Loading...
1// SPDX-License-Identifier: GPL-2.0
2#include <math.h>
3#include <stdio.h>
4#include "evsel.h"
5#include "stat.h"
6#include "color.h"
7#include "debug.h"
8#include "pmu.h"
9#include "rblist.h"
10#include "evlist.h"
11#include "expr.h"
12#include "metricgroup.h"
13#include "cgroup.h"
14#include "units.h"
15#include <linux/zalloc.h>
16#include "iostat.h"
17#include "util/hashmap.h"
18
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}
1// SPDX-License-Identifier: GPL-2.0
2#include <stdio.h>
3#include "evsel.h"
4#include "stat.h"
5#include "color.h"
6#include "pmu.h"
7#include "rblist.h"
8#include "evlist.h"
9#include "expr.h"
10#include "metricgroup.h"
11#include "cgroup.h"
12#include "units.h"
13#include <linux/zalloc.h>
14#include "iostat.h"
15
16/*
17 * AGGR_GLOBAL: Use CPU 0
18 * AGGR_SOCKET: Use first CPU of socket
19 * AGGR_DIE: Use first CPU of die
20 * AGGR_CORE: Use first CPU of core
21 * AGGR_NONE: Use matching CPU
22 * AGGR_THREAD: Not supported?
23 */
24
25struct runtime_stat rt_stat;
26struct stats walltime_nsecs_stats;
27
28struct saved_value {
29 struct rb_node rb_node;
30 struct evsel *evsel;
31 enum stat_type type;
32 int ctx;
33 int cpu;
34 struct cgroup *cgrp;
35 struct runtime_stat *stat;
36 struct stats stats;
37 u64 metric_total;
38 int metric_other;
39};
40
41static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
42{
43 struct saved_value *a = container_of(rb_node,
44 struct saved_value,
45 rb_node);
46 const struct saved_value *b = entry;
47
48 if (a->cpu != b->cpu)
49 return a->cpu - b->cpu;
50
51 /*
52 * Previously the rbtree was used to link generic metrics.
53 * The keys were evsel/cpu. Now the rbtree is extended to support
54 * per-thread shadow stats. For shadow stats case, the keys
55 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
56 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
57 */
58 if (a->type != b->type)
59 return a->type - b->type;
60
61 if (a->ctx != b->ctx)
62 return a->ctx - b->ctx;
63
64 if (a->cgrp != b->cgrp)
65 return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1;
66
67 if (a->evsel == NULL && b->evsel == NULL) {
68 if (a->stat == b->stat)
69 return 0;
70
71 if ((char *)a->stat < (char *)b->stat)
72 return -1;
73
74 return 1;
75 }
76
77 if (a->evsel == b->evsel)
78 return 0;
79 if ((char *)a->evsel < (char *)b->evsel)
80 return -1;
81 return +1;
82}
83
84static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
85 const void *entry)
86{
87 struct saved_value *nd = malloc(sizeof(struct saved_value));
88
89 if (!nd)
90 return NULL;
91 memcpy(nd, entry, sizeof(struct saved_value));
92 return &nd->rb_node;
93}
94
95static void saved_value_delete(struct rblist *rblist __maybe_unused,
96 struct rb_node *rb_node)
97{
98 struct saved_value *v;
99
100 BUG_ON(!rb_node);
101 v = container_of(rb_node, struct saved_value, rb_node);
102 free(v);
103}
104
105static struct saved_value *saved_value_lookup(struct evsel *evsel,
106 int cpu,
107 bool create,
108 enum stat_type type,
109 int ctx,
110 struct runtime_stat *st,
111 struct cgroup *cgrp)
112{
113 struct rblist *rblist;
114 struct rb_node *nd;
115 struct saved_value dm = {
116 .cpu = cpu,
117 .evsel = evsel,
118 .type = type,
119 .ctx = ctx,
120 .stat = st,
121 .cgrp = cgrp,
122 };
123
124 rblist = &st->value_list;
125
126 /* don't use context info for clock events */
127 if (type == STAT_NSECS)
128 dm.ctx = 0;
129
130 nd = rblist__find(rblist, &dm);
131 if (nd)
132 return container_of(nd, struct saved_value, rb_node);
133 if (create) {
134 rblist__add_node(rblist, &dm);
135 nd = rblist__find(rblist, &dm);
136 if (nd)
137 return container_of(nd, struct saved_value, rb_node);
138 }
139 return NULL;
140}
141
142void runtime_stat__init(struct runtime_stat *st)
143{
144 struct rblist *rblist = &st->value_list;
145
146 rblist__init(rblist);
147 rblist->node_cmp = saved_value_cmp;
148 rblist->node_new = saved_value_new;
149 rblist->node_delete = saved_value_delete;
150}
151
152void runtime_stat__exit(struct runtime_stat *st)
153{
154 rblist__exit(&st->value_list);
155}
156
157void perf_stat__init_shadow_stats(void)
158{
159 runtime_stat__init(&rt_stat);
160}
161
162static int evsel_context(struct evsel *evsel)
163{
164 int ctx = 0;
165
166 if (evsel->core.attr.exclude_kernel)
167 ctx |= CTX_BIT_KERNEL;
168 if (evsel->core.attr.exclude_user)
169 ctx |= CTX_BIT_USER;
170 if (evsel->core.attr.exclude_hv)
171 ctx |= CTX_BIT_HV;
172 if (evsel->core.attr.exclude_host)
173 ctx |= CTX_BIT_HOST;
174 if (evsel->core.attr.exclude_idle)
175 ctx |= CTX_BIT_IDLE;
176
177 return ctx;
178}
179
180static void reset_stat(struct runtime_stat *st)
181{
182 struct rblist *rblist;
183 struct rb_node *pos, *next;
184
185 rblist = &st->value_list;
186 next = rb_first_cached(&rblist->entries);
187 while (next) {
188 pos = next;
189 next = rb_next(pos);
190 memset(&container_of(pos, struct saved_value, rb_node)->stats,
191 0,
192 sizeof(struct stats));
193 }
194}
195
196void perf_stat__reset_shadow_stats(void)
