB_CACHE,
47 STAT_L1D_MISS,
48 STAT_L1I_MISS,
49 STAT_LL_MISS,
50 STAT_DTLB_MISS,
51 STAT_ITLB_MISS,
52 STAT_MAX
53};
54
55static int evsel_context(const struct evsel *evsel)
56{
57 int ctx = 0;
58
59 if (evsel->core.attr.exclude_kernel)
60 ctx |= CTX_BIT_KERNEL;
61 if (evsel->core.attr.exclude_user)
62 ctx |= CTX_BIT_USER;
63 if (evsel->core.attr.exclude_hv)
64 ctx |= CTX_BIT_HV;
65 if (evsel->core.attr.exclude_host)
66 ctx |= CTX_BIT_HOST;
67 if (evsel->core.attr.exclude_idle)
68 ctx |= CTX_BIT_IDLE;
69
70 return ctx;
71}
72
73void perf_stat__reset_shadow_stats(void)
74{
75 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
76 memset(&ru_stats, 0, sizeof(ru_stats));
77}
78
79static enum stat_type evsel__stat_type(const struct evsel *evsel)
80{
81 /* Fake perf_hw_cache_op_id values for use with evsel__match. */
82 u64 PERF_COUNT_hw_cache_l1d_miss = PERF_COUNT_HW_CACHE_L1D |
83 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
84 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
85 u64 PERF_COUNT_hw_cache_l1i_miss = PERF_COUNT_HW_CACHE_L1I |
86 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
87 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
88 u64 PERF_COUNT_hw_cache_ll_miss = PERF_COUNT_HW_CACHE_LL |
89 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
90 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
91 u64 PERF_COUNT_hw_cache_dtlb_miss = PERF_COUNT_HW_CACHE_DTLB |
92 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
93 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
94 u64 PERF_COUNT_hw_cache_itlb_miss = PERF_COUNT_HW_CACHE_ITLB |
95 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
96 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
97
98 if (evsel__is_clock(evsel))
99 return STAT_NSECS;
100 else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES))
101 return STAT_CYCLES;
102 else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS))
103 return STAT_INSTRUCTIONS;
104 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
105 return STAT_STALLED_CYCLES_FRONT;
106 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND))
107 return STAT_STALLED_CYCLES_BACK;
108 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_INSTRUCTIONS))
109 return STAT_BRANCHES;
110 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES))
111 return STAT_BRANCH_MISS;
112 else if (evsel__match(evsel, HARDWARE, HW_CACHE_REFERENCES))
113 return STAT_CACHE_REFS;
114 else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES))
115 return STAT_CACHE_MISSES;
116 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1D))
117 return STAT_L1_DCACHE;
118 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1I))
119 return STAT_L1_ICACHE;
120 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_LL))
121 return STAT_LL_CACHE;
122 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_DTLB))
123 return STAT_DTLB_CACHE;
124 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_ITLB))
125 return STAT_ITLB_CACHE;
126 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1d_miss))
127 return STAT_L1D_MISS;
128 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1i_miss))
129 return STAT_L1I_MISS;
130 else if (evsel__match(evsel, HW_CACHE, hw_cache_ll_miss))
131 return STAT_LL_MISS;
132 else if (evsel__match(evsel, HW_CACHE, hw_cache_dtlb_miss))
133 return STAT_DTLB_MISS;
134 else if (evsel__match(evsel, HW_CACHE, hw_cache_itlb_miss))
135 return STAT_ITLB_MISS;
136 return STAT_NONE;
137}
138
139static const char *get_ratio_color(const double ratios[3], double val)
140{
141 const char *color = PERF_COLOR_NORMAL;
142
143 if (val > ratios[0])
144 color = PERF_COLOR_RED;
145 else if (val > ratios[1])
146 color = PERF_COLOR_MAGENTA;
147 else if (val > ratios[2])
148 color = PERF_COLOR_YELLOW;
149
150 return color;
151}
152
153static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type type)
154{
155 const struct evsel *cur;
156 int evsel_ctx = evsel_context(evsel);
157
158 evlist__for_each_entry(evsel->evlist, cur) {
159 struct perf_stat_aggr *aggr;
160
161 /* Ignore the evsel that is being searched from. */
162 if (evsel == cur)
163 continue;
164
165 /* Ignore evsels that are part of different groups. */
166 if (evsel->core.leader->nr_members > 1 &&
167 evsel->core.leader != cur->core.leader)
168 continue;
169 /* Ignore evsels with mismatched modifiers. */
170 if (evsel_ctx != evsel_context(cur))
171 continue;
172 /* Ignore if not the cgroup we're looking for. */
173 if (evsel->cgrp != cur->cgrp)
174 continue;
175 /* Ignore if not the stat we're looking for. */
176 if (type != evsel__stat_type(cur))
177 continue;
178
179 aggr = &cur->stats->aggr[aggr_idx];
180 if (type == STAT_NSECS)
181 return aggr->counts.val;
182 return aggr->counts.val * cur->scale;
183 }
184 return 0.0;
185}
186
187static void print_ratio(struct perf_stat_config *config,
188 const struct evsel *evsel, int aggr_idx,
189 double numerator, struct perf_stat_output_ctx *out,
190 enum stat_type denominator_type,
191 const double color_ratios[3], const char *unit)
192{
193 double denominator = find_stat(evsel, aggr_idx, denominator_type);
194
195 if (numerator && denominator) {
196 double ratio = numerator / denominator * 100.0;
197 const char *color = get_ratio_color(color_ratios, ratio);
198
199 out->print_metric(config, out->ctx, color, "%7.2f%%", unit, ratio);
200 } else
201 out->print_metric(config, out->ctx, NULL, NULL, unit, 0);
202}
203
204static void print_stalled_cycles_front(struct perf_stat_config *config,
205 const struct evsel *evsel,
206 int aggr_idx, double stalled,
207 struct perf_stat_output_ctx *out)
208{
209 static const double color_ratios[3] = {50.0, 30.0, 10.0};
210
211 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, color_ratios,
212 "frontend cycles idle");
213}
214
215static void print_stalled_cycles_back(struct perf_stat_config *config,
216 const struct evsel *evsel,
217 int aggr_idx, double stalled,
