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
18 /*
19 * AGGR_GLOBAL: Use CPU 0
20 * AGGR_SOCKET: Use first CPU of socket
21 * AGGR_DIE: Use first CPU of die
22 * AGGR_CORE: Use first CPU of core
23 * AGGR_NONE: Use matching CPU
24 * AGGR_THREAD: Not supported?
25 */
26
27 struct runtime_stat rt_stat;
28 struct stats walltime_nsecs_stats;
29 struct rusage_stats ru_stats;
30
31 struct saved_value {
32 struct rb_node rb_node;
33 struct evsel *evsel;
34 enum stat_type type;
35 int ctx;
36 int map_idx; /* cpu or thread map index */
37 struct cgroup *cgrp;
38 struct stats stats;
39 u64 metric_total;
40 int metric_other;
41 };
42
saved_value_cmp(struct rb_node * rb_node,const void * entry)43 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
44 {
45 struct saved_value *a = container_of(rb_node,
46 struct saved_value,
47 rb_node);
48 const struct saved_value *b = entry;
49
50 if (a->map_idx != b->map_idx)
51 return a->map_idx - b->map_idx;
52
53 /*
54 * Previously the rbtree was used to link generic metrics.
55 * The keys were evsel/cpu. Now the rbtree is extended to support
56 * per-thread shadow stats. For shadow stats case, the keys
57 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
58 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
59 */
60 if (a->type != b->type)
61 return a->type - b->type;
62
63 if (a->ctx != b->ctx)
64 return a->ctx - b->ctx;
65
66 if (a->cgrp != b->cgrp)
67 return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1;
68
69 if (a->evsel == b->evsel)
70 return 0;
71 if ((char *)a->evsel < (char *)b->evsel)
72 return -1;
73 return +1;
74 }
75
saved_value_new(struct rblist * rblist __maybe_unused,const void * entry)76 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
77 const void *entry)
78 {
79 struct saved_value *nd = malloc(sizeof(struct saved_value));
80
81 if (!nd)
82 return NULL;
83 memcpy(nd, entry, sizeof(struct saved_value));
84 return &nd->rb_node;
85 }
86
saved_value_delete(struct rblist * rblist __maybe_unused,struct rb_node * rb_node)87 static void saved_value_delete(struct rblist *rblist __maybe_unused,
88 struct rb_node *rb_node)
89 {
90 struct saved_value *v;
91
92 BUG_ON(!rb_node);
93 v = container_of(rb_node, struct saved_value, rb_node);
94 free(v);
95 }
96
saved_value_lookup(struct evsel * evsel,int map_idx,bool create,enum stat_type type,int ctx,struct runtime_stat * st,struct cgroup * cgrp)97 static struct saved_value *saved_value_lookup(struct evsel *evsel,
98 int map_idx,
99 bool create,
100 enum stat_type type,
101 int ctx,
102 struct runtime_stat *st,
103 struct cgroup *cgrp)
104 {
105 struct rblist *rblist;
106 struct rb_node *nd;
107 struct saved_value dm = {
108 .map_idx = map_idx,
109 .evsel = evsel,
110 .type = type,
111 .ctx = ctx,
112 .cgrp = cgrp,
113 };
114
115 rblist = &st->value_list;
116
117 /* don't use context info for clock events */
118 if (type == STAT_NSECS)
119 dm.ctx = 0;
120
121 nd = rblist__find(rblist, &dm);
122 if (nd)
123 return container_of(nd, struct saved_value, rb_node);
124 if (create) {
125 rblist__add_node(rblist, &dm);
126 nd = rblist__find(rblist, &dm);
127 if (nd)
128 return container_of(nd, struct saved_value, rb_node);
129 }
130 return NULL;
131 }
132
runtime_stat__init(struct runtime_stat * st)133 void runtime_stat__init(struct runtime_stat *st)
134 {
135 struct rblist *rblist = &st->value_list;
136
137 rblist__init(rblist);
138 rblist->node_cmp = saved_value_cmp;
139 rblist->node_new = saved_value_new;
140 rblist->node_delete = saved_value_delete;
141 }
142
runtime_stat__exit(struct runtime_stat * st)143 void runtime_stat__exit(struct runtime_stat *st)
144 {
145 rblist__exit(&st->value_list);
146 }
147
perf_stat__init_shadow_stats(void)148 void perf_stat__init_shadow_stats(void)
149 {
150 runtime_stat__init(&rt_stat);
151 }
152
evsel_context(struct evsel * evsel)153 static int evsel_context(struct evsel *evsel)
154 {
155 int ctx = 0;
156
157 if (evsel->core.attr.exclude_kernel)
158 ctx |= CTX_BIT_KERNEL;
159 if (evsel->core.attr.exclude_user)
160 ctx |= CTX_BIT_USER;
161 if (evsel->core.attr.exclude_hv)
162 ctx |= CTX_BIT_HV;
163 if (evsel->core.attr.exclude_host)
164 ctx |= CTX_BIT_HOST;
165 if (evsel->core.attr.exclude_idle)
166 ctx |= CTX_BIT_IDLE;
167
168 return ctx;
169 }
170
reset_stat(struct runtime_stat * st)171 static void reset_stat(struct runtime_stat *st)
172 {
173 struct rblist *rblist;
174 struct rb_node *pos, *next;
175
176 rblist = &st->value_list;
177 next = rb_first_cached(&rblist->entries);
178 while (next) {
179 pos = next;
180 next = rb_next(pos);
181 memset(&container_of(pos, struct saved_value, rb_node)->stats,
182 0,
183 sizeof(struct stats));
184 }
185 }
186
perf_stat__reset_shadow_stats(void)187 void perf_stat__reset_shadow_stats(void)
188 {
189 reset_stat(&rt_stat);
190 memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
191 memset(&ru_stats, 0, sizeof(ru_stats));
192 }
193
perf_stat__reset_shadow_per_stat(struct runtime_stat * st)194 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
195 {
196 reset_stat(st);
197 }
198
199 struct runtime_stat_data {
200 int ctx;
201 struct cgroup *cgrp;
202 };
203
update_runtime_stat(struct runtime_stat * st,enum stat_type type,int map_idx,u64 count,struct runtime_stat_data * rsd)204 static void update_runtime_stat(struct runtime_stat *st,
205 enum stat_type type,
206 int map_idx, u64 count,
207 struct runtime_stat_data *rsd)
208 {
209 struct saved_value *v = saved_value_lookup(NULL, map_idx, true, type,
210 rsd->ctx, st, rsd->cgrp);
211
212 if (v)
213 update_stats(&v->stats, count);
214 }
215
216 /*
217 * Update various tracking values we maintain to print
218 * more semantic information such as miss/hit ratios,
219 * instruction rates, etc:
220 */
perf_stat__update_shadow_stats(struct evsel * counter,u64 count,int map_idx,struct runtime_stat * st)221 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
222 int map_idx, struct runtime_stat *st)
223 {
224 u64 count_ns = count;
225 struct saved_value *v;
