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