218 struct perf_stat_output_ctx *out)
219{
220 static const double color_ratios[3] = {75.0, 50.0, 20.0};
221
222 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, color_ratios,
223 "backend cycles idle");
224}
225
226static void print_branch_miss(struct perf_stat_config *config,
227 const struct evsel *evsel,
228 int aggr_idx, double misses,
229 struct perf_stat_output_ctx *out)
230{
231 static const double color_ratios[3] = {20.0, 10.0, 5.0};
232
233 print_ratio(config, evsel, aggr_idx, misses, out, STAT_BRANCHES, color_ratios,
234 "of all branches");
235}
236
237static void print_l1d_miss(struct perf_stat_config *config,
238 const struct evsel *evsel,
239 int aggr_idx, double misses,
240 struct perf_stat_output_ctx *out)
241{
242 static const double color_ratios[3] = {20.0, 10.0, 5.0};
243
244 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_DCACHE, color_ratios,
245 "of all L1-dcache accesses");
246}
247
248static void print_l1i_miss(struct perf_stat_config *config,
249 const struct evsel *evsel,
250 int aggr_idx, double misses,
251 struct perf_stat_output_ctx *out)
252{
253 static const double color_ratios[3] = {20.0, 10.0, 5.0};
254
255 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_ICACHE, color_ratios,
256 "of all L1-icache accesses");
257}
258
259static void print_ll_miss(struct perf_stat_config *config,
260 const struct evsel *evsel,
261 int aggr_idx, double misses,
262 struct perf_stat_output_ctx *out)
263{
264 static const double color_ratios[3] = {20.0, 10.0, 5.0};
265
266 print_ratio(config, evsel, aggr_idx, misses, out, STAT_LL_CACHE, color_ratios,
267 "of all LL-cache accesses");
268}
269
270static void print_dtlb_miss(struct perf_stat_config *config,
271 const struct evsel *evsel,
272 int aggr_idx, double misses,
273 struct perf_stat_output_ctx *out)
274{
275 static const double color_ratios[3] = {20.0, 10.0, 5.0};
276
277 print_ratio(config, evsel, aggr_idx, misses, out, STAT_DTLB_CACHE, color_ratios,
278 "of all dTLB cache accesses");
279}
280
281static void print_itlb_miss(struct perf_stat_config *config,
282 const struct evsel *evsel,
283 int aggr_idx, double misses,
284 struct perf_stat_output_ctx *out)
285{
286 static const double color_ratios[3] = {20.0, 10.0, 5.0};
287
288 print_ratio(config, evsel, aggr_idx, misses, out, STAT_ITLB_CACHE, color_ratios,
289 "of all iTLB cache accesses");
290}
291
292static void print_cache_miss(struct perf_stat_config *config,
293 const struct evsel *evsel,
294 int aggr_idx, double misses,
295 struct perf_stat_output_ctx *out)
296{
297 static const double color_ratios[3] = {20.0, 10.0, 5.0};
298
299 print_ratio(config, evsel, aggr_idx, misses, out, STAT_CACHE_REFS, color_ratios,
300 "of all cache refs");
301}
302
303static void print_instructions(struct perf_stat_config *config,
304 const struct evsel *evsel,
305 int aggr_idx, double instructions,
306 struct perf_stat_output_ctx *out)
307{
308 print_metric_t print_metric = out->print_metric;
309 void *ctxp = out->ctx;
310 double cycles = find_stat(evsel, aggr_idx, STAT_CYCLES);
311 double max_stalled = max(find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_FRONT),
312 find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_BACK));
313
314 if (cycles) {
315 print_metric(config, ctxp, NULL, "%7.2f ", "insn per cycle",
316 instructions / cycles);
317 } else
318 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
319
320 if (max_stalled && instructions) {
321 out->new_line(config, ctxp);
322 print_metric(config, ctxp, NULL, "%7.2f ", "stalled cycles per insn",
323 max_stalled / instructions);
324 }
325}
326
327static void print_cycles(struct perf_stat_config *config,
328 const struct evsel *evsel,
329 int aggr_idx, double cycles,
330 struct perf_stat_output_ctx *out)
331{
332 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
333
334 if (cycles && nsecs) {
335 double ratio = cycles / nsecs;
336
337 out->print_metric(config, out->ctx, NULL, "%8.3f", "GHz", ratio);
338 } else
339 out->print_metric(config, out->ctx, NULL, NULL, "GHz", 0);
340}
341
342static void print_nsecs(struct perf_stat_config *config,
343 const struct evsel *evsel,
344 int aggr_idx __maybe_unused, double nsecs,
345 struct perf_stat_output_ctx *out)
346{
347 print_metric_t print_metric = out->print_metric;
348 void *ctxp = out->ctx;
349 double wall_time = avg_stats(&walltime_nsecs_stats);
350
351 if (wall_time) {
352 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
353 nsecs / (wall_time * evsel->scale));
354 } else
355 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
356}
357
358static int prepare_metric(const struct metric_expr *mexp,
359 const struct evsel *evsel,
360 struct expr_parse_ctx *pctx,
361 int aggr_idx)
362{
363 struct evsel * const *metric_events = mexp->metric_events;
364 struct metric_ref *metric_refs = mexp->metric_refs;
365 int i;
366
367 for (i = 0; metric_events[i]; i++) {
368 char *n;
369 double val;
370 int source_count = 0;
371
372 if (evsel__is_tool(metric_events[i])) {
373 struct stats *stats;
374 double scale;
375
376 switch (metric_events[i]->tool_event) {
377 case PERF_TOOL_DURATION_TIME:
378 stats = &walltime_nsecs_stats;
379 scale = 1e-9;
380 break;
381 case PERF_TOOL_USER_TIME:
382 stats = &ru_stats.ru_utime_usec_stat;
383 scale = 1e-6;
384 break;
385 case PERF_TOOL_SYSTEM_TIME:
386 stats = &ru_stats.ru_stime_usec_stat;
387 scale = 1e-6;
388 break;
389 case PERF_TOOL_NONE:
390 pr_err("Invalid tool event 'none'");
391 abort();
392 case PERF_TOOL_MAX:
393 pr_err("Invalid tool event 'max'");
394 abort();
395 default:
396 pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
397 abort();
398 }
399 val = avg_stats(stats) * scale;
400 source_count = 1;
401 } else {
402 struct perf_stat_evsel *ps = metric_events[i]->stats;
403 struct perf_stat_aggr *aggr;
404
405 /*
406 * If there are multiple uncore PMUs and we're not
407 * reading the leader's stats, determine the stats for
408 * the appropriate uncore PMU.
