xref: /linux/tools/perf/util/stat-shadow.c (revision 8a79db5e83a5d52c74e6f3c40d6f312cf899213e)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <stdio.h>
3 #include "evsel.h"
4 #include "stat.h"
5 #include "color.h"
6 #include "pmu.h"
7 #include "rblist.h"
8 #include "evlist.h"
9 #include "expr.h"
10 #include "metricgroup.h"
11 #include <linux/zalloc.h>
12 
13 /*
14  * AGGR_GLOBAL: Use CPU 0
15  * AGGR_SOCKET: Use first CPU of socket
16  * AGGR_DIE: Use first CPU of die
17  * AGGR_CORE: Use first CPU of core
18  * AGGR_NONE: Use matching CPU
19  * AGGR_THREAD: Not supported?
20  */
21 static bool have_frontend_stalled;
22 
23 struct runtime_stat rt_stat;
24 struct stats walltime_nsecs_stats;
25 
26 struct saved_value {
27 	struct rb_node rb_node;
28 	struct evsel *evsel;
29 	enum stat_type type;
30 	int ctx;
31 	int cpu;
32 	struct runtime_stat *stat;
33 	struct stats stats;
34 	u64 metric_total;
35 	int metric_other;
36 };
37 
38 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
39 {
40 	struct saved_value *a = container_of(rb_node,
41 					     struct saved_value,
42 					     rb_node);
43 	const struct saved_value *b = entry;
44 
45 	if (a->cpu != b->cpu)
46 		return a->cpu - b->cpu;
47 
48 	/*
49 	 * Previously the rbtree was used to link generic metrics.
50 	 * The keys were evsel/cpu. Now the rbtree is extended to support
51 	 * per-thread shadow stats. For shadow stats case, the keys
52 	 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
53 	 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
54 	 */
55 	if (a->type != b->type)
56 		return a->type - b->type;
57 
58 	if (a->ctx != b->ctx)
59 		return a->ctx - b->ctx;
60 
61 	if (a->evsel == NULL && b->evsel == NULL) {
62 		if (a->stat == b->stat)
63 			return 0;
64 
65 		if ((char *)a->stat < (char *)b->stat)
66 			return -1;
67 
68 		return 1;
69 	}
70 
71 	if (a->evsel == b->evsel)
72 		return 0;
73 	if ((char *)a->evsel < (char *)b->evsel)
74 		return -1;
75 	return +1;
76 }
77 
78 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
79 				     const void *entry)
80 {
81 	struct saved_value *nd = malloc(sizeof(struct saved_value));
82 
83 	if (!nd)
84 		return NULL;
85 	memcpy(nd, entry, sizeof(struct saved_value));
86 	return &nd->rb_node;
87 }
88 
89 static void saved_value_delete(struct rblist *rblist __maybe_unused,
90 			       struct rb_node *rb_node)
91 {
92 	struct saved_value *v;
93 
94 	BUG_ON(!rb_node);
95 	v = container_of(rb_node, struct saved_value, rb_node);
96 	free(v);
97 }
98 
99 static struct saved_value *saved_value_lookup(struct evsel *evsel,
100 					      int cpu,
101 					      bool create,
102 					      enum stat_type type,
103 					      int ctx,
104 					      struct runtime_stat *st)
105 {
106 	struct rblist *rblist;
107 	struct rb_node *nd;
108 	struct saved_value dm = {
109 		.cpu = cpu,
110 		.evsel = evsel,
111 		.type = type,
112 		.ctx = ctx,
113 		.stat = st,
114 	};
115 
116 	rblist = &st->value_list;
117 
118 	nd = rblist__find(rblist, &dm);
119 	if (nd)
120 		return container_of(nd, struct saved_value, rb_node);
121 	if (create) {
122 		rblist__add_node(rblist, &dm);
123 		nd = rblist__find(rblist, &dm);
124 		if (nd)
125 			return container_of(nd, struct saved_value, rb_node);
126 	}
127 	return NULL;
128 }
129 
130 void runtime_stat__init(struct runtime_stat *st)
131 {
132 	struct rblist *rblist = &st->value_list;
133 
134 	rblist__init(rblist);
135 	rblist->node_cmp = saved_value_cmp;
136 	rblist->node_new = saved_value_new;
137 	rblist->node_delete = saved_value_delete;
138 }
139 
140 void runtime_stat__exit(struct runtime_stat *st)
141 {
142 	rblist__exit(&st->value_list);
143 }
144 
145 void perf_stat__init_shadow_stats(void)
146 {
147 	have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
148 	runtime_stat__init(&rt_stat);
149 }
150 
151 static int evsel_context(struct evsel *evsel)
152 {
