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 <stdio.h>
3#include "evsel.h"
4#include "stat.h"
5#include "color.h"
6#include "pmu.h"
7#include "rblist.h"
8#include "evlist.h"
9#include "expr.h"
10#include "metricgroup.h"
11#include <linux/zalloc.h>
12
13/*
14 * AGGR_GLOBAL: Use CPU 0
15 * AGGR_SOCKET: Use first CPU of socket
16 * AGGR_DIE: Use first CPU of die
17 * AGGR_CORE: Use first CPU of core
18 * AGGR_NONE: Use matching CPU
19 * AGGR_THREAD: Not supported?
20 */
21
22struct runtime_stat rt_stat;
23struct stats walltime_nsecs_stats;
24
25struct saved_value {
26 struct rb_node rb_node;
27 struct evsel *evsel;
28 enum stat_type type;
29 int ctx;
30 int cpu;
31 struct runtime_stat *stat;
32 struct stats stats;
33 u64 metric_total;
34 int metric_other;
35};
36
37static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
38{
39 struct saved_value *a = container_of(rb_node,
40 struct saved_value,
41 rb_node);
42 const struct saved_value *b = entry;
43
44 if (a->cpu != b->cpu)
45 return a->cpu - b->cpu;
46
47 /*
48 * Previously the rbtree was used to link generic metrics.
49 * The keys were evsel/cpu. Now the rbtree is extended to support
50 * per-thread shadow stats. For shadow stats case, the keys
51 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
52 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
53 */
54 if (a->type != b->type)
55 return a->type - b->type;
56
57 if (a->ctx != b->ctx)
58 return a->ctx - b->ctx;
59
60 if (a->evsel == NULL && b->evsel == NULL) {
61 if (a->stat == b->stat)
62 return 0;
63
64 if ((char *)a->stat < (char *)b->stat)
65 return -1;
66
67 return 1;
68 }
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 cpu,
100 bool create,
101 enum stat_type type,
102 int ctx,
103 struct runtime_stat *st)
104{
105 struct rblist *rblist;
106 struct rb_node *nd;
107 struct saved_value dm = {
108 .cpu = cpu,
109 .evsel = evsel,
110 .type = type,
111 .ctx = ctx,
112 .stat = st,
113 };
114
115 rblist = &st->value_list;
116
117 nd = rblist__find(rblist, &dm);
118 if (nd)
119 return container_of(nd, struct saved_value, rb_node);
120 if (create) {
121 rblist__add_node(rblist, &dm);
122 nd = rblist__find(rblist, &dm);
123 if (nd)
124 return container_of(nd, struct saved_value, rb_node);
125 }
126 return NULL;
127}
128
129void runtime_stat__init(struct runtime_stat *st)
130{
131 struct rblist *rblist = &st->value_list;
132
133 rblist__init(rblist);
134 rblist->node_cmp = saved_value_cmp;
135 rblist->node_new = saved_value_new;
136 rblist->node_delete = saved_value_delete;
137}
138
139void runtime_stat__exit(struct runtime_stat *st)
140{
141 rblist__exit(&st->value_list);
142}
143
144void perf_stat__init_shadow_stats(void)
145{
146 runtime_stat__init(&rt_stat);
147}
148
149static int evsel_context(struct evsel *evsel)
150{
151 int ctx = 0;
152
153 if (evsel->core.attr.exclude_kernel)
154 ctx |= CTX_BIT_KERNEL;
155 if (evsel->core.attr.exclude_user)
156 ctx |= CTX_BIT_USER;
157 if (evsel->core.attr.exclude_hv)
158 ctx |= CTX_BIT_HV;
159 if (evsel->core.attr.exclude_host)
160 ctx |= CTX_BIT_HOST;
161 if (evsel->core.attr.exclude_idle)
162 ctx |= CTX_BIT_IDLE;
163
164 return ctx;
165}
166
167static void reset_stat(struct runtime_stat *st)
168{
169 struct rblist *rblist;
170 struct rb_node *pos, *next;
171
172 rblist = &st->value_list;
173 next = rb_first_cached(&rblist->entries);
174 while (next) {
175 pos = next;
176 next = rb_next(pos);
177 memset(&container_of(pos, struct saved_value, rb_node)->stats,
178 0,
179 sizeof(struct stats));
180 }
181}
182
183void perf_stat__reset_shadow_stats(void)
184{
185 reset_stat(&rt_stat);
186 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
187}
188
189void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
190{
191 reset_stat(st);
192}
193
194static void update_runtime_stat(struct runtime_stat *st,
195 enum stat_type type,
196 int ctx, int cpu, u64 count)
197{
198 struct saved_value *v = saved_value_lookup(NULL, cpu, true,
199 type, ctx, st);
200
201 if (v)
202 update_stats(&v->stats, count);
203}
204
205/*
206 * Update various tracking values we maintain to print
207 * more semantic information such as miss/hit ratios,
208 * instruction rates, etc:
209 */
210void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
211 int cpu, struct runtime_stat *st)
212{
213 int ctx = evsel_context(counter);
