Loading...
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
12/*
13 * AGGR_GLOBAL: Use CPU 0
14 * AGGR_SOCKET: Use first CPU of socket
15 * AGGR_CORE: Use first CPU of core
16 * AGGR_NONE: Use matching CPU
17 * AGGR_THREAD: Not supported?
18 */
19static bool have_frontend_stalled;
20
21struct runtime_stat rt_stat;
22struct stats walltime_nsecs_stats;
23
24struct saved_value {
25 struct rb_node rb_node;
26 struct perf_evsel *evsel;
27 enum stat_type type;
28 int ctx;
29 int cpu;
30 struct runtime_stat *stat;
31 struct stats stats;
32};
33
34static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
35{
36 struct saved_value *a = container_of(rb_node,
37 struct saved_value,
38 rb_node);
39 const struct saved_value *b = entry;
40
41 if (a->cpu != b->cpu)
42 return a->cpu - b->cpu;
43
44 /*
45 * Previously the rbtree was used to link generic metrics.
46 * The keys were evsel/cpu. Now the rbtree is extended to support
47 * per-thread shadow stats. For shadow stats case, the keys
48 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
49 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
50 */
51 if (a->type != b->type)
52 return a->type - b->type;
53
54 if (a->ctx != b->ctx)
55 return a->ctx - b->ctx;
56
57 if (a->evsel == NULL && b->evsel == NULL) {
58 if (a->stat == b->stat)
59 return 0;
60
61 if ((char *)a->stat < (char *)b->stat)
62 return -1;
63
64 return 1;
65 }
66
67 if (a->evsel == b->evsel)
68 return 0;
69 if ((char *)a->evsel < (char *)b->evsel)
70 return -1;
71 return +1;
72}
73
74static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
75 const void *entry)
76{
77 struct saved_value *nd = malloc(sizeof(struct saved_value));
78
79 if (!nd)
80 return NULL;
81 memcpy(nd, entry, sizeof(struct saved_value));
82 return &nd->rb_node;
83}
84
85static void saved_value_delete(struct rblist *rblist __maybe_unused,
86 struct rb_node *rb_node)
87{
88 struct saved_value *v;
89
90 BUG_ON(!rb_node);
91 v = container_of(rb_node, struct saved_value, rb_node);
92 free(v);
93}
94
95static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
96 int cpu,
97 bool create,
98 enum stat_type type,
99 int ctx,
100 struct runtime_stat *st)
101{
102 struct rblist *rblist;
103 struct rb_node *nd;
104 struct saved_value dm = {
105 .cpu = cpu,
106 .evsel = evsel,
107 .type = type,
108 .ctx = ctx,
109 .stat = st,
110 };
111
112 rblist = &st->value_list;
113
114 nd = rblist__find(rblist, &dm);
115 if (nd)
116 return container_of(nd, struct saved_value, rb_node);
117 if (create) {
118 rblist__add_node(rblist, &dm);
119 nd = rblist__find(rblist, &dm);
120 if (nd)
121 return container_of(nd, struct saved_value, rb_node);
122 }
123 return NULL;
124}
125
126void runtime_stat__init(struct runtime_stat *st)
127{
128 struct rblist *rblist = &st->value_list;
129
130 rblist__init(rblist);
131 rblist->node_cmp = saved_value_cmp;
132 rblist->node_new = saved_value_new;
133 rblist->node_delete = saved_value_delete;
134}
135
136void runtime_stat__exit(struct runtime_stat *st)
137{
138 rblist__exit(&st->value_list);
139}
140
141void perf_stat__init_shadow_stats(void)
142{
143 have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
144 runtime_stat__init(&rt_stat);
145}
146
147static int evsel_context(struct perf_evsel *evsel)
148{
149 int ctx = 0;
150
151 if (evsel->attr.exclude_kernel)
152 ctx |= CTX_BIT_KERNEL;
153 if (evsel->attr.exclude_user)
154 ctx |= CTX_BIT_USER;
155 if (evsel->attr.exclude_hv)
156 ctx |= CTX_BIT_HV;
157 if (evsel->attr.exclude_host)
158 ctx |= CTX_BIT_HOST;
159 if (evsel->attr.exclude_idle)
160 ctx |= CTX_BIT_IDLE;
161
162 return ctx;
163}
164
165static void reset_stat(struct runtime_stat *st)
166{
167 struct rblist *rblist;
168 struct rb_node *pos, *next;
169
170 rblist = &st->value_list;
171 next = rb_first(&rblist->entries);
172 while (next) {
173 pos = next;
174 next = rb_next(pos);
175 memset(&container_of(pos, struct saved_value, rb_node)->stats,
176 0,
177 sizeof(struct stats));
178 }
179}
180
181void perf_stat__reset_shadow_stats(void)
182{
183 reset_stat(&rt_stat);
184 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
185}
186
187void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
188{
189 reset_stat(st);
190}
191
192static void update_runtime_stat(struct runtime_stat *st,
193 enum stat_type type,
194 int ctx, int cpu, u64 count)
195{
196 struct saved_value *v = saved_value_lookup(NULL, cpu, true,
197 type, ctx, st);
198
199 if (v)
200 update_stats(&v->stats, count);
201}
202
203/*
204 * Update various tracking values we maintain to print
205 * more semantic information such as miss/hit ratios,
206 * instruction rates, etc:
207 */
208void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 count,
209 int cpu, struct runtime_stat *st)
210{
211 int ctx = evsel_context(counter);
212
213 count *= counter->scale;
214
215 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
216 perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
217 update_runtime_stat(st, STAT_NSECS, 0, cpu, count);
218 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
219 update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count);
220 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
221 update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count);
222 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
223 update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count);
224 else if (perf_stat_evsel__is(counter, ELISION_START))
225 update_runtime_stat(st, STAT_ELISION, ctx, cpu, count);
226 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
227 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
228 ctx, cpu, count);
229 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
230 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
231 ctx, cpu, count);
232 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
233 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
234 ctx, cpu, count);
235 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
236 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
237 ctx, cpu, count);
238 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
239 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
240 ctx, cpu, count);
241 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
242 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
243 ctx, cpu, count);
244 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
245 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
246 ctx, cpu, count);
247 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
248 update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
249 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
250 update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
251 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
252 update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
253 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
254 update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
255 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
256 update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
257 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
258 update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
259 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
260 update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count);
261 else if (perf_stat_evsel__is(counter, SMI_NUM))
262 update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count);
263 else if (perf_stat_evsel__is(counter, APERF))
264 update_runtime_stat(st, STAT_APERF, ctx, cpu, count);
265
266 if (counter->collect_stat) {
267 struct saved_value *v = saved_value_lookup(counter, cpu, true,
268 STAT_NONE, 0, st);
269 update_stats(&v->stats, count);
270 }
271}
272
273/* used for get_ratio_color() */
274enum grc_type {