153 	int ctx = 0;
154 
155 	if (evsel->core.attr.exclude_kernel)
156 		ctx |= CTX_BIT_KERNEL;
157 	if (evsel->core.attr.exclude_user)
158 		ctx |= CTX_BIT_USER;
159 	if (evsel->core.attr.exclude_hv)
160 		ctx |= CTX_BIT_HV;
161 	if (evsel->core.attr.exclude_host)
162 		ctx |= CTX_BIT_HOST;
163 	if (evsel->core.attr.exclude_idle)
164 		ctx |= CTX_BIT_IDLE;
165 
166 	return ctx;
167 }
168 
169 static void reset_stat(struct runtime_stat *st)
170 {
171 	struct rblist *rblist;
172 	struct rb_node *pos, *next;
173 
174 	rblist = &st->value_list;
175 	next = rb_first_cached(&rblist->entries);
176 	while (next) {
177 		pos = next;
178 		next = rb_next(pos);
179 		memset(&container_of(pos, struct saved_value, rb_node)->stats,
180 		       0,
181 		       sizeof(struct stats));
182 	}
183 }
184 
185 void perf_stat__reset_shadow_stats(void)
186 {
187 	reset_stat(&rt_stat);
188 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
189 }
190 
191 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
192 {
193 	reset_stat(st);
194 }
195 
196 static void update_runtime_stat(struct runtime_stat *st,
197 				enum stat_type type,
198 				int ctx, int cpu, u64 count)
199 {
200 	struct saved_value *v = saved_value_lookup(NULL, cpu, true,
201 						   type, ctx, st);
202 
203 	if (v)
204 		update_stats(&v->stats, count);
205 }
206 
207 /*
208  * Update various tracking values we maintain to print
209  * more semantic information such as miss/hit ratios,
210  * instruction rates, etc:
211  */
212 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
213 				    int cpu, struct runtime_stat *st)
214 {
215 	int ctx = evsel_context(counter);
216 	u64 count_ns = count;
217 	struct saved_value *v;
218 
219 	count *= counter->scale;
220 
221 	if (perf_evsel__is_clock(counter))
222 		update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns);
223 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
224 		update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count);
225 	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
226 		update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count);
227 	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
228 		update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count);
229 	else if (perf_stat_evsel__is(counter, ELISION_START))
230 		update_runtime_stat(st, STAT_ELISION, ctx, cpu, count);
231 	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
232 		update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
233 				    ctx, cpu, count);
234 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
235 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
236 				    ctx, cpu, count);
237 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
238 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
239 				    ctx, cpu, count);
240 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
241 		update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
242 				    ctx, cpu, count);
243 	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
244 		update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
245 				    ctx, cpu, count);
246 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
247 		update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
248 				    ctx, cpu, count);
249 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
250 		update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
251 				    ctx, cpu, count);
252 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
253 		update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
254 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
255 		update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
256 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
257 		update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
258 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