214 u64 count_ns = count;
215 struct saved_value *v;
216
217 count *= counter->scale;
218
219 if (evsel__is_clock(counter))
220 update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns);
221 else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
222 update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count);
223 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
224 update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count);
225 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
226 update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count);
227 else if (perf_stat_evsel__is(counter, ELISION_START))
228 update_runtime_stat(st, STAT_ELISION, ctx, cpu, count);
229 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
230 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
231 ctx, cpu, count);
232 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
233 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
234 ctx, cpu, count);
235 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
236 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
237 ctx, cpu, count);
238 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
239 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
240 ctx, cpu, count);
241 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
242 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
243 ctx, cpu, count);
244 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
245 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
246 ctx, cpu, count);
247 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
248 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
249 ctx, cpu, count);
250 else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
251 update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
252 else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
253 update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
254 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
255 update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
256 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
257 update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
258 else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
259 update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
260 else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
261 update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
262 else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
263 update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count);
264 else if (perf_stat_evsel__is(counter, SMI_NUM))
265 update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count);
266 else if (perf_stat_evsel__is(counter, APERF))
267 update_runtime_stat(st, STAT_APERF, ctx, cpu, count);
268
269 if (counter->collect_stat) {
270 v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st);
271 update_stats(&v->stats, count);
272 if (counter->metric_leader)
273 v->metric_total += count;
274 } else if (counter->metric_leader) {
275 v = saved_value_lookup(counter->metric_leader,
276 cpu, true, STAT_NONE, 0, st);
277 v->metric_total += count;
278 v->metric_other++;
279 }
280}
281
282/* used for get_ratio_color() */
283enum grc_type {
284 GRC_STALLED_CYCLES_FE,
285 GRC_STALLED_CYCLES_BE,
286 GRC_CACHE_MISSES,
287 GRC_MAX_NR
288};
289
290static const char *get_ratio_color(enum grc_type type, double ratio)
291{
292 static const double grc_table[GRC_MAX_NR][3] = {
293 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
294 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
295 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
296 };
297 const char *color = PERF_COLOR_NORMAL;
298
299 if (ratio > grc_table[type][0])
300 color = PERF_COLOR_RED;
301 else if (ratio > grc_table[type][1])
302 color = PERF_COLOR_MAGENTA;
303 else if (ratio > grc_table[type][2])
304 color = PERF_COLOR_YELLOW;
305
306 return color;
307}
308
309static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
310 const char *name)
311{
312 struct evsel *c2;
313
314 evlist__for_each_entry (evsel_list, c2) {
315 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
316 return c2;
317 }
318 return NULL;
319}
320
321/* Mark MetricExpr target events and link events using them to them. */
322void perf_stat__collect_metric_expr(struct evlist *evsel_list)
323{
