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(struct evsel **metric_events,
359 struct metric_ref *metric_refs,
360 struct expr_parse_ctx *pctx,
361 int aggr_idx)
362{
363 int i;
364
365 for (i = 0; metric_events[i]; i++) {
366 char *n;
367 double val;
368 int source_count = 0;
369
370 if (evsel__is_tool(metric_events[i])) {
371 struct stats *stats;
372 double scale;
373
374 switch (metric_events[i]->tool_event) {
375 case PERF_TOOL_DURATION_TIME:
376 stats = &walltime_nsecs_stats;
377 scale = 1e-9;
378 break;
379 case PERF_TOOL_USER_TIME:
380 stats = &ru_stats.ru_utime_usec_stat;
381 scale = 1e-6;
382 break;
383 case PERF_TOOL_SYSTEM_TIME:
384 stats = &ru_stats.ru_stime_usec_stat;
385 scale = 1e-6;
386 break;
387 case PERF_TOOL_NONE:
388 pr_err("Invalid tool event 'none'");
389 abort();
390 case PERF_TOOL_MAX:
391 pr_err("Invalid tool event 'max'");
392 abort();
393 default:
394 pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
395 abort();
396 }
397 val = avg_stats(stats) * scale;
398 source_count = 1;
399 } else {
400 struct perf_stat_evsel *ps = metric_events[i]->stats;
401 struct perf_stat_aggr *aggr = &ps->aggr[aggr_idx];
402
403 if (!aggr)
404 break;
405
406 if (!metric_events[i]->supported) {
407 /*
408 * Not supported events will have a count of 0,
409 * which can be confusing in a
410 * metric. Explicitly set the value to NAN. Not
411 * counted events (enable time of 0) are read as
412 * 0.
413 */
414 val = NAN;
415 source_count = 0;
416 } else {
417 /*
418 * If an event was scaled during stat gathering,
419 * reverse the scale before computing the
420 * metric.
421 */
422 val = aggr->counts.val * (1.0 / metric_events[i]->scale);
423 source_count = evsel__source_count(metric_events[i]);
424 }
425 }
426 n = strdup(evsel__metric_id(metric_events[i]));
427 if (!n)
428 return -ENOMEM;
429
430 expr__add_id_val_source_count(pctx, n, val, source_count);
431 }
432
433 for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
434 int ret = expr__add_ref(pctx, &metric_refs[j]);
435
436 if (ret)
437 return ret;
438 }
439
440 return i;
441}
442
443static void generic_metric(struct perf_stat_config *config,
444 const char *metric_expr,
445 const char *metric_threshold,
446 struct evsel **metric_events,
447 struct metric_ref *metric_refs,
448 char *name,
449 const char *metric_name,
450 const char *metric_unit,
451 int runtime,
452 int aggr_idx,
453 struct perf_stat_output_ctx *out)
454{
455 print_metric_t print_metric = out->print_metric;
456 struct expr_parse_ctx *pctx;
457 double ratio, scale, threshold;
458 int i;
459 void *ctxp = out->ctx;
460 const char *color = NULL;
461
462 pctx = expr__ctx_new();
463 if (!pctx)
464 return;
465
466 if (config->user_requested_cpu_list)
467 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
468 pctx->sctx.runtime = runtime;
469 pctx->sctx.system_wide = config->system_wide;
470 i = prepare_metric(metric_events, metric_refs, pctx, aggr_idx);
471 if (i < 0) {
472 expr__ctx_free(pctx);
473 return;
474 }
475 if (!metric_events[i]) {
476 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
477 char *unit;
478 char metric_bf[64];
479
480 if (metric_threshold &&
481 expr__parse(&threshold, pctx, metric_threshold) == 0 &&
482 !isnan(threshold)) {
483 color = fpclassify(threshold) == FP_ZERO
484 ? PERF_COLOR_GREEN : PERF_COLOR_RED;
485 }
486
487 if (metric_unit && metric_name) {
488 if (perf_pmu__convert_scale(metric_unit,
489 &unit, &scale) >= 0) {
490 ratio *= scale;
491 }
492 if (strstr(metric_expr, "?"))
493 scnprintf(metric_bf, sizeof(metric_bf),
494 "%s %s_%d", unit, metric_name, runtime);
495 else
496 scnprintf(metric_bf, sizeof(metric_bf),
497 "%s %s", unit, metric_name);
498
499 print_metric(config, ctxp, color, "%8.1f",
500 metric_bf, ratio);
501 } else {
502 print_metric(config, ctxp, color, "%8.2f",
503 metric_name ?
504 metric_name :
505 out->force_header ? name : "",
506 ratio);
507 }
508 } else {
509 print_metric(config, ctxp, color, /*unit=*/NULL,
510 out->force_header ?
511 (metric_name ? metric_name : name) : "", 0);
512 }
513 } else {
514 print_metric(config, ctxp, color, /*unit=*/NULL,
515 out->force_header ?
516 (metric_name ? metric_name : name) : "", 0);
517 }
518
519 expr__ctx_free(pctx);
520}
521
522double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
523{
524 struct expr_parse_ctx *pctx;
525 double ratio = 0.0;
526
527 pctx = expr__ctx_new();
528 if (!pctx)
529 return NAN;
530
531 if (prepare_metric(mexp->metric_events, mexp->metric_refs, pctx, aggr_idx) < 0)
532 goto out;
533
534 if (expr__parse(&ratio, pctx, mexp->metric_expr))
535 ratio = 0.0;
536
537out:
538 expr__ctx_free(pctx);
539 return ratio;
540}
541
542static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
543 struct evsel *evsel,
544 void *ctxp,
545 const char *name,
546 struct perf_stat_output_ctx *out)
547{
548 bool need_full_name = perf_pmus__num_core_pmus() > 1;
549 static const char *last_name;
550 static const char *last_pmu;
551 char full_name[64];
552
553 /*
554 * A metricgroup may have several metric events,
555 * e.g.,TopdownL1 on e-core of ADL.
556 * The name has been output by the first metric
557 * event. Only align with other metics from
558 * different metric events.
559 */
560 if (last_name && !strcmp(last_name, name)) {
561 if (!need_full_name || !strcmp(last_pmu, evsel->pmu_name)) {
562 out->print_metricgroup_header(config, ctxp, NULL);
563 return;
564 }
565 }
566
567 if (need_full_name)
568 scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu_name);
569 else
570 scnprintf(full_name, sizeof(full_name), "%s", name);
571
572 out->print_metricgroup_header(config, ctxp, full_name);
573
574 last_name = name;
575 last_pmu = evsel->pmu_name;
576}
577
578/**
579 * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
580 * For the non-default, all metrics associated
581 * with the evsel are printed.