259 		update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
260 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
261 		update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
262 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
263 		update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
264 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
265 		update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count);
266 	else if (perf_stat_evsel__is(counter, SMI_NUM))
267 		update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count);
268 	else if (perf_stat_evsel__is(counter, APERF))
269 		update_runtime_stat(st, STAT_APERF, ctx, cpu, count);
270 
271 	if (counter->collect_stat) {
272 		v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st);
273 		update_stats(&v->stats, count);
274 		if (counter->metric_leader)
275 			v->metric_total += count;
276 	} else if (counter->metric_leader) {
277 		v = saved_value_lookup(counter->metric_leader,
278 				       cpu, true, STAT_NONE, 0, st);
279 		v->metric_total += count;
280 		v->metric_other++;
281 	}
282 }
283 
284 /* used for get_ratio_color() */
285 enum grc_type {
286 	GRC_STALLED_CYCLES_FE,
287 	GRC_STALLED_CYCLES_BE,
288 	GRC_CACHE_MISSES,
289 	GRC_MAX_NR
290 };
291 
292 static const char *get_ratio_color(enum grc_type type, double ratio)
293 {
294 	static const double grc_table[GRC_MAX_NR][3] = {
295 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
296 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
297 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
298 	};
299 	const char *color = PERF_COLOR_NORMAL;
300 
301 	if (ratio > grc_table[type][0])
302 		color = PERF_COLOR_RED;
303 	else if (ratio > grc_table[type][1])
304 		color = PERF_COLOR_MAGENTA;
305 	else if (ratio > grc_table[type][2])
306 		color = PERF_COLOR_YELLOW;
307 
308 	return color;
309 }
310 
311 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
312 						const char *name)
313 {
314 	struct evsel *c2;
315 
316 	evlist__for_each_entry (evsel_list, c2) {
317 		if (!strcasecmp(c2->name, name) && !c2->collect_stat)
318 			return c2;
319 	}
320 	return NULL;
321 }
322 
323 /* Mark MetricExpr target events and link events using them to them. */
324 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
325 {
326 	struct evsel *counter, *leader, **metric_events, *oc;
327 	bool found;
328 	const char **metric_names;
329 	int i;
330 	int num_metric_names;
331 
332 	evlist__for_each_entry(evsel_list, counter) {
333 		bool invalid = false;
334 
335 		leader = counter->leader;
336 		if (!counter->metric_expr)
337 			continue;
338 		metric_events = counter->metric_events;
339 		if (!metric_events) {
340 			if (expr__find_other(counter->metric_expr, counter->name,
341 						&metric_names, &num_metric_names) < 0)
342 				continue;
343 
344 			metric_events = calloc(sizeof(struct evsel *),
345 					       num_metric_names + 1);
346 			if (!metric_events)
347 				return;
348 			counter->metric_events = metric_events;
349 		}
350 
351 		for (i = 0; i < num_metric_names; i++) {
352 			found = false;
353 			if (leader) {
354 				/* Search in group */
355 				for_each_group_member (oc, leader) {
356 					if (!strcasecmp(oc->name, metric_names[i]) &&
357 						!oc->collect_stat) {
358 						found = true;
359 						break;
360 					}
361 				}
362 			}
363 			if (!found) {
364 				/* Search ignoring groups */
365 				oc = perf_stat__find_event(evsel_list, metric_names[i]);
366 			}
367 			if (!oc) {
368 				/* Deduping one is good enough to handle duplicated PMUs. */
369 				static char *printed;
370 
371 				/*
372 				 * Adding events automatically would be difficult, because
373 				 * it would risk creating groups that are not schedulable.
374 				 * perf stat doesn't understand all the scheduling constraints
375 				 * of events. So we ask the user instead to add the missing
376 				 * events.