226 struct runtime_stat_data rsd = {
227 .ctx = evsel_context(counter),
228 .cgrp = counter->cgrp,
229 };
230
231 count *= counter->scale;
232
233 if (evsel__is_clock(counter))
234 update_runtime_stat(st, STAT_NSECS, map_idx, count_ns, &rsd);
235 else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
236 update_runtime_stat(st, STAT_CYCLES, map_idx, count, &rsd);
237 else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
238 update_runtime_stat(st, STAT_CYCLES_IN_TX, map_idx, count, &rsd);
239 else if (perf_stat_evsel__is(counter, TRANSACTION_START))
240 update_runtime_stat(st, STAT_TRANSACTION, map_idx, count, &rsd);
241 else if (perf_stat_evsel__is(counter, ELISION_START))
242 update_runtime_stat(st, STAT_ELISION, map_idx, count, &rsd);
243 else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
244 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
245 map_idx, count, &rsd);
246 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
247 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
248 map_idx, count, &rsd);
249 else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
250 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
251 map_idx, count, &rsd);
252 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
253 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
254 map_idx, count, &rsd);
255 else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
256 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
257 map_idx, count, &rsd);
258 else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
259 update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
260 map_idx, count, &rsd);
261 else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
262 update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
263 map_idx, count, &rsd);
264 else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
265 update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
266 map_idx, count, &rsd);
267 else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
268 update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
269 map_idx, count, &rsd);
270 else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS))
271 update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS,
272 map_idx, count, &rsd);
273 else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT))
274 update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT,
275 map_idx, count, &rsd);
276 else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT))
277 update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT,
278 map_idx, count, &rsd);
279 else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND))
280 update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND,
281 map_idx, count, &rsd);
282 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
283 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
284 map_idx, count, &rsd);
285 else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
286 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
287 map_idx, count, &rsd);
288 else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
289 update_runtime_stat(st, STAT_BRANCHES, map_idx, count, &rsd);
290 else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
291 update_runtime_stat(st, STAT_CACHEREFS, map_idx, count, &rsd);
292 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
293 update_runtime_stat(st, STAT_L1_DCACHE, map_idx, count, &rsd);
294 else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
295 update_runtime_stat(st, STAT_L1_ICACHE, map_idx, count, &rsd);
296 else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
297 update_runtime_stat(st, STAT_LL_CACHE, map_idx, count, &rsd);
298 else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
299 update_runtime_stat(st, STAT_DTLB_CACHE, map_idx, count, &rsd);
300 else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
301 update_runtime_stat(st, STAT_ITLB_CACHE, map_idx, count, &rsd);
302 else if (perf_stat_evsel__is(counter, SMI_NUM))
303 update_runtime_stat(st, STAT_SMI_NUM, map_idx, count, &rsd);
304 else if (perf_stat_evsel__is(counter, APERF))
305 update_runtime_stat(st, STAT_APERF, map_idx, count, &rsd);
306
307 if (counter->collect_stat) {
308 v = saved_value_lookup(counter, map_idx, true, STAT_NONE, 0, st,
309 rsd.cgrp);
310 update_stats(&v->stats, count);
311 if (counter->metric_leader)
312 v->metric_total += count;
313 } else if (counter->metric_leader) {
314 v = saved_value_lookup(counter->metric_leader,
315 map_idx, true, STAT_NONE, 0, st, rsd.cgrp);
316 v->metric_total += count;
317 v->metric_other++;
318 }
319 }
320
321 /* used for get_ratio_color() */
322 enum grc_type {
323 GRC_STALLED_CYCLES_FE,
324 GRC_STALLED_CYCLES_BE,
325 GRC_CACHE_MISSES,
326 GRC_MAX_NR
327 };
328
get_ratio_color(enum grc_type type,double ratio)329 static const char *get_ratio_color(enum grc_type type, double ratio)
330 {
331 static const double grc_table[GRC_MAX_NR][3] = {
332 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
333 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
334 [GRC_CACHE_MISSES] = { 20.0, 10.0, 5.0 },
335 };
336 const char *color = PERF_COLOR_NORMAL;
337
338 if (ratio > grc_table[type][0])
339 color = PERF_COLOR_RED;
340 else if (ratio > grc_table[type][1])
341 color = PERF_COLOR_MAGENTA;
342 else if (ratio > grc_table[type][2])
343 color = PERF_COLOR_YELLOW;
344
345 return color;
346 }
347
perf_stat__find_event(struct evlist * evsel_list,const char * name)348 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
349 const char *name)
350 {
351 struct evsel *c2;
352
353 evlist__for_each_entry (evsel_list, c2) {
354 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
355 return c2;
356 }
357 return NULL;
358 }
359
360 /* Mark MetricExpr target events and link events using them to them. */
perf_stat__collect_metric_expr(struct evlist * evsel_list)361 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
362 {
363 struct evsel *counter, *leader, **metric_events, *oc;
364 bool found;
365 struct expr_parse_ctx *ctx;