582 * For the default mode, only the metrics from
583 * the same metricgroup and the name of the
584 * metricgroup are printed. To print the metrics
585 * from the next metricgroup (if available),
586 * invoke the function with correspoinding
587 * metric_expr.
588 */
589void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
590 struct evsel *evsel,
591 int aggr_idx,
592 int *num,
593 void *from,
594 struct perf_stat_output_ctx *out,
595 struct rblist *metric_events)
596{
597 struct metric_event *me;
598 struct metric_expr *mexp = from;
599 void *ctxp = out->ctx;
600 bool header_printed = false;
601 const char *name = NULL;
602
603 me = metricgroup__lookup(metric_events, evsel, false);
604 if (me == NULL)
605 return NULL;
606
607 if (!mexp)
608 mexp = list_first_entry(&me->head, typeof(*mexp), nd);
609
610 list_for_each_entry_from(mexp, &me->head, nd) {
611 /* Print the display name of the Default metricgroup */
612 if (!config->metric_only && me->is_default) {
613 if (!name)
614 name = mexp->default_metricgroup_name;
615 /*
616 * Two or more metricgroup may share the same metric
617 * event, e.g., TopdownL1 and TopdownL2 on SPR.
618 * Return and print the prefix, e.g., noise, running
619 * for the next metricgroup.
620 */
621 if (strcmp(name, mexp->default_metricgroup_name))
622 return (void *)mexp;
623 /* Only print the name of the metricgroup once */
624 if (!header_printed) {
625 header_printed = true;
626 perf_stat__print_metricgroup_header(config, evsel, ctxp,
627 name, out);
628 }
629 }
630
631 if ((*num)++ > 0)
632 out->new_line(config, ctxp);
633 generic_metric(config, mexp->metric_expr, mexp->metric_threshold,
634 mexp->metric_events, mexp->metric_refs, evsel->name,
635 mexp->metric_name, mexp->metric_unit, mexp->runtime,
636 aggr_idx, out);
637 }
638
639 return NULL;
640}
641
642void perf_stat__print_shadow_stats(struct perf_stat_config *config,
643 struct evsel *evsel,
644 double avg, int aggr_idx,
645 struct perf_stat_output_ctx *out,
646 struct rblist *metric_events)
647{
648 typedef void (*stat_print_function_t)(struct perf_stat_config *config,
649 const struct evsel *evsel,
650 int aggr_idx, double misses,
651 struct perf_stat_output_ctx *out);
652 static const stat_print_function_t stat_print_function[STAT_MAX] = {
653 [STAT_INSTRUCTIONS] = print_instructions,
654 [STAT_BRANCH_MISS] = print_branch_miss,
655 [STAT_L1D_MISS] = print_l1d_miss,
656 [STAT_L1I_MISS] = print_l1i_miss,
657 [STAT_DTLB_MISS] = print_dtlb_miss,
658 [STAT_ITLB_MISS] = print_itlb_miss,
659 [STAT_LL_MISS] = print_ll_miss,
660 [STAT_CACHE_MISSES] = print_cache_miss,
661 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
662 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
663 [STAT_CYCLES] = print_cycles,
664 [STAT_NSECS] = print_nsecs,
665 };
666 print_metric_t print_metric = out->print_metric;
667 void *ctxp = out->ctx;
668 int num = 1;
669
670 if (config->iostat_run) {
671 iostat_print_metric(config, evsel, out);
672 } else {
673 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
674
675 if (fn)
676 fn(config, evsel, aggr_idx, avg, out);
677 else {
678 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
679
680 if (nsecs) {
681 char unit = ' ';
682 char unit_buf[10] = "/sec";
683 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
684 &unit);
685
686 if (unit != ' ')
687 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
688 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
689 } else
690 num = 0;
691 }
692 }
693
694 perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
695 &num, NULL, out, metric_events);
696
697 if (num == 0)
698 print_metric(config, ctxp, NULL, NULL, NULL, 0);
699}
700
701/**
702 * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
703 * if it's not running or not the metric event.
704 */
705bool perf_stat__skip_metric_event(struct evsel *evsel,
706 struct rblist *metric_events,
707 u64 ena, u64 run)
708{
709 if (!evsel->default_metricgroup)
710 return false;
711
712 if (!ena || !run)
713 return true;
714
715 return !metricgroup__lookup(metric_events, evsel, false);
716}
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
19/*
20 * AGGR_GLOBAL: Use CPU 0
21 * AGGR_SOCKET: Use first CPU of socket
22 * AGGR_DIE: Use first CPU of die
23 * AGGR_CORE: Use first CPU of core
24 * AGGR_NONE: Use matching CPU
25 * AGGR_THREAD: Not supported?
26 */
27
28struct runtime_stat rt_stat;
29struct stats walltime_nsecs_stats;
30struct rusage_stats ru_stats;
31
32struct saved_value {
33 struct rb_node rb_node;
34 struct evsel *evsel;
35 enum stat_type type;
36 int ctx;
37 int map_idx; /* cpu or thread map index */
38 struct cgroup *cgrp;
39 struct stats stats;
40 u64 metric_total;
41 int metric_other;
42};
43
44static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
45{
46 struct saved_value *a = container_of(rb_node,
47 struct saved_value,
48 rb_node);
49 const struct saved_value *b = entry;
50
51 if (a->map_idx != b->map_idx)
52 return a->map_idx - b->map_idx;
53
54 /*
55 * Previously the rbtree was used to link generic metrics.