377 				 */
378 				if (!printed || strcasecmp(printed, metric_names[i])) {
379 					fprintf(stderr,
380 						"Add %s event to groups to get metric expression for %s\n",
381 						metric_names[i],
382 						counter->name);
383 					printed = strdup(metric_names[i]);
384 				}
385 				invalid = true;
386 				continue;
387 			}
388 			metric_events[i] = oc;
389 			oc->collect_stat = true;
390 		}
391 		metric_events[i] = NULL;
392 		free(metric_names);
393 		if (invalid) {
394 			free(metric_events);
395 			counter->metric_events = NULL;
396 			counter->metric_expr = NULL;
397 		}
398 	}
399 }
400 
401 static double runtime_stat_avg(struct runtime_stat *st,
402 			       enum stat_type type, int ctx, int cpu)
403 {
404 	struct saved_value *v;
405 
406 	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
407 	if (!v)
408 		return 0.0;
409 
410 	return avg_stats(&v->stats);
411 }
412 
413 static double runtime_stat_n(struct runtime_stat *st,
414 			     enum stat_type type, int ctx, int cpu)
415 {
416 	struct saved_value *v;
417 
418 	v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
419 	if (!v)
420 		return 0.0;
421 
422 	return v->stats.n;
423 }
424 
425 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
426 					  int cpu,
427 					  struct evsel *evsel, double avg,
428 					  struct perf_stat_output_ctx *out,
429 					  struct runtime_stat *st)
430 {
431 	double total, ratio = 0.0;
432 	const char *color;
433 	int ctx = evsel_context(evsel);
434 
435 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
436 
437 	if (total)
438 		ratio = avg / total * 100.0;
439 
440 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
441 
442 	if (ratio)
443 		out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
444 				  ratio);
445 	else
446 		out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
447 }
448 
449 static void print_stalled_cycles_backend(struct perf_stat_config *config,
450 					 int cpu,
451 					 struct evsel *evsel, double avg,
452 					 struct perf_stat_output_ctx *out,
453 					 struct runtime_stat *st)
454 {
455 	double total, ratio = 0.0;
456 	const char *color;
457 	int ctx = evsel_context(evsel);
458 
459 	total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
460 
461 	if (total)
462 		ratio = avg / total * 100.0;
463 
464 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
465 
466 	out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
467 }
468 
469 static void print_branch_misses(struct perf_stat_config *config,
470 				int cpu,
471 				struct evsel *evsel,
472 				double avg,
473 				struct perf_stat_output_ctx *out,
474 				struct runtime_stat *st)
475 {
476 	double total, ratio = 0.0;
477 	const char *color;
478 	int ctx = evsel_context(evsel);
479 
480 	total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
481 
482 	if (total)
483 		ratio = avg / total * 100.0;
484 
485 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
486 
487 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
488 }
489 
490 static void print_l1_dcache_misses(struct perf_stat_config *config,
491 				   int cpu,
492 				   struct evsel *evsel,
493 				   double avg,
494 				   struct perf_stat_output_ctx *out,
495 				   struct runtime_stat *st)
496 
497 {
498 	double total, ratio = 0.0;
499 	const char *color;
500 	int ctx = evsel_context(evsel);
501 
502 	total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
503 
504 	if (total)
505 		ratio = avg / total * 100.0;
506 
507 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
508 
509 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
510 }
511 
512 static void print_l1_icache_misses(struct perf_stat_config *config,
513 				   int cpu,
514 				   struct evsel *evsel,
515 				   double avg,
516 				   struct perf_stat_output_ctx *out,
517 				   struct runtime_stat *st)
518 
519 {
520 	double total, ratio = 0.0;
521 	const char *color;
522 	int ctx = evsel_context(evsel);
523 
524 	total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
525 
526 	if (total)
527 		ratio = avg / total * 100.0;
528 
529 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
530 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