197{
198 reset_stat(&rt_stat);
199 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
200}
201
202void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
203{
204 reset_stat(st);
205}
206
207struct runtime_stat_data {
208 int ctx;
209 struct cgroup *cgrp;
210};
211
212static void update_runtime_stat(struct runtime_stat *st,
213 enum stat_type type,
214 int cpu, u64 count,
215 struct runtime_stat_data *rsd)
216{
217 struct saved_value *v = saved_value_lookup(NULL, cpu, true, type,
218 rsd->ctx, st, rsd->cgrp);
219
220 if (v)
221 update_stats(&v->stats, count);
222}
223
224/*
225 * Update various tracking values we maintain to print
226 * more semantic information such as miss/hit ratios,
227 * instruction rates, etc:
228 */
229void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
230 int cpu, struct runtime_stat *st)
231{
232 u64 count_ns = count;
233 struct saved_value *v;
234 struct runtime_stat_data rsd = {
235 .ctx = evsel_context(counter),
236 .cgrp = counter->cgrp,
237 };
238
239 count *= counter->scale;
240
241 if (evsel__is_clock(counter))
242 update_runtime_stat(st, STAT_NSECS, cpu, count_ns, &rsd);
243 else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
244 update_runtime_stat(st, STAT_CYCLES, cpu, count, &rsd);
245 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
246 update_runtime_stat(st, STAT_CYCLES_IN_TX, cpu, count, &rsd);
247 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
248 update_runtime_stat(st, STAT_TRANSACTION, cpu, count, &rsd);
249 else if (perf_stat_evsel__is(counter, ELISION_START))
250 update_runtime_stat(st, STAT_ELISION, cpu, count, &rsd);
251 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
252 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
253 cpu, count, &rsd);
254 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
255 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
256 cpu, count, &rsd);
257 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
258 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
259 cpu, count, &rsd);
260 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
261 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
262 cpu, count, &rsd);
263 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
264 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
265 cpu, count, &rsd);
266 else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
267 update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
268 cpu, count, &rsd);
269 else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
270 update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
271 cpu, count, &rsd);
272 else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
273 update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
274 cpu, count, &rsd);
275 else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
276 update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
277 cpu, count, &rsd);
278 else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS))
279 update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS,
280 cpu, count, &rsd);
281 else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT))
282 update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT,
283 cpu, count, &rsd);
284 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT))
285 update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT,
286 cpu, count, &rsd);
287 else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND))
288 update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND,
289 cpu, count, &rsd);
290 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
291 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
292 cpu, count, &rsd);
293 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
294 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
295 cpu, count, &rsd);
296 else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
297 update_runtime_stat(st, STAT_BRANCHES, cpu, count, &rsd);
298 else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
299 update_runtime_stat(st, STAT_CACHEREFS, cpu, count, &rsd);
300 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
301 update_runtime_stat(st, STAT_L1_DCACHE, cpu, count, &rsd);
302 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
303 update_runtime_stat(st, STAT_L1_ICACHE, cpu, count, &rsd);
304 else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
305 update_runtime_stat(st, STAT_LL_CACHE, cpu, count, &rsd);
306 else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
307 update_runtime_stat(st, STAT_DTLB_CACHE, cpu, count, &rsd);
308 else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
309 update_runtime_stat(st, STAT_ITLB_CACHE, cpu, count, &rsd);
310 else if (perf_stat_evsel__is(counter, SMI_NUM))
311 update_runtime_stat(st, STAT_SMI_NUM, cpu, count, &rsd);
312 else if (perf_stat_evsel__is(counter, APERF))
313 update_runtime_stat(st, STAT_APERF, cpu, count, &rsd);
314
315 if (counter->collect_stat) {
316 v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st,