56 * The keys were evsel/cpu. Now the rbtree is extended to support
57 * per-thread shadow stats. For shadow stats case, the keys
58 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
59 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
60 */
61 if (a->type != b->type)
62 return a->type - b->type;
63
64 if (a->ctx != b->ctx)
65 return a->ctx - b->ctx;
66
67 if (a->cgrp != b->cgrp)
68 return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1;
69
70 if (a->evsel == b->evsel)
71 return 0;
72 if ((char *)a->evsel < (char *)b->evsel)
73 return -1;
74 return +1;
75}
76
77static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
78 const void *entry)
79{
80 struct saved_value *nd = malloc(sizeof(struct saved_value));
81
82 if (!nd)
83 return NULL;
84 memcpy(nd, entry, sizeof(struct saved_value));
85 return &nd->rb_node;
86}
87
88static void saved_value_delete(struct rblist *rblist __maybe_unused,
89 struct rb_node *rb_node)
90{
91 struct saved_value *v;
92
93 BUG_ON(!rb_node);
94 v = container_of(rb_node, struct saved_value, rb_node);
95 free(v);
96}
97
98static struct saved_value *saved_value_lookup(struct evsel *evsel,
99 int map_idx,
100 bool create,
101 enum stat_type type,
102 int ctx,
103 struct runtime_stat *st,
104 struct cgroup *cgrp)
105{
106 struct rblist *rblist;
107 struct rb_node *nd;
108 struct saved_value dm = {
109 .map_idx = map_idx,
110 .evsel = evsel,
111 .type = type,
112 .ctx = ctx,
113 .cgrp = cgrp,
114 };
115
116 rblist = &st->value_list;
117
118 /* don't use context info for clock events */
119 if (type == STAT_NSECS)
120 dm.ctx = 0;
121
122 nd = rblist__find(rblist, &dm);
123 if (nd)
124 return container_of(nd, struct saved_value, rb_node);
125 if (create) {
126 rblist__add_node(rblist, &dm);
127 nd = rblist__find(rblist, &dm);
128 if (nd)
129 return container_of(nd, struct saved_value, rb_node);
130 }
131 return NULL;
132}
133
134void runtime_stat__init(struct runtime_stat *st)
135{
136 struct rblist *rblist = &st->value_list;
137
138 rblist__init(rblist);
139 rblist->node_cmp = saved_value_cmp;
140 rblist->node_new = saved_value_new;
141 rblist->node_delete = saved_value_delete;
142}
143
144void runtime_stat__exit(struct runtime_stat *st)
145{
146 rblist__exit(&st->value_list);
147}
148
149void perf_stat__init_shadow_stats(void)
150{
151 runtime_stat__init(&rt_stat);
152}
153
154static int evsel_context(struct evsel *evsel)
155{
156 int ctx = 0;
157
158 if (evsel->core.attr.exclude_kernel)
159 ctx |= CTX_BIT_KERNEL;
160 if (evsel->core.attr.exclude_user)
161 ctx |= CTX_BIT_USER;
162 if (evsel->core.attr.exclude_hv)
163 ctx |= CTX_BIT_HV;
164 if (evsel->core.attr.exclude_host)
165 ctx |= CTX_BIT_HOST;
166 if (evsel->core.attr.exclude_idle)
167 ctx |= CTX_BIT_IDLE;
168
169 return ctx;
170}
171
172static void reset_stat(struct runtime_stat *st)
173{
174 struct rblist *rblist;
175 struct rb_node *pos, *next;
176
177 rblist = &st->value_list;
178 next = rb_first_cached(&rblist->entries);
179 while (next) {
180 pos = next;
181 next = rb_next(pos);
182 memset(&container_of(pos, struct saved_value, rb_node)->stats,
183 0,
184 sizeof(struct stats));
185 }
186}
187
188void perf_stat__reset_shadow_stats(void)
189{
190 reset_stat(&rt_stat);
191 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
192 memset(&ru_stats, 0, sizeof(ru_stats));
193}
194
195void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
196{
197 reset_stat(st);
198}
199
200struct runtime_stat_data {
201 int ctx;
202 struct cgroup *cgrp;
203};
204
205static void update_runtime_stat(struct runtime_stat *st,
206 enum stat_type type,
207 int map_idx, u64 count,
208 struct runtime_stat_data *rsd)
209{
210 struct saved_value *v = saved_value_lookup(NULL, map_idx, true, type,
211 rsd->ctx, st, rsd->cgrp);
212
213 if (v)
214 update_stats(&v->stats, count);
215}
216
217/*
218 * Update various tracking values we maintain to print
219 * more semantic information such as miss/hit ratios,
220 * instruction rates, etc:
221 */
222void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
223 int map_idx, struct runtime_stat *st)
224{
225 u64 count_ns = count;
226 struct saved_value *v;
227 struct runtime_stat_data rsd = {
228 .ctx = evsel_context(counter),
229 .cgrp = counter->cgrp,
230 };
231
232 count *= counter->scale;
233
234 if (evsel__is_clock(counter))
235 update_runtime_stat(st, STAT_NSECS, map_idx, count_ns, &rsd);
236 else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
237 update_runtime_stat(st, STAT_CYCLES, map_idx, count, &rsd);
238 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
239 update_runtime_stat(st, STAT_CYCLES_IN_TX, map_idx, count, &rsd);
240 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
241 update_runtime_stat(st, STAT_TRANSACTION, map_idx, count, &rsd);
242 else if (perf_stat_evsel__is(counter, ELISION_START))
243 update_runtime_stat(st, STAT_ELISION, map_idx, count, &rsd);
244 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
245 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
246 map_idx, count, &rsd);
247 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
248 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
249 map_idx, count, &rsd);
250 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
251 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
252 map_idx, count, &rsd);
253 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
254 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
255 map_idx, count, &rsd);
256 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
257 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
258 map_idx, count, &rsd);
259 else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
260 update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
261 map_idx, count, &rsd);
262 else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
263 update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
264 map_idx, count, &rsd);
265 else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
266 update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
267 map_idx, count, &rsd);
268 else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
269 update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
270 map_idx, count, &rsd);
271 else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS))
272 update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS,
273 map_idx, count, &rsd);
274 else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT))
275 update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT,
276 map_idx, count, &rsd);
277 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT))
278 update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT,
279 map_idx, count, &rsd);
280 else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND))
281 update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND,
282 map_idx, count, &rsd);
283 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
284 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
285 map_idx, count, &rsd);
286 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