366 struct hashmap_entry *cur;
367 size_t bkt;
368 int i;
369
370 ctx = expr__ctx_new();
371 if (!ctx) {
372 pr_debug("expr__ctx_new failed");
373 return;
374 }
375 evlist__for_each_entry(evsel_list, counter) {
376 bool invalid = false;
377
378 leader = evsel__leader(counter);
379 if (!counter->metric_expr)
380 continue;
381
382 expr__ctx_clear(ctx);
383 metric_events = counter->metric_events;
384 if (!metric_events) {
385 if (expr__find_ids(counter->metric_expr,
386 counter->name,
387 ctx) < 0)
388 continue;
389
390 metric_events = calloc(sizeof(struct evsel *),
391 hashmap__size(ctx->ids) + 1);
392 if (!metric_events) {
393 expr__ctx_free(ctx);
394 return;
395 }
396 counter->metric_events = metric_events;
397 }
398
399 i = 0;
400 hashmap__for_each_entry(ctx->ids, cur, bkt) {
401 const char *metric_name = (const char *)cur->key;
402
403 found = false;
404 if (leader) {
405 /* Search in group */
406 for_each_group_member (oc, leader) {
407 if (!strcasecmp(oc->name,
408 metric_name) &&
409 !oc->collect_stat) {
410 found = true;
411 break;
412 }
413 }
414 }
415 if (!found) {
416 /* Search ignoring groups */
417 oc = perf_stat__find_event(evsel_list,
418 metric_name);
419 }
420 if (!oc) {
421 /* Deduping one is good enough to handle duplicated PMUs. */
422 static char *printed;
423
424 /*
425 * Adding events automatically would be difficult, because
426 * it would risk creating groups that are not schedulable.
427 * perf stat doesn't understand all the scheduling constraints
428 * of events. So we ask the user instead to add the missing
429 * events.
430 */
431 if (!printed ||
432 strcasecmp(printed, metric_name)) {
433 fprintf(stderr,
434 "Add %s event to groups to get metric expression for %s\n",
435 metric_name,
436 counter->name);
437 free(printed);
438 printed = strdup(metric_name);
439 }
440 invalid = true;
441 continue;
442 }
443 metric_events[i++] = oc;
444 oc->collect_stat = true;
445 }
446 metric_events[i] = NULL;
447 if (invalid) {
448 free(metric_events);
449 counter->metric_events = NULL;
450 counter->metric_expr = NULL;
451 }
452 }
453 expr__ctx_free(ctx);
454 }
455
runtime_stat_avg(struct runtime_stat * st,enum stat_type type,int map_idx,struct runtime_stat_data * rsd)456 static double runtime_stat_avg(struct runtime_stat *st,
457 enum stat_type type, int map_idx,
458 struct runtime_stat_data *rsd)
459 {
460 struct saved_value *v;
461
462 v = saved_value_lookup(NULL, map_idx, false, type, rsd->ctx, st, rsd->cgrp);
463 if (!v)
464 return 0.0;
465
466 return avg_stats(&v->stats);
467 }
468
runtime_stat_n(struct runtime_stat * st,enum stat_type type,int map_idx,struct runtime_stat_data * rsd)469 static double runtime_stat_n(struct runtime_stat *st,
470 enum stat_type type, int map_idx,
471 struct runtime_stat_data *rsd)
472 {
473 struct saved_value *v;
474
475 v = saved_value_lookup(NULL, map_idx, false, type, rsd->ctx, st, rsd->cgrp);
476 if (!v)
477 return 0.0;
478
479 return v->stats.n;
480 }
481
print_stalled_cycles_frontend(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)482 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
483 int map_idx, double avg,
484 struct perf_stat_output_ctx *out,
485 struct runtime_stat *st,
486 struct runtime_stat_data *rsd)
487 {
488 double total, ratio = 0.0;
489 const char *color;
490
491 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
492
493 if (total)
494 ratio = avg / total * 100.0;
495
496 color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
497
498 if (ratio)
499 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
500 ratio);
501 else
502 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
503 }
504
print_stalled_cycles_backend(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)505 static void print_stalled_cycles_backend(struct perf_stat_config *config,
506 int map_idx, double avg,
507 struct perf_stat_output_ctx *out,
508 struct runtime_stat *st,
509 struct runtime_stat_data *rsd)
510 {
511 double total, ratio = 0.0;
512 const char *color;
513
514 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
515
516 if (total)
517 ratio = avg / total * 100.0;
518
519 color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
520
521 out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
522 }
523
print_branch_misses(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)524 static void print_branch_misses(struct perf_stat_config *config,
525 int map_idx, double avg,
526 struct perf_stat_output_ctx *out,
527 struct runtime_stat *st,
528 struct runtime_stat_data *rsd)
529 {
530 double total, ratio = 0.0;
531 const char *color;
532
533 total = runtime_stat_avg(st, STAT_BRANCHES, map_idx, rsd);
534
535 if (total)
536 ratio = avg / total * 100.0;
537
538 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
539
540 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
541 }
542
print_l1_dcache_misses(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)543 static void print_l1_dcache_misses(struct perf_stat_config *config,
544 int map_idx, double avg,
545 struct perf_stat_output_ctx *out,
546 struct runtime_stat *st,
547 struct runtime_stat_data *rsd)
548 {
549 double total, ratio = 0.0;
550 const char *color;
551
552 total = runtime_stat_avg(st, STAT_L1_DCACHE, map_idx, rsd);
553
554 if (total)
555 ratio = avg / total * 100.0;
556
557 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
558
559 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
560 }
561
print_l1_icache_misses(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)562 static void print_l1_icache_misses(struct perf_stat_config *config,
563 int map_idx, double avg,
564 struct perf_stat_output_ctx *out,
565 struct runtime_stat *st,
566 struct runtime_stat_data *rsd)
567 {
568 double total, ratio = 0.0;
569 const char *color;
570
571 total = runtime_stat_avg(st, STAT_L1_ICACHE, map_idx, rsd);
572
573 if (total)
574 ratio = avg / total * 100.0;
575
576 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
577 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