317 rsd.cgrp);
318 update_stats(&v->stats, count);
319 if (counter->metric_leader)
320 v->metric_total += count;
321 } else if (counter->metric_leader) {
322 v = saved_value_lookup(counter->metric_leader,
323 cpu, true, STAT_NONE, 0, st, rsd.cgrp);
324 v->metric_total += count;
325 v->metric_other++;
326 }
327}
328
329/* used for get_ratio_color() */
330enum grc_type {
331 GRC_STALLED_CYCLES_FE,
332 GRC_STALLED_CYCLES_BE,
333 GRC_CACHE_MISSES,
334 GRC_MAX_NR
335};
336
337static const char *get_ratio_color(enum grc_type type, double ratio)
338{
339 static const double grc_table[GRC_MAX_NR][3] = {
340 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
341 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
342 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
343 };
344 const char *color = PERF_COLOR_NORMAL;
345
346 if (ratio > grc_table[type][0])
347 color = PERF_COLOR_RED;
348 else if (ratio > grc_table[type][1])
349 color = PERF_COLOR_MAGENTA;
350 else if (ratio > grc_table[type][2])
351 color = PERF_COLOR_YELLOW;
352
353 return color;
354}
355
356static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
357 const char *name)
358{
359 struct evsel *c2;
360
361 evlist__for_each_entry (evsel_list, c2) {
362 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
363 return c2;
364 }
365 return NULL;
366}
367
368/* Mark MetricExpr target events and link events using them to them. */
369void perf_stat__collect_metric_expr(struct evlist *evsel_list)
370{
371 struct evsel *counter, *leader, **metric_events, *oc;
372 bool found;
373 struct expr_parse_ctx ctx;
374 struct hashmap_entry *cur;
375 size_t bkt;
376 int i;
377
378 expr__ctx_init(&ctx);
379 evlist__for_each_entry(evsel_list, counter) {
380 bool invalid = false;
381
382 leader = evsel__leader(counter);
383 if (!counter->metric_expr)
384 continue;
385
386 expr__ctx_clear(&ctx);
387 metric_events = counter->metric_events;
388 if (!metric_events) {
389 if (expr__find_other(counter->metric_expr,
390 counter->name,
391 &ctx, 1) < 0)
392 continue;
393
394 metric_events = calloc(sizeof(struct evsel *),
395 hashmap__size(&ctx.ids) + 1);
396 if (!metric_events) {
397 expr__ctx_clear(&ctx);
398 return;
399 }
400 counter->metric_events = metric_events;
401 }
402
403 i = 0;
404 hashmap__for_each_entry((&ctx.ids), cur, bkt) {
405 const char *metric_name = (const char *)cur->key;
406
407 found = false;
408 if (leader) {
409 /* Search in group */
410 for_each_group_member (oc, leader) {
411 if (!strcasecmp(oc->name,
412 metric_name) &&
413 !oc->collect_stat) {
414 found = true;
415 break;
416 }
417 }
418 }
419 if (!found) {
420 /* Search ignoring groups */
421 oc = perf_stat__find_event(evsel_list,
422 metric_name);
423 }
424 if (!oc) {
425 /* Deduping one is good enough to handle duplicated PMUs. */
426 static char *printed;
427
428 /*
429 * Adding events automatically would be difficult, because
430 * it would risk creating groups that are not schedulable.
431 * perf stat doesn't understand all the scheduling constraints
432 * of events. So we ask the user instead to add the missing
433 * events.
434 */
435 if (!printed ||
436 strcasecmp(printed, metric_name)) {
437 fprintf(stderr,
438 "Add %s event to groups to get metric expression for %s\n",
439 metric_name,
440 counter->name);
441 printed = strdup(metric_name);
442 }
443 invalid = true;
444 continue;
445 }
446 metric_events[i++] = oc;
447 oc->collect_stat = true;
448 }
449 metric_events[i] = NULL;
450 if (invalid) {
451 free(metric_events);
452 counter->metric_events = NULL;
453 counter->metric_expr = NULL;
454 }
455 }
456 expr__ctx_clear(&ctx);
457}
458
459static double runtime_stat_avg(struct runtime_stat *st,
460 enum stat_type type, int cpu,
461 struct runtime_stat_data *rsd)
462{
463 struct saved_value *v;
464
465 v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
466 if (!v)
467 return 0.0;
468
469 return avg_stats(&v->stats);
470}
471
472static double runtime_stat_n(struct runtime_stat *st,
473 enum stat_type type, int cpu,
474 struct runtime_stat_data *rsd)
475{
476 struct saved_value *v;
477
478 v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
479 if (!v)
480 return 0.0;
481
482 return v->stats.n;
483}
484
485static void print_stalled_cycles_frontend(struct perf_stat_config *config,
486 int cpu, double avg,
487 struct perf_stat_output_ctx *out,
488 struct runtime_stat *st,
489 struct runtime_stat_data *rsd)
490{
491 double total, ratio = 0.0;
492 const char *color;
493
494 total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
495
496 if (total)
497 ratio = avg / total * 100.0;
498
499 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
500
501 if (ratio)
502 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
503 ratio);
504 else
505 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
506}
507
508static void print_stalled_cycles_backend(struct perf_stat_config *config,
509 int cpu, double avg,
510 struct perf_stat_output_ctx *out,
511 struct runtime_stat *st,