409 */
410 if (evsel && evsel->metric_leader &&
411 evsel->pmu != evsel->metric_leader->pmu &&
412 mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
413 struct evsel *pos;
414
415 evlist__for_each_entry(evsel->evlist, pos) {
416 if (pos->pmu != evsel->pmu)
417 continue;
418 if (pos->metric_leader != mexp->metric_events[i])
419 continue;
420 ps = pos->stats;
421 source_count = 1;
422 break;
423 }
424 }
425 aggr = &ps->aggr[aggr_idx];
426 if (!aggr)
427 break;
428
429 if (!metric_events[i]->supported) {
430 /*
431 * Not supported events will have a count of 0,
432 * which can be confusing in a
433 * metric. Explicitly set the value to NAN. Not
434 * counted events (enable time of 0) are read as
435 * 0.
436 */
437 val = NAN;
438 source_count = 0;
439 } else {
440 val = aggr->counts.val;
441 if (!source_count)
442 source_count = evsel__source_count(metric_events[i]);
443 }
444 }
445 n = strdup(evsel__metric_id(metric_events[i]));
446 if (!n)
447 return -ENOMEM;
448
449 expr__add_id_val_source_count(pctx, n, val, source_count);
450 }
451
452 for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
453 int ret = expr__add_ref(pctx, &metric_refs[j]);
454
455 if (ret)
456 return ret;
457 }
458
459 return i;
460}
461
462static void generic_metric(struct perf_stat_config *config,
463 struct metric_expr *mexp,
464 struct evsel *evsel,
465 int aggr_idx,
466 struct perf_stat_output_ctx *out)
467{
468 print_metric_t print_metric = out->print_metric;
469 const char *metric_name = mexp->metric_name;
470 const char *metric_expr = mexp->metric_expr;
471 const char *metric_threshold = mexp->metric_threshold;
472 const char *metric_unit = mexp->metric_unit;
473 struct evsel * const *metric_events = mexp->metric_events;
474 int runtime = mexp->runtime;
475 struct expr_parse_ctx *pctx;
476 double ratio, scale, threshold;
477 int i;
478 void *ctxp = out->ctx;
479 const char *color = NULL;
480
481 pctx = expr__ctx_new();
482 if (!pctx)
483 return;
484
485 if (config->user_requested_cpu_list)
486 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
487 pctx->sctx.runtime = runtime;
488 pctx->sctx.system_wide = config->system_wide;
489 i = prepare_metric(mexp, evsel, pctx, aggr_idx);
490 if (i < 0) {
491 expr__ctx_free(pctx);
492 return;
493 }
494 if (!metric_events[i]) {
495 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
496 char *unit;
497 char metric_bf[64];
498
499 if (metric_threshold &&
500 expr__parse(&threshold, pctx, metric_threshold) == 0 &&
501 !isnan(threshold)) {
502 color = fpclassify(threshold) == FP_ZERO
503 ? PERF_COLOR_GREEN : PERF_COLOR_RED;
504 }
505
506 if (metric_unit && metric_name) {
507 if (perf_pmu__convert_scale(metric_unit,
508 &unit, &scale) >= 0) {
509 ratio *= scale;
510 }
511 if (strstr(metric_expr, "?"))
512 scnprintf(metric_bf, sizeof(metric_bf),
513 "%s %s_%d", unit, metric_name, runtime);
514 else
515 scnprintf(metric_bf, sizeof(metric_bf),
516 "%s %s", unit, metric_name);
517
518 print_metric(config, ctxp, color, "%8.1f",
519 metric_bf, ratio);
520 } else {
521 print_metric(config, ctxp, color, "%8.2f",
522 metric_name ?
523 metric_name :
524 out->force_header ? evsel->name : "",
525 ratio);
526 }
527 } else {
528 print_metric(config, ctxp, color, /*unit=*/NULL,
529 out->force_header ?
530 (metric_name ?: evsel->name) : "", 0);
531 }
532 } else {
533 print_metric(config, ctxp, color, /*unit=*/NULL,
534 out->force_header ?
535 (metric_name ?: evsel->name) : "", 0);
536 }
537
538 expr__ctx_free(pctx);
539}
540
541double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
542{
543 struct expr_parse_ctx *pctx;
544 double ratio = 0.0;
545
546 pctx = expr__ctx_new();
547 if (!pctx)
548 return NAN;
549
550 if (prepare_metric(mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
551 goto out;
552
553 if (expr__parse(&ratio, pctx, mexp->metric_expr))
554 ratio = 0.0;
555
556out:
557 expr__ctx_free(pctx);
558 return ratio;
559}
560
561static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
562 struct evsel *evsel,
563 void *ctxp,
564 const char *name,
565 struct perf_stat_output_ctx *out)
566{
567 bool need_full_name = perf_pmus__num_core_pmus() > 1;
568 static const char *last_name;
569 static const char *last_pmu;
570 char full_name[64];
571
572 /*
573 * A metricgroup may have several metric events,
574 * e.g.,TopdownL1 on e-core of ADL.
575 * The name has been output by the first metric
576 * event. Only align with other metics from
577 * different metric events.