275 GRC_STALLED_CYCLES_FE,
276 GRC_STALLED_CYCLES_BE,
277 GRC_CACHE_MISSES,
278 GRC_MAX_NR
279};
280
281static const char *get_ratio_color(enum grc_type type, double ratio)
282{
283 static const double grc_table[GRC_MAX_NR][3] = {
284 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
285 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
286 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
287 };
288 const char *color = PERF_COLOR_NORMAL;
289
290 if (ratio > grc_table[type][0])
291 color = PERF_COLOR_RED;
292 else if (ratio > grc_table[type][1])
293 color = PERF_COLOR_MAGENTA;
294 else if (ratio > grc_table[type][2])
295 color = PERF_COLOR_YELLOW;
296
297 return color;
298}
299
300static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
301 const char *name)
302{
303 struct perf_evsel *c2;
304
305 evlist__for_each_entry (evsel_list, c2) {
306 if (!strcasecmp(c2->name, name))
307 return c2;
308 }
309 return NULL;
310}
311
312/* Mark MetricExpr target events and link events using them to them. */
313void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
314{
315 struct perf_evsel *counter, *leader, **metric_events, *oc;
316 bool found;
317 const char **metric_names;
318 int i;
319 int num_metric_names;
320
321 evlist__for_each_entry(evsel_list, counter) {
322 bool invalid = false;
323
324 leader = counter->leader;
325 if (!counter->metric_expr)
326 continue;
327 metric_events = counter->metric_events;
328 if (!metric_events) {
329 if (expr__find_other(counter->metric_expr, counter->name,
330 &metric_names, &num_metric_names) < 0)
331 continue;
332
333 metric_events = calloc(sizeof(struct perf_evsel *),
334 num_metric_names + 1);
335 if (!metric_events)
336 return;
337 counter->metric_events = metric_events;
338 }
339
340 for (i = 0; i < num_metric_names; i++) {
341 found = false;
342 if (leader) {
343 /* Search in group */
344 for_each_group_member (oc, leader) {
345 if (!strcasecmp(oc->name, metric_names[i])) {
346 found = true;
347 break;
348 }
349 }
350 }
351 if (!found) {
352 /* Search ignoring groups */
353 oc = perf_stat__find_event(evsel_list, metric_names[i]);
354 }
355 if (!oc) {
356 /* Deduping one is good enough to handle duplicated PMUs. */
357 static char *printed;
358
359 /*
360 * Adding events automatically would be difficult, because
361 * it would risk creating groups that are not schedulable.
362 * perf stat doesn't understand all the scheduling constraints
363 * of events. So we ask the user instead to add the missing
364 * events.
365 */
366 if (!printed || strcasecmp(printed, metric_names[i])) {
367 fprintf(stderr,
368 "Add %s event to groups to get metric expression for %s\n",
369 metric_names[i],
370 counter->name);
371 printed = strdup(metric_names[i]);
372 }
373 invalid = true;
374 continue;
375 }
376 metric_events[i] = oc;
377 oc->collect_stat = true;
378 }
379 metric_events[i] = NULL;
380 free(metric_names);
381 if (invalid) {
382 free(metric_events);
383 counter->metric_events = NULL;
384 counter->metric_expr = NULL;
385 }
386 }
387}
388
389static double runtime_stat_avg(struct runtime_stat *st,
390 enum stat_type type, int ctx, int cpu)
391{
392 struct saved_value *v;
393
394 v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
395 if (!v)
396 return 0.0;
397
398 return avg_stats(&v->stats);
399}
400
401static double runtime_stat_n(struct runtime_stat *st,
402 enum stat_type type, int ctx, int cpu)
403{
404 struct saved_value *v;
405
406 v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
407 if (!v)
408 return 0.0;
409
410 return v->stats.n;
411}
412
413static void print_stalled_cycles_frontend(int cpu,
414 struct perf_evsel *evsel, double avg,
415 struct perf_stat_output_ctx *out,
416 struct runtime_stat *st)
417{
418 double total, ratio = 0.0;
419 const char *color;
420 int ctx = evsel_context(evsel);
421
422 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
423
424 if (total)
425 ratio = avg / total * 100.0;
426
427 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
428
429 if (ratio)
430 out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
431 ratio);
432 else
433 out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
434}
435
436static void print_stalled_cycles_backend(int cpu,
437 struct perf_evsel *evsel, double avg,
438 struct perf_stat_output_ctx *out,
439 struct runtime_stat *st)
440{
441 double total, ratio = 0.0;
442 const char *color;
443 int ctx = evsel_context(evsel);
444
445 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
446
447 if (total)
448 ratio = avg / total * 100.0;
449
450 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
451
452 out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
453}
454
455static void print_branch_misses(int cpu,
456 struct perf_evsel *evsel,
457 double avg,
458 struct perf_stat_output_ctx *out,
459 struct runtime_stat *st)
460{
461 double total, ratio = 0.0;
462 const char *color;
463 int ctx = evsel_context(evsel);
464
465 total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
466
467 if (total)
468 ratio = avg / total * 100.0;
469
470 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
471
472 out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
473}
474
475static void print_l1_dcache_misses(int cpu,
476 struct perf_evsel *evsel,
477 double avg,
478 struct perf_stat_output_ctx *out,
479 struct runtime_stat *st)
480
481{
482 double total, ratio = 0.0;
483 const char *color;
484 int ctx = evsel_context(evsel);
485
486 total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
487
488 if (total)
489 ratio = avg / total * 100.0;
490
491 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
492
493 out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
494}
495
496static void print_l1_icache_misses(int cpu,
497 struct perf_evsel *evsel,
498 double avg,
499 struct perf_stat_output_ctx *out,
500 struct runtime_stat *st)
501
502{
503 double total, ratio = 0.0;
504 const char *color;
505 int ctx = evsel_context(evsel);
506
507 total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
508
509 if (total)
510 ratio = avg / total * 100.0;
511
512 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
513 out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
514}
515
516static void print_dtlb_cache_misses(int cpu,
517 struct perf_evsel *evsel,
518 double avg,
519 struct perf_stat_output_ctx *out,
520 struct runtime_stat *st)
521{
522 double total, ratio = 0.0;
523 const char *color;
524 int ctx = evsel_context(evsel);
525
526 total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
527
528 if (total)
529 ratio = avg / total * 100.0;
530
531 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
532 out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
533}
534
535static void print_itlb_cache_misses(int cpu,
536 struct perf_evsel *evsel,
537 double avg,
538 struct perf_stat_output_ctx *out,
539 struct runtime_stat *st)
540{
541 double total, ratio = 0.0;
542 const char *color;
543 int ctx = evsel_context(evsel);
544
545 total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
546
547 if (total)
548 ratio = avg / total * 100.0;
549
550 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
551 out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
552}
553
554static void print_ll_cache_misses(int cpu,
555 struct perf_evsel *evsel,
556 double avg,
557 struct perf_stat_output_ctx *out,
558 struct runtime_stat *st)
559{
560 double total, ratio = 0.0;
561 const char *color;
562 int ctx = evsel_context(evsel);
563
564 total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
565
566 if (total)
567 ratio = avg / total * 100.0;
568
569 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
570 out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
571}
572
573/*
574 * High level "TopDown" CPU core pipe line bottleneck break down.