324 struct evsel *counter, *leader, **metric_events, *oc;
325 bool found;
326 struct expr_parse_ctx ctx;
327 struct hashmap_entry *cur;
328 size_t bkt;
329 int i;
330
331 expr__ctx_init(&ctx);
332 evlist__for_each_entry(evsel_list, counter) {
333 bool invalid = false;
334
335 leader = counter->leader;
336 if (!counter->metric_expr)
337 continue;
338
339 expr__ctx_clear(&ctx);
340 metric_events = counter->metric_events;
341 if (!metric_events) {
342 if (expr__find_other(counter->metric_expr,
343 counter->name,
344 &ctx, 1) < 0)
345 continue;
346
347 metric_events = calloc(sizeof(struct evsel *),
348 hashmap__size(&ctx.ids) + 1);
349 if (!metric_events) {
350 expr__ctx_clear(&ctx);
351 return;
352 }
353 counter->metric_events = metric_events;
354 }
355
356 i = 0;
357 hashmap__for_each_entry((&ctx.ids), cur, bkt) {
358 const char *metric_name = (const char *)cur->key;
359
360 found = false;
361 if (leader) {
362 /* Search in group */
363 for_each_group_member (oc, leader) {
364 if (!strcasecmp(oc->name,
365 metric_name) &&
366 !oc->collect_stat) {
367 found = true;
368 break;
369 }
370 }
371 }
372 if (!found) {
373 /* Search ignoring groups */
374 oc = perf_stat__find_event(evsel_list,
375 metric_name);
376 }
377 if (!oc) {
378 /* Deduping one is good enough to handle duplicated PMUs. */
379 static char *printed;
380
381 /*
382 * Adding events automatically would be difficult, because
383 * it would risk creating groups that are not schedulable.
384 * perf stat doesn't understand all the scheduling constraints
385 * of events. So we ask the user instead to add the missing
386 * events.
387 */
388 if (!printed ||
389 strcasecmp(printed, metric_name)) {
390 fprintf(stderr,
391 "Add %s event to groups to get metric expression for %s\n",
392 metric_name,
393 counter->name);
394 printed = strdup(metric_name);
395 }
396 invalid = true;
397 continue;
398 }
399 metric_events[i++] = oc;
400 oc->collect_stat = true;
401 }
402 metric_events[i] = NULL;
403 if (invalid) {
404 free(metric_events);
405 counter->metric_events = NULL;
406 counter->metric_expr = NULL;
407 }
408 }
409 expr__ctx_clear(&ctx);
410}
411
412static double runtime_stat_avg(struct runtime_stat *st,
413 enum stat_type type, int ctx, int cpu)
414{
415 struct saved_value *v;
416
417 v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
418 if (!v)
419 return 0.0;
420
421 return avg_stats(&v->stats);
422}
423
424static double runtime_stat_n(struct runtime_stat *st,
425 enum stat_type type, int ctx, int cpu)
426{
427 struct saved_value *v;
428
429 v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
430 if (!v)
431 return 0.0;
432
433 return v->stats.n;
434}
435
436static void print_stalled_cycles_frontend(struct perf_stat_config *config,
437 int cpu,
438 struct evsel *evsel, double avg,
439 struct perf_stat_output_ctx *out,
440 struct runtime_stat *st)
441{
442 double total, ratio = 0.0;
443 const char *color;
444 int ctx = evsel_context(evsel);
445
446 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
447
448 if (total)
449 ratio = avg / total * 100.0;
450
451 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
452
453 if (ratio)
454 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
455 ratio);
456 else
457 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
458}
459
460static void print_stalled_cycles_backend(struct perf_stat_config *config,
461 int cpu,
462 struct evsel *evsel, double avg,
463 struct perf_stat_output_ctx *out,
464 struct runtime_stat *st)
465{
466 double total, ratio = 0.0;
467 const char *color;
468 int ctx = evsel_context(evsel);
469
470 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
471
472 if (total)
473 ratio = avg / total * 100.0;
474
475 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
476
477 out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
478}
479
480static void print_branch_misses(struct perf_stat_config *config,
481 int cpu,
482 struct evsel *evsel,
483 double avg,
484 struct perf_stat_output_ctx *out,
485 struct runtime_stat *st)
486{
487 double total, ratio = 0.0;
488 const char *color;
489 int ctx = evsel_context(evsel);
490
491 total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
492
493 if (total)
494 ratio = avg / total * 100.0;
495
496 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
497
498 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