512 struct runtime_stat_data *rsd)
513{
514 double total, ratio = 0.0;
515 const char *color;
516
517 total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
518
519 if (total)
520 ratio = avg / total * 100.0;
521
522 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
523
524 out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
525}
526
527static void print_branch_misses(struct perf_stat_config *config,
528 int cpu, double avg,
529 struct perf_stat_output_ctx *out,
530 struct runtime_stat *st,
531 struct runtime_stat_data *rsd)
532{
533 double total, ratio = 0.0;
534 const char *color;
535
536 total = runtime_stat_avg(st, STAT_BRANCHES, cpu, rsd);
537
538 if (total)
539 ratio = avg / total * 100.0;
540
541 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
542
543 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
544}
545
546static void print_l1_dcache_misses(struct perf_stat_config *config,
547 int cpu, double avg,
548 struct perf_stat_output_ctx *out,
549 struct runtime_stat *st,
550 struct runtime_stat_data *rsd)
551{
552 double total, ratio = 0.0;
553 const char *color;
554
555 total = runtime_stat_avg(st, STAT_L1_DCACHE, cpu, rsd);
556
557 if (total)
558 ratio = avg / total * 100.0;
559
560 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
561
562 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
563}
564
565static void print_l1_icache_misses(struct perf_stat_config *config,
566 int cpu, double avg,
567 struct perf_stat_output_ctx *out,
568 struct runtime_stat *st,
569 struct runtime_stat_data *rsd)
570{
571 double total, ratio = 0.0;
572 const char *color;
573
574 total = runtime_stat_avg(st, STAT_L1_ICACHE, cpu, rsd);
575
576 if (total)
577 ratio = avg / total * 100.0;
578
579 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
580 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
581}
582
583static void print_dtlb_cache_misses(struct perf_stat_config *config,
584 int cpu, double avg,
585 struct perf_stat_output_ctx *out,
586 struct runtime_stat *st,
587 struct runtime_stat_data *rsd)
588{
589 double total, ratio = 0.0;
590 const char *color;
591
592 total = runtime_stat_avg(st, STAT_DTLB_CACHE, cpu, rsd);
593
594 if (total)
595 ratio = avg / total * 100.0;
596
597 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
598 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
599}
600
601static void print_itlb_cache_misses(struct perf_stat_config *config,
602 int cpu, double avg,
603 struct perf_stat_output_ctx *out,
604 struct runtime_stat *st,
605 struct runtime_stat_data *rsd)
606{
607 double total, ratio = 0.0;
608 const char *color;
609
610 total = runtime_stat_avg(st, STAT_ITLB_CACHE, cpu, rsd);
611
612 if (total)
613 ratio = avg / total * 100.0;
614
615 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
616 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
617}
618
619static void print_ll_cache_misses(struct perf_stat_config *config,
620 int cpu, double avg,
621 struct perf_stat_output_ctx *out,
622 struct runtime_stat *st,
623 struct runtime_stat_data *rsd)
624{
625 double total, ratio = 0.0;
626 const char *color;
627
628 total = runtime_stat_avg(st, STAT_LL_CACHE, cpu, rsd);
629
630 if (total)
631 ratio = avg / total * 100.0;
632
633 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
634 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
635}
636
637/*
638 * High level "TopDown" CPU core pipe line bottleneck break down.
639 *
640 * Basic concept following
641 * Yasin, A Top Down Method for Performance analysis and Counter architecture
642 * ISPASS14
643 *
644 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
645 *
646 * Frontend -> Backend -> Retiring
647 * BadSpeculation in addition means out of order execution that is thrown away
648 * (for example branch mispredictions)
649 * Frontend is instruction decoding.
650 * Backend is execution, like computation and accessing data in memory
651 * Retiring is good execution that is not directly bottlenecked
652 *
653 * The formulas are computed in slots.
654 * A slot is an entry in the pipeline each for the pipeline width
655 * (for example a 4-wide pipeline has 4 slots for each cycle)
656 *
657 * Formulas:
658 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
659 * TotalSlots
660 * Retiring = SlotsRetired / TotalSlots
661 * FrontendBound = FetchBubbles / TotalSlots
662 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
663 *
664 * The kernel provides the mapping to the low level CPU events and any scaling
665 * needed for the CPU pipeline width, for example:
666 *
667 * TotalSlots = Cycles * 4
668 *
669 * The scaling factor is communicated in the sysfs unit.
670 *
671 * In some cases the CPU may not be able to measure all the formulas due to
672 * missing events. In this case multiple formulas are combined, as possible.
673 *
674 * Full TopDown supports more levels to sub-divide each area: for example
675 * BackendBound into computing bound and memory bound. For now we only
676 * support Level 1 TopDown.