575 *
576 * Basic concept following
577 * Yasin, A Top Down Method for Performance analysis and Counter architecture
578 * ISPASS14
579 *
580 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
581 *
582 * Frontend -> Backend -> Retiring
583 * BadSpeculation in addition means out of order execution that is thrown away
584 * (for example branch mispredictions)
585 * Frontend is instruction decoding.
586 * Backend is execution, like computation and accessing data in memory
587 * Retiring is good execution that is not directly bottlenecked
588 *
589 * The formulas are computed in slots.
590 * A slot is an entry in the pipeline each for the pipeline width
591 * (for example a 4-wide pipeline has 4 slots for each cycle)
592 *
593 * Formulas:
594 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
595 * TotalSlots
596 * Retiring = SlotsRetired / TotalSlots
597 * FrontendBound = FetchBubbles / TotalSlots
598 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
599 *
600 * The kernel provides the mapping to the low level CPU events and any scaling
601 * needed for the CPU pipeline width, for example:
602 *
603 * TotalSlots = Cycles * 4
604 *
605 * The scaling factor is communicated in the sysfs unit.
606 *
607 * In some cases the CPU may not be able to measure all the formulas due to
608 * missing events. In this case multiple formulas are combined, as possible.
609 *
610 * Full TopDown supports more levels to sub-divide each area: for example
611 * BackendBound into computing bound and memory bound. For now we only
612 * support Level 1 TopDown.
613 */
614
615static double sanitize_val(double x)
616{
617 if (x < 0 && x >= -0.02)
618 return 0.0;
619 return x;
620}
621
622static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
623{
624 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
625}
626
627static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
628{
629 double bad_spec = 0;
630 double total_slots;
631 double total;
632
633 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
634 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
635 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
636
637 total_slots = td_total_slots(ctx, cpu, st);
638 if (total_slots)
639 bad_spec = total / total_slots;
640 return sanitize_val(bad_spec);
641}
642
643static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
644{
645 double retiring = 0;
646 double total_slots = td_total_slots(ctx, cpu, st);
647 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
648 ctx, cpu);
649
650 if (total_slots)
651 retiring = ret_slots / total_slots;
652 return retiring;
653}
654
655static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
656{
657 double fe_bound = 0;
658 double total_slots = td_total_slots(ctx, cpu, st);
659 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
660 ctx, cpu);
661
662 if (total_slots)
663 fe_bound = fetch_bub / total_slots;
664 return fe_bound;
665}
666
667static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
668{
669 double sum = (td_fe_bound(ctx, cpu, st) +
670 td_bad_spec(ctx, cpu, st) +
671 td_retiring(ctx, cpu, st));
672 if (sum == 0)
673 return 0;
674 return sanitize_val(1.0 - sum);
675}
676
677static void print_smi_cost(int cpu, struct perf_evsel *evsel,
678 struct perf_stat_output_ctx *out,
679 struct runtime_stat *st)
680{
681 double smi_num, aperf, cycles, cost = 0.0;
682 int ctx = evsel_context(evsel);
683 const char *color = NULL;
684
685 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
686 aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
687 cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
688
689 if ((cycles == 0) || (aperf == 0))
690 return;
691
692 if (smi_num)
693 cost = (aperf - cycles) / aperf * 100.00;
694
695 if (cost > 10)
696 color = PERF_COLOR_RED;
697 out->print_metric(out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
698 out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
699}
700
701static void generic_metric(const char *metric_expr,
702 struct perf_evsel **metric_events,
703 char *name,
704 const char *metric_name,
705 double avg,
706 int cpu,
707 struct perf_stat_output_ctx *out,
708 struct runtime_stat *st)
709{
710 print_metric_t print_metric = out->print_metric;
711 struct parse_ctx pctx;
712 double ratio;
713 int i;
714 void *ctxp = out->ctx;
715
716 expr__ctx_init(&pctx);
717 expr__add_id(&pctx, name, avg);
718 for (i = 0; metric_events[i]; i++) {
719 struct saved_value *v;
720 struct stats *stats;
721 double scale;
722
723 if (!strcmp(metric_events[i]->name, "duration_time")) {
724 stats = &walltime_nsecs_stats;
725 scale = 1e-9;
726 } else {
727 v = saved_value_lookup(metric_events[i], cpu, false,
728 STAT_NONE, 0, st);
729 if (!v)
730 break;
731 stats = &v->stats;
732 scale = 1.0;
733 }
734 expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
735 }
736 if (!metric_events[i]) {
737 const char *p = metric_expr;
738
739 if (expr__parse(&ratio, &pctx, &p) == 0)
740 print_metric(ctxp, NULL, "%8.1f",
741 metric_name ?
742 metric_name :
743 out->force_header ? name : "",
744 ratio);
745 else
746 print_metric(ctxp, NULL, NULL,
747 out->force_header ?