499}
500
501static void print_l1_dcache_misses(struct perf_stat_config *config,
502 int cpu,
503 struct evsel *evsel,
504 double avg,
505 struct perf_stat_output_ctx *out,
506 struct runtime_stat *st)
507
508{
509 double total, ratio = 0.0;
510 const char *color;
511 int ctx = evsel_context(evsel);
512
513 total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
514
515 if (total)
516 ratio = avg / total * 100.0;
517
518 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
519
520 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
521}
522
523static void print_l1_icache_misses(struct perf_stat_config *config,
524 int cpu,
525 struct evsel *evsel,
526 double avg,
527 struct perf_stat_output_ctx *out,
528 struct runtime_stat *st)
529
530{
531 double total, ratio = 0.0;
532 const char *color;
533 int ctx = evsel_context(evsel);
534
535 total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
536
537 if (total)
538 ratio = avg / total * 100.0;
539
540 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
541 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
542}
543
544static void print_dtlb_cache_misses(struct perf_stat_config *config,
545 int cpu,
546 struct evsel *evsel,
547 double avg,
548 struct perf_stat_output_ctx *out,
549 struct runtime_stat *st)
550{
551 double total, ratio = 0.0;
552 const char *color;
553 int ctx = evsel_context(evsel);
554
555 total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
556
557 if (total)
558 ratio = avg / total * 100.0;
559
560 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
561 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
562}
563
564static void print_itlb_cache_misses(struct perf_stat_config *config,
565 int cpu,
566 struct evsel *evsel,
567 double avg,
568 struct perf_stat_output_ctx *out,
569 struct runtime_stat *st)
570{
571 double total, ratio = 0.0;
572 const char *color;
573 int ctx = evsel_context(evsel);
574
575 total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
576
577 if (total)
578 ratio = avg / total * 100.0;
579
580 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
581 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
582}
583
584static void print_ll_cache_misses(struct perf_stat_config *config,
585 int cpu,
586 struct evsel *evsel,
587 double avg,
588 struct perf_stat_output_ctx *out,
589 struct runtime_stat *st)
590{
591 double total, ratio = 0.0;
592 const char *color;
593 int ctx = evsel_context(evsel);
594
595 total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
596
597 if (total)
598 ratio = avg / total * 100.0;
599
600 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
601 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
602}
603
604/*
605 * High level "TopDown" CPU core pipe line bottleneck break down.
606 *
607 * Basic concept following
608 * Yasin, A Top Down Method for Performance analysis and Counter architecture
609 * ISPASS14
610 *
611 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
612 *
613 * Frontend -> Backend -> Retiring
614 * BadSpeculation in addition means out of order execution that is thrown away
615 * (for example branch mispredictions)
616 * Frontend is instruction decoding.
617 * Backend is execution, like computation and accessing data in memory
618 * Retiring is good execution that is not directly bottlenecked
619 *
620 * The formulas are computed in slots.
621 * A slot is an entry in the pipeline each for the pipeline width
622 * (for example a 4-wide pipeline has 4 slots for each cycle)
623 *
624 * Formulas:
625 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
626 * TotalSlots
627 * Retiring = SlotsRetired / TotalSlots
628 * FrontendBound = FetchBubbles / TotalSlots
629 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
630 *
631 * The kernel provides the mapping to the low level CPU events and any scaling
632 * needed for the CPU pipeline width, for example:
633 *
634 * TotalSlots = Cycles * 4
635 *
636 * The scaling factor is communicated in the sysfs unit.
637 *
638 * In some cases the CPU may not be able to measure all the formulas due to
639 * missing events. In this case multiple formulas are combined, as possible.
640 *
641 * Full TopDown supports more levels to sub-divide each area: for example
642 * BackendBound into computing bound and memory bound. For now we only
643 * support Level 1 TopDown.