578 }
579
print_dtlb_cache_misses(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)580 static void print_dtlb_cache_misses(struct perf_stat_config *config,
581 int map_idx, double avg,
582 struct perf_stat_output_ctx *out,
583 struct runtime_stat *st,
584 struct runtime_stat_data *rsd)
585 {
586 double total, ratio = 0.0;
587 const char *color;
588
589 total = runtime_stat_avg(st, STAT_DTLB_CACHE, map_idx, rsd);
590
591 if (total)
592 ratio = avg / total * 100.0;
593
594 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
595 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
596 }
597
print_itlb_cache_misses(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)598 static void print_itlb_cache_misses(struct perf_stat_config *config,
599 int map_idx, double avg,
600 struct perf_stat_output_ctx *out,
601 struct runtime_stat *st,
602 struct runtime_stat_data *rsd)
603 {
604 double total, ratio = 0.0;
605 const char *color;
606
607 total = runtime_stat_avg(st, STAT_ITLB_CACHE, map_idx, rsd);
608
609 if (total)
610 ratio = avg / total * 100.0;
611
612 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
613 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
614 }
615
print_ll_cache_misses(struct perf_stat_config * config,int map_idx,double avg,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)616 static void print_ll_cache_misses(struct perf_stat_config *config,
617 int map_idx, double avg,
618 struct perf_stat_output_ctx *out,
619 struct runtime_stat *st,
620 struct runtime_stat_data *rsd)
621 {
622 double total, ratio = 0.0;
623 const char *color;
624
625 total = runtime_stat_avg(st, STAT_LL_CACHE, map_idx, rsd);
626
627 if (total)
628 ratio = avg / total * 100.0;
629
630 color = get_ratio_color(GRC_CACHE_MISSES, ratio);
631 out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
632 }
633
634 /*
635 * High level "TopDown" CPU core pipe line bottleneck break down.
636 *
637 * Basic concept following
638 * Yasin, A Top Down Method for Performance analysis and Counter architecture
639 * ISPASS14
640 *
641 * The CPU pipeline is divided into 4 areas that can be bottlenecks:
642 *
643 * Frontend -> Backend -> Retiring
644 * BadSpeculation in addition means out of order execution that is thrown away
645 * (for example branch mispredictions)
646 * Frontend is instruction decoding.
647 * Backend is execution, like computation and accessing data in memory
648 * Retiring is good execution that is not directly bottlenecked
649 *
650 * The formulas are computed in slots.
651 * A slot is an entry in the pipeline each for the pipeline width
652 * (for example a 4-wide pipeline has 4 slots for each cycle)
653 *
654 * Formulas:
655 * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
656 * TotalSlots
657 * Retiring = SlotsRetired / TotalSlots
658 * FrontendBound = FetchBubbles / TotalSlots
659 * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
660 *
661 * The kernel provides the mapping to the low level CPU events and any scaling
662 * needed for the CPU pipeline width, for example:
663 *
664 * TotalSlots = Cycles * 4
665 *
666 * The scaling factor is communicated in the sysfs unit.
667 *
668 * In some cases the CPU may not be able to measure all the formulas due to
669 * missing events. In this case multiple formulas are combined, as possible.
670 *
671 * Full TopDown supports more levels to sub-divide each area: for example
672 * BackendBound into computing bound and memory bound. For now we only
673 * support Level 1 TopDown.
674 */
675
sanitize_val(double x)676 static double sanitize_val(double x)
677 {
678 if (x < 0 && x >= -0.02)
679 return 0.0;
680 return x;
681 }
682
td_total_slots(int map_idx,struct runtime_stat * st,struct runtime_stat_data * rsd)683 static double td_total_slots(int map_idx, struct runtime_stat *st,
684 struct runtime_stat_data *rsd)
685 {
686 return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, map_idx, rsd);
687 }
688
td_bad_spec(int map_idx,struct runtime_stat * st,struct runtime_stat_data * rsd)689 static double td_bad_spec(int map_idx, struct runtime_stat *st,
690 struct runtime_stat_data *rsd)
691 {
692 double bad_spec = 0;
693 double total_slots;
694 double total;
695
696 total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, map_idx, rsd) -
697 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, map_idx, rsd) +
698 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, map_idx, rsd);
699
700 total_slots = td_total_slots(map_idx, st, rsd);
701 if (total_slots)
702 bad_spec = total / total_slots;
703 return sanitize_val(bad_spec);
704 }
705
td_retiring(int map_idx,struct runtime_stat * st,struct runtime_stat_data * rsd)706 static double td_retiring(int map_idx, struct runtime_stat *st,
707 struct runtime_stat_data *rsd)
708 {
709 double retiring = 0;
710 double total_slots = td_total_slots(map_idx, st, rsd);
711 double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
712 map_idx, rsd);
713
714 if (total_slots)
715 retiring = ret_slots / total_slots;
716 return retiring;
717 }
718
td_fe_bound(int map_idx,struct runtime_stat * st,struct runtime_stat_data * rsd)719 static double td_fe_bound(int map_idx, struct runtime_stat *st,
720 struct runtime_stat_data *rsd)
721 {
722 double fe_bound = 0;
723 double total_slots = td_total_slots(map_idx, st, rsd);
724 double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
725 map_idx, rsd);
726
727 if (total_slots)
728 fe_bound = fetch_bub / total_slots;
729 return fe_bound;
730 }
731
td_be_bound(int map_idx,struct runtime_stat * st,struct runtime_stat_data * rsd)732 static double td_be_bound(int map_idx, struct runtime_stat *st,
733 struct runtime_stat_data *rsd)
734 {
735 double sum = (td_fe_bound(map_idx, st, rsd) +
736 td_bad_spec(map_idx, st, rsd) +
737 td_retiring(map_idx, st, rsd));
738 if (sum == 0)
739 return 0;
740 return sanitize_val(1.0 - sum);
741 }
742
743 /*
744 * Kernel reports metrics multiplied with slots. To get back
745 * the ratios we need to recreate the sum.