748 (metric_name ? metric_name : name) : "", 0);
749 } else
750 print_metric(ctxp, NULL, NULL, "", 0);
751}
752
753void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
754 double avg, int cpu,
755 struct perf_stat_output_ctx *out,
756 struct rblist *metric_events,
757 struct runtime_stat *st)
758{
759 void *ctxp = out->ctx;
760 print_metric_t print_metric = out->print_metric;
761 double total, ratio = 0.0, total2;
762 const char *color = NULL;
763 int ctx = evsel_context(evsel);
764 struct metric_event *me;
765 int num = 1;
766
767 if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
768 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
769
770 if (total) {
771 ratio = avg / total;
772 print_metric(ctxp, NULL, "%7.2f ",
773 "insn per cycle", ratio);
774 } else {
775 print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
776 }
777
778 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
779 ctx, cpu);
780
781 total = max(total, runtime_stat_avg(st,
782 STAT_STALLED_CYCLES_BACK,
783 ctx, cpu));
784
785 if (total && avg) {
786 out->new_line(ctxp);
787 ratio = total / avg;
788 print_metric(ctxp, NULL, "%7.2f ",
789 "stalled cycles per insn",
790 ratio);
791 } else if (have_frontend_stalled) {
792 print_metric(ctxp, NULL, NULL,
793 "stalled cycles per insn", 0);
794 }
795 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
796 if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
797 print_branch_misses(cpu, evsel, avg, out, st);
798 else
799 print_metric(ctxp, NULL, NULL, "of all branches", 0);
800 } else if (
801 evsel->attr.type == PERF_TYPE_HW_CACHE &&
802 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
803 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
804 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
805
806 if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
807 print_l1_dcache_misses(cpu, evsel, avg, out, st);
808 else
809 print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
810 } else if (
811 evsel->attr.type == PERF_TYPE_HW_CACHE &&
812 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
813 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
814 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
815
816 if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
817 print_l1_icache_misses(cpu, evsel, avg, out, st);
818 else
819 print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
820 } else if (
821 evsel->attr.type == PERF_TYPE_HW_CACHE &&
822 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
823 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
824 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
825
826 if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
827 print_dtlb_cache_misses(cpu, evsel, avg, out, st);
828 else
829 print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
830 } else if (
831 evsel->attr.type == PERF_TYPE_HW_CACHE &&
832 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
833 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
834 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
835
836 if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
837 print_itlb_cache_misses(cpu, evsel, avg, out, st);
838 else
839 print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
840 } else if (
841 evsel->attr.type == PERF_TYPE_HW_CACHE &&
842 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
843 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
844 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
845
846 if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
847 print_ll_cache_misses(cpu, evsel, avg, out, st);
848 else
849 print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
850 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
851 total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
852
853 if (total)
854 ratio = avg * 100 / total;
855
856 if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
857 print_metric(ctxp, NULL, "%8.3f %%",
858 "of all cache refs", ratio);
859 else
860 print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
861 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
862 print_stalled_cycles_frontend(cpu, evsel, avg, out, st);
863 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
864 print_stalled_cycles_backend(cpu, evsel, avg, out, st);
865 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
866 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
867
868 if (total) {
869 ratio = avg / total;
870 print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
871 } else {
872 print_metric(ctxp, NULL, NULL, "Ghz", 0);
873 }
874 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
875 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
876
877 if (total)
878 print_metric(ctxp, NULL,
879 "%7.2f%%", "transactional cycles",
880 100.0 * (avg / total));
881 else
882 print_metric(ctxp, NULL, NULL, "transactional cycles",
883 0);
884 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
885 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
886 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
887
888 if (total2 < avg)
889 total2 = avg;
890 if (total)
891 print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
892 100.0 * ((total2-avg) / total));
893 else
894 print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
895 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
896 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
897 ctx, cpu);
898
899 if (avg)
900 ratio = total / avg;
901
902 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
903 print_metric(ctxp, NULL, "%8.0f",
904 "cycles / transaction", ratio);
905 else
906 print_metric(ctxp, NULL, NULL, "cycles / transaction",
907 0);
908 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
909 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
910 ctx, cpu);
911
912 if (avg)
913 ratio = total / avg;
914
915 print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
916 } else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK) ||
917 perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK)) {
918 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
919 print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
920 avg / ratio);
921 else
922 print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
923 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
924 double fe_bound = td_fe_bound(ctx, cpu, st);
925
926 if (fe_bound > 0.2)
927 color = PERF_COLOR_RED;
928 print_metric(ctxp, color, "%8.1f%%", "frontend bound",
929 fe_bound * 100.);
930 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
931 double retiring = td_retiring(ctx, cpu, st);
932
933 if (retiring > 0.7)
934 color = PERF_COLOR_GREEN;
935 print_metric(ctxp, color, "%8.1f%%", "retiring",
936 retiring * 100.);
937 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
938 double bad_spec = td_bad_spec(ctx, cpu, st);
939
940 if (bad_spec > 0.1)
941 color = PERF_COLOR_RED;
942 print_metric(ctxp, color, "%8.1f%%", "bad speculation",
943 bad_spec * 100.);
944 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
945 double be_bound = td_be_bound(ctx, cpu, st);
946 const char *name = "backend bound";
947 static int have_recovery_bubbles = -1;
948
949 /* In case the CPU does not support topdown-recovery-bubbles */