531 }
532 
533 static void print_dtlb_cache_misses(struct perf_stat_config *config,
534 				    int cpu,
535 				    struct evsel *evsel,
536 				    double avg,
537 				    struct perf_stat_output_ctx *out,
538 				    struct runtime_stat *st)
539 {
540 	double total, ratio = 0.0;
541 	const char *color;
542 	int ctx = evsel_context(evsel);
543 
544 	total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
545 
546 	if (total)
547 		ratio = avg / total * 100.0;
548 
549 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
550 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
551 }
552 
553 static void print_itlb_cache_misses(struct perf_stat_config *config,
554 				    int cpu,
555 				    struct evsel *evsel,
556 				    double avg,
557 				    struct perf_stat_output_ctx *out,
558 				    struct runtime_stat *st)
559 {
560 	double total, ratio = 0.0;
561 	const char *color;
562 	int ctx = evsel_context(evsel);
563 
564 	total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
565 
566 	if (total)
567 		ratio = avg / total * 100.0;
568 
569 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
570 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
571 }
572 
573 static void print_ll_cache_misses(struct perf_stat_config *config,
574 				  int cpu,
575 				  struct evsel *evsel,
576 				  double avg,
577 				  struct perf_stat_output_ctx *out,
578 				  struct runtime_stat *st)
579 {
580 	double total, ratio = 0.0;
581 	const char *color;
582 	int ctx = evsel_context(evsel);
583 
584 	total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
585 
586 	if (total)
587 		ratio = avg / total * 100.0;
588 
589 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
590 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
591 }
592 
593 /*
594  * High level "TopDown" CPU core pipe line bottleneck break down.
595  *
596  * Basic concept following
597  * Yasin, A Top Down Method for Performance analysis and Counter architecture
598  * ISPASS14
599  *
600  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
601  *
602  * Frontend -> Backend -> Retiring
603  * BadSpeculation in addition means out of order execution that is thrown away
604  * (for example branch mispredictions)
605  * Frontend is instruction decoding.
606  * Backend is execution, like computation and accessing data in memory
607  * Retiring is good execution that is not directly bottlenecked
608  *
609  * The formulas are computed in slots.
610  * A slot is an entry in the pipeline each for the pipeline width
611  * (for example a 4-wide pipeline has 4 slots for each cycle)
612  *
613  * Formulas:
614  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
615  *			TotalSlots
616  * Retiring = SlotsRetired / TotalSlots
617  * FrontendBound = FetchBubbles / TotalSlots
618  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
619  *
620  * The kernel provides the mapping to the low level CPU events and any scaling
621  * needed for the CPU pipeline width, for example:
622  *
623  * TotalSlots = Cycles * 4
624  *
625  * The scaling factor is communicated in the sysfs unit.
626  *
627  * In some cases the CPU may not be able to measure all the formulas due to
628  * missing events. In this case multiple formulas are combined, as possible.
629  *
630  * Full TopDown supports more levels to sub-divide each area: for example
631  * BackendBound into computing bound and memory bound. For now we only
632  * support Level 1 TopDown.
633  */
634 
635 static double sanitize_val(double x)
636 {
637 	if (x < 0 && x >= -0.02)
638 		return 0.0;
639 	return x;
640 }
641 
642 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
643 {
644 	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
645 }
646 
647 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
648 {
649 	double bad_spec = 0;
650 	double total_slots;
651 	double total;
652 
653 	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
654 		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
655 		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
656 
657 	total_slots = td_total_slots(ctx, cpu, st);
658 	if (total_slots)
659 		bad_spec = total / total_slots;