578 */
579 if (last_name && !strcmp(last_name, name)) {
580 if (!need_full_name || !strcmp(last_pmu, evsel->pmu_name)) {
581 out->print_metricgroup_header(config, ctxp, NULL);
582 return;
583 }
584 }
585
586 if (need_full_name)
587 scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu_name);
588 else
589 scnprintf(full_name, sizeof(full_name), "%s", name);
590
591 out->print_metricgroup_header(config, ctxp, full_name);
592
593 last_name = name;
594 last_pmu = evsel->pmu_name;
595}
596
597/**
598 * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
599 * For the non-default, all metrics associated
600 * with the evsel are printed.
601 * For the default mode, only the metrics from
602 * the same metricgroup and the name of the
603 * metricgroup are printed. To print the metrics
604 * from the next metricgroup (if available),
605 * invoke the function with correspoinding
606 * metric_expr.
607 */
608void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
609 struct evsel *evsel,
610 int aggr_idx,
611 int *num,
612 void *from,
613 struct perf_stat_output_ctx *out,
614 struct rblist *metric_events)
615{
616 struct metric_event *me;
617 struct metric_expr *mexp = from;
618 void *ctxp = out->ctx;
619 bool header_printed = false;
620 const char *name = NULL;
621
622 me = metricgroup__lookup(metric_events, evsel, false);
623 if (me == NULL)
624 return NULL;
625
626 if (!mexp)
627 mexp = list_first_entry(&me->head, typeof(*mexp), nd);
628
629 list_for_each_entry_from(mexp, &me->head, nd) {
630 /* Print the display name of the Default metricgroup */
631 if (!config->metric_only && me->is_default) {
632 if (!name)
633 name = mexp->default_metricgroup_name;
634 /*
635 * Two or more metricgroup may share the same metric
636 * event, e.g., TopdownL1 and TopdownL2 on SPR.
637 * Return and print the prefix, e.g., noise, running
638 * for the next metricgroup.
639 */
640 if (strcmp(name, mexp->default_metricgroup_name))
641 return (void *)mexp;
642 /* Only print the name of the metricgroup once */
643 if (!header_printed) {
644 header_printed = true;
645 perf_stat__print_metricgroup_header(config, evsel, ctxp,
646 name, out);
647 }
648 }
649
650 if ((*num)++ > 0)
651 out->new_line(config, ctxp);
652 generic_metric(config, mexp, evsel, aggr_idx, out);
653 }
654
655 return NULL;
656}
657
658void perf_stat__print_shadow_stats(struct perf_stat_config *config,
659 struct evsel *evsel,
660 double avg, int aggr_idx,
661 struct perf_stat_output_ctx *out,
662 struct rblist *metric_events)
663{
664 typedef void (*stat_print_function_t)(struct perf_stat_config *config,
665 const struct evsel *evsel,
666 int aggr_idx, double misses,
667 struct perf_stat_output_ctx *out);
668 static const stat_print_function_t stat_print_function[STAT_MAX] = {
669 [STAT_INSTRUCTIONS] = print_instructions,
670 [STAT_BRANCH_MISS] = print_branch_miss,
671 [STAT_L1D_MISS] = print_l1d_miss,
672 [STAT_L1I_MISS] = print_l1i_miss,
673 [STAT_DTLB_MISS] = print_dtlb_miss,
674 [STAT_ITLB_MISS] = print_itlb_miss,
675 [STAT_LL_MISS] = print_ll_miss,
676 [STAT_CACHE_MISSES] = print_cache_miss,
677 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
678 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
679 [STAT_CYCLES] = print_cycles,
680 [STAT_NSECS] = print_nsecs,
681 };
682 print_metric_t print_metric = out->print_metric;
683 void *ctxp = out->ctx;
684 int num = 1;
685
686 if (config->iostat_run) {
687 iostat_print_metric(config, evsel, out);
688 } else {
689 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
690
691 if (fn)
692 fn(config, evsel, aggr_idx, avg, out);
693 else {
694 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
695
696 if (nsecs) {
697 char unit = ' ';
698 char unit_buf[10] = "/sec";
699 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
700 &unit);
701
702 if (unit != ' ')
703 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
704 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
705 } else
706 num = 0;
707 }
708 }
709
710 perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
711 &num, NULL, out, metric_events);
712
713 if (num == 0)
714 print_metric(config, ctxp, NULL, NULL, NULL, 0);
715}
716
717/**
718 * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
719 * if it's not running or not the metric event.