950 if (have_recovery_bubbles < 0)
951 have_recovery_bubbles = pmu_have_event("cpu",
952 "topdown-recovery-bubbles");
953 if (!have_recovery_bubbles)
954 name = "backend bound/bad spec";
955
956 if (be_bound > 0.2)
957 color = PERF_COLOR_RED;
958 if (td_total_slots(ctx, cpu, st) > 0)
959 print_metric(ctxp, color, "%8.1f%%", name,
960 be_bound * 100.);
961 else
962 print_metric(ctxp, NULL, NULL, name, 0);
963 } else if (evsel->metric_expr) {
964 generic_metric(evsel->metric_expr, evsel->metric_events, evsel->name,
965 evsel->metric_name, avg, cpu, out, st);
966 } else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
967 char unit = 'M';
968 char unit_buf[10];
969
970 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
971
972 if (total)
973 ratio = 1000.0 * avg / total;
974 if (ratio < 0.001) {
975 ratio *= 1000;
976 unit = 'K';
977 }
978 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
979 print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
980 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
981 print_smi_cost(cpu, evsel, out, st);
982 } else {
983 num = 0;
984 }
985
986 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
987 struct metric_expr *mexp;
988
989 list_for_each_entry (mexp, &me->head, nd) {
990 if (num++ > 0)
991 out->new_line(ctxp);
992 generic_metric(mexp->metric_expr, mexp->metric_events,
993 evsel->name, mexp->metric_name,
994 avg, cpu, out, st);
995 }
996 }
997 if (num == 0)
998 print_metric(ctxp, NULL, NULL, NULL, 0);
999}
1// SPDX-License-Identifier: GPL-2.0
2#include <math.h>
3#include <stdio.h>
4#include "evsel.h"
5#include "stat.h"
6#include "color.h"
7#include "debug.h"
8#include "pmu.h"
9#include "rblist.h"
10#include "evlist.h"
11#include "expr.h"
12#include "metricgroup.h"
13#include "cgroup.h"
14#include "units.h"
15#include <linux/zalloc.h>
16#include "iostat.h"
17#include "util/hashmap.h"
18#include "tool_pmu.h"
19
20struct stats walltime_nsecs_stats;
21struct rusage_stats ru_stats;
22
23enum {
24 CTX_BIT_USER = 1 << 0,
25 CTX_BIT_KERNEL = 1 << 1,
26 CTX_BIT_HV = 1 << 2,
27 CTX_BIT_HOST = 1 << 3,
28 CTX_BIT_IDLE = 1 << 4,
29 CTX_BIT_MAX = 1 << 5,
30};
31
32enum stat_type {
33 STAT_NONE = 0,
34 STAT_NSECS,
35 STAT_CYCLES,
36 STAT_INSTRUCTIONS,
37 STAT_STALLED_CYCLES_FRONT,
38 STAT_STALLED_CYCLES_BACK,
39 STAT_BRANCHES,
40 STAT_BRANCH_MISS,
41 STAT_CACHE_REFS,
42 STAT_CACHE_MISSES,
43 STAT_L1_DCACHE,
44 STAT_L1_ICACHE,
45 STAT_LL_CACHE,
46 STAT_ITLB_CACHE,
47 STAT_DTLB_CACHE,
48 STAT_L1D_MISS,
49 STAT_L1I_MISS,
50 STAT_LL_MISS,
51 STAT_DTLB_MISS,
52 STAT_ITLB_MISS,
53 STAT_MAX
54};
55
56static int evsel_context(const struct evsel *evsel)
57{
58 int ctx = 0;
59
60 if (evsel->core.attr.exclude_kernel)
61 ctx |= CTX_BIT_KERNEL;
62 if (evsel->core.attr.exclude_user)
63 ctx |= CTX_BIT_USER;
64 if (evsel->core.attr.exclude_hv)
65 ctx |= CTX_BIT_HV;
66 if (evsel->core.attr.exclude_host)
67 ctx |= CTX_BIT_HOST;
68 if (evsel->core.attr.exclude_idle)
69 ctx |= CTX_BIT_IDLE;
70
71 return ctx;
72}
73
74void perf_stat__reset_shadow_stats(void)
75{
76 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
77 memset(&ru_stats, 0, sizeof(ru_stats));
78}
79
80static enum stat_type evsel__stat_type(struct evsel *evsel)
81{
82 /* Fake perf_hw_cache_op_id values for use with evsel__match. */
83 u64 PERF_COUNT_hw_cache_l1d_miss = PERF_COUNT_HW_CACHE_L1D |
84 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
85 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
86 u64 PERF_COUNT_hw_cache_l1i_miss = PERF_COUNT_HW_CACHE_L1I |
87 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
88 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
89 u64 PERF_COUNT_hw_cache_ll_miss = PERF_COUNT_HW_CACHE_LL |
90 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
91 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
92 u64 PERF_COUNT_hw_cache_dtlb_miss = PERF_COUNT_HW_CACHE_DTLB |
93 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
94 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
95 u64 PERF_COUNT_hw_cache_itlb_miss = PERF_COUNT_HW_CACHE_ITLB |
96 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
97 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16);
98
99 if (evsel__is_clock(evsel))
100 return STAT_NSECS;
101 else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES))
102 return STAT_CYCLES;
103 else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS))
104 return STAT_INSTRUCTIONS;
105 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
106 return STAT_STALLED_CYCLES_FRONT;
107 else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND))
108 return STAT_STALLED_CYCLES_BACK;
109 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_INSTRUCTIONS))
110 return STAT_BRANCHES;
111 else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES))
112 return STAT_BRANCH_MISS;
113 else if (evsel__match(evsel, HARDWARE, HW_CACHE_REFERENCES))
114 return STAT_CACHE_REFS;
115 else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES))
116 return STAT_CACHE_MISSES;
117 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1D))
118 return STAT_L1_DCACHE;
119 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_L1I))
120 return STAT_L1_ICACHE;
121 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_LL))
122 return STAT_LL_CACHE;
123 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_DTLB))
124 return STAT_DTLB_CACHE;
125 else if (evsel__match(evsel, HW_CACHE, HW_CACHE_ITLB))
126 return STAT_ITLB_CACHE;
127 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1d_miss))
128 return STAT_L1D_MISS;
129 else if (evsel__match(evsel, HW_CACHE, hw_cache_l1i_miss))
130 return STAT_L1I_MISS;
131 else if (evsel__match(evsel, HW_CACHE, hw_cache_ll_miss))
132 return STAT_LL_MISS;
133 else if (evsel__match(evsel, HW_CACHE, hw_cache_dtlb_miss))
134 return STAT_DTLB_MISS;
135 else if (evsel__match(evsel, HW_CACHE, hw_cache_itlb_miss))
136 return STAT_ITLB_MISS;
137 return STAT_NONE;
138}
139
140static enum metric_threshold_classify get_ratio_thresh(const double ratios[3], double val)
141{
142 assert(ratios[0] > ratios[1]);
143 assert(ratios[1] > ratios[2]);
144
145 return val > ratios[1]
146 ? (val > ratios[0] ? METRIC_THRESHOLD_BAD : METRIC_THRESHOLD_NEARLY_BAD)
147 : (val > ratios[2] ? METRIC_THRESHOLD_LESS_GOOD : METRIC_THRESHOLD_GOOD);
148}
149
150static double find_stat(const struct evsel *evsel, int aggr_idx, enum stat_type type)
151{
152 struct evsel *cur;
153 int evsel_ctx = evsel_context(evsel);
154
155 evlist__for_each_entry(evsel->evlist, cur) {
156 struct perf_stat_aggr *aggr;
157
158 /* Ignore the evsel that is being searched from. */
159 if (evsel == cur)
160 continue;
161
162 /* Ignore evsels that are part of different groups. */
163 if (evsel->core.leader->nr_members > 1 &&
164 evsel->core.leader != cur->core.leader)
165 continue;
166 /* Ignore evsels with mismatched modifiers. */
167 if (evsel_ctx != evsel_context(cur))
168 continue;
169 /* Ignore if not the cgroup we're looking for. */
170 if (evsel->cgrp != cur->cgrp)
171 continue;
172 /* Ignore if not the stat we're looking for. */
173 if (type != evsel__stat_type(cur))
174 continue;
175
176 /*
177 * Except the SW CLOCK events,
178 * ignore if not the PMU we're looking for.