644 */
645
646static double sanitize_val(double x)
647{
648 if (x < 0 && x >= -0.02)
649 return 0.0;
650 return x;
651}
652
653static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
654{
655 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
656}
657
658static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
659{
660 double bad_spec = 0;
661 double total_slots;
662 double total;
663
664 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
665 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
666 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
667
668 total_slots = td_total_slots(ctx, cpu, st);
669 if (total_slots)
670 bad_spec = total / total_slots;
671 return sanitize_val(bad_spec);
672}
673
674static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
675{
676 double retiring = 0;
677 double total_slots = td_total_slots(ctx, cpu, st);
678 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
679 ctx, cpu);
680
681 if (total_slots)
682 retiring = ret_slots / total_slots;
683 return retiring;
684}
685
686static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
687{
688 double fe_bound = 0;
689 double total_slots = td_total_slots(ctx, cpu, st);
690 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
691 ctx, cpu);
692
693 if (total_slots)
694 fe_bound = fetch_bub / total_slots;
695 return fe_bound;
696}
697
698static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
699{
700 double sum = (td_fe_bound(ctx, cpu, st) +
701 td_bad_spec(ctx, cpu, st) +
702 td_retiring(ctx, cpu, st));
703 if (sum == 0)
704 return 0;
705 return sanitize_val(1.0 - sum);
706}
707
708static void print_smi_cost(struct perf_stat_config *config,
709 int cpu, struct evsel *evsel,
710 struct perf_stat_output_ctx *out,
711 struct runtime_stat *st)
712{
713 double smi_num, aperf, cycles, cost = 0.0;
714 int ctx = evsel_context(evsel);
715 const char *color = NULL;
716
717 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
718 aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
719 cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
720
721 if ((cycles == 0) || (aperf == 0))
722 return;
723
724 if (smi_num)
725 cost = (aperf - cycles) / aperf * 100.00;
726
727 if (cost > 10)
728 color = PERF_COLOR_RED;
729 out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
730 out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
731}
732
733static int prepare_metric(struct evsel **metric_events,
734 struct metric_ref *metric_refs,
735 struct expr_parse_ctx *pctx,
736 int cpu,
737 struct runtime_stat *st)
738{
739 double scale;
740 char *n, *pn;
741 int i, j, ret;
742
743 expr__ctx_init(pctx);
744 for (i = 0; metric_events[i]; i++) {
745 struct saved_value *v;
746 struct stats *stats;
747 u64 metric_total = 0;
748
749 if (!strcmp(metric_events[i]->name, "duration_time")) {
750 stats = &walltime_nsecs_stats;
751 scale = 1e-9;
752 } else {
753 v = saved_value_lookup(metric_events[i], cpu, false,
754 STAT_NONE, 0, st);
755 if (!v)
756 break;
757 stats = &v->stats;
758 scale = 1.0;
759
760 if (v->metric_other)
761 metric_total = v->metric_total;
762 }
763
764 n = strdup(metric_events[i]->name);
765 if (!n)
766 return -ENOMEM;
767 /*
768 * This display code with --no-merge adds [cpu] postfixes.
769 * These are not supported by the parser. Remove everything
770 * after the space.
771 */
772 pn = strchr(n, ' ');
773 if (pn)
774 *pn = 0;
775
776 if (metric_total)
777 expr__add_id_val(pctx, n, metric_total);
778 else
779 expr__add_id_val(pctx, n, avg_stats(stats)*scale);
780 }
781
782 for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
783 ret = expr__add_ref(pctx, &metric_refs[j]);
784 if (ret)
785 return ret;
786 }
787
788 return i;
789}
790
791static void generic_metric(struct perf_stat_config *config,
792 const char *metric_expr,
793 struct evsel **metric_events,
794 struct metric_ref *metric_refs,
795 char *name,
796 const char *metric_name,
797 const char *metric_unit,
798 int runtime,
799 int cpu,
800 struct perf_stat_output_ctx *out,
801 struct runtime_stat *st)
802{
803 print_metric_t print_metric = out->print_metric;
804 struct expr_parse_ctx pctx;
805 double ratio, scale;
806 int i;
807 void *ctxp = out->ctx;
808
809 i = prepare_metric(metric_events, metric_refs, &pctx, cpu, st);
810 if (i < 0)
811 return;
812
813 if (!metric_events[i]) {
814 if (expr__parse(&ratio, &pctx, metric_expr, runtime) == 0) {
815 char *unit;
816 char metric_bf[64];
817
818 if (metric_unit && metric_name) {
819 if (perf_pmu__convert_scale(metric_unit,
820 &unit, &scale) >= 0) {
821 ratio *= scale;
822 }
823 if (strstr(metric_expr, "?"))