720 */
721bool perf_stat__skip_metric_event(struct evsel *evsel,
722 struct rblist *metric_events,
723 u64 ena, u64 run)
724{
725 if (!evsel->default_metricgroup)
726 return false;
727
728 if (!ena || !run)
729 return true;
730
731 return !metricgroup__lookup(metric_events, evsel, false);
732}
1// SPDX-License-Identifier: GPL-2.0
2#include <math.h>
3#include <stdio.h>
4#include "evsel.h"
5#include "stat.h"
6#include "color.h"
7#include "debug.h"
8#include "pmu.h"
9#include "rblist.h"
10#include "evlist.h"
11#include "expr.h"
12#include "metricgroup.h"
13#include "cgroup.h"
14#include "units.h"
15#include <linux/zalloc.h>
16#include "iostat.h"
17#include "util/hashmap.h"
18#include "tool_pmu.h"
19
20struct stats walltime_nsecs_stats;
21struct rusage_stats ru_stats;
22
23enum {
24 CTX_BIT_USER = 1 << 0,
25 CTX_BIT_KERNEL = 1 << 1,
26 CTX_BIT_HV = 1 << 2,
27 CTX_BIT_HOST = 1 << 3,
28 CTX_BIT_IDLE = 1 << 4,
29 CTX_BIT_MAX = 1 << 5,
30};
31
32enum stat_type {
33 STAT_NONE = 0,
34 STAT_NSECS,
35 STAT_CYCLES,
36 STAT_INSTRUCTIONS,
37 STAT_STALLED_CYCLES_FRONT,
38 STAT_STALLED_CYCLES_BACK,
39 STAT_BRANCHES,
40 STAT_BRANCH_MISS,
41 STAT_CACHE_REFS,
42 STAT_CACHE_MISSES,
43 STAT_L1_DCACHE,
44 STAT_L1_ICACHE,
45 STAT_LL_CACHE,
46 STAT_ITLB_CACHE,
47 STAT_DTLB_CACHE,
48 STAT_L1D_MISS,
49 STAT_L1I_MISS,
50 STAT_LL_MISS,
51 STAT_DTLB_MISS,
52 STAT_ITLB_MISS,
53 STAT_MAX
54};
55
56static int evsel_context(const struct evsel *evsel)
57{
58 int ctx = 0;
59
60 if (evsel->core.attr.exclude_kernel)
61 ctx |= CTX_BIT_KERNEL;
62 if (evsel->core.attr.exclude_user)
63 ctx |= CTX_BIT_USER;
64 if (evsel->core.attr.exclude_hv)
65 ctx |= CTX_BIT_HV;
66 if (evsel->core.attr.exclude_host)
67 ctx |= CTX_BIT_HOST;
68 if (evsel->core.attr.exclude_idle)
69 ctx |= CTX_BIT_IDLE;
70
71 return ctx;
72}
73
74void perf_stat__reset_shadow_stats(void)
75{
76 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
77 memset(&ru_stats, 0, sizeof(ru_stats));
78}
79
80static enum stat_type evsel__stat_type(struct evsel *evsel)
81{
82 /* Fake perf_hw_cache_op_id values for use with evsel__match. */
83 u64 PERF_COUNT_hw_cache_l1d_miss = PERF_COUNT_HW_CACHE_L1D |
84 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
85 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
86 u64 PERF_COUNT_hw_cache_l1i_miss = PERF_COUNT_HW_CACHE_L1I |
87 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
88 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
89 u64 PERF_COUNT_hw_cache_ll_miss = PERF_COUNT_HW_CACHE_LL |
90 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
91 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
92 u64 PERF_COUNT_hw_cache_dtlb_miss = PERF_COUNT_HW_CACHE_DTLB |
93 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
94 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
95 u64 PERF_COUNT_hw_cache_itlb_miss = PERF_COUNT_HW_CACHE_ITLB |
96 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
97 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
98
99 if (evsel__is_clock(evsel))
100 return STAT_NSECS;
101 else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES))
102 return STAT_CYCLES;
103 else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS))
104 return STAT_INSTRUCTIONS;
105 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
106 return STAT_STALLED_CYCLES_FRONT;
107 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND))
108 return STAT_STALLED_CYCLES_BACK;
109 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_INSTRUCTIONS))
110 return STAT_BRANCHES;
111 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES))
112 return STAT_BRANCH_MISS;
113 else if (evsel__match(evsel, HARDWARE, HW_CACHE_REFERENCES))
114 return STAT_CACHE_REFS;
115 else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES))
116 return STAT_CACHE_MISSES;
117 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1D))
118 return STAT_L1_DCACHE;
119 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1I))
120 return STAT_L1_ICACHE;
121 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_LL))
122 return STAT_LL_CACHE;
123 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_DTLB))
124 return STAT_DTLB_CACHE;
125 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_ITLB))
126 return STAT_ITLB_CACHE;
127 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1d_miss))
128 return STAT_L1D_MISS;
129 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1i_miss))
130 return STAT_L1I_MISS;
131 else if (evsel__match(evsel, HW_CACHE, hw_cache_ll_miss))
132 return STAT_LL_MISS;
133 else if (evsel__match(evsel, HW_CACHE, hw_cache_dtlb_miss))
134 return STAT_DTLB_MISS;
135 else if (evsel__match(evsel, HW_CACHE, hw_cache_itlb_miss))
136 return STAT_ITLB_MISS;
137 return STAT_NONE;
138}
139
140static enum metric_threshold_classify get_ratio_thresh(const double ratios[3], double val)
141{
142 assert(ratios[0] > ratios[1]);
143 assert(ratios[1] > ratios[2]);
144
145 return val > ratios[1]
146 ? (val > ratios[0] ? METRIC_THRESHOLD_BAD : METRIC_THRESHOLD_NEARLY_BAD)
147 : (val > ratios[2] ? METRIC_THRESHOLD_LESS_GOOD : METRIC_THRESHOLD_GOOD);
148}
149
150static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type type)
151{
152 struct evsel *cur;
153 int evsel_ctx = evsel_context(evsel);
154
155 evlist__for_each_entry(evsel->evlist, cur) {
156 struct perf_stat_aggr *aggr;
157
158 /* Ignore the evsel that is being searched from. */
159 if (evsel == cur)
160 continue;
161
162 /* Ignore evsels that are part of different groups. */
163 if (evsel->core.leader->nr_members > 1 &&
164 evsel->core.leader != cur->core.leader)
165 continue;
166 /* Ignore evsels with mismatched modifiers. */
167 if (evsel_ctx != evsel_context(cur))
168 continue;
169 /* Ignore if not the cgroup we're looking for. */
170 if (evsel->cgrp != cur->cgrp)
171 continue;
172 /* Ignore if not the stat we're looking for. */
173 if (type != evsel__stat_type(cur))
174 continue;
175
176 /*
177 * Except the SW CLOCK events,
178 * ignore if not the PMU we're looking for.