179 */
180 if ((type != STAT_NSECS) && (evsel->pmu != cur->pmu))
181 continue;
182
183 aggr = &cur->stats->aggr[aggr_idx];
184 if (type == STAT_NSECS)
185 return aggr->counts.val;
186 return aggr->counts.val * cur->scale;
187 }
188 return 0.0;
189}
190
191static void print_ratio(struct perf_stat_config *config,
192 const struct evsel *evsel, int aggr_idx,
193 double numerator, struct perf_stat_output_ctx *out,
194 enum stat_type denominator_type,
195 const double thresh_ratios[3], const char *_unit)
196{
197 double denominator = find_stat(evsel, aggr_idx, denominator_type);
198 double ratio = 0;
199 enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
200 const char *fmt = NULL;
201 const char *unit = NULL;
202
203 if (numerator && denominator) {
204 ratio = numerator / denominator * 100.0;
205 thresh = get_ratio_thresh(thresh_ratios, ratio);
206 fmt = "%7.2f%%";
207 unit = _unit;
208 }
209 out->print_metric(config, out->ctx, thresh, fmt, unit, ratio);
210}
211
212static void print_stalled_cycles_front(struct perf_stat_config *config,
213 const struct evsel *evsel,
214 int aggr_idx, double stalled,
215 struct perf_stat_output_ctx *out)
216{
217 const double thresh_ratios[3] = {50.0, 30.0, 10.0};
218
219 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
220 "frontend cycles idle");
221}
222
223static void print_stalled_cycles_back(struct perf_stat_config *config,
224 const struct evsel *evsel,
225 int aggr_idx, double stalled,
226 struct perf_stat_output_ctx *out)
227{
228 const double thresh_ratios[3] = {75.0, 50.0, 20.0};
229
230 print_ratio(config, evsel, aggr_idx, stalled, out, STAT_CYCLES, thresh_ratios,
231 "backend cycles idle");
232}
233
234static void print_branch_miss(struct perf_stat_config *config,
235 const struct evsel *evsel,
236 int aggr_idx, double misses,
237 struct perf_stat_output_ctx *out)
238{
239 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
240
241 print_ratio(config, evsel, aggr_idx, misses, out, STAT_BRANCHES, thresh_ratios,
242 "of all branches");
243}
244
245static void print_l1d_miss(struct perf_stat_config *config,
246 const struct evsel *evsel,
247 int aggr_idx, double misses,
248 struct perf_stat_output_ctx *out)
249{
250 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
251
252 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_DCACHE, thresh_ratios,
253 "of all L1-dcache accesses");
254}
255
256static void print_l1i_miss(struct perf_stat_config *config,
257 const struct evsel *evsel,
258 int aggr_idx, double misses,
259 struct perf_stat_output_ctx *out)
260{
261 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
262
263 print_ratio(config, evsel, aggr_idx, misses, out, STAT_L1_ICACHE, thresh_ratios,
264 "of all L1-icache accesses");
265}
266
267static void print_ll_miss(struct perf_stat_config *config,
268 const struct evsel *evsel,
269 int aggr_idx, double misses,
270 struct perf_stat_output_ctx *out)
271{
272 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
273
274 print_ratio(config, evsel, aggr_idx, misses, out, STAT_LL_CACHE, thresh_ratios,
275 "of all LL-cache accesses");
276}
277
278static void print_dtlb_miss(struct perf_stat_config *config,
279 const struct evsel *evsel,
280 int aggr_idx, double misses,
281 struct perf_stat_output_ctx *out)
282{
283 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
284
285 print_ratio(config, evsel, aggr_idx, misses, out, STAT_DTLB_CACHE, thresh_ratios,
286 "of all dTLB cache accesses");
287}
288
289static void print_itlb_miss(struct perf_stat_config *config,
290 const struct evsel *evsel,
291 int aggr_idx, double misses,
292 struct perf_stat_output_ctx *out)
293{
294 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
295
296 print_ratio(config, evsel, aggr_idx, misses, out, STAT_ITLB_CACHE, thresh_ratios,
297 "of all iTLB cache accesses");
298}
299
300static void print_cache_miss(struct perf_stat_config *config,
301 const struct evsel *evsel,
302 int aggr_idx, double misses,
303 struct perf_stat_output_ctx *out)
304{
305 const double thresh_ratios[3] = {20.0, 10.0, 5.0};
306
307 print_ratio(config, evsel, aggr_idx, misses, out, STAT_CACHE_REFS, thresh_ratios,
308 "of all cache refs");
309}
310
311static void print_instructions(struct perf_stat_config *config,
312 const struct evsel *evsel,
313 int aggr_idx, double instructions,
314 struct perf_stat_output_ctx *out)
315{
316 print_metric_t print_metric = out->print_metric;
317 void *ctxp = out->ctx;
318 double cycles = find_stat(evsel, aggr_idx, STAT_CYCLES);
319 double max_stalled = max(find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_FRONT),
320 find_stat(evsel, aggr_idx, STAT_STALLED_CYCLES_BACK));
321
322 if (cycles) {
323 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
324 "insn per cycle", instructions / cycles);
325 } else {
326 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
327 "insn per cycle", 0);
328 }
329 if (max_stalled && instructions) {
330 out->new_line(config, ctxp);
331 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%7.2f ",
332 "stalled cycles per insn", max_stalled / instructions);
333 }
334}
335
336static void print_cycles(struct perf_stat_config *config,
337 const struct evsel *evsel,
338 int aggr_idx, double cycles,
339 struct perf_stat_output_ctx *out)
340{
341 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
342
343 if (cycles && nsecs) {
344 double ratio = cycles / nsecs;
345
346 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
347 "GHz", ratio);
348 } else {
349 out->print_metric(config, out->ctx, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
350 "GHz", 0);
351 }
352}
353
354static void print_nsecs(struct perf_stat_config *config,
355 const struct evsel *evsel,
356 int aggr_idx __maybe_unused, double nsecs,
357 struct perf_stat_output_ctx *out)
358{
359 print_metric_t print_metric = out->print_metric;
360 void *ctxp = out->ctx;
361 double wall_time = avg_stats(&walltime_nsecs_stats);
362
363 if (wall_time) {
364 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f", "CPUs utilized",
365 nsecs / (wall_time * evsel->scale));
366 } else {
367 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, /*fmt=*/NULL,
368 "CPUs utilized", 0);
369 }
370}
371
372static int prepare_metric(const struct metric_expr *mexp,
373 const struct evsel *evsel,
374 struct expr_parse_ctx *pctx,
375 int aggr_idx)
376{
377 struct evsel * const *metric_events = mexp->metric_events;
378 struct metric_ref *metric_refs = mexp->metric_refs;
379 int i;
380
381 for (i = 0; metric_events[i]; i++) {
382 char *n;
383 double val;
384 int source_count = 0;
385
386 if (evsel__is_tool(metric_events[i])) {
387 struct stats *stats;
388 double scale;
389
390 switch (evsel__tool_event(metric_events[i])) {
391 case TOOL_PMU__EVENT_DURATION_TIME:
392 stats = &walltime_nsecs_stats;
393 scale = 1e-9;
394 break;
395 case TOOL_PMU__EVENT_USER_TIME:
396 stats = &ru_stats.ru_utime_usec_stat;
397 scale = 1e-6;
398 break;
399 case TOOL_PMU__EVENT_SYSTEM_TIME:
400 stats = &ru_stats.ru_stime_usec_stat;
401 scale = 1e-6;
402 break;
403 case TOOL_PMU__EVENT_NONE:
404 pr_err("Invalid tool event 'none'");
405 abort();
406 case TOOL_PMU__EVENT_MAX:
407 pr_err("Invalid tool event 'max'");
408 abort();
409 case TOOL_PMU__EVENT_HAS_PMEM:
410 case TOOL_PMU__EVENT_NUM_CORES:
411 case TOOL_PMU__EVENT_NUM_CPUS:
412 case TOOL_PMU__EVENT_NUM_CPUS_ONLINE:
413 case TOOL_PMU__EVENT_NUM_DIES:
414 case TOOL_PMU__EVENT_NUM_PACKAGES:
415 case TOOL_PMU__EVENT_SLOTS:
416 case TOOL_PMU__EVENT_SMT_ON:
417 case TOOL_PMU__EVENT_SYSTEM_TSC_FREQ:
418 default:
419 pr_err("Unexpected tool event '%s'", evsel__name(metric_events[i]));
420 abort();
421 }
422 val = avg_stats(stats) * scale;
423 source_count = 1;
424 } else {
425 struct perf_stat_evsel *ps = metric_events[i]->stats;
426 struct perf_stat_aggr *aggr;
427
428 /*
429 * If there are multiple uncore PMUs and we're not
430 * reading the leader's stats, determine the stats for
431 * the appropriate uncore PMU.