677 */
678
679static double sanitize_val(double x)
680{
681 if (x < 0 && x >= -0.02)
682 return 0.0;
683 return x;
684}
685
686static double td_total_slots(int cpu, struct runtime_stat *st,
687 struct runtime_stat_data *rsd)
688{
689 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, cpu, rsd);
690}
691
692static double td_bad_spec(int cpu, struct runtime_stat *st,
693 struct runtime_stat_data *rsd)
694{
695 double bad_spec = 0;
696 double total_slots;
697 double total;
698
699 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, cpu, rsd) -
700 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, cpu, rsd) +
701 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, cpu, rsd);
702
703 total_slots = td_total_slots(cpu, st, rsd);
704 if (total_slots)
705 bad_spec = total / total_slots;
706 return sanitize_val(bad_spec);
707}
708
709static double td_retiring(int cpu, struct runtime_stat *st,
710 struct runtime_stat_data *rsd)
711{
712 double retiring = 0;
713 double total_slots = td_total_slots(cpu, st, rsd);
714 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
715 cpu, rsd);
716
717 if (total_slots)
718 retiring = ret_slots / total_slots;
719 return retiring;
720}
721
722static double td_fe_bound(int cpu, struct runtime_stat *st,
723 struct runtime_stat_data *rsd)
724{
725 double fe_bound = 0;
726 double total_slots = td_total_slots(cpu, st, rsd);
727 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
728 cpu, rsd);
729
730 if (total_slots)
731 fe_bound = fetch_bub / total_slots;
732 return fe_bound;
733}
734
735static double td_be_bound(int cpu, struct runtime_stat *st,
736 struct runtime_stat_data *rsd)
737{
738 double sum = (td_fe_bound(cpu, st, rsd) +
739 td_bad_spec(cpu, st, rsd) +
740 td_retiring(cpu, st, rsd));
741 if (sum == 0)
742 return 0;
743 return sanitize_val(1.0 - sum);
744}
745
746/*
747 * Kernel reports metrics multiplied with slots. To get back
748 * the ratios we need to recreate the sum.
749 */
750
751static double td_metric_ratio(int cpu, enum stat_type type,
752 struct runtime_stat *stat,
753 struct runtime_stat_data *rsd)
754{
755 double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) +
756 runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) +
757 runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) +
758 runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd);
759 double d = runtime_stat_avg(stat, type, cpu, rsd);
760
761 if (sum)
762 return d / sum;
763 return 0;
764}
765
766/*
767 * ... but only if most of the values are actually available.
768 * We allow two missing.
769 */
770
771static bool full_td(int cpu, struct runtime_stat *stat,
772 struct runtime_stat_data *rsd)
773{
774 int c = 0;
775
776 if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) > 0)
777 c++;
778 if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) > 0)
779 c++;
780 if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) > 0)
781 c++;
782 if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd) > 0)
783 c++;
784 return c >= 2;
785}
786
787static void print_smi_cost(struct perf_stat_config *config, int cpu,
788 struct perf_stat_output_ctx *out,
789 struct runtime_stat *st,
790 struct runtime_stat_data *rsd)
791{
792 double smi_num, aperf, cycles, cost = 0.0;
793 const char *color = NULL;
794
795 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, cpu, rsd);
796 aperf = runtime_stat_avg(st, STAT_APERF, cpu, rsd);
797 cycles = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
798
799 if ((cycles == 0) || (aperf == 0))
800 return;
801
802 if (smi_num)
803 cost = (aperf - cycles) / aperf * 100.00;
804
805 if (cost > 10)
806 color = PERF_COLOR_RED;
807 out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
808 out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
809}
810
811static int prepare_metric(struct evsel **metric_events,
812 struct metric_ref *metric_refs,
813 struct expr_parse_ctx *pctx,
814 int cpu,
815 struct runtime_stat *st)
816{
817 double scale;
818 char *n, *pn;
819 int i, j, ret;
820
821 expr__ctx_init(pctx);
822 for (i = 0; metric_events[i]; i++) {
823 struct saved_value *v;
824 struct stats *stats;
825 u64 metric_total = 0;
826
827 if (!strcmp(metric_events[i]->name, "duration_time")) {
828 stats = &walltime_nsecs_stats;
829 scale = 1e-9;
830 } else {
831 v = saved_value_lookup(metric_events[i], cpu, false,
832 STAT_NONE, 0, st,
833 metric_events[i]->cgrp);
834 if (!v)
835 break;
836 stats = &v->stats;
837 scale = 1.0;
838
839 if (v->metric_other)
840 metric_total = v->metric_total;
841 }
842
843 n = strdup(metric_events[i]->name);
844 if (!n)
845 return -ENOMEM;
846 /*
847 * This display code with --no-merge adds [cpu] postfixes.
848 * These are not supported by the parser. Remove everything
849 * after the space.