179 */
180 if ((type != STAT_NSECS) && (evsel->pmu != cur->pmu))
181 continue;
182
183 aggr = &cur->stats->aggr[aggr_idx];
184 if (type == STAT_NSECS)
185 return aggr->counts.val;
186 return aggr->counts.val * cur->scale;
187 }
188 return 0.0;
189}
190
191static void print_ratio(struct perf_stat_config *config,
192 const struct evsel *evsel, int aggr_idx,
193 double numerator, struct perf_stat_output_ctx *out,
194 enum stat_type denominator_type,
195 const double thresh_ratios[3], const char *_unit)
196{
197 double denominator = find_stat(evsel, aggr_idx, denominator_type);
198 double ratio = 0;
199 enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
200 const char *fmt = NULL;
201 const char *unit = NULL;
202
203 if (numerator && denominator) {
204 ratio = numerator / denominator * 100.0;
205 thresh = get_ratio_thresh(thresh_ratios, ratio);
206 fmt = "%7.2f%%";
207 unit = _unit;
208 }
209 out->print_metric(config, out->ctx, thresh, fmt, unit, ratio);
210}
211
212static void print_stalled_cycles_front(struct perf_stat_config *config,
213 const struct evsel *evsel,
214 int aggr_idx, double stalled,
215 struct perf_stat_output_ctx *out)
216{
217 const double thresh_ratios[3] = {50.0, 30.0, 10.0};
218
219 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
220 "frontend cycles idle");
221}
222
223static void print_stalled_cycles_back(struct perf_stat_config *config,
224 const struct evsel *evsel,
225 int aggr_idx, double stalled,
226 struct perf_stat_output_ctx *out)
227{
228 const double thresh_ratios[3] = {75.0, 50.0, 20.0};
229
230 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
231 "backend cycles idle");
232}
233
234static void print_branch_miss(struct perf_stat_config *config,
235 const struct evsel *evsel,
236 int aggr_idx, double misses,
237 struct perf_stat_output_ctx *out)
238{
239 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
240
241 print_ratio(config, evsel, aggr_idx, misses, out, STAT_BRANCHES, thresh_ratios,
242 "of all branches");
243}
244
245static void print_l1d_miss(struct perf_stat_config *config,
246 const struct evsel *evsel,
247 int aggr_idx, double misses,
248 struct perf_stat_output_ctx *out)
249{
250 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
251
252 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_DCACHE, thresh_ratios,
253 "of all L1-dcache accesses");
254}
255
256static void print_l1i_miss(struct perf_stat_config *config,
257 const struct evsel *evsel,
258 int aggr_idx, double misses,
259 struct perf_stat_output_ctx *out)
260{
261 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
262
263 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_ICACHE, thresh_ratios,
264 "of all L1-icache accesses");
265}
266
267static void print_ll_miss(struct perf_stat_config *config,
268 const struct evsel *evsel,
269 int aggr_idx, double misses,
270 struct perf_stat_output_ctx *out)
271{
272 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
273
274 print_ratio(config, evsel, aggr_idx, misses, out, STAT_LL_CACHE, thresh_ratios,
275 "of all LL-cache accesses");
276}
277
278static void print_dtlb_miss(struct perf_stat_config *config,
279 const struct evsel *evsel,
280 int aggr_idx, double misses,
281 struct perf_stat_output_ctx *out)
282{
283 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
284
285 print_ratio(config, evsel, aggr_idx, misses, out, STAT_DTLB_CACHE, thresh_ratios,
286 "of all dTLB cache accesses");
287}
288
289static void print_itlb_miss(struct perf_stat_config *config,
290 const struct evsel *evsel,
291 int aggr_idx, double misses,
292 struct perf_stat_output_ctx *out)
293{
294 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
295
296 print_ratio(config, evsel, aggr_idx, misses, out, STAT_ITLB_CACHE, thresh_ratios,
297 "of all iTLB cache accesses");
298}
299
300static void print_cache_miss(struct perf_stat_config *config,
301 const struct evsel *evsel,
302 int aggr_idx, double misses,
303 struct perf_stat_output_ctx *out)
304{
305 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
306
307 print_ratio(config, evsel, aggr_idx, misses, out, STAT_CACHE_REFS, thresh_ratios,
308 "of all cache refs");
309}
310
311static void print_instructions(struct perf_stat_config *config,
312 const struct evsel *evsel,
313 int aggr_idx, double instructions,
314 struct perf_stat_output_ctx *out)
315{
316 print_metric_t print_metric = out->print_metric;
317 void *ctxp = out->ctx;
318 double cycles = find_stat(evsel, aggr_idx, STAT_CYCLES);
319 double max_stalled = max(find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_FRONT),
320 find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_BACK));
321
322 if (cycles) {
323 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
324 "insn per cycle", instructions / cycles);
325 } else {
326 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
327 "insn per cycle", 0);
328 }
329 if (max_stalled && instructions) {
330 out->new_line(config, ctxp);
331 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
332 "stalled cycles per insn", max_stalled / instructions);
333 }
334}
335
336static void print_cycles(struct perf_stat_config *config,
337 const struct evsel *evsel,
338 int aggr_idx, double cycles,
339 struct perf_stat_output_ctx *out)
340{
341 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
342
343 if (cycles && nsecs) {
344 double ratio = cycles / nsecs;
345
346 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
347 "GHz", ratio);
348 } else {
349 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
350 "GHz", 0);
351 }
352}
353
354static void print_nsecs(struct perf_stat_config *config,
355 const struct evsel *evsel,
356 int aggr_idx __maybe_unused, double nsecs,