432 */
433 if (evsel && evsel->metric_leader &&
434 evsel->pmu != evsel->metric_leader->pmu &&
435 mexp->metric_events[i]->pmu == evsel->metric_leader->pmu) {
436 struct evsel *pos;
437
438 evlist__for_each_entry(evsel->evlist, pos) {
439 if (pos->pmu != evsel->pmu)
440 continue;
441 if (pos->metric_leader != mexp->metric_events[i])
442 continue;
443 ps = pos->stats;
444 source_count = 1;
445 break;
446 }
447 }
448 aggr = &ps->aggr[aggr_idx];
449 if (!aggr)
450 break;
451
452 if (!metric_events[i]->supported) {
453 /*
454 * Not supported events will have a count of 0,
455 * which can be confusing in a
456 * metric. Explicitly set the value to NAN. Not
457 * counted events (enable time of 0) are read as
458 * 0.
459 */
460 val = NAN;
461 source_count = 0;
462 } else {
463 val = aggr->counts.val;
464 if (!source_count)
465 source_count = evsel__source_count(metric_events[i]);
466 }
467 }
468 n = strdup(evsel__metric_id(metric_events[i]));
469 if (!n)
470 return -ENOMEM;
471
472 expr__add_id_val_source_count(pctx, n, val, source_count);
473 }
474
475 for (int j = 0; metric_refs && metric_refs[j].metric_name; j++) {
476 int ret = expr__add_ref(pctx, &metric_refs[j]);
477
478 if (ret)
479 return ret;
480 }
481
482 return i;
483}
484
485static void generic_metric(struct perf_stat_config *config,
486 struct metric_expr *mexp,
487 struct evsel *evsel,
488 int aggr_idx,
489 struct perf_stat_output_ctx *out)
490{
491 print_metric_t print_metric = out->print_metric;
492 const char *metric_name = mexp->metric_name;
493 const char *metric_expr = mexp->metric_expr;
494 const char *metric_threshold = mexp->metric_threshold;
495 const char *metric_unit = mexp->metric_unit;
496 struct evsel * const *metric_events = mexp->metric_events;
497 int runtime = mexp->runtime;
498 struct expr_parse_ctx *pctx;
499 double ratio, scale, threshold;
500 int i;
501 void *ctxp = out->ctx;
502 enum metric_threshold_classify thresh = METRIC_THRESHOLD_UNKNOWN;
503
504 pctx = expr__ctx_new();
505 if (!pctx)
506 return;
507
508 if (config->user_requested_cpu_list)
509 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
510 pctx->sctx.runtime = runtime;
511 pctx->sctx.system_wide = config->system_wide;
512 i = prepare_metric(mexp, evsel, pctx, aggr_idx);
513 if (i < 0) {
514 expr__ctx_free(pctx);
515 return;
516 }
517 if (!metric_events[i]) {
518 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
519 char *unit;
520 char metric_bf[128];
521
522 if (metric_threshold &&
523 expr__parse(&threshold, pctx, metric_threshold) == 0 &&
524 !isnan(threshold)) {
525 thresh = fpclassify(threshold) == FP_ZERO
526 ? METRIC_THRESHOLD_GOOD : METRIC_THRESHOLD_BAD;
527 }
528
529 if (metric_unit && metric_name) {
530 if (perf_pmu__convert_scale(metric_unit,
531 &unit, &scale) >= 0) {
532 ratio *= scale;
533 }
534 if (strstr(metric_expr, "?"))
535 scnprintf(metric_bf, sizeof(metric_bf),
536 "%s %s_%d", unit, metric_name, runtime);
537 else
538 scnprintf(metric_bf, sizeof(metric_bf),
539 "%s %s", unit, metric_name);
540
541 print_metric(config, ctxp, thresh, "%8.1f",
542 metric_bf, ratio);
543 } else {
544 print_metric(config, ctxp, thresh, "%8.2f",
545 metric_name ?
546 metric_name :
547 out->force_header ? evsel->name : "",
548 ratio);
549 }
550 } else {
551 print_metric(config, ctxp, thresh, /*fmt=*/NULL,
552 out->force_header ?
553 (metric_name ?: evsel->name) : "", 0);
554 }
555 } else {
556 print_metric(config, ctxp, thresh, /*fmt=*/NULL,
557 out->force_header ?