850 */
851 pn = strchr(n, ' ');
852 if (pn)
853 *pn = 0;
854
855 if (metric_total)
856 expr__add_id_val(pctx, n, metric_total);
857 else
858 expr__add_id_val(pctx, n, avg_stats(stats)*scale);
859 }
860
861 for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
862 ret = expr__add_ref(pctx, &metric_refs[j]);
863 if (ret)
864 return ret;
865 }
866
867 return i;
868}
869
870static void generic_metric(struct perf_stat_config *config,
871 const char *metric_expr,
872 struct evsel **metric_events,
873 struct metric_ref *metric_refs,
874 char *name,
875 const char *metric_name,
876 const char *metric_unit,
877 int runtime,
878 int cpu,
879 struct perf_stat_output_ctx *out,
880 struct runtime_stat *st)
881{
882 print_metric_t print_metric = out->print_metric;
883 struct expr_parse_ctx pctx;
884 double ratio, scale;
885 int i;
886 void *ctxp = out->ctx;
887
888 i = prepare_metric(metric_events, metric_refs, &pctx, cpu, st);
889 if (i < 0)
890 return;
891
892 if (!metric_events[i]) {
893 if (expr__parse(&ratio, &pctx, metric_expr, runtime) == 0) {
894 char *unit;
895 char metric_bf[64];
896
897 if (metric_unit && metric_name) {
898 if (perf_pmu__convert_scale(metric_unit,
899 &unit, &scale) >= 0) {
900 ratio *= scale;
901 }
902 if (strstr(metric_expr, "?"))
903 scnprintf(metric_bf, sizeof(metric_bf),
904 "%s %s_%d", unit, metric_name, runtime);
905 else
906 scnprintf(metric_bf, sizeof(metric_bf),
907 "%s %s", unit, metric_name);
908
909 print_metric(config, ctxp, NULL, "%8.1f",
910 metric_bf, ratio);
911 } else {
912 print_metric(config, ctxp, NULL, "%8.2f",
913 metric_name ?
914 metric_name :
915 out->force_header ? name : "",
916 ratio);
917 }
918 } else {
919 print_metric(config, ctxp, NULL, NULL,
920 out->force_header ?
921 (metric_name ? metric_name : name) : "", 0);
922 }
923 } else {
924 print_metric(config, ctxp, NULL, NULL,
925 out->force_header ?
926 (metric_name ? metric_name : name) : "", 0);
927 }
928
929 expr__ctx_clear(&pctx);
930}
931
932double test_generic_metric(struct metric_expr *mexp, int cpu, struct runtime_stat *st)
933{
934 struct expr_parse_ctx pctx;
935 double ratio = 0.0;
936
937 if (prepare_metric(mexp->metric_events, mexp->metric_refs, &pctx, cpu, st) < 0)
938 goto out;
939
940 if (expr__parse(&ratio, &pctx, mexp->metric_expr, 1))
941 ratio = 0.0;
942
943out:
944 expr__ctx_clear(&pctx);
945 return ratio;
946}
947
948void perf_stat__print_shadow_stats(struct perf_stat_config *config,
949 struct evsel *evsel,
950 double avg, int cpu,
951 struct perf_stat_output_ctx *out,
952 struct rblist *metric_events,
953 struct runtime_stat *st)
954{
955 void *ctxp = out->ctx;
956 print_metric_t print_metric = out->print_metric;
957 double total, ratio = 0.0, total2;
958 const char *color = NULL;
959 struct runtime_stat_data rsd = {
960 .ctx = evsel_context(evsel),
961 .cgrp = evsel->cgrp,
962 };
963 struct metric_event *me;
964 int num = 1;
965
966 if (config->iostat_run) {
967 iostat_print_metric(config, evsel, out);
968 } else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
969 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
970
971 if (total) {
972 ratio = avg / total;
973 print_metric(config, ctxp, NULL, "%7.2f ",
974 "insn per cycle", ratio);
975 } else {
976 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
977 }
978
979 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, cpu, &rsd);
980
981 total = max(total, runtime_stat_avg(st,
982 STAT_STALLED_CYCLES_BACK,
983 cpu, &rsd));
984
985 if (total && avg) {
986 out->new_line(config, ctxp);
987 ratio = total / avg;
988 print_metric(config, ctxp, NULL, "%7.2f ",
989 "stalled cycles per insn",
990 ratio);
991 }
992 } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
993 if (runtime_stat_n(st, STAT_BRANCHES, cpu, &rsd) != 0)
994 print_branch_misses(config, cpu, avg, out, st, &rsd);
995 else
996 print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
997 } else if (
998 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
999 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1000 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1001 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1002
1003 if (runtime_stat_n(st, STAT_L1_DCACHE, cpu, &rsd) != 0)
1004 print_l1_dcache_misses(config, cpu, avg, out, st, &rsd);
1005 else
1006 print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
1007 } else if (
1008 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1009 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1010 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1011 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1012
1013 if (runtime_stat_n(st, STAT_L1_ICACHE, cpu, &rsd) != 0)
1014 print_l1_icache_misses(config, cpu, avg, out, st, &rsd);
1015 else
1016 print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
1017 } else if (
1018 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1019 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1020 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1021 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1022
1023 if (runtime_stat_n(st, STAT_DTLB_CACHE, cpu, &rsd) != 0)
1024 print_dtlb_cache_misses(config, cpu, avg, out, st, &rsd);