660 	return sanitize_val(bad_spec);
661 }
662 
663 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
664 {
665 	double retiring = 0;
666 	double total_slots = td_total_slots(ctx, cpu, st);
667 	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
668 					    ctx, cpu);
669 
670 	if (total_slots)
671 		retiring = ret_slots / total_slots;
672 	return retiring;
673 }
674 
675 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
676 {
677 	double fe_bound = 0;
678 	double total_slots = td_total_slots(ctx, cpu, st);
679 	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
680 					    ctx, cpu);
681 
682 	if (total_slots)
683 		fe_bound = fetch_bub / total_slots;
684 	return fe_bound;
685 }
686 
687 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
688 {
689 	double sum = (td_fe_bound(ctx, cpu, st) +
690 		      td_bad_spec(ctx, cpu, st) +
691 		      td_retiring(ctx, cpu, st));
692 	if (sum == 0)
693 		return 0;
694 	return sanitize_val(1.0 - sum);
695 }
696 
697 static void print_smi_cost(struct perf_stat_config *config,
698 			   int cpu, struct evsel *evsel,
699 			   struct perf_stat_output_ctx *out,
700 			   struct runtime_stat *st)
701 {
702 	double smi_num, aperf, cycles, cost = 0.0;
703 	int ctx = evsel_context(evsel);
704 	const char *color = NULL;
705 
706 	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
707 	aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
708 	cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
709 
710 	if ((cycles == 0) || (aperf == 0))
711 		return;
712 
713 	if (smi_num)
714 		cost = (aperf - cycles) / aperf * 100.00;
715 
716 	if (cost > 10)
717 		color = PERF_COLOR_RED;
718 	out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
719 	out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
720 }
721 
722 static void generic_metric(struct perf_stat_config *config,
723 			   const char *metric_expr,
724 			   struct evsel **metric_events,
725 			   char *name,
726 			   const char *metric_name,
727 			   const char *metric_unit,
728 			   double avg,
729 			   int cpu,
730 			   struct perf_stat_output_ctx *out,
731 			   struct runtime_stat *st)
732 {
733 	print_metric_t print_metric = out->print_metric;
734 	struct parse_ctx pctx;
735 	double ratio, scale;
736 	int i;
737 	void *ctxp = out->ctx;
738 	char *n, *pn;
739 
740 	expr__ctx_init(&pctx);
741 	/* Must be first id entry */
742 	expr__add_id(&pctx, name, avg);
743 	for (i = 0; metric_events[i]; i++) {
744 		struct saved_value *v;
745 		struct stats *stats;
746 		u64 metric_total = 0;
747 
748 		if (!strcmp(metric_events[i]->name, "duration_time")) {
749 			stats = &walltime_nsecs_stats;
750 			scale = 1e-9;
751 		} else {
752 			v = saved_value_lookup(metric_events[i], cpu, false,
753 					       STAT_NONE, 0, st);
754 			if (!v)
755 				break;
756 			stats = &v->stats;
757 			scale = 1.0;
758 
759 			if (v->metric_other)
760 				metric_total = v->metric_total;
761 		}
762 
763 		n = strdup(metric_events[i]->name);
764 		if (!n)
765 			return;
766 		/*
767 		 * This display code with --no-merge adds [cpu] postfixes.
768 		 * These are not supported by the parser. Remove everything
769 		 * after the space.
770 		 */
771 		pn = strchr(n, ' ');
772 		if (pn)
773 			*pn = 0;
774 
775 		if (metric_total)
776 			expr__add_id(&pctx, n, metric_total);
777 		else
778 			expr__add_id(&pctx, n, avg_stats(stats)*scale);
779 	}
780 
781 	if (!metric_events[i]) {
782 		const char *p = metric_expr;
783 
784 		if (expr__parse(&ratio, &pctx, &p) == 0) {
785 			char *unit;
786 			char metric_bf[64];
787 
788 			if (metric_unit && metric_name) {
789 				if (perf_pmu__convert_scale(metric_unit,
790 					&unit, &scale) >= 0) {
791 					ratio *= scale;
792 				}
793 
794 				scnprintf(metric_bf, sizeof(metric_bf),
795 					  "%s  %s", unit, metric_name);
796 				print_metric(config, ctxp, NULL, "%8.1f",
797 					     metric_bf, ratio);
798 			} else {
799 				print_metric(config, ctxp, NULL, "%8.1f",
800 					metric_name ?
801 					metric_name :
802 					out->force_header ?  name : "",
803 					ratio);
804 			}
805 		} else {
806 			print_metric(config, ctxp, NULL, NULL,
807 				     out->force_header ?