558 (metric_name ?: evsel->name) : "", 0);
559 }
560
561 expr__ctx_free(pctx);
562}
563
564double test_generic_metric(struct metric_expr *mexp, int aggr_idx)
565{
566 struct expr_parse_ctx *pctx;
567 double ratio = 0.0;
568
569 pctx = expr__ctx_new();
570 if (!pctx)
571 return NAN;
572
573 if (prepare_metric(mexp, /*evsel=*/NULL, pctx, aggr_idx) < 0)
574 goto out;
575
576 if (expr__parse(&ratio, pctx, mexp->metric_expr))
577 ratio = 0.0;
578
579out:
580 expr__ctx_free(pctx);
581 return ratio;
582}
583
584static void perf_stat__print_metricgroup_header(struct perf_stat_config *config,
585 struct evsel *evsel,
586 void *ctxp,
587 const char *name,
588 struct perf_stat_output_ctx *out)
589{
590 bool need_full_name = perf_pmus__num_core_pmus() > 1;
591 static const char *last_name;
592 static const struct perf_pmu *last_pmu;
593 char full_name[64];
594
595 /*
596 * A metricgroup may have several metric events,
597 * e.g.,TopdownL1 on e-core of ADL.
598 * The name has been output by the first metric
599 * event. Only align with other metics from
600 * different metric events.
601 */
602 if (last_name && !strcmp(last_name, name)) {
603 if (!need_full_name || last_pmu != evsel->pmu) {
604 out->print_metricgroup_header(config, ctxp, NULL);
605 return;
606 }
607 }
608
609 if (need_full_name && evsel->pmu)
610 scnprintf(full_name, sizeof(full_name), "%s (%s)", name, evsel->pmu->name);
611 else
612 scnprintf(full_name, sizeof(full_name), "%s", name);
613
614 out->print_metricgroup_header(config, ctxp, full_name);
615
616 last_name = name;
617 last_pmu = evsel->pmu;
618}
619
620/**
621 * perf_stat__print_shadow_stats_metricgroup - Print out metrics associated with the evsel
622 * For the non-default, all metrics associated
623 * with the evsel are printed.
624 * For the default mode, only the metrics from
625 * the same metricgroup and the name of the
626 * metricgroup are printed. To print the metrics
627 * from the next metricgroup (if available),
628 * invoke the function with correspoinding
629 * metric_expr.
630 */
631void *perf_stat__print_shadow_stats_metricgroup(struct perf_stat_config *config,
632 struct evsel *evsel,
633 int aggr_idx,
634 int *num,
635 void *from,
636 struct perf_stat_output_ctx *out,
637 struct rblist *metric_events)
638{
639 struct metric_event *me;
640 struct metric_expr *mexp = from;
641 void *ctxp = out->ctx;
642 bool header_printed = false;
643 const char *name = NULL;
644
645 me = metricgroup__lookup(metric_events, evsel, false);
646 if (me == NULL)
647 return NULL;
648
649 if (!mexp)
650 mexp = list_first_entry(&me->head, typeof(*mexp), nd);
651
652 list_for_each_entry_from(mexp, &me->head, nd) {
653 /* Print the display name of the Default metricgroup */
654 if (!config->metric_only && me->is_default) {
655 if (!name)
656 name = mexp->default_metricgroup_name;
657 /*
658 * Two or more metricgroup may share the same metric
659 * event, e.g., TopdownL1 and TopdownL2 on SPR.
660 * Return and print the prefix, e.g., noise, running
661 * for the next metricgroup.
662 */
663 if (strcmp(name, mexp->default_metricgroup_name))
664 return (void *)mexp;
665 /* Only print the name of the metricgroup once */
666 if (!header_printed) {
667 header_printed = true;
668 perf_stat__print_metricgroup_header(config, evsel, ctxp,
669 name, out);
670 }
671 }
672
673 if ((*num)++ > 0)
674 out->new_line(config, ctxp);
675 generic_metric(config, mexp, evsel, aggr_idx, out);
676 }
677
678 return NULL;
679}
680
681void perf_stat__print_shadow_stats(struct perf_stat_config *config,
682 struct evsel *evsel,
683 double avg, int aggr_idx,
684 struct perf_stat_output_ctx *out,
685 struct rblist *metric_events)
686{
687 typedef void (*stat_print_function_t)(struct perf_stat_config *config,
688 const struct evsel *evsel,
689 int aggr_idx, double misses,
690 struct perf_stat_output_ctx *out);
691 static const stat_print_function_t stat_print_function[STAT_MAX] = {
692 [STAT_INSTRUCTIONS] = print_instructions,
693 [STAT_BRANCH_MISS] = print_branch_miss,
694 [STAT_L1D_MISS] = print_l1d_miss,
695 [STAT_L1I_MISS] = print_l1i_miss,
696 [STAT_DTLB_MISS] = print_dtlb_miss,
697 [STAT_ITLB_MISS] = print_itlb_miss,
698 [STAT_LL_MISS] = print_ll_miss,
699 [STAT_CACHE_MISSES] = print_cache_miss,
700 [STAT_STALLED_CYCLES_FRONT] = print_stalled_cycles_front,
701 [STAT_STALLED_CYCLES_BACK] = print_stalled_cycles_back,
702 [STAT_CYCLES] = print_cycles,
703 [STAT_NSECS] = print_nsecs,
704 };
705 print_metric_t print_metric = out->print_metric;
706 void *ctxp = out->ctx;
707 int num = 1;
708
709 if (config->iostat_run) {
710 iostat_print_metric(config, evsel, out);
711 } else {
712 stat_print_function_t fn = stat_print_function[evsel__stat_type(evsel)];
713
714 if (fn)
715 fn(config, evsel, aggr_idx, avg, out);
716 else {
717 double nsecs = find_stat(evsel, aggr_idx, STAT_NSECS);
718
719 if (nsecs) {
720 char unit = ' ';
721 char unit_buf[10] = "/sec";
722 double ratio = convert_unit_double(1000000000.0 * avg / nsecs,
723 &unit);
724
725 if (unit != ' ')
726 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
727 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN, "%8.3f",
728 unit_buf, ratio);
729 } else {
730 num = 0;
731 }
732 }
733 }
734
735 perf_stat__print_shadow_stats_metricgroup(config, evsel, aggr_idx,
736 &num, NULL, out, metric_events);
737
738 if (num == 0) {
739 print_metric(config, ctxp, METRIC_THRESHOLD_UNKNOWN,
740 /*fmt=*/NULL, /*unit=*/NULL, 0);
741 }
742}
743
744/**
745 * perf_stat__skip_metric_event - Skip the evsel in the Default metricgroup,
746 * if it's not running or not the metric event.
747 */
748bool perf_stat__skip_metric_event(struct evsel *evsel,
749 struct rblist *metric_events,
750 u64 ena, u64 run)
751{
752 if (!evsel->default_metricgroup)
753 return false;
754
755 if (!ena || !run)
756 return true;
757
758 return !metricgroup__lookup(metric_events, evsel, false);
759}