1025 else
1026 print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
1027 } else if (
1028 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1029 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1030 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1031 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1032
1033 if (runtime_stat_n(st, STAT_ITLB_CACHE, cpu, &rsd) != 0)
1034 print_itlb_cache_misses(config, cpu, avg, out, st, &rsd);
1035 else
1036 print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
1037 } else if (
1038 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1039 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL |
1040 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1041 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1042
1043 if (runtime_stat_n(st, STAT_LL_CACHE, cpu, &rsd) != 0)
1044 print_ll_cache_misses(config, cpu, avg, out, st, &rsd);
1045 else
1046 print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
1047 } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
1048 total = runtime_stat_avg(st, STAT_CACHEREFS, cpu, &rsd);
1049
1050 if (total)
1051 ratio = avg * 100 / total;
1052
1053 if (runtime_stat_n(st, STAT_CACHEREFS, cpu, &rsd) != 0)
1054 print_metric(config, ctxp, NULL, "%8.3f %%",
1055 "of all cache refs", ratio);
1056 else
1057 print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
1058 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1059 print_stalled_cycles_frontend(config, cpu, avg, out, st, &rsd);
1060 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1061 print_stalled_cycles_backend(config, cpu, avg, out, st, &rsd);
1062 } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1063 total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1064
1065 if (total) {
1066 ratio = avg / total;
1067 print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
1068 } else {
1069 print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
1070 }
1071 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1072 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1073
1074 if (total)
1075 print_metric(config, ctxp, NULL,
1076 "%7.2f%%", "transactional cycles",
1077 100.0 * (avg / total));
1078 else
1079 print_metric(config, ctxp, NULL, NULL, "transactional cycles",
1080 0);
1081 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1082 total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1083 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1084
1085 if (total2 < avg)
1086 total2 = avg;
1087 if (total)
1088 print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1089 100.0 * ((total2-avg) / total));
1090 else
1091 print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1092 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1093 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1094
1095 if (avg)
1096 ratio = total / avg;
1097
1098 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, cpu, &rsd) != 0)
1099 print_metric(config, ctxp, NULL, "%8.0f",
1100 "cycles / transaction", ratio);
1101 else
1102 print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1103 0);
1104 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1105 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1106
1107 if (avg)
1108 ratio = total / avg;
1109
1110 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1111 } else if (evsel__is_clock(evsel)) {
1112 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1113 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1114 avg / (ratio * evsel->scale));
1115 else
1116 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1117 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1118 double fe_bound = td_fe_bound(cpu, st, &rsd);
1119
1120 if (fe_bound > 0.2)
1121 color = PERF_COLOR_RED;
1122 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1123 fe_bound * 100.);
1124 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1125 double retiring = td_retiring(cpu, st, &rsd);
1126
1127 if (retiring > 0.7)
1128 color = PERF_COLOR_GREEN;
1129 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1130 retiring * 100.);
1131 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1132 double bad_spec = td_bad_spec(cpu, st, &rsd);
1133
1134 if (bad_spec > 0.1)
1135 color = PERF_COLOR_RED;
1136 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1137 bad_spec * 100.);
1138 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1139 double be_bound = td_be_bound(cpu, st, &rsd);
1140 const char *name = "backend bound";
1141 static int have_recovery_bubbles = -1;
1142
1143 /* In case the CPU does not support topdown-recovery-bubbles */
1144 if (have_recovery_bubbles < 0)
1145 have_recovery_bubbles = pmu_have_event("cpu",
1146 "topdown-recovery-bubbles");
1147 if (!have_recovery_bubbles)
1148 name = "backend bound/bad spec";
1149
1150 if (be_bound > 0.2)
1151 color = PERF_COLOR_RED;
1152 if (td_total_slots(cpu, st, &rsd) > 0)
1153 print_metric(config, ctxp, color, "%8.1f%%", name,
1154 be_bound * 100.);
1155 else
1156 print_metric(config, ctxp, NULL, NULL, name, 0);
1157 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1158 full_td(cpu, st, &rsd)) {
1159 double retiring = td_metric_ratio(cpu,