287 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
288 map_idx, count, &rsd);
289 else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
290 update_runtime_stat(st, STAT_BRANCHES, map_idx, count, &rsd);
291 else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
292 update_runtime_stat(st, STAT_CACHEREFS, map_idx, count, &rsd);
293 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
294 update_runtime_stat(st, STAT_L1_DCACHE, map_idx, count, &rsd);
295 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
296 update_runtime_stat(st, STAT_L1_ICACHE, map_idx, count, &rsd);
297 else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
298 update_runtime_stat(st, STAT_LL_CACHE, map_idx, count, &rsd);
299 else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
300 update_runtime_stat(st, STAT_DTLB_CACHE, map_idx, count, &rsd);
301 else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
302 update_runtime_stat(st, STAT_ITLB_CACHE, map_idx, count, &rsd);
303 else if (perf_stat_evsel__is(counter, SMI_NUM))
304 update_runtime_stat(st, STAT_SMI_NUM, map_idx, count, &rsd);
305 else if (perf_stat_evsel__is(counter, APERF))
306 update_runtime_stat(st, STAT_APERF, map_idx, count, &rsd);
307
308 if (counter->collect_stat) {
309 v = saved_value_lookup(counter, map_idx, true, STAT_NONE, 0, st,
310 rsd.cgrp);
311 update_stats(&v->stats, count);
312 if (counter->metric_leader)
313 v->metric_total += count;
314 } else if (counter->metric_leader) {
315 v = saved_value_lookup(counter->metric_leader,
316 map_idx, true, STAT_NONE, 0, st, rsd.cgrp);
317 v->metric_total += count;
318 v->metric_other++;
319 }
320}
321
322/* used for get_ratio_color() */
323enum grc_type {
324 GRC_STALLED_CYCLES_FE,
325 GRC_STALLED_CYCLES_BE,
326 GRC_CACHE_MISSES,
327 GRC_MAX_NR
328};
329
330static const char *get_ratio_color(enum grc_type type, double ratio)
331{
332 static const double grc_table[GRC_MAX_NR][3] = {
333 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
334 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
335 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
336 };
337 const char *color = PERF_COLOR_NORMAL;
338
339 if (ratio > grc_table[type][0])
340 color = PERF_COLOR_RED;
341 else if (ratio > grc_table[type][1])
342 color = PERF_COLOR_MAGENTA;
343 else if (ratio > grc_table[type][2])
344 color = PERF_COLOR_YELLOW;
345
346 return color;
347}
348
349static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
350 const char *name)
351{
352 struct evsel *c2;
353
354 evlist__for_each_entry (evsel_list, c2) {
355 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
356 return c2;
357 }
358 return NULL;
359}
360
361/* Mark MetricExpr target events and link events using them to them. */
362void perf_stat__collect_metric_expr(struct evlist *evsel_list)
363{
364 struct evsel *counter, *leader, **metric_events, *oc;
365 bool found;
366 struct expr_parse_ctx *ctx;
367 struct hashmap_entry *cur;
368 size_t bkt;
369 int i;
370
371 ctx = expr__ctx_new();
372 if (!ctx) {
373 pr_debug("expr__ctx_new failed");
374 return;
375 }
376 evlist__for_each_entry(evsel_list, counter) {
377 bool invalid = false;
378
379 leader = evsel__leader(counter);
380 if (!counter->metric_expr)
381 continue;
382
383 expr__ctx_clear(ctx);
384 metric_events = counter->metric_events;
385 if (!metric_events) {
386 if (expr__find_ids(counter->metric_expr,
387 counter->name,
388 ctx) < 0)
389 continue;
390
391 metric_events = calloc(sizeof(struct evsel *),
392 hashmap__size(ctx->ids) + 1);
393 if (!metric_events) {
394 expr__ctx_free(ctx);
395 return;
396 }
397 counter->metric_events = metric_events;
398 }
399
400 i = 0;
401 hashmap__for_each_entry(ctx->ids, cur, bkt) {
402 const char *metric_name = cur->pkey;
403
404 found = false;
405 if (leader) {
406 /* Search in group */
407 for_each_group_member (oc, leader) {
408 if (!strcasecmp(oc->name,
409 metric_name) &&
410 !oc->collect_stat) {
411 found = true;
412 break;
413 }
414 }
415 }
416 if (!found) {
417 /* Search ignoring groups */
418 oc = perf_stat__find_event(evsel_list,
419 metric_name);
420 }
421 if (!oc) {
422 /* Deduping one is good enough to handle duplicated PMUs. */
423 static char *printed;
424
425 /*
426 * Adding events automatically would be difficult, because
427 * it would risk creating groups that are not schedulable.
428 * perf stat doesn't understand all the scheduling constraints
429 * of events. So we ask the user instead to add the missing
430 * events.
431 */
432 if (!printed ||
433 strcasecmp(printed, metric_name)) {
434 fprintf(stderr,
435 "Add %s event to groups to get metric expression for %s\n",
436 metric_name,
437 counter->name);
438 free(printed);
439 printed = strdup(metric_name);
440 }
441 invalid = true;
442 continue;
443 }
444 metric_events[i++] = oc;
445 oc->collect_stat = true;
446 }
447 metric_events[i] = NULL;
448 if (invalid) {
449 free(metric_events);
450 counter->metric_events = NULL;
451 counter->metric_expr = NULL;
452 }
453 }
454 expr__ctx_free(ctx);
455}
456
457static double runtime_stat_avg(struct runtime_stat *st,
458 enum stat_type type, int map_idx,
459 struct runtime_stat_data *rsd)
460{
461 struct saved_value *v;
462
463 v = saved_value_lookup(NULL, map_idx, false, type, rsd->ctx, st, rsd->cgrp);
464 if (!v)
465 return 0.0;
466
467 return avg_stats(&v->stats);
468}
469
470static double runtime_stat_n(struct runtime_stat *st,
471 enum stat_type type, int map_idx,
472 struct runtime_stat_data *rsd)
473{
474 struct saved_value *v;
475
476 v = saved_value_lookup(NULL, map_idx, false, type, rsd->ctx, st, rsd->cgrp);
477 if (!v)
478 return 0.0;
479
480 return v->stats.n;
481}
482
483static void print_stalled_cycles_frontend(struct perf_stat_config *config,
484 int map_idx, double avg,
485 struct perf_stat_output_ctx *out,
486 struct runtime_stat *st,
487 struct runtime_stat_data *rsd)
488{
489 double total, ratio = 0.0;
490 const char *color;
491
492 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
493
494 if (total)
495 ratio = avg / total * 100.0;
496
497 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
498
499 if (ratio)
500 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
501 ratio);
502 else
503 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
504}
505
506static void print_stalled_cycles_backend(struct perf_stat_config *config,
507 int map_idx, double avg,
508 struct perf_stat_output_ctx *out,
509 struct runtime_stat *st,
510 struct runtime_stat_data *rsd)
511{
512 double total, ratio = 0.0;
513 const char *color;
514
515 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
516
517 if (total)
518 ratio = avg / total * 100.0;
519
520 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
521
522 out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
523}
524
525static void print_branch_misses(struct perf_stat_config *config,
526 int map_idx, double avg,
527 struct perf_stat_output_ctx *out,
528 struct runtime_stat *st,
529 struct runtime_stat_data *rsd)
530{
531 double total, ratio = 0.0;
532 const char *color;
533
534 total = runtime_stat_avg(st, STAT_BRANCHES, map_idx, rsd);