824 scnprintf(metric_bf, sizeof(metric_bf),
825 "%s %s_%d", unit, metric_name, runtime);
826 else
827 scnprintf(metric_bf, sizeof(metric_bf),
828 "%s %s", unit, metric_name);
829
830 print_metric(config, ctxp, NULL, "%8.1f",
831 metric_bf, ratio);
832 } else {
833 print_metric(config, ctxp, NULL, "%8.2f",
834 metric_name ?
835 metric_name :
836 out->force_header ? name : "",
837 ratio);
838 }
839 } else {
840 print_metric(config, ctxp, NULL, NULL,
841 out->force_header ?
842 (metric_name ? metric_name : name) : "", 0);
843 }
844 } else {
845 print_metric(config, ctxp, NULL, NULL,
846 out->force_header ?
847 (metric_name ? metric_name : name) : "", 0);
848 }
849
850 expr__ctx_clear(&pctx);
851}
852
853double test_generic_metric(struct metric_expr *mexp, int cpu, struct runtime_stat *st)
854{
855 struct expr_parse_ctx pctx;
856 double ratio = 0.0;
857
858 if (prepare_metric(mexp->metric_events, mexp->metric_refs, &pctx, cpu, st) < 0)
859 goto out;
860
861 if (expr__parse(&ratio, &pctx, mexp->metric_expr, 1))
862 ratio = 0.0;
863
864out:
865 expr__ctx_clear(&pctx);
866 return ratio;
867}
868
869void perf_stat__print_shadow_stats(struct perf_stat_config *config,
870 struct evsel *evsel,
871 double avg, int cpu,
872 struct perf_stat_output_ctx *out,
873 struct rblist *metric_events,
874 struct runtime_stat *st)
875{
876 void *ctxp = out->ctx;
877 print_metric_t print_metric = out->print_metric;
878 double total, ratio = 0.0, total2;
879 const char *color = NULL;
880 int ctx = evsel_context(evsel);
881 struct metric_event *me;
882 int num = 1;
883
884 if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
885 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
886
887 if (total) {
888 ratio = avg / total;
889 print_metric(config, ctxp, NULL, "%7.2f ",
890 "insn per cycle", ratio);
891 } else {
892 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
893 }
894
895 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
896 ctx, cpu);
897
898 total = max(total, runtime_stat_avg(st,
899 STAT_STALLED_CYCLES_BACK,
900 ctx, cpu));
901
902 if (total && avg) {
903 out->new_line(config, ctxp);
904 ratio = total / avg;
905 print_metric(config, ctxp, NULL, "%7.2f ",
906 "stalled cycles per insn",
907 ratio);
908 }
909 } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
910 if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
911 print_branch_misses(config, cpu, evsel, avg, out, st);
912 else
913 print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
914 } else if (
915 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
916 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D |
917 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
918 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
919
920 if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
921 print_l1_dcache_misses(config, cpu, evsel, avg, out, st);
922 else
923 print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
924 } else if (
925 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
926 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I |
927 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
928 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
929
930 if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
931 print_l1_icache_misses(config, cpu, evsel, avg, out, st);
932 else
933 print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
934 } else if (
935 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
936 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
937 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
938 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
939
940 if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
941 print_dtlb_cache_misses(config, cpu, evsel, avg, out, st);
942 else
943 print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
944 } else if (
945 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
946 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
947 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
948 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
949
950 if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
951 print_itlb_cache_misses(config, cpu, evsel, avg, out, st);
952 else
953 print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
954 } else if (
955 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
956 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL |
957 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
958 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
959
960 if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
961 print_ll_cache_misses(config, cpu, evsel, avg, out, st);
962 else
963 print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
964 } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
965 total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
966
967 if (total)
968 ratio = avg * 100 / total;
969
970 if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