1160 STAT_TOPDOWN_RETIRING, st,
1161 &rsd);
1162 if (retiring > 0.7)
1163 color = PERF_COLOR_GREEN;
1164 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1165 retiring * 100.);
1166 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1167 full_td(cpu, st, &rsd)) {
1168 double fe_bound = td_metric_ratio(cpu,
1169 STAT_TOPDOWN_FE_BOUND, st,
1170 &rsd);
1171 if (fe_bound > 0.2)
1172 color = PERF_COLOR_RED;
1173 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1174 fe_bound * 100.);
1175 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1176 full_td(cpu, st, &rsd)) {
1177 double be_bound = td_metric_ratio(cpu,
1178 STAT_TOPDOWN_BE_BOUND, st,
1179 &rsd);
1180 if (be_bound > 0.2)
1181 color = PERF_COLOR_RED;
1182 print_metric(config, ctxp, color, "%8.1f%%", "backend bound",
1183 be_bound * 100.);
1184 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1185 full_td(cpu, st, &rsd)) {
1186 double bad_spec = td_metric_ratio(cpu,
1187 STAT_TOPDOWN_BAD_SPEC, st,
1188 &rsd);
1189 if (bad_spec > 0.1)
1190 color = PERF_COLOR_RED;
1191 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1192 bad_spec * 100.);
1193 } else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) &&
1194 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1195 double retiring = td_metric_ratio(cpu,
1196 STAT_TOPDOWN_RETIRING, st,
1197 &rsd);
1198 double heavy_ops = td_metric_ratio(cpu,
1199 STAT_TOPDOWN_HEAVY_OPS, st,
1200 &rsd);
1201 double light_ops = retiring - heavy_ops;
1202
1203 if (retiring > 0.7 && heavy_ops > 0.1)
1204 color = PERF_COLOR_GREEN;
1205 print_metric(config, ctxp, color, "%8.1f%%", "heavy operations",
1206 heavy_ops * 100.);
1207 if (retiring > 0.7 && light_ops > 0.6)
1208 color = PERF_COLOR_GREEN;
1209 else
1210 color = NULL;
1211 print_metric(config, ctxp, color, "%8.1f%%", "light operations",
1212 light_ops * 100.);
1213 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) &&
1214 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1215 double bad_spec = td_metric_ratio(cpu,
1216 STAT_TOPDOWN_BAD_SPEC, st,
1217 &rsd);
1218 double br_mis = td_metric_ratio(cpu,
1219 STAT_TOPDOWN_BR_MISPREDICT, st,
1220 &rsd);
1221 double m_clears = bad_spec - br_mis;
1222
1223 if (bad_spec > 0.1 && br_mis > 0.05)
1224 color = PERF_COLOR_RED;
1225 print_metric(config, ctxp, color, "%8.1f%%", "branch mispredict",
1226 br_mis * 100.);
1227 if (bad_spec > 0.1 && m_clears > 0.05)
1228 color = PERF_COLOR_RED;
1229 else
1230 color = NULL;
1231 print_metric(config, ctxp, color, "%8.1f%%", "machine clears",
1232 m_clears * 100.);
1233 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) &&
1234 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1235 double fe_bound = td_metric_ratio(cpu,
1236 STAT_TOPDOWN_FE_BOUND, st,
1237 &rsd);
1238 double fetch_lat = td_metric_ratio(cpu,
1239 STAT_TOPDOWN_FETCH_LAT, st,
1240 &rsd);
1241 double fetch_bw = fe_bound - fetch_lat;
1242
1243 if (fe_bound > 0.2 && fetch_lat > 0.15)
1244 color = PERF_COLOR_RED;
1245 print_metric(config, ctxp, color, "%8.1f%%", "fetch latency",
1246 fetch_lat * 100.);
1247 if (fe_bound > 0.2 && fetch_bw > 0.1)
1248 color = PERF_COLOR_RED;
1249 else
1250 color = NULL;
1251 print_metric(config, ctxp, color, "%8.1f%%", "fetch bandwidth",
1252 fetch_bw * 100.);
1253 } else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) &&
1254 full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1255 double be_bound = td_metric_ratio(cpu,
1256 STAT_TOPDOWN_BE_BOUND, st,
1257 &rsd);
1258 double mem_bound = td_metric_ratio(cpu,
1259 STAT_TOPDOWN_MEM_BOUND, st,
1260 &rsd);
1261 double core_bound = be_bound - mem_bound;
1262
1263 if (be_bound > 0.2 && mem_bound > 0.2)
1264 color = PERF_COLOR_RED;
1265 print_metric(config, ctxp, color, "%8.1f%%", "memory bound",
1266 mem_bound * 100.);
1267 if (be_bound > 0.2 && core_bound > 0.1)
1268 color = PERF_COLOR_RED;
1269 else
1270 color = NULL;
1271 print_metric(config, ctxp, color, "%8.1f%%", "Core bound",
1272 core_bound * 100.);
1273 } else if (evsel->metric_expr) {
1274 generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1275 evsel->name, evsel->metric_name, NULL, 1, cpu, out, st);
1276 } else if (runtime_stat_n(st, STAT_NSECS, cpu, &rsd) != 0) {
1277 char unit = ' ';
1278 char unit_buf[10] = "/sec";
1279
1280 total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1281 if (total)
1282 ratio = convert_unit_double(1000000000.0 * avg / total, &unit);
1283
1284 if (unit != ' ')
1285 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1286 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1287 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1288 print_smi_cost(config, cpu, out, st, &rsd);
1289 } else {
1290 num = 0;
1291 }
1292
1293 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1294 struct metric_expr *mexp;
1295
1296 list_for_each_entry (mexp, &me->head, nd) {
1297 if (num++ > 0)
1298 out->new_line(config, ctxp);
1299 generic_metric(config, mexp->metric_expr, mexp->metric_events,
1300 mexp->metric_refs, evsel->name, mexp->metric_name,
1301 mexp->metric_unit, mexp->runtime, cpu, out, st);
1302 }
1303 }
1304 if (num == 0)
1305 print_metric(config, ctxp, NULL, NULL, NULL, 0);
1306}