746 */
747
td_metric_ratio(int map_idx,enum stat_type type,struct runtime_stat * stat,struct runtime_stat_data * rsd)748 static double td_metric_ratio(int map_idx, enum stat_type type,
749 struct runtime_stat *stat,
750 struct runtime_stat_data *rsd)
751 {
752 double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, map_idx, rsd) +
753 runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, map_idx, rsd) +
754 runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, map_idx, rsd) +
755 runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, map_idx, rsd);
756 double d = runtime_stat_avg(stat, type, map_idx, rsd);
757
758 if (sum)
759 return d / sum;
760 return 0;
761 }
762
763 /*
764 * ... but only if most of the values are actually available.
765 * We allow two missing.
766 */
767
full_td(int map_idx,struct runtime_stat * stat,struct runtime_stat_data * rsd)768 static bool full_td(int map_idx, struct runtime_stat *stat,
769 struct runtime_stat_data *rsd)
770 {
771 int c = 0;
772
773 if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, map_idx, rsd) > 0)
774 c++;
775 if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, map_idx, rsd) > 0)
776 c++;
777 if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, map_idx, rsd) > 0)
778 c++;
779 if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, map_idx, rsd) > 0)
780 c++;
781 return c >= 2;
782 }
783
print_smi_cost(struct perf_stat_config * config,int map_idx,struct perf_stat_output_ctx * out,struct runtime_stat * st,struct runtime_stat_data * rsd)784 static void print_smi_cost(struct perf_stat_config *config, int map_idx,
785 struct perf_stat_output_ctx *out,
786 struct runtime_stat *st,
787 struct runtime_stat_data *rsd)
788 {
789 double smi_num, aperf, cycles, cost = 0.0;
790 const char *color = NULL;
791
792 smi_num = runtime_stat_avg(st, STAT_SMI_NUM, map_idx, rsd);
793 aperf = runtime_stat_avg(st, STAT_APERF, map_idx, rsd);
794 cycles = runtime_stat_avg(st, STAT_CYCLES, map_idx, rsd);
795
796 if ((cycles == 0) || (aperf == 0))
797 return;
798
799 if (smi_num)
800 cost = (aperf - cycles) / aperf * 100.00;
801
802 if (cost > 10)
803 color = PERF_COLOR_RED;
804 out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
805 out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
806 }
807
prepare_metric(struct evsel ** metric_events,struct metric_ref * metric_refs,struct expr_parse_ctx * pctx,int map_idx,struct runtime_stat * st)808 static int prepare_metric(struct evsel **metric_events,
809 struct metric_ref *metric_refs,
810 struct expr_parse_ctx *pctx,
811 int map_idx,
812 struct runtime_stat *st)
813 {
814 double scale;
815 char *n;
816 int i, j, ret;
817
818 for (i = 0; metric_events[i]; i++) {
819 struct saved_value *v;
820 struct stats *stats;
821 u64 metric_total = 0;
822 int source_count;
823
824 if (evsel__is_tool(metric_events[i])) {
825 source_count = 1;
826 switch (metric_events[i]->tool_event) {
827 case PERF_TOOL_DURATION_TIME:
828 stats = &walltime_nsecs_stats;
829 scale = 1e-9;
830 break;
831 case PERF_TOOL_USER_TIME:
832 stats = &ru_stats.ru_utime_usec_stat;
833 scale = 1e-6;
834 break;
835 case PERF_TOOL_SYSTEM_TIME:
836 stats = &ru_stats.ru_stime_usec_stat;
837 scale = 1e-6;
838 break;
839 case PERF_TOOL_NONE:
840 pr_err("Invalid tool event 'none'");
841 abort();
842 case PERF_TOOL_MAX:
843 pr_err("Invalid tool event 'max'");
844 abort();
845 default:
846 pr_err("Unknown tool event '%s'", evsel__name(metric_events[i]));
847 abort();
848 }
849 } else {
850 v = saved_value_lookup(metric_events[i], map_idx, false,
851 STAT_NONE, 0, st,
852 metric_events[i]->cgrp);
853 if (!v)
854 break;
855 stats = &v->stats;
856 /*
857 * If an event was scaled during stat gathering, reverse
858 * the scale before computing the metric.