357 struct perf_stat_output_ctx *out)
358{
359 print_metric_t print_metric = out->print_metric;
360 void *ctxp = out->ctx;
361 double wall_time = avg_stats(&walltime_nsecs_stats);
362
363 if (wall_time) {
364 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f", "CPUs utilized",
365 nsecs / (wall_time * evsel->scale));
366 } else {
367 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
368 "CPUs utilized", 0);
369 }
370}
371
372static int prepare_metric(const struct metric_expr *mexp,
373 const struct evsel *evsel,
374 struct expr_parse_ctx *pctx,
375 int aggr_idx)
376{
377 struct evsel * const *metric_events = mexp->metric_events;
378 struct metric_ref *metric_refs = mexp->metric_refs;
379 int i;
380
381 for (i = 0; metric_events[i]; i++) {
382 char *n;
383 double val;
384 int source_count = 0;
385
386 if (evsel__is_tool(metric_events[i])) {
387 struct stats *stats;
388 double scale;
389
390 switch (evsel__tool_event(metric_events[i])) {
391 case TOOL_PMU__EVENT_DURATION_TIME:
392 stats = &walltime_nsecs_stats;
393 scale = 1e-9;
394 break;
395 case TOOL_PMU__EVENT_USER_TIME:
396 stats = &ru_stats.ru_utime_usec_stat;
397 scale = 1e-6;
398 break;
399 case TOOL_PMU__EVENT_SYSTEM_TIME:
400 stats = &ru_stats.ru_stime_usec_stat;
401 scale = 1e-6;
402 break;
403 case TOOL_PMU__EVENT_NONE:
404 pr_err("Invalid tool event 'none'");
405 abort();
406 case TOOL_PMU__EVENT_MAX:
407 pr_err("Invalid tool event 'max'");
408 abort();
409 case TOOL_PMU__EVENT_HAS_PMEM:
410 case TOOL_PMU__EVENT_NUM_CORES:
411 case TOOL_PMU__EVENT_NUM_CPUS:
412 case TOOL_PMU__EVENT_NUM_CPUS_ONLINE:
413 case TOOL_PMU__EVENT_NUM_DIES:
414 case TOOL_PMU__EVENT_NUM_PACKAGES:
415 case TOOL_PMU__EVENT_SLOTS:
416 case TOOL_PMU__EVENT_SMT_ON:
417 case TOOL_PMU__EVENT_SYSTEM_TSC_FREQ:
418 default:
419 pr_err("Unexpected tool event '%s'", evsel__name(metric_events[i]));
420 abort();
421 }
422 val = avg_stats(stats) * scale;
423 source_count = 1;
424 } else {
425 struct perf_stat_evsel *ps = metric_events[i]->stats;
426 struct perf_stat_aggr *aggr;
427
428 /*
429 * If there are multiple uncore PMUs and we're not
430 * reading the leader's stats, determine the stats for
431 * the appropriate uncore PMU.
432 */
433 if (evsel && evsel->metric_leader &&
434 evsel->pmu != evsel->metric_leader->pmu &&
435 mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
436 struct evsel *pos;
437
438 evlist__for_each_entry(evsel->evlist, pos) {
439 if (pos->pmu != evsel->pmu)
440 continue;
441 if (pos->metric_leader != mexp->metric_events[i])
442 continue;
443 ps = pos->stats;
444 source_count = 1;
445 break;
446 }
447 }
448 aggr = &ps->aggr[aggr_idx];
449 if (!aggr)
450 break;
451
452 if (!metric_events[i]->supported) {
453 /*
454 * Not supported events will have a count of 0,
455 * which can be confusing in a
456 * metric. Explicitly set the value to NAN. Not
457 * counted events (enable time of 0) are read as
458 * 0.
459 */
460 val = NAN;
461 source_count = 0;
462 } else {
463 val = aggr->counts.val;
464 if (!source_count)
465 source_count = evsel__source_count(metric_events[i]);
466 }
467 }
468 n = strdup(evsel__metric_id(metric_events[i]));
469 if (!n)
470 return -ENOMEM;
471
472 expr__add_id_val_source_count(pctx, n, val, source_count);
473 }
474
475 for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
476 int ret = expr__add_ref(pctx, &metric_refs[j]);
477
478 if (ret)
479 return ret;
480 }
481
482 return i;
483}
484
485static void generic_metric(struct perf_stat_config *config,
486 struct metric_expr *mexp,
487 struct evsel *evsel,
488 int aggr_idx,
489 struct perf_stat_output_ctx *out)
490{
491 print_metric_t print_metric = out->print_metric;
492 const char *metric_name = mexp->metric_name;
493 const char *metric_expr = mexp->metric_expr;
494 const char *metric_threshold = mexp->metric_threshold;
495 const char *metric_unit = mexp->metric_unit;
496 struct evsel * const *metric_events = mexp->metric_events;
497 int runtime = mexp->runtime;
498 struct expr_parse_ctx *pctx;
499 double ratio, scale, threshold;
500 int i;
501 void *ctxp = out->ctx;
502 enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
503
504 pctx = expr__ctx_new();
505 if (!pctx)
506 return;
507
508 if (config->user_requested_cpu_list)
509 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
510 pctx->sctx.runtime = runtime;
511 pctx->sctx.system_wide = config->system_wide;
512 i = prepare_metric(mexp, evsel, pctx, aggr_idx);
513 if (i < 0) {
514 expr__ctx_free(pctx);
515 return;
516 }
517 if (!metric_events[i]) {
518 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
519 char *unit;
520 char metric_bf[128];
521
522 if (metric_threshold &&
523 expr__parse(&threshold, pctx, metric_threshold) == 0 &&
524 !isnan(threshold)) {
525 thresh = fpclassify(threshold) == FP_ZERO
526 ? METRIC_THRESHOLD_GOOD : METRIC_THRESHOLD_BAD;
527 }
528
529 if (metric_unit && metric_name) {
530 if (perf_pmu__convert_scale(metric_unit,
531 &unit, &scale) >= 0) {
532 ratio *= scale;
533 }
534 if (strstr(metric_expr, "?"))
535 scnprintf(metric_bf, sizeof(metric_bf),
536 "%s %s_%d", unit, metric_name, runtime);
537 else
538 scnprintf(metric_bf, sizeof(metric_bf),
539 "%s %s", unit, metric_name);
540
541 print_metric(config, ctxp, thresh, "%8.1f",
542 metric_bf, ratio);
543 } else {
544 print_metric(config, ctxp, thresh, "%8.2f",
545 metric_name ?
546 metric_name :
547 out->force_header ? evsel->name : "",
548 ratio);
549 }
550 } else {
551 print_metric(config, ctxp, thresh, /*fmt=*/NULL,
552 out->force_header ?
553 (metric_name ?: evsel->name) : "", 0);
554 }
555 } else {
556 print_metric(config, ctxp, thresh, /*fmt=*/NULL,
557 out->force_header ?