535
536 if (total)
537 ratio = avg / total * 100.0;
538
539 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
540
541 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
542}
543
544static void print_l1_dcache_misses(struct perf_stat_config *config,
545 int map_idx, double avg,
546 struct perf_stat_output_ctx *out,
547 struct runtime_stat *st,
548 struct runtime_stat_data *rsd)
549{
550 double total, ratio = 0.0;
551 const char *color;
552
553 total = runtime_stat_avg(st, STAT_L1_DCACHE, map_idx, rsd);
554
555 if (total)
556 ratio = avg / total * 100.0;
557
558 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
559
560 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
561}
562
563static void print_l1_icache_misses(struct perf_stat_config *config,
564 int map_idx, double avg,
565 struct perf_stat_output_ctx *out,
566 struct runtime_stat *st,
567 struct runtime_stat_data *rsd)
568{
569 double total, ratio = 0.0;
570 const char *color;
571
572 total = runtime_stat_avg(st, STAT_L1_ICACHE, map_idx, rsd);
573
574 if (total)
575 ratio = avg / total * 100.0;
576
577 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
578 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
579}
580
581static void print_dtlb_cache_misses(struct perf_stat_config *config,
582 int map_idx, double avg,
583 struct perf_stat_output_ctx *out,
584 struct runtime_stat *st,
585 struct runtime_stat_data *rsd)
586{
587 double total, ratio = 0.0;
588 const char *color;
589
590 total = runtime_stat_avg(st, STAT_DTLB_CACHE, map_idx, rsd);
591
592 if (total)
593 ratio = avg / total * 100.0;
594
595 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
596 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
597}
598
599static void print_itlb_cache_misses(struct perf_stat_config *config,
600 int map_idx, double avg,
601 struct perf_stat_output_ctx *out,
602 struct runtime_stat *st,
603 struct runtime_stat_data *rsd)
604{
605 double total, ratio = 0.0;
606 const char *color;
607
608 total = runtime_stat_avg(st, STAT_ITLB_CACHE, map_idx, rsd);
609
610 if (total)
611 ratio = avg / total * 100.0;
612
613 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
614 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
615}
616
617static void print_ll_cache_misses(struct perf_stat_config *config,
618 int map_idx, double avg,
619 struct perf_stat_output_ctx *out,
620 struct runtime_stat *st,
621 struct runtime_stat_data *rsd)
622{
623 double total, ratio = 0.0;
624 const char *color;
625
626 total = runtime_stat_avg(st, STAT_LL_CACHE, map_idx, rsd);
627
628 if (total)
629 ratio = avg / total * 100.0;
630
631 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
632 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
633}
634
635/*
636 * High level "TopDown" CPU core pipe line bottleneck break down.
637 *
638 * Basic concept following
639 * Yasin, A Top Down Method for Performance analysis and Counter architecture
640 * ISPASS14
641 *
642 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
643 *
644 * Frontend -> Backend -> Retiring
645 * BadSpeculation in addition means out of order execution that is thrown away
646 * (for example branch mispredictions)
647 * Frontend is instruction decoding.
648 * Backend is execution, like computation and accessing data in memory
649 * Retiring is good execution that is not directly bottlenecked
650 *
651 * The formulas are computed in slots.
652 * A slot is an entry in the pipeline each for the pipeline width
653 * (for example a 4-wide pipeline has 4 slots for each cycle)
654 *
655 * Formulas:
656 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
657 * TotalSlots
658 * Retiring = SlotsRetired / TotalSlots
659 * FrontendBound = FetchBubbles / TotalSlots
660 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
661 *
662 * The kernel provides the mapping to the low level CPU events and any scaling
663 * needed for the CPU pipeline width, for example:
664 *
665 * TotalSlots = Cycles * 4
666 *
667 * The scaling factor is communicated in the sysfs unit.
668 *
669 * In some cases the CPU may not be able to measure all the formulas due to
670 * missing events. In this case multiple formulas are combined, as possible.
671 *
672 * Full TopDown supports more levels to sub-divide each area: for example
673 * BackendBound into computing bound and memory bound. For now we only
674 * support Level 1 TopDown.
675 */
676
677static double sanitize_val(double x)
678{
679 if (x < 0 && x >= -0.02)
680 return 0.0;
681 return x;
682}
683
684static double td_total_slots(int map_idx, struct runtime_stat *st,
685 struct runtime_stat_data *rsd)
686{
687 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, map_idx, rsd);
688}
689
690static double td_bad_spec(int map_idx, struct runtime_stat *st,
691 struct runtime_stat_data *rsd)
692{
693 double bad_spec = 0;
694 double total_slots;
695 double total;
696
697 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, map_idx, rsd) -
698 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, map_idx, rsd) +
699 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, map_idx, rsd);
700
701 total_slots = td_total_slots(map_idx, st, rsd);
702 if (total_slots)
703 bad_spec = total / total_slots;
704 return sanitize_val(bad_spec);
705}
706
707static double td_retiring(int map_idx, struct runtime_stat *st,
708 struct runtime_stat_data *rsd)
709{
710 double retiring = 0;
711 double total_slots = td_total_slots(map_idx, st, rsd);
712 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
713 map_idx, rsd);
714
715 if (total_slots)
716 retiring = ret_slots / total_slots;
717 return retiring;
718}
719
720static double td_fe_bound(int map_idx, struct runtime_stat *st,
721 struct runtime_stat_data *rsd)
722{
723 double fe_bound = 0;
724 double total_slots = td_total_slots(map_idx, st, rsd);
725 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
726 map_idx, rsd);
727
728 if (total_slots)
729 fe_bound = fetch_bub / total_slots;
730 return fe_bound;
731}
732
733static double td_be_bound(int map_idx, struct runtime_stat *st,
734 struct runtime_stat_data *rsd)
735{
736 double sum = (td_fe_bound(map_idx, st, rsd) +
737 td_bad_spec(map_idx, st, rsd) +
738 td_retiring(map_idx, st, rsd));
739 if (sum == 0)
740 return 0;
741 return sanitize_val(1.0 - sum);
742}
743
744/*
745 * Kernel reports metrics multiplied with slots. To get back
746 * the ratios we need to recreate the sum.
747 */
748
749static double td_metric_ratio(int map_idx, enum stat_type type,
750 struct runtime_stat *stat,
751 struct runtime_stat_data *rsd)
752{
753 double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, map_idx, rsd) +
754 runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, map_idx, rsd) +
755 runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, map_idx, rsd) +
756 runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, map_idx, rsd);
757 double d = runtime_stat_avg(stat, type, map_idx, rsd);
758
759 if (sum)
760 return d / sum;
761 return 0;
762}
763
764/*
765 * ... but only if most of the values are actually available.