859 */
860 scale = 1.0 / metric_events[i]->scale;
861
862 source_count = evsel__source_count(metric_events[i]);
863
864 if (v->metric_other)
865 metric_total = v->metric_total * scale;
866 }
867 n = strdup(evsel__metric_id(metric_events[i]));
868 if (!n)
869 return -ENOMEM;
870
871 expr__add_id_val_source_count(pctx, n,
872 metric_total ? : avg_stats(stats) * scale,
873 source_count);
874 }
875
876 for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
877 ret = expr__add_ref(pctx, &metric_refs[j]);
878 if (ret)
879 return ret;
880 }
881
882 return i;
883 }
884
generic_metric(struct perf_stat_config * config,const char * metric_expr,struct evsel ** metric_events,struct metric_ref * metric_refs,char * name,const char * metric_name,const char * metric_unit,int runtime,int map_idx,struct perf_stat_output_ctx * out,struct runtime_stat * st)885 static void generic_metric(struct perf_stat_config *config,
886 const char *metric_expr,
887 struct evsel **metric_events,
888 struct metric_ref *metric_refs,
889 char *name,
890 const char *metric_name,
891 const char *metric_unit,
892 int runtime,
893 int map_idx,
894 struct perf_stat_output_ctx *out,
895 struct runtime_stat *st)
896 {
897 print_metric_t print_metric = out->print_metric;
898 struct expr_parse_ctx *pctx;
899 double ratio, scale;
900 int i;
901 void *ctxp = out->ctx;
902
903 pctx = expr__ctx_new();
904 if (!pctx)
905 return;
906
907 if (config->user_requested_cpu_list)
908 pctx->sctx.user_requested_cpu_list = strdup(config->user_requested_cpu_list);
909 pctx->sctx.runtime = runtime;
910 pctx->sctx.system_wide = config->system_wide;
911 i = prepare_metric(metric_events, metric_refs, pctx, map_idx, st);
912 if (i < 0) {
913 expr__ctx_free(pctx);
914 return;
915 }
916 if (!metric_events[i]) {
917 if (expr__parse(&ratio, pctx, metric_expr) == 0) {
918 char *unit;
919 char metric_bf[64];
920
921 if (metric_unit && metric_name) {
922 if (perf_pmu__convert_scale(metric_unit,
923 &unit, &scale) >= 0) {
924 ratio *= scale;
925 }
926 if (strstr(metric_expr, "?"))
927 scnprintf(metric_bf, sizeof(metric_bf),
928 "%s %s_%d", unit, metric_name, runtime);
929 else
930 scnprintf(metric_bf, sizeof(metric_bf),
931 "%s %s", unit, metric_name);
932
933 print_metric(config, ctxp, NULL, "%8.1f",
934 metric_bf, ratio);
935 } else {
936 print_metric(config, ctxp, NULL, "%8.2f",
937 metric_name ?
938 metric_name :
939 out->force_header ? name : "",
940 ratio);
941 }
942 } else {
943 print_metric(config, ctxp, NULL, NULL,
944 out->force_header ?
945 (metric_name ? metric_name : name) : "", 0);
946 }
947 } else {
948 print_metric(config, ctxp, NULL, NULL,
949 out->force_header ?
950 (metric_name ? metric_name : name) : "", 0);
951 }
952
953 expr__ctx_free(pctx);
954 }
955
test_generic_metric(struct metric_expr * mexp,int map_idx,struct runtime_stat * st)956 double test_generic_metric(struct metric_expr *mexp, int map_idx, struct runtime_stat *st)
957 {
958 struct expr_parse_ctx *pctx;
959 double ratio = 0.0;
960
961 pctx = expr__ctx_new();
962 if (!pctx)
963 return NAN;
964
965 if (prepare_metric(mexp->metric_events, mexp->metric_refs, pctx, map_idx, st) < 0)
966 goto out;
967
968 if (expr__parse(&ratio, pctx, mexp->metric_expr))
969 ratio = 0.0;
970
971 out:
972 expr__ctx_free(pctx);
973 return ratio;
974 }
975
perf_stat__print_shadow_stats(struct perf_stat_config * config,struct evsel * evsel,double avg,int map_idx,struct perf_stat_output_ctx * out,struct rblist * metric_events,struct runtime_stat * st)976 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
977 struct evsel *evsel,
978 double avg, int map_idx,
979 struct perf_stat_output_ctx *out,
980 struct rblist *metric_events,
981 struct runtime_stat *st)
982 {
983 void *ctxp = out->ctx;
984 print_metric_t print_metric = out->print_metric;
985 double total, ratio = 0.0, total2;
986 const char *color = NULL;
987 struct runtime_stat_data rsd = {
988 .ctx = evsel_context(evsel),
989 .cgrp = evsel->cgrp,
990 };
991 struct metric_event *me;
992 int num = 1;
993
994 if (config->iostat_run) {
995 iostat_print_metric(config, evsel, out);
996 } else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
997 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
998
999 if (total) {
1000 ratio = avg / total;
1001 print_metric(config, ctxp, NULL, "%7.2f ",
1002 "insn per cycle", ratio);
1003 } else {
1004 print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
1005 }
1006
1007 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, map_idx, &rsd);
1008
1009 total = max(total, runtime_stat_avg(st,
1010 STAT_STALLED_CYCLES_BACK,
1011 map_idx, &rsd));
1012
1013 if (total && avg) {
1014 out->new_line(config, ctxp);
1015 ratio = total / avg;
1016 print_metric(config, ctxp, NULL, "%7.2f ",
1017 "stalled cycles per insn",
1018 ratio);
1019 }
1020 } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
1021 if (runtime_stat_n(st, STAT_BRANCHES, map_idx, &rsd) != 0)
1022 print_branch_misses(config, map_idx, avg, out, st, &rsd);
1023 else
1024 print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
1025 } else if (
1026 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1027 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1D |
1028 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1029 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1030
1031 if (runtime_stat_n(st, STAT_L1_DCACHE, map_idx, &rsd) != 0)
1032 print_l1_dcache_misses(config, map_idx, avg, out, st, &rsd);
1033 else
1034 print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
1035 } else if (
1036 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1037 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_L1I |