558 (metric_name ?: evsel->name) : "", 0);
559 }
560
561 expr__ctx_free(pctx);
562}
563
564double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
565{
566 struct expr_parse_ctx *pctx;
567 double ratio = 0.0;
568
569 pctx = expr__ctx_new();
570 if (!pctx)
571 return NAN;
572
573 if (prepare_metric(mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
574 goto out;
575
576 if (expr__parse(&ratio, pctx, mexp->metric_expr))
577 ratio = 0.0;
578
579out:
580 expr__ctx_free(pctx);
581 return ratio;
582}
583
584static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
585 struct evsel *evsel,
586 void *ctxp,
587 const char *name,
588 struct perf_stat_output_ctx *out)
589{
590 bool need_full_name = perf_pmus__num_core_pmus() > 1;
591 static const char *last_name;
592 static const struct perf_pmu *last_pmu;
593 char full_name[64];
594
595 /*
596 * A metricgroup may have several metric events,
597 * e.g.,TopdownL1 on e-core of ADL.
598 * The name has been output by the first metric
599 * event. Only align with other metics from
600 * different metric events.
601 */
602 if (last_name && !strcmp(last_name, name)) {
603 if (!need_full_name || last_pmu != evsel->pmu) {
604 out->print_metricgroup_header(config, ctxp, NULL);
605 return;
606 }
607 }
608
609 if (need_full_name && evsel->pmu)
610 scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu->name);
611 else
612 scnprintf(full_name, sizeof(full_name), "%s", name);
613
614 out->print_metricgroup_header(config, ctxp, full_name);
615
616 last_name = name;
617 last_pmu = evsel->pmu;
618}
619
620/**
621 * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
622 * For the non-default, all metrics associated
623 * with the evsel are printed.
624 * For the default mode, only the metrics from
625 * the same metricgroup and the name of the
626 * metricgroup are printed. To print the metrics
627 * from the next metricgroup (if available),
628 * invoke the function with correspoinding
629 * metric_expr.
630 */
631void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
632 struct evsel *evsel,
633 int aggr_idx,
634 int *num,
635 void *from,
636 struct perf_stat_output_ctx *out,
637 struct rblist *metric_events)
638{
639 struct metric_event *me;
640 struct metric_expr *mexp = from;
641 void *ctxp = out->ctx;
642 bool header_printed = false;
643 const char *name = NULL;
644
645 me = metricgroup__lookup(metric_events, evsel, false);
646 if (me == NULL)
647 return NULL;
648
649 if (!mexp)
650 mexp = list_first_entry(&me->head, typeof(*mexp), nd);
651
652 list_for_each_entry_from(mexp, &me->head, nd) {
653 /* Print the display name of the Default metricgroup */
654 if (!config->metric_only && me->is_default) {
655 if (!name)
656 name = mexp->default_metricgroup_name;
657 /*
658 * Two or more metricgroup may share the same metric
659 * event, e.g., TopdownL1 and TopdownL2 on SPR.
660 * Return and print the prefix, e.g., noise, running
661 * for the next metricgroup.
662 */
663 if (strcmp(name, mexp->default_metricgroup_name))
664 return (void *)mexp;
665 /* Only print the name of the metricgroup once */
666 if (!header_printed) {
667 header_printed = true;
668 perf_stat__print_metricgroup_header(config, evsel, ctxp,
669 name, out);
670 }
671 }
672
673 if ((*num)++ > 0)
674 out->new_line(config, ctxp);
675 generic_metric(config, mexp, evsel, aggr_idx, out);
676 }
677
678 return NULL;
679}
680
681void perf_stat__print_shadow_stats(struct perf_stat_config *config,
682 struct evsel *evsel,
683 double avg, int aggr_idx,
684 struct perf_stat_output_ctx *out,
685 struct rblist *metric_events)
686{
687 typedef void (*stat_print_function_t)(struct perf_stat_config *config,
688 const struct evsel *evsel,
689 int aggr_idx, double misses,
690 struct perf_stat_output_ctx *out);
691 static const stat_print_function_t stat_print_function[STAT_MAX] = {
692 [STAT_INSTRUCTIONS] = print_instructions,
693 [STAT_BRANCH_MISS] = print_branch_miss,
694 [STAT_L1D_MISS] = print_l1d_miss,
695 [STAT_L1I_MISS] = print_l1i_miss,
696 [STAT_DTLB_MISS] = print_dtlb_miss,
697 [STAT_ITLB_MISS] = print_itlb_miss,
698 [STAT_LL_MISS] = print_ll_miss,
699 [STAT_CACHE_MISSES] = print_cache_miss,
700 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
701 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
702 [STAT_CYCLES] = print_cycles,
703 [STAT_NSECS] = print_nsecs,
704 };
705 print_metric_t print_metric = out->print_metric;
706 void *ctxp = out->ctx;
707 int num = 1;
708
709 if (config->iostat_run) {
710 iostat_print_metric(config, evsel, out);
711 } else {
712 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
713
714 if (fn)
715 fn(config, evsel, aggr_idx, avg, out);
716 else {
717 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
718
719 if (nsecs) {
720 char unit = ' ';
721 char unit_buf[10] = "/sec";
722 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
723 &unit);
724
725 if (unit != ' ')
726 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
727 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
728 unit_buf, ratio);
729 } else {
730 num = 0;
731 }
732 }
733 }
734
735 perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
736 &num, NULL, out, metric_events);
737
738 if (num == 0) {
739 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN,
740 /*fmt=*/NULL, /*unit=*/NULL, 0);
741 }
742}
743
744/**
745 * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
746 * if it's not running or not the metric event.
747 */
748bool perf_stat__skip_metric_event(struct evsel *evsel,
749 struct rblist *metric_events,
750 u64 ena, u64 run)
751{
752 if (!evsel->default_metricgroup)
753 return false;
754
755 if (!ena || !run)
756 return true;
757
758 return !metricgroup__lookup(metric_events, evsel, false);
759}