766 * We allow two missing.
767 */
768
769static bool full_td(int map_idx, struct runtime_stat *stat,
770 struct runtime_stat_data *rsd)
771{
772 int c = 0;
773
774 if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, map_idx, rsd) > 0)
775 c++;
776 if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, map_idx, rsd) > 0)
777 c++;
778 if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, map_idx, rsd) > 0)
779 c++;
780 if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, map_idx, rsd) > 0)
781 c++;
782 return c >= 2;
783}
784
785static void print_smi_cost(struct perf_stat_config *config, int map_idx,
786 struct perf_stat_output_ctx *out,
787 struct runtime_stat *st,
788 struct runtime_stat_data *rsd)
789{
790 double smi_num, aperf, cycles, cost = 0.0;
791 const char *color = NULL;
792
793 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, map_idx, rsd);
794 aperf = runtime_stat_avg(st, STAT_APERF, map_idx, rsd);
795 cycles = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
796
797 if ((cycles == 0) || (aperf == 0))
798 return;
799
800 if (smi_num)
801 cost = (aperf - cycles) / aperf * 100.00;
802
803 if (cost > 10)
804 color = PERF_COLOR_RED;
805 out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
806 out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
807}
808
809static int prepare_metric(struct evsel **metric_events,
810 struct metric_ref *metric_refs,
811 struct expr_parse_ctx *pctx,
812 int map_idx,
813 struct runtime_stat *st)
814{
815 double scale;
816 char *n;
817 int i, j, ret;
818
819 for (i = 0; metric_events[i]; i++) {
820 struct saved_value *v;
821 struct stats *stats;
822 u64 metric_total = 0;
823 int source_count;
824
825 if (evsel__is_tool(metric_events[i])) {
826 source_count = 1;
827 switch (metric_events[i]->tool_event) {
828 case PERF_TOOL_DURATION_TIME:
829 stats = &walltime_nsecs_stats;
830 scale = 1e-9;
831 break;
832 case PERF_TOOL_USER_TIME:
833 stats = &ru_stats.ru_utime_usec_stat;
834 scale = 1e-6;
835 break;
836 case PERF_TOOL_SYSTEM_TIME:
837 stats = &ru_stats.ru_stime_usec_stat;
838 scale = 1e-6;
839 break;
840 case PERF_TOOL_NONE:
841 pr_err("Invalid tool event 'none'");
842 abort();
843 case PERF_TOOL_MAX:
844 pr_err("Invalid tool event 'max'");
845 abort();
846 default:
847 pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
848 abort();
849 }
850 } else {
851 v = saved_value_lookup(metric_events[i], map_idx, false,
852 STAT_NONE, 0, st,
853 metric_events[i]->cgrp);
854 if (!v)
855 break;
856 stats = &v->stats;
857 /*
858 * If an event was scaled during stat gathering, reverse
859 * the scale before computing the metric.
860 */
861 scale = 1.0 / metric_events[i]->scale;
862
863 source_count = evsel__source_count(metric_events[i]);
864
865 if (v->metric_other)
866 metric_total = v->metric_total * scale;
867 }
868 n = strdup(evsel__metric_id(metric_events[i]));
869 if (!n)
870 return -ENOMEM;
871
872 expr__add_id_val_source_count(pctx, n,
873 metric_total ? : avg_stats(stats) * scale,
874 source_count);
875 }
876
877 for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
878 ret = expr__add_ref(pctx, &metric_refs[j]);
879 if (ret)
880 return ret;
881 }
882
883 return i;
884}
885
886static void generic_metric(struct perf_stat_config *config,
887 const char *metric_expr,
888 struct evsel **metric_events,
889 struct metric_ref *metric_refs,
890 char *name,
891 const char *metric_name,
892 const char *metric_unit,
893 int runtime,
894 int map_idx,
895 struct perf_stat_output_ctx *out,
896 struct runtime_stat *st)
897{
898 print_metric_t print_metric = out->print_metric;
899 struct expr_parse_ctx *pctx;
900 double ratio, scale;
901 int i;
902 void *ctxp = out->ctx;
903
904 pctx = expr__ctx_new();
905 if (!pctx)
906 return;
907
908 if (config->user_requested_cpu_list)
909 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
910 pctx->sctx.runtime = runtime;
911 pctx->sctx.system_wide = config->system_wide;
912 i = prepare_metric(metric_events, metric_refs, pctx, map_idx, st);
913 if (i < 0) {
914 expr__ctx_free(pctx);
915 return;
916 }
917 if (!metric_events[i]) {
918 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
919 char *unit;
920 char metric_bf[64];
921
922 if (metric_unit && metric_name) {
923 if (perf_pmu__convert_scale(metric_unit,
924 &unit, &scale) >= 0) {
925 ratio *= scale;
926 }
927 if (strstr(metric_expr, "?"))
928 scnprintf(metric_bf, sizeof(metric_bf),
929 "%s %s_%d", unit, metric_name, runtime);
930 else
931 scnprintf(metric_bf, sizeof(metric_bf),
932 "%s %s", unit, metric_name);
933
934 print_metric(config, ctxp, NULL, "%8.1f",
935 metric_bf, ratio);
936 } else {
937 print_metric(config, ctxp, NULL, "%8.2f",
938 metric_name ?
939 metric_name :
940 out->force_header ? name : "",
941 ratio);
942 }
943 } else {
944 print_metric(config, ctxp, NULL, NULL,
945 out->force_header ?
946 (metric_name ? metric_name : name) : "", 0);
947 }
948 } else {
949 print_metric(config, ctxp, NULL, NULL,
950 out->force_header ?