1038 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1039 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1040
1041 if (runtime_stat_n(st, STAT_L1_ICACHE, map_idx, &rsd) != 0)
1042 print_l1_icache_misses(config, map_idx, avg, out, st, &rsd);
1043 else
1044 print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
1045 } else if (
1046 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1047 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
1048 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1049 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1050
1051 if (runtime_stat_n(st, STAT_DTLB_CACHE, map_idx, &rsd) != 0)
1052 print_dtlb_cache_misses(config, map_idx, avg, out, st, &rsd);
1053 else
1054 print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
1055 } else if (
1056 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1057 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
1058 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1059 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1060
1061 if (runtime_stat_n(st, STAT_ITLB_CACHE, map_idx, &rsd) != 0)
1062 print_itlb_cache_misses(config, map_idx, avg, out, st, &rsd);
1063 else
1064 print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
1065 } else if (
1066 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1067 evsel->core.attr.config == ( PERF_COUNT_HW_CACHE_LL |
1068 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1069 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1070
1071 if (runtime_stat_n(st, STAT_LL_CACHE, map_idx, &rsd) != 0)
1072 print_ll_cache_misses(config, map_idx, avg, out, st, &rsd);
1073 else
1074 print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
1075 } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
1076 total = runtime_stat_avg(st, STAT_CACHEREFS, map_idx, &rsd);
1077
1078 if (total)
1079 ratio = avg * 100 / total;
1080
1081 if (runtime_stat_n(st, STAT_CACHEREFS, map_idx, &rsd) != 0)
1082 print_metric(config, ctxp, NULL, "%8.3f %%",
1083 "of all cache refs", ratio);
1084 else
1085 print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
1086 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1087 print_stalled_cycles_frontend(config, map_idx, avg, out, st, &rsd);
1088 } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1089 print_stalled_cycles_backend(config, map_idx, avg, out, st, &rsd);
1090 } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1091 total = runtime_stat_avg(st, STAT_NSECS, map_idx, &rsd);
1092
1093 if (total) {
1094 ratio = avg / total;
1095 print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
1096 } else {
1097 print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
1098 }
1099 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1100 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
1101
1102 if (total)
1103 print_metric(config, ctxp, NULL,
1104 "%7.2f%%", "transactional cycles",
1105 100.0 * (avg / total));
1106 else
1107 print_metric(config, ctxp, NULL, NULL, "transactional cycles",
1108 0);
1109 } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1110 total = runtime_stat_avg(st, STAT_CYCLES, map_idx, &rsd);
1111 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1112
1113 if (total2 < avg)
1114 total2 = avg;
1115 if (total)
1116 print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1117 100.0 * ((total2-avg) / total));
1118 else
1119 print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1120 } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1121 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1122
1123 if (avg)
1124 ratio = total / avg;
1125
1126 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, map_idx, &rsd) != 0)
1127 print_metric(config, ctxp, NULL, "%8.0f",
1128 "cycles / transaction", ratio);
1129 else
1130 print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1131 0);
1132 } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1133 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, map_idx, &rsd);
1134
1135 if (avg)
1136 ratio = total / avg;
1137
1138 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1139 } else if (evsel__is_clock(evsel)) {
1140 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1141 print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1142 avg / (ratio * evsel->scale));
1143 else
1144 print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1145 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1146 double fe_bound = td_fe_bound(map_idx, st, &rsd);
1147
1148 if (fe_bound > 0.2)
1149 color = PERF_COLOR_RED;
1150 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1151 fe_bound * 100.);
1152 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1153 double retiring = td_retiring(map_idx, st, &rsd);
1154
1155 if (retiring > 0.7)
1156 color = PERF_COLOR_GREEN;
1157 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1158 retiring * 100.);
1159 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1160 double bad_spec = td_bad_spec(map_idx, st, &rsd);
1161
1162 if (bad_spec > 0.1)
1163 color = PERF_COLOR_RED;
1164 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1165 bad_spec * 100.);
1166 } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1167 double be_bound = td_be_bound(map_idx, st, &rsd);
1168 const char *name = "backend bound";
1169 static int have_recovery_bubbles = -1;
1170
1171 /* In case the CPU does not support topdown-recovery-bubbles */
1172 if (have_recovery_bubbles < 0)
1173 have_recovery_bubbles = pmu_have_event("cpu",
1174 "topdown-recovery-bubbles");
1175 if (!have_recovery_bubbles)
1176 name = "backend bound/bad spec";
1177
1178 if (be_bound > 0.2)