808 				     (metric_name ? metric_name : name) : "", 0);
809 		}
810 	} else
811 		print_metric(config, ctxp, NULL, NULL, "", 0);
812 
813 	for (i = 1; i < pctx.num_ids; i++)
814 		zfree(&pctx.ids[i].name);
815 }
816 
817 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
818 				   struct evsel *evsel,
819 				   double avg, int cpu,
820 				   struct perf_stat_output_ctx *out,
821 				   struct rblist *metric_events,
822 				   struct runtime_stat *st)
823 {
824 	void *ctxp = out->ctx;
825 	print_metric_t print_metric = out->print_metric;
826 	double total, ratio = 0.0, total2;
827 	const char *color = NULL;
828 	int ctx = evsel_context(evsel);
829 	struct metric_event *me;
830 	int num = 1;
831 
832 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
833 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
834 
835 		if (total) {
836 			ratio = avg / total;
837 			print_metric(config, ctxp, NULL, "%7.2f ",
838 					"insn per cycle", ratio);
839 		} else {
840 			print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
841 		}
842 
843 		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
844 					 ctx, cpu);
845 
846 		total = max(total, runtime_stat_avg(st,
847 						    STAT_STALLED_CYCLES_BACK,
848 						    ctx, cpu));
849 
850 		if (total && avg) {
851 			out->new_line(config, ctxp);
852 			ratio = total / avg;
853 			print_metric(config, ctxp, NULL, "%7.2f ",
854 					"stalled cycles per insn",
855 					ratio);
856 		} else if (have_frontend_stalled) {
857 			out->new_line(config, ctxp);
858 			print_metric(config, ctxp, NULL, "%7.2f ",
859 				     "stalled cycles per insn", 0);
860 		}
861 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
862 		if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
863 			print_branch_misses(config, cpu, evsel, avg, out, st);
864 		else
865 			print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
866 	} else if (
867 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
868 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
869 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
870 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
871 
872 		if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
873 			print_l1_dcache_misses(config, cpu, evsel, avg, out, st);
874 		else
875 			print_metric(config, ctxp, NULL, NULL, "of all L1-dcache hits", 0);
876 	} else if (
877 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
878 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
879 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
880 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
881 
882 		if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
883 			print_l1_icache_misses(config, cpu, evsel, avg, out, st);
884 		else
885 			print_metric(config, ctxp, NULL, NULL, "of all L1-icache hits", 0);
886 	} else if (
887 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
888 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
889 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
890 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
891 
892 		if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
893 			print_dtlb_cache_misses(config, cpu, evsel, avg, out, st);
894 		else
895 			print_metric(config, ctxp, NULL, NULL, "of all dTLB cache hits", 0);
896 	} else if (
897 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
898 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
899 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
900 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
901 
902 		if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
903 			print_itlb_cache_misses(config, cpu, evsel, avg, out, st);
904 		else
905 			print_metric(config, ctxp, NULL, NULL, "of all iTLB cache hits", 0);
906 	} else if (
907 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
908 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
909 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
910 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
911 
912 		if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
913 			print_ll_cache_misses(config, cpu, evsel, avg, out, st);
914 		else
915 			print_metric(config, ctxp, NULL, NULL, "of all LL-cache hits", 0);
916 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
917 		total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
918 
919 		if (total)
920 			ratio = avg * 100 / total;
921 
922 		if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
923 			print_metric(config, ctxp, NULL, "%8.3f %%",
924 				     "of all cache refs", ratio);
925 		else
926 			print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
927 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
928 		print_stalled_cycles_frontend(config, cpu, evsel, avg, out, st);
929 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
930 		print_stalled_cycles_backend(config, cpu, evsel, avg, out, st);