951 (metric_name ? metric_name : name) : "", 0);
952 }
953
954 expr__ctx_free(pctx);
955}
956
957double test_generic_metric(struct metric_expr *mexp, int map_idx, struct runtime_stat *st)
958{
959 struct expr_parse_ctx *pctx;
960 double ratio = 0.0;
961
962 pctx = expr__ctx_new();
963 if (!pctx)
964 return NAN;
965
966 if (prepare_metric(mexp->metric_events, mexp->metric_refs, pctx, map_idx, st) < 0)
967 goto out;
968
969 if (expr__parse(&ratio, pctx, mexp->metric_expr))
970 ratio = 0.0;
971
972out:
973 expr__ctx_free(pctx);
974 return ratio;
975}
976
977void perf_stat__print_shadow_stats(struct perf_stat_config *config,
978 struct evsel *evsel,
979 double avg, int map_idx,
980 struct perf_stat_output_ctx *out,
981 struct rblist *metric_events,
982 struct runtime_stat *st)
983{
984 void *ctxp = out->ctx;
985 print_metric_t print_metric = out->print_metric;
986 double total, ratio = 0.0, total2;
987 const char *color = NULL;
988 struct runtime_stat_data rsd = {
989 .ctx = evsel_context(evsel),
990 .cgrp = evsel->cgrp,
991 };
992 struct metric_event *me;
993 int num = 1;
994
995 if (config->iostat_run) {
996 iostat_print_metric(config, evsel, out);
997 } else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
998 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
999
1000 if (total) {
1001 ratio = avg / total;
1002 print_metric(config, ctxp, NULL, "%7.2f ",
1003 "insn per cycle", ratio);
1004 } else {
1005 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
1006 }
1007
1008 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, map_idx, &rsd);
1009
1010 total = max(total, runtime_stat_avg(st,
1011 STAT_STALLED_CYCLES_BACK,
1012 map_idx, &rsd));
1013
1014 if (total && avg) {
1015 out->new_line(config, ctxp);
1016 ratio = total / avg;
1017 print_metric(config, ctxp, NULL, "%7.2f ",
1018 "stalled cycles per insn",
1019 ratio);
1020 }
1021 } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
1022 if (runtime_stat_n(st, STAT_BRANCHES, map_idx, &rsd) != 0)
1023 print_branch_misses(config, map_idx, avg, out, st, &rsd);
1024 else
1025 print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
1026 } else if (
1027 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1028 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1029 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1030 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1031
1032 if (runtime_stat_n(st, STAT_L1_DCACHE, map_idx, &rsd) != 0)
1033 print_l1_dcache_misses(config, map_idx, avg, out, st, &rsd);
1034 else
1035 print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
1036 } else if (
1037 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1038 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1039 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1040 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1041
1042 if (runtime_stat_n(st, STAT_L1_ICACHE, map_idx, &rsd) != 0)
1043 print_l1_icache_misses(config, map_idx, avg, out, st, &rsd);
1044 else
1045 print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
1046 } else if (
1047 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1048 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1049 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1050 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1051
1052 if (runtime_stat_n(st, STAT_DTLB_CACHE, map_idx, &rsd) != 0)
1053 print_dtlb_cache_misses(config, map_idx, avg, out, st, &rsd);
1054 else
1055 print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
1056 } else if (
1057 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1058 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1059 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1060 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1061
1062 if (runtime_stat_n(st, STAT_ITLB_CACHE, map_idx, &rsd) != 0)
1063 print_itlb_cache_misses(config, map_idx, avg, out, st, &rsd);
1064 else
1065 print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
1066 } else if (
1067 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1068 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL |
1069 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1070 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1071
1072 if (runtime_stat_n(st, STAT_LL_CACHE, map_idx, &rsd) != 0)
1073 print_ll_cache_misses(config, map_idx, avg, out, st, &rsd);
1074 else
1075 print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
1076 } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
1077 total = runtime_stat_avg(st, STAT_CACHEREFS, map_idx, &rsd);
1078
1079 if (total)
1080 ratio = avg * 100 / total;
1081
1082 if (runtime_stat_n(st, STAT_CACHEREFS, map_idx, &rsd) != 0)
1083 print_metric(config, ctxp, NULL, "%8.3f %%",
1084 "of all cache refs", ratio);
1085 else
1086 print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
1087 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1088 print_stalled_cycles_frontend(config, map_idx, avg, out, st, &rsd);
1089 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1090 print_stalled_cycles_backend(config, map_idx, avg, out, st, &rsd);
1091 } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1092 total = runtime_stat_avg(st, STAT_NSECS, map_idx, &rsd);
1093
1094 if (total) {
1095 ratio = avg / total;
1096 print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
1097 } else {
1098 print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
1099 }
1100 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1101 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
1102
1103 if (total)
1104 print_metric(config, ctxp, NULL,
1105 "%7.2f%%", "transactional cycles",
1106 100.0 * (avg / total));
1107 else
1108 print_metric(config, ctxp, NULL, NULL, "transactional cycles",
1109 0);
1110 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1111 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
1112 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1113
1114 if (total2 < avg)
1115 total2 = avg;
1116 if (total)
1117 print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1118 100.0 * ((total2-avg) / total));
1119 else
1120 print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1121 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1122 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1123
1124 if (avg)
1125 ratio = total / avg;
1126
1127 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, map_idx, &rsd) != 0)
1128 print_metric(config, ctxp, NULL, "%8.0f",
1129 "cycles / transaction", ratio);
1130 else
1131 print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1132 0);
1133 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1134 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1135
1136 if (avg)
1137 ratio = total / avg;
1138
1139 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1140 } else if (evsel__is_clock(evsel)) {
1141 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1142 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1143 avg / (ratio * evsel->scale));