971 print_metric(config, ctxp, NULL, "%8.3f %%",
972 "of all cache refs", ratio);
973 else
974 print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
975 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
976 print_stalled_cycles_frontend(config, cpu, evsel, avg, out, st);
977 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
978 print_stalled_cycles_backend(config, cpu, evsel, avg, out, st);
979 } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
980 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
981
982 if (total) {
983 ratio = avg / total;
984 print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
985 } else {
986 print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
987 }
988 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
989 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
990
991 if (total)
992 print_metric(config, ctxp, NULL,
993 "%7.2f%%", "transactional cycles",
994 100.0 * (avg / total));
995 else
996 print_metric(config, ctxp, NULL, NULL, "transactional cycles",
997 0);
998 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
999 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
1000 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
1001
1002 if (total2 < avg)
1003 total2 = avg;
1004 if (total)
1005 print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1006 100.0 * ((total2-avg) / total));
1007 else
1008 print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1009 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1010 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
1011 ctx, cpu);
1012
1013 if (avg)
1014 ratio = total / avg;
1015
1016 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
1017 print_metric(config, ctxp, NULL, "%8.0f",
1018 "cycles / transaction", ratio);
1019 else
1020 print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1021 0);
1022 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1023 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
1024 ctx, cpu);
1025
1026 if (avg)
1027 ratio = total / avg;
1028
1029 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1030 } else if (evsel__is_clock(evsel)) {
1031 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1032 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1033 avg / (ratio * evsel->scale));
1034 else
1035 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1036 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1037 double fe_bound = td_fe_bound(ctx, cpu, st);
1038
1039 if (fe_bound > 0.2)
1040 color = PERF_COLOR_RED;
1041 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1042 fe_bound * 100.);
1043 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1044 double retiring = td_retiring(ctx, cpu, st);
1045
1046 if (retiring > 0.7)
1047 color = PERF_COLOR_GREEN;
1048 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1049 retiring * 100.);
1050 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1051 double bad_spec = td_bad_spec(ctx, cpu, st);
1052
1053 if (bad_spec > 0.1)
1054 color = PERF_COLOR_RED;
1055 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1056 bad_spec * 100.);
1057 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1058 double be_bound = td_be_bound(ctx, cpu, st);
1059 const char *name = "backend bound";
1060 static int have_recovery_bubbles = -1;
1061
1062 /* In case the CPU does not support topdown-recovery-bubbles */
1063 if (have_recovery_bubbles < 0)
1064 have_recovery_bubbles = pmu_have_event("cpu",
1065 "topdown-recovery-bubbles");
1066 if (!have_recovery_bubbles)
1067 name = "backend bound/bad spec";
1068
1069 if (be_bound > 0.2)
1070 color = PERF_COLOR_RED;
1071 if (td_total_slots(ctx, cpu, st) > 0)
1072 print_metric(config, ctxp, color, "%8.1f%%", name,
1073 be_bound * 100.);
1074 else
1075 print_metric(config, ctxp, NULL, NULL, name, 0);
1076 } else if (evsel->metric_expr) {
1077 generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1078 evsel->name, evsel->metric_name, NULL, 1, cpu, out, st);
1079 } else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
1080 char unit = 'M';
1081 char unit_buf[10];
1082
1083 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
1084
1085 if (total)
1086 ratio = 1000.0 * avg / total;
1087 if (ratio < 0.001) {
1088 ratio *= 1000;
1089 unit = 'K';
1090 }
1091 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1092 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1093 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1094 print_smi_cost(config, cpu, evsel, out, st);
1095 } else {
1096 num = 0;
1097 }
1098
1099 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1100 struct metric_expr *mexp;
1101
1102 list_for_each_entry (mexp, &me->head, nd) {
1103 if (num++ > 0)
1104 out->new_line(config, ctxp);
1105 generic_metric(config, mexp->metric_expr, mexp->metric_events,
1106 mexp->metric_refs, evsel->name, mexp->metric_name,
1107 mexp->metric_unit, mexp->runtime, cpu, out, st);
1108 }
1109 }
1110 if (num == 0)
1111 print_metric(config, ctxp, NULL, NULL, NULL, 0);
1112}