931 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
932 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
933 
934 		if (total) {
935 			ratio = avg / total;
936 			print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
937 		} else {
938 			print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
939 		}
940 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
941 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
942 
943 		if (total)
944 			print_metric(config, ctxp, NULL,
945 					"%7.2f%%", "transactional cycles",
946 					100.0 * (avg / total));
947 		else
948 			print_metric(config, ctxp, NULL, NULL, "transactional cycles",
949 				     0);
950 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
951 		total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
952 		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
953 
954 		if (total2 < avg)
955 			total2 = avg;
956 		if (total)
957 			print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
958 				100.0 * ((total2-avg) / total));
959 		else
960 			print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
961 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
962 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
963 					 ctx, cpu);
964 
965 		if (avg)
966 			ratio = total / avg;
967 
968 		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
969 			print_metric(config, ctxp, NULL, "%8.0f",
970 				     "cycles / transaction", ratio);
971 		else
972 			print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
973 				      0);
974 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
975 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
976 					 ctx, cpu);
977 
978 		if (avg)
979 			ratio = total / avg;
980 
981 		print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
982 	} else if (perf_evsel__is_clock(evsel)) {
983 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
984 			print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
985 				     avg / (ratio * evsel->scale));
986 		else
987 			print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
988 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
989 		double fe_bound = td_fe_bound(ctx, cpu, st);
990 
991 		if (fe_bound > 0.2)
992 			color = PERF_COLOR_RED;
993 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
994 				fe_bound * 100.);
995 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
996 		double retiring = td_retiring(ctx, cpu, st);
997 
998 		if (retiring > 0.7)
999 			color = PERF_COLOR_GREEN;
1000 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1001 				retiring * 100.);
1002 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1003 		double bad_spec = td_bad_spec(ctx, cpu, st);
1004 
1005 		if (bad_spec > 0.1)
1006 			color = PERF_COLOR_RED;
1007 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1008 				bad_spec * 100.);
1009 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1010 		double be_bound = td_be_bound(ctx, cpu, st);
1011 		const char *name = "backend bound";
1012 		static int have_recovery_bubbles = -1;
1013 
1014 		/* In case the CPU does not support topdown-recovery-bubbles */
1015 		if (have_recovery_bubbles < 0)
1016 			have_recovery_bubbles = pmu_have_event("cpu",
1017 					"topdown-recovery-bubbles");
1018 		if (!have_recovery_bubbles)
1019 			name = "backend bound/bad spec";
1020 
1021 		if (be_bound > 0.2)
1022 			color = PERF_COLOR_RED;
1023 		if (td_total_slots(ctx, cpu, st) > 0)
1024 			print_metric(config, ctxp, color, "%8.1f%%", name,
1025 					be_bound * 100.);
1026 		else
1027 			print_metric(config, ctxp, NULL, NULL, name, 0);
1028 	} else if (evsel->metric_expr) {
1029 		generic_metric(config, evsel->metric_expr, evsel->metric_events, evsel->name,
1030 				evsel->metric_name, NULL, avg, cpu, out, st);
1031 	} else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
1032 		char unit = 'M';
1033 		char unit_buf[10];
1034 
1035 		total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
1036 
1037 		if (total)
1038 			ratio = 1000.0 * avg / total;
1039 		if (ratio < 0.001) {
1040 			ratio *= 1000;
1041 			unit = 'K';
1042 		}
1043 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1044 		print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1045 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1046 		print_smi_cost(config, cpu, evsel, out, st);
1047 	} else {
1048 		num = 0;
1049 	}
1050 
1051 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1052 		struct metric_expr *mexp;
1053 
1054 		list_for_each_entry (mexp, &me->head, nd) {
1055 			if (num++ > 0)
1056 				out->new_line(config, ctxp);
1057 			generic_metric(config, mexp->metric_expr, mexp->metric_events,
1058 					evsel->name, mexp->metric_name,
1059 					mexp->metric_unit, avg, cpu, out, st);
1060 		}
1061 	}
1062 	if (num == 0)
1063 		print_metric(config, ctxp, NULL, NULL, NULL, 0);
1064 }
1065