1144 else
1145 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1146 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1147 double fe_bound = td_fe_bound(map_idx, st, &rsd);
1148
1149 if (fe_bound > 0.2)
1150 color = PERF_COLOR_RED;
1151 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1152 fe_bound * 100.);
1153 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1154 double retiring = td_retiring(map_idx, st, &rsd);
1155
1156 if (retiring > 0.7)
1157 color = PERF_COLOR_GREEN;
1158 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1159 retiring * 100.);
1160 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1161 double bad_spec = td_bad_spec(map_idx, st, &rsd);
1162
1163 if (bad_spec > 0.1)
1164 color = PERF_COLOR_RED;
1165 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1166 bad_spec * 100.);
1167 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1168 double be_bound = td_be_bound(map_idx, st, &rsd);
1169 const char *name = "backend bound";
1170 static int have_recovery_bubbles = -1;
1171
1172 /* In case the CPU does not support topdown-recovery-bubbles */
1173 if (have_recovery_bubbles < 0)
1174 have_recovery_bubbles = pmu_have_event("cpu",
1175 "topdown-recovery-bubbles");
1176 if (!have_recovery_bubbles)
1177 name = "backend bound/bad spec";
1178
1179 if (be_bound > 0.2)
1180 color = PERF_COLOR_RED;
1181 if (td_total_slots(map_idx, st, &rsd) > 0)
1182 print_metric(config, ctxp, color, "%8.1f%%", name,
1183 be_bound * 100.);
1184 else
1185 print_metric(config, ctxp, NULL, NULL, name, 0);
1186 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1187 full_td(map_idx, st, &rsd)) {
1188 double retiring = td_metric_ratio(map_idx,
1189 STAT_TOPDOWN_RETIRING, st,
1190 &rsd);
1191 if (retiring > 0.7)
1192 color = PERF_COLOR_GREEN;
1193 print_metric(config, ctxp, color, "%8.1f%%", "Retiring",
1194 retiring * 100.);
1195 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1196 full_td(map_idx, st, &rsd)) {
1197 double fe_bound = td_metric_ratio(map_idx,
1198 STAT_TOPDOWN_FE_BOUND, st,
1199 &rsd);
1200 if (fe_bound > 0.2)
1201 color = PERF_COLOR_RED;
1202 print_metric(config, ctxp, color, "%8.1f%%", "Frontend Bound",
1203 fe_bound * 100.);
1204 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1205 full_td(map_idx, st, &rsd)) {
1206 double be_bound = td_metric_ratio(map_idx,
1207 STAT_TOPDOWN_BE_BOUND, st,
1208 &rsd);
1209 if (be_bound > 0.2)
1210 color = PERF_COLOR_RED;
1211 print_metric(config, ctxp, color, "%8.1f%%", "Backend Bound",
1212 be_bound * 100.);
1213 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1214 full_td(map_idx, st, &rsd)) {
1215 double bad_spec = td_metric_ratio(map_idx,
1216 STAT_TOPDOWN_BAD_SPEC, st,
1217 &rsd);
1218 if (bad_spec > 0.1)
1219 color = PERF_COLOR_RED;
1220 print_metric(config, ctxp, color, "%8.1f%%", "Bad Speculation",
1221 bad_spec * 100.);
1222 } else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) &&
1223 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1224 double retiring = td_metric_ratio(map_idx,
1225 STAT_TOPDOWN_RETIRING, st,
1226 &rsd);
1227 double heavy_ops = td_metric_ratio(map_idx,
1228 STAT_TOPDOWN_HEAVY_OPS, st,
1229 &rsd);
1230 double light_ops = retiring - heavy_ops;
1231
1232 if (retiring > 0.7 && heavy_ops > 0.1)
1233 color = PERF_COLOR_GREEN;
1234 print_metric(config, ctxp, color, "%8.1f%%", "Heavy Operations",
1235 heavy_ops * 100.);
1236 if (retiring > 0.7 && light_ops > 0.6)
1237 color = PERF_COLOR_GREEN;
1238 else
1239 color = NULL;
1240 print_metric(config, ctxp, color, "%8.1f%%", "Light Operations",
1241 light_ops * 100.);
1242 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) &&
1243 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1244 double bad_spec = td_metric_ratio(map_idx,
1245 STAT_TOPDOWN_BAD_SPEC, st,
1246 &rsd);
1247 double br_mis = td_metric_ratio(map_idx,
1248 STAT_TOPDOWN_BR_MISPREDICT, st,
1249 &rsd);
1250 double m_clears = bad_spec - br_mis;
1251
1252 if (bad_spec > 0.1 && br_mis > 0.05)
1253 color = PERF_COLOR_RED;
1254 print_metric(config, ctxp, color, "%8.1f%%", "Branch Mispredict",
1255 br_mis * 100.);
1256 if (bad_spec > 0.1 && m_clears > 0.05)
1257 color = PERF_COLOR_RED;
1258 else
1259 color = NULL;
1260 print_metric(config, ctxp, color, "%8.1f%%", "Machine Clears",
1261 m_clears * 100.);
1262 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) &&
1263 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1264 double fe_bound = td_metric_ratio(map_idx,
1265 STAT_TOPDOWN_FE_BOUND, st,
1266 &rsd);
1267 double fetch_lat = td_metric_ratio(map_idx,
1268 STAT_TOPDOWN_FETCH_LAT, st,
1269 &rsd);
1270 double fetch_bw = fe_bound - fetch_lat;
1271
1272 if (fe_bound > 0.2 && fetch_lat > 0.15)
1273 color = PERF_COLOR_RED;
1274 print_metric(config, ctxp, color, "%8.1f%%", "Fetch Latency",
1275 fetch_lat * 100.);
1276 if (fe_bound > 0.2 && fetch_bw > 0.1)
1277 color = PERF_COLOR_RED;
1278 else
1279 color = NULL;
1280 print_metric(config, ctxp, color, "%8.1f%%", "Fetch Bandwidth",
1281 fetch_bw * 100.);
1282 } else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) &&
1283 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1284 double be_bound = td_metric_ratio(map_idx,
1285 STAT_TOPDOWN_BE_BOUND, st,
1286 &rsd);
1287 double mem_bound = td_metric_ratio(map_idx,
1288 STAT_TOPDOWN_MEM_BOUND, st,
1289 &rsd);
1290 double core_bound = be_bound - mem_bound;
1291
1292 if (be_bound > 0.2 && mem_bound > 0.2)
1293 color = PERF_COLOR_RED;
1294 print_metric(config, ctxp, color, "%8.1f%%", "Memory Bound",
1295 mem_bound * 100.);
1296 if (be_bound > 0.2 && core_bound > 0.1)
1297 color = PERF_COLOR_RED;
1298 else
1299 color = NULL;
1300 print_metric(config, ctxp, color, "%8.1f%%", "Core Bound",
1301 core_bound * 100.);
1302 } else if (evsel->metric_expr) {
1303 generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1304 evsel->name, evsel->metric_name, NULL, 1,
1305 map_idx, out, st);
1306 } else if (runtime_stat_n(st, STAT_NSECS, map_idx, &rsd) != 0) {
1307 char unit = ' ';
1308 char unit_buf[10] = "/sec";
1309
1310 total = runtime_stat_avg(st, STAT_NSECS, map_idx, &rsd);
1311 if (total)
1312 ratio = convert_unit_double(1000000000.0 * avg / total, &unit);
1313
1314 if (unit != ' ')
1315 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1316 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1317 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1318 print_smi_cost(config, map_idx, out, st, &rsd);
1319 } else {
1320 num = 0;
1321 }
1322
1323 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1324 struct metric_expr *mexp;
1325
1326 list_for_each_entry (mexp, &me->head, nd) {
1327 if (num++ > 0)
1328 out->new_line(config, ctxp);
1329 generic_metric(config, mexp->metric_expr, mexp->metric_events,
1330 mexp->metric_refs, evsel->name, mexp->metric_name,
1331 mexp->metric_unit, mexp->runtime,
1332 map_idx, out, st);
1333 }
1334 }
1335 if (num == 0)
1336 print_metric(config, ctxp, NULL, NULL, NULL, 0);
1337}