1179 color = PERF_COLOR_RED;
1180 if (td_total_slots(map_idx, st, &rsd) > 0)
1181 print_metric(config, ctxp, color, "%8.1f%%", name,
1182 be_bound * 100.);
1183 else
1184 print_metric(config, ctxp, NULL, NULL, name, 0);
1185 } else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1186 full_td(map_idx, st, &rsd)) {
1187 double retiring = td_metric_ratio(map_idx,
1188 STAT_TOPDOWN_RETIRING, st,
1189 &rsd);
1190 if (retiring > 0.7)
1191 color = PERF_COLOR_GREEN;
1192 print_metric(config, ctxp, color, "%8.1f%%", "Retiring",
1193 retiring * 100.);
1194 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1195 full_td(map_idx, st, &rsd)) {
1196 double fe_bound = td_metric_ratio(map_idx,
1197 STAT_TOPDOWN_FE_BOUND, st,
1198 &rsd);
1199 if (fe_bound > 0.2)
1200 color = PERF_COLOR_RED;
1201 print_metric(config, ctxp, color, "%8.1f%%", "Frontend Bound",
1202 fe_bound * 100.);
1203 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1204 full_td(map_idx, st, &rsd)) {
1205 double be_bound = td_metric_ratio(map_idx,
1206 STAT_TOPDOWN_BE_BOUND, st,
1207 &rsd);
1208 if (be_bound > 0.2)
1209 color = PERF_COLOR_RED;
1210 print_metric(config, ctxp, color, "%8.1f%%", "Backend Bound",
1211 be_bound * 100.);
1212 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1213 full_td(map_idx, st, &rsd)) {
1214 double bad_spec = td_metric_ratio(map_idx,
1215 STAT_TOPDOWN_BAD_SPEC, st,
1216 &rsd);
1217 if (bad_spec > 0.1)
1218 color = PERF_COLOR_RED;
1219 print_metric(config, ctxp, color, "%8.1f%%", "Bad Speculation",
1220 bad_spec * 100.);
1221 } else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) &&
1222 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1223 double retiring = td_metric_ratio(map_idx,
1224 STAT_TOPDOWN_RETIRING, st,
1225 &rsd);
1226 double heavy_ops = td_metric_ratio(map_idx,
1227 STAT_TOPDOWN_HEAVY_OPS, st,
1228 &rsd);
1229 double light_ops = retiring - heavy_ops;
1230
1231 if (retiring > 0.7 && heavy_ops > 0.1)
1232 color = PERF_COLOR_GREEN;
1233 print_metric(config, ctxp, color, "%8.1f%%", "Heavy Operations",
1234 heavy_ops * 100.);
1235 if (retiring > 0.7 && light_ops > 0.6)
1236 color = PERF_COLOR_GREEN;
1237 else
1238 color = NULL;
1239 print_metric(config, ctxp, color, "%8.1f%%", "Light Operations",
1240 light_ops * 100.);
1241 } else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) &&
1242 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1243 double bad_spec = td_metric_ratio(map_idx,
1244 STAT_TOPDOWN_BAD_SPEC, st,
1245 &rsd);
1246 double br_mis = td_metric_ratio(map_idx,
1247 STAT_TOPDOWN_BR_MISPREDICT, st,
1248 &rsd);
1249 double m_clears = bad_spec - br_mis;
1250
1251 if (bad_spec > 0.1 && br_mis > 0.05)
1252 color = PERF_COLOR_RED;
1253 print_metric(config, ctxp, color, "%8.1f%%", "Branch Mispredict",
1254 br_mis * 100.);
1255 if (bad_spec > 0.1 && m_clears > 0.05)
1256 color = PERF_COLOR_RED;
1257 else
1258 color = NULL;
1259 print_metric(config, ctxp, color, "%8.1f%%", "Machine Clears",
1260 m_clears * 100.);
1261 } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) &&
1262 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1263 double fe_bound = td_metric_ratio(map_idx,
1264 STAT_TOPDOWN_FE_BOUND, st,
1265 &rsd);
1266 double fetch_lat = td_metric_ratio(map_idx,
1267 STAT_TOPDOWN_FETCH_LAT, st,
1268 &rsd);
1269 double fetch_bw = fe_bound - fetch_lat;
1270
1271 if (fe_bound > 0.2 && fetch_lat > 0.15)
1272 color = PERF_COLOR_RED;
1273 print_metric(config, ctxp, color, "%8.1f%%", "Fetch Latency",
1274 fetch_lat * 100.);
1275 if (fe_bound > 0.2 && fetch_bw > 0.1)
1276 color = PERF_COLOR_RED;
1277 else
1278 color = NULL;
1279 print_metric(config, ctxp, color, "%8.1f%%", "Fetch Bandwidth",
1280 fetch_bw * 100.);
1281 } else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) &&
1282 full_td(map_idx, st, &rsd) && (config->topdown_level > 1)) {
1283 double be_bound = td_metric_ratio(map_idx,
1284 STAT_TOPDOWN_BE_BOUND, st,
1285 &rsd);
1286 double mem_bound = td_metric_ratio(map_idx,
1287 STAT_TOPDOWN_MEM_BOUND, st,
1288 &rsd);
1289 double core_bound = be_bound - mem_bound;
1290
1291 if (be_bound > 0.2 && mem_bound > 0.2)
1292 color = PERF_COLOR_RED;
1293 print_metric(config, ctxp, color, "%8.1f%%", "Memory Bound",
1294 mem_bound * 100.);
1295 if (be_bound > 0.2 && core_bound > 0.1)
1296 color = PERF_COLOR_RED;
1297 else
1298 color = NULL;
1299 print_metric(config, ctxp, color, "%8.1f%%", "Core Bound",
1300 core_bound * 100.);
1301 } else if (evsel->metric_expr) {
1302 generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1303 evsel->name, evsel->metric_name, NULL, 1,
1304 map_idx, out, st);
1305 } else if (runtime_stat_n(st, STAT_NSECS, map_idx, &rsd) != 0) {
1306 char unit = ' ';
1307 char unit_buf[10] = "/sec";
1308
1309 total = runtime_stat_avg(st, STAT_NSECS, map_idx, &rsd);
1310 if (total)
1311 ratio = convert_unit_double(1000000000.0 * avg / total, &unit);
1312
1313 if (unit != ' ')
1314 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1315 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1316 } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1317 print_smi_cost(config, map_idx, out, st, &rsd);
1318 } else {
1319 num = 0;
1320 }
1321
1322 if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1323 struct metric_expr *mexp;
1324
1325 list_for_each_entry (mexp, &me->head, nd) {
1326 if (num++ > 0)
1327 out->new_line(config, ctxp);
1328 generic_metric(config, mexp->metric_expr, mexp->metric_events,
1329 mexp->metric_refs, evsel->name, mexp->metric_name,
1330 mexp->metric_unit, mexp->runtime,
1331 map_idx, out, st);
1332 }
1333 }
1334 if (num == 0)
1335 print_metric(config, ctxp, NULL, NULL, NULL, 0);
